^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1) // SPDX-License-Identifier: GPL-2.0
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3) * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5) * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7) * Interactivity improvements by Mike Galbraith
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8) * (C) 2007 Mike Galbraith <efault@gmx.de>
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10) * Various enhancements by Dmitry Adamushko.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11) * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 12) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 13) * Group scheduling enhancements by Srivatsa Vaddagiri
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 14) * Copyright IBM Corporation, 2007
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 15) * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 16) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 17) * Scaled math optimizations by Thomas Gleixner
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 18) * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 19) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 20) * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 21) * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 22) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 23) #include "sched.h"
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 24)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 25) #include <trace/hooks/sched.h>
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 26)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 27) EXPORT_TRACEPOINT_SYMBOL_GPL(sched_stat_runtime);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 28)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 29) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 30) * Targeted preemption latency for CPU-bound tasks:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 31) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 32) * NOTE: this latency value is not the same as the concept of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 33) * 'timeslice length' - timeslices in CFS are of variable length
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 34) * and have no persistent notion like in traditional, time-slice
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 35) * based scheduling concepts.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 36) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 37) * (to see the precise effective timeslice length of your workload,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 38) * run vmstat and monitor the context-switches (cs) field)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 39) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 40) * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 41) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 42) unsigned int sysctl_sched_latency = 6000000ULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 43) EXPORT_SYMBOL_GPL(sysctl_sched_latency);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 44) static unsigned int normalized_sysctl_sched_latency = 6000000ULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 45)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 46) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 47) * The initial- and re-scaling of tunables is configurable
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 48) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 49) * Options are:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 50) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 51) * SCHED_TUNABLESCALING_NONE - unscaled, always *1
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 52) * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 53) * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 54) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 55) * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 56) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 57) enum sched_tunable_scaling sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 58)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 59) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 60) * Minimal preemption granularity for CPU-bound tasks:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 61) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 62) * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 63) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 64) unsigned int sysctl_sched_min_granularity = 750000ULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 65) EXPORT_SYMBOL_GPL(sysctl_sched_min_granularity);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 66) static unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 67)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 68) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 69) * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 70) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 71) static unsigned int sched_nr_latency = 8;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 72)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 73) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 74) * After fork, child runs first. If set to 0 (default) then
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 75) * parent will (try to) run first.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 76) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 77) unsigned int sysctl_sched_child_runs_first __read_mostly;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 78)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 79) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 80) * SCHED_OTHER wake-up granularity.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 81) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 82) * This option delays the preemption effects of decoupled workloads
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 83) * and reduces their over-scheduling. Synchronous workloads will still
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 84) * have immediate wakeup/sleep latencies.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 85) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 86) * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 87) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 88) unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 89) static unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 90)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 91) const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 92)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 93) int sched_thermal_decay_shift;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 94) static int __init setup_sched_thermal_decay_shift(char *str)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 95) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 96) int _shift = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 97)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 98) if (kstrtoint(str, 0, &_shift))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 99) pr_warn("Unable to set scheduler thermal pressure decay shift parameter\n");
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 100)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 101) sched_thermal_decay_shift = clamp(_shift, 0, 10);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 102) return 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 103) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 104) __setup("sched_thermal_decay_shift=", setup_sched_thermal_decay_shift);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 105)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 106) #ifdef CONFIG_SMP
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 107) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 108) * For asym packing, by default the lower numbered CPU has higher priority.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 109) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 110) int __weak arch_asym_cpu_priority(int cpu)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 111) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 112) return -cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 113) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 114)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 115) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 116) * The margin used when comparing utilization with CPU capacity.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 117) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 118) * (default: ~20%)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 119) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 120) #define fits_capacity(cap, max) ((cap) * 1280 < (max) * 1024)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 121)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 122) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 123)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 124) #ifdef CONFIG_CFS_BANDWIDTH
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 125) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 126) * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 127) * each time a cfs_rq requests quota.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 128) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 129) * Note: in the case that the slice exceeds the runtime remaining (either due
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 130) * to consumption or the quota being specified to be smaller than the slice)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 131) * we will always only issue the remaining available time.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 132) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 133) * (default: 5 msec, units: microseconds)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 134) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 135) unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 136) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 137)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 138) static inline void update_load_add(struct load_weight *lw, unsigned long inc)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 139) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 140) lw->weight += inc;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 141) lw->inv_weight = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 142) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 143)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 144) static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 145) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 146) lw->weight -= dec;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 147) lw->inv_weight = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 148) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 149)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 150) static inline void update_load_set(struct load_weight *lw, unsigned long w)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 151) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 152) lw->weight = w;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 153) lw->inv_weight = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 154) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 155)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 156) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 157) * Increase the granularity value when there are more CPUs,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 158) * because with more CPUs the 'effective latency' as visible
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 159) * to users decreases. But the relationship is not linear,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 160) * so pick a second-best guess by going with the log2 of the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 161) * number of CPUs.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 162) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 163) * This idea comes from the SD scheduler of Con Kolivas:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 164) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 165) static unsigned int get_update_sysctl_factor(void)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 166) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 167) unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 168) unsigned int factor;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 169)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 170) switch (sysctl_sched_tunable_scaling) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 171) case SCHED_TUNABLESCALING_NONE:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 172) factor = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 173) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 174) case SCHED_TUNABLESCALING_LINEAR:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 175) factor = cpus;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 176) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 177) case SCHED_TUNABLESCALING_LOG:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 178) default:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 179) factor = 1 + ilog2(cpus);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 180) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 181) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 182)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 183) return factor;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 184) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 185)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 186) static void update_sysctl(void)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 187) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 188) unsigned int factor = get_update_sysctl_factor();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 189)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 190) #define SET_SYSCTL(name) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 191) (sysctl_##name = (factor) * normalized_sysctl_##name)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 192) SET_SYSCTL(sched_min_granularity);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 193) SET_SYSCTL(sched_latency);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 194) SET_SYSCTL(sched_wakeup_granularity);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 195) #undef SET_SYSCTL
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 196) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 197)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 198) void __init sched_init_granularity(void)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 199) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 200) update_sysctl();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 201) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 202)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 203) #define WMULT_CONST (~0U)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 204) #define WMULT_SHIFT 32
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 205)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 206) static void __update_inv_weight(struct load_weight *lw)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 207) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 208) unsigned long w;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 209)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 210) if (likely(lw->inv_weight))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 211) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 212)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 213) w = scale_load_down(lw->weight);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 214)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 215) if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 216) lw->inv_weight = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 217) else if (unlikely(!w))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 218) lw->inv_weight = WMULT_CONST;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 219) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 220) lw->inv_weight = WMULT_CONST / w;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 221) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 222)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 223) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 224) * delta_exec * weight / lw.weight
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 225) * OR
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 226) * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 227) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 228) * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 229) * we're guaranteed shift stays positive because inv_weight is guaranteed to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 230) * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 231) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 232) * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 233) * weight/lw.weight <= 1, and therefore our shift will also be positive.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 234) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 235) static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 236) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 237) u64 fact = scale_load_down(weight);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 238) int shift = WMULT_SHIFT;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 239)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 240) __update_inv_weight(lw);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 241)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 242) if (unlikely(fact >> 32)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 243) while (fact >> 32) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 244) fact >>= 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 245) shift--;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 246) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 247) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 248)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 249) fact = mul_u32_u32(fact, lw->inv_weight);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 250)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 251) while (fact >> 32) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 252) fact >>= 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 253) shift--;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 254) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 255)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 256) return mul_u64_u32_shr(delta_exec, fact, shift);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 257) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 258)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 259)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 260) const struct sched_class fair_sched_class;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 261)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 262) /**************************************************************
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 263) * CFS operations on generic schedulable entities:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 264) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 265)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 266) #ifdef CONFIG_FAIR_GROUP_SCHED
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 267) static inline struct task_struct *task_of(struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 268) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 269) SCHED_WARN_ON(!entity_is_task(se));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 270) return container_of(se, struct task_struct, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 271) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 272)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 273) /* Walk up scheduling entities hierarchy */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 274) #define for_each_sched_entity(se) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 275) for (; se; se = se->parent)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 276)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 277) static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 278) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 279) return p->se.cfs_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 280) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 281)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 282) /* runqueue on which this entity is (to be) queued */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 283) static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 284) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 285) return se->cfs_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 286) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 287)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 288) /* runqueue "owned" by this group */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 289) static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 290) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 291) return grp->my_q;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 292) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 293)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 294) static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 295) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 296) if (!path)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 297) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 298)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 299) if (cfs_rq && task_group_is_autogroup(cfs_rq->tg))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 300) autogroup_path(cfs_rq->tg, path, len);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 301) else if (cfs_rq && cfs_rq->tg->css.cgroup)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 302) cgroup_path(cfs_rq->tg->css.cgroup, path, len);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 303) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 304) strlcpy(path, "(null)", len);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 305) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 306)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 307) static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 308) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 309) struct rq *rq = rq_of(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 310) int cpu = cpu_of(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 311)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 312) if (cfs_rq->on_list)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 313) return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 314)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 315) cfs_rq->on_list = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 316)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 317) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 318) * Ensure we either appear before our parent (if already
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 319) * enqueued) or force our parent to appear after us when it is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 320) * enqueued. The fact that we always enqueue bottom-up
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 321) * reduces this to two cases and a special case for the root
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 322) * cfs_rq. Furthermore, it also means that we will always reset
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 323) * tmp_alone_branch either when the branch is connected
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 324) * to a tree or when we reach the top of the tree
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 325) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 326) if (cfs_rq->tg->parent &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 327) cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 328) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 329) * If parent is already on the list, we add the child
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 330) * just before. Thanks to circular linked property of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 331) * the list, this means to put the child at the tail
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 332) * of the list that starts by parent.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 333) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 334) list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 335) &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 336) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 337) * The branch is now connected to its tree so we can
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 338) * reset tmp_alone_branch to the beginning of the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 339) * list.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 340) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 341) rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 342) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 343) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 344)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 345) if (!cfs_rq->tg->parent) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 346) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 347) * cfs rq without parent should be put
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 348) * at the tail of the list.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 349) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 350) list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 351) &rq->leaf_cfs_rq_list);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 352) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 353) * We have reach the top of a tree so we can reset
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 354) * tmp_alone_branch to the beginning of the list.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 355) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 356) rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 357) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 358) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 359)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 360) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 361) * The parent has not already been added so we want to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 362) * make sure that it will be put after us.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 363) * tmp_alone_branch points to the begin of the branch
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 364) * where we will add parent.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 365) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 366) list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 367) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 368) * update tmp_alone_branch to points to the new begin
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 369) * of the branch
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 370) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 371) rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 372) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 373) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 374)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 375) static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 376) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 377) if (cfs_rq->on_list) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 378) struct rq *rq = rq_of(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 379)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 380) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 381) * With cfs_rq being unthrottled/throttled during an enqueue,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 382) * it can happen the tmp_alone_branch points the a leaf that
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 383) * we finally want to del. In this case, tmp_alone_branch moves
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 384) * to the prev element but it will point to rq->leaf_cfs_rq_list
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 385) * at the end of the enqueue.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 386) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 387) if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 388) rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 389)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 390) list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 391) cfs_rq->on_list = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 392) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 393) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 394)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 395) static inline void assert_list_leaf_cfs_rq(struct rq *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 396) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 397) SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 398) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 399)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 400) /* Iterate thr' all leaf cfs_rq's on a runqueue */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 401) #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 402) list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 403) leaf_cfs_rq_list)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 404)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 405) /* Do the two (enqueued) entities belong to the same group ? */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 406) static inline struct cfs_rq *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 407) is_same_group(struct sched_entity *se, struct sched_entity *pse)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 408) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 409) if (se->cfs_rq == pse->cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 410) return se->cfs_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 411)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 412) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 413) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 414)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 415) static inline struct sched_entity *parent_entity(struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 416) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 417) return se->parent;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 418) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 419)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 420) static void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 421) find_matching_se(struct sched_entity **se, struct sched_entity **pse)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 422) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 423) int se_depth, pse_depth;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 424)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 425) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 426) * preemption test can be made between sibling entities who are in the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 427) * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 428) * both tasks until we find their ancestors who are siblings of common
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 429) * parent.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 430) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 431)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 432) /* First walk up until both entities are at same depth */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 433) se_depth = (*se)->depth;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 434) pse_depth = (*pse)->depth;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 435)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 436) while (se_depth > pse_depth) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 437) se_depth--;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 438) *se = parent_entity(*se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 439) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 440)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 441) while (pse_depth > se_depth) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 442) pse_depth--;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 443) *pse = parent_entity(*pse);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 444) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 445)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 446) while (!is_same_group(*se, *pse)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 447) *se = parent_entity(*se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 448) *pse = parent_entity(*pse);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 449) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 450) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 451)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 452) #else /* !CONFIG_FAIR_GROUP_SCHED */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 453)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 454) static inline struct task_struct *task_of(struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 455) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 456) return container_of(se, struct task_struct, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 457) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 458)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 459) #define for_each_sched_entity(se) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 460) for (; se; se = NULL)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 461)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 462) static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 463) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 464) return &task_rq(p)->cfs;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 465) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 466)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 467) static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 468) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 469) struct task_struct *p = task_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 470) struct rq *rq = task_rq(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 471)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 472) return &rq->cfs;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 473) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 474)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 475) /* runqueue "owned" by this group */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 476) static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 477) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 478) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 479) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 480)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 481) static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 482) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 483) if (path)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 484) strlcpy(path, "(null)", len);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 485) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 486)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 487) static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 488) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 489) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 490) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 491)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 492) static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 493) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 494) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 495)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 496) static inline void assert_list_leaf_cfs_rq(struct rq *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 497) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 498) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 499)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 500) #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 501) for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 502)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 503) static inline struct sched_entity *parent_entity(struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 504) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 505) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 506) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 507)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 508) static inline void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 509) find_matching_se(struct sched_entity **se, struct sched_entity **pse)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 510) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 511) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 512)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 513) #endif /* CONFIG_FAIR_GROUP_SCHED */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 514)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 515) static __always_inline
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 516) void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 517)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 518) /**************************************************************
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 519) * Scheduling class tree data structure manipulation methods:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 520) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 521)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 522) static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 523) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 524) s64 delta = (s64)(vruntime - max_vruntime);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 525) if (delta > 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 526) max_vruntime = vruntime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 527)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 528) return max_vruntime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 529) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 530)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 531) static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 532) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 533) s64 delta = (s64)(vruntime - min_vruntime);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 534) if (delta < 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 535) min_vruntime = vruntime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 536)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 537) return min_vruntime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 538) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 539)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 540) static inline int entity_before(struct sched_entity *a,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 541) struct sched_entity *b)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 542) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 543) return (s64)(a->vruntime - b->vruntime) < 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 544) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 545)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 546) static void update_min_vruntime(struct cfs_rq *cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 547) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 548) struct sched_entity *curr = cfs_rq->curr;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 549) struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 550)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 551) u64 vruntime = cfs_rq->min_vruntime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 552)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 553) if (curr) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 554) if (curr->on_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 555) vruntime = curr->vruntime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 556) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 557) curr = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 558) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 559)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 560) if (leftmost) { /* non-empty tree */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 561) struct sched_entity *se;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 562) se = rb_entry(leftmost, struct sched_entity, run_node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 563)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 564) if (!curr)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 565) vruntime = se->vruntime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 566) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 567) vruntime = min_vruntime(vruntime, se->vruntime);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 568) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 569)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 570) /* ensure we never gain time by being placed backwards. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 571) cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 572) #ifndef CONFIG_64BIT
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 573) smp_wmb();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 574) cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 575) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 576) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 577)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 578) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 579) * Enqueue an entity into the rb-tree:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 580) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 581) static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 582) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 583) struct rb_node **link = &cfs_rq->tasks_timeline.rb_root.rb_node;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 584) struct rb_node *parent = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 585) struct sched_entity *entry;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 586) bool leftmost = true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 587)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 588) trace_android_rvh_enqueue_entity(cfs_rq, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 589) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 590) * Find the right place in the rbtree:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 591) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 592) while (*link) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 593) parent = *link;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 594) entry = rb_entry(parent, struct sched_entity, run_node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 595) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 596) * We dont care about collisions. Nodes with
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 597) * the same key stay together.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 598) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 599) if (entity_before(se, entry)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 600) link = &parent->rb_left;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 601) } else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 602) link = &parent->rb_right;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 603) leftmost = false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 604) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 605) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 606)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 607) rb_link_node(&se->run_node, parent, link);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 608) rb_insert_color_cached(&se->run_node,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 609) &cfs_rq->tasks_timeline, leftmost);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 610) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 611)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 612) static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 613) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 614) trace_android_rvh_dequeue_entity(cfs_rq, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 615) rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 616) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 617)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 618) struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 619) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 620) struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 621)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 622) if (!left)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 623) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 624)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 625) return rb_entry(left, struct sched_entity, run_node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 626) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 627)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 628) static struct sched_entity *__pick_next_entity(struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 629) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 630) struct rb_node *next = rb_next(&se->run_node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 631)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 632) if (!next)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 633) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 634)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 635) return rb_entry(next, struct sched_entity, run_node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 636) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 637)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 638) #ifdef CONFIG_SCHED_DEBUG
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 639) struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 640) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 641) struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 642)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 643) if (!last)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 644) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 645)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 646) return rb_entry(last, struct sched_entity, run_node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 647) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 648)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 649) /**************************************************************
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 650) * Scheduling class statistics methods:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 651) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 652)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 653) int sched_proc_update_handler(struct ctl_table *table, int write,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 654) void *buffer, size_t *lenp, loff_t *ppos)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 655) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 656) int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 657) unsigned int factor = get_update_sysctl_factor();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 658)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 659) if (ret || !write)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 660) return ret;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 661)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 662) sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 663) sysctl_sched_min_granularity);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 664)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 665) #define WRT_SYSCTL(name) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 666) (normalized_sysctl_##name = sysctl_##name / (factor))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 667) WRT_SYSCTL(sched_min_granularity);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 668) WRT_SYSCTL(sched_latency);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 669) WRT_SYSCTL(sched_wakeup_granularity);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 670) #undef WRT_SYSCTL
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 671)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 672) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 673) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 674) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 675)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 676) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 677) * delta /= w
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 678) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 679) static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 680) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 681) if (unlikely(se->load.weight != NICE_0_LOAD))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 682) delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 683)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 684) return delta;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 685) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 686)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 687) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 688) * The idea is to set a period in which each task runs once.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 689) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 690) * When there are too many tasks (sched_nr_latency) we have to stretch
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 691) * this period because otherwise the slices get too small.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 692) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 693) * p = (nr <= nl) ? l : l*nr/nl
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 694) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 695) static u64 __sched_period(unsigned long nr_running)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 696) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 697) if (unlikely(nr_running > sched_nr_latency))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 698) return nr_running * sysctl_sched_min_granularity;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 699) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 700) return sysctl_sched_latency;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 701) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 702)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 703) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 704) * We calculate the wall-time slice from the period by taking a part
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 705) * proportional to the weight.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 706) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 707) * s = p*P[w/rw]
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 708) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 709) static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 710) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 711) unsigned int nr_running = cfs_rq->nr_running;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 712) u64 slice;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 713)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 714) if (sched_feat(ALT_PERIOD))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 715) nr_running = rq_of(cfs_rq)->cfs.h_nr_running;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 716)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 717) slice = __sched_period(nr_running + !se->on_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 718)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 719) for_each_sched_entity(se) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 720) struct load_weight *load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 721) struct load_weight lw;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 722)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 723) cfs_rq = cfs_rq_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 724) load = &cfs_rq->load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 725)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 726) if (unlikely(!se->on_rq)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 727) lw = cfs_rq->load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 728)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 729) update_load_add(&lw, se->load.weight);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 730) load = &lw;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 731) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 732) slice = __calc_delta(slice, se->load.weight, load);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 733) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 734)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 735) if (sched_feat(BASE_SLICE))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 736) slice = max(slice, (u64)sysctl_sched_min_granularity);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 737)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 738) return slice;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 739) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 740)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 741) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 742) * We calculate the vruntime slice of a to-be-inserted task.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 743) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 744) * vs = s/w
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 745) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 746) static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 747) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 748) return calc_delta_fair(sched_slice(cfs_rq, se), se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 749) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 750)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 751) #include "pelt.h"
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 752) #ifdef CONFIG_SMP
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 753)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 754) static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 755) static unsigned long task_h_load(struct task_struct *p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 756) static unsigned long capacity_of(int cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 757)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 758) /* Give new sched_entity start runnable values to heavy its load in infant time */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 759) void init_entity_runnable_average(struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 760) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 761) struct sched_avg *sa = &se->avg;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 762)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 763) memset(sa, 0, sizeof(*sa));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 764)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 765) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 766) * Tasks are initialized with full load to be seen as heavy tasks until
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 767) * they get a chance to stabilize to their real load level.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 768) * Group entities are initialized with zero load to reflect the fact that
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 769) * nothing has been attached to the task group yet.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 770) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 771) if (entity_is_task(se))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 772) sa->load_avg = scale_load_down(se->load.weight);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 773)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 774) /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 775) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 776)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 777) static void attach_entity_cfs_rq(struct sched_entity *se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 778)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 779) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 780) * With new tasks being created, their initial util_avgs are extrapolated
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 781) * based on the cfs_rq's current util_avg:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 782) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 783) * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 784) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 785) * However, in many cases, the above util_avg does not give a desired
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 786) * value. Moreover, the sum of the util_avgs may be divergent, such
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 787) * as when the series is a harmonic series.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 788) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 789) * To solve this problem, we also cap the util_avg of successive tasks to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 790) * only 1/2 of the left utilization budget:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 791) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 792) * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 793) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 794) * where n denotes the nth task and cpu_scale the CPU capacity.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 795) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 796) * For example, for a CPU with 1024 of capacity, a simplest series from
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 797) * the beginning would be like:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 798) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 799) * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 800) * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 801) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 802) * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 803) * if util_avg > util_avg_cap.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 804) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 805) void post_init_entity_util_avg(struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 806) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 807) struct sched_entity *se = &p->se;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 808) struct cfs_rq *cfs_rq = cfs_rq_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 809) struct sched_avg *sa = &se->avg;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 810) long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq)));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 811) long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 812)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 813) if (cap > 0) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 814) if (cfs_rq->avg.util_avg != 0) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 815) sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 816) sa->util_avg /= (cfs_rq->avg.load_avg + 1);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 817)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 818) if (sa->util_avg > cap)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 819) sa->util_avg = cap;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 820) } else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 821) sa->util_avg = cap;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 822) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 823) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 824)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 825) sa->runnable_avg = sa->util_avg;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 826)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 827) if (p->sched_class != &fair_sched_class) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 828) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 829) * For !fair tasks do:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 830) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 831) update_cfs_rq_load_avg(now, cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 832) attach_entity_load_avg(cfs_rq, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 833) switched_from_fair(rq, p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 834) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 835) * such that the next switched_to_fair() has the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 836) * expected state.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 837) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 838) se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 839) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 840) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 841)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 842) /* Hook before this se's util is attached to cfs_rq's util */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 843) trace_android_rvh_post_init_entity_util_avg(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 844) attach_entity_cfs_rq(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 845) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 846)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 847) #else /* !CONFIG_SMP */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 848) void init_entity_runnable_average(struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 849) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 850) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 851) void post_init_entity_util_avg(struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 852) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 853) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 854) static void update_tg_load_avg(struct cfs_rq *cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 855) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 856) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 857) #endif /* CONFIG_SMP */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 858)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 859) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 860) * Update the current task's runtime statistics.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 861) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 862) static void update_curr(struct cfs_rq *cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 863) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 864) struct sched_entity *curr = cfs_rq->curr;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 865) u64 now = rq_clock_task(rq_of(cfs_rq));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 866) u64 delta_exec;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 867)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 868) if (unlikely(!curr))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 869) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 870)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 871) delta_exec = now - curr->exec_start;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 872) if (unlikely((s64)delta_exec <= 0))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 873) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 874)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 875) curr->exec_start = now;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 876)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 877) schedstat_set(curr->statistics.exec_max,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 878) max(delta_exec, curr->statistics.exec_max));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 879)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 880) curr->sum_exec_runtime += delta_exec;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 881) schedstat_add(cfs_rq->exec_clock, delta_exec);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 882)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 883) curr->vruntime += calc_delta_fair(delta_exec, curr);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 884) update_min_vruntime(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 885)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 886) if (entity_is_task(curr)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 887) struct task_struct *curtask = task_of(curr);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 888)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 889) trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 890) cgroup_account_cputime(curtask, delta_exec);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 891) account_group_exec_runtime(curtask, delta_exec);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 892) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 893)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 894) account_cfs_rq_runtime(cfs_rq, delta_exec);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 895) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 896)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 897) static void update_curr_fair(struct rq *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 898) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 899) update_curr(cfs_rq_of(&rq->curr->se));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 900) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 901)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 902) static inline void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 903) update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 904) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 905) u64 wait_start, prev_wait_start;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 906)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 907) if (!schedstat_enabled())
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 908) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 909)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 910) wait_start = rq_clock(rq_of(cfs_rq));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 911) prev_wait_start = schedstat_val(se->statistics.wait_start);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 912)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 913) if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 914) likely(wait_start > prev_wait_start))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 915) wait_start -= prev_wait_start;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 916)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 917) __schedstat_set(se->statistics.wait_start, wait_start);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 918) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 919)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 920) static inline void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 921) update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 922) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 923) struct task_struct *p;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 924) u64 delta;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 925)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 926) if (!schedstat_enabled())
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 927) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 928)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 929) delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 930)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 931) if (entity_is_task(se)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 932) p = task_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 933) if (task_on_rq_migrating(p)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 934) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 935) * Preserve migrating task's wait time so wait_start
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 936) * time stamp can be adjusted to accumulate wait time
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 937) * prior to migration.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 938) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 939) __schedstat_set(se->statistics.wait_start, delta);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 940) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 941) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 942) trace_sched_stat_wait(p, delta);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 943) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 944)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 945) __schedstat_set(se->statistics.wait_max,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 946) max(schedstat_val(se->statistics.wait_max), delta));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 947) __schedstat_inc(se->statistics.wait_count);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 948) __schedstat_add(se->statistics.wait_sum, delta);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 949) __schedstat_set(se->statistics.wait_start, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 950) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 951)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 952) static inline void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 953) update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 954) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 955) struct task_struct *tsk = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 956) u64 sleep_start, block_start;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 957)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 958) if (!schedstat_enabled())
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 959) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 960)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 961) sleep_start = schedstat_val(se->statistics.sleep_start);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 962) block_start = schedstat_val(se->statistics.block_start);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 963)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 964) if (entity_is_task(se))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 965) tsk = task_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 966)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 967) if (sleep_start) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 968) u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 969)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 970) if ((s64)delta < 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 971) delta = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 972)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 973) if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 974) __schedstat_set(se->statistics.sleep_max, delta);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 975)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 976) __schedstat_set(se->statistics.sleep_start, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 977) __schedstat_add(se->statistics.sum_sleep_runtime, delta);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 978)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 979) if (tsk) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 980) account_scheduler_latency(tsk, delta >> 10, 1);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 981) trace_sched_stat_sleep(tsk, delta);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 982) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 983) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 984) if (block_start) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 985) u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 986)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 987) if ((s64)delta < 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 988) delta = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 989)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 990) if (unlikely(delta > schedstat_val(se->statistics.block_max)))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 991) __schedstat_set(se->statistics.block_max, delta);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 992)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 993) __schedstat_set(se->statistics.block_start, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 994) __schedstat_add(se->statistics.sum_sleep_runtime, delta);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 995)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 996) if (tsk) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 997) if (tsk->in_iowait) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 998) __schedstat_add(se->statistics.iowait_sum, delta);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 999) __schedstat_inc(se->statistics.iowait_count);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1000) trace_sched_stat_iowait(tsk, delta);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1001) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1002)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1003) trace_sched_stat_blocked(tsk, delta);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1004)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1005) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1006) * Blocking time is in units of nanosecs, so shift by
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1007) * 20 to get a milliseconds-range estimation of the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1008) * amount of time that the task spent sleeping:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1009) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1010) if (unlikely(prof_on == SLEEP_PROFILING)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1011) profile_hits(SLEEP_PROFILING,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1012) (void *)get_wchan(tsk),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1013) delta >> 20);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1014) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1015) account_scheduler_latency(tsk, delta >> 10, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1016) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1017) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1018) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1019)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1020) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1021) * Task is being enqueued - update stats:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1022) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1023) static inline void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1024) update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1025) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1026) if (!schedstat_enabled())
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1027) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1028)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1029) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1030) * Are we enqueueing a waiting task? (for current tasks
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1031) * a dequeue/enqueue event is a NOP)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1032) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1033) if (se != cfs_rq->curr)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1034) update_stats_wait_start(cfs_rq, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1035)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1036) if (flags & ENQUEUE_WAKEUP)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1037) update_stats_enqueue_sleeper(cfs_rq, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1038) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1039)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1040) static inline void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1041) update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1042) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1043)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1044) if (!schedstat_enabled())
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1045) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1046)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1047) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1048) * Mark the end of the wait period if dequeueing a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1049) * waiting task:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1050) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1051) if (se != cfs_rq->curr)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1052) update_stats_wait_end(cfs_rq, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1053)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1054) if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1055) struct task_struct *tsk = task_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1056)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1057) if (tsk->state & TASK_INTERRUPTIBLE)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1058) __schedstat_set(se->statistics.sleep_start,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1059) rq_clock(rq_of(cfs_rq)));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1060) if (tsk->state & TASK_UNINTERRUPTIBLE)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1061) __schedstat_set(se->statistics.block_start,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1062) rq_clock(rq_of(cfs_rq)));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1063) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1064) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1065)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1066) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1067) * We are picking a new current task - update its stats:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1068) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1069) static inline void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1070) update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1071) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1072) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1073) * We are starting a new run period:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1074) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1075) se->exec_start = rq_clock_task(rq_of(cfs_rq));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1076) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1077)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1078) /**************************************************
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1079) * Scheduling class queueing methods:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1080) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1081)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1082) #ifdef CONFIG_NUMA_BALANCING
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1083) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1084) * Approximate time to scan a full NUMA task in ms. The task scan period is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1085) * calculated based on the tasks virtual memory size and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1086) * numa_balancing_scan_size.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1087) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1088) unsigned int sysctl_numa_balancing_scan_period_min = 1000;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1089) unsigned int sysctl_numa_balancing_scan_period_max = 60000;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1090)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1091) /* Portion of address space to scan in MB */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1092) unsigned int sysctl_numa_balancing_scan_size = 256;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1093)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1094) /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1095) unsigned int sysctl_numa_balancing_scan_delay = 1000;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1096)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1097) struct numa_group {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1098) refcount_t refcount;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1099)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1100) spinlock_t lock; /* nr_tasks, tasks */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1101) int nr_tasks;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1102) pid_t gid;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1103) int active_nodes;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1104)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1105) struct rcu_head rcu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1106) unsigned long total_faults;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1107) unsigned long max_faults_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1108) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1109) * Faults_cpu is used to decide whether memory should move
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1110) * towards the CPU. As a consequence, these stats are weighted
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1111) * more by CPU use than by memory faults.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1112) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1113) unsigned long *faults_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1114) unsigned long faults[];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1115) };
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1116)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1117) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1118) * For functions that can be called in multiple contexts that permit reading
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1119) * ->numa_group (see struct task_struct for locking rules).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1120) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1121) static struct numa_group *deref_task_numa_group(struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1122) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1123) return rcu_dereference_check(p->numa_group, p == current ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1124) (lockdep_is_held(&task_rq(p)->lock) && !READ_ONCE(p->on_cpu)));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1125) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1126)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1127) static struct numa_group *deref_curr_numa_group(struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1128) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1129) return rcu_dereference_protected(p->numa_group, p == current);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1130) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1131)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1132) static inline unsigned long group_faults_priv(struct numa_group *ng);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1133) static inline unsigned long group_faults_shared(struct numa_group *ng);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1134)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1135) static unsigned int task_nr_scan_windows(struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1136) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1137) unsigned long rss = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1138) unsigned long nr_scan_pages;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1139)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1140) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1141) * Calculations based on RSS as non-present and empty pages are skipped
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1142) * by the PTE scanner and NUMA hinting faults should be trapped based
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1143) * on resident pages
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1144) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1145) nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1146) rss = get_mm_rss(p->mm);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1147) if (!rss)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1148) rss = nr_scan_pages;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1149)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1150) rss = round_up(rss, nr_scan_pages);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1151) return rss / nr_scan_pages;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1152) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1153)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1154) /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1155) #define MAX_SCAN_WINDOW 2560
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1156)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1157) static unsigned int task_scan_min(struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1158) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1159) unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1160) unsigned int scan, floor;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1161) unsigned int windows = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1162)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1163) if (scan_size < MAX_SCAN_WINDOW)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1164) windows = MAX_SCAN_WINDOW / scan_size;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1165) floor = 1000 / windows;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1166)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1167) scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1168) return max_t(unsigned int, floor, scan);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1169) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1170)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1171) static unsigned int task_scan_start(struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1172) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1173) unsigned long smin = task_scan_min(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1174) unsigned long period = smin;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1175) struct numa_group *ng;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1176)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1177) /* Scale the maximum scan period with the amount of shared memory. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1178) rcu_read_lock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1179) ng = rcu_dereference(p->numa_group);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1180) if (ng) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1181) unsigned long shared = group_faults_shared(ng);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1182) unsigned long private = group_faults_priv(ng);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1183)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1184) period *= refcount_read(&ng->refcount);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1185) period *= shared + 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1186) period /= private + shared + 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1187) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1188) rcu_read_unlock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1189)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1190) return max(smin, period);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1191) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1192)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1193) static unsigned int task_scan_max(struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1194) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1195) unsigned long smin = task_scan_min(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1196) unsigned long smax;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1197) struct numa_group *ng;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1198)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1199) /* Watch for min being lower than max due to floor calculations */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1200) smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1201)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1202) /* Scale the maximum scan period with the amount of shared memory. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1203) ng = deref_curr_numa_group(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1204) if (ng) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1205) unsigned long shared = group_faults_shared(ng);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1206) unsigned long private = group_faults_priv(ng);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1207) unsigned long period = smax;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1208)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1209) period *= refcount_read(&ng->refcount);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1210) period *= shared + 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1211) period /= private + shared + 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1212)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1213) smax = max(smax, period);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1214) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1215)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1216) return max(smin, smax);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1217) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1218)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1219) static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1220) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1221) rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1222) rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1223) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1224)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1225) static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1226) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1227) rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1228) rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1229) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1230)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1231) /* Shared or private faults. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1232) #define NR_NUMA_HINT_FAULT_TYPES 2
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1233)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1234) /* Memory and CPU locality */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1235) #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1236)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1237) /* Averaged statistics, and temporary buffers. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1238) #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1239)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1240) pid_t task_numa_group_id(struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1241) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1242) struct numa_group *ng;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1243) pid_t gid = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1244)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1245) rcu_read_lock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1246) ng = rcu_dereference(p->numa_group);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1247) if (ng)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1248) gid = ng->gid;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1249) rcu_read_unlock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1250)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1251) return gid;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1252) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1253)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1254) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1255) * The averaged statistics, shared & private, memory & CPU,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1256) * occupy the first half of the array. The second half of the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1257) * array is for current counters, which are averaged into the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1258) * first set by task_numa_placement.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1259) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1260) static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1261) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1262) return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1263) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1264)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1265) static inline unsigned long task_faults(struct task_struct *p, int nid)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1266) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1267) if (!p->numa_faults)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1268) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1269)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1270) return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1271) p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1272) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1273)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1274) static inline unsigned long group_faults(struct task_struct *p, int nid)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1275) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1276) struct numa_group *ng = deref_task_numa_group(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1277)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1278) if (!ng)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1279) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1280)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1281) return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1282) ng->faults[task_faults_idx(NUMA_MEM, nid, 1)];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1283) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1284)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1285) static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1286) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1287) return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1288) group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1289) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1290)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1291) static inline unsigned long group_faults_priv(struct numa_group *ng)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1292) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1293) unsigned long faults = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1294) int node;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1295)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1296) for_each_online_node(node) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1297) faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1298) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1299)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1300) return faults;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1301) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1302)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1303) static inline unsigned long group_faults_shared(struct numa_group *ng)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1304) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1305) unsigned long faults = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1306) int node;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1307)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1308) for_each_online_node(node) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1309) faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1310) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1311)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1312) return faults;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1313) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1314)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1315) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1316) * A node triggering more than 1/3 as many NUMA faults as the maximum is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1317) * considered part of a numa group's pseudo-interleaving set. Migrations
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1318) * between these nodes are slowed down, to allow things to settle down.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1319) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1320) #define ACTIVE_NODE_FRACTION 3
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1321)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1322) static bool numa_is_active_node(int nid, struct numa_group *ng)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1323) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1324) return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1325) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1326)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1327) /* Handle placement on systems where not all nodes are directly connected. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1328) static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1329) int maxdist, bool task)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1330) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1331) unsigned long score = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1332) int node;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1333)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1334) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1335) * All nodes are directly connected, and the same distance
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1336) * from each other. No need for fancy placement algorithms.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1337) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1338) if (sched_numa_topology_type == NUMA_DIRECT)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1339) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1340)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1341) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1342) * This code is called for each node, introducing N^2 complexity,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1343) * which should be ok given the number of nodes rarely exceeds 8.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1344) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1345) for_each_online_node(node) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1346) unsigned long faults;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1347) int dist = node_distance(nid, node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1348)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1349) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1350) * The furthest away nodes in the system are not interesting
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1351) * for placement; nid was already counted.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1352) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1353) if (dist == sched_max_numa_distance || node == nid)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1354) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1355)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1356) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1357) * On systems with a backplane NUMA topology, compare groups
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1358) * of nodes, and move tasks towards the group with the most
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1359) * memory accesses. When comparing two nodes at distance
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1360) * "hoplimit", only nodes closer by than "hoplimit" are part
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1361) * of each group. Skip other nodes.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1362) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1363) if (sched_numa_topology_type == NUMA_BACKPLANE &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1364) dist >= maxdist)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1365) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1366)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1367) /* Add up the faults from nearby nodes. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1368) if (task)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1369) faults = task_faults(p, node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1370) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1371) faults = group_faults(p, node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1372)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1373) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1374) * On systems with a glueless mesh NUMA topology, there are
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1375) * no fixed "groups of nodes". Instead, nodes that are not
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1376) * directly connected bounce traffic through intermediate
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1377) * nodes; a numa_group can occupy any set of nodes.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1378) * The further away a node is, the less the faults count.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1379) * This seems to result in good task placement.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1380) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1381) if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1382) faults *= (sched_max_numa_distance - dist);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1383) faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1384) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1385)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1386) score += faults;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1387) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1388)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1389) return score;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1390) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1391)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1392) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1393) * These return the fraction of accesses done by a particular task, or
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1394) * task group, on a particular numa node. The group weight is given a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1395) * larger multiplier, in order to group tasks together that are almost
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1396) * evenly spread out between numa nodes.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1397) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1398) static inline unsigned long task_weight(struct task_struct *p, int nid,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1399) int dist)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1400) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1401) unsigned long faults, total_faults;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1402)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1403) if (!p->numa_faults)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1404) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1405)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1406) total_faults = p->total_numa_faults;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1407)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1408) if (!total_faults)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1409) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1410)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1411) faults = task_faults(p, nid);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1412) faults += score_nearby_nodes(p, nid, dist, true);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1413)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1414) return 1000 * faults / total_faults;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1415) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1416)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1417) static inline unsigned long group_weight(struct task_struct *p, int nid,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1418) int dist)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1419) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1420) struct numa_group *ng = deref_task_numa_group(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1421) unsigned long faults, total_faults;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1422)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1423) if (!ng)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1424) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1425)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1426) total_faults = ng->total_faults;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1427)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1428) if (!total_faults)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1429) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1430)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1431) faults = group_faults(p, nid);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1432) faults += score_nearby_nodes(p, nid, dist, false);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1433)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1434) return 1000 * faults / total_faults;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1435) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1436)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1437) bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1438) int src_nid, int dst_cpu)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1439) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1440) struct numa_group *ng = deref_curr_numa_group(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1441) int dst_nid = cpu_to_node(dst_cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1442) int last_cpupid, this_cpupid;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1443)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1444) this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1445) last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1446)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1447) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1448) * Allow first faults or private faults to migrate immediately early in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1449) * the lifetime of a task. The magic number 4 is based on waiting for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1450) * two full passes of the "multi-stage node selection" test that is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1451) * executed below.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1452) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1453) if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1454) (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1455) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1456)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1457) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1458) * Multi-stage node selection is used in conjunction with a periodic
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1459) * migration fault to build a temporal task<->page relation. By using
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1460) * a two-stage filter we remove short/unlikely relations.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1461) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1462) * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1463) * a task's usage of a particular page (n_p) per total usage of this
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1464) * page (n_t) (in a given time-span) to a probability.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1465) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1466) * Our periodic faults will sample this probability and getting the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1467) * same result twice in a row, given these samples are fully
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1468) * independent, is then given by P(n)^2, provided our sample period
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1469) * is sufficiently short compared to the usage pattern.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1470) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1471) * This quadric squishes small probabilities, making it less likely we
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1472) * act on an unlikely task<->page relation.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1473) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1474) if (!cpupid_pid_unset(last_cpupid) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1475) cpupid_to_nid(last_cpupid) != dst_nid)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1476) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1477)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1478) /* Always allow migrate on private faults */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1479) if (cpupid_match_pid(p, last_cpupid))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1480) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1481)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1482) /* A shared fault, but p->numa_group has not been set up yet. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1483) if (!ng)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1484) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1485)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1486) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1487) * Destination node is much more heavily used than the source
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1488) * node? Allow migration.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1489) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1490) if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1491) ACTIVE_NODE_FRACTION)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1492) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1493)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1494) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1495) * Distribute memory according to CPU & memory use on each node,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1496) * with 3/4 hysteresis to avoid unnecessary memory migrations:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1497) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1498) * faults_cpu(dst) 3 faults_cpu(src)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1499) * --------------- * - > ---------------
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1500) * faults_mem(dst) 4 faults_mem(src)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1501) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1502) return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1503) group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1504) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1505)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1506) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1507) * 'numa_type' describes the node at the moment of load balancing.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1508) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1509) enum numa_type {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1510) /* The node has spare capacity that can be used to run more tasks. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1511) node_has_spare = 0,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1512) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1513) * The node is fully used and the tasks don't compete for more CPU
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1514) * cycles. Nevertheless, some tasks might wait before running.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1515) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1516) node_fully_busy,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1517) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1518) * The node is overloaded and can't provide expected CPU cycles to all
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1519) * tasks.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1520) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1521) node_overloaded
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1522) };
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1523)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1524) /* Cached statistics for all CPUs within a node */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1525) struct numa_stats {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1526) unsigned long load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1527) unsigned long runnable;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1528) unsigned long util;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1529) /* Total compute capacity of CPUs on a node */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1530) unsigned long compute_capacity;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1531) unsigned int nr_running;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1532) unsigned int weight;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1533) enum numa_type node_type;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1534) int idle_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1535) };
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1536)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1537) static inline bool is_core_idle(int cpu)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1538) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1539) #ifdef CONFIG_SCHED_SMT
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1540) int sibling;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1541)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1542) for_each_cpu(sibling, cpu_smt_mask(cpu)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1543) if (cpu == sibling)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1544) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1545)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1546) if (!idle_cpu(sibling))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1547) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1548) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1549) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1550)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1551) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1552) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1553)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1554) struct task_numa_env {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1555) struct task_struct *p;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1556)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1557) int src_cpu, src_nid;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1558) int dst_cpu, dst_nid;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1559)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1560) struct numa_stats src_stats, dst_stats;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1561)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1562) int imbalance_pct;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1563) int dist;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1564)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1565) struct task_struct *best_task;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1566) long best_imp;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1567) int best_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1568) };
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1569)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1570) static unsigned long cpu_load(struct rq *rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1571) static unsigned long cpu_runnable(struct rq *rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1572) static unsigned long cpu_util(int cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1573) static inline long adjust_numa_imbalance(int imbalance, int nr_running);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1574)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1575) static inline enum
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1576) numa_type numa_classify(unsigned int imbalance_pct,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1577) struct numa_stats *ns)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1578) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1579) if ((ns->nr_running > ns->weight) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1580) (((ns->compute_capacity * 100) < (ns->util * imbalance_pct)) ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1581) ((ns->compute_capacity * imbalance_pct) < (ns->runnable * 100))))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1582) return node_overloaded;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1583)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1584) if ((ns->nr_running < ns->weight) ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1585) (((ns->compute_capacity * 100) > (ns->util * imbalance_pct)) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1586) ((ns->compute_capacity * imbalance_pct) > (ns->runnable * 100))))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1587) return node_has_spare;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1588)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1589) return node_fully_busy;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1590) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1591)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1592) #ifdef CONFIG_SCHED_SMT
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1593) /* Forward declarations of select_idle_sibling helpers */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1594) static inline bool test_idle_cores(int cpu, bool def);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1595) static inline int numa_idle_core(int idle_core, int cpu)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1596) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1597) if (!static_branch_likely(&sched_smt_present) ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1598) idle_core >= 0 || !test_idle_cores(cpu, false))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1599) return idle_core;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1600)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1601) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1602) * Prefer cores instead of packing HT siblings
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1603) * and triggering future load balancing.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1604) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1605) if (is_core_idle(cpu))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1606) idle_core = cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1607)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1608) return idle_core;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1609) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1610) #else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1611) static inline int numa_idle_core(int idle_core, int cpu)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1612) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1613) return idle_core;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1614) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1615) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1616)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1617) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1618) * Gather all necessary information to make NUMA balancing placement
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1619) * decisions that are compatible with standard load balancer. This
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1620) * borrows code and logic from update_sg_lb_stats but sharing a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1621) * common implementation is impractical.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1622) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1623) static void update_numa_stats(struct task_numa_env *env,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1624) struct numa_stats *ns, int nid,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1625) bool find_idle)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1626) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1627) int cpu, idle_core = -1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1628)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1629) memset(ns, 0, sizeof(*ns));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1630) ns->idle_cpu = -1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1631)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1632) rcu_read_lock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1633) for_each_cpu(cpu, cpumask_of_node(nid)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1634) struct rq *rq = cpu_rq(cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1635)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1636) ns->load += cpu_load(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1637) ns->runnable += cpu_runnable(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1638) ns->util += cpu_util(cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1639) ns->nr_running += rq->cfs.h_nr_running;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1640) ns->compute_capacity += capacity_of(cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1641)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1642) if (find_idle && !rq->nr_running && idle_cpu(cpu)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1643) if (READ_ONCE(rq->numa_migrate_on) ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1644) !cpumask_test_cpu(cpu, env->p->cpus_ptr))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1645) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1646)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1647) if (ns->idle_cpu == -1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1648) ns->idle_cpu = cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1649)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1650) idle_core = numa_idle_core(idle_core, cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1651) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1652) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1653) rcu_read_unlock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1654)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1655) ns->weight = cpumask_weight(cpumask_of_node(nid));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1656)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1657) ns->node_type = numa_classify(env->imbalance_pct, ns);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1658)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1659) if (idle_core >= 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1660) ns->idle_cpu = idle_core;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1661) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1662)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1663) static void task_numa_assign(struct task_numa_env *env,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1664) struct task_struct *p, long imp)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1665) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1666) struct rq *rq = cpu_rq(env->dst_cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1667)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1668) /* Check if run-queue part of active NUMA balance. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1669) if (env->best_cpu != env->dst_cpu && xchg(&rq->numa_migrate_on, 1)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1670) int cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1671) int start = env->dst_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1672)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1673) /* Find alternative idle CPU. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1674) for_each_cpu_wrap(cpu, cpumask_of_node(env->dst_nid), start) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1675) if (cpu == env->best_cpu || !idle_cpu(cpu) ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1676) !cpumask_test_cpu(cpu, env->p->cpus_ptr)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1677) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1678) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1679)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1680) env->dst_cpu = cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1681) rq = cpu_rq(env->dst_cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1682) if (!xchg(&rq->numa_migrate_on, 1))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1683) goto assign;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1684) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1685)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1686) /* Failed to find an alternative idle CPU */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1687) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1688) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1689)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1690) assign:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1691) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1692) * Clear previous best_cpu/rq numa-migrate flag, since task now
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1693) * found a better CPU to move/swap.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1694) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1695) if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1696) rq = cpu_rq(env->best_cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1697) WRITE_ONCE(rq->numa_migrate_on, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1698) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1699)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1700) if (env->best_task)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1701) put_task_struct(env->best_task);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1702) if (p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1703) get_task_struct(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1704)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1705) env->best_task = p;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1706) env->best_imp = imp;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1707) env->best_cpu = env->dst_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1708) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1709)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1710) static bool load_too_imbalanced(long src_load, long dst_load,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1711) struct task_numa_env *env)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1712) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1713) long imb, old_imb;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1714) long orig_src_load, orig_dst_load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1715) long src_capacity, dst_capacity;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1716)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1717) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1718) * The load is corrected for the CPU capacity available on each node.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1719) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1720) * src_load dst_load
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1721) * ------------ vs ---------
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1722) * src_capacity dst_capacity
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1723) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1724) src_capacity = env->src_stats.compute_capacity;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1725) dst_capacity = env->dst_stats.compute_capacity;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1726)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1727) imb = abs(dst_load * src_capacity - src_load * dst_capacity);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1728)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1729) orig_src_load = env->src_stats.load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1730) orig_dst_load = env->dst_stats.load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1731)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1732) old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1733)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1734) /* Would this change make things worse? */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1735) return (imb > old_imb);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1736) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1737)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1738) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1739) * Maximum NUMA importance can be 1998 (2*999);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1740) * SMALLIMP @ 30 would be close to 1998/64.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1741) * Used to deter task migration.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1742) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1743) #define SMALLIMP 30
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1744)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1745) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1746) * This checks if the overall compute and NUMA accesses of the system would
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1747) * be improved if the source tasks was migrated to the target dst_cpu taking
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1748) * into account that it might be best if task running on the dst_cpu should
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1749) * be exchanged with the source task
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1750) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1751) static bool task_numa_compare(struct task_numa_env *env,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1752) long taskimp, long groupimp, bool maymove)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1753) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1754) struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1755) struct rq *dst_rq = cpu_rq(env->dst_cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1756) long imp = p_ng ? groupimp : taskimp;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1757) struct task_struct *cur;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1758) long src_load, dst_load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1759) int dist = env->dist;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1760) long moveimp = imp;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1761) long load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1762) bool stopsearch = false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1763)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1764) if (READ_ONCE(dst_rq->numa_migrate_on))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1765) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1766)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1767) rcu_read_lock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1768) cur = rcu_dereference(dst_rq->curr);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1769) if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1770) cur = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1771)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1772) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1773) * Because we have preemption enabled we can get migrated around and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1774) * end try selecting ourselves (current == env->p) as a swap candidate.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1775) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1776) if (cur == env->p) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1777) stopsearch = true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1778) goto unlock;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1779) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1780)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1781) if (!cur) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1782) if (maymove && moveimp >= env->best_imp)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1783) goto assign;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1784) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1785) goto unlock;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1786) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1787)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1788) /* Skip this swap candidate if cannot move to the source cpu. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1789) if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1790) goto unlock;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1791)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1792) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1793) * Skip this swap candidate if it is not moving to its preferred
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1794) * node and the best task is.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1795) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1796) if (env->best_task &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1797) env->best_task->numa_preferred_nid == env->src_nid &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1798) cur->numa_preferred_nid != env->src_nid) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1799) goto unlock;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1800) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1801)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1802) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1803) * "imp" is the fault differential for the source task between the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1804) * source and destination node. Calculate the total differential for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1805) * the source task and potential destination task. The more negative
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1806) * the value is, the more remote accesses that would be expected to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1807) * be incurred if the tasks were swapped.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1808) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1809) * If dst and source tasks are in the same NUMA group, or not
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1810) * in any group then look only at task weights.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1811) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1812) cur_ng = rcu_dereference(cur->numa_group);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1813) if (cur_ng == p_ng) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1814) imp = taskimp + task_weight(cur, env->src_nid, dist) -
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1815) task_weight(cur, env->dst_nid, dist);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1816) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1817) * Add some hysteresis to prevent swapping the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1818) * tasks within a group over tiny differences.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1819) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1820) if (cur_ng)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1821) imp -= imp / 16;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1822) } else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1823) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1824) * Compare the group weights. If a task is all by itself
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1825) * (not part of a group), use the task weight instead.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1826) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1827) if (cur_ng && p_ng)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1828) imp += group_weight(cur, env->src_nid, dist) -
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1829) group_weight(cur, env->dst_nid, dist);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1830) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1831) imp += task_weight(cur, env->src_nid, dist) -
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1832) task_weight(cur, env->dst_nid, dist);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1833) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1834)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1835) /* Discourage picking a task already on its preferred node */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1836) if (cur->numa_preferred_nid == env->dst_nid)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1837) imp -= imp / 16;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1838)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1839) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1840) * Encourage picking a task that moves to its preferred node.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1841) * This potentially makes imp larger than it's maximum of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1842) * 1998 (see SMALLIMP and task_weight for why) but in this
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1843) * case, it does not matter.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1844) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1845) if (cur->numa_preferred_nid == env->src_nid)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1846) imp += imp / 8;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1847)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1848) if (maymove && moveimp > imp && moveimp > env->best_imp) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1849) imp = moveimp;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1850) cur = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1851) goto assign;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1852) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1853)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1854) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1855) * Prefer swapping with a task moving to its preferred node over a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1856) * task that is not.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1857) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1858) if (env->best_task && cur->numa_preferred_nid == env->src_nid &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1859) env->best_task->numa_preferred_nid != env->src_nid) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1860) goto assign;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1861) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1862)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1863) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1864) * If the NUMA importance is less than SMALLIMP,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1865) * task migration might only result in ping pong
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1866) * of tasks and also hurt performance due to cache
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1867) * misses.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1868) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1869) if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1870) goto unlock;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1871)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1872) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1873) * In the overloaded case, try and keep the load balanced.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1874) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1875) load = task_h_load(env->p) - task_h_load(cur);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1876) if (!load)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1877) goto assign;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1878)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1879) dst_load = env->dst_stats.load + load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1880) src_load = env->src_stats.load - load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1881)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1882) if (load_too_imbalanced(src_load, dst_load, env))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1883) goto unlock;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1884)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1885) assign:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1886) /* Evaluate an idle CPU for a task numa move. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1887) if (!cur) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1888) int cpu = env->dst_stats.idle_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1889)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1890) /* Nothing cached so current CPU went idle since the search. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1891) if (cpu < 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1892) cpu = env->dst_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1893)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1894) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1895) * If the CPU is no longer truly idle and the previous best CPU
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1896) * is, keep using it.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1897) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1898) if (!idle_cpu(cpu) && env->best_cpu >= 0 &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1899) idle_cpu(env->best_cpu)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1900) cpu = env->best_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1901) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1902)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1903) env->dst_cpu = cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1904) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1905)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1906) task_numa_assign(env, cur, imp);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1907)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1908) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1909) * If a move to idle is allowed because there is capacity or load
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1910) * balance improves then stop the search. While a better swap
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1911) * candidate may exist, a search is not free.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1912) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1913) if (maymove && !cur && env->best_cpu >= 0 && idle_cpu(env->best_cpu))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1914) stopsearch = true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1915)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1916) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1917) * If a swap candidate must be identified and the current best task
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1918) * moves its preferred node then stop the search.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1919) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1920) if (!maymove && env->best_task &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1921) env->best_task->numa_preferred_nid == env->src_nid) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1922) stopsearch = true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1923) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1924) unlock:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1925) rcu_read_unlock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1926)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1927) return stopsearch;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1928) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1929)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1930) static void task_numa_find_cpu(struct task_numa_env *env,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1931) long taskimp, long groupimp)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1932) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1933) bool maymove = false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1934) int cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1935)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1936) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1937) * If dst node has spare capacity, then check if there is an
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1938) * imbalance that would be overruled by the load balancer.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1939) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1940) if (env->dst_stats.node_type == node_has_spare) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1941) unsigned int imbalance;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1942) int src_running, dst_running;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1943)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1944) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1945) * Would movement cause an imbalance? Note that if src has
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1946) * more running tasks that the imbalance is ignored as the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1947) * move improves the imbalance from the perspective of the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1948) * CPU load balancer.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1949) * */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1950) src_running = env->src_stats.nr_running - 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1951) dst_running = env->dst_stats.nr_running + 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1952) imbalance = max(0, dst_running - src_running);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1953) imbalance = adjust_numa_imbalance(imbalance, dst_running);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1954)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1955) /* Use idle CPU if there is no imbalance */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1956) if (!imbalance) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1957) maymove = true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1958) if (env->dst_stats.idle_cpu >= 0) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1959) env->dst_cpu = env->dst_stats.idle_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1960) task_numa_assign(env, NULL, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1961) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1962) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1963) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1964) } else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1965) long src_load, dst_load, load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1966) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1967) * If the improvement from just moving env->p direction is better
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1968) * than swapping tasks around, check if a move is possible.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1969) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1970) load = task_h_load(env->p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1971) dst_load = env->dst_stats.load + load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1972) src_load = env->src_stats.load - load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1973) maymove = !load_too_imbalanced(src_load, dst_load, env);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1974) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1975)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1976) for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1977) /* Skip this CPU if the source task cannot migrate */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1978) if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1979) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1980)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1981) env->dst_cpu = cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1982) if (task_numa_compare(env, taskimp, groupimp, maymove))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1983) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1984) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1985) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1986)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1987) static int task_numa_migrate(struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1988) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1989) struct task_numa_env env = {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1990) .p = p,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1991)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1992) .src_cpu = task_cpu(p),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1993) .src_nid = task_node(p),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1994)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1995) .imbalance_pct = 112,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1996)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1997) .best_task = NULL,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1998) .best_imp = 0,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1999) .best_cpu = -1,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2000) };
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2001) unsigned long taskweight, groupweight;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2002) struct sched_domain *sd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2003) long taskimp, groupimp;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2004) struct numa_group *ng;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2005) struct rq *best_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2006) int nid, ret, dist;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2007)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2008) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2009) * Pick the lowest SD_NUMA domain, as that would have the smallest
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2010) * imbalance and would be the first to start moving tasks about.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2011) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2012) * And we want to avoid any moving of tasks about, as that would create
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2013) * random movement of tasks -- counter the numa conditions we're trying
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2014) * to satisfy here.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2015) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2016) rcu_read_lock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2017) sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2018) if (sd)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2019) env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2020) rcu_read_unlock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2021)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2022) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2023) * Cpusets can break the scheduler domain tree into smaller
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2024) * balance domains, some of which do not cross NUMA boundaries.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2025) * Tasks that are "trapped" in such domains cannot be migrated
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2026) * elsewhere, so there is no point in (re)trying.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2027) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2028) if (unlikely(!sd)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2029) sched_setnuma(p, task_node(p));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2030) return -EINVAL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2031) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2032)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2033) env.dst_nid = p->numa_preferred_nid;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2034) dist = env.dist = node_distance(env.src_nid, env.dst_nid);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2035) taskweight = task_weight(p, env.src_nid, dist);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2036) groupweight = group_weight(p, env.src_nid, dist);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2037) update_numa_stats(&env, &env.src_stats, env.src_nid, false);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2038) taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2039) groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2040) update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2041)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2042) /* Try to find a spot on the preferred nid. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2043) task_numa_find_cpu(&env, taskimp, groupimp);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2044)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2045) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2046) * Look at other nodes in these cases:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2047) * - there is no space available on the preferred_nid
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2048) * - the task is part of a numa_group that is interleaved across
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2049) * multiple NUMA nodes; in order to better consolidate the group,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2050) * we need to check other locations.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2051) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2052) ng = deref_curr_numa_group(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2053) if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2054) for_each_online_node(nid) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2055) if (nid == env.src_nid || nid == p->numa_preferred_nid)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2056) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2057)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2058) dist = node_distance(env.src_nid, env.dst_nid);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2059) if (sched_numa_topology_type == NUMA_BACKPLANE &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2060) dist != env.dist) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2061) taskweight = task_weight(p, env.src_nid, dist);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2062) groupweight = group_weight(p, env.src_nid, dist);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2063) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2064)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2065) /* Only consider nodes where both task and groups benefit */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2066) taskimp = task_weight(p, nid, dist) - taskweight;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2067) groupimp = group_weight(p, nid, dist) - groupweight;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2068) if (taskimp < 0 && groupimp < 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2069) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2070)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2071) env.dist = dist;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2072) env.dst_nid = nid;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2073) update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2074) task_numa_find_cpu(&env, taskimp, groupimp);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2075) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2076) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2077)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2078) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2079) * If the task is part of a workload that spans multiple NUMA nodes,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2080) * and is migrating into one of the workload's active nodes, remember
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2081) * this node as the task's preferred numa node, so the workload can
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2082) * settle down.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2083) * A task that migrated to a second choice node will be better off
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2084) * trying for a better one later. Do not set the preferred node here.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2085) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2086) if (ng) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2087) if (env.best_cpu == -1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2088) nid = env.src_nid;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2089) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2090) nid = cpu_to_node(env.best_cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2091)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2092) if (nid != p->numa_preferred_nid)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2093) sched_setnuma(p, nid);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2094) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2095)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2096) /* No better CPU than the current one was found. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2097) if (env.best_cpu == -1) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2098) trace_sched_stick_numa(p, env.src_cpu, NULL, -1);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2099) return -EAGAIN;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2100) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2101)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2102) best_rq = cpu_rq(env.best_cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2103) if (env.best_task == NULL) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2104) ret = migrate_task_to(p, env.best_cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2105) WRITE_ONCE(best_rq->numa_migrate_on, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2106) if (ret != 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2107) trace_sched_stick_numa(p, env.src_cpu, NULL, env.best_cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2108) return ret;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2109) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2110)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2111) ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2112) WRITE_ONCE(best_rq->numa_migrate_on, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2113)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2114) if (ret != 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2115) trace_sched_stick_numa(p, env.src_cpu, env.best_task, env.best_cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2116) put_task_struct(env.best_task);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2117) return ret;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2118) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2119)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2120) /* Attempt to migrate a task to a CPU on the preferred node. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2121) static void numa_migrate_preferred(struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2122) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2123) unsigned long interval = HZ;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2124)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2125) /* This task has no NUMA fault statistics yet */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2126) if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2127) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2128)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2129) /* Periodically retry migrating the task to the preferred node */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2130) interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2131) p->numa_migrate_retry = jiffies + interval;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2132)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2133) /* Success if task is already running on preferred CPU */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2134) if (task_node(p) == p->numa_preferred_nid)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2135) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2136)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2137) /* Otherwise, try migrate to a CPU on the preferred node */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2138) task_numa_migrate(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2139) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2140)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2141) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2142) * Find out how many nodes on the workload is actively running on. Do this by
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2143) * tracking the nodes from which NUMA hinting faults are triggered. This can
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2144) * be different from the set of nodes where the workload's memory is currently
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2145) * located.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2146) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2147) static void numa_group_count_active_nodes(struct numa_group *numa_group)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2148) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2149) unsigned long faults, max_faults = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2150) int nid, active_nodes = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2151)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2152) for_each_online_node(nid) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2153) faults = group_faults_cpu(numa_group, nid);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2154) if (faults > max_faults)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2155) max_faults = faults;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2156) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2157)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2158) for_each_online_node(nid) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2159) faults = group_faults_cpu(numa_group, nid);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2160) if (faults * ACTIVE_NODE_FRACTION > max_faults)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2161) active_nodes++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2162) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2163)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2164) numa_group->max_faults_cpu = max_faults;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2165) numa_group->active_nodes = active_nodes;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2166) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2167)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2168) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2169) * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2170) * increments. The more local the fault statistics are, the higher the scan
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2171) * period will be for the next scan window. If local/(local+remote) ratio is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2172) * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2173) * the scan period will decrease. Aim for 70% local accesses.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2174) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2175) #define NUMA_PERIOD_SLOTS 10
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2176) #define NUMA_PERIOD_THRESHOLD 7
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2177)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2178) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2179) * Increase the scan period (slow down scanning) if the majority of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2180) * our memory is already on our local node, or if the majority of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2181) * the page accesses are shared with other processes.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2182) * Otherwise, decrease the scan period.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2183) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2184) static void update_task_scan_period(struct task_struct *p,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2185) unsigned long shared, unsigned long private)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2186) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2187) unsigned int period_slot;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2188) int lr_ratio, ps_ratio;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2189) int diff;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2190)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2191) unsigned long remote = p->numa_faults_locality[0];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2192) unsigned long local = p->numa_faults_locality[1];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2193)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2194) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2195) * If there were no record hinting faults then either the task is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2196) * completely idle or all activity is areas that are not of interest
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2197) * to automatic numa balancing. Related to that, if there were failed
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2198) * migration then it implies we are migrating too quickly or the local
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2199) * node is overloaded. In either case, scan slower
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2200) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2201) if (local + shared == 0 || p->numa_faults_locality[2]) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2202) p->numa_scan_period = min(p->numa_scan_period_max,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2203) p->numa_scan_period << 1);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2204)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2205) p->mm->numa_next_scan = jiffies +
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2206) msecs_to_jiffies(p->numa_scan_period);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2207)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2208) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2209) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2210)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2211) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2212) * Prepare to scale scan period relative to the current period.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2213) * == NUMA_PERIOD_THRESHOLD scan period stays the same
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2214) * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2215) * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2216) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2217) period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2218) lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2219) ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2220)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2221) if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2222) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2223) * Most memory accesses are local. There is no need to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2224) * do fast NUMA scanning, since memory is already local.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2225) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2226) int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2227) if (!slot)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2228) slot = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2229) diff = slot * period_slot;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2230) } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2231) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2232) * Most memory accesses are shared with other tasks.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2233) * There is no point in continuing fast NUMA scanning,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2234) * since other tasks may just move the memory elsewhere.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2235) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2236) int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2237) if (!slot)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2238) slot = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2239) diff = slot * period_slot;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2240) } else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2241) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2242) * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2243) * yet they are not on the local NUMA node. Speed up
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2244) * NUMA scanning to get the memory moved over.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2245) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2246) int ratio = max(lr_ratio, ps_ratio);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2247) diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2248) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2249)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2250) p->numa_scan_period = clamp(p->numa_scan_period + diff,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2251) task_scan_min(p), task_scan_max(p));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2252) memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2253) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2254)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2255) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2256) * Get the fraction of time the task has been running since the last
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2257) * NUMA placement cycle. The scheduler keeps similar statistics, but
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2258) * decays those on a 32ms period, which is orders of magnitude off
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2259) * from the dozens-of-seconds NUMA balancing period. Use the scheduler
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2260) * stats only if the task is so new there are no NUMA statistics yet.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2261) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2262) static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2263) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2264) u64 runtime, delta, now;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2265) /* Use the start of this time slice to avoid calculations. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2266) now = p->se.exec_start;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2267) runtime = p->se.sum_exec_runtime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2268)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2269) if (p->last_task_numa_placement) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2270) delta = runtime - p->last_sum_exec_runtime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2271) *period = now - p->last_task_numa_placement;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2272)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2273) /* Avoid time going backwards, prevent potential divide error: */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2274) if (unlikely((s64)*period < 0))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2275) *period = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2276) } else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2277) delta = p->se.avg.load_sum;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2278) *period = LOAD_AVG_MAX;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2279) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2280)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2281) p->last_sum_exec_runtime = runtime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2282) p->last_task_numa_placement = now;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2283)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2284) return delta;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2285) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2286)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2287) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2288) * Determine the preferred nid for a task in a numa_group. This needs to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2289) * be done in a way that produces consistent results with group_weight,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2290) * otherwise workloads might not converge.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2291) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2292) static int preferred_group_nid(struct task_struct *p, int nid)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2293) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2294) nodemask_t nodes;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2295) int dist;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2296)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2297) /* Direct connections between all NUMA nodes. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2298) if (sched_numa_topology_type == NUMA_DIRECT)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2299) return nid;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2300)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2301) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2302) * On a system with glueless mesh NUMA topology, group_weight
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2303) * scores nodes according to the number of NUMA hinting faults on
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2304) * both the node itself, and on nearby nodes.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2305) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2306) if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2307) unsigned long score, max_score = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2308) int node, max_node = nid;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2309)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2310) dist = sched_max_numa_distance;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2311)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2312) for_each_online_node(node) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2313) score = group_weight(p, node, dist);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2314) if (score > max_score) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2315) max_score = score;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2316) max_node = node;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2317) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2318) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2319) return max_node;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2320) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2321)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2322) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2323) * Finding the preferred nid in a system with NUMA backplane
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2324) * interconnect topology is more involved. The goal is to locate
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2325) * tasks from numa_groups near each other in the system, and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2326) * untangle workloads from different sides of the system. This requires
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2327) * searching down the hierarchy of node groups, recursively searching
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2328) * inside the highest scoring group of nodes. The nodemask tricks
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2329) * keep the complexity of the search down.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2330) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2331) nodes = node_online_map;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2332) for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2333) unsigned long max_faults = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2334) nodemask_t max_group = NODE_MASK_NONE;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2335) int a, b;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2336)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2337) /* Are there nodes at this distance from each other? */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2338) if (!find_numa_distance(dist))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2339) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2340)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2341) for_each_node_mask(a, nodes) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2342) unsigned long faults = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2343) nodemask_t this_group;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2344) nodes_clear(this_group);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2345)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2346) /* Sum group's NUMA faults; includes a==b case. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2347) for_each_node_mask(b, nodes) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2348) if (node_distance(a, b) < dist) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2349) faults += group_faults(p, b);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2350) node_set(b, this_group);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2351) node_clear(b, nodes);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2352) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2353) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2354)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2355) /* Remember the top group. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2356) if (faults > max_faults) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2357) max_faults = faults;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2358) max_group = this_group;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2359) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2360) * subtle: at the smallest distance there is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2361) * just one node left in each "group", the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2362) * winner is the preferred nid.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2363) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2364) nid = a;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2365) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2366) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2367) /* Next round, evaluate the nodes within max_group. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2368) if (!max_faults)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2369) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2370) nodes = max_group;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2371) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2372) return nid;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2373) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2374)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2375) static void task_numa_placement(struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2376) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2377) int seq, nid, max_nid = NUMA_NO_NODE;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2378) unsigned long max_faults = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2379) unsigned long fault_types[2] = { 0, 0 };
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2380) unsigned long total_faults;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2381) u64 runtime, period;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2382) spinlock_t *group_lock = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2383) struct numa_group *ng;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2384)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2385) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2386) * The p->mm->numa_scan_seq field gets updated without
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2387) * exclusive access. Use READ_ONCE() here to ensure
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2388) * that the field is read in a single access:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2389) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2390) seq = READ_ONCE(p->mm->numa_scan_seq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2391) if (p->numa_scan_seq == seq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2392) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2393) p->numa_scan_seq = seq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2394) p->numa_scan_period_max = task_scan_max(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2395)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2396) total_faults = p->numa_faults_locality[0] +
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2397) p->numa_faults_locality[1];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2398) runtime = numa_get_avg_runtime(p, &period);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2399)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2400) /* If the task is part of a group prevent parallel updates to group stats */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2401) ng = deref_curr_numa_group(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2402) if (ng) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2403) group_lock = &ng->lock;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2404) spin_lock_irq(group_lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2405) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2406)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2407) /* Find the node with the highest number of faults */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2408) for_each_online_node(nid) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2409) /* Keep track of the offsets in numa_faults array */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2410) int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2411) unsigned long faults = 0, group_faults = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2412) int priv;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2413)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2414) for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2415) long diff, f_diff, f_weight;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2416)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2417) mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2418) membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2419) cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2420) cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2421)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2422) /* Decay existing window, copy faults since last scan */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2423) diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2424) fault_types[priv] += p->numa_faults[membuf_idx];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2425) p->numa_faults[membuf_idx] = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2426)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2427) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2428) * Normalize the faults_from, so all tasks in a group
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2429) * count according to CPU use, instead of by the raw
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2430) * number of faults. Tasks with little runtime have
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2431) * little over-all impact on throughput, and thus their
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2432) * faults are less important.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2433) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2434) f_weight = div64_u64(runtime << 16, period + 1);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2435) f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2436) (total_faults + 1);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2437) f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2438) p->numa_faults[cpubuf_idx] = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2439)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2440) p->numa_faults[mem_idx] += diff;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2441) p->numa_faults[cpu_idx] += f_diff;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2442) faults += p->numa_faults[mem_idx];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2443) p->total_numa_faults += diff;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2444) if (ng) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2445) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2446) * safe because we can only change our own group
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2447) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2448) * mem_idx represents the offset for a given
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2449) * nid and priv in a specific region because it
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2450) * is at the beginning of the numa_faults array.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2451) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2452) ng->faults[mem_idx] += diff;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2453) ng->faults_cpu[mem_idx] += f_diff;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2454) ng->total_faults += diff;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2455) group_faults += ng->faults[mem_idx];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2456) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2457) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2458)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2459) if (!ng) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2460) if (faults > max_faults) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2461) max_faults = faults;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2462) max_nid = nid;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2463) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2464) } else if (group_faults > max_faults) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2465) max_faults = group_faults;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2466) max_nid = nid;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2467) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2468) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2469)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2470) if (ng) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2471) numa_group_count_active_nodes(ng);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2472) spin_unlock_irq(group_lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2473) max_nid = preferred_group_nid(p, max_nid);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2474) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2475)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2476) if (max_faults) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2477) /* Set the new preferred node */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2478) if (max_nid != p->numa_preferred_nid)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2479) sched_setnuma(p, max_nid);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2480) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2481)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2482) update_task_scan_period(p, fault_types[0], fault_types[1]);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2483) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2484)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2485) static inline int get_numa_group(struct numa_group *grp)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2486) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2487) return refcount_inc_not_zero(&grp->refcount);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2488) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2489)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2490) static inline void put_numa_group(struct numa_group *grp)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2491) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2492) if (refcount_dec_and_test(&grp->refcount))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2493) kfree_rcu(grp, rcu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2494) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2495)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2496) static void task_numa_group(struct task_struct *p, int cpupid, int flags,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2497) int *priv)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2498) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2499) struct numa_group *grp, *my_grp;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2500) struct task_struct *tsk;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2501) bool join = false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2502) int cpu = cpupid_to_cpu(cpupid);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2503) int i;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2504)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2505) if (unlikely(!deref_curr_numa_group(p))) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2506) unsigned int size = sizeof(struct numa_group) +
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2507) 4*nr_node_ids*sizeof(unsigned long);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2508)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2509) grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2510) if (!grp)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2511) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2512)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2513) refcount_set(&grp->refcount, 1);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2514) grp->active_nodes = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2515) grp->max_faults_cpu = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2516) spin_lock_init(&grp->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2517) grp->gid = p->pid;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2518) /* Second half of the array tracks nids where faults happen */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2519) grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2520) nr_node_ids;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2521)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2522) for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2523) grp->faults[i] = p->numa_faults[i];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2524)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2525) grp->total_faults = p->total_numa_faults;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2526)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2527) grp->nr_tasks++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2528) rcu_assign_pointer(p->numa_group, grp);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2529) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2530)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2531) rcu_read_lock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2532) tsk = READ_ONCE(cpu_rq(cpu)->curr);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2533)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2534) if (!cpupid_match_pid(tsk, cpupid))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2535) goto no_join;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2536)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2537) grp = rcu_dereference(tsk->numa_group);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2538) if (!grp)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2539) goto no_join;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2540)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2541) my_grp = deref_curr_numa_group(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2542) if (grp == my_grp)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2543) goto no_join;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2544)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2545) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2546) * Only join the other group if its bigger; if we're the bigger group,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2547) * the other task will join us.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2548) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2549) if (my_grp->nr_tasks > grp->nr_tasks)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2550) goto no_join;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2551)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2552) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2553) * Tie-break on the grp address.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2554) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2555) if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2556) goto no_join;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2557)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2558) /* Always join threads in the same process. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2559) if (tsk->mm == current->mm)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2560) join = true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2561)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2562) /* Simple filter to avoid false positives due to PID collisions */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2563) if (flags & TNF_SHARED)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2564) join = true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2565)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2566) /* Update priv based on whether false sharing was detected */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2567) *priv = !join;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2568)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2569) if (join && !get_numa_group(grp))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2570) goto no_join;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2571)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2572) rcu_read_unlock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2573)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2574) if (!join)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2575) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2576)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2577) BUG_ON(irqs_disabled());
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2578) double_lock_irq(&my_grp->lock, &grp->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2579)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2580) for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2581) my_grp->faults[i] -= p->numa_faults[i];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2582) grp->faults[i] += p->numa_faults[i];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2583) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2584) my_grp->total_faults -= p->total_numa_faults;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2585) grp->total_faults += p->total_numa_faults;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2586)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2587) my_grp->nr_tasks--;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2588) grp->nr_tasks++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2589)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2590) spin_unlock(&my_grp->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2591) spin_unlock_irq(&grp->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2592)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2593) rcu_assign_pointer(p->numa_group, grp);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2594)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2595) put_numa_group(my_grp);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2596) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2597)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2598) no_join:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2599) rcu_read_unlock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2600) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2601) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2602)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2603) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2604) * Get rid of NUMA staticstics associated with a task (either current or dead).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2605) * If @final is set, the task is dead and has reached refcount zero, so we can
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2606) * safely free all relevant data structures. Otherwise, there might be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2607) * concurrent reads from places like load balancing and procfs, and we should
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2608) * reset the data back to default state without freeing ->numa_faults.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2609) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2610) void task_numa_free(struct task_struct *p, bool final)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2611) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2612) /* safe: p either is current or is being freed by current */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2613) struct numa_group *grp = rcu_dereference_raw(p->numa_group);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2614) unsigned long *numa_faults = p->numa_faults;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2615) unsigned long flags;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2616) int i;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2617)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2618) if (!numa_faults)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2619) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2620)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2621) if (grp) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2622) spin_lock_irqsave(&grp->lock, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2623) for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2624) grp->faults[i] -= p->numa_faults[i];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2625) grp->total_faults -= p->total_numa_faults;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2626)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2627) grp->nr_tasks--;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2628) spin_unlock_irqrestore(&grp->lock, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2629) RCU_INIT_POINTER(p->numa_group, NULL);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2630) put_numa_group(grp);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2631) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2632)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2633) if (final) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2634) p->numa_faults = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2635) kfree(numa_faults);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2636) } else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2637) p->total_numa_faults = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2638) for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2639) numa_faults[i] = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2640) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2641) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2642)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2643) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2644) * Got a PROT_NONE fault for a page on @node.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2645) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2646) void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2647) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2648) struct task_struct *p = current;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2649) bool migrated = flags & TNF_MIGRATED;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2650) int cpu_node = task_node(current);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2651) int local = !!(flags & TNF_FAULT_LOCAL);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2652) struct numa_group *ng;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2653) int priv;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2654)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2655) if (!static_branch_likely(&sched_numa_balancing))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2656) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2657)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2658) /* for example, ksmd faulting in a user's mm */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2659) if (!p->mm)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2660) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2661)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2662) /* Allocate buffer to track faults on a per-node basis */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2663) if (unlikely(!p->numa_faults)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2664) int size = sizeof(*p->numa_faults) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2665) NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2666)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2667) p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2668) if (!p->numa_faults)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2669) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2670)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2671) p->total_numa_faults = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2672) memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2673) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2674)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2675) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2676) * First accesses are treated as private, otherwise consider accesses
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2677) * to be private if the accessing pid has not changed
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2678) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2679) if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2680) priv = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2681) } else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2682) priv = cpupid_match_pid(p, last_cpupid);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2683) if (!priv && !(flags & TNF_NO_GROUP))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2684) task_numa_group(p, last_cpupid, flags, &priv);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2685) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2686)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2687) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2688) * If a workload spans multiple NUMA nodes, a shared fault that
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2689) * occurs wholly within the set of nodes that the workload is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2690) * actively using should be counted as local. This allows the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2691) * scan rate to slow down when a workload has settled down.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2692) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2693) ng = deref_curr_numa_group(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2694) if (!priv && !local && ng && ng->active_nodes > 1 &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2695) numa_is_active_node(cpu_node, ng) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2696) numa_is_active_node(mem_node, ng))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2697) local = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2698)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2699) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2700) * Retry to migrate task to preferred node periodically, in case it
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2701) * previously failed, or the scheduler moved us.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2702) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2703) if (time_after(jiffies, p->numa_migrate_retry)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2704) task_numa_placement(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2705) numa_migrate_preferred(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2706) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2707)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2708) if (migrated)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2709) p->numa_pages_migrated += pages;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2710) if (flags & TNF_MIGRATE_FAIL)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2711) p->numa_faults_locality[2] += pages;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2712)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2713) p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2714) p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2715) p->numa_faults_locality[local] += pages;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2716) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2717)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2718) static void reset_ptenuma_scan(struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2719) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2720) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2721) * We only did a read acquisition of the mmap sem, so
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2722) * p->mm->numa_scan_seq is written to without exclusive access
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2723) * and the update is not guaranteed to be atomic. That's not
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2724) * much of an issue though, since this is just used for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2725) * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2726) * expensive, to avoid any form of compiler optimizations:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2727) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2728) WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2729) p->mm->numa_scan_offset = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2730) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2731)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2732) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2733) * The expensive part of numa migration is done from task_work context.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2734) * Triggered from task_tick_numa().
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2735) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2736) static void task_numa_work(struct callback_head *work)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2737) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2738) unsigned long migrate, next_scan, now = jiffies;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2739) struct task_struct *p = current;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2740) struct mm_struct *mm = p->mm;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2741) u64 runtime = p->se.sum_exec_runtime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2742) struct vm_area_struct *vma;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2743) unsigned long start, end;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2744) unsigned long nr_pte_updates = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2745) long pages, virtpages;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2746)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2747) SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2748)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2749) work->next = work;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2750) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2751) * Who cares about NUMA placement when they're dying.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2752) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2753) * NOTE: make sure not to dereference p->mm before this check,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2754) * exit_task_work() happens _after_ exit_mm() so we could be called
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2755) * without p->mm even though we still had it when we enqueued this
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2756) * work.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2757) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2758) if (p->flags & PF_EXITING)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2759) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2760)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2761) if (!mm->numa_next_scan) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2762) mm->numa_next_scan = now +
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2763) msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2764) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2765)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2766) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2767) * Enforce maximal scan/migration frequency..
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2768) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2769) migrate = mm->numa_next_scan;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2770) if (time_before(now, migrate))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2771) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2772)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2773) if (p->numa_scan_period == 0) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2774) p->numa_scan_period_max = task_scan_max(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2775) p->numa_scan_period = task_scan_start(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2776) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2777)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2778) next_scan = now + msecs_to_jiffies(p->numa_scan_period);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2779) if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2780) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2781)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2782) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2783) * Delay this task enough that another task of this mm will likely win
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2784) * the next time around.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2785) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2786) p->node_stamp += 2 * TICK_NSEC;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2787)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2788) start = mm->numa_scan_offset;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2789) pages = sysctl_numa_balancing_scan_size;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2790) pages <<= 20 - PAGE_SHIFT; /* MB in pages */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2791) virtpages = pages * 8; /* Scan up to this much virtual space */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2792) if (!pages)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2793) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2794)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2795)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2796) if (!mmap_read_trylock(mm))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2797) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2798) vma = find_vma(mm, start);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2799) if (!vma) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2800) reset_ptenuma_scan(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2801) start = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2802) vma = mm->mmap;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2803) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2804) for (; vma; vma = vma->vm_next) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2805) if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2806) is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2807) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2808) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2809)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2810) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2811) * Shared library pages mapped by multiple processes are not
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2812) * migrated as it is expected they are cache replicated. Avoid
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2813) * hinting faults in read-only file-backed mappings or the vdso
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2814) * as migrating the pages will be of marginal benefit.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2815) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2816) if (!vma->vm_mm ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2817) (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2818) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2819)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2820) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2821) * Skip inaccessible VMAs to avoid any confusion between
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2822) * PROT_NONE and NUMA hinting ptes
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2823) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2824) if (!vma_is_accessible(vma))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2825) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2826)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2827) do {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2828) start = max(start, vma->vm_start);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2829) end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2830) end = min(end, vma->vm_end);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2831) nr_pte_updates = change_prot_numa(vma, start, end);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2832)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2833) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2834) * Try to scan sysctl_numa_balancing_size worth of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2835) * hpages that have at least one present PTE that
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2836) * is not already pte-numa. If the VMA contains
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2837) * areas that are unused or already full of prot_numa
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2838) * PTEs, scan up to virtpages, to skip through those
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2839) * areas faster.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2840) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2841) if (nr_pte_updates)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2842) pages -= (end - start) >> PAGE_SHIFT;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2843) virtpages -= (end - start) >> PAGE_SHIFT;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2844)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2845) start = end;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2846) if (pages <= 0 || virtpages <= 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2847) goto out;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2848)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2849) cond_resched();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2850) } while (end != vma->vm_end);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2851) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2852)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2853) out:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2854) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2855) * It is possible to reach the end of the VMA list but the last few
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2856) * VMAs are not guaranteed to the vma_migratable. If they are not, we
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2857) * would find the !migratable VMA on the next scan but not reset the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2858) * scanner to the start so check it now.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2859) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2860) if (vma)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2861) mm->numa_scan_offset = start;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2862) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2863) reset_ptenuma_scan(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2864) mmap_read_unlock(mm);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2865)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2866) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2867) * Make sure tasks use at least 32x as much time to run other code
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2868) * than they used here, to limit NUMA PTE scanning overhead to 3% max.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2869) * Usually update_task_scan_period slows down scanning enough; on an
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2870) * overloaded system we need to limit overhead on a per task basis.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2871) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2872) if (unlikely(p->se.sum_exec_runtime != runtime)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2873) u64 diff = p->se.sum_exec_runtime - runtime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2874) p->node_stamp += 32 * diff;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2875) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2876) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2877)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2878) void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2879) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2880) int mm_users = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2881) struct mm_struct *mm = p->mm;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2882)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2883) if (mm) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2884) mm_users = atomic_read(&mm->mm_users);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2885) if (mm_users == 1) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2886) mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2887) mm->numa_scan_seq = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2888) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2889) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2890) p->node_stamp = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2891) p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2892) p->numa_scan_period = sysctl_numa_balancing_scan_delay;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2893) /* Protect against double add, see task_tick_numa and task_numa_work */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2894) p->numa_work.next = &p->numa_work;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2895) p->numa_faults = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2896) RCU_INIT_POINTER(p->numa_group, NULL);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2897) p->last_task_numa_placement = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2898) p->last_sum_exec_runtime = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2899)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2900) init_task_work(&p->numa_work, task_numa_work);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2901)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2902) /* New address space, reset the preferred nid */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2903) if (!(clone_flags & CLONE_VM)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2904) p->numa_preferred_nid = NUMA_NO_NODE;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2905) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2906) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2907)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2908) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2909) * New thread, keep existing numa_preferred_nid which should be copied
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2910) * already by arch_dup_task_struct but stagger when scans start.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2911) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2912) if (mm) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2913) unsigned int delay;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2914)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2915) delay = min_t(unsigned int, task_scan_max(current),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2916) current->numa_scan_period * mm_users * NSEC_PER_MSEC);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2917) delay += 2 * TICK_NSEC;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2918) p->node_stamp = delay;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2919) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2920) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2921)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2922) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2923) * Drive the periodic memory faults..
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2924) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2925) static void task_tick_numa(struct rq *rq, struct task_struct *curr)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2926) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2927) struct callback_head *work = &curr->numa_work;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2928) u64 period, now;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2929)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2930) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2931) * We don't care about NUMA placement if we don't have memory.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2932) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2933) if ((curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2934) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2935)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2936) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2937) * Using runtime rather than walltime has the dual advantage that
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2938) * we (mostly) drive the selection from busy threads and that the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2939) * task needs to have done some actual work before we bother with
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2940) * NUMA placement.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2941) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2942) now = curr->se.sum_exec_runtime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2943) period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2944)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2945) if (now > curr->node_stamp + period) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2946) if (!curr->node_stamp)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2947) curr->numa_scan_period = task_scan_start(curr);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2948) curr->node_stamp += period;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2949)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2950) if (!time_before(jiffies, curr->mm->numa_next_scan))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2951) task_work_add(curr, work, TWA_RESUME);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2952) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2953) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2954)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2955) static void update_scan_period(struct task_struct *p, int new_cpu)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2956) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2957) int src_nid = cpu_to_node(task_cpu(p));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2958) int dst_nid = cpu_to_node(new_cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2959)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2960) if (!static_branch_likely(&sched_numa_balancing))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2961) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2962)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2963) if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2964) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2965)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2966) if (src_nid == dst_nid)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2967) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2968)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2969) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2970) * Allow resets if faults have been trapped before one scan
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2971) * has completed. This is most likely due to a new task that
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2972) * is pulled cross-node due to wakeups or load balancing.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2973) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2974) if (p->numa_scan_seq) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2975) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2976) * Avoid scan adjustments if moving to the preferred
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2977) * node or if the task was not previously running on
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2978) * the preferred node.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2979) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2980) if (dst_nid == p->numa_preferred_nid ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2981) (p->numa_preferred_nid != NUMA_NO_NODE &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2982) src_nid != p->numa_preferred_nid))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2983) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2984) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2985)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2986) p->numa_scan_period = task_scan_start(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2987) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2988)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2989) #else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2990) static void task_tick_numa(struct rq *rq, struct task_struct *curr)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2991) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2992) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2993)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2994) static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2995) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2996) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2997)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2998) static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2999) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3000) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3001)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3002) static inline void update_scan_period(struct task_struct *p, int new_cpu)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3003) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3004) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3005)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3006) #endif /* CONFIG_NUMA_BALANCING */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3007)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3008) static void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3009) account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3010) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3011) update_load_add(&cfs_rq->load, se->load.weight);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3012) #ifdef CONFIG_SMP
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3013) if (entity_is_task(se)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3014) struct rq *rq = rq_of(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3015)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3016) account_numa_enqueue(rq, task_of(se));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3017) list_add(&se->group_node, &rq->cfs_tasks);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3018) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3019) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3020) cfs_rq->nr_running++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3021) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3022)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3023) static void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3024) account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3025) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3026) update_load_sub(&cfs_rq->load, se->load.weight);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3027) #ifdef CONFIG_SMP
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3028) if (entity_is_task(se)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3029) account_numa_dequeue(rq_of(cfs_rq), task_of(se));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3030) list_del_init(&se->group_node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3031) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3032) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3033) cfs_rq->nr_running--;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3034) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3035)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3036) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3037) * Signed add and clamp on underflow.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3038) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3039) * Explicitly do a load-store to ensure the intermediate value never hits
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3040) * memory. This allows lockless observations without ever seeing the negative
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3041) * values.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3042) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3043) #define add_positive(_ptr, _val) do { \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3044) typeof(_ptr) ptr = (_ptr); \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3045) typeof(_val) val = (_val); \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3046) typeof(*ptr) res, var = READ_ONCE(*ptr); \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3047) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3048) res = var + val; \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3049) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3050) if (val < 0 && res > var) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3051) res = 0; \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3052) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3053) WRITE_ONCE(*ptr, res); \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3054) } while (0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3055)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3056) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3057) * Unsigned subtract and clamp on underflow.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3058) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3059) * Explicitly do a load-store to ensure the intermediate value never hits
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3060) * memory. This allows lockless observations without ever seeing the negative
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3061) * values.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3062) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3063) #define sub_positive(_ptr, _val) do { \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3064) typeof(_ptr) ptr = (_ptr); \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3065) typeof(*ptr) val = (_val); \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3066) typeof(*ptr) res, var = READ_ONCE(*ptr); \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3067) res = var - val; \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3068) if (res > var) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3069) res = 0; \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3070) WRITE_ONCE(*ptr, res); \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3071) } while (0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3072)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3073) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3074) * Remove and clamp on negative, from a local variable.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3075) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3076) * A variant of sub_positive(), which does not use explicit load-store
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3077) * and is thus optimized for local variable updates.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3078) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3079) #define lsub_positive(_ptr, _val) do { \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3080) typeof(_ptr) ptr = (_ptr); \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3081) *ptr -= min_t(typeof(*ptr), *ptr, _val); \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3082) } while (0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3083)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3084) #ifdef CONFIG_SMP
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3085) static inline void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3086) enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3087) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3088) cfs_rq->avg.load_avg += se->avg.load_avg;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3089) cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3090) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3091)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3092) static inline void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3093) dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3094) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3095) sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3096) sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3097) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3098) #else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3099) static inline void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3100) enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3101) static inline void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3102) dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3103) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3104)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3105) static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3106) unsigned long weight)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3107) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3108) if (se->on_rq) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3109) /* commit outstanding execution time */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3110) if (cfs_rq->curr == se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3111) update_curr(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3112) update_load_sub(&cfs_rq->load, se->load.weight);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3113) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3114) dequeue_load_avg(cfs_rq, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3115)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3116) update_load_set(&se->load, weight);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3117)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3118) #ifdef CONFIG_SMP
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3119) do {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3120) u32 divider = get_pelt_divider(&se->avg);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3121)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3122) se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3123) } while (0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3124) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3125)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3126) enqueue_load_avg(cfs_rq, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3127) if (se->on_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3128) update_load_add(&cfs_rq->load, se->load.weight);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3129)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3130) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3131)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3132) void reweight_task(struct task_struct *p, int prio)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3133) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3134) struct sched_entity *se = &p->se;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3135) struct cfs_rq *cfs_rq = cfs_rq_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3136) struct load_weight *load = &se->load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3137) unsigned long weight = scale_load(sched_prio_to_weight[prio]);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3138)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3139) reweight_entity(cfs_rq, se, weight);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3140) load->inv_weight = sched_prio_to_wmult[prio];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3141) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3142)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3143) #ifdef CONFIG_FAIR_GROUP_SCHED
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3144) #ifdef CONFIG_SMP
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3145) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3146) * All this does is approximate the hierarchical proportion which includes that
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3147) * global sum we all love to hate.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3148) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3149) * That is, the weight of a group entity, is the proportional share of the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3150) * group weight based on the group runqueue weights. That is:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3151) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3152) * tg->weight * grq->load.weight
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3153) * ge->load.weight = ----------------------------- (1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3154) * \Sum grq->load.weight
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3155) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3156) * Now, because computing that sum is prohibitively expensive to compute (been
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3157) * there, done that) we approximate it with this average stuff. The average
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3158) * moves slower and therefore the approximation is cheaper and more stable.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3159) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3160) * So instead of the above, we substitute:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3161) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3162) * grq->load.weight -> grq->avg.load_avg (2)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3163) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3164) * which yields the following:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3165) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3166) * tg->weight * grq->avg.load_avg
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3167) * ge->load.weight = ------------------------------ (3)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3168) * tg->load_avg
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3169) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3170) * Where: tg->load_avg ~= \Sum grq->avg.load_avg
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3171) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3172) * That is shares_avg, and it is right (given the approximation (2)).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3173) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3174) * The problem with it is that because the average is slow -- it was designed
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3175) * to be exactly that of course -- this leads to transients in boundary
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3176) * conditions. In specific, the case where the group was idle and we start the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3177) * one task. It takes time for our CPU's grq->avg.load_avg to build up,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3178) * yielding bad latency etc..
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3179) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3180) * Now, in that special case (1) reduces to:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3181) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3182) * tg->weight * grq->load.weight
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3183) * ge->load.weight = ----------------------------- = tg->weight (4)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3184) * grp->load.weight
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3185) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3186) * That is, the sum collapses because all other CPUs are idle; the UP scenario.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3187) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3188) * So what we do is modify our approximation (3) to approach (4) in the (near)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3189) * UP case, like:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3190) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3191) * ge->load.weight =
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3192) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3193) * tg->weight * grq->load.weight
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3194) * --------------------------------------------------- (5)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3195) * tg->load_avg - grq->avg.load_avg + grq->load.weight
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3196) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3197) * But because grq->load.weight can drop to 0, resulting in a divide by zero,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3198) * we need to use grq->avg.load_avg as its lower bound, which then gives:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3199) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3200) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3201) * tg->weight * grq->load.weight
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3202) * ge->load.weight = ----------------------------- (6)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3203) * tg_load_avg'
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3204) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3205) * Where:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3206) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3207) * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3208) * max(grq->load.weight, grq->avg.load_avg)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3209) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3210) * And that is shares_weight and is icky. In the (near) UP case it approaches
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3211) * (4) while in the normal case it approaches (3). It consistently
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3212) * overestimates the ge->load.weight and therefore:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3213) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3214) * \Sum ge->load.weight >= tg->weight
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3215) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3216) * hence icky!
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3217) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3218) static long calc_group_shares(struct cfs_rq *cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3219) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3220) long tg_weight, tg_shares, load, shares;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3221) struct task_group *tg = cfs_rq->tg;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3222)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3223) tg_shares = READ_ONCE(tg->shares);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3224)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3225) load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3226)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3227) tg_weight = atomic_long_read(&tg->load_avg);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3228)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3229) /* Ensure tg_weight >= load */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3230) tg_weight -= cfs_rq->tg_load_avg_contrib;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3231) tg_weight += load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3232)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3233) shares = (tg_shares * load);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3234) if (tg_weight)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3235) shares /= tg_weight;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3236)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3237) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3238) * MIN_SHARES has to be unscaled here to support per-CPU partitioning
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3239) * of a group with small tg->shares value. It is a floor value which is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3240) * assigned as a minimum load.weight to the sched_entity representing
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3241) * the group on a CPU.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3242) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3243) * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3244) * on an 8-core system with 8 tasks each runnable on one CPU shares has
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3245) * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3246) * case no task is runnable on a CPU MIN_SHARES=2 should be returned
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3247) * instead of 0.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3248) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3249) return clamp_t(long, shares, MIN_SHARES, tg_shares);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3250) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3251) #endif /* CONFIG_SMP */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3252)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3253) static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3254)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3255) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3256) * Recomputes the group entity based on the current state of its group
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3257) * runqueue.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3258) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3259) static void update_cfs_group(struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3260) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3261) struct cfs_rq *gcfs_rq = group_cfs_rq(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3262) long shares;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3263)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3264) if (!gcfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3265) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3266)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3267) if (throttled_hierarchy(gcfs_rq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3268) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3269)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3270) #ifndef CONFIG_SMP
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3271) shares = READ_ONCE(gcfs_rq->tg->shares);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3272)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3273) if (likely(se->load.weight == shares))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3274) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3275) #else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3276) shares = calc_group_shares(gcfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3277) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3278)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3279) reweight_entity(cfs_rq_of(se), se, shares);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3280) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3281)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3282) #else /* CONFIG_FAIR_GROUP_SCHED */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3283) static inline void update_cfs_group(struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3284) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3285) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3286) #endif /* CONFIG_FAIR_GROUP_SCHED */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3287)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3288) static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3289) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3290) struct rq *rq = rq_of(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3291)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3292) if (&rq->cfs == cfs_rq) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3293) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3294) * There are a few boundary cases this might miss but it should
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3295) * get called often enough that that should (hopefully) not be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3296) * a real problem.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3297) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3298) * It will not get called when we go idle, because the idle
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3299) * thread is a different class (!fair), nor will the utilization
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3300) * number include things like RT tasks.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3301) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3302) * As is, the util number is not freq-invariant (we'd have to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3303) * implement arch_scale_freq_capacity() for that).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3304) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3305) * See cpu_util().
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3306) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3307) cpufreq_update_util(rq, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3308) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3309) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3310)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3311) #ifdef CONFIG_SMP
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3312) #ifdef CONFIG_FAIR_GROUP_SCHED
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3313) /**
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3314) * update_tg_load_avg - update the tg's load avg
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3315) * @cfs_rq: the cfs_rq whose avg changed
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3316) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3317) * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3318) * However, because tg->load_avg is a global value there are performance
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3319) * considerations.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3320) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3321) * In order to avoid having to look at the other cfs_rq's, we use a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3322) * differential update where we store the last value we propagated. This in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3323) * turn allows skipping updates if the differential is 'small'.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3324) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3325) * Updating tg's load_avg is necessary before update_cfs_share().
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3326) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3327) static inline void update_tg_load_avg(struct cfs_rq *cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3328) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3329) long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3330)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3331) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3332) * No need to update load_avg for root_task_group as it is not used.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3333) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3334) if (cfs_rq->tg == &root_task_group)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3335) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3336)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3337) if (abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3338) atomic_long_add(delta, &cfs_rq->tg->load_avg);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3339) cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3340) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3341) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3342)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3343) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3344) * Called within set_task_rq() right before setting a task's CPU. The
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3345) * caller only guarantees p->pi_lock is held; no other assumptions,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3346) * including the state of rq->lock, should be made.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3347) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3348) void set_task_rq_fair(struct sched_entity *se,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3349) struct cfs_rq *prev, struct cfs_rq *next)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3350) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3351) u64 p_last_update_time;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3352) u64 n_last_update_time;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3353)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3354) if (!sched_feat(ATTACH_AGE_LOAD))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3355) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3356)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3357) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3358) * We are supposed to update the task to "current" time, then its up to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3359) * date and ready to go to new CPU/cfs_rq. But we have difficulty in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3360) * getting what current time is, so simply throw away the out-of-date
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3361) * time. This will result in the wakee task is less decayed, but giving
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3362) * the wakee more load sounds not bad.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3363) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3364) if (!(se->avg.last_update_time && prev))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3365) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3366)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3367) #ifndef CONFIG_64BIT
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3368) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3369) u64 p_last_update_time_copy;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3370) u64 n_last_update_time_copy;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3371)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3372) do {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3373) p_last_update_time_copy = prev->load_last_update_time_copy;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3374) n_last_update_time_copy = next->load_last_update_time_copy;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3375)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3376) smp_rmb();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3377)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3378) p_last_update_time = prev->avg.last_update_time;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3379) n_last_update_time = next->avg.last_update_time;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3380)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3381) } while (p_last_update_time != p_last_update_time_copy ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3382) n_last_update_time != n_last_update_time_copy);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3383) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3384) #else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3385) p_last_update_time = prev->avg.last_update_time;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3386) n_last_update_time = next->avg.last_update_time;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3387) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3388) __update_load_avg_blocked_se(p_last_update_time, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3389) se->avg.last_update_time = n_last_update_time;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3390) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3391)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3392) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3393) * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3394) * propagate its contribution. The key to this propagation is the invariant
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3395) * that for each group:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3396) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3397) * ge->avg == grq->avg (1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3398) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3399) * _IFF_ we look at the pure running and runnable sums. Because they
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3400) * represent the very same entity, just at different points in the hierarchy.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3401) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3402) * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3403) * and simply copies the running/runnable sum over (but still wrong, because
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3404) * the group entity and group rq do not have their PELT windows aligned).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3405) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3406) * However, update_tg_cfs_load() is more complex. So we have:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3407) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3408) * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3409) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3410) * And since, like util, the runnable part should be directly transferable,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3411) * the following would _appear_ to be the straight forward approach:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3412) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3413) * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3414) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3415) * And per (1) we have:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3416) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3417) * ge->avg.runnable_avg == grq->avg.runnable_avg
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3418) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3419) * Which gives:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3420) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3421) * ge->load.weight * grq->avg.load_avg
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3422) * ge->avg.load_avg = ----------------------------------- (4)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3423) * grq->load.weight
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3424) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3425) * Except that is wrong!
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3426) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3427) * Because while for entities historical weight is not important and we
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3428) * really only care about our future and therefore can consider a pure
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3429) * runnable sum, runqueues can NOT do this.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3430) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3431) * We specifically want runqueues to have a load_avg that includes
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3432) * historical weights. Those represent the blocked load, the load we expect
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3433) * to (shortly) return to us. This only works by keeping the weights as
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3434) * integral part of the sum. We therefore cannot decompose as per (3).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3435) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3436) * Another reason this doesn't work is that runnable isn't a 0-sum entity.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3437) * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3438) * rq itself is runnable anywhere between 2/3 and 1 depending on how the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3439) * runnable section of these tasks overlap (or not). If they were to perfectly
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3440) * align the rq as a whole would be runnable 2/3 of the time. If however we
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3441) * always have at least 1 runnable task, the rq as a whole is always runnable.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3442) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3443) * So we'll have to approximate.. :/
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3444) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3445) * Given the constraint:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3446) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3447) * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3448) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3449) * We can construct a rule that adds runnable to a rq by assuming minimal
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3450) * overlap.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3451) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3452) * On removal, we'll assume each task is equally runnable; which yields:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3453) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3454) * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3455) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3456) * XXX: only do this for the part of runnable > running ?
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3457) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3458) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3459) static inline void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3460) update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3461) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3462) long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3463) u32 divider;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3464)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3465) /* Nothing to update */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3466) if (!delta)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3467) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3468)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3469) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3470) * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3471) * See ___update_load_avg() for details.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3472) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3473) divider = get_pelt_divider(&cfs_rq->avg);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3474)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3475) /* Set new sched_entity's utilization */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3476) se->avg.util_avg = gcfs_rq->avg.util_avg;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3477) se->avg.util_sum = se->avg.util_avg * divider;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3478)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3479) /* Update parent cfs_rq utilization */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3480) add_positive(&cfs_rq->avg.util_avg, delta);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3481) cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * divider;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3482) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3483)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3484) static inline void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3485) update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3486) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3487) long delta = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3488) u32 divider;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3489)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3490) /* Nothing to update */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3491) if (!delta)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3492) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3493)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3494) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3495) * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3496) * See ___update_load_avg() for details.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3497) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3498) divider = get_pelt_divider(&cfs_rq->avg);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3499)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3500) /* Set new sched_entity's runnable */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3501) se->avg.runnable_avg = gcfs_rq->avg.runnable_avg;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3502) se->avg.runnable_sum = se->avg.runnable_avg * divider;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3503)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3504) /* Update parent cfs_rq runnable */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3505) add_positive(&cfs_rq->avg.runnable_avg, delta);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3506) cfs_rq->avg.runnable_sum = cfs_rq->avg.runnable_avg * divider;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3507) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3508)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3509) static inline void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3510) update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3511) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3512) long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3513) unsigned long load_avg;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3514) u64 load_sum = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3515) s64 delta_sum;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3516) u32 divider;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3517)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3518) if (!runnable_sum)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3519) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3520)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3521) gcfs_rq->prop_runnable_sum = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3522)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3523) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3524) * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3525) * See ___update_load_avg() for details.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3526) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3527) divider = get_pelt_divider(&cfs_rq->avg);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3528)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3529) if (runnable_sum >= 0) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3530) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3531) * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3532) * the CPU is saturated running == runnable.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3533) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3534) runnable_sum += se->avg.load_sum;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3535) runnable_sum = min_t(long, runnable_sum, divider);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3536) } else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3537) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3538) * Estimate the new unweighted runnable_sum of the gcfs_rq by
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3539) * assuming all tasks are equally runnable.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3540) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3541) if (scale_load_down(gcfs_rq->load.weight)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3542) load_sum = div_s64(gcfs_rq->avg.load_sum,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3543) scale_load_down(gcfs_rq->load.weight));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3544) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3545)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3546) /* But make sure to not inflate se's runnable */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3547) runnable_sum = min(se->avg.load_sum, load_sum);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3548) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3549)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3550) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3551) * runnable_sum can't be lower than running_sum
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3552) * Rescale running sum to be in the same range as runnable sum
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3553) * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3554) * runnable_sum is in [0 : LOAD_AVG_MAX]
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3555) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3556) running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3557) runnable_sum = max(runnable_sum, running_sum);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3558)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3559) load_sum = (s64)se_weight(se) * runnable_sum;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3560) load_avg = div_s64(load_sum, divider);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3561)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3562) delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3563) delta_avg = load_avg - se->avg.load_avg;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3564)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3565) se->avg.load_sum = runnable_sum;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3566) se->avg.load_avg = load_avg;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3567) add_positive(&cfs_rq->avg.load_avg, delta_avg);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3568) add_positive(&cfs_rq->avg.load_sum, delta_sum);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3569) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3570)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3571) static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3572) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3573) cfs_rq->propagate = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3574) cfs_rq->prop_runnable_sum += runnable_sum;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3575) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3576)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3577) /* Update task and its cfs_rq load average */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3578) static inline int propagate_entity_load_avg(struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3579) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3580) struct cfs_rq *cfs_rq, *gcfs_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3581)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3582) if (entity_is_task(se))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3583) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3584)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3585) gcfs_rq = group_cfs_rq(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3586) if (!gcfs_rq->propagate)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3587) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3588)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3589) gcfs_rq->propagate = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3590)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3591) cfs_rq = cfs_rq_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3592)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3593) add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3594)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3595) update_tg_cfs_util(cfs_rq, se, gcfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3596) update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3597) update_tg_cfs_load(cfs_rq, se, gcfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3598)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3599) trace_pelt_cfs_tp(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3600) trace_pelt_se_tp(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3601)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3602) return 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3603) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3604)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3605) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3606) * Check if we need to update the load and the utilization of a blocked
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3607) * group_entity:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3608) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3609) static inline bool skip_blocked_update(struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3610) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3611) struct cfs_rq *gcfs_rq = group_cfs_rq(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3612)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3613) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3614) * If sched_entity still have not zero load or utilization, we have to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3615) * decay it:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3616) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3617) if (se->avg.load_avg || se->avg.util_avg)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3618) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3619)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3620) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3621) * If there is a pending propagation, we have to update the load and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3622) * the utilization of the sched_entity:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3623) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3624) if (gcfs_rq->propagate)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3625) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3626)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3627) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3628) * Otherwise, the load and the utilization of the sched_entity is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3629) * already zero and there is no pending propagation, so it will be a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3630) * waste of time to try to decay it:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3631) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3632) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3633) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3634)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3635) #else /* CONFIG_FAIR_GROUP_SCHED */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3636)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3637) static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) {}
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3638)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3639) static inline int propagate_entity_load_avg(struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3640) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3641) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3642) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3643)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3644) static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3645)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3646) #endif /* CONFIG_FAIR_GROUP_SCHED */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3647)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3648) /**
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3649) * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3650) * @now: current time, as per cfs_rq_clock_pelt()
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3651) * @cfs_rq: cfs_rq to update
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3652) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3653) * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3654) * avg. The immediate corollary is that all (fair) tasks must be attached, see
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3655) * post_init_entity_util_avg().
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3656) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3657) * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3658) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3659) * Returns true if the load decayed or we removed load.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3660) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3661) * Since both these conditions indicate a changed cfs_rq->avg.load we should
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3662) * call update_tg_load_avg() when this function returns true.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3663) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3664) static inline int
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3665) update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3666) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3667) unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3668) struct sched_avg *sa = &cfs_rq->avg;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3669) int decayed = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3670)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3671) if (cfs_rq->removed.nr) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3672) unsigned long r;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3673) u32 divider = get_pelt_divider(&cfs_rq->avg);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3674)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3675) raw_spin_lock(&cfs_rq->removed.lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3676) swap(cfs_rq->removed.util_avg, removed_util);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3677) swap(cfs_rq->removed.load_avg, removed_load);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3678) swap(cfs_rq->removed.runnable_avg, removed_runnable);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3679) cfs_rq->removed.nr = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3680) raw_spin_unlock(&cfs_rq->removed.lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3681)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3682) r = removed_load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3683) sub_positive(&sa->load_avg, r);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3684) sub_positive(&sa->load_sum, r * divider);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3685)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3686) r = removed_util;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3687) sub_positive(&sa->util_avg, r);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3688) sub_positive(&sa->util_sum, r * divider);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3689) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3690) * Because of rounding, se->util_sum might ends up being +1 more than
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3691) * cfs->util_sum. Although this is not a problem by itself, detaching
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3692) * a lot of tasks with the rounding problem between 2 updates of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3693) * util_avg (~1ms) can make cfs->util_sum becoming null whereas
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3694) * cfs_util_avg is not.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3695) * Check that util_sum is still above its lower bound for the new
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3696) * util_avg. Given that period_contrib might have moved since the last
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3697) * sync, we are only sure that util_sum must be above or equal to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3698) * util_avg * minimum possible divider
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3699) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3700) sa->util_sum = max_t(u32, sa->util_sum, sa->util_avg * PELT_MIN_DIVIDER);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3701)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3702) r = removed_runnable;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3703) sub_positive(&sa->runnable_avg, r);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3704) sub_positive(&sa->runnable_sum, r * divider);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3705)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3706) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3707) * removed_runnable is the unweighted version of removed_load so we
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3708) * can use it to estimate removed_load_sum.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3709) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3710) add_tg_cfs_propagate(cfs_rq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3711) -(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3712)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3713) decayed = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3714) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3715)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3716) decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3717)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3718) #ifndef CONFIG_64BIT
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3719) smp_wmb();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3720) cfs_rq->load_last_update_time_copy = sa->last_update_time;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3721) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3722)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3723) return decayed;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3724) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3725)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3726) /**
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3727) * attach_entity_load_avg - attach this entity to its cfs_rq load avg
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3728) * @cfs_rq: cfs_rq to attach to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3729) * @se: sched_entity to attach
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3730) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3731) * Must call update_cfs_rq_load_avg() before this, since we rely on
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3732) * cfs_rq->avg.last_update_time being current.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3733) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3734) static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3735) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3736) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3737) * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3738) * See ___update_load_avg() for details.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3739) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3740) u32 divider = get_pelt_divider(&cfs_rq->avg);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3741)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3742) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3743) * When we attach the @se to the @cfs_rq, we must align the decay
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3744) * window because without that, really weird and wonderful things can
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3745) * happen.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3746) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3747) * XXX illustrate
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3748) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3749) se->avg.last_update_time = cfs_rq->avg.last_update_time;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3750) se->avg.period_contrib = cfs_rq->avg.period_contrib;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3751)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3752) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3753) * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3754) * period_contrib. This isn't strictly correct, but since we're
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3755) * entirely outside of the PELT hierarchy, nobody cares if we truncate
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3756) * _sum a little.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3757) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3758) se->avg.util_sum = se->avg.util_avg * divider;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3759)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3760) se->avg.runnable_sum = se->avg.runnable_avg * divider;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3761)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3762) se->avg.load_sum = divider;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3763) if (se_weight(se)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3764) se->avg.load_sum =
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3765) div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3766) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3767)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3768) enqueue_load_avg(cfs_rq, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3769) cfs_rq->avg.util_avg += se->avg.util_avg;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3770) cfs_rq->avg.util_sum += se->avg.util_sum;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3771) cfs_rq->avg.runnable_avg += se->avg.runnable_avg;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3772) cfs_rq->avg.runnable_sum += se->avg.runnable_sum;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3773)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3774) add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3775)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3776) cfs_rq_util_change(cfs_rq, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3777)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3778) trace_pelt_cfs_tp(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3779) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3780)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3781) /**
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3782) * detach_entity_load_avg - detach this entity from its cfs_rq load avg
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3783) * @cfs_rq: cfs_rq to detach from
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3784) * @se: sched_entity to detach
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3785) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3786) * Must call update_cfs_rq_load_avg() before this, since we rely on
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3787) * cfs_rq->avg.last_update_time being current.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3788) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3789) static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3790) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3791) dequeue_load_avg(cfs_rq, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3792) sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3793) sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3794) sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3795) sub_positive(&cfs_rq->avg.runnable_sum, se->avg.runnable_sum);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3796)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3797) add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3798)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3799) cfs_rq_util_change(cfs_rq, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3800)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3801) trace_pelt_cfs_tp(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3802) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3803)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3804) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3805) * Optional action to be done while updating the load average
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3806) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3807) #define UPDATE_TG 0x1
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3808) #define SKIP_AGE_LOAD 0x2
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3809) #define DO_ATTACH 0x4
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3810)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3811) /* Update task and its cfs_rq load average */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3812) static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3813) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3814) u64 now = cfs_rq_clock_pelt(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3815) int decayed;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3816)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3817) trace_android_vh_prepare_update_load_avg_se(se, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3818) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3819) * Track task load average for carrying it to new CPU after migrated, and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3820) * track group sched_entity load average for task_h_load calc in migration
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3821) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3822) if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3823) __update_load_avg_se(now, cfs_rq, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3824)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3825) trace_android_vh_finish_update_load_avg_se(se, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3826)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3827) decayed = update_cfs_rq_load_avg(now, cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3828) decayed |= propagate_entity_load_avg(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3829)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3830) if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3831)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3832) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3833) * DO_ATTACH means we're here from enqueue_entity().
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3834) * !last_update_time means we've passed through
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3835) * migrate_task_rq_fair() indicating we migrated.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3836) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3837) * IOW we're enqueueing a task on a new CPU.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3838) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3839) attach_entity_load_avg(cfs_rq, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3840) update_tg_load_avg(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3841)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3842) } else if (decayed) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3843) cfs_rq_util_change(cfs_rq, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3844)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3845) if (flags & UPDATE_TG)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3846) update_tg_load_avg(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3847) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3848) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3849)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3850) #ifndef CONFIG_64BIT
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3851) static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3852) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3853) u64 last_update_time_copy;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3854) u64 last_update_time;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3855)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3856) do {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3857) last_update_time_copy = cfs_rq->load_last_update_time_copy;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3858) smp_rmb();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3859) last_update_time = cfs_rq->avg.last_update_time;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3860) } while (last_update_time != last_update_time_copy);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3861)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3862) return last_update_time;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3863) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3864) #else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3865) static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3866) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3867) return cfs_rq->avg.last_update_time;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3868) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3869) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3870)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3871) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3872) * Synchronize entity load avg of dequeued entity without locking
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3873) * the previous rq.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3874) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3875) static void sync_entity_load_avg(struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3876) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3877) struct cfs_rq *cfs_rq = cfs_rq_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3878) u64 last_update_time;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3879)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3880) last_update_time = cfs_rq_last_update_time(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3881) trace_android_vh_prepare_update_load_avg_se(se, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3882) __update_load_avg_blocked_se(last_update_time, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3883) trace_android_vh_finish_update_load_avg_se(se, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3884) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3885)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3886) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3887) * Task first catches up with cfs_rq, and then subtract
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3888) * itself from the cfs_rq (task must be off the queue now).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3889) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3890) static void remove_entity_load_avg(struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3891) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3892) struct cfs_rq *cfs_rq = cfs_rq_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3893) unsigned long flags;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3894)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3895) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3896) * tasks cannot exit without having gone through wake_up_new_task() ->
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3897) * post_init_entity_util_avg() which will have added things to the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3898) * cfs_rq, so we can remove unconditionally.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3899) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3900)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3901) sync_entity_load_avg(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3902)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3903) raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3904) ++cfs_rq->removed.nr;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3905) cfs_rq->removed.util_avg += se->avg.util_avg;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3906) cfs_rq->removed.load_avg += se->avg.load_avg;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3907) cfs_rq->removed.runnable_avg += se->avg.runnable_avg;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3908) raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3909) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3910)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3911) static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq *cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3912) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3913) return cfs_rq->avg.runnable_avg;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3914) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3915)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3916) static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3917) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3918) return cfs_rq->avg.load_avg;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3919) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3920)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3921) static int newidle_balance(struct rq *this_rq, struct rq_flags *rf);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3922)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3923) static inline unsigned long task_util(struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3924) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3925) return READ_ONCE(p->se.avg.util_avg);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3926) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3927)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3928) static inline unsigned long _task_util_est(struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3929) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3930) struct util_est ue = READ_ONCE(p->se.avg.util_est);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3931)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3932) return max(ue.ewma, (ue.enqueued & ~UTIL_AVG_UNCHANGED));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3933) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3934)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3935) static inline unsigned long task_util_est(struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3936) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3937) return max(task_util(p), _task_util_est(p));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3938) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3939)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3940) #ifdef CONFIG_UCLAMP_TASK
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3941) static inline unsigned long uclamp_task_util(struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3942) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3943) return clamp(task_util_est(p),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3944) uclamp_eff_value(p, UCLAMP_MIN),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3945) uclamp_eff_value(p, UCLAMP_MAX));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3946) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3947) #else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3948) static inline unsigned long uclamp_task_util(struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3949) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3950) return task_util_est(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3951) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3952) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3953)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3954) static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3955) struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3956) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3957) unsigned int enqueued;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3958)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3959) if (!sched_feat(UTIL_EST))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3960) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3961)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3962) /* Update root cfs_rq's estimated utilization */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3963) enqueued = cfs_rq->avg.util_est.enqueued;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3964) enqueued += _task_util_est(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3965) WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3966)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3967) trace_sched_util_est_cfs_tp(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3968) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3969)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3970) static inline void util_est_dequeue(struct cfs_rq *cfs_rq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3971) struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3972) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3973) unsigned int enqueued;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3974)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3975) if (!sched_feat(UTIL_EST))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3976) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3977)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3978) /* Update root cfs_rq's estimated utilization */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3979) enqueued = cfs_rq->avg.util_est.enqueued;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3980) enqueued -= min_t(unsigned int, enqueued, _task_util_est(p));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3981) WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3982)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3983) trace_sched_util_est_cfs_tp(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3984) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3985)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3986) #define UTIL_EST_MARGIN (SCHED_CAPACITY_SCALE / 100)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3987)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3988) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3989) * Check if a (signed) value is within a specified (unsigned) margin,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3990) * based on the observation that:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3991) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3992) * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3993) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3994) * NOTE: this only works when value + maring < INT_MAX.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3995) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3996) static inline bool within_margin(int value, int margin)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3997) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3998) return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3999) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4000)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4001) static inline void util_est_update(struct cfs_rq *cfs_rq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4002) struct task_struct *p,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4003) bool task_sleep)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4004) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4005) long last_ewma_diff, last_enqueued_diff;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4006) struct util_est ue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4007) int ret = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4008)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4009) trace_android_rvh_util_est_update(cfs_rq, p, task_sleep, &ret);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4010) if (ret)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4011) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4012)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4013) if (!sched_feat(UTIL_EST))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4014) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4015)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4016) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4017) * Skip update of task's estimated utilization when the task has not
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4018) * yet completed an activation, e.g. being migrated.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4019) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4020) if (!task_sleep)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4021) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4022)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4023) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4024) * If the PELT values haven't changed since enqueue time,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4025) * skip the util_est update.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4026) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4027) ue = p->se.avg.util_est;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4028) if (ue.enqueued & UTIL_AVG_UNCHANGED)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4029) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4030)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4031) last_enqueued_diff = ue.enqueued;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4032)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4033) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4034) * Reset EWMA on utilization increases, the moving average is used only
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4035) * to smooth utilization decreases.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4036) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4037) ue.enqueued = task_util(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4038) if (sched_feat(UTIL_EST_FASTUP)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4039) if (ue.ewma < ue.enqueued) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4040) ue.ewma = ue.enqueued;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4041) goto done;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4042) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4043) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4044)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4045) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4046) * Skip update of task's estimated utilization when its members are
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4047) * already ~1% close to its last activation value.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4048) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4049) last_ewma_diff = ue.enqueued - ue.ewma;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4050) last_enqueued_diff -= ue.enqueued;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4051) if (within_margin(last_ewma_diff, UTIL_EST_MARGIN)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4052) if (!within_margin(last_enqueued_diff, UTIL_EST_MARGIN))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4053) goto done;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4054)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4055) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4056) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4057)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4058) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4059) * To avoid overestimation of actual task utilization, skip updates if
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4060) * we cannot grant there is idle time in this CPU.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4061) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4062) if (task_util(p) > capacity_orig_of(cpu_of(rq_of(cfs_rq))))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4063) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4064)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4065) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4066) * Update Task's estimated utilization
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4067) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4068) * When *p completes an activation we can consolidate another sample
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4069) * of the task size. This is done by storing the current PELT value
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4070) * as ue.enqueued and by using this value to update the Exponential
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4071) * Weighted Moving Average (EWMA):
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4072) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4073) * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4074) * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4075) * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4076) * = w * ( last_ewma_diff ) + ewma(t-1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4077) * = w * (last_ewma_diff + ewma(t-1) / w)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4078) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4079) * Where 'w' is the weight of new samples, which is configured to be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4080) * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4081) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4082) ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4083) ue.ewma += last_ewma_diff;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4084) ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4085) done:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4086) ue.enqueued |= UTIL_AVG_UNCHANGED;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4087) WRITE_ONCE(p->se.avg.util_est, ue);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4088)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4089) trace_sched_util_est_se_tp(&p->se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4090) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4091)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4092) static inline int task_fits_capacity(struct task_struct *p, long capacity)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4093) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4094) return fits_capacity(uclamp_task_util(p), capacity);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4095) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4096)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4097) static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4098) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4099) bool need_update = true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4100)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4101) trace_android_rvh_update_misfit_status(p, rq, &need_update);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4102) if (!static_branch_unlikely(&sched_asym_cpucapacity) || !need_update)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4103) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4104)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4105) if (!p || p->nr_cpus_allowed == 1) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4106) rq->misfit_task_load = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4107) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4108) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4109)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4110) if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4111) rq->misfit_task_load = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4112) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4113) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4114)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4115) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4116) * Make sure that misfit_task_load will not be null even if
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4117) * task_h_load() returns 0.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4118) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4119) rq->misfit_task_load = max_t(unsigned long, task_h_load(p), 1);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4120) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4121)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4122) #else /* CONFIG_SMP */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4123)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4124) #define UPDATE_TG 0x0
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4125) #define SKIP_AGE_LOAD 0x0
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4126) #define DO_ATTACH 0x0
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4127)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4128) static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4129) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4130) cfs_rq_util_change(cfs_rq, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4131) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4132)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4133) static inline void remove_entity_load_avg(struct sched_entity *se) {}
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4134)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4135) static inline void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4136) attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4137) static inline void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4138) detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4139)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4140) static inline int newidle_balance(struct rq *rq, struct rq_flags *rf)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4141) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4142) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4143) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4144)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4145) static inline void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4146) util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4147)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4148) static inline void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4149) util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4150)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4151) static inline void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4152) util_est_update(struct cfs_rq *cfs_rq, struct task_struct *p,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4153) bool task_sleep) {}
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4154) static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4155)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4156) #endif /* CONFIG_SMP */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4157)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4158) static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4159) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4160) #ifdef CONFIG_SCHED_DEBUG
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4161) s64 d = se->vruntime - cfs_rq->min_vruntime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4162)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4163) if (d < 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4164) d = -d;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4165)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4166) if (d > 3*sysctl_sched_latency)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4167) schedstat_inc(cfs_rq->nr_spread_over);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4168) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4169) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4170)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4171) static void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4172) place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4173) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4174) u64 vruntime = cfs_rq->min_vruntime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4175)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4176) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4177) * The 'current' period is already promised to the current tasks,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4178) * however the extra weight of the new task will slow them down a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4179) * little, place the new task so that it fits in the slot that
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4180) * stays open at the end.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4181) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4182) if (initial && sched_feat(START_DEBIT))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4183) vruntime += sched_vslice(cfs_rq, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4184)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4185) /* sleeps up to a single latency don't count. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4186) if (!initial) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4187) unsigned long thresh = sysctl_sched_latency;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4188)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4189) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4190) * Halve their sleep time's effect, to allow
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4191) * for a gentler effect of sleepers:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4192) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4193) if (sched_feat(GENTLE_FAIR_SLEEPERS))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4194) thresh >>= 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4195)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4196) vruntime -= thresh;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4197) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4198)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4199) /* ensure we never gain time by being placed backwards. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4200) se->vruntime = max_vruntime(se->vruntime, vruntime);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4201) trace_android_rvh_place_entity(cfs_rq, se, initial, vruntime);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4202) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4203)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4204) static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4205)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4206) static inline void check_schedstat_required(void)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4207) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4208) #ifdef CONFIG_SCHEDSTATS
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4209) if (schedstat_enabled())
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4210) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4211)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4212) /* Force schedstat enabled if a dependent tracepoint is active */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4213) if (trace_sched_stat_wait_enabled() ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4214) trace_sched_stat_sleep_enabled() ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4215) trace_sched_stat_iowait_enabled() ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4216) trace_sched_stat_blocked_enabled() ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4217) trace_sched_stat_runtime_enabled()) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4218) printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4219) "stat_blocked and stat_runtime require the "
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4220) "kernel parameter schedstats=enable or "
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4221) "kernel.sched_schedstats=1\n");
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4222) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4223) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4224) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4225)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4226) static inline bool cfs_bandwidth_used(void);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4227)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4228) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4229) * MIGRATION
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4230) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4231) * dequeue
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4232) * update_curr()
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4233) * update_min_vruntime()
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4234) * vruntime -= min_vruntime
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4235) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4236) * enqueue
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4237) * update_curr()
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4238) * update_min_vruntime()
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4239) * vruntime += min_vruntime
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4240) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4241) * this way the vruntime transition between RQs is done when both
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4242) * min_vruntime are up-to-date.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4243) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4244) * WAKEUP (remote)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4245) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4246) * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4247) * vruntime -= min_vruntime
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4248) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4249) * enqueue
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4250) * update_curr()
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4251) * update_min_vruntime()
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4252) * vruntime += min_vruntime
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4253) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4254) * this way we don't have the most up-to-date min_vruntime on the originating
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4255) * CPU and an up-to-date min_vruntime on the destination CPU.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4256) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4257)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4258) static void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4259) enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4260) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4261) bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4262) bool curr = cfs_rq->curr == se;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4263)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4264) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4265) * If we're the current task, we must renormalise before calling
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4266) * update_curr().
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4267) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4268) if (renorm && curr)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4269) se->vruntime += cfs_rq->min_vruntime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4270)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4271) update_curr(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4272)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4273) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4274) * Otherwise, renormalise after, such that we're placed at the current
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4275) * moment in time, instead of some random moment in the past. Being
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4276) * placed in the past could significantly boost this task to the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4277) * fairness detriment of existing tasks.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4278) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4279) if (renorm && !curr)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4280) se->vruntime += cfs_rq->min_vruntime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4281)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4282) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4283) * When enqueuing a sched_entity, we must:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4284) * - Update loads to have both entity and cfs_rq synced with now.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4285) * - Add its load to cfs_rq->runnable_avg
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4286) * - For group_entity, update its weight to reflect the new share of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4287) * its group cfs_rq
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4288) * - Add its new weight to cfs_rq->load.weight
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4289) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4290) update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4291) se_update_runnable(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4292) update_cfs_group(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4293) account_entity_enqueue(cfs_rq, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4294)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4295) if (flags & ENQUEUE_WAKEUP)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4296) place_entity(cfs_rq, se, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4297)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4298) check_schedstat_required();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4299) update_stats_enqueue(cfs_rq, se, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4300) check_spread(cfs_rq, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4301) if (!curr)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4302) __enqueue_entity(cfs_rq, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4303) se->on_rq = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4304)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4305) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4306) * When bandwidth control is enabled, cfs might have been removed
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4307) * because of a parent been throttled but cfs->nr_running > 1. Try to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4308) * add it unconditionnally.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4309) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4310) if (cfs_rq->nr_running == 1 || cfs_bandwidth_used())
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4311) list_add_leaf_cfs_rq(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4312)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4313) if (cfs_rq->nr_running == 1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4314) check_enqueue_throttle(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4315) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4316)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4317) static void __clear_buddies_last(struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4318) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4319) for_each_sched_entity(se) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4320) struct cfs_rq *cfs_rq = cfs_rq_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4321) if (cfs_rq->last != se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4322) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4323)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4324) cfs_rq->last = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4325) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4326) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4327)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4328) static void __clear_buddies_next(struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4329) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4330) for_each_sched_entity(se) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4331) struct cfs_rq *cfs_rq = cfs_rq_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4332) if (cfs_rq->next != se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4333) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4334)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4335) cfs_rq->next = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4336) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4337) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4338)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4339) static void __clear_buddies_skip(struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4340) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4341) for_each_sched_entity(se) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4342) struct cfs_rq *cfs_rq = cfs_rq_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4343) if (cfs_rq->skip != se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4344) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4345)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4346) cfs_rq->skip = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4347) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4348) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4349)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4350) static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4351) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4352) if (cfs_rq->last == se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4353) __clear_buddies_last(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4354)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4355) if (cfs_rq->next == se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4356) __clear_buddies_next(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4357)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4358) if (cfs_rq->skip == se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4359) __clear_buddies_skip(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4360) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4361)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4362) static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4363)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4364) static void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4365) dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4366) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4367) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4368) * Update run-time statistics of the 'current'.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4369) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4370) update_curr(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4371)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4372) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4373) * When dequeuing a sched_entity, we must:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4374) * - Update loads to have both entity and cfs_rq synced with now.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4375) * - Subtract its load from the cfs_rq->runnable_avg.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4376) * - Subtract its previous weight from cfs_rq->load.weight.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4377) * - For group entity, update its weight to reflect the new share
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4378) * of its group cfs_rq.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4379) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4380) update_load_avg(cfs_rq, se, UPDATE_TG);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4381) se_update_runnable(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4382)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4383) update_stats_dequeue(cfs_rq, se, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4384)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4385) clear_buddies(cfs_rq, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4386)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4387) if (se != cfs_rq->curr)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4388) __dequeue_entity(cfs_rq, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4389) se->on_rq = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4390) account_entity_dequeue(cfs_rq, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4391)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4392) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4393) * Normalize after update_curr(); which will also have moved
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4394) * min_vruntime if @se is the one holding it back. But before doing
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4395) * update_min_vruntime() again, which will discount @se's position and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4396) * can move min_vruntime forward still more.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4397) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4398) if (!(flags & DEQUEUE_SLEEP))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4399) se->vruntime -= cfs_rq->min_vruntime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4400)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4401) /* return excess runtime on last dequeue */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4402) return_cfs_rq_runtime(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4403)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4404) update_cfs_group(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4405)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4406) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4407) * Now advance min_vruntime if @se was the entity holding it back,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4408) * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4409) * put back on, and if we advance min_vruntime, we'll be placed back
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4410) * further than we started -- ie. we'll be penalized.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4411) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4412) if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4413) update_min_vruntime(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4414) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4415)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4416) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4417) * Preempt the current task with a newly woken task if needed:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4418) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4419) static void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4420) check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4421) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4422) unsigned long ideal_runtime, delta_exec;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4423) struct sched_entity *se;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4424) s64 delta;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4425) bool skip_preempt = false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4426)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4427) ideal_runtime = sched_slice(cfs_rq, curr);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4428) delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4429) trace_android_rvh_check_preempt_tick(current, &ideal_runtime, &skip_preempt,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4430) delta_exec, cfs_rq, curr, sysctl_sched_min_granularity);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4431) if (skip_preempt)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4432) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4433) if (delta_exec > ideal_runtime) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4434) resched_curr(rq_of(cfs_rq));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4435) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4436) * The current task ran long enough, ensure it doesn't get
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4437) * re-elected due to buddy favours.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4438) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4439) clear_buddies(cfs_rq, curr);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4440) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4441) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4442)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4443) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4444) * Ensure that a task that missed wakeup preemption by a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4445) * narrow margin doesn't have to wait for a full slice.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4446) * This also mitigates buddy induced latencies under load.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4447) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4448) if (delta_exec < sysctl_sched_min_granularity)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4449) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4450)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4451) se = __pick_first_entity(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4452) delta = curr->vruntime - se->vruntime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4453)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4454) if (delta < 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4455) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4456)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4457) if (delta > ideal_runtime)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4458) resched_curr(rq_of(cfs_rq));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4459) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4460)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4461) void set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4462) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4463) /* 'current' is not kept within the tree. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4464) if (se->on_rq) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4465) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4466) * Any task has to be enqueued before it get to execute on
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4467) * a CPU. So account for the time it spent waiting on the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4468) * runqueue.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4469) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4470) update_stats_wait_end(cfs_rq, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4471) __dequeue_entity(cfs_rq, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4472) update_load_avg(cfs_rq, se, UPDATE_TG);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4473) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4474)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4475) update_stats_curr_start(cfs_rq, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4476) cfs_rq->curr = se;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4477)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4478) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4479) * Track our maximum slice length, if the CPU's load is at
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4480) * least twice that of our own weight (i.e. dont track it
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4481) * when there are only lesser-weight tasks around):
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4482) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4483) if (schedstat_enabled() &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4484) rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4485) schedstat_set(se->statistics.slice_max,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4486) max((u64)schedstat_val(se->statistics.slice_max),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4487) se->sum_exec_runtime - se->prev_sum_exec_runtime));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4488) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4489)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4490) se->prev_sum_exec_runtime = se->sum_exec_runtime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4491) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4492) EXPORT_SYMBOL_GPL(set_next_entity);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4493)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4494)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4495) static int
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4496) wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4497)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4498) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4499) * Pick the next process, keeping these things in mind, in this order:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4500) * 1) keep things fair between processes/task groups
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4501) * 2) pick the "next" process, since someone really wants that to run
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4502) * 3) pick the "last" process, for cache locality
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4503) * 4) do not run the "skip" process, if something else is available
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4504) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4505) static struct sched_entity *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4506) pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4507) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4508) struct sched_entity *left = __pick_first_entity(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4509) struct sched_entity *se = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4510)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4511) trace_android_rvh_pick_next_entity(cfs_rq, curr, &se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4512) if (se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4513) goto done;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4514)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4515) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4516) * If curr is set we have to see if its left of the leftmost entity
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4517) * still in the tree, provided there was anything in the tree at all.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4518) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4519) if (!left || (curr && entity_before(curr, left)))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4520) left = curr;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4521)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4522) se = left; /* ideally we run the leftmost entity */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4523)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4524) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4525) * Avoid running the skip buddy, if running something else can
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4526) * be done without getting too unfair.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4527) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4528) if (cfs_rq->skip == se) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4529) struct sched_entity *second;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4530)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4531) if (se == curr) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4532) second = __pick_first_entity(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4533) } else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4534) second = __pick_next_entity(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4535) if (!second || (curr && entity_before(curr, second)))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4536) second = curr;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4537) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4538)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4539) if (second && wakeup_preempt_entity(second, left) < 1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4540) se = second;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4541) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4542)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4543) if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4544) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4545) * Someone really wants this to run. If it's not unfair, run it.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4546) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4547) se = cfs_rq->next;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4548) } else if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4549) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4550) * Prefer last buddy, try to return the CPU to a preempted task.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4551) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4552) se = cfs_rq->last;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4553) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4554)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4555) done:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4556) clear_buddies(cfs_rq, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4557)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4558) return se;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4559) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4560)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4561) static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4562)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4563) static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4564) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4565) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4566) * If still on the runqueue then deactivate_task()
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4567) * was not called and update_curr() has to be done:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4568) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4569) if (prev->on_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4570) update_curr(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4571)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4572) /* throttle cfs_rqs exceeding runtime */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4573) check_cfs_rq_runtime(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4574)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4575) check_spread(cfs_rq, prev);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4576)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4577) if (prev->on_rq) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4578) update_stats_wait_start(cfs_rq, prev);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4579) /* Put 'current' back into the tree. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4580) __enqueue_entity(cfs_rq, prev);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4581) /* in !on_rq case, update occurred at dequeue */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4582) update_load_avg(cfs_rq, prev, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4583) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4584) cfs_rq->curr = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4585) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4586)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4587) static void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4588) entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4589) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4590) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4591) * Update run-time statistics of the 'current'.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4592) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4593) update_curr(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4594)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4595) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4596) * Ensure that runnable average is periodically updated.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4597) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4598) update_load_avg(cfs_rq, curr, UPDATE_TG);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4599) update_cfs_group(curr);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4600)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4601) #ifdef CONFIG_SCHED_HRTICK
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4602) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4603) * queued ticks are scheduled to match the slice, so don't bother
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4604) * validating it and just reschedule.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4605) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4606) if (queued) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4607) resched_curr(rq_of(cfs_rq));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4608) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4609) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4610) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4611) * don't let the period tick interfere with the hrtick preemption
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4612) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4613) if (!sched_feat(DOUBLE_TICK) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4614) hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4615) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4616) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4617)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4618) if (cfs_rq->nr_running > 1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4619) check_preempt_tick(cfs_rq, curr);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4620) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4621)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4622)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4623) /**************************************************
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4624) * CFS bandwidth control machinery
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4625) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4626)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4627) #ifdef CONFIG_CFS_BANDWIDTH
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4628)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4629) #ifdef CONFIG_JUMP_LABEL
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4630) static struct static_key __cfs_bandwidth_used;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4631)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4632) static inline bool cfs_bandwidth_used(void)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4633) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4634) return static_key_false(&__cfs_bandwidth_used);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4635) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4636)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4637) void cfs_bandwidth_usage_inc(void)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4638) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4639) static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4640) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4641)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4642) void cfs_bandwidth_usage_dec(void)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4643) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4644) static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4645) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4646) #else /* CONFIG_JUMP_LABEL */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4647) static bool cfs_bandwidth_used(void)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4648) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4649) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4650) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4651)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4652) void cfs_bandwidth_usage_inc(void) {}
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4653) void cfs_bandwidth_usage_dec(void) {}
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4654) #endif /* CONFIG_JUMP_LABEL */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4655)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4656) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4657) * default period for cfs group bandwidth.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4658) * default: 0.1s, units: nanoseconds
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4659) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4660) static inline u64 default_cfs_period(void)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4661) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4662) return 100000000ULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4663) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4664)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4665) static inline u64 sched_cfs_bandwidth_slice(void)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4666) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4667) return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4668) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4669)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4670) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4671) * Replenish runtime according to assigned quota. We use sched_clock_cpu
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4672) * directly instead of rq->clock to avoid adding additional synchronization
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4673) * around rq->lock.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4674) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4675) * requires cfs_b->lock
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4676) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4677) void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4678) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4679) if (cfs_b->quota != RUNTIME_INF)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4680) cfs_b->runtime = cfs_b->quota;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4681) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4682)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4683) static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4684) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4685) return &tg->cfs_bandwidth;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4686) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4687)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4688) /* returns 0 on failure to allocate runtime */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4689) static int __assign_cfs_rq_runtime(struct cfs_bandwidth *cfs_b,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4690) struct cfs_rq *cfs_rq, u64 target_runtime)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4691) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4692) u64 min_amount, amount = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4693)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4694) lockdep_assert_held(&cfs_b->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4695)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4696) /* note: this is a positive sum as runtime_remaining <= 0 */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4697) min_amount = target_runtime - cfs_rq->runtime_remaining;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4698)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4699) if (cfs_b->quota == RUNTIME_INF)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4700) amount = min_amount;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4701) else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4702) start_cfs_bandwidth(cfs_b);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4703)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4704) if (cfs_b->runtime > 0) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4705) amount = min(cfs_b->runtime, min_amount);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4706) cfs_b->runtime -= amount;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4707) cfs_b->idle = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4708) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4709) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4710)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4711) cfs_rq->runtime_remaining += amount;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4712)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4713) return cfs_rq->runtime_remaining > 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4714) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4715)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4716) /* returns 0 on failure to allocate runtime */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4717) static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4718) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4719) struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4720) int ret;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4721)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4722) raw_spin_lock(&cfs_b->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4723) ret = __assign_cfs_rq_runtime(cfs_b, cfs_rq, sched_cfs_bandwidth_slice());
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4724) raw_spin_unlock(&cfs_b->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4725)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4726) return ret;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4727) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4728)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4729) static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4730) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4731) /* dock delta_exec before expiring quota (as it could span periods) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4732) cfs_rq->runtime_remaining -= delta_exec;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4733)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4734) if (likely(cfs_rq->runtime_remaining > 0))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4735) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4736)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4737) if (cfs_rq->throttled)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4738) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4739) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4740) * if we're unable to extend our runtime we resched so that the active
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4741) * hierarchy can be throttled
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4742) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4743) if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4744) resched_curr(rq_of(cfs_rq));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4745) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4746)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4747) static __always_inline
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4748) void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4749) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4750) if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4751) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4752)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4753) __account_cfs_rq_runtime(cfs_rq, delta_exec);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4754) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4755)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4756) static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4757) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4758) return cfs_bandwidth_used() && cfs_rq->throttled;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4759) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4760)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4761) /* check whether cfs_rq, or any parent, is throttled */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4762) static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4763) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4764) return cfs_bandwidth_used() && cfs_rq->throttle_count;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4765) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4766)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4767) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4768) * Ensure that neither of the group entities corresponding to src_cpu or
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4769) * dest_cpu are members of a throttled hierarchy when performing group
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4770) * load-balance operations.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4771) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4772) static inline int throttled_lb_pair(struct task_group *tg,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4773) int src_cpu, int dest_cpu)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4774) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4775) struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4776)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4777) src_cfs_rq = tg->cfs_rq[src_cpu];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4778) dest_cfs_rq = tg->cfs_rq[dest_cpu];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4779)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4780) return throttled_hierarchy(src_cfs_rq) ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4781) throttled_hierarchy(dest_cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4782) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4783)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4784) static int tg_unthrottle_up(struct task_group *tg, void *data)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4785) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4786) struct rq *rq = data;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4787) struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4788)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4789) cfs_rq->throttle_count--;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4790) if (!cfs_rq->throttle_count) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4791) cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4792) cfs_rq->throttled_clock_task;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4793)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4794) /* Add cfs_rq with already running entity in the list */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4795) if (cfs_rq->nr_running >= 1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4796) list_add_leaf_cfs_rq(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4797) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4798)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4799) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4800) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4801)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4802) static int tg_throttle_down(struct task_group *tg, void *data)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4803) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4804) struct rq *rq = data;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4805) struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4806)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4807) /* group is entering throttled state, stop time */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4808) if (!cfs_rq->throttle_count) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4809) cfs_rq->throttled_clock_task = rq_clock_task(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4810) list_del_leaf_cfs_rq(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4811) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4812) cfs_rq->throttle_count++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4813)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4814) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4815) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4816)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4817) static bool throttle_cfs_rq(struct cfs_rq *cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4818) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4819) struct rq *rq = rq_of(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4820) struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4821) struct sched_entity *se;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4822) long task_delta, idle_task_delta, dequeue = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4823)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4824) raw_spin_lock(&cfs_b->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4825) /* This will start the period timer if necessary */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4826) if (__assign_cfs_rq_runtime(cfs_b, cfs_rq, 1)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4827) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4828) * We have raced with bandwidth becoming available, and if we
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4829) * actually throttled the timer might not unthrottle us for an
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4830) * entire period. We additionally needed to make sure that any
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4831) * subsequent check_cfs_rq_runtime calls agree not to throttle
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4832) * us, as we may commit to do cfs put_prev+pick_next, so we ask
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4833) * for 1ns of runtime rather than just check cfs_b.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4834) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4835) dequeue = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4836) } else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4837) list_add_tail_rcu(&cfs_rq->throttled_list,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4838) &cfs_b->throttled_cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4839) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4840) raw_spin_unlock(&cfs_b->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4841)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4842) if (!dequeue)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4843) return false; /* Throttle no longer required. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4844)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4845) se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4846)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4847) /* freeze hierarchy runnable averages while throttled */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4848) rcu_read_lock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4849) walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4850) rcu_read_unlock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4851)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4852) task_delta = cfs_rq->h_nr_running;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4853) idle_task_delta = cfs_rq->idle_h_nr_running;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4854) for_each_sched_entity(se) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4855) struct cfs_rq *qcfs_rq = cfs_rq_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4856) /* throttled entity or throttle-on-deactivate */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4857) if (!se->on_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4858) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4859)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4860) if (dequeue) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4861) dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4862) } else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4863) update_load_avg(qcfs_rq, se, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4864) se_update_runnable(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4865) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4866)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4867) qcfs_rq->h_nr_running -= task_delta;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4868) qcfs_rq->idle_h_nr_running -= idle_task_delta;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4869)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4870) if (qcfs_rq->load.weight)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4871) dequeue = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4872) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4873)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4874) if (!se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4875) sub_nr_running(rq, task_delta);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4876)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4877) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4878) * Note: distribution will already see us throttled via the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4879) * throttled-list. rq->lock protects completion.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4880) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4881) cfs_rq->throttled = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4882) cfs_rq->throttled_clock = rq_clock(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4883) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4884) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4885)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4886) void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4887) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4888) struct rq *rq = rq_of(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4889) struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4890) struct sched_entity *se;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4891) long task_delta, idle_task_delta;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4892)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4893) se = cfs_rq->tg->se[cpu_of(rq)];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4894)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4895) cfs_rq->throttled = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4896)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4897) update_rq_clock(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4898)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4899) raw_spin_lock(&cfs_b->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4900) cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4901) list_del_rcu(&cfs_rq->throttled_list);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4902) raw_spin_unlock(&cfs_b->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4903)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4904) /* update hierarchical throttle state */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4905) walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4906)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4907) if (!cfs_rq->load.weight)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4908) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4909)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4910) task_delta = cfs_rq->h_nr_running;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4911) idle_task_delta = cfs_rq->idle_h_nr_running;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4912) for_each_sched_entity(se) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4913) if (se->on_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4914) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4915) cfs_rq = cfs_rq_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4916) enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4917)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4918) cfs_rq->h_nr_running += task_delta;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4919) cfs_rq->idle_h_nr_running += idle_task_delta;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4920)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4921) /* end evaluation on encountering a throttled cfs_rq */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4922) if (cfs_rq_throttled(cfs_rq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4923) goto unthrottle_throttle;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4924) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4925)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4926) for_each_sched_entity(se) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4927) cfs_rq = cfs_rq_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4928)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4929) update_load_avg(cfs_rq, se, UPDATE_TG);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4930) se_update_runnable(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4931)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4932) cfs_rq->h_nr_running += task_delta;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4933) cfs_rq->idle_h_nr_running += idle_task_delta;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4934)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4935)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4936) /* end evaluation on encountering a throttled cfs_rq */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4937) if (cfs_rq_throttled(cfs_rq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4938) goto unthrottle_throttle;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4939)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4940) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4941) * One parent has been throttled and cfs_rq removed from the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4942) * list. Add it back to not break the leaf list.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4943) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4944) if (throttled_hierarchy(cfs_rq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4945) list_add_leaf_cfs_rq(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4946) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4947)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4948) /* At this point se is NULL and we are at root level*/
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4949) add_nr_running(rq, task_delta);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4950)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4951) unthrottle_throttle:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4952) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4953) * The cfs_rq_throttled() breaks in the above iteration can result in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4954) * incomplete leaf list maintenance, resulting in triggering the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4955) * assertion below.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4956) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4957) for_each_sched_entity(se) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4958) cfs_rq = cfs_rq_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4959)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4960) if (list_add_leaf_cfs_rq(cfs_rq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4961) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4962) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4963)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4964) assert_list_leaf_cfs_rq(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4965)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4966) /* Determine whether we need to wake up potentially idle CPU: */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4967) if (rq->curr == rq->idle && rq->cfs.nr_running)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4968) resched_curr(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4969) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4970)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4971) static void distribute_cfs_runtime(struct cfs_bandwidth *cfs_b)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4972) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4973) struct cfs_rq *cfs_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4974) u64 runtime, remaining = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4975)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4976) rcu_read_lock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4977) list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4978) throttled_list) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4979) struct rq *rq = rq_of(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4980) struct rq_flags rf;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4981)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4982) rq_lock_irqsave(rq, &rf);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4983) if (!cfs_rq_throttled(cfs_rq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4984) goto next;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4985)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4986) /* By the above check, this should never be true */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4987) SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4988)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4989) raw_spin_lock(&cfs_b->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4990) runtime = -cfs_rq->runtime_remaining + 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4991) if (runtime > cfs_b->runtime)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4992) runtime = cfs_b->runtime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4993) cfs_b->runtime -= runtime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4994) remaining = cfs_b->runtime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4995) raw_spin_unlock(&cfs_b->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4996)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4997) cfs_rq->runtime_remaining += runtime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4998)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4999) /* we check whether we're throttled above */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5000) if (cfs_rq->runtime_remaining > 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5001) unthrottle_cfs_rq(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5002)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5003) next:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5004) rq_unlock_irqrestore(rq, &rf);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5005)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5006) if (!remaining)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5007) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5008) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5009) rcu_read_unlock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5010) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5011)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5012) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5013) * Responsible for refilling a task_group's bandwidth and unthrottling its
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5014) * cfs_rqs as appropriate. If there has been no activity within the last
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5015) * period the timer is deactivated until scheduling resumes; cfs_b->idle is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5016) * used to track this state.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5017) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5018) static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5019) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5020) int throttled;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5021)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5022) /* no need to continue the timer with no bandwidth constraint */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5023) if (cfs_b->quota == RUNTIME_INF)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5024) goto out_deactivate;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5025)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5026) throttled = !list_empty(&cfs_b->throttled_cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5027) cfs_b->nr_periods += overrun;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5028)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5029) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5030) * idle depends on !throttled (for the case of a large deficit), and if
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5031) * we're going inactive then everything else can be deferred
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5032) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5033) if (cfs_b->idle && !throttled)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5034) goto out_deactivate;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5035)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5036) __refill_cfs_bandwidth_runtime(cfs_b);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5037)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5038) if (!throttled) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5039) /* mark as potentially idle for the upcoming period */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5040) cfs_b->idle = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5041) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5042) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5043)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5044) /* account preceding periods in which throttling occurred */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5045) cfs_b->nr_throttled += overrun;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5046)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5047) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5048) * This check is repeated as we release cfs_b->lock while we unthrottle.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5049) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5050) while (throttled && cfs_b->runtime > 0) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5051) raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5052) /* we can't nest cfs_b->lock while distributing bandwidth */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5053) distribute_cfs_runtime(cfs_b);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5054) raw_spin_lock_irqsave(&cfs_b->lock, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5055)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5056) throttled = !list_empty(&cfs_b->throttled_cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5057) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5058)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5059) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5060) * While we are ensured activity in the period following an
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5061) * unthrottle, this also covers the case in which the new bandwidth is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5062) * insufficient to cover the existing bandwidth deficit. (Forcing the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5063) * timer to remain active while there are any throttled entities.)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5064) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5065) cfs_b->idle = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5066)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5067) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5068)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5069) out_deactivate:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5070) return 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5071) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5072)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5073) /* a cfs_rq won't donate quota below this amount */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5074) static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5075) /* minimum remaining period time to redistribute slack quota */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5076) static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5077) /* how long we wait to gather additional slack before distributing */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5078) static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5079)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5080) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5081) * Are we near the end of the current quota period?
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5082) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5083) * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5084) * hrtimer base being cleared by hrtimer_start. In the case of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5085) * migrate_hrtimers, base is never cleared, so we are fine.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5086) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5087) static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5088) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5089) struct hrtimer *refresh_timer = &cfs_b->period_timer;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5090) s64 remaining;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5091)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5092) /* if the call-back is running a quota refresh is already occurring */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5093) if (hrtimer_callback_running(refresh_timer))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5094) return 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5095)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5096) /* is a quota refresh about to occur? */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5097) remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5098) if (remaining < (s64)min_expire)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5099) return 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5100)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5101) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5102) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5103)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5104) static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5105) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5106) u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5107)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5108) /* if there's a quota refresh soon don't bother with slack */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5109) if (runtime_refresh_within(cfs_b, min_left))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5110) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5111)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5112) /* don't push forwards an existing deferred unthrottle */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5113) if (cfs_b->slack_started)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5114) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5115) cfs_b->slack_started = true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5116)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5117) hrtimer_start(&cfs_b->slack_timer,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5118) ns_to_ktime(cfs_bandwidth_slack_period),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5119) HRTIMER_MODE_REL);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5120) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5121)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5122) /* we know any runtime found here is valid as update_curr() precedes return */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5123) static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5124) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5125) struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5126) s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5127)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5128) if (slack_runtime <= 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5129) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5130)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5131) raw_spin_lock(&cfs_b->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5132) if (cfs_b->quota != RUNTIME_INF) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5133) cfs_b->runtime += slack_runtime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5134)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5135) /* we are under rq->lock, defer unthrottling using a timer */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5136) if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5137) !list_empty(&cfs_b->throttled_cfs_rq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5138) start_cfs_slack_bandwidth(cfs_b);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5139) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5140) raw_spin_unlock(&cfs_b->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5141)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5142) /* even if it's not valid for return we don't want to try again */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5143) cfs_rq->runtime_remaining -= slack_runtime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5144) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5145)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5146) static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5147) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5148) if (!cfs_bandwidth_used())
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5149) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5150)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5151) if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5152) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5153)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5154) __return_cfs_rq_runtime(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5155) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5156)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5157) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5158) * This is done with a timer (instead of inline with bandwidth return) since
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5159) * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5160) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5161) static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5162) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5163) u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5164) unsigned long flags;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5165)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5166) /* confirm we're still not at a refresh boundary */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5167) raw_spin_lock_irqsave(&cfs_b->lock, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5168) cfs_b->slack_started = false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5169)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5170) if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5171) raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5172) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5173) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5174)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5175) if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5176) runtime = cfs_b->runtime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5177)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5178) raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5179)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5180) if (!runtime)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5181) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5182)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5183) distribute_cfs_runtime(cfs_b);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5184)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5185) raw_spin_lock_irqsave(&cfs_b->lock, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5186) raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5187) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5188)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5189) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5190) * When a group wakes up we want to make sure that its quota is not already
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5191) * expired/exceeded, otherwise it may be allowed to steal additional ticks of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5192) * runtime as update_curr() throttling can not trigger until it's on-rq.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5193) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5194) static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5195) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5196) if (!cfs_bandwidth_used())
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5197) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5198)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5199) /* an active group must be handled by the update_curr()->put() path */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5200) if (!cfs_rq->runtime_enabled || cfs_rq->curr)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5201) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5202)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5203) /* ensure the group is not already throttled */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5204) if (cfs_rq_throttled(cfs_rq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5205) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5206)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5207) /* update runtime allocation */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5208) account_cfs_rq_runtime(cfs_rq, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5209) if (cfs_rq->runtime_remaining <= 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5210) throttle_cfs_rq(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5211) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5212)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5213) static void sync_throttle(struct task_group *tg, int cpu)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5214) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5215) struct cfs_rq *pcfs_rq, *cfs_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5216)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5217) if (!cfs_bandwidth_used())
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5218) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5219)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5220) if (!tg->parent)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5221) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5222)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5223) cfs_rq = tg->cfs_rq[cpu];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5224) pcfs_rq = tg->parent->cfs_rq[cpu];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5225)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5226) cfs_rq->throttle_count = pcfs_rq->throttle_count;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5227) cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5228) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5229)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5230) /* conditionally throttle active cfs_rq's from put_prev_entity() */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5231) static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5232) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5233) if (!cfs_bandwidth_used())
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5234) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5235)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5236) if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5237) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5238)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5239) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5240) * it's possible for a throttled entity to be forced into a running
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5241) * state (e.g. set_curr_task), in this case we're finished.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5242) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5243) if (cfs_rq_throttled(cfs_rq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5244) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5245)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5246) return throttle_cfs_rq(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5247) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5248)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5249) static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5250) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5251) struct cfs_bandwidth *cfs_b =
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5252) container_of(timer, struct cfs_bandwidth, slack_timer);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5253)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5254) do_sched_cfs_slack_timer(cfs_b);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5255)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5256) return HRTIMER_NORESTART;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5257) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5258)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5259) extern const u64 max_cfs_quota_period;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5260)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5261) static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5262) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5263) struct cfs_bandwidth *cfs_b =
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5264) container_of(timer, struct cfs_bandwidth, period_timer);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5265) unsigned long flags;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5266) int overrun;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5267) int idle = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5268) int count = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5269)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5270) raw_spin_lock_irqsave(&cfs_b->lock, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5271) for (;;) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5272) overrun = hrtimer_forward_now(timer, cfs_b->period);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5273) if (!overrun)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5274) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5275)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5276) idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5277)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5278) if (++count > 3) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5279) u64 new, old = ktime_to_ns(cfs_b->period);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5280)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5281) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5282) * Grow period by a factor of 2 to avoid losing precision.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5283) * Precision loss in the quota/period ratio can cause __cfs_schedulable
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5284) * to fail.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5285) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5286) new = old * 2;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5287) if (new < max_cfs_quota_period) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5288) cfs_b->period = ns_to_ktime(new);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5289) cfs_b->quota *= 2;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5290)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5291) pr_warn_ratelimited(
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5292) "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5293) smp_processor_id(),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5294) div_u64(new, NSEC_PER_USEC),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5295) div_u64(cfs_b->quota, NSEC_PER_USEC));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5296) } else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5297) pr_warn_ratelimited(
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5298) "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5299) smp_processor_id(),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5300) div_u64(old, NSEC_PER_USEC),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5301) div_u64(cfs_b->quota, NSEC_PER_USEC));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5302) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5303)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5304) /* reset count so we don't come right back in here */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5305) count = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5306) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5307) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5308) if (idle)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5309) cfs_b->period_active = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5310) raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5311)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5312) return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5313) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5314)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5315) void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5316) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5317) raw_spin_lock_init(&cfs_b->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5318) cfs_b->runtime = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5319) cfs_b->quota = RUNTIME_INF;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5320) cfs_b->period = ns_to_ktime(default_cfs_period());
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5321)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5322) INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5323) hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5324) cfs_b->period_timer.function = sched_cfs_period_timer;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5325) hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5326) cfs_b->slack_timer.function = sched_cfs_slack_timer;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5327) cfs_b->slack_started = false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5328) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5329)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5330) static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5331) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5332) cfs_rq->runtime_enabled = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5333) INIT_LIST_HEAD(&cfs_rq->throttled_list);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5334) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5335)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5336) void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5337) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5338) lockdep_assert_held(&cfs_b->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5339)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5340) if (cfs_b->period_active)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5341) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5342)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5343) cfs_b->period_active = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5344) hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5345) hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5346) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5347)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5348) static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5349) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5350) /* init_cfs_bandwidth() was not called */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5351) if (!cfs_b->throttled_cfs_rq.next)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5352) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5353)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5354) hrtimer_cancel(&cfs_b->period_timer);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5355) hrtimer_cancel(&cfs_b->slack_timer);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5356) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5357)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5358) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5359) * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5360) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5361) * The race is harmless, since modifying bandwidth settings of unhooked group
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5362) * bits doesn't do much.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5363) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5364)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5365) /* cpu online calback */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5366) static void __maybe_unused update_runtime_enabled(struct rq *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5367) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5368) struct task_group *tg;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5369)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5370) lockdep_assert_held(&rq->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5371)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5372) rcu_read_lock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5373) list_for_each_entry_rcu(tg, &task_groups, list) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5374) struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5375) struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5376)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5377) raw_spin_lock(&cfs_b->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5378) cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5379) raw_spin_unlock(&cfs_b->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5380) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5381) rcu_read_unlock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5382) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5383)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5384) /* cpu offline callback */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5385) static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5386) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5387) struct task_group *tg;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5388)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5389) lockdep_assert_held(&rq->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5390)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5391) rcu_read_lock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5392) list_for_each_entry_rcu(tg, &task_groups, list) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5393) struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5394)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5395) if (!cfs_rq->runtime_enabled)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5396) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5397)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5398) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5399) * clock_task is not advancing so we just need to make sure
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5400) * there's some valid quota amount
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5401) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5402) cfs_rq->runtime_remaining = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5403) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5404) * Offline rq is schedulable till CPU is completely disabled
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5405) * in take_cpu_down(), so we prevent new cfs throttling here.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5406) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5407) cfs_rq->runtime_enabled = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5408)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5409) if (cfs_rq_throttled(cfs_rq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5410) unthrottle_cfs_rq(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5411) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5412) rcu_read_unlock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5413) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5414)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5415) #else /* CONFIG_CFS_BANDWIDTH */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5416)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5417) static inline bool cfs_bandwidth_used(void)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5418) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5419) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5420) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5421)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5422) static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5423) static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5424) static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5425) static inline void sync_throttle(struct task_group *tg, int cpu) {}
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5426) static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5427)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5428) static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5429) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5430) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5431) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5432)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5433) static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5434) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5435) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5436) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5437)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5438) static inline int throttled_lb_pair(struct task_group *tg,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5439) int src_cpu, int dest_cpu)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5440) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5441) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5442) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5443)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5444) void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5445)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5446) #ifdef CONFIG_FAIR_GROUP_SCHED
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5447) static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5448) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5449)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5450) static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5451) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5452) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5453) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5454) static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5455) static inline void update_runtime_enabled(struct rq *rq) {}
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5456) static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5457)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5458) #endif /* CONFIG_CFS_BANDWIDTH */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5459)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5460) /**************************************************
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5461) * CFS operations on tasks:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5462) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5463)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5464) #ifdef CONFIG_SCHED_HRTICK
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5465) static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5466) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5467) struct sched_entity *se = &p->se;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5468) struct cfs_rq *cfs_rq = cfs_rq_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5469)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5470) SCHED_WARN_ON(task_rq(p) != rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5471)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5472) if (rq->cfs.h_nr_running > 1) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5473) u64 slice = sched_slice(cfs_rq, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5474) u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5475) s64 delta = slice - ran;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5476)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5477) if (delta < 0) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5478) if (rq->curr == p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5479) resched_curr(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5480) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5481) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5482) hrtick_start(rq, delta);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5483) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5484) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5485)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5486) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5487) * called from enqueue/dequeue and updates the hrtick when the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5488) * current task is from our class and nr_running is low enough
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5489) * to matter.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5490) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5491) static void hrtick_update(struct rq *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5492) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5493) struct task_struct *curr = rq->curr;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5494)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5495) if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5496) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5497)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5498) if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5499) hrtick_start_fair(rq, curr);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5500) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5501) #else /* !CONFIG_SCHED_HRTICK */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5502) static inline void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5503) hrtick_start_fair(struct rq *rq, struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5504) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5505) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5506)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5507) static inline void hrtick_update(struct rq *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5508) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5509) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5510) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5511)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5512) #ifdef CONFIG_SMP
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5513) static inline unsigned long cpu_util(int cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5514)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5515) static inline bool cpu_overutilized(int cpu)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5516) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5517) int overutilized = -1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5518)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5519) trace_android_rvh_cpu_overutilized(cpu, &overutilized);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5520) if (overutilized != -1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5521) return overutilized;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5522)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5523) return !fits_capacity(cpu_util(cpu), capacity_of(cpu));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5524) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5525)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5526) static inline void update_overutilized_status(struct rq *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5527) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5528) if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5529) WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5530) trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5531) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5532) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5533) #else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5534) static inline void update_overutilized_status(struct rq *rq) { }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5535) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5536)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5537) /* Runqueue only has SCHED_IDLE tasks enqueued */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5538) static int sched_idle_rq(struct rq *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5539) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5540) return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5541) rq->nr_running);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5542) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5543)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5544) #ifdef CONFIG_SMP
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5545) static int sched_idle_cpu(int cpu)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5546) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5547) return sched_idle_rq(cpu_rq(cpu));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5548) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5549) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5550)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5551) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5552) * The enqueue_task method is called before nr_running is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5553) * increased. Here we update the fair scheduling stats and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5554) * then put the task into the rbtree:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5555) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5556) static void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5557) enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5558) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5559) struct cfs_rq *cfs_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5560) struct sched_entity *se = &p->se;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5561) int idle_h_nr_running = task_has_idle_policy(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5562) int task_new = !(flags & ENQUEUE_WAKEUP);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5563) int should_iowait_boost;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5564)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5565) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5566) * The code below (indirectly) updates schedutil which looks at
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5567) * the cfs_rq utilization to select a frequency.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5568) * Let's add the task's estimated utilization to the cfs_rq's
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5569) * estimated utilization, before we update schedutil.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5570) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5571) util_est_enqueue(&rq->cfs, p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5572)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5573) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5574) * If in_iowait is set, the code below may not trigger any cpufreq
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5575) * utilization updates, so do it here explicitly with the IOWAIT flag
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5576) * passed.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5577) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5578) should_iowait_boost = p->in_iowait;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5579) trace_android_rvh_set_iowait(p, &should_iowait_boost);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5580) if (should_iowait_boost)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5581) cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5582)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5583) for_each_sched_entity(se) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5584) if (se->on_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5585) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5586) cfs_rq = cfs_rq_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5587) enqueue_entity(cfs_rq, se, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5588)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5589) cfs_rq->h_nr_running++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5590) cfs_rq->idle_h_nr_running += idle_h_nr_running;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5591)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5592) /* end evaluation on encountering a throttled cfs_rq */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5593) if (cfs_rq_throttled(cfs_rq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5594) goto enqueue_throttle;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5595)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5596) flags = ENQUEUE_WAKEUP;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5597) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5598)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5599) trace_android_rvh_enqueue_task_fair(rq, p, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5600) for_each_sched_entity(se) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5601) cfs_rq = cfs_rq_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5602)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5603) update_load_avg(cfs_rq, se, UPDATE_TG);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5604) se_update_runnable(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5605) update_cfs_group(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5606)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5607) cfs_rq->h_nr_running++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5608) cfs_rq->idle_h_nr_running += idle_h_nr_running;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5609)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5610) /* end evaluation on encountering a throttled cfs_rq */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5611) if (cfs_rq_throttled(cfs_rq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5612) goto enqueue_throttle;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5613)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5614) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5615) * One parent has been throttled and cfs_rq removed from the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5616) * list. Add it back to not break the leaf list.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5617) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5618) if (throttled_hierarchy(cfs_rq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5619) list_add_leaf_cfs_rq(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5620) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5621)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5622) /* At this point se is NULL and we are at root level*/
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5623) add_nr_running(rq, 1);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5624)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5625) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5626) * Since new tasks are assigned an initial util_avg equal to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5627) * half of the spare capacity of their CPU, tiny tasks have the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5628) * ability to cross the overutilized threshold, which will
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5629) * result in the load balancer ruining all the task placement
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5630) * done by EAS. As a way to mitigate that effect, do not account
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5631) * for the first enqueue operation of new tasks during the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5632) * overutilized flag detection.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5633) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5634) * A better way of solving this problem would be to wait for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5635) * the PELT signals of tasks to converge before taking them
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5636) * into account, but that is not straightforward to implement,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5637) * and the following generally works well enough in practice.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5638) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5639) if (!task_new)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5640) update_overutilized_status(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5641)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5642) enqueue_throttle:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5643) if (cfs_bandwidth_used()) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5644) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5645) * When bandwidth control is enabled; the cfs_rq_throttled()
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5646) * breaks in the above iteration can result in incomplete
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5647) * leaf list maintenance, resulting in triggering the assertion
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5648) * below.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5649) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5650) for_each_sched_entity(se) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5651) cfs_rq = cfs_rq_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5652)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5653) if (list_add_leaf_cfs_rq(cfs_rq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5654) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5655) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5656) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5657)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5658) assert_list_leaf_cfs_rq(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5659)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5660) hrtick_update(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5661) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5662)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5663) static void set_next_buddy(struct sched_entity *se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5664)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5665) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5666) * The dequeue_task method is called before nr_running is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5667) * decreased. We remove the task from the rbtree and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5668) * update the fair scheduling stats:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5669) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5670) static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5671) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5672) struct cfs_rq *cfs_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5673) struct sched_entity *se = &p->se;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5674) int task_sleep = flags & DEQUEUE_SLEEP;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5675) int idle_h_nr_running = task_has_idle_policy(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5676) bool was_sched_idle = sched_idle_rq(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5677)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5678) util_est_dequeue(&rq->cfs, p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5679)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5680) for_each_sched_entity(se) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5681) cfs_rq = cfs_rq_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5682) dequeue_entity(cfs_rq, se, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5683)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5684) cfs_rq->h_nr_running--;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5685) cfs_rq->idle_h_nr_running -= idle_h_nr_running;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5686)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5687) /* end evaluation on encountering a throttled cfs_rq */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5688) if (cfs_rq_throttled(cfs_rq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5689) goto dequeue_throttle;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5690)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5691) /* Don't dequeue parent if it has other entities besides us */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5692) if (cfs_rq->load.weight) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5693) /* Avoid re-evaluating load for this entity: */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5694) se = parent_entity(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5695) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5696) * Bias pick_next to pick a task from this cfs_rq, as
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5697) * p is sleeping when it is within its sched_slice.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5698) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5699) if (task_sleep && se && !throttled_hierarchy(cfs_rq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5700) set_next_buddy(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5701) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5702) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5703) flags |= DEQUEUE_SLEEP;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5704) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5705)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5706) trace_android_rvh_dequeue_task_fair(rq, p, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5707) for_each_sched_entity(se) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5708) cfs_rq = cfs_rq_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5709)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5710) update_load_avg(cfs_rq, se, UPDATE_TG);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5711) se_update_runnable(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5712) update_cfs_group(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5713)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5714) cfs_rq->h_nr_running--;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5715) cfs_rq->idle_h_nr_running -= idle_h_nr_running;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5716)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5717) /* end evaluation on encountering a throttled cfs_rq */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5718) if (cfs_rq_throttled(cfs_rq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5719) goto dequeue_throttle;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5720)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5721) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5722)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5723) /* At this point se is NULL and we are at root level*/
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5724) sub_nr_running(rq, 1);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5725)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5726) /* balance early to pull high priority tasks */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5727) if (unlikely(!was_sched_idle && sched_idle_rq(rq)))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5728) rq->next_balance = jiffies;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5729)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5730) dequeue_throttle:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5731) util_est_update(&rq->cfs, p, task_sleep);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5732) hrtick_update(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5733) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5734)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5735) #ifdef CONFIG_SMP
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5736)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5737) /* Working cpumask for: load_balance, load_balance_newidle. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5738) DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5739) DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5740)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5741) #ifdef CONFIG_NO_HZ_COMMON
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5742)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5743) static struct {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5744) cpumask_var_t idle_cpus_mask;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5745) atomic_t nr_cpus;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5746) int has_blocked; /* Idle CPUS has blocked load */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5747) unsigned long next_balance; /* in jiffy units */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5748) unsigned long next_blocked; /* Next update of blocked load in jiffies */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5749) } nohz ____cacheline_aligned;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5750)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5751) #endif /* CONFIG_NO_HZ_COMMON */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5752)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5753) static unsigned long cpu_load(struct rq *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5754) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5755) return cfs_rq_load_avg(&rq->cfs);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5756) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5757)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5758) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5759) * cpu_load_without - compute CPU load without any contributions from *p
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5760) * @cpu: the CPU which load is requested
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5761) * @p: the task which load should be discounted
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5762) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5763) * The load of a CPU is defined by the load of tasks currently enqueued on that
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5764) * CPU as well as tasks which are currently sleeping after an execution on that
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5765) * CPU.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5766) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5767) * This method returns the load of the specified CPU by discounting the load of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5768) * the specified task, whenever the task is currently contributing to the CPU
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5769) * load.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5770) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5771) static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5772) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5773) struct cfs_rq *cfs_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5774) unsigned int load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5775)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5776) /* Task has no contribution or is new */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5777) if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5778) return cpu_load(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5779)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5780) cfs_rq = &rq->cfs;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5781) load = READ_ONCE(cfs_rq->avg.load_avg);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5782)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5783) /* Discount task's util from CPU's util */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5784) lsub_positive(&load, task_h_load(p));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5785)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5786) return load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5787) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5788)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5789) static unsigned long cpu_runnable(struct rq *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5790) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5791) return cfs_rq_runnable_avg(&rq->cfs);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5792) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5793)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5794) static unsigned long cpu_runnable_without(struct rq *rq, struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5795) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5796) struct cfs_rq *cfs_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5797) unsigned int runnable;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5798)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5799) /* Task has no contribution or is new */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5800) if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5801) return cpu_runnable(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5802)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5803) cfs_rq = &rq->cfs;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5804) runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5805)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5806) /* Discount task's runnable from CPU's runnable */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5807) lsub_positive(&runnable, p->se.avg.runnable_avg);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5808)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5809) return runnable;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5810) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5811)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5812) static unsigned long capacity_of(int cpu)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5813) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5814) return cpu_rq(cpu)->cpu_capacity;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5815) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5816)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5817) static void record_wakee(struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5818) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5819) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5820) * Only decay a single time; tasks that have less then 1 wakeup per
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5821) * jiffy will not have built up many flips.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5822) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5823) if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5824) current->wakee_flips >>= 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5825) current->wakee_flip_decay_ts = jiffies;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5826) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5827)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5828) if (current->last_wakee != p) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5829) current->last_wakee = p;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5830) current->wakee_flips++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5831) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5832) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5833)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5834) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5835) * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5836) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5837) * A waker of many should wake a different task than the one last awakened
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5838) * at a frequency roughly N times higher than one of its wakees.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5839) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5840) * In order to determine whether we should let the load spread vs consolidating
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5841) * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5842) * partner, and a factor of lls_size higher frequency in the other.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5843) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5844) * With both conditions met, we can be relatively sure that the relationship is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5845) * non-monogamous, with partner count exceeding socket size.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5846) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5847) * Waker/wakee being client/server, worker/dispatcher, interrupt source or
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5848) * whatever is irrelevant, spread criteria is apparent partner count exceeds
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5849) * socket size.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5850) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5851) static int wake_wide(struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5852) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5853) unsigned int master = current->wakee_flips;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5854) unsigned int slave = p->wakee_flips;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5855) int factor = __this_cpu_read(sd_llc_size);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5856)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5857) if (master < slave)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5858) swap(master, slave);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5859) if (slave < factor || master < slave * factor)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5860) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5861) return 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5862) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5863)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5864) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5865) * The purpose of wake_affine() is to quickly determine on which CPU we can run
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5866) * soonest. For the purpose of speed we only consider the waking and previous
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5867) * CPU.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5868) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5869) * wake_affine_idle() - only considers 'now', it check if the waking CPU is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5870) * cache-affine and is (or will be) idle.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5871) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5872) * wake_affine_weight() - considers the weight to reflect the average
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5873) * scheduling latency of the CPUs. This seems to work
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5874) * for the overloaded case.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5875) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5876) static int
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5877) wake_affine_idle(int this_cpu, int prev_cpu, int sync)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5878) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5879) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5880) * If this_cpu is idle, it implies the wakeup is from interrupt
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5881) * context. Only allow the move if cache is shared. Otherwise an
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5882) * interrupt intensive workload could force all tasks onto one
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5883) * node depending on the IO topology or IRQ affinity settings.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5884) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5885) * If the prev_cpu is idle and cache affine then avoid a migration.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5886) * There is no guarantee that the cache hot data from an interrupt
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5887) * is more important than cache hot data on the prev_cpu and from
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5888) * a cpufreq perspective, it's better to have higher utilisation
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5889) * on one CPU.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5890) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5891) if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5892) return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5893)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5894) if (sync && cpu_rq(this_cpu)->nr_running == 1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5895) return this_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5896)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5897) return nr_cpumask_bits;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5898) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5899)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5900) static int
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5901) wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5902) int this_cpu, int prev_cpu, int sync)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5903) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5904) s64 this_eff_load, prev_eff_load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5905) unsigned long task_load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5906)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5907) this_eff_load = cpu_load(cpu_rq(this_cpu));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5908)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5909) if (sync) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5910) unsigned long current_load = task_h_load(current);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5911)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5912) if (current_load > this_eff_load)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5913) return this_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5914)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5915) this_eff_load -= current_load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5916) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5917)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5918) task_load = task_h_load(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5919)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5920) this_eff_load += task_load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5921) if (sched_feat(WA_BIAS))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5922) this_eff_load *= 100;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5923) this_eff_load *= capacity_of(prev_cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5924)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5925) prev_eff_load = cpu_load(cpu_rq(prev_cpu));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5926) prev_eff_load -= task_load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5927) if (sched_feat(WA_BIAS))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5928) prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5929) prev_eff_load *= capacity_of(this_cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5930)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5931) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5932) * If sync, adjust the weight of prev_eff_load such that if
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5933) * prev_eff == this_eff that select_idle_sibling() will consider
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5934) * stacking the wakee on top of the waker if no other CPU is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5935) * idle.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5936) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5937) if (sync)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5938) prev_eff_load += 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5939)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5940) return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5941) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5942)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5943) static int wake_affine(struct sched_domain *sd, struct task_struct *p,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5944) int this_cpu, int prev_cpu, int sync)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5945) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5946) int target = nr_cpumask_bits;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5947)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5948) if (sched_feat(WA_IDLE))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5949) target = wake_affine_idle(this_cpu, prev_cpu, sync);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5950)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5951) if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5952) target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5953)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5954) schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5955) if (target == nr_cpumask_bits)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5956) return prev_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5957)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5958) schedstat_inc(sd->ttwu_move_affine);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5959) schedstat_inc(p->se.statistics.nr_wakeups_affine);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5960) return target;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5961) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5962)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5963) static struct sched_group *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5964) find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5965)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5966) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5967) * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5968) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5969) static int
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5970) find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5971) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5972) unsigned long load, min_load = ULONG_MAX;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5973) unsigned int min_exit_latency = UINT_MAX;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5974) u64 latest_idle_timestamp = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5975) int least_loaded_cpu = this_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5976) int shallowest_idle_cpu = -1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5977) int i;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5978)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5979) /* Check if we have any choice: */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5980) if (group->group_weight == 1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5981) return cpumask_first(sched_group_span(group));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5982)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5983) /* Traverse only the allowed CPUs */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5984) for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5985) if (sched_idle_cpu(i))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5986) return i;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5987)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5988) if (available_idle_cpu(i)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5989) struct rq *rq = cpu_rq(i);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5990) struct cpuidle_state *idle = idle_get_state(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5991) if (idle && idle->exit_latency < min_exit_latency) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5992) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5993) * We give priority to a CPU whose idle state
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5994) * has the smallest exit latency irrespective
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5995) * of any idle timestamp.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5996) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5997) min_exit_latency = idle->exit_latency;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5998) latest_idle_timestamp = rq->idle_stamp;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5999) shallowest_idle_cpu = i;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6000) } else if ((!idle || idle->exit_latency == min_exit_latency) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6001) rq->idle_stamp > latest_idle_timestamp) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6002) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6003) * If equal or no active idle state, then
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6004) * the most recently idled CPU might have
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6005) * a warmer cache.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6006) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6007) latest_idle_timestamp = rq->idle_stamp;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6008) shallowest_idle_cpu = i;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6009) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6010) } else if (shallowest_idle_cpu == -1) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6011) load = cpu_load(cpu_rq(i));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6012) if (load < min_load) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6013) min_load = load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6014) least_loaded_cpu = i;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6015) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6016) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6017) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6018)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6019) return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6020) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6021)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6022) static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6023) int cpu, int prev_cpu, int sd_flag)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6024) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6025) int new_cpu = cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6026)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6027) if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6028) return prev_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6029)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6030) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6031) * We need task's util for cpu_util_without, sync it up to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6032) * prev_cpu's last_update_time.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6033) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6034) if (!(sd_flag & SD_BALANCE_FORK))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6035) sync_entity_load_avg(&p->se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6036)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6037) while (sd) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6038) struct sched_group *group;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6039) struct sched_domain *tmp;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6040) int weight;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6041)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6042) if (!(sd->flags & sd_flag)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6043) sd = sd->child;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6044) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6045) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6046)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6047) group = find_idlest_group(sd, p, cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6048) if (!group) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6049) sd = sd->child;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6050) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6051) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6052)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6053) new_cpu = find_idlest_group_cpu(group, p, cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6054) if (new_cpu == cpu) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6055) /* Now try balancing at a lower domain level of 'cpu': */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6056) sd = sd->child;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6057) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6058) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6059)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6060) /* Now try balancing at a lower domain level of 'new_cpu': */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6061) cpu = new_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6062) weight = sd->span_weight;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6063) sd = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6064) for_each_domain(cpu, tmp) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6065) if (weight <= tmp->span_weight)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6066) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6067) if (tmp->flags & sd_flag)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6068) sd = tmp;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6069) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6070) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6071)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6072) return new_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6073) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6074)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6075) #ifdef CONFIG_SCHED_SMT
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6076) DEFINE_STATIC_KEY_FALSE(sched_smt_present);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6077) EXPORT_SYMBOL_GPL(sched_smt_present);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6078)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6079) static inline void set_idle_cores(int cpu, int val)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6080) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6081) struct sched_domain_shared *sds;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6082)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6083) sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6084) if (sds)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6085) WRITE_ONCE(sds->has_idle_cores, val);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6086) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6087)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6088) static inline bool test_idle_cores(int cpu, bool def)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6089) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6090) struct sched_domain_shared *sds;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6091)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6092) sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6093) if (sds)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6094) return READ_ONCE(sds->has_idle_cores);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6095)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6096) return def;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6097) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6098)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6099) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6100) * Scans the local SMT mask to see if the entire core is idle, and records this
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6101) * information in sd_llc_shared->has_idle_cores.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6102) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6103) * Since SMT siblings share all cache levels, inspecting this limited remote
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6104) * state should be fairly cheap.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6105) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6106) void __update_idle_core(struct rq *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6107) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6108) int core = cpu_of(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6109) int cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6110)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6111) rcu_read_lock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6112) if (test_idle_cores(core, true))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6113) goto unlock;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6114)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6115) for_each_cpu(cpu, cpu_smt_mask(core)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6116) if (cpu == core)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6117) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6118)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6119) if (!available_idle_cpu(cpu))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6120) goto unlock;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6121) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6122)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6123) set_idle_cores(core, 1);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6124) unlock:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6125) rcu_read_unlock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6126) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6127)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6128) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6129) * Scan the entire LLC domain for idle cores; this dynamically switches off if
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6130) * there are no idle cores left in the system; tracked through
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6131) * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6132) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6133) static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6134) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6135) struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6136) int core, cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6137)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6138) if (!static_branch_likely(&sched_smt_present))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6139) return -1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6140)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6141) if (!test_idle_cores(target, false))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6142) return -1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6143)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6144) cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6145)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6146) for_each_cpu_wrap(core, cpus, target) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6147) bool idle = true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6148)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6149) for_each_cpu(cpu, cpu_smt_mask(core)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6150) if (!available_idle_cpu(cpu)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6151) idle = false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6152) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6153) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6154) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6155) cpumask_andnot(cpus, cpus, cpu_smt_mask(core));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6156)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6157) if (idle)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6158) return core;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6159) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6160)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6161) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6162) * Failed to find an idle core; stop looking for one.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6163) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6164) set_idle_cores(target, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6165)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6166) return -1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6167) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6168)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6169) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6170) * Scan the local SMT mask for idle CPUs.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6171) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6172) static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6173) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6174) int cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6175)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6176) if (!static_branch_likely(&sched_smt_present))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6177) return -1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6178)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6179) for_each_cpu(cpu, cpu_smt_mask(target)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6180) if (!cpumask_test_cpu(cpu, p->cpus_ptr) ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6181) !cpumask_test_cpu(cpu, sched_domain_span(sd)))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6182) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6183) if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6184) return cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6185) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6186)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6187) return -1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6188) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6189)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6190) #else /* CONFIG_SCHED_SMT */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6191)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6192) static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6193) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6194) return -1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6195) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6196)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6197) static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6198) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6199) return -1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6200) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6201)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6202) #endif /* CONFIG_SCHED_SMT */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6203)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6204) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6205) * Scan the LLC domain for idle CPUs; this is dynamically regulated by
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6206) * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6207) * average idle time for this rq (as found in rq->avg_idle).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6208) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6209) static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6210) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6211) struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6212) struct sched_domain *this_sd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6213) u64 avg_cost, avg_idle;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6214) u64 time;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6215) int this = smp_processor_id();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6216) int cpu, nr = INT_MAX;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6217)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6218) this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6219) if (!this_sd)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6220) return -1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6221)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6222) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6223) * Due to large variance we need a large fuzz factor; hackbench in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6224) * particularly is sensitive here.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6225) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6226) avg_idle = this_rq()->avg_idle / 512;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6227) avg_cost = this_sd->avg_scan_cost + 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6228)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6229) if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6230) return -1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6231)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6232) if (sched_feat(SIS_PROP)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6233) u64 span_avg = sd->span_weight * avg_idle;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6234) if (span_avg > 4*avg_cost)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6235) nr = div_u64(span_avg, avg_cost);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6236) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6237) nr = 4;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6238) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6239)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6240) time = cpu_clock(this);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6241)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6242) cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6243)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6244) for_each_cpu_wrap(cpu, cpus, target) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6245) if (!--nr)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6246) return -1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6247) if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6248) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6249) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6250)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6251) time = cpu_clock(this) - time;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6252) update_avg(&this_sd->avg_scan_cost, time);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6253)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6254) return cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6255) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6256)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6257) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6258) * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6259) * the task fits. If no CPU is big enough, but there are idle ones, try to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6260) * maximize capacity.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6261) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6262) static int
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6263) select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6264) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6265) unsigned long task_util, best_cap = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6266) int cpu, best_cpu = -1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6267) struct cpumask *cpus;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6268)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6269) cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6270) cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6271)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6272) task_util = uclamp_task_util(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6273)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6274) for_each_cpu_wrap(cpu, cpus, target) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6275) unsigned long cpu_cap = capacity_of(cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6276)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6277) if (!available_idle_cpu(cpu) && !sched_idle_cpu(cpu))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6278) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6279) if (fits_capacity(task_util, cpu_cap))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6280) return cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6281)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6282) if (cpu_cap > best_cap) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6283) best_cap = cpu_cap;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6284) best_cpu = cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6285) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6286) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6287)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6288) return best_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6289) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6290)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6291) static inline bool asym_fits_capacity(int task_util, int cpu)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6292) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6293) if (static_branch_unlikely(&sched_asym_cpucapacity))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6294) return fits_capacity(task_util, capacity_of(cpu));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6295)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6296) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6297) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6298)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6299) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6300) * Try and locate an idle core/thread in the LLC cache domain.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6301) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6302) static int select_idle_sibling(struct task_struct *p, int prev, int target)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6303) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6304) struct sched_domain *sd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6305) unsigned long task_util;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6306) int i, recent_used_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6307)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6308) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6309) * On asymmetric system, update task utilization because we will check
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6310) * that the task fits with cpu's capacity.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6311) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6312) if (static_branch_unlikely(&sched_asym_cpucapacity)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6313) sync_entity_load_avg(&p->se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6314) task_util = uclamp_task_util(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6315) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6316)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6317) if ((available_idle_cpu(target) || sched_idle_cpu(target)) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6318) asym_fits_capacity(task_util, target))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6319) return target;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6320)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6321) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6322) * If the previous CPU is cache affine and idle, don't be stupid:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6323) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6324) if (prev != target && cpus_share_cache(prev, target) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6325) (available_idle_cpu(prev) || sched_idle_cpu(prev)) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6326) asym_fits_capacity(task_util, prev))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6327) return prev;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6328)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6329) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6330) * Allow a per-cpu kthread to stack with the wakee if the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6331) * kworker thread and the tasks previous CPUs are the same.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6332) * The assumption is that the wakee queued work for the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6333) * per-cpu kthread that is now complete and the wakeup is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6334) * essentially a sync wakeup. An obvious example of this
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6335) * pattern is IO completions.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6336) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6337) if (is_per_cpu_kthread(current) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6338) in_task() &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6339) prev == smp_processor_id() &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6340) this_rq()->nr_running <= 1 &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6341) asym_fits_capacity(task_util, prev)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6342) return prev;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6343) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6344)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6345) /* Check a recently used CPU as a potential idle candidate: */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6346) recent_used_cpu = p->recent_used_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6347) if (recent_used_cpu != prev &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6348) recent_used_cpu != target &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6349) cpus_share_cache(recent_used_cpu, target) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6350) (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6351) cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6352) asym_fits_capacity(task_util, recent_used_cpu)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6353) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6354) * Replace recent_used_cpu with prev as it is a potential
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6355) * candidate for the next wake:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6356) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6357) p->recent_used_cpu = prev;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6358) return recent_used_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6359) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6360)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6361) if (IS_ENABLED(CONFIG_ROCKCHIP_PERFORMANCE)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6362) if (rockchip_perf_get_level() == ROCKCHIP_PERFORMANCE_HIGH)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6363) goto sd_llc;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6364) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6365)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6366) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6367) * For asymmetric CPU capacity systems, our domain of interest is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6368) * sd_asym_cpucapacity rather than sd_llc.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6369) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6370) if (static_branch_unlikely(&sched_asym_cpucapacity)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6371) sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, target));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6372) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6373) * On an asymmetric CPU capacity system where an exclusive
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6374) * cpuset defines a symmetric island (i.e. one unique
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6375) * capacity_orig value through the cpuset), the key will be set
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6376) * but the CPUs within that cpuset will not have a domain with
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6377) * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6378) * capacity path.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6379) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6380) if (sd) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6381) i = select_idle_capacity(p, sd, target);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6382) return ((unsigned)i < nr_cpumask_bits) ? i : target;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6383) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6384) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6385)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6386) sd_llc:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6387) sd = rcu_dereference(per_cpu(sd_llc, target));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6388) if (!sd)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6389) return target;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6390)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6391) i = select_idle_core(p, sd, target);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6392) if ((unsigned)i < nr_cpumask_bits)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6393) return i;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6394)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6395) i = select_idle_cpu(p, sd, target);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6396) if ((unsigned)i < nr_cpumask_bits)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6397) return i;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6398)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6399) i = select_idle_smt(p, sd, target);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6400) if ((unsigned)i < nr_cpumask_bits)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6401) return i;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6402)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6403) return target;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6404) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6405)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6406) /**
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6407) * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6408) * @cpu: the CPU to get the utilization of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6409) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6410) * The unit of the return value must be the one of capacity so we can compare
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6411) * the utilization with the capacity of the CPU that is available for CFS task
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6412) * (ie cpu_capacity).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6413) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6414) * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6415) * recent utilization of currently non-runnable tasks on a CPU. It represents
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6416) * the amount of utilization of a CPU in the range [0..capacity_orig] where
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6417) * capacity_orig is the cpu_capacity available at the highest frequency
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6418) * (arch_scale_freq_capacity()).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6419) * The utilization of a CPU converges towards a sum equal to or less than the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6420) * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6421) * the running time on this CPU scaled by capacity_curr.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6422) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6423) * The estimated utilization of a CPU is defined to be the maximum between its
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6424) * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6425) * currently RUNNABLE on that CPU.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6426) * This allows to properly represent the expected utilization of a CPU which
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6427) * has just got a big task running since a long sleep period. At the same time
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6428) * however it preserves the benefits of the "blocked utilization" in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6429) * describing the potential for other tasks waking up on the same CPU.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6430) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6431) * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6432) * higher than capacity_orig because of unfortunate rounding in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6433) * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6434) * the average stabilizes with the new running time. We need to check that the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6435) * utilization stays within the range of [0..capacity_orig] and cap it if
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6436) * necessary. Without utilization capping, a group could be seen as overloaded
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6437) * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6438) * available capacity. We allow utilization to overshoot capacity_curr (but not
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6439) * capacity_orig) as it useful for predicting the capacity required after task
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6440) * migrations (scheduler-driven DVFS).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6441) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6442) * Return: the (estimated) utilization for the specified CPU
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6443) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6444) static inline unsigned long cpu_util(int cpu)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6445) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6446) struct cfs_rq *cfs_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6447) unsigned int util;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6448)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6449) cfs_rq = &cpu_rq(cpu)->cfs;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6450) util = READ_ONCE(cfs_rq->avg.util_avg);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6451)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6452) if (sched_feat(UTIL_EST))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6453) util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6454)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6455) return min_t(unsigned long, util, capacity_orig_of(cpu));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6456) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6457)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6458) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6459) * cpu_util_without: compute cpu utilization without any contributions from *p
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6460) * @cpu: the CPU which utilization is requested
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6461) * @p: the task which utilization should be discounted
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6462) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6463) * The utilization of a CPU is defined by the utilization of tasks currently
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6464) * enqueued on that CPU as well as tasks which are currently sleeping after an
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6465) * execution on that CPU.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6466) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6467) * This method returns the utilization of the specified CPU by discounting the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6468) * utilization of the specified task, whenever the task is currently
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6469) * contributing to the CPU utilization.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6470) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6471) static unsigned long cpu_util_without(int cpu, struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6472) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6473) struct cfs_rq *cfs_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6474) unsigned int util;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6475)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6476) /* Task has no contribution or is new */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6477) if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6478) return cpu_util(cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6479)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6480) cfs_rq = &cpu_rq(cpu)->cfs;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6481) util = READ_ONCE(cfs_rq->avg.util_avg);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6482)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6483) /* Discount task's util from CPU's util */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6484) lsub_positive(&util, task_util(p));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6485)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6486) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6487) * Covered cases:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6488) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6489) * a) if *p is the only task sleeping on this CPU, then:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6490) * cpu_util (== task_util) > util_est (== 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6491) * and thus we return:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6492) * cpu_util_without = (cpu_util - task_util) = 0
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6493) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6494) * b) if other tasks are SLEEPING on this CPU, which is now exiting
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6495) * IDLE, then:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6496) * cpu_util >= task_util
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6497) * cpu_util > util_est (== 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6498) * and thus we discount *p's blocked utilization to return:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6499) * cpu_util_without = (cpu_util - task_util) >= 0
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6500) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6501) * c) if other tasks are RUNNABLE on that CPU and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6502) * util_est > cpu_util
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6503) * then we use util_est since it returns a more restrictive
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6504) * estimation of the spare capacity on that CPU, by just
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6505) * considering the expected utilization of tasks already
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6506) * runnable on that CPU.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6507) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6508) * Cases a) and b) are covered by the above code, while case c) is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6509) * covered by the following code when estimated utilization is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6510) * enabled.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6511) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6512) if (sched_feat(UTIL_EST)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6513) unsigned int estimated =
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6514) READ_ONCE(cfs_rq->avg.util_est.enqueued);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6515)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6516) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6517) * Despite the following checks we still have a small window
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6518) * for a possible race, when an execl's select_task_rq_fair()
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6519) * races with LB's detach_task():
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6520) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6521) * detach_task()
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6522) * p->on_rq = TASK_ON_RQ_MIGRATING;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6523) * ---------------------------------- A
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6524) * deactivate_task() \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6525) * dequeue_task() + RaceTime
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6526) * util_est_dequeue() /
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6527) * ---------------------------------- B
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6528) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6529) * The additional check on "current == p" it's required to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6530) * properly fix the execl regression and it helps in further
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6531) * reducing the chances for the above race.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6532) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6533) if (unlikely(task_on_rq_queued(p) || current == p))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6534) lsub_positive(&estimated, _task_util_est(p));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6535)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6536) util = max(util, estimated);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6537) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6538)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6539) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6540) * Utilization (estimated) can exceed the CPU capacity, thus let's
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6541) * clamp to the maximum CPU capacity to ensure consistency with
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6542) * the cpu_util call.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6543) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6544) return min_t(unsigned long, util, capacity_orig_of(cpu));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6545) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6546)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6547) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6548) * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6549) * to @dst_cpu.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6550) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6551) static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6552) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6553) struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6554) unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6555)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6556) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6557) * If @p migrates from @cpu to another, remove its contribution. Or,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6558) * if @p migrates from another CPU to @cpu, add its contribution. In
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6559) * the other cases, @cpu is not impacted by the migration, so the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6560) * util_avg should already be correct.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6561) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6562) if (task_cpu(p) == cpu && dst_cpu != cpu)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6563) sub_positive(&util, task_util(p));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6564) else if (task_cpu(p) != cpu && dst_cpu == cpu)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6565) util += task_util(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6566)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6567) if (sched_feat(UTIL_EST)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6568) util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6569)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6570) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6571) * During wake-up, the task isn't enqueued yet and doesn't
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6572) * appear in the cfs_rq->avg.util_est.enqueued of any rq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6573) * so just add it (if needed) to "simulate" what will be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6574) * cpu_util() after the task has been enqueued.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6575) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6576) if (dst_cpu == cpu)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6577) util_est += _task_util_est(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6578)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6579) util = max(util, util_est);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6580) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6581)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6582) return min(util, capacity_orig_of(cpu));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6583) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6584)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6585) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6586) * compute_energy(): Estimates the energy that @pd would consume if @p was
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6587) * migrated to @dst_cpu. compute_energy() predicts what will be the utilization
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6588) * landscape of @pd's CPUs after the task migration, and uses the Energy Model
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6589) * to compute what would be the energy if we decided to actually migrate that
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6590) * task.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6591) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6592) static long
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6593) compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6594) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6595) struct cpumask *pd_mask = perf_domain_span(pd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6596) unsigned long cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6597) unsigned long max_util = 0, sum_util = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6598) unsigned long energy = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6599) int cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6600)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6601) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6602) * The capacity state of CPUs of the current rd can be driven by CPUs
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6603) * of another rd if they belong to the same pd. So, account for the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6604) * utilization of these CPUs too by masking pd with cpu_online_mask
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6605) * instead of the rd span.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6606) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6607) * If an entire pd is outside of the current rd, it will not appear in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6608) * its pd list and will not be accounted by compute_energy().
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6609) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6610) for_each_cpu_and(cpu, pd_mask, cpu_online_mask) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6611) unsigned long cpu_util, util_cfs = cpu_util_next(cpu, p, dst_cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6612) struct task_struct *tsk = cpu == dst_cpu ? p : NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6613)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6614) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6615) * Busy time computation: utilization clamping is not
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6616) * required since the ratio (sum_util / cpu_capacity)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6617) * is already enough to scale the EM reported power
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6618) * consumption at the (eventually clamped) cpu_capacity.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6619) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6620) sum_util += schedutil_cpu_util(cpu, util_cfs, cpu_cap,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6621) ENERGY_UTIL, NULL);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6622)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6623) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6624) * Performance domain frequency: utilization clamping
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6625) * must be considered since it affects the selection
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6626) * of the performance domain frequency.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6627) * NOTE: in case RT tasks are running, by default the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6628) * FREQUENCY_UTIL's utilization can be max OPP.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6629) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6630) cpu_util = schedutil_cpu_util(cpu, util_cfs, cpu_cap,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6631) FREQUENCY_UTIL, tsk);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6632) max_util = max(max_util, cpu_util);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6633) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6634)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6635) trace_android_vh_em_cpu_energy(pd->em_pd, max_util, sum_util, &energy);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6636) if (!energy)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6637) energy = em_cpu_energy(pd->em_pd, max_util, sum_util);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6638)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6639) return energy;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6640) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6641)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6642) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6643) * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6644) * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6645) * spare capacity in each performance domain and uses it as a potential
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6646) * candidate to execute the task. Then, it uses the Energy Model to figure
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6647) * out which of the CPU candidates is the most energy-efficient.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6648) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6649) * The rationale for this heuristic is as follows. In a performance domain,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6650) * all the most energy efficient CPU candidates (according to the Energy
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6651) * Model) are those for which we'll request a low frequency. When there are
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6652) * several CPUs for which the frequency request will be the same, we don't
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6653) * have enough data to break the tie between them, because the Energy Model
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6654) * only includes active power costs. With this model, if we assume that
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6655) * frequency requests follow utilization (e.g. using schedutil), the CPU with
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6656) * the maximum spare capacity in a performance domain is guaranteed to be among
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6657) * the best candidates of the performance domain.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6658) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6659) * In practice, it could be preferable from an energy standpoint to pack
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6660) * small tasks on a CPU in order to let other CPUs go in deeper idle states,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6661) * but that could also hurt our chances to go cluster idle, and we have no
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6662) * ways to tell with the current Energy Model if this is actually a good
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6663) * idea or not. So, find_energy_efficient_cpu() basically favors
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6664) * cluster-packing, and spreading inside a cluster. That should at least be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6665) * a good thing for latency, and this is consistent with the idea that most
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6666) * of the energy savings of EAS come from the asymmetry of the system, and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6667) * not so much from breaking the tie between identical CPUs. That's also the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6668) * reason why EAS is enabled in the topology code only for systems where
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6669) * SD_ASYM_CPUCAPACITY is set.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6670) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6671) * NOTE: Forkees are not accepted in the energy-aware wake-up path because
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6672) * they don't have any useful utilization data yet and it's not possible to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6673) * forecast their impact on energy consumption. Consequently, they will be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6674) * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6675) * to be energy-inefficient in some use-cases. The alternative would be to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6676) * bias new tasks towards specific types of CPUs first, or to try to infer
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6677) * their util_avg from the parent task, but those heuristics could hurt
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6678) * other use-cases too. So, until someone finds a better way to solve this,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6679) * let's keep things simple by re-using the existing slow path.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6680) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6681) static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu, int sync)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6682) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6683) unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6684) struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6685) int max_spare_cap_cpu_ls = prev_cpu, best_idle_cpu = -1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6686) unsigned long max_spare_cap_ls = 0, target_cap;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6687) unsigned long cpu_cap, util, base_energy = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6688) bool boosted, latency_sensitive = false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6689) unsigned int min_exit_lat = UINT_MAX;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6690) int cpu, best_energy_cpu = prev_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6691) struct cpuidle_state *idle;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6692) struct sched_domain *sd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6693) struct perf_domain *pd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6694) int new_cpu = INT_MAX;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6695)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6696) sync_entity_load_avg(&p->se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6697) trace_android_rvh_find_energy_efficient_cpu(p, prev_cpu, sync, &new_cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6698) if (new_cpu != INT_MAX)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6699) return new_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6700)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6701) rcu_read_lock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6702) pd = rcu_dereference(rd->pd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6703) if (!pd || READ_ONCE(rd->overutilized))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6704) goto fail;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6705)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6706) cpu = smp_processor_id();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6707) if (sync && cpu_rq(cpu)->nr_running == 1 &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6708) cpumask_test_cpu(cpu, p->cpus_ptr) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6709) task_fits_capacity(p, capacity_of(cpu))) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6710) rcu_read_unlock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6711) return cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6712) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6713)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6714) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6715) * Energy-aware wake-up happens on the lowest sched_domain starting
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6716) * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6717) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6718) sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6719) while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6720) sd = sd->parent;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6721) if (!sd)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6722) goto fail;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6723)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6724) if (!task_util_est(p))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6725) goto unlock;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6726)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6727) latency_sensitive = uclamp_latency_sensitive(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6728) boosted = uclamp_boosted(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6729) target_cap = boosted ? 0 : ULONG_MAX;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6730)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6731) for (; pd; pd = pd->next) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6732) unsigned long cur_delta, spare_cap, max_spare_cap = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6733) unsigned long base_energy_pd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6734) int max_spare_cap_cpu = -1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6735)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6736) /* Compute the 'base' energy of the pd, without @p */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6737) base_energy_pd = compute_energy(p, -1, pd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6738) base_energy += base_energy_pd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6739)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6740) for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6741) if (!cpumask_test_cpu(cpu, p->cpus_ptr))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6742) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6743)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6744) util = cpu_util_next(cpu, p, cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6745) cpu_cap = capacity_of(cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6746) spare_cap = cpu_cap;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6747) lsub_positive(&spare_cap, util);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6748)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6749) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6750) * Skip CPUs that cannot satisfy the capacity request.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6751) * IOW, placing the task there would make the CPU
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6752) * overutilized. Take uclamp into account to see how
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6753) * much capacity we can get out of the CPU; this is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6754) * aligned with schedutil_cpu_util().
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6755) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6756) util = uclamp_rq_util_with(cpu_rq(cpu), util, p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6757) if (!fits_capacity(util, cpu_cap))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6758) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6759)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6760) /* Always use prev_cpu as a candidate. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6761) if (!latency_sensitive && cpu == prev_cpu) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6762) prev_delta = compute_energy(p, prev_cpu, pd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6763) prev_delta -= base_energy_pd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6764) best_delta = min(best_delta, prev_delta);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6765) if (IS_ENABLED(CONFIG_ROCKCHIP_PERFORMANCE)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6766) if (prev_delta == best_delta)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6767) best_energy_cpu = prev_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6768) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6769) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6770)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6771) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6772) * Find the CPU with the maximum spare capacity in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6773) * the performance domain
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6774) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6775) if (spare_cap > max_spare_cap) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6776) max_spare_cap = spare_cap;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6777) max_spare_cap_cpu = cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6778) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6779)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6780) if (!latency_sensitive)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6781) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6782)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6783) if (idle_cpu(cpu)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6784) cpu_cap = capacity_orig_of(cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6785) if (boosted && cpu_cap < target_cap)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6786) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6787) if (!boosted && cpu_cap > target_cap)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6788) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6789) idle = idle_get_state(cpu_rq(cpu));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6790) if (idle && idle->exit_latency > min_exit_lat &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6791) cpu_cap == target_cap)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6792) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6793)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6794) if (idle)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6795) min_exit_lat = idle->exit_latency;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6796) target_cap = cpu_cap;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6797) best_idle_cpu = cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6798) } else if (spare_cap > max_spare_cap_ls) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6799) max_spare_cap_ls = spare_cap;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6800) max_spare_cap_cpu_ls = cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6801) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6802) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6803)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6804) /* Evaluate the energy impact of using this CPU. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6805) if (!latency_sensitive && max_spare_cap_cpu >= 0 &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6806) max_spare_cap_cpu != prev_cpu) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6807) cur_delta = compute_energy(p, max_spare_cap_cpu, pd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6808) cur_delta -= base_energy_pd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6809) if (cur_delta < best_delta) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6810) best_delta = cur_delta;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6811) best_energy_cpu = max_spare_cap_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6812) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6813) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6814) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6815) unlock:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6816) rcu_read_unlock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6817)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6818) if (latency_sensitive)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6819) return best_idle_cpu >= 0 ? best_idle_cpu : max_spare_cap_cpu_ls;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6820)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6821) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6822) * Pick the best CPU if prev_cpu cannot be used, or if it saves at
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6823) * least 6% of the energy used by prev_cpu.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6824) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6825) if (prev_delta == ULONG_MAX)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6826) return best_energy_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6827)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6828) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6829) * when select ROCKCHIP_PERFORMANCE_LOW:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6830) * Pick best_energy_cpu immediately if prev_cpu is big cpu and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6831) * best_energy_cpu is little cpu, so that tasks can migrate from
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6832) * big cpu to little cpu easier to save power.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6833) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6834) if (IS_ENABLED(CONFIG_ROCKCHIP_PERFORMANCE)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6835) struct cpumask *cpul_mask = rockchip_perf_get_cpul_mask();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6836) struct cpumask *cpub_mask = rockchip_perf_get_cpub_mask();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6837) int level = rockchip_perf_get_level();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6838)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6839) if ((level == ROCKCHIP_PERFORMANCE_LOW) && cpul_mask &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6840) cpub_mask && cpumask_test_cpu(prev_cpu, cpub_mask) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6841) cpumask_test_cpu(best_energy_cpu, cpul_mask)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6842) return best_energy_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6843) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6844) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6845)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6846) if ((prev_delta - best_delta) > ((prev_delta + base_energy) >> 4))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6847) return best_energy_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6848)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6849) return prev_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6850)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6851) fail:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6852) rcu_read_unlock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6853)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6854) return -1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6855) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6856)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6857) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6858) * select_task_rq_fair: Select target runqueue for the waking task in domains
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6859) * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6860) * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6861) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6862) * Balances load by selecting the idlest CPU in the idlest group, or under
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6863) * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6864) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6865) * Returns the target CPU number.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6866) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6867) * preempt must be disabled.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6868) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6869) static int
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6870) select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6871) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6872) struct sched_domain *tmp, *sd = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6873) int cpu = smp_processor_id();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6874) int new_cpu = prev_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6875) int want_affine = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6876) int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6877) int target_cpu = -1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6878)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6879) if (trace_android_rvh_select_task_rq_fair_enabled() &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6880) !(sd_flag & SD_BALANCE_FORK))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6881) sync_entity_load_avg(&p->se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6882) trace_android_rvh_select_task_rq_fair(p, prev_cpu, sd_flag,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6883) wake_flags, &target_cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6884) if (target_cpu >= 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6885) return target_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6886)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6887) if (sd_flag & SD_BALANCE_WAKE) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6888) record_wakee(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6889)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6890) if (IS_ENABLED(CONFIG_ROCKCHIP_PERFORMANCE)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6891) if (rockchip_perf_get_level() == ROCKCHIP_PERFORMANCE_HIGH)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6892) goto no_eas;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6893) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6894)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6895) if (sched_energy_enabled()) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6896) new_cpu = find_energy_efficient_cpu(p, prev_cpu, sync);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6897) if (new_cpu >= 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6898) return new_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6899) new_cpu = prev_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6900) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6901)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6902) no_eas:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6903) want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6904) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6905)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6906) rcu_read_lock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6907) for_each_domain(cpu, tmp) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6908) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6909) * If both 'cpu' and 'prev_cpu' are part of this domain,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6910) * cpu is a valid SD_WAKE_AFFINE target.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6911) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6912) if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6913) cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6914) if (cpu != prev_cpu)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6915) new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6916)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6917) sd = NULL; /* Prefer wake_affine over balance flags */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6918) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6919) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6920)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6921) if (tmp->flags & sd_flag)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6922) sd = tmp;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6923) else if (!want_affine)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6924) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6925) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6926)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6927) if (unlikely(sd)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6928) /* Slow path */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6929) new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6930) } else if (sd_flag & SD_BALANCE_WAKE) { /* XXX always ? */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6931) /* Fast path */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6932)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6933) new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6934)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6935) if (IS_ENABLED(CONFIG_ROCKCHIP_PERFORMANCE)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6936) struct root_domain *rd = cpu_rq(cpu)->rd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6937) struct cpumask *cpul_mask = rockchip_perf_get_cpul_mask();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6938) struct cpumask *cpub_mask = rockchip_perf_get_cpub_mask();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6939) int level = rockchip_perf_get_level();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6940)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6941) if ((level == ROCKCHIP_PERFORMANCE_HIGH) && !READ_ONCE(rd->overutilized) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6942) cpul_mask && cpub_mask && cpumask_intersects(p->cpus_ptr, cpub_mask) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6943) cpumask_test_cpu(new_cpu, cpul_mask)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6944) for_each_domain(cpu, tmp) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6945) sd = tmp;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6946) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6947) if (sd)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6948) new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6949) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6950) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6951)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6952) if (want_affine)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6953) current->recent_used_cpu = cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6954) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6955) rcu_read_unlock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6956)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6957) return new_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6958) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6959)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6960) static void detach_entity_cfs_rq(struct sched_entity *se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6961)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6962) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6963) * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6964) * cfs_rq_of(p) references at time of call are still valid and identify the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6965) * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6966) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6967) static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6968) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6969) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6970) * As blocked tasks retain absolute vruntime the migration needs to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6971) * deal with this by subtracting the old and adding the new
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6972) * min_vruntime -- the latter is done by enqueue_entity() when placing
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6973) * the task on the new runqueue.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6974) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6975) if (p->state == TASK_WAKING) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6976) struct sched_entity *se = &p->se;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6977) struct cfs_rq *cfs_rq = cfs_rq_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6978) u64 min_vruntime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6979)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6980) #ifndef CONFIG_64BIT
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6981) u64 min_vruntime_copy;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6982)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6983) do {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6984) min_vruntime_copy = cfs_rq->min_vruntime_copy;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6985) smp_rmb();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6986) min_vruntime = cfs_rq->min_vruntime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6987) } while (min_vruntime != min_vruntime_copy);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6988) #else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6989) min_vruntime = cfs_rq->min_vruntime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6990) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6991)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6992) se->vruntime -= min_vruntime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6993) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6994)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6995) if (p->on_rq == TASK_ON_RQ_MIGRATING) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6996) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6997) * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6998) * rq->lock and can modify state directly.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6999) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7000) lockdep_assert_held(&task_rq(p)->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7001) detach_entity_cfs_rq(&p->se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7002)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7003) } else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7004) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7005) * We are supposed to update the task to "current" time, then
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7006) * its up to date and ready to go to new CPU/cfs_rq. But we
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7007) * have difficulty in getting what current time is, so simply
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7008) * throw away the out-of-date time. This will result in the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7009) * wakee task is less decayed, but giving the wakee more load
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7010) * sounds not bad.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7011) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7012) remove_entity_load_avg(&p->se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7013) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7014)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7015) /* Tell new CPU we are migrated */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7016) p->se.avg.last_update_time = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7017)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7018) /* We have migrated, no longer consider this task hot */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7019) p->se.exec_start = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7020)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7021) update_scan_period(p, new_cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7022) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7023)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7024) static void task_dead_fair(struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7025) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7026) remove_entity_load_avg(&p->se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7027) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7028)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7029) static int
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7030) balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7031) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7032) if (rq->nr_running)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7033) return 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7034)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7035) return newidle_balance(rq, rf) != 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7036) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7037) #endif /* CONFIG_SMP */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7038)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7039) static unsigned long wakeup_gran(struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7040) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7041) unsigned long gran = sysctl_sched_wakeup_granularity;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7042)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7043) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7044) * Since its curr running now, convert the gran from real-time
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7045) * to virtual-time in his units.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7046) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7047) * By using 'se' instead of 'curr' we penalize light tasks, so
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7048) * they get preempted easier. That is, if 'se' < 'curr' then
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7049) * the resulting gran will be larger, therefore penalizing the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7050) * lighter, if otoh 'se' > 'curr' then the resulting gran will
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7051) * be smaller, again penalizing the lighter task.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7052) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7053) * This is especially important for buddies when the leftmost
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7054) * task is higher priority than the buddy.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7055) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7056) return calc_delta_fair(gran, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7057) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7058)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7059) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7060) * Should 'se' preempt 'curr'.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7061) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7062) * |s1
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7063) * |s2
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7064) * |s3
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7065) * g
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7066) * |<--->|c
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7067) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7068) * w(c, s1) = -1
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7069) * w(c, s2) = 0
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7070) * w(c, s3) = 1
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7071) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7072) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7073) static int
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7074) wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7075) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7076) s64 gran, vdiff = curr->vruntime - se->vruntime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7077)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7078) if (vdiff <= 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7079) return -1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7080)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7081) gran = wakeup_gran(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7082) if (vdiff > gran)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7083) return 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7084)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7085) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7086) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7087)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7088) static void set_last_buddy(struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7089) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7090) if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7091) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7092)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7093) for_each_sched_entity(se) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7094) if (SCHED_WARN_ON(!se->on_rq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7095) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7096) cfs_rq_of(se)->last = se;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7097) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7098) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7099)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7100) static void set_next_buddy(struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7101) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7102) if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7103) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7104)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7105) for_each_sched_entity(se) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7106) if (SCHED_WARN_ON(!se->on_rq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7107) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7108) cfs_rq_of(se)->next = se;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7109) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7110) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7111)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7112) static void set_skip_buddy(struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7113) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7114) for_each_sched_entity(se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7115) cfs_rq_of(se)->skip = se;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7116) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7117)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7118) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7119) * Preempt the current task with a newly woken task if needed:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7120) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7121) static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7122) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7123) struct task_struct *curr = rq->curr;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7124) struct sched_entity *se = &curr->se, *pse = &p->se;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7125) struct cfs_rq *cfs_rq = task_cfs_rq(curr);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7126) int scale = cfs_rq->nr_running >= sched_nr_latency;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7127) int next_buddy_marked = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7128) bool preempt = false, nopreempt = false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7129)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7130) if (unlikely(se == pse))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7131) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7132)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7133) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7134) * This is possible from callers such as attach_tasks(), in which we
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7135) * unconditionally check_prempt_curr() after an enqueue (which may have
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7136) * lead to a throttle). This both saves work and prevents false
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7137) * next-buddy nomination below.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7138) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7139) if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7140) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7141)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7142) if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7143) set_next_buddy(pse);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7144) next_buddy_marked = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7145) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7146)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7147) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7148) * We can come here with TIF_NEED_RESCHED already set from new task
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7149) * wake up path.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7150) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7151) * Note: this also catches the edge-case of curr being in a throttled
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7152) * group (e.g. via set_curr_task), since update_curr() (in the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7153) * enqueue of curr) will have resulted in resched being set. This
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7154) * prevents us from potentially nominating it as a false LAST_BUDDY
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7155) * below.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7156) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7157) if (test_tsk_need_resched(curr))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7158) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7159)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7160) /* Idle tasks are by definition preempted by non-idle tasks. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7161) if (unlikely(task_has_idle_policy(curr)) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7162) likely(!task_has_idle_policy(p)))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7163) goto preempt;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7164)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7165) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7166) * Batch and idle tasks do not preempt non-idle tasks (their preemption
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7167) * is driven by the tick):
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7168) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7169) if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7170) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7171)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7172) find_matching_se(&se, &pse);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7173) update_curr(cfs_rq_of(se));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7174) trace_android_rvh_check_preempt_wakeup(rq, p, &preempt, &nopreempt,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7175) wake_flags, se, pse, next_buddy_marked, sysctl_sched_wakeup_granularity);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7176) if (preempt)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7177) goto preempt;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7178) if (nopreempt)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7179) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7180) BUG_ON(!pse);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7181) if (wakeup_preempt_entity(se, pse) == 1) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7182) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7183) * Bias pick_next to pick the sched entity that is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7184) * triggering this preemption.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7185) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7186) if (!next_buddy_marked)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7187) set_next_buddy(pse);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7188) goto preempt;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7189) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7190)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7191) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7192)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7193) preempt:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7194) resched_curr(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7195) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7196) * Only set the backward buddy when the current task is still
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7197) * on the rq. This can happen when a wakeup gets interleaved
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7198) * with schedule on the ->pre_schedule() or idle_balance()
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7199) * point, either of which can * drop the rq lock.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7200) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7201) * Also, during early boot the idle thread is in the fair class,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7202) * for obvious reasons its a bad idea to schedule back to it.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7203) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7204) if (unlikely(!se->on_rq || curr == rq->idle))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7205) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7206)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7207) if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7208) set_last_buddy(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7209) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7210)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7211) struct task_struct *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7212) pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7213) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7214) struct cfs_rq *cfs_rq = &rq->cfs;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7215) struct sched_entity *se = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7216) struct task_struct *p = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7217) int new_tasks;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7218) bool repick = false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7219)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7220) again:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7221) if (!sched_fair_runnable(rq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7222) goto idle;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7223)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7224) #ifdef CONFIG_FAIR_GROUP_SCHED
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7225) if (!prev || prev->sched_class != &fair_sched_class)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7226) goto simple;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7227)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7228) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7229) * Because of the set_next_buddy() in dequeue_task_fair() it is rather
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7230) * likely that a next task is from the same cgroup as the current.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7231) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7232) * Therefore attempt to avoid putting and setting the entire cgroup
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7233) * hierarchy, only change the part that actually changes.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7234) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7235)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7236) do {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7237) struct sched_entity *curr = cfs_rq->curr;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7238)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7239) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7240) * Since we got here without doing put_prev_entity() we also
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7241) * have to consider cfs_rq->curr. If it is still a runnable
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7242) * entity, update_curr() will update its vruntime, otherwise
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7243) * forget we've ever seen it.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7244) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7245) if (curr) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7246) if (curr->on_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7247) update_curr(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7248) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7249) curr = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7250)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7251) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7252) * This call to check_cfs_rq_runtime() will do the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7253) * throttle and dequeue its entity in the parent(s).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7254) * Therefore the nr_running test will indeed
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7255) * be correct.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7256) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7257) if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7258) cfs_rq = &rq->cfs;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7259)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7260) if (!cfs_rq->nr_running)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7261) goto idle;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7262)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7263) goto simple;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7264) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7265) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7266)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7267) se = pick_next_entity(cfs_rq, curr);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7268) cfs_rq = group_cfs_rq(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7269) } while (cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7270)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7271) p = task_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7272) trace_android_rvh_replace_next_task_fair(rq, &p, &se, &repick, false, prev);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7273) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7274) * Since we haven't yet done put_prev_entity and if the selected task
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7275) * is a different task than we started out with, try and touch the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7276) * least amount of cfs_rqs.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7277) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7278) if (prev != p) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7279) struct sched_entity *pse = &prev->se;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7280)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7281) while (!(cfs_rq = is_same_group(se, pse))) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7282) int se_depth = se->depth;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7283) int pse_depth = pse->depth;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7284)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7285) if (se_depth <= pse_depth) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7286) put_prev_entity(cfs_rq_of(pse), pse);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7287) pse = parent_entity(pse);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7288) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7289) if (se_depth >= pse_depth) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7290) set_next_entity(cfs_rq_of(se), se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7291) se = parent_entity(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7292) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7293) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7294)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7295) put_prev_entity(cfs_rq, pse);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7296) set_next_entity(cfs_rq, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7297) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7298)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7299) goto done;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7300) simple:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7301) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7302) if (prev)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7303) put_prev_task(rq, prev);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7304)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7305) trace_android_rvh_replace_next_task_fair(rq, &p, &se, &repick, true, prev);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7306) if (repick) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7307) for_each_sched_entity(se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7308) set_next_entity(cfs_rq_of(se), se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7309) goto done;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7310) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7311)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7312) do {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7313) se = pick_next_entity(cfs_rq, NULL);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7314) set_next_entity(cfs_rq, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7315) cfs_rq = group_cfs_rq(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7316) } while (cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7317)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7318) p = task_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7319)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7320) done: __maybe_unused;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7321) #ifdef CONFIG_SMP
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7322) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7323) * Move the next running task to the front of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7324) * the list, so our cfs_tasks list becomes MRU
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7325) * one.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7326) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7327) list_move(&p->se.group_node, &rq->cfs_tasks);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7328) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7329)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7330) if (hrtick_enabled(rq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7331) hrtick_start_fair(rq, p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7332)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7333) update_misfit_status(p, rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7334)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7335) return p;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7336)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7337) idle:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7338) if (!rf)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7339) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7340)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7341) new_tasks = newidle_balance(rq, rf);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7342)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7343) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7344) * Because newidle_balance() releases (and re-acquires) rq->lock, it is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7345) * possible for any higher priority task to appear. In that case we
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7346) * must re-start the pick_next_entity() loop.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7347) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7348) if (new_tasks < 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7349) return RETRY_TASK;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7350)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7351) if (new_tasks > 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7352) goto again;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7353)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7354) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7355) * rq is about to be idle, check if we need to update the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7356) * lost_idle_time of clock_pelt
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7357) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7358) update_idle_rq_clock_pelt(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7359)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7360) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7361) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7362)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7363) static struct task_struct *__pick_next_task_fair(struct rq *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7364) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7365) return pick_next_task_fair(rq, NULL, NULL);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7366) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7367)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7368) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7369) * Account for a descheduled task:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7370) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7371) static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7372) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7373) struct sched_entity *se = &prev->se;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7374) struct cfs_rq *cfs_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7375)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7376) for_each_sched_entity(se) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7377) cfs_rq = cfs_rq_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7378) put_prev_entity(cfs_rq, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7379) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7380) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7381)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7382) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7383) * sched_yield() is very simple
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7384) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7385) * The magic of dealing with the ->skip buddy is in pick_next_entity.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7386) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7387) static void yield_task_fair(struct rq *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7388) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7389) struct task_struct *curr = rq->curr;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7390) struct cfs_rq *cfs_rq = task_cfs_rq(curr);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7391) struct sched_entity *se = &curr->se;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7392)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7393) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7394) * Are we the only task in the tree?
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7395) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7396) if (unlikely(rq->nr_running == 1))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7397) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7398)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7399) clear_buddies(cfs_rq, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7400)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7401) if (curr->policy != SCHED_BATCH) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7402) update_rq_clock(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7403) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7404) * Update run-time statistics of the 'current'.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7405) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7406) update_curr(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7407) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7408) * Tell update_rq_clock() that we've just updated,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7409) * so we don't do microscopic update in schedule()
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7410) * and double the fastpath cost.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7411) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7412) rq_clock_skip_update(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7413) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7414)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7415) set_skip_buddy(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7416) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7417)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7418) static bool yield_to_task_fair(struct rq *rq, struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7419) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7420) struct sched_entity *se = &p->se;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7421)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7422) /* throttled hierarchies are not runnable */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7423) if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7424) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7425)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7426) /* Tell the scheduler that we'd really like pse to run next. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7427) set_next_buddy(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7428)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7429) yield_task_fair(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7430)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7431) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7432) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7433)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7434) #ifdef CONFIG_SMP
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7435) /**************************************************
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7436) * Fair scheduling class load-balancing methods.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7437) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7438) * BASICS
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7439) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7440) * The purpose of load-balancing is to achieve the same basic fairness the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7441) * per-CPU scheduler provides, namely provide a proportional amount of compute
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7442) * time to each task. This is expressed in the following equation:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7443) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7444) * W_i,n/P_i == W_j,n/P_j for all i,j (1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7445) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7446) * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7447) * W_i,0 is defined as:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7448) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7449) * W_i,0 = \Sum_j w_i,j (2)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7450) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7451) * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7452) * is derived from the nice value as per sched_prio_to_weight[].
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7453) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7454) * The weight average is an exponential decay average of the instantaneous
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7455) * weight:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7456) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7457) * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7458) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7459) * C_i is the compute capacity of CPU i, typically it is the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7460) * fraction of 'recent' time available for SCHED_OTHER task execution. But it
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7461) * can also include other factors [XXX].
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7462) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7463) * To achieve this balance we define a measure of imbalance which follows
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7464) * directly from (1):
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7465) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7466) * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7467) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7468) * We them move tasks around to minimize the imbalance. In the continuous
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7469) * function space it is obvious this converges, in the discrete case we get
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7470) * a few fun cases generally called infeasible weight scenarios.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7471) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7472) * [XXX expand on:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7473) * - infeasible weights;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7474) * - local vs global optima in the discrete case. ]
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7475) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7476) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7477) * SCHED DOMAINS
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7478) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7479) * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7480) * for all i,j solution, we create a tree of CPUs that follows the hardware
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7481) * topology where each level pairs two lower groups (or better). This results
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7482) * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7483) * tree to only the first of the previous level and we decrease the frequency
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7484) * of load-balance at each level inv. proportional to the number of CPUs in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7485) * the groups.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7486) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7487) * This yields:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7488) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7489) * log_2 n 1 n
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7490) * \Sum { --- * --- * 2^i } = O(n) (5)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7491) * i = 0 2^i 2^i
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7492) * `- size of each group
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7493) * | | `- number of CPUs doing load-balance
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7494) * | `- freq
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7495) * `- sum over all levels
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7496) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7497) * Coupled with a limit on how many tasks we can migrate every balance pass,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7498) * this makes (5) the runtime complexity of the balancer.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7499) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7500) * An important property here is that each CPU is still (indirectly) connected
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7501) * to every other CPU in at most O(log n) steps:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7502) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7503) * The adjacency matrix of the resulting graph is given by:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7504) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7505) * log_2 n
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7506) * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7507) * k = 0
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7508) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7509) * And you'll find that:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7510) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7511) * A^(log_2 n)_i,j != 0 for all i,j (7)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7512) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7513) * Showing there's indeed a path between every CPU in at most O(log n) steps.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7514) * The task movement gives a factor of O(m), giving a convergence complexity
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7515) * of:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7516) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7517) * O(nm log n), n := nr_cpus, m := nr_tasks (8)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7518) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7519) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7520) * WORK CONSERVING
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7521) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7522) * In order to avoid CPUs going idle while there's still work to do, new idle
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7523) * balancing is more aggressive and has the newly idle CPU iterate up the domain
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7524) * tree itself instead of relying on other CPUs to bring it work.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7525) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7526) * This adds some complexity to both (5) and (8) but it reduces the total idle
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7527) * time.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7528) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7529) * [XXX more?]
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7530) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7531) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7532) * CGROUPS
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7533) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7534) * Cgroups make a horror show out of (2), instead of a simple sum we get:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7535) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7536) * s_k,i
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7537) * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7538) * S_k
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7539) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7540) * Where
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7541) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7542) * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7543) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7544) * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7545) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7546) * The big problem is S_k, its a global sum needed to compute a local (W_i)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7547) * property.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7548) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7549) * [XXX write more on how we solve this.. _after_ merging pjt's patches that
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7550) * rewrite all of this once again.]
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7551) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7552)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7553) unsigned long __read_mostly max_load_balance_interval = HZ/10;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7554) EXPORT_SYMBOL_GPL(max_load_balance_interval);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7555)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7556) enum fbq_type { regular, remote, all };
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7557)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7558) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7559) * 'group_type' describes the group of CPUs at the moment of load balancing.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7560) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7561) * The enum is ordered by pulling priority, with the group with lowest priority
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7562) * first so the group_type can simply be compared when selecting the busiest
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7563) * group. See update_sd_pick_busiest().
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7564) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7565) enum group_type {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7566) /* The group has spare capacity that can be used to run more tasks. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7567) group_has_spare = 0,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7568) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7569) * The group is fully used and the tasks don't compete for more CPU
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7570) * cycles. Nevertheless, some tasks might wait before running.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7571) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7572) group_fully_busy,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7573) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7574) * SD_ASYM_CPUCAPACITY only: One task doesn't fit with CPU's capacity
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7575) * and must be migrated to a more powerful CPU.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7576) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7577) group_misfit_task,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7578) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7579) * SD_ASYM_PACKING only: One local CPU with higher capacity is available,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7580) * and the task should be migrated to it instead of running on the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7581) * current CPU.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7582) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7583) group_asym_packing,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7584) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7585) * The tasks' affinity constraints previously prevented the scheduler
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7586) * from balancing the load across the system.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7587) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7588) group_imbalanced,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7589) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7590) * The CPU is overloaded and can't provide expected CPU cycles to all
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7591) * tasks.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7592) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7593) group_overloaded
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7594) };
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7595)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7596) enum migration_type {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7597) migrate_load = 0,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7598) migrate_util,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7599) migrate_task,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7600) migrate_misfit
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7601) };
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7602)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7603) #define LBF_ALL_PINNED 0x01
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7604) #define LBF_NEED_BREAK 0x02
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7605) #define LBF_DST_PINNED 0x04
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7606) #define LBF_SOME_PINNED 0x08
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7607) #define LBF_NOHZ_STATS 0x10
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7608) #define LBF_NOHZ_AGAIN 0x20
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7609)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7610) struct lb_env {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7611) struct sched_domain *sd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7612)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7613) struct rq *src_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7614) int src_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7615)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7616) int dst_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7617) struct rq *dst_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7618)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7619) struct cpumask *dst_grpmask;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7620) int new_dst_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7621) enum cpu_idle_type idle;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7622) long imbalance;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7623) /* The set of CPUs under consideration for load-balancing */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7624) struct cpumask *cpus;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7625)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7626) unsigned int flags;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7627)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7628) unsigned int loop;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7629) unsigned int loop_break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7630) unsigned int loop_max;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7631)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7632) enum fbq_type fbq_type;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7633) enum migration_type migration_type;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7634) struct list_head tasks;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7635) struct rq_flags *src_rq_rf;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7636) };
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7637)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7638) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7639) * Is this task likely cache-hot:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7640) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7641) static int task_hot(struct task_struct *p, struct lb_env *env)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7642) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7643) s64 delta;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7644)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7645) lockdep_assert_held(&env->src_rq->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7646)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7647) if (p->sched_class != &fair_sched_class)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7648) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7649)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7650) if (unlikely(task_has_idle_policy(p)))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7651) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7652)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7653) /* SMT siblings share cache */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7654) if (env->sd->flags & SD_SHARE_CPUCAPACITY)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7655) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7656)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7657) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7658) * Buddy candidates are cache hot:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7659) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7660) if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7661) (&p->se == cfs_rq_of(&p->se)->next ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7662) &p->se == cfs_rq_of(&p->se)->last))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7663) return 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7664)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7665) if (sysctl_sched_migration_cost == -1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7666) return 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7667) if (sysctl_sched_migration_cost == 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7668) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7669)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7670) delta = rq_clock_task(env->src_rq) - p->se.exec_start;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7671)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7672) return delta < (s64)sysctl_sched_migration_cost;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7673) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7674)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7675) #ifdef CONFIG_NUMA_BALANCING
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7676) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7677) * Returns 1, if task migration degrades locality
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7678) * Returns 0, if task migration improves locality i.e migration preferred.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7679) * Returns -1, if task migration is not affected by locality.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7680) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7681) static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7682) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7683) struct numa_group *numa_group = rcu_dereference(p->numa_group);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7684) unsigned long src_weight, dst_weight;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7685) int src_nid, dst_nid, dist;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7686)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7687) if (!static_branch_likely(&sched_numa_balancing))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7688) return -1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7689)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7690) if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7691) return -1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7692)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7693) src_nid = cpu_to_node(env->src_cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7694) dst_nid = cpu_to_node(env->dst_cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7695)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7696) if (src_nid == dst_nid)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7697) return -1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7698)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7699) /* Migrating away from the preferred node is always bad. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7700) if (src_nid == p->numa_preferred_nid) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7701) if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7702) return 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7703) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7704) return -1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7705) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7706)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7707) /* Encourage migration to the preferred node. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7708) if (dst_nid == p->numa_preferred_nid)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7709) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7710)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7711) /* Leaving a core idle is often worse than degrading locality. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7712) if (env->idle == CPU_IDLE)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7713) return -1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7714)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7715) dist = node_distance(src_nid, dst_nid);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7716) if (numa_group) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7717) src_weight = group_weight(p, src_nid, dist);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7718) dst_weight = group_weight(p, dst_nid, dist);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7719) } else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7720) src_weight = task_weight(p, src_nid, dist);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7721) dst_weight = task_weight(p, dst_nid, dist);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7722) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7723)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7724) return dst_weight < src_weight;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7725) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7726)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7727) #else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7728) static inline int migrate_degrades_locality(struct task_struct *p,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7729) struct lb_env *env)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7730) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7731) return -1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7732) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7733) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7734)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7735) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7736) * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7737) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7738) static
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7739) int can_migrate_task(struct task_struct *p, struct lb_env *env)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7740) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7741) int tsk_cache_hot;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7742) int can_migrate = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7743)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7744) lockdep_assert_held(&env->src_rq->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7745)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7746) trace_android_rvh_can_migrate_task(p, env->dst_cpu, &can_migrate);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7747) if (!can_migrate)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7748) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7749)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7750) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7751) * We do not migrate tasks that are:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7752) * 1) throttled_lb_pair, or
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7753) * 2) cannot be migrated to this CPU due to cpus_ptr, or
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7754) * 3) running (obviously), or
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7755) * 4) are cache-hot on their current CPU.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7756) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7757) if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7758) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7759)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7760) /* Disregard pcpu kthreads; they are where they need to be. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7761) if (kthread_is_per_cpu(p))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7762) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7763)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7764) if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7765) int cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7766)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7767) schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7768)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7769) env->flags |= LBF_SOME_PINNED;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7770)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7771) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7772) * Remember if this task can be migrated to any other CPU in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7773) * our sched_group. We may want to revisit it if we couldn't
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7774) * meet load balance goals by pulling other tasks on src_cpu.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7775) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7776) * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7777) * already computed one in current iteration.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7778) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7779) if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7780) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7781)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7782) /* Prevent to re-select dst_cpu via env's CPUs: */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7783) for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7784) if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7785) env->flags |= LBF_DST_PINNED;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7786) env->new_dst_cpu = cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7787) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7788) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7789) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7790)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7791) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7792) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7793)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7794) /* Record that we found atleast one task that could run on dst_cpu */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7795) env->flags &= ~LBF_ALL_PINNED;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7796)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7797) if (task_running(env->src_rq, p)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7798) schedstat_inc(p->se.statistics.nr_failed_migrations_running);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7799) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7800) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7801)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7802) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7803) * Aggressive migration if:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7804) * 1) destination numa is preferred
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7805) * 2) task is cache cold, or
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7806) * 3) too many balance attempts have failed.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7807) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7808) tsk_cache_hot = migrate_degrades_locality(p, env);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7809) if (tsk_cache_hot == -1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7810) tsk_cache_hot = task_hot(p, env);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7811)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7812) if (tsk_cache_hot <= 0 ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7813) env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7814) if (tsk_cache_hot == 1) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7815) schedstat_inc(env->sd->lb_hot_gained[env->idle]);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7816) schedstat_inc(p->se.statistics.nr_forced_migrations);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7817) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7818) return 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7819) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7820)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7821) schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7822) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7823) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7824)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7825) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7826) * detach_task() -- detach the task for the migration specified in env
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7827) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7828) static void detach_task(struct task_struct *p, struct lb_env *env)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7829) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7830) int detached = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7831)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7832) lockdep_assert_held(&env->src_rq->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7833)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7834) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7835) * The vendor hook may drop the lock temporarily, so
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7836) * pass the rq flags to unpin lock. We expect the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7837) * rq lock to be held after return.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7838) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7839) trace_android_rvh_migrate_queued_task(env->src_rq, env->src_rq_rf, p,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7840) env->dst_cpu, &detached);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7841) if (detached)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7842) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7843)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7844) deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7845) set_task_cpu(p, env->dst_cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7846) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7847)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7848) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7849) * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7850) * part of active balancing operations within "domain".
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7851) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7852) * Returns a task if successful and NULL otherwise.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7853) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7854) static struct task_struct *detach_one_task(struct lb_env *env)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7855) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7856) struct task_struct *p;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7857)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7858) lockdep_assert_held(&env->src_rq->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7859)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7860) list_for_each_entry_reverse(p,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7861) &env->src_rq->cfs_tasks, se.group_node) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7862) if (!can_migrate_task(p, env))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7863) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7864)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7865) detach_task(p, env);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7866)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7867) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7868) * Right now, this is only the second place where
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7869) * lb_gained[env->idle] is updated (other is detach_tasks)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7870) * so we can safely collect stats here rather than
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7871) * inside detach_tasks().
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7872) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7873) schedstat_inc(env->sd->lb_gained[env->idle]);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7874) return p;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7875) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7876) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7877) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7878)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7879) static const unsigned int sched_nr_migrate_break = 32;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7880)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7881) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7882) * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7883) * busiest_rq, as part of a balancing operation within domain "sd".
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7884) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7885) * Returns number of detached tasks if successful and 0 otherwise.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7886) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7887) static int detach_tasks(struct lb_env *env)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7888) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7889) struct list_head *tasks = &env->src_rq->cfs_tasks;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7890) unsigned long util, load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7891) struct task_struct *p;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7892) int detached = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7893)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7894) lockdep_assert_held(&env->src_rq->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7895)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7896) if (env->imbalance <= 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7897) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7898)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7899) while (!list_empty(tasks)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7900) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7901) * We don't want to steal all, otherwise we may be treated likewise,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7902) * which could at worst lead to a livelock crash.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7903) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7904) if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7905) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7906)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7907) p = list_last_entry(tasks, struct task_struct, se.group_node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7908)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7909) env->loop++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7910) /* We've more or less seen every task there is, call it quits */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7911) if (env->loop > env->loop_max)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7912) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7913)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7914) /* take a breather every nr_migrate tasks */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7915) if (env->loop > env->loop_break) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7916) env->loop_break += sched_nr_migrate_break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7917) env->flags |= LBF_NEED_BREAK;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7918) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7919) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7920)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7921) if (!can_migrate_task(p, env))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7922) goto next;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7923)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7924) switch (env->migration_type) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7925) case migrate_load:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7926) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7927) * Depending of the number of CPUs and tasks and the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7928) * cgroup hierarchy, task_h_load() can return a null
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7929) * value. Make sure that env->imbalance decreases
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7930) * otherwise detach_tasks() will stop only after
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7931) * detaching up to loop_max tasks.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7932) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7933) load = max_t(unsigned long, task_h_load(p), 1);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7934)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7935) if (sched_feat(LB_MIN) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7936) load < 16 && !env->sd->nr_balance_failed)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7937) goto next;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7938)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7939) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7940) * Make sure that we don't migrate too much load.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7941) * Nevertheless, let relax the constraint if
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7942) * scheduler fails to find a good waiting task to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7943) * migrate.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7944) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7945) if (shr_bound(load, env->sd->nr_balance_failed) > env->imbalance)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7946) goto next;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7947)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7948) env->imbalance -= load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7949) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7950)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7951) case migrate_util:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7952) util = task_util_est(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7953)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7954) if (util > env->imbalance)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7955) goto next;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7956)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7957) env->imbalance -= util;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7958) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7959)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7960) case migrate_task:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7961) env->imbalance--;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7962) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7963)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7964) case migrate_misfit:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7965) /* This is not a misfit task */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7966) if (task_fits_capacity(p, capacity_of(env->src_cpu)))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7967) goto next;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7968)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7969) env->imbalance = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7970) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7971) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7972)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7973) detach_task(p, env);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7974) list_add(&p->se.group_node, &env->tasks);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7975)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7976) detached++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7977)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7978) #ifdef CONFIG_PREEMPTION
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7979) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7980) * NEWIDLE balancing is a source of latency, so preemptible
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7981) * kernels will stop after the first task is detached to minimize
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7982) * the critical section.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7983) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7984) if (env->idle == CPU_NEWLY_IDLE)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7985) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7986) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7987)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7988) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7989) * We only want to steal up to the prescribed amount of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7990) * load/util/tasks.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7991) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7992) if (env->imbalance <= 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7993) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7994)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7995) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7996) next:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7997) list_move(&p->se.group_node, tasks);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7998) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7999)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8000) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8001) * Right now, this is one of only two places we collect this stat
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8002) * so we can safely collect detach_one_task() stats here rather
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8003) * than inside detach_one_task().
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8004) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8005) schedstat_add(env->sd->lb_gained[env->idle], detached);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8006)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8007) return detached;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8008) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8009)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8010) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8011) * attach_task() -- attach the task detached by detach_task() to its new rq.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8012) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8013) static void attach_task(struct rq *rq, struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8014) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8015) lockdep_assert_held(&rq->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8016)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8017) BUG_ON(task_rq(p) != rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8018) activate_task(rq, p, ENQUEUE_NOCLOCK);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8019) check_preempt_curr(rq, p, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8020) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8021)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8022) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8023) * attach_one_task() -- attaches the task returned from detach_one_task() to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8024) * its new rq.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8025) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8026) static void attach_one_task(struct rq *rq, struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8027) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8028) struct rq_flags rf;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8029)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8030) rq_lock(rq, &rf);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8031) update_rq_clock(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8032) attach_task(rq, p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8033) rq_unlock(rq, &rf);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8034) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8035)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8036) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8037) * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8038) * new rq.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8039) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8040) static void attach_tasks(struct lb_env *env)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8041) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8042) struct list_head *tasks = &env->tasks;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8043) struct task_struct *p;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8044) struct rq_flags rf;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8045)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8046) rq_lock(env->dst_rq, &rf);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8047) update_rq_clock(env->dst_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8048)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8049) while (!list_empty(tasks)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8050) p = list_first_entry(tasks, struct task_struct, se.group_node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8051) list_del_init(&p->se.group_node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8052)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8053) attach_task(env->dst_rq, p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8054) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8055)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8056) rq_unlock(env->dst_rq, &rf);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8057) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8058)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8059) #ifdef CONFIG_NO_HZ_COMMON
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8060) static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8061) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8062) if (cfs_rq->avg.load_avg)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8063) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8064)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8065) if (cfs_rq->avg.util_avg)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8066) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8067)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8068) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8069) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8070)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8071) static inline bool others_have_blocked(struct rq *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8072) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8073) if (READ_ONCE(rq->avg_rt.util_avg))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8074) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8075)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8076) if (READ_ONCE(rq->avg_dl.util_avg))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8077) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8078)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8079) if (thermal_load_avg(rq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8080) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8081)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8082) #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8083) if (READ_ONCE(rq->avg_irq.util_avg))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8084) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8085) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8086)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8087) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8088) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8089)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8090) static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8091) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8092) rq->last_blocked_load_update_tick = jiffies;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8093)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8094) if (!has_blocked)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8095) rq->has_blocked_load = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8096) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8097) #else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8098) static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8099) static inline bool others_have_blocked(struct rq *rq) { return false; }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8100) static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8101) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8102)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8103) static bool __update_blocked_others(struct rq *rq, bool *done)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8104) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8105) const struct sched_class *curr_class;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8106) u64 now = rq_clock_pelt(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8107) unsigned long thermal_pressure;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8108) bool decayed;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8109)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8110) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8111) * update_load_avg() can call cpufreq_update_util(). Make sure that RT,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8112) * DL and IRQ signals have been updated before updating CFS.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8113) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8114) curr_class = rq->curr->sched_class;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8115)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8116) thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8117)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8118) decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) |
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8119) update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) |
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8120) update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure) |
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8121) update_irq_load_avg(rq, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8122)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8123) if (others_have_blocked(rq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8124) *done = false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8125)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8126) return decayed;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8127) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8128)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8129) #ifdef CONFIG_FAIR_GROUP_SCHED
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8130)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8131) static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8132) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8133) if (cfs_rq->load.weight)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8134) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8135)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8136) if (cfs_rq->avg.load_sum)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8137) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8138)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8139) if (cfs_rq->avg.util_sum)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8140) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8141)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8142) if (cfs_rq->avg.runnable_sum)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8143) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8144)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8145) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8146) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8147)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8148) static bool __update_blocked_fair(struct rq *rq, bool *done)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8149) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8150) struct cfs_rq *cfs_rq, *pos;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8151) bool decayed = false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8152) int cpu = cpu_of(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8153)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8154) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8155) * Iterates the task_group tree in a bottom up fashion, see
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8156) * list_add_leaf_cfs_rq() for details.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8157) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8158) for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8159) struct sched_entity *se;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8160)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8161) if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8162) update_tg_load_avg(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8163)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8164) if (cfs_rq == &rq->cfs)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8165) decayed = true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8166) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8167)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8168) /* Propagate pending load changes to the parent, if any: */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8169) se = cfs_rq->tg->se[cpu];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8170) if (se && !skip_blocked_update(se))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8171) update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8172)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8173) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8174) * There can be a lot of idle CPU cgroups. Don't let fully
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8175) * decayed cfs_rqs linger on the list.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8176) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8177) if (cfs_rq_is_decayed(cfs_rq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8178) list_del_leaf_cfs_rq(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8179)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8180) /* Don't need periodic decay once load/util_avg are null */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8181) if (cfs_rq_has_blocked(cfs_rq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8182) *done = false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8183) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8184)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8185) return decayed;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8186) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8187)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8188) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8189) * Compute the hierarchical load factor for cfs_rq and all its ascendants.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8190) * This needs to be done in a top-down fashion because the load of a child
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8191) * group is a fraction of its parents load.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8192) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8193) static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8194) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8195) struct rq *rq = rq_of(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8196) struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8197) unsigned long now = jiffies;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8198) unsigned long load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8199)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8200) if (cfs_rq->last_h_load_update == now)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8201) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8202)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8203) WRITE_ONCE(cfs_rq->h_load_next, NULL);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8204) for_each_sched_entity(se) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8205) cfs_rq = cfs_rq_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8206) WRITE_ONCE(cfs_rq->h_load_next, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8207) if (cfs_rq->last_h_load_update == now)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8208) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8209) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8210)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8211) if (!se) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8212) cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8213) cfs_rq->last_h_load_update = now;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8214) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8215)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8216) while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8217) load = cfs_rq->h_load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8218) load = div64_ul(load * se->avg.load_avg,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8219) cfs_rq_load_avg(cfs_rq) + 1);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8220) cfs_rq = group_cfs_rq(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8221) cfs_rq->h_load = load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8222) cfs_rq->last_h_load_update = now;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8223) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8224) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8225)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8226) static unsigned long task_h_load(struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8227) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8228) struct cfs_rq *cfs_rq = task_cfs_rq(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8229)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8230) update_cfs_rq_h_load(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8231) return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8232) cfs_rq_load_avg(cfs_rq) + 1);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8233) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8234) #else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8235) static bool __update_blocked_fair(struct rq *rq, bool *done)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8236) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8237) struct cfs_rq *cfs_rq = &rq->cfs;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8238) bool decayed;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8239)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8240) decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8241) if (cfs_rq_has_blocked(cfs_rq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8242) *done = false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8243)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8244) return decayed;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8245) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8246)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8247) static unsigned long task_h_load(struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8248) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8249) return p->se.avg.load_avg;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8250) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8251) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8252)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8253) static void update_blocked_averages(int cpu)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8254) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8255) bool decayed = false, done = true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8256) struct rq *rq = cpu_rq(cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8257) struct rq_flags rf;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8258)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8259) rq_lock_irqsave(rq, &rf);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8260) update_rq_clock(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8261)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8262) decayed |= __update_blocked_others(rq, &done);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8263) decayed |= __update_blocked_fair(rq, &done);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8264)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8265) update_blocked_load_status(rq, !done);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8266) if (decayed)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8267) cpufreq_update_util(rq, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8268) rq_unlock_irqrestore(rq, &rf);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8269) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8270)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8271) /********** Helpers for find_busiest_group ************************/
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8272)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8273) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8274) * sg_lb_stats - stats of a sched_group required for load_balancing
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8275) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8276) struct sg_lb_stats {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8277) unsigned long avg_load; /*Avg load across the CPUs of the group */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8278) unsigned long group_load; /* Total load over the CPUs of the group */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8279) unsigned long group_capacity;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8280) unsigned long group_util; /* Total utilization over the CPUs of the group */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8281) unsigned long group_runnable; /* Total runnable time over the CPUs of the group */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8282) unsigned int sum_nr_running; /* Nr of tasks running in the group */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8283) unsigned int sum_h_nr_running; /* Nr of CFS tasks running in the group */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8284) unsigned int idle_cpus;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8285) unsigned int group_weight;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8286) enum group_type group_type;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8287) unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8288) unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8289) #ifdef CONFIG_NUMA_BALANCING
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8290) unsigned int nr_numa_running;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8291) unsigned int nr_preferred_running;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8292) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8293) };
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8294)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8295) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8296) * sd_lb_stats - Structure to store the statistics of a sched_domain
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8297) * during load balancing.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8298) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8299) struct sd_lb_stats {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8300) struct sched_group *busiest; /* Busiest group in this sd */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8301) struct sched_group *local; /* Local group in this sd */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8302) unsigned long total_load; /* Total load of all groups in sd */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8303) unsigned long total_capacity; /* Total capacity of all groups in sd */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8304) unsigned long avg_load; /* Average load across all groups in sd */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8305) unsigned int prefer_sibling; /* tasks should go to sibling first */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8306)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8307) struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8308) struct sg_lb_stats local_stat; /* Statistics of the local group */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8309) };
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8310)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8311) static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8312) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8313) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8314) * Skimp on the clearing to avoid duplicate work. We can avoid clearing
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8315) * local_stat because update_sg_lb_stats() does a full clear/assignment.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8316) * We must however set busiest_stat::group_type and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8317) * busiest_stat::idle_cpus to the worst busiest group because
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8318) * update_sd_pick_busiest() reads these before assignment.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8319) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8320) *sds = (struct sd_lb_stats){
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8321) .busiest = NULL,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8322) .local = NULL,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8323) .total_load = 0UL,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8324) .total_capacity = 0UL,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8325) .busiest_stat = {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8326) .idle_cpus = UINT_MAX,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8327) .group_type = group_has_spare,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8328) },
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8329) };
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8330) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8331)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8332) static unsigned long scale_rt_capacity(int cpu)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8333) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8334) struct rq *rq = cpu_rq(cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8335) unsigned long max = arch_scale_cpu_capacity(cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8336) unsigned long used, free;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8337) unsigned long irq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8338)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8339) irq = cpu_util_irq(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8340)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8341) if (unlikely(irq >= max))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8342) return 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8343)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8344) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8345) * avg_rt.util_avg and avg_dl.util_avg track binary signals
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8346) * (running and not running) with weights 0 and 1024 respectively.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8347) * avg_thermal.load_avg tracks thermal pressure and the weighted
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8348) * average uses the actual delta max capacity(load).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8349) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8350) used = READ_ONCE(rq->avg_rt.util_avg);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8351) used += READ_ONCE(rq->avg_dl.util_avg);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8352) used += thermal_load_avg(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8353)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8354) if (unlikely(used >= max))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8355) return 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8356)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8357) free = max - used;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8358)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8359) return scale_irq_capacity(free, irq, max);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8360) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8361)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8362) static void update_cpu_capacity(struct sched_domain *sd, int cpu)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8363) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8364) unsigned long capacity = scale_rt_capacity(cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8365) struct sched_group *sdg = sd->groups;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8366)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8367) cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8368)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8369) if (!capacity)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8370) capacity = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8371)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8372) trace_android_rvh_update_cpu_capacity(cpu, &capacity);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8373) cpu_rq(cpu)->cpu_capacity = capacity;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8374) trace_sched_cpu_capacity_tp(cpu_rq(cpu));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8375)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8376) sdg->sgc->capacity = capacity;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8377) sdg->sgc->min_capacity = capacity;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8378) sdg->sgc->max_capacity = capacity;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8379) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8380)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8381) void update_group_capacity(struct sched_domain *sd, int cpu)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8382) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8383) struct sched_domain *child = sd->child;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8384) struct sched_group *group, *sdg = sd->groups;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8385) unsigned long capacity, min_capacity, max_capacity;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8386) unsigned long interval;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8387)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8388) interval = msecs_to_jiffies(sd->balance_interval);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8389) interval = clamp(interval, 1UL, max_load_balance_interval);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8390) sdg->sgc->next_update = jiffies + interval;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8391)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8392) if (!child) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8393) update_cpu_capacity(sd, cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8394) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8395) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8396)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8397) capacity = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8398) min_capacity = ULONG_MAX;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8399) max_capacity = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8400)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8401) if (child->flags & SD_OVERLAP) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8402) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8403) * SD_OVERLAP domains cannot assume that child groups
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8404) * span the current group.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8405) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8406)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8407) for_each_cpu(cpu, sched_group_span(sdg)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8408) unsigned long cpu_cap = capacity_of(cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8409)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8410) capacity += cpu_cap;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8411) min_capacity = min(cpu_cap, min_capacity);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8412) max_capacity = max(cpu_cap, max_capacity);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8413) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8414) } else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8415) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8416) * !SD_OVERLAP domains can assume that child groups
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8417) * span the current group.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8418) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8419)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8420) group = child->groups;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8421) do {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8422) struct sched_group_capacity *sgc = group->sgc;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8423)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8424) capacity += sgc->capacity;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8425) min_capacity = min(sgc->min_capacity, min_capacity);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8426) max_capacity = max(sgc->max_capacity, max_capacity);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8427) group = group->next;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8428) } while (group != child->groups);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8429) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8430)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8431) sdg->sgc->capacity = capacity;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8432) sdg->sgc->min_capacity = min_capacity;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8433) sdg->sgc->max_capacity = max_capacity;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8434) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8435)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8436) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8437) * Check whether the capacity of the rq has been noticeably reduced by side
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8438) * activity. The imbalance_pct is used for the threshold.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8439) * Return true is the capacity is reduced
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8440) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8441) static inline int
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8442) check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8443) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8444) return ((rq->cpu_capacity * sd->imbalance_pct) <
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8445) (rq->cpu_capacity_orig * 100));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8446) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8447)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8448) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8449) * Check whether a rq has a misfit task and if it looks like we can actually
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8450) * help that task: we can migrate the task to a CPU of higher capacity, or
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8451) * the task's current CPU is heavily pressured.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8452) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8453) static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8454) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8455) return rq->misfit_task_load &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8456) (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8457) check_cpu_capacity(rq, sd));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8458) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8459)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8460) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8461) * Group imbalance indicates (and tries to solve) the problem where balancing
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8462) * groups is inadequate due to ->cpus_ptr constraints.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8463) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8464) * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8465) * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8466) * Something like:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8467) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8468) * { 0 1 2 3 } { 4 5 6 7 }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8469) * * * * *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8470) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8471) * If we were to balance group-wise we'd place two tasks in the first group and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8472) * two tasks in the second group. Clearly this is undesired as it will overload
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8473) * cpu 3 and leave one of the CPUs in the second group unused.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8474) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8475) * The current solution to this issue is detecting the skew in the first group
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8476) * by noticing the lower domain failed to reach balance and had difficulty
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8477) * moving tasks due to affinity constraints.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8478) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8479) * When this is so detected; this group becomes a candidate for busiest; see
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8480) * update_sd_pick_busiest(). And calculate_imbalance() and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8481) * find_busiest_group() avoid some of the usual balance conditions to allow it
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8482) * to create an effective group imbalance.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8483) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8484) * This is a somewhat tricky proposition since the next run might not find the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8485) * group imbalance and decide the groups need to be balanced again. A most
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8486) * subtle and fragile situation.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8487) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8488)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8489) static inline int sg_imbalanced(struct sched_group *group)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8490) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8491) return group->sgc->imbalance;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8492) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8493)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8494) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8495) * group_has_capacity returns true if the group has spare capacity that could
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8496) * be used by some tasks.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8497) * We consider that a group has spare capacity if the * number of task is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8498) * smaller than the number of CPUs or if the utilization is lower than the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8499) * available capacity for CFS tasks.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8500) * For the latter, we use a threshold to stabilize the state, to take into
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8501) * account the variance of the tasks' load and to return true if the available
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8502) * capacity in meaningful for the load balancer.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8503) * As an example, an available capacity of 1% can appear but it doesn't make
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8504) * any benefit for the load balance.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8505) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8506) static inline bool
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8507) group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8508) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8509) if (sgs->sum_nr_running < sgs->group_weight)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8510) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8511)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8512) if ((sgs->group_capacity * imbalance_pct) <
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8513) (sgs->group_runnable * 100))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8514) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8515)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8516) if ((sgs->group_capacity * 100) >
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8517) (sgs->group_util * imbalance_pct))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8518) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8519)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8520) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8521) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8522)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8523) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8524) * group_is_overloaded returns true if the group has more tasks than it can
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8525) * handle.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8526) * group_is_overloaded is not equals to !group_has_capacity because a group
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8527) * with the exact right number of tasks, has no more spare capacity but is not
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8528) * overloaded so both group_has_capacity and group_is_overloaded return
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8529) * false.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8530) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8531) static inline bool
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8532) group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8533) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8534) if (sgs->sum_nr_running <= sgs->group_weight)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8535) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8536)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8537) if ((sgs->group_capacity * 100) <
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8538) (sgs->group_util * imbalance_pct))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8539) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8540)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8541) if ((sgs->group_capacity * imbalance_pct) <
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8542) (sgs->group_runnable * 100))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8543) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8544)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8545) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8546) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8547)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8548) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8549) * group_smaller_min_cpu_capacity: Returns true if sched_group sg has smaller
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8550) * per-CPU capacity than sched_group ref.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8551) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8552) static inline bool
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8553) group_smaller_min_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8554) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8555) return fits_capacity(sg->sgc->min_capacity, ref->sgc->min_capacity);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8556) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8557)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8558) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8559) * group_smaller_max_cpu_capacity: Returns true if sched_group sg has smaller
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8560) * per-CPU capacity_orig than sched_group ref.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8561) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8562) static inline bool
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8563) group_smaller_max_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8564) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8565) return fits_capacity(sg->sgc->max_capacity, ref->sgc->max_capacity);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8566) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8567)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8568) static inline enum
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8569) group_type group_classify(unsigned int imbalance_pct,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8570) struct sched_group *group,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8571) struct sg_lb_stats *sgs)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8572) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8573) if (group_is_overloaded(imbalance_pct, sgs))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8574) return group_overloaded;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8575)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8576) if (sg_imbalanced(group))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8577) return group_imbalanced;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8578)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8579) if (sgs->group_asym_packing)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8580) return group_asym_packing;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8581)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8582) if (sgs->group_misfit_task_load)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8583) return group_misfit_task;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8584)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8585) if (!group_has_capacity(imbalance_pct, sgs))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8586) return group_fully_busy;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8587)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8588) return group_has_spare;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8589) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8590)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8591) static bool update_nohz_stats(struct rq *rq, bool force)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8592) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8593) #ifdef CONFIG_NO_HZ_COMMON
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8594) unsigned int cpu = rq->cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8595)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8596) if (!rq->has_blocked_load)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8597) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8598)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8599) if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8600) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8601)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8602) if (!force && !time_after(jiffies, rq->last_blocked_load_update_tick))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8603) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8604)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8605) update_blocked_averages(cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8606)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8607) return rq->has_blocked_load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8608) #else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8609) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8610) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8611) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8612)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8613) /**
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8614) * update_sg_lb_stats - Update sched_group's statistics for load balancing.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8615) * @env: The load balancing environment.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8616) * @group: sched_group whose statistics are to be updated.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8617) * @sgs: variable to hold the statistics for this group.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8618) * @sg_status: Holds flag indicating the status of the sched_group
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8619) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8620) static inline void update_sg_lb_stats(struct lb_env *env,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8621) struct sched_group *group,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8622) struct sg_lb_stats *sgs,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8623) int *sg_status)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8624) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8625) int i, nr_running, local_group;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8626)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8627) memset(sgs, 0, sizeof(*sgs));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8628)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8629) local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(group));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8630)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8631) for_each_cpu_and(i, sched_group_span(group), env->cpus) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8632) struct rq *rq = cpu_rq(i);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8633)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8634) if ((env->flags & LBF_NOHZ_STATS) && update_nohz_stats(rq, false))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8635) env->flags |= LBF_NOHZ_AGAIN;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8636)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8637) sgs->group_load += cpu_load(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8638) sgs->group_util += cpu_util(i);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8639) sgs->group_runnable += cpu_runnable(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8640) sgs->sum_h_nr_running += rq->cfs.h_nr_running;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8641)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8642) nr_running = rq->nr_running;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8643) sgs->sum_nr_running += nr_running;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8644)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8645) if (nr_running > 1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8646) *sg_status |= SG_OVERLOAD;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8647)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8648) if (cpu_overutilized(i))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8649) *sg_status |= SG_OVERUTILIZED;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8650)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8651) #ifdef CONFIG_NUMA_BALANCING
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8652) sgs->nr_numa_running += rq->nr_numa_running;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8653) sgs->nr_preferred_running += rq->nr_preferred_running;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8654) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8655) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8656) * No need to call idle_cpu() if nr_running is not 0
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8657) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8658) if (!nr_running && idle_cpu(i)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8659) sgs->idle_cpus++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8660) /* Idle cpu can't have misfit task */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8661) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8662) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8663)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8664) if (local_group)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8665) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8666)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8667) /* Check for a misfit task on the cpu */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8668) if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8669) sgs->group_misfit_task_load < rq->misfit_task_load) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8670) sgs->group_misfit_task_load = rq->misfit_task_load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8671) *sg_status |= SG_OVERLOAD;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8672) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8673) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8674)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8675) /* Check if dst CPU is idle and preferred to this group */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8676) if (env->sd->flags & SD_ASYM_PACKING &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8677) env->idle != CPU_NOT_IDLE &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8678) sgs->sum_h_nr_running &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8679) sched_asym_prefer(env->dst_cpu, group->asym_prefer_cpu)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8680) sgs->group_asym_packing = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8681) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8682)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8683) sgs->group_capacity = group->sgc->capacity;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8684)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8685) sgs->group_weight = group->group_weight;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8686)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8687) sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8688)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8689) /* Computing avg_load makes sense only when group is overloaded */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8690) if (sgs->group_type == group_overloaded)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8691) sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8692) sgs->group_capacity;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8693) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8694)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8695) /**
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8696) * update_sd_pick_busiest - return 1 on busiest group
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8697) * @env: The load balancing environment.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8698) * @sds: sched_domain statistics
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8699) * @sg: sched_group candidate to be checked for being the busiest
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8700) * @sgs: sched_group statistics
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8701) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8702) * Determine if @sg is a busier group than the previously selected
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8703) * busiest group.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8704) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8705) * Return: %true if @sg is a busier group than the previously selected
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8706) * busiest group. %false otherwise.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8707) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8708) static bool update_sd_pick_busiest(struct lb_env *env,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8709) struct sd_lb_stats *sds,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8710) struct sched_group *sg,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8711) struct sg_lb_stats *sgs)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8712) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8713) struct sg_lb_stats *busiest = &sds->busiest_stat;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8714)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8715) /* Make sure that there is at least one task to pull */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8716) if (!sgs->sum_h_nr_running)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8717) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8718)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8719) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8720) * Don't try to pull misfit tasks we can't help.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8721) * We can use max_capacity here as reduction in capacity on some
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8722) * CPUs in the group should either be possible to resolve
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8723) * internally or be covered by avg_load imbalance (eventually).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8724) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8725) if (sgs->group_type == group_misfit_task &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8726) (!group_smaller_max_cpu_capacity(sg, sds->local) ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8727) sds->local_stat.group_type != group_has_spare))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8728) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8729)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8730) if (sgs->group_type > busiest->group_type)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8731) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8732)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8733) if (sgs->group_type < busiest->group_type)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8734) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8735)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8736) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8737) * The candidate and the current busiest group are the same type of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8738) * group. Let check which one is the busiest according to the type.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8739) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8740)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8741) switch (sgs->group_type) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8742) case group_overloaded:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8743) /* Select the overloaded group with highest avg_load. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8744) if (sgs->avg_load <= busiest->avg_load)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8745) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8746) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8747)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8748) case group_imbalanced:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8749) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8750) * Select the 1st imbalanced group as we don't have any way to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8751) * choose one more than another.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8752) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8753) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8754)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8755) case group_asym_packing:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8756) /* Prefer to move from lowest priority CPU's work */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8757) if (sched_asym_prefer(sg->asym_prefer_cpu, sds->busiest->asym_prefer_cpu))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8758) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8759) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8760)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8761) case group_misfit_task:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8762) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8763) * If we have more than one misfit sg go with the biggest
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8764) * misfit.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8765) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8766) if (sgs->group_misfit_task_load < busiest->group_misfit_task_load)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8767) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8768) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8769)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8770) case group_fully_busy:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8771) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8772) * Select the fully busy group with highest avg_load. In
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8773) * theory, there is no need to pull task from such kind of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8774) * group because tasks have all compute capacity that they need
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8775) * but we can still improve the overall throughput by reducing
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8776) * contention when accessing shared HW resources.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8777) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8778) * XXX for now avg_load is not computed and always 0 so we
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8779) * select the 1st one.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8780) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8781) if (sgs->avg_load <= busiest->avg_load)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8782) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8783) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8784)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8785) case group_has_spare:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8786) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8787) * Select not overloaded group with lowest number of idle cpus
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8788) * and highest number of running tasks. We could also compare
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8789) * the spare capacity which is more stable but it can end up
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8790) * that the group has less spare capacity but finally more idle
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8791) * CPUs which means less opportunity to pull tasks.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8792) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8793) if (sgs->idle_cpus > busiest->idle_cpus)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8794) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8795) else if ((sgs->idle_cpus == busiest->idle_cpus) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8796) (sgs->sum_nr_running <= busiest->sum_nr_running))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8797) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8798)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8799) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8800) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8801)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8802) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8803) * Candidate sg has no more than one task per CPU and has higher
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8804) * per-CPU capacity. Migrating tasks to less capable CPUs may harm
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8805) * throughput. Maximize throughput, power/energy consequences are not
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8806) * considered.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8807) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8808) if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8809) (sgs->group_type <= group_fully_busy) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8810) (group_smaller_min_cpu_capacity(sds->local, sg)))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8811) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8812)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8813) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8814) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8815)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8816) #ifdef CONFIG_NUMA_BALANCING
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8817) static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8818) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8819) if (sgs->sum_h_nr_running > sgs->nr_numa_running)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8820) return regular;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8821) if (sgs->sum_h_nr_running > sgs->nr_preferred_running)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8822) return remote;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8823) return all;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8824) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8825)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8826) static inline enum fbq_type fbq_classify_rq(struct rq *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8827) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8828) if (rq->nr_running > rq->nr_numa_running)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8829) return regular;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8830) if (rq->nr_running > rq->nr_preferred_running)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8831) return remote;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8832) return all;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8833) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8834) #else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8835) static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8836) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8837) return all;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8838) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8839)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8840) static inline enum fbq_type fbq_classify_rq(struct rq *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8841) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8842) return regular;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8843) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8844) #endif /* CONFIG_NUMA_BALANCING */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8845)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8846)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8847) struct sg_lb_stats;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8848)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8849) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8850) * task_running_on_cpu - return 1 if @p is running on @cpu.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8851) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8852)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8853) static unsigned int task_running_on_cpu(int cpu, struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8854) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8855) /* Task has no contribution or is new */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8856) if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8857) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8858)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8859) if (task_on_rq_queued(p))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8860) return 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8861)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8862) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8863) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8864)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8865) /**
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8866) * idle_cpu_without - would a given CPU be idle without p ?
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8867) * @cpu: the processor on which idleness is tested.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8868) * @p: task which should be ignored.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8869) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8870) * Return: 1 if the CPU would be idle. 0 otherwise.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8871) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8872) static int idle_cpu_without(int cpu, struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8873) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8874) struct rq *rq = cpu_rq(cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8875)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8876) if (rq->curr != rq->idle && rq->curr != p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8877) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8878)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8879) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8880) * rq->nr_running can't be used but an updated version without the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8881) * impact of p on cpu must be used instead. The updated nr_running
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8882) * be computed and tested before calling idle_cpu_without().
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8883) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8884)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8885) #ifdef CONFIG_SMP
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8886) if (rq->ttwu_pending)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8887) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8888) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8889)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8890) return 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8891) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8892)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8893) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8894) * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8895) * @sd: The sched_domain level to look for idlest group.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8896) * @group: sched_group whose statistics are to be updated.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8897) * @sgs: variable to hold the statistics for this group.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8898) * @p: The task for which we look for the idlest group/CPU.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8899) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8900) static inline void update_sg_wakeup_stats(struct sched_domain *sd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8901) struct sched_group *group,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8902) struct sg_lb_stats *sgs,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8903) struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8904) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8905) int i, nr_running;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8906)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8907) memset(sgs, 0, sizeof(*sgs));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8908)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8909) for_each_cpu(i, sched_group_span(group)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8910) struct rq *rq = cpu_rq(i);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8911) unsigned int local;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8912)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8913) sgs->group_load += cpu_load_without(rq, p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8914) sgs->group_util += cpu_util_without(i, p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8915) sgs->group_runnable += cpu_runnable_without(rq, p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8916) local = task_running_on_cpu(i, p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8917) sgs->sum_h_nr_running += rq->cfs.h_nr_running - local;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8918)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8919) nr_running = rq->nr_running - local;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8920) sgs->sum_nr_running += nr_running;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8921)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8922) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8923) * No need to call idle_cpu_without() if nr_running is not 0
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8924) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8925) if (!nr_running && idle_cpu_without(i, p))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8926) sgs->idle_cpus++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8927)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8928) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8929)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8930) /* Check if task fits in the group */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8931) if (sd->flags & SD_ASYM_CPUCAPACITY &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8932) !task_fits_capacity(p, group->sgc->max_capacity)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8933) sgs->group_misfit_task_load = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8934) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8935)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8936) sgs->group_capacity = group->sgc->capacity;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8937)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8938) sgs->group_weight = group->group_weight;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8939)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8940) sgs->group_type = group_classify(sd->imbalance_pct, group, sgs);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8941)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8942) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8943) * Computing avg_load makes sense only when group is fully busy or
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8944) * overloaded
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8945) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8946) if (sgs->group_type == group_fully_busy ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8947) sgs->group_type == group_overloaded)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8948) sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8949) sgs->group_capacity;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8950) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8951)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8952) static bool update_pick_idlest(struct sched_group *idlest,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8953) struct sg_lb_stats *idlest_sgs,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8954) struct sched_group *group,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8955) struct sg_lb_stats *sgs)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8956) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8957) if (sgs->group_type < idlest_sgs->group_type)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8958) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8959)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8960) if (sgs->group_type > idlest_sgs->group_type)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8961) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8962)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8963) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8964) * The candidate and the current idlest group are the same type of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8965) * group. Let check which one is the idlest according to the type.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8966) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8967)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8968) switch (sgs->group_type) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8969) case group_overloaded:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8970) case group_fully_busy:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8971) /* Select the group with lowest avg_load. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8972) if (idlest_sgs->avg_load <= sgs->avg_load)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8973) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8974) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8975)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8976) case group_imbalanced:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8977) case group_asym_packing:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8978) /* Those types are not used in the slow wakeup path */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8979) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8980)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8981) case group_misfit_task:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8982) /* Select group with the highest max capacity */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8983) if (idlest->sgc->max_capacity >= group->sgc->max_capacity)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8984) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8985) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8986)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8987) case group_has_spare:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8988) /* Select group with most idle CPUs */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8989) if (idlest_sgs->idle_cpus > sgs->idle_cpus)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8990) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8991)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8992) /* Select group with lowest group_util */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8993) if (idlest_sgs->idle_cpus == sgs->idle_cpus &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8994) idlest_sgs->group_util <= sgs->group_util)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8995) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8996)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8997) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8998) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8999)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9000) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9001) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9002)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9003) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9004) * find_idlest_group() finds and returns the least busy CPU group within the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9005) * domain.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9006) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9007) * Assumes p is allowed on at least one CPU in sd.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9008) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9009) static struct sched_group *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9010) find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9011) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9012) struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9013) struct sg_lb_stats local_sgs, tmp_sgs;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9014) struct sg_lb_stats *sgs;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9015) unsigned long imbalance;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9016) struct sg_lb_stats idlest_sgs = {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9017) .avg_load = UINT_MAX,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9018) .group_type = group_overloaded,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9019) };
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9020)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9021) imbalance = scale_load_down(NICE_0_LOAD) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9022) (sd->imbalance_pct-100) / 100;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9023)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9024) do {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9025) int local_group;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9026)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9027) if (IS_ENABLED(CONFIG_ROCKCHIP_PERFORMANCE)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9028) struct root_domain *rd = cpu_rq(this_cpu)->rd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9029) struct cpumask *cpub_mask = rockchip_perf_get_cpub_mask();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9030) int level = rockchip_perf_get_level();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9031)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9032) if ((level == ROCKCHIP_PERFORMANCE_HIGH) && !READ_ONCE(rd->overutilized) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9033) cpub_mask && cpumask_intersects(p->cpus_ptr, cpub_mask) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9034) !cpumask_intersects(sched_group_span(group), cpub_mask))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9035) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9036) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9037)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9038) /* Skip over this group if it has no CPUs allowed */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9039) if (!cpumask_intersects(sched_group_span(group),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9040) p->cpus_ptr))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9041) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9042)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9043) local_group = cpumask_test_cpu(this_cpu,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9044) sched_group_span(group));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9045)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9046) if (local_group) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9047) sgs = &local_sgs;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9048) local = group;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9049) } else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9050) sgs = &tmp_sgs;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9051) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9052)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9053) update_sg_wakeup_stats(sd, group, sgs, p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9054)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9055) if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9056) idlest = group;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9057) idlest_sgs = *sgs;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9058) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9059)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9060) } while (group = group->next, group != sd->groups);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9061)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9062)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9063) /* There is no idlest group to push tasks to */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9064) if (!idlest)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9065) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9066)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9067) /* The local group has been skipped because of CPU affinity */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9068) if (!local)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9069) return idlest;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9070)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9071) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9072) * If the local group is idler than the selected idlest group
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9073) * don't try and push the task.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9074) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9075) if (local_sgs.group_type < idlest_sgs.group_type)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9076) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9077)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9078) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9079) * If the local group is busier than the selected idlest group
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9080) * try and push the task.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9081) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9082) if (local_sgs.group_type > idlest_sgs.group_type)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9083) return idlest;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9084)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9085) switch (local_sgs.group_type) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9086) case group_overloaded:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9087) case group_fully_busy:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9088) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9089) * When comparing groups across NUMA domains, it's possible for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9090) * the local domain to be very lightly loaded relative to the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9091) * remote domains but "imbalance" skews the comparison making
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9092) * remote CPUs look much more favourable. When considering
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9093) * cross-domain, add imbalance to the load on the remote node
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9094) * and consider staying local.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9095) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9096)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9097) if ((sd->flags & SD_NUMA) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9098) ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9099) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9100)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9101) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9102) * If the local group is less loaded than the selected
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9103) * idlest group don't try and push any tasks.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9104) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9105) if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9106) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9107)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9108) if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9109) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9110) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9111)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9112) case group_imbalanced:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9113) case group_asym_packing:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9114) /* Those type are not used in the slow wakeup path */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9115) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9116)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9117) case group_misfit_task:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9118) /* Select group with the highest max capacity */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9119) if (local->sgc->max_capacity >= idlest->sgc->max_capacity)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9120) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9121) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9122)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9123) case group_has_spare:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9124) if (sd->flags & SD_NUMA) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9125) #ifdef CONFIG_NUMA_BALANCING
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9126) int idlest_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9127) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9128) * If there is spare capacity at NUMA, try to select
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9129) * the preferred node
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9130) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9131) if (cpu_to_node(this_cpu) == p->numa_preferred_nid)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9132) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9133)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9134) idlest_cpu = cpumask_first(sched_group_span(idlest));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9135) if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9136) return idlest;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9137) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9138) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9139) * Otherwise, keep the task on this node to stay close
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9140) * its wakeup source and improve locality. If there is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9141) * a real need of migration, periodic load balance will
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9142) * take care of it.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9143) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9144) if (local_sgs.idle_cpus)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9145) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9146) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9147)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9148) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9149) * Select group with highest number of idle CPUs. We could also
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9150) * compare the utilization which is more stable but it can end
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9151) * up that the group has less spare capacity but finally more
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9152) * idle CPUs which means more opportunity to run task.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9153) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9154) if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9155) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9156) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9157) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9158)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9159) return idlest;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9160) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9161)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9162) /**
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9163) * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9164) * @env: The load balancing environment.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9165) * @sds: variable to hold the statistics for this sched_domain.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9166) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9167)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9168) static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9169) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9170) struct sched_domain *child = env->sd->child;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9171) struct sched_group *sg = env->sd->groups;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9172) struct sg_lb_stats *local = &sds->local_stat;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9173) struct sg_lb_stats tmp_sgs;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9174) int sg_status = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9175)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9176) #ifdef CONFIG_NO_HZ_COMMON
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9177) if (env->idle == CPU_NEWLY_IDLE && READ_ONCE(nohz.has_blocked))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9178) env->flags |= LBF_NOHZ_STATS;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9179) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9180)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9181) do {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9182) struct sg_lb_stats *sgs = &tmp_sgs;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9183) int local_group;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9184)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9185) local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9186) if (local_group) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9187) sds->local = sg;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9188) sgs = local;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9189)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9190) if (env->idle != CPU_NEWLY_IDLE ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9191) time_after_eq(jiffies, sg->sgc->next_update))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9192) update_group_capacity(env->sd, env->dst_cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9193) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9194)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9195) update_sg_lb_stats(env, sg, sgs, &sg_status);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9196)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9197) if (local_group)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9198) goto next_group;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9199)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9200)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9201) if (update_sd_pick_busiest(env, sds, sg, sgs)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9202) sds->busiest = sg;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9203) sds->busiest_stat = *sgs;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9204) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9205)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9206) next_group:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9207) /* Now, start updating sd_lb_stats */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9208) sds->total_load += sgs->group_load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9209) sds->total_capacity += sgs->group_capacity;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9210)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9211) sg = sg->next;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9212) } while (sg != env->sd->groups);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9213)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9214) /* Tag domain that child domain prefers tasks go to siblings first */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9215) sds->prefer_sibling = child && child->flags & SD_PREFER_SIBLING;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9216)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9217) #ifdef CONFIG_NO_HZ_COMMON
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9218) if ((env->flags & LBF_NOHZ_AGAIN) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9219) cpumask_subset(nohz.idle_cpus_mask, sched_domain_span(env->sd))) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9220)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9221) WRITE_ONCE(nohz.next_blocked,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9222) jiffies + msecs_to_jiffies(LOAD_AVG_PERIOD));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9223) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9224) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9225)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9226) if (env->sd->flags & SD_NUMA)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9227) env->fbq_type = fbq_classify_group(&sds->busiest_stat);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9228)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9229) if (!env->sd->parent) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9230) struct root_domain *rd = env->dst_rq->rd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9231)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9232) /* update overload indicator if we are at root domain */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9233) WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9234)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9235) /* Update over-utilization (tipping point, U >= 0) indicator */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9236) WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9237) trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9238) } else if (sg_status & SG_OVERUTILIZED) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9239) struct root_domain *rd = env->dst_rq->rd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9240)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9241) WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9242) trace_sched_overutilized_tp(rd, SG_OVERUTILIZED);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9243) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9244) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9245)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9246) static inline long adjust_numa_imbalance(int imbalance, int nr_running)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9247) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9248) unsigned int imbalance_min;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9249)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9250) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9251) * Allow a small imbalance based on a simple pair of communicating
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9252) * tasks that remain local when the source domain is almost idle.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9253) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9254) imbalance_min = 2;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9255) if (nr_running <= imbalance_min)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9256) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9257)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9258) return imbalance;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9259) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9260)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9261) /**
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9262) * calculate_imbalance - Calculate the amount of imbalance present within the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9263) * groups of a given sched_domain during load balance.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9264) * @env: load balance environment
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9265) * @sds: statistics of the sched_domain whose imbalance is to be calculated.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9266) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9267) static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9268) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9269) struct sg_lb_stats *local, *busiest;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9270)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9271) local = &sds->local_stat;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9272) busiest = &sds->busiest_stat;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9273)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9274) if (busiest->group_type == group_misfit_task) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9275) /* Set imbalance to allow misfit tasks to be balanced. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9276) env->migration_type = migrate_misfit;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9277) env->imbalance = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9278) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9279) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9280)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9281) if (busiest->group_type == group_asym_packing) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9282) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9283) * In case of asym capacity, we will try to migrate all load to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9284) * the preferred CPU.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9285) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9286) env->migration_type = migrate_task;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9287) env->imbalance = busiest->sum_h_nr_running;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9288) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9289) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9290)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9291) if (busiest->group_type == group_imbalanced) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9292) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9293) * In the group_imb case we cannot rely on group-wide averages
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9294) * to ensure CPU-load equilibrium, try to move any task to fix
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9295) * the imbalance. The next load balance will take care of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9296) * balancing back the system.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9297) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9298) env->migration_type = migrate_task;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9299) env->imbalance = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9300) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9301) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9302)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9303) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9304) * Try to use spare capacity of local group without overloading it or
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9305) * emptying busiest.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9306) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9307) if (local->group_type == group_has_spare) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9308) if ((busiest->group_type > group_fully_busy) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9309) !(env->sd->flags & SD_SHARE_PKG_RESOURCES)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9310) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9311) * If busiest is overloaded, try to fill spare
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9312) * capacity. This might end up creating spare capacity
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9313) * in busiest or busiest still being overloaded but
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9314) * there is no simple way to directly compute the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9315) * amount of load to migrate in order to balance the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9316) * system.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9317) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9318) env->migration_type = migrate_util;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9319) env->imbalance = max(local->group_capacity, local->group_util) -
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9320) local->group_util;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9321)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9322) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9323) * In some cases, the group's utilization is max or even
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9324) * higher than capacity because of migrations but the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9325) * local CPU is (newly) idle. There is at least one
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9326) * waiting task in this overloaded busiest group. Let's
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9327) * try to pull it.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9328) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9329) if (env->idle != CPU_NOT_IDLE && env->imbalance == 0) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9330) env->migration_type = migrate_task;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9331) env->imbalance = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9332) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9333)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9334) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9335) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9336)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9337) if (busiest->group_weight == 1 || sds->prefer_sibling) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9338) unsigned int nr_diff = busiest->sum_nr_running;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9339) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9340) * When prefer sibling, evenly spread running tasks on
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9341) * groups.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9342) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9343) env->migration_type = migrate_task;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9344) lsub_positive(&nr_diff, local->sum_nr_running);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9345) env->imbalance = nr_diff >> 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9346) } else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9347)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9348) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9349) * If there is no overload, we just want to even the number of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9350) * idle cpus.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9351) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9352) env->migration_type = migrate_task;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9353) env->imbalance = max_t(long, 0, (local->idle_cpus -
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9354) busiest->idle_cpus) >> 1);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9355) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9356)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9357) /* Consider allowing a small imbalance between NUMA groups */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9358) if (env->sd->flags & SD_NUMA)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9359) env->imbalance = adjust_numa_imbalance(env->imbalance,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9360) busiest->sum_nr_running);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9361)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9362) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9363) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9364)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9365) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9366) * Local is fully busy but has to take more load to relieve the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9367) * busiest group
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9368) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9369) if (local->group_type < group_overloaded) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9370) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9371) * Local will become overloaded so the avg_load metrics are
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9372) * finally needed.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9373) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9374)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9375) local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) /
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9376) local->group_capacity;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9377)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9378) sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) /
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9379) sds->total_capacity;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9380) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9381) * If the local group is more loaded than the selected
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9382) * busiest group don't try to pull any tasks.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9383) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9384) if (local->avg_load >= busiest->avg_load) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9385) env->imbalance = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9386) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9387) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9388) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9389)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9390) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9391) * Both group are or will become overloaded and we're trying to get all
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9392) * the CPUs to the average_load, so we don't want to push ourselves
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9393) * above the average load, nor do we wish to reduce the max loaded CPU
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9394) * below the average load. At the same time, we also don't want to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9395) * reduce the group load below the group capacity. Thus we look for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9396) * the minimum possible imbalance.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9397) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9398) env->migration_type = migrate_load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9399) env->imbalance = min(
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9400) (busiest->avg_load - sds->avg_load) * busiest->group_capacity,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9401) (sds->avg_load - local->avg_load) * local->group_capacity
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9402) ) / SCHED_CAPACITY_SCALE;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9403) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9404)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9405) /******* find_busiest_group() helpers end here *********************/
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9406)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9407) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9408) * Decision matrix according to the local and busiest group type:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9409) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9410) * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9411) * has_spare nr_idle balanced N/A N/A balanced balanced
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9412) * fully_busy nr_idle nr_idle N/A N/A balanced balanced
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9413) * misfit_task force N/A N/A N/A force force
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9414) * asym_packing force force N/A N/A force force
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9415) * imbalanced force force N/A N/A force force
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9416) * overloaded force force N/A N/A force avg_load
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9417) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9418) * N/A : Not Applicable because already filtered while updating
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9419) * statistics.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9420) * balanced : The system is balanced for these 2 groups.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9421) * force : Calculate the imbalance as load migration is probably needed.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9422) * avg_load : Only if imbalance is significant enough.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9423) * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9424) * different in groups.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9425) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9426)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9427) /**
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9428) * find_busiest_group - Returns the busiest group within the sched_domain
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9429) * if there is an imbalance.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9430) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9431) * Also calculates the amount of runnable load which should be moved
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9432) * to restore balance.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9433) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9434) * @env: The load balancing environment.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9435) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9436) * Return: - The busiest group if imbalance exists.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9437) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9438) static struct sched_group *find_busiest_group(struct lb_env *env)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9439) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9440) struct sg_lb_stats *local, *busiest;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9441) struct sd_lb_stats sds;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9442)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9443) init_sd_lb_stats(&sds);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9444)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9445) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9446) * Compute the various statistics relevant for load balancing at
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9447) * this level.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9448) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9449) update_sd_lb_stats(env, &sds);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9450)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9451) if (sched_energy_enabled()) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9452) struct root_domain *rd = env->dst_rq->rd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9453) int out_balance = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9454)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9455) trace_android_rvh_find_busiest_group(sds.busiest, env->dst_rq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9456) &out_balance);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9457) if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9458) && out_balance)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9459) goto out_balanced;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9460) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9461)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9462) local = &sds.local_stat;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9463) busiest = &sds.busiest_stat;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9464)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9465) /* There is no busy sibling group to pull tasks from */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9466) if (!sds.busiest)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9467) goto out_balanced;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9468)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9469) /* Misfit tasks should be dealt with regardless of the avg load */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9470) if (busiest->group_type == group_misfit_task)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9471) goto force_balance;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9472)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9473) /* ASYM feature bypasses nice load balance check */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9474) if (busiest->group_type == group_asym_packing)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9475) goto force_balance;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9476)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9477) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9478) * If the busiest group is imbalanced the below checks don't
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9479) * work because they assume all things are equal, which typically
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9480) * isn't true due to cpus_ptr constraints and the like.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9481) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9482) if (busiest->group_type == group_imbalanced)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9483) goto force_balance;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9484)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9485) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9486) * If the local group is busier than the selected busiest group
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9487) * don't try and pull any tasks.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9488) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9489) if (local->group_type > busiest->group_type)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9490) goto out_balanced;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9491)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9492) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9493) * When groups are overloaded, use the avg_load to ensure fairness
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9494) * between tasks.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9495) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9496) if (local->group_type == group_overloaded) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9497) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9498) * If the local group is more loaded than the selected
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9499) * busiest group don't try to pull any tasks.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9500) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9501) if (local->avg_load >= busiest->avg_load)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9502) goto out_balanced;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9503)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9504) /* XXX broken for overlapping NUMA groups */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9505) sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) /
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9506) sds.total_capacity;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9507)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9508) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9509) * Don't pull any tasks if this group is already above the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9510) * domain average load.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9511) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9512) if (local->avg_load >= sds.avg_load)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9513) goto out_balanced;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9514)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9515) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9516) * If the busiest group is more loaded, use imbalance_pct to be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9517) * conservative.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9518) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9519) if (100 * busiest->avg_load <=
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9520) env->sd->imbalance_pct * local->avg_load)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9521) goto out_balanced;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9522) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9523)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9524) /* Try to move all excess tasks to child's sibling domain */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9525) if (sds.prefer_sibling && local->group_type == group_has_spare &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9526) busiest->sum_nr_running > local->sum_nr_running + 1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9527) goto force_balance;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9528)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9529) if (busiest->group_type != group_overloaded) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9530) if (env->idle == CPU_NOT_IDLE)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9531) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9532) * If the busiest group is not overloaded (and as a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9533) * result the local one too) but this CPU is already
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9534) * busy, let another idle CPU try to pull task.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9535) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9536) goto out_balanced;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9537)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9538) if (busiest->group_weight > 1 &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9539) local->idle_cpus <= (busiest->idle_cpus + 1))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9540) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9541) * If the busiest group is not overloaded
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9542) * and there is no imbalance between this and busiest
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9543) * group wrt idle CPUs, it is balanced. The imbalance
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9544) * becomes significant if the diff is greater than 1
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9545) * otherwise we might end up to just move the imbalance
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9546) * on another group. Of course this applies only if
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9547) * there is more than 1 CPU per group.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9548) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9549) goto out_balanced;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9550)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9551) if (busiest->sum_h_nr_running == 1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9552) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9553) * busiest doesn't have any tasks waiting to run
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9554) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9555) goto out_balanced;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9556) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9557)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9558) force_balance:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9559) /* Looks like there is an imbalance. Compute it */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9560) calculate_imbalance(env, &sds);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9561) return env->imbalance ? sds.busiest : NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9562)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9563) out_balanced:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9564) env->imbalance = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9565) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9566) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9567)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9568) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9569) * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9570) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9571) static struct rq *find_busiest_queue(struct lb_env *env,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9572) struct sched_group *group)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9573) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9574) struct rq *busiest = NULL, *rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9575) unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9576) unsigned int busiest_nr = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9577) int i, done = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9578)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9579) trace_android_rvh_find_busiest_queue(env->dst_cpu, group, env->cpus,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9580) &busiest, &done);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9581) if (done)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9582) return busiest;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9583)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9584) for_each_cpu_and(i, sched_group_span(group), env->cpus) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9585) unsigned long capacity, load, util;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9586) unsigned int nr_running;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9587) enum fbq_type rt;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9588)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9589) rq = cpu_rq(i);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9590) rt = fbq_classify_rq(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9591)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9592) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9593) * We classify groups/runqueues into three groups:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9594) * - regular: there are !numa tasks
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9595) * - remote: there are numa tasks that run on the 'wrong' node
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9596) * - all: there is no distinction
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9597) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9598) * In order to avoid migrating ideally placed numa tasks,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9599) * ignore those when there's better options.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9600) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9601) * If we ignore the actual busiest queue to migrate another
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9602) * task, the next balance pass can still reduce the busiest
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9603) * queue by moving tasks around inside the node.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9604) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9605) * If we cannot move enough load due to this classification
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9606) * the next pass will adjust the group classification and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9607) * allow migration of more tasks.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9608) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9609) * Both cases only affect the total convergence complexity.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9610) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9611) if (rt > env->fbq_type)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9612) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9613)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9614) capacity = capacity_of(i);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9615) nr_running = rq->cfs.h_nr_running;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9616)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9617) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9618) * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9619) * eventually lead to active_balancing high->low capacity.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9620) * Higher per-CPU capacity is considered better than balancing
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9621) * average load.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9622) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9623) if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9624) capacity_of(env->dst_cpu) < capacity &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9625) nr_running == 1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9626) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9627)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9628) switch (env->migration_type) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9629) case migrate_load:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9630) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9631) * When comparing with load imbalance, use cpu_load()
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9632) * which is not scaled with the CPU capacity.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9633) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9634) load = cpu_load(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9635)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9636) if (nr_running == 1 && load > env->imbalance &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9637) !check_cpu_capacity(rq, env->sd))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9638) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9639)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9640) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9641) * For the load comparisons with the other CPUs,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9642) * consider the cpu_load() scaled with the CPU
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9643) * capacity, so that the load can be moved away
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9644) * from the CPU that is potentially running at a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9645) * lower capacity.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9646) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9647) * Thus we're looking for max(load_i / capacity_i),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9648) * crosswise multiplication to rid ourselves of the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9649) * division works out to:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9650) * load_i * capacity_j > load_j * capacity_i;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9651) * where j is our previous maximum.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9652) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9653) if (load * busiest_capacity > busiest_load * capacity) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9654) busiest_load = load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9655) busiest_capacity = capacity;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9656) busiest = rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9657) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9658) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9659)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9660) case migrate_util:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9661) util = cpu_util(cpu_of(rq));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9662)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9663) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9664) * Don't try to pull utilization from a CPU with one
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9665) * running task. Whatever its utilization, we will fail
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9666) * detach the task.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9667) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9668) if (nr_running <= 1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9669) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9670)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9671) if (busiest_util < util) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9672) busiest_util = util;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9673) busiest = rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9674) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9675) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9676)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9677) case migrate_task:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9678) if (busiest_nr < nr_running) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9679) busiest_nr = nr_running;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9680) busiest = rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9681) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9682) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9683)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9684) case migrate_misfit:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9685) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9686) * For ASYM_CPUCAPACITY domains with misfit tasks we
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9687) * simply seek the "biggest" misfit task.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9688) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9689) if (rq->misfit_task_load > busiest_load) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9690) busiest_load = rq->misfit_task_load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9691) busiest = rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9692) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9693)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9694) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9695)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9696) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9697) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9698)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9699) return busiest;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9700) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9701)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9702) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9703) * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9704) * so long as it is large enough.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9705) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9706) #define MAX_PINNED_INTERVAL 512
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9707)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9708) static inline bool
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9709) asym_active_balance(struct lb_env *env)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9710) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9711) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9712) * ASYM_PACKING needs to force migrate tasks from busy but
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9713) * lower priority CPUs in order to pack all tasks in the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9714) * highest priority CPUs.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9715) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9716) return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9717) sched_asym_prefer(env->dst_cpu, env->src_cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9718) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9719)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9720) static inline bool
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9721) voluntary_active_balance(struct lb_env *env)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9722) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9723) struct sched_domain *sd = env->sd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9724)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9725) if (asym_active_balance(env))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9726) return 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9727)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9728) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9729) * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9730) * It's worth migrating the task if the src_cpu's capacity is reduced
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9731) * because of other sched_class or IRQs if more capacity stays
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9732) * available on dst_cpu.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9733) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9734) if ((env->idle != CPU_NOT_IDLE) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9735) (env->src_rq->cfs.h_nr_running == 1)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9736) if ((check_cpu_capacity(env->src_rq, sd)) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9737) (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9738) return 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9739) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9740)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9741) if (env->migration_type == migrate_misfit)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9742) return 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9743)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9744) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9745) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9746)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9747) static int need_active_balance(struct lb_env *env)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9748) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9749) struct sched_domain *sd = env->sd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9750)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9751) if (voluntary_active_balance(env))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9752) return 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9753)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9754) return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9755) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9756)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9757) static int active_load_balance_cpu_stop(void *data);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9758)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9759) static int should_we_balance(struct lb_env *env)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9760) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9761) struct sched_group *sg = env->sd->groups;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9762) int cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9763)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9764) if (IS_ENABLED(CONFIG_ROCKCHIP_PERFORMANCE)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9765) struct root_domain *rd = env->dst_rq->rd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9766) struct cpumask *cpul_mask = rockchip_perf_get_cpul_mask();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9767) int level = rockchip_perf_get_level();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9768)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9769) if ((level == ROCKCHIP_PERFORMANCE_HIGH) && !READ_ONCE(rd->overutilized) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9770) cpul_mask && cpumask_test_cpu(env->dst_cpu, cpul_mask))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9771) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9772) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9773)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9774) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9775) * Ensure the balancing environment is consistent; can happen
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9776) * when the softirq triggers 'during' hotplug.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9777) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9778) if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9779) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9780)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9781) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9782) * In the newly idle case, we will allow all the CPUs
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9783) * to do the newly idle load balance.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9784) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9785) if (env->idle == CPU_NEWLY_IDLE)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9786) return 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9787)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9788) /* Try to find first idle CPU */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9789) for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9790) if (!idle_cpu(cpu))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9791) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9792)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9793) /* Are we the first idle CPU? */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9794) return cpu == env->dst_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9795) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9796)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9797) /* Are we the first CPU of this group ? */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9798) return group_balance_cpu(sg) == env->dst_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9799) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9800)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9801) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9802) * Check this_cpu to ensure it is balanced within domain. Attempt to move
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9803) * tasks if there is an imbalance.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9804) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9805) static int load_balance(int this_cpu, struct rq *this_rq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9806) struct sched_domain *sd, enum cpu_idle_type idle,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9807) int *continue_balancing)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9808) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9809) int ld_moved, cur_ld_moved, active_balance = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9810) struct sched_domain *sd_parent = sd->parent;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9811) struct sched_group *group;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9812) struct rq *busiest;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9813) struct rq_flags rf;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9814) struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9815)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9816) struct lb_env env = {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9817) .sd = sd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9818) .dst_cpu = this_cpu,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9819) .dst_rq = this_rq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9820) .dst_grpmask = sched_group_span(sd->groups),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9821) .idle = idle,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9822) .loop_break = sched_nr_migrate_break,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9823) .cpus = cpus,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9824) .fbq_type = all,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9825) .tasks = LIST_HEAD_INIT(env.tasks),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9826) };
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9827)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9828) cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9829)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9830) schedstat_inc(sd->lb_count[idle]);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9831)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9832) redo:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9833) if (!should_we_balance(&env)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9834) *continue_balancing = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9835) goto out_balanced;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9836) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9837)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9838) group = find_busiest_group(&env);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9839) if (!group) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9840) schedstat_inc(sd->lb_nobusyg[idle]);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9841) goto out_balanced;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9842) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9843)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9844) busiest = find_busiest_queue(&env, group);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9845) if (!busiest) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9846) schedstat_inc(sd->lb_nobusyq[idle]);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9847) goto out_balanced;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9848) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9849)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9850) BUG_ON(busiest == env.dst_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9851)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9852) schedstat_add(sd->lb_imbalance[idle], env.imbalance);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9853)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9854) env.src_cpu = busiest->cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9855) env.src_rq = busiest;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9856)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9857) ld_moved = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9858) if (busiest->nr_running > 1) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9859) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9860) * Attempt to move tasks. If find_busiest_group has found
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9861) * an imbalance but busiest->nr_running <= 1, the group is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9862) * still unbalanced. ld_moved simply stays zero, so it is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9863) * correctly treated as an imbalance.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9864) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9865) env.flags |= LBF_ALL_PINNED;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9866) env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9867)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9868) more_balance:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9869) rq_lock_irqsave(busiest, &rf);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9870) env.src_rq_rf = &rf;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9871) update_rq_clock(busiest);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9872)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9873) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9874) * cur_ld_moved - load moved in current iteration
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9875) * ld_moved - cumulative load moved across iterations
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9876) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9877) cur_ld_moved = detach_tasks(&env);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9878)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9879) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9880) * We've detached some tasks from busiest_rq. Every
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9881) * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9882) * unlock busiest->lock, and we are able to be sure
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9883) * that nobody can manipulate the tasks in parallel.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9884) * See task_rq_lock() family for the details.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9885) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9886)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9887) rq_unlock(busiest, &rf);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9888)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9889) if (cur_ld_moved) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9890) attach_tasks(&env);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9891) ld_moved += cur_ld_moved;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9892) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9893)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9894) local_irq_restore(rf.flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9895)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9896) if (env.flags & LBF_NEED_BREAK) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9897) env.flags &= ~LBF_NEED_BREAK;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9898) goto more_balance;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9899) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9900)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9901) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9902) * Revisit (affine) tasks on src_cpu that couldn't be moved to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9903) * us and move them to an alternate dst_cpu in our sched_group
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9904) * where they can run. The upper limit on how many times we
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9905) * iterate on same src_cpu is dependent on number of CPUs in our
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9906) * sched_group.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9907) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9908) * This changes load balance semantics a bit on who can move
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9909) * load to a given_cpu. In addition to the given_cpu itself
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9910) * (or a ilb_cpu acting on its behalf where given_cpu is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9911) * nohz-idle), we now have balance_cpu in a position to move
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9912) * load to given_cpu. In rare situations, this may cause
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9913) * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9914) * _independently_ and at _same_ time to move some load to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9915) * given_cpu) causing exceess load to be moved to given_cpu.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9916) * This however should not happen so much in practice and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9917) * moreover subsequent load balance cycles should correct the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9918) * excess load moved.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9919) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9920) if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9921)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9922) /* Prevent to re-select dst_cpu via env's CPUs */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9923) __cpumask_clear_cpu(env.dst_cpu, env.cpus);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9924)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9925) env.dst_rq = cpu_rq(env.new_dst_cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9926) env.dst_cpu = env.new_dst_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9927) env.flags &= ~LBF_DST_PINNED;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9928) env.loop = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9929) env.loop_break = sched_nr_migrate_break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9930)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9931) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9932) * Go back to "more_balance" rather than "redo" since we
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9933) * need to continue with same src_cpu.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9934) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9935) goto more_balance;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9936) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9937)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9938) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9939) * We failed to reach balance because of affinity.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9940) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9941) if (sd_parent) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9942) int *group_imbalance = &sd_parent->groups->sgc->imbalance;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9943)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9944) if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9945) *group_imbalance = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9946) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9947)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9948) /* All tasks on this runqueue were pinned by CPU affinity */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9949) if (unlikely(env.flags & LBF_ALL_PINNED)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9950) __cpumask_clear_cpu(cpu_of(busiest), cpus);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9951) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9952) * Attempting to continue load balancing at the current
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9953) * sched_domain level only makes sense if there are
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9954) * active CPUs remaining as possible busiest CPUs to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9955) * pull load from which are not contained within the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9956) * destination group that is receiving any migrated
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9957) * load.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9958) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9959) if (!cpumask_subset(cpus, env.dst_grpmask)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9960) env.loop = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9961) env.loop_break = sched_nr_migrate_break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9962) goto redo;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9963) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9964) goto out_all_pinned;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9965) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9966) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9967)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9968) if (!ld_moved) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9969) schedstat_inc(sd->lb_failed[idle]);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9970) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9971) * Increment the failure counter only on periodic balance.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9972) * We do not want newidle balance, which can be very
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9973) * frequent, pollute the failure counter causing
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9974) * excessive cache_hot migrations and active balances.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9975) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9976) if (idle != CPU_NEWLY_IDLE)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9977) sd->nr_balance_failed++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9978)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9979) if (need_active_balance(&env)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9980) unsigned long flags;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9981)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9982) raw_spin_lock_irqsave(&busiest->lock, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9983)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9984) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9985) * Don't kick the active_load_balance_cpu_stop,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9986) * if the curr task on busiest CPU can't be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9987) * moved to this_cpu:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9988) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9989) if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9990) raw_spin_unlock_irqrestore(&busiest->lock,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9991) flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9992) env.flags |= LBF_ALL_PINNED;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9993) goto out_one_pinned;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9994) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9995)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9996) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9997) * ->active_balance synchronizes accesses to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9998) * ->active_balance_work. Once set, it's cleared
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9999) * only after active load balance is finished.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10000) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10001) if (!busiest->active_balance) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10002) busiest->active_balance = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10003) busiest->push_cpu = this_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10004) active_balance = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10005) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10006) raw_spin_unlock_irqrestore(&busiest->lock, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10007)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10008) if (active_balance) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10009) stop_one_cpu_nowait(cpu_of(busiest),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10010) active_load_balance_cpu_stop, busiest,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10011) &busiest->active_balance_work);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10012) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10013)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10014) /* We've kicked active balancing, force task migration. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10015) sd->nr_balance_failed = sd->cache_nice_tries+1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10016) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10017) } else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10018) sd->nr_balance_failed = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10019)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10020) if (likely(!active_balance) || voluntary_active_balance(&env)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10021) /* We were unbalanced, so reset the balancing interval */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10022) sd->balance_interval = sd->min_interval;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10023) } else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10024) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10025) * If we've begun active balancing, start to back off. This
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10026) * case may not be covered by the all_pinned logic if there
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10027) * is only 1 task on the busy runqueue (because we don't call
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10028) * detach_tasks).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10029) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10030) if (sd->balance_interval < sd->max_interval)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10031) sd->balance_interval *= 2;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10032) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10033)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10034) goto out;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10035)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10036) out_balanced:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10037) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10038) * We reach balance although we may have faced some affinity
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10039) * constraints. Clear the imbalance flag only if other tasks got
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10040) * a chance to move and fix the imbalance.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10041) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10042) if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10043) int *group_imbalance = &sd_parent->groups->sgc->imbalance;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10044)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10045) if (*group_imbalance)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10046) *group_imbalance = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10047) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10048)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10049) out_all_pinned:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10050) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10051) * We reach balance because all tasks are pinned at this level so
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10052) * we can't migrate them. Let the imbalance flag set so parent level
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10053) * can try to migrate them.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10054) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10055) schedstat_inc(sd->lb_balanced[idle]);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10056)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10057) sd->nr_balance_failed = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10058)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10059) out_one_pinned:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10060) ld_moved = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10061)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10062) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10063) * newidle_balance() disregards balance intervals, so we could
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10064) * repeatedly reach this code, which would lead to balance_interval
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10065) * skyrocketting in a short amount of time. Skip the balance_interval
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10066) * increase logic to avoid that.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10067) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10068) if (env.idle == CPU_NEWLY_IDLE)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10069) goto out;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10070)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10071) /* tune up the balancing interval */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10072) if ((env.flags & LBF_ALL_PINNED &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10073) sd->balance_interval < MAX_PINNED_INTERVAL) ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10074) sd->balance_interval < sd->max_interval)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10075) sd->balance_interval *= 2;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10076) out:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10077) return ld_moved;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10078) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10079)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10080) static inline unsigned long
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10081) get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10082) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10083) unsigned long interval = sd->balance_interval;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10084)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10085) if (cpu_busy)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10086) interval *= sd->busy_factor;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10087)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10088) /* scale ms to jiffies */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10089) interval = msecs_to_jiffies(interval);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10090)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10091) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10092) * Reduce likelihood of busy balancing at higher domains racing with
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10093) * balancing at lower domains by preventing their balancing periods
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10094) * from being multiples of each other.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10095) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10096) if (cpu_busy)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10097) interval -= 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10098)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10099) interval = clamp(interval, 1UL, max_load_balance_interval);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10100)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10101) return interval;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10102) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10103)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10104) static inline void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10105) update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10106) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10107) unsigned long interval, next;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10108)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10109) /* used by idle balance, so cpu_busy = 0 */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10110) interval = get_sd_balance_interval(sd, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10111) next = sd->last_balance + interval;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10112)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10113) if (time_after(*next_balance, next))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10114) *next_balance = next;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10115) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10116)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10117) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10118) * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10119) * running tasks off the busiest CPU onto idle CPUs. It requires at
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10120) * least 1 task to be running on each physical CPU where possible, and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10121) * avoids physical / logical imbalances.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10122) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10123) static int active_load_balance_cpu_stop(void *data)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10124) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10125) struct rq *busiest_rq = data;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10126) int busiest_cpu = cpu_of(busiest_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10127) int target_cpu = busiest_rq->push_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10128) struct rq *target_rq = cpu_rq(target_cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10129) struct sched_domain *sd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10130) struct task_struct *p = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10131) struct rq_flags rf;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10132)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10133) rq_lock_irq(busiest_rq, &rf);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10134) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10135) * Between queueing the stop-work and running it is a hole in which
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10136) * CPUs can become inactive. We should not move tasks from or to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10137) * inactive CPUs.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10138) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10139) if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10140) goto out_unlock;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10141)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10142) /* Make sure the requested CPU hasn't gone down in the meantime: */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10143) if (unlikely(busiest_cpu != smp_processor_id() ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10144) !busiest_rq->active_balance))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10145) goto out_unlock;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10146)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10147) /* Is there any task to move? */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10148) if (busiest_rq->nr_running <= 1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10149) goto out_unlock;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10150)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10151) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10152) * This condition is "impossible", if it occurs
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10153) * we need to fix it. Originally reported by
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10154) * Bjorn Helgaas on a 128-CPU setup.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10155) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10156) BUG_ON(busiest_rq == target_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10157)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10158) /* Search for an sd spanning us and the target CPU. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10159) rcu_read_lock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10160) for_each_domain(target_cpu, sd) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10161) if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10162) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10163) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10164)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10165) if (likely(sd)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10166) struct lb_env env = {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10167) .sd = sd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10168) .dst_cpu = target_cpu,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10169) .dst_rq = target_rq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10170) .src_cpu = busiest_rq->cpu,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10171) .src_rq = busiest_rq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10172) .idle = CPU_IDLE,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10173) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10174) * can_migrate_task() doesn't need to compute new_dst_cpu
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10175) * for active balancing. Since we have CPU_IDLE, but no
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10176) * @dst_grpmask we need to make that test go away with lying
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10177) * about DST_PINNED.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10178) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10179) .flags = LBF_DST_PINNED,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10180) .src_rq_rf = &rf,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10181) };
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10182)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10183) schedstat_inc(sd->alb_count);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10184) update_rq_clock(busiest_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10185)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10186) p = detach_one_task(&env);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10187) if (p) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10188) schedstat_inc(sd->alb_pushed);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10189) /* Active balancing done, reset the failure counter. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10190) sd->nr_balance_failed = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10191) } else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10192) schedstat_inc(sd->alb_failed);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10193) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10194) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10195) rcu_read_unlock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10196) out_unlock:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10197) busiest_rq->active_balance = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10198) rq_unlock(busiest_rq, &rf);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10199)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10200) if (p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10201) attach_one_task(target_rq, p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10202)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10203) local_irq_enable();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10204)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10205) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10206) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10207)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10208) static DEFINE_SPINLOCK(balancing);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10209)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10210) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10211) * Scale the max load_balance interval with the number of CPUs in the system.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10212) * This trades load-balance latency on larger machines for less cross talk.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10213) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10214) void update_max_interval(void)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10215) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10216) max_load_balance_interval = HZ*num_active_cpus()/10;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10217) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10218)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10219) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10220) * It checks each scheduling domain to see if it is due to be balanced,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10221) * and initiates a balancing operation if so.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10222) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10223) * Balancing parameters are set up in init_sched_domains.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10224) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10225) static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10226) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10227) int continue_balancing = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10228) int cpu = rq->cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10229) int busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10230) unsigned long interval;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10231) struct sched_domain *sd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10232) /* Earliest time when we have to do rebalance again */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10233) unsigned long next_balance = jiffies + 60*HZ;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10234) int update_next_balance = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10235) int need_serialize, need_decay = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10236) u64 max_cost = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10237)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10238) trace_android_rvh_sched_rebalance_domains(rq, &continue_balancing);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10239) if (!continue_balancing)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10240) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10241)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10242) rcu_read_lock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10243) for_each_domain(cpu, sd) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10244) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10245) * Decay the newidle max times here because this is a regular
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10246) * visit to all the domains. Decay ~1% per second.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10247) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10248) if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10249) sd->max_newidle_lb_cost =
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10250) (sd->max_newidle_lb_cost * 253) / 256;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10251) sd->next_decay_max_lb_cost = jiffies + HZ;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10252) need_decay = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10253) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10254) max_cost += sd->max_newidle_lb_cost;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10255)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10256) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10257) * Stop the load balance at this level. There is another
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10258) * CPU in our sched group which is doing load balancing more
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10259) * actively.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10260) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10261) if (!continue_balancing) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10262) if (need_decay)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10263) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10264) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10265) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10266)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10267) interval = get_sd_balance_interval(sd, busy);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10268)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10269) need_serialize = sd->flags & SD_SERIALIZE;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10270) if (need_serialize) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10271) if (!spin_trylock(&balancing))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10272) goto out;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10273) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10274)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10275) if (time_after_eq(jiffies, sd->last_balance + interval)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10276) if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10277) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10278) * The LBF_DST_PINNED logic could have changed
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10279) * env->dst_cpu, so we can't know our idle
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10280) * state even if we migrated tasks. Update it.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10281) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10282) idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10283) busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10284) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10285) sd->last_balance = jiffies;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10286) interval = get_sd_balance_interval(sd, busy);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10287) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10288) if (need_serialize)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10289) spin_unlock(&balancing);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10290) out:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10291) if (time_after(next_balance, sd->last_balance + interval)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10292) next_balance = sd->last_balance + interval;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10293) update_next_balance = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10294) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10295) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10296) if (need_decay) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10297) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10298) * Ensure the rq-wide value also decays but keep it at a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10299) * reasonable floor to avoid funnies with rq->avg_idle.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10300) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10301) rq->max_idle_balance_cost =
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10302) max((u64)sysctl_sched_migration_cost, max_cost);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10303) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10304) rcu_read_unlock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10305)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10306) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10307) * next_balance will be updated only when there is a need.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10308) * When the cpu is attached to null domain for ex, it will not be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10309) * updated.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10310) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10311) if (likely(update_next_balance)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10312) rq->next_balance = next_balance;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10313)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10314) #ifdef CONFIG_NO_HZ_COMMON
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10315) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10316) * If this CPU has been elected to perform the nohz idle
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10317) * balance. Other idle CPUs have already rebalanced with
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10318) * nohz_idle_balance() and nohz.next_balance has been
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10319) * updated accordingly. This CPU is now running the idle load
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10320) * balance for itself and we need to update the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10321) * nohz.next_balance accordingly.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10322) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10323) if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10324) nohz.next_balance = rq->next_balance;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10325) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10326) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10327) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10328)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10329) static inline int on_null_domain(struct rq *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10330) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10331) return unlikely(!rcu_dereference_sched(rq->sd));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10332) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10333)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10334) #ifdef CONFIG_NO_HZ_COMMON
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10335) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10336) * idle load balancing details
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10337) * - When one of the busy CPUs notice that there may be an idle rebalancing
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10338) * needed, they will kick the idle load balancer, which then does idle
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10339) * load balancing for all the idle CPUs.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10340) * - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10341) * anywhere yet.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10342) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10343)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10344) static inline int find_new_ilb(void)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10345) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10346) int ilb = -1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10347)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10348) trace_android_rvh_find_new_ilb(nohz.idle_cpus_mask, &ilb);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10349) if (ilb >= 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10350) return ilb;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10351)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10352) for_each_cpu_and(ilb, nohz.idle_cpus_mask,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10353) housekeeping_cpumask(HK_FLAG_MISC)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10354) if (idle_cpu(ilb))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10355) return ilb;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10356) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10357)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10358) return nr_cpu_ids;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10359) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10360)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10361) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10362) * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10363) * idle CPU in the HK_FLAG_MISC housekeeping set (if there is one).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10364) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10365) static void kick_ilb(unsigned int flags)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10366) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10367) int ilb_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10368)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10369) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10370) * Increase nohz.next_balance only when if full ilb is triggered but
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10371) * not if we only update stats.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10372) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10373) if (flags & NOHZ_BALANCE_KICK)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10374) nohz.next_balance = jiffies+1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10375)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10376) ilb_cpu = find_new_ilb();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10377)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10378) if (ilb_cpu >= nr_cpu_ids)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10379) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10380)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10381) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10382) * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10383) * the first flag owns it; cleared by nohz_csd_func().
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10384) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10385) flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10386) if (flags & NOHZ_KICK_MASK)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10387) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10388)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10389) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10390) * This way we generate an IPI on the target CPU which
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10391) * is idle. And the softirq performing nohz idle load balance
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10392) * will be run before returning from the IPI.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10393) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10394) smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10395) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10396)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10397) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10398) * Current decision point for kicking the idle load balancer in the presence
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10399) * of idle CPUs in the system.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10400) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10401) static void nohz_balancer_kick(struct rq *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10402) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10403) unsigned long now = jiffies;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10404) struct sched_domain_shared *sds;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10405) struct sched_domain *sd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10406) int nr_busy, i, cpu = rq->cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10407) unsigned int flags = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10408) int done = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10409)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10410) if (unlikely(rq->idle_balance))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10411) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10412)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10413) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10414) * We may be recently in ticked or tickless idle mode. At the first
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10415) * busy tick after returning from idle, we will update the busy stats.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10416) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10417) nohz_balance_exit_idle(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10418)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10419) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10420) * None are in tickless mode and hence no need for NOHZ idle load
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10421) * balancing.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10422) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10423) if (likely(!atomic_read(&nohz.nr_cpus)))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10424) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10425)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10426) if (READ_ONCE(nohz.has_blocked) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10427) time_after(now, READ_ONCE(nohz.next_blocked)))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10428) flags = NOHZ_STATS_KICK;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10429)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10430) if (time_before(now, nohz.next_balance))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10431) goto out;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10432)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10433) trace_android_rvh_sched_nohz_balancer_kick(rq, &flags, &done);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10434) if (done)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10435) goto out;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10436)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10437) if (rq->nr_running >= 2) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10438) flags = NOHZ_KICK_MASK;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10439) goto out;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10440) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10441)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10442) rcu_read_lock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10443)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10444) sd = rcu_dereference(rq->sd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10445) if (sd) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10446) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10447) * If there's a CFS task and the current CPU has reduced
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10448) * capacity; kick the ILB to see if there's a better CPU to run
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10449) * on.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10450) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10451) if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10452) flags = NOHZ_KICK_MASK;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10453) goto unlock;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10454) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10455) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10456)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10457) sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10458) if (sd) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10459) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10460) * When ASYM_PACKING; see if there's a more preferred CPU
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10461) * currently idle; in which case, kick the ILB to move tasks
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10462) * around.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10463) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10464) for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10465) if (sched_asym_prefer(i, cpu)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10466) flags = NOHZ_KICK_MASK;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10467) goto unlock;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10468) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10469) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10470) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10471)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10472) sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10473) if (sd) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10474) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10475) * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10476) * to run the misfit task on.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10477) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10478) if (check_misfit_status(rq, sd)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10479) flags = NOHZ_KICK_MASK;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10480) goto unlock;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10481) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10482)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10483) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10484) * For asymmetric systems, we do not want to nicely balance
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10485) * cache use, instead we want to embrace asymmetry and only
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10486) * ensure tasks have enough CPU capacity.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10487) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10488) * Skip the LLC logic because it's not relevant in that case.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10489) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10490) goto unlock;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10491) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10492)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10493) sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10494) if (sds) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10495) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10496) * If there is an imbalance between LLC domains (IOW we could
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10497) * increase the overall cache use), we need some less-loaded LLC
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10498) * domain to pull some load. Likewise, we may need to spread
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10499) * load within the current LLC domain (e.g. packed SMT cores but
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10500) * other CPUs are idle). We can't really know from here how busy
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10501) * the others are - so just get a nohz balance going if it looks
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10502) * like this LLC domain has tasks we could move.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10503) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10504) nr_busy = atomic_read(&sds->nr_busy_cpus);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10505) if (nr_busy > 1) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10506) flags = NOHZ_KICK_MASK;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10507) goto unlock;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10508) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10509) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10510) unlock:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10511) rcu_read_unlock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10512) out:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10513) if (flags)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10514) kick_ilb(flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10515) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10516)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10517) static void set_cpu_sd_state_busy(int cpu)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10518) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10519) struct sched_domain *sd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10520)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10521) rcu_read_lock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10522) sd = rcu_dereference(per_cpu(sd_llc, cpu));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10523)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10524) if (!sd || !sd->nohz_idle)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10525) goto unlock;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10526) sd->nohz_idle = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10527)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10528) atomic_inc(&sd->shared->nr_busy_cpus);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10529) unlock:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10530) rcu_read_unlock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10531) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10532)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10533) void nohz_balance_exit_idle(struct rq *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10534) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10535) SCHED_WARN_ON(rq != this_rq());
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10536)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10537) if (likely(!rq->nohz_tick_stopped))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10538) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10539)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10540) rq->nohz_tick_stopped = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10541) cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10542) atomic_dec(&nohz.nr_cpus);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10543)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10544) set_cpu_sd_state_busy(rq->cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10545) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10546)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10547) static void set_cpu_sd_state_idle(int cpu)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10548) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10549) struct sched_domain *sd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10550)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10551) rcu_read_lock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10552) sd = rcu_dereference(per_cpu(sd_llc, cpu));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10553)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10554) if (!sd || sd->nohz_idle)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10555) goto unlock;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10556) sd->nohz_idle = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10557)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10558) atomic_dec(&sd->shared->nr_busy_cpus);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10559) unlock:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10560) rcu_read_unlock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10561) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10562)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10563) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10564) * This routine will record that the CPU is going idle with tick stopped.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10565) * This info will be used in performing idle load balancing in the future.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10566) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10567) void nohz_balance_enter_idle(int cpu)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10568) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10569) struct rq *rq = cpu_rq(cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10570)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10571) SCHED_WARN_ON(cpu != smp_processor_id());
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10572)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10573) if (!cpu_active(cpu)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10574) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10575) * A CPU can be paused while it is idle with it's tick
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10576) * stopped. nohz_balance_exit_idle() should be called
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10577) * from the local CPU, so it can't be called during
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10578) * pause. This results in paused CPU participating in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10579) * the nohz idle balance, which should be avoided.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10580) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10581) * When the paused CPU exits idle and enters again,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10582) * exempt the paused CPU from nohz_balance_exit_idle.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10583) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10584) nohz_balance_exit_idle(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10585) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10586) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10587)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10588) /* Spare idle load balancing on CPUs that don't want to be disturbed: */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10589) if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10590) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10591)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10592) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10593) * Can be set safely without rq->lock held
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10594) * If a clear happens, it will have evaluated last additions because
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10595) * rq->lock is held during the check and the clear
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10596) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10597) rq->has_blocked_load = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10598)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10599) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10600) * The tick is still stopped but load could have been added in the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10601) * meantime. We set the nohz.has_blocked flag to trig a check of the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10602) * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10603) * of nohz.has_blocked can only happen after checking the new load
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10604) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10605) if (rq->nohz_tick_stopped)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10606) goto out;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10607)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10608) /* If we're a completely isolated CPU, we don't play: */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10609) if (on_null_domain(rq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10610) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10611)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10612) rq->nohz_tick_stopped = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10613)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10614) cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10615) atomic_inc(&nohz.nr_cpus);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10616)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10617) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10618) * Ensures that if nohz_idle_balance() fails to observe our
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10619) * @idle_cpus_mask store, it must observe the @has_blocked
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10620) * store.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10621) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10622) smp_mb__after_atomic();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10623)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10624) set_cpu_sd_state_idle(cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10625)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10626) out:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10627) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10628) * Each time a cpu enter idle, we assume that it has blocked load and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10629) * enable the periodic update of the load of idle cpus
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10630) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10631) WRITE_ONCE(nohz.has_blocked, 1);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10632) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10633)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10634) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10635) * Internal function that runs load balance for all idle cpus. The load balance
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10636) * can be a simple update of blocked load or a complete load balance with
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10637) * tasks movement depending of flags.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10638) * The function returns false if the loop has stopped before running
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10639) * through all idle CPUs.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10640) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10641) static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10642) enum cpu_idle_type idle)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10643) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10644) /* Earliest time when we have to do rebalance again */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10645) unsigned long now = jiffies;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10646) unsigned long next_balance = now + 60*HZ;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10647) bool has_blocked_load = false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10648) int update_next_balance = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10649) int this_cpu = this_rq->cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10650) int balance_cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10651) int ret = false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10652) struct rq *rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10653)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10654) SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10655)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10656) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10657) * We assume there will be no idle load after this update and clear
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10658) * the has_blocked flag. If a cpu enters idle in the mean time, it will
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10659) * set the has_blocked flag and trig another update of idle load.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10660) * Because a cpu that becomes idle, is added to idle_cpus_mask before
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10661) * setting the flag, we are sure to not clear the state and not
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10662) * check the load of an idle cpu.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10663) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10664) WRITE_ONCE(nohz.has_blocked, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10665)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10666) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10667) * Ensures that if we miss the CPU, we must see the has_blocked
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10668) * store from nohz_balance_enter_idle().
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10669) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10670) smp_mb();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10671)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10672) for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10673) if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10674) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10675)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10676) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10677) * If this CPU gets work to do, stop the load balancing
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10678) * work being done for other CPUs. Next load
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10679) * balancing owner will pick it up.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10680) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10681) if (need_resched()) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10682) has_blocked_load = true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10683) goto abort;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10684) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10685)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10686) rq = cpu_rq(balance_cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10687)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10688) has_blocked_load |= update_nohz_stats(rq, true);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10689)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10690) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10691) * If time for next balance is due,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10692) * do the balance.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10693) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10694) if (time_after_eq(jiffies, rq->next_balance)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10695) struct rq_flags rf;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10696)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10697) rq_lock_irqsave(rq, &rf);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10698) update_rq_clock(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10699) rq_unlock_irqrestore(rq, &rf);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10700)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10701) if (flags & NOHZ_BALANCE_KICK)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10702) rebalance_domains(rq, CPU_IDLE);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10703) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10704)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10705) if (time_after(next_balance, rq->next_balance)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10706) next_balance = rq->next_balance;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10707) update_next_balance = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10708) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10709) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10710)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10711) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10712) * next_balance will be updated only when there is a need.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10713) * When the CPU is attached to null domain for ex, it will not be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10714) * updated.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10715) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10716) if (likely(update_next_balance))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10717) nohz.next_balance = next_balance;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10718)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10719) /* Newly idle CPU doesn't need an update */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10720) if (idle != CPU_NEWLY_IDLE) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10721) update_blocked_averages(this_cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10722) has_blocked_load |= this_rq->has_blocked_load;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10723) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10724)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10725) if (flags & NOHZ_BALANCE_KICK)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10726) rebalance_domains(this_rq, CPU_IDLE);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10727)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10728) WRITE_ONCE(nohz.next_blocked,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10729) now + msecs_to_jiffies(LOAD_AVG_PERIOD));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10730)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10731) /* The full idle balance loop has been done */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10732) ret = true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10733)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10734) abort:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10735) /* There is still blocked load, enable periodic update */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10736) if (has_blocked_load)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10737) WRITE_ONCE(nohz.has_blocked, 1);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10738)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10739) return ret;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10740) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10741)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10742) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10743) * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10744) * rebalancing for all the cpus for whom scheduler ticks are stopped.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10745) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10746) static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10747) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10748) unsigned int flags = this_rq->nohz_idle_balance;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10749)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10750) if (!flags)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10751) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10752)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10753) this_rq->nohz_idle_balance = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10754)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10755) if (idle != CPU_IDLE)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10756) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10757)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10758) _nohz_idle_balance(this_rq, flags, idle);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10759)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10760) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10761) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10762)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10763) static void nohz_newidle_balance(struct rq *this_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10764) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10765) int this_cpu = this_rq->cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10766)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10767) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10768) * This CPU doesn't want to be disturbed by scheduler
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10769) * housekeeping
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10770) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10771) if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10772) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10773)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10774) /* Will wake up very soon. No time for doing anything else*/
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10775) if (this_rq->avg_idle < sysctl_sched_migration_cost)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10776) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10777)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10778) /* Don't need to update blocked load of idle CPUs*/
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10779) if (!READ_ONCE(nohz.has_blocked) ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10780) time_before(jiffies, READ_ONCE(nohz.next_blocked)))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10781) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10782)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10783) raw_spin_unlock(&this_rq->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10784) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10785) * This CPU is going to be idle and blocked load of idle CPUs
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10786) * need to be updated. Run the ilb locally as it is a good
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10787) * candidate for ilb instead of waking up another idle CPU.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10788) * Kick an normal ilb if we failed to do the update.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10789) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10790) if (!_nohz_idle_balance(this_rq, NOHZ_STATS_KICK, CPU_NEWLY_IDLE))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10791) kick_ilb(NOHZ_STATS_KICK);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10792) raw_spin_lock(&this_rq->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10793) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10794)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10795) #else /* !CONFIG_NO_HZ_COMMON */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10796) static inline void nohz_balancer_kick(struct rq *rq) { }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10797)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10798) static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10799) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10800) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10801) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10802)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10803) static inline void nohz_newidle_balance(struct rq *this_rq) { }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10804) #endif /* CONFIG_NO_HZ_COMMON */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10805)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10806) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10807) * idle_balance is called by schedule() if this_cpu is about to become
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10808) * idle. Attempts to pull tasks from other CPUs.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10809) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10810) * Returns:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10811) * < 0 - we released the lock and there are !fair tasks present
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10812) * 0 - failed, no new tasks
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10813) * > 0 - success, new (fair) tasks present
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10814) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10815) static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10816) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10817) unsigned long next_balance = jiffies + HZ;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10818) int this_cpu = this_rq->cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10819) struct sched_domain *sd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10820) int pulled_task = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10821) u64 curr_cost = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10822) int done = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10823)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10824) trace_android_rvh_sched_newidle_balance(this_rq, rf, &pulled_task, &done);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10825) if (done)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10826) return pulled_task;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10827)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10828) update_misfit_status(NULL, this_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10829) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10830) * We must set idle_stamp _before_ calling idle_balance(), such that we
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10831) * measure the duration of idle_balance() as idle time.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10832) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10833) this_rq->idle_stamp = rq_clock(this_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10834)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10835) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10836) * Do not pull tasks towards !active CPUs...
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10837) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10838) if (!cpu_active(this_cpu))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10839) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10840)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10841) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10842) * This is OK, because current is on_cpu, which avoids it being picked
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10843) * for load-balance and preemption/IRQs are still disabled avoiding
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10844) * further scheduler activity on it and we're being very careful to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10845) * re-start the picking loop.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10846) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10847) rq_unpin_lock(this_rq, rf);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10848)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10849) if (this_rq->avg_idle < sysctl_sched_migration_cost ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10850) !READ_ONCE(this_rq->rd->overload)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10851)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10852) rcu_read_lock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10853) sd = rcu_dereference_check_sched_domain(this_rq->sd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10854) if (sd)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10855) update_next_balance(sd, &next_balance);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10856) rcu_read_unlock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10857)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10858) nohz_newidle_balance(this_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10859)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10860) goto out;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10861) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10862)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10863) raw_spin_unlock(&this_rq->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10864)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10865) update_blocked_averages(this_cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10866) rcu_read_lock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10867) for_each_domain(this_cpu, sd) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10868) int continue_balancing = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10869) u64 t0, domain_cost;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10870)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10871) if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10872) update_next_balance(sd, &next_balance);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10873) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10874) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10875)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10876) if (sd->flags & SD_BALANCE_NEWIDLE) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10877) t0 = sched_clock_cpu(this_cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10878)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10879) pulled_task = load_balance(this_cpu, this_rq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10880) sd, CPU_NEWLY_IDLE,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10881) &continue_balancing);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10882)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10883) domain_cost = sched_clock_cpu(this_cpu) - t0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10884) if (domain_cost > sd->max_newidle_lb_cost)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10885) sd->max_newidle_lb_cost = domain_cost;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10886)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10887) curr_cost += domain_cost;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10888) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10889)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10890) update_next_balance(sd, &next_balance);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10891)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10892) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10893) * Stop searching for tasks to pull if there are
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10894) * now runnable tasks on this rq.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10895) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10896) if (pulled_task || this_rq->nr_running > 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10897) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10898) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10899) rcu_read_unlock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10900)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10901) raw_spin_lock(&this_rq->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10902)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10903) if (curr_cost > this_rq->max_idle_balance_cost)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10904) this_rq->max_idle_balance_cost = curr_cost;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10905)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10906) out:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10907) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10908) * While browsing the domains, we released the rq lock, a task could
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10909) * have been enqueued in the meantime. Since we're not going idle,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10910) * pretend we pulled a task.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10911) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10912) if (this_rq->cfs.h_nr_running && !pulled_task)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10913) pulled_task = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10914)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10915) /* Move the next balance forward */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10916) if (time_after(this_rq->next_balance, next_balance))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10917) this_rq->next_balance = next_balance;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10918)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10919) /* Is there a task of a high priority class? */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10920) if (this_rq->nr_running != this_rq->cfs.h_nr_running)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10921) pulled_task = -1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10922)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10923) if (pulled_task)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10924) this_rq->idle_stamp = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10925)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10926) rq_repin_lock(this_rq, rf);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10927)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10928) return pulled_task;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10929) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10930)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10931) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10932) * run_rebalance_domains is triggered when needed from the scheduler tick.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10933) * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10934) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10935) static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10936) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10937) struct rq *this_rq = this_rq();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10938) enum cpu_idle_type idle = this_rq->idle_balance ?
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10939) CPU_IDLE : CPU_NOT_IDLE;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10940)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10941) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10942) * If this CPU has a pending nohz_balance_kick, then do the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10943) * balancing on behalf of the other idle CPUs whose ticks are
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10944) * stopped. Do nohz_idle_balance *before* rebalance_domains to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10945) * give the idle CPUs a chance to load balance. Else we may
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10946) * load balance only within the local sched_domain hierarchy
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10947) * and abort nohz_idle_balance altogether if we pull some load.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10948) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10949) if (nohz_idle_balance(this_rq, idle))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10950) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10951)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10952) /* normal load balance */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10953) update_blocked_averages(this_rq->cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10954) rebalance_domains(this_rq, idle);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10955) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10956)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10957) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10958) * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10959) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10960) void trigger_load_balance(struct rq *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10961) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10962) /* Don't need to rebalance while attached to NULL domain */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10963) if (unlikely(on_null_domain(rq)))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10964) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10965)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10966) if (time_after_eq(jiffies, rq->next_balance))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10967) raise_softirq(SCHED_SOFTIRQ);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10968)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10969) nohz_balancer_kick(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10970) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10971)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10972) static void rq_online_fair(struct rq *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10973) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10974) update_sysctl();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10975)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10976) update_runtime_enabled(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10977) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10978)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10979) static void rq_offline_fair(struct rq *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10980) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10981) update_sysctl();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10982)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10983) /* Ensure any throttled groups are reachable by pick_next_task */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10984) unthrottle_offline_cfs_rqs(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10985) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10986)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10987) #endif /* CONFIG_SMP */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10988)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10989) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10990) * scheduler tick hitting a task of our scheduling class.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10991) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10992) * NOTE: This function can be called remotely by the tick offload that
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10993) * goes along full dynticks. Therefore no local assumption can be made
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10994) * and everything must be accessed through the @rq and @curr passed in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10995) * parameters.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10996) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10997) static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10998) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10999) struct cfs_rq *cfs_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11000) struct sched_entity *se = &curr->se;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11001)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11002) for_each_sched_entity(se) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11003) cfs_rq = cfs_rq_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11004) entity_tick(cfs_rq, se, queued);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11005) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11006)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11007) if (static_branch_unlikely(&sched_numa_balancing))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11008) task_tick_numa(rq, curr);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11009)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11010) update_misfit_status(curr, rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11011) update_overutilized_status(task_rq(curr));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11012) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11013)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11014) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11015) * called on fork with the child task as argument from the parent's context
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11016) * - child not yet on the tasklist
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11017) * - preemption disabled
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11018) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11019) static void task_fork_fair(struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11020) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11021) struct cfs_rq *cfs_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11022) struct sched_entity *se = &p->se, *curr;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11023) struct rq *rq = this_rq();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11024) struct rq_flags rf;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11025)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11026) rq_lock(rq, &rf);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11027) update_rq_clock(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11028)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11029) cfs_rq = task_cfs_rq(current);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11030) curr = cfs_rq->curr;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11031) if (curr) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11032) update_curr(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11033) se->vruntime = curr->vruntime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11034) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11035) place_entity(cfs_rq, se, 1);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11036)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11037) if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11038) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11039) * Upon rescheduling, sched_class::put_prev_task() will place
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11040) * 'current' within the tree based on its new key value.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11041) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11042) swap(curr->vruntime, se->vruntime);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11043) resched_curr(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11044) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11045)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11046) se->vruntime -= cfs_rq->min_vruntime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11047) rq_unlock(rq, &rf);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11048) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11049)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11050) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11051) * Priority of the task has changed. Check to see if we preempt
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11052) * the current task.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11053) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11054) static void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11055) prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11056) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11057) if (!task_on_rq_queued(p))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11058) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11059)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11060) if (rq->cfs.nr_running == 1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11061) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11062)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11063) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11064) * Reschedule if we are currently running on this runqueue and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11065) * our priority decreased, or if we are not currently running on
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11066) * this runqueue and our priority is higher than the current's
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11067) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11068) if (rq->curr == p) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11069) if (p->prio > oldprio)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11070) resched_curr(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11071) } else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11072) check_preempt_curr(rq, p, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11073) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11074)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11075) static inline bool vruntime_normalized(struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11076) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11077) struct sched_entity *se = &p->se;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11078)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11079) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11080) * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11081) * the dequeue_entity(.flags=0) will already have normalized the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11082) * vruntime.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11083) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11084) if (p->on_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11085) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11086)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11087) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11088) * When !on_rq, vruntime of the task has usually NOT been normalized.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11089) * But there are some cases where it has already been normalized:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11090) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11091) * - A forked child which is waiting for being woken up by
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11092) * wake_up_new_task().
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11093) * - A task which has been woken up by try_to_wake_up() and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11094) * waiting for actually being woken up by sched_ttwu_pending().
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11095) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11096) if (!se->sum_exec_runtime ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11097) (p->state == TASK_WAKING && p->sched_remote_wakeup))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11098) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11099)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11100) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11101) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11102)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11103) #ifdef CONFIG_FAIR_GROUP_SCHED
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11104) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11105) * Propagate the changes of the sched_entity across the tg tree to make it
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11106) * visible to the root
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11107) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11108) static void propagate_entity_cfs_rq(struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11109) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11110) struct cfs_rq *cfs_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11111)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11112) list_add_leaf_cfs_rq(cfs_rq_of(se));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11113)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11114) /* Start to propagate at parent */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11115) se = se->parent;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11116)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11117) for_each_sched_entity(se) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11118) cfs_rq = cfs_rq_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11119)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11120) if (!cfs_rq_throttled(cfs_rq)){
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11121) update_load_avg(cfs_rq, se, UPDATE_TG);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11122) list_add_leaf_cfs_rq(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11123) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11124) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11125)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11126) if (list_add_leaf_cfs_rq(cfs_rq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11127) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11128) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11129) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11130) #else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11131) static void propagate_entity_cfs_rq(struct sched_entity *se) { }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11132) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11133)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11134) static void detach_entity_cfs_rq(struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11135) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11136) struct cfs_rq *cfs_rq = cfs_rq_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11137)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11138) /* Catch up with the cfs_rq and remove our load when we leave */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11139) update_load_avg(cfs_rq, se, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11140) detach_entity_load_avg(cfs_rq, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11141) update_tg_load_avg(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11142) propagate_entity_cfs_rq(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11143) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11144)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11145) static void attach_entity_cfs_rq(struct sched_entity *se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11146) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11147) struct cfs_rq *cfs_rq = cfs_rq_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11148)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11149) #ifdef CONFIG_FAIR_GROUP_SCHED
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11150) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11151) * Since the real-depth could have been changed (only FAIR
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11152) * class maintain depth value), reset depth properly.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11153) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11154) se->depth = se->parent ? se->parent->depth + 1 : 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11155) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11156)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11157) /* Synchronize entity with its cfs_rq */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11158) update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11159) attach_entity_load_avg(cfs_rq, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11160) update_tg_load_avg(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11161) propagate_entity_cfs_rq(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11162) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11163)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11164) static void detach_task_cfs_rq(struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11165) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11166) struct sched_entity *se = &p->se;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11167) struct cfs_rq *cfs_rq = cfs_rq_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11168)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11169) if (!vruntime_normalized(p)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11170) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11171) * Fix up our vruntime so that the current sleep doesn't
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11172) * cause 'unlimited' sleep bonus.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11173) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11174) place_entity(cfs_rq, se, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11175) se->vruntime -= cfs_rq->min_vruntime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11176) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11177)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11178) detach_entity_cfs_rq(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11179) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11180)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11181) static void attach_task_cfs_rq(struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11182) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11183) struct sched_entity *se = &p->se;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11184) struct cfs_rq *cfs_rq = cfs_rq_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11185)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11186) attach_entity_cfs_rq(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11187)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11188) if (!vruntime_normalized(p))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11189) se->vruntime += cfs_rq->min_vruntime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11190) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11191)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11192) static void switched_from_fair(struct rq *rq, struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11193) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11194) detach_task_cfs_rq(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11195) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11196)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11197) static void switched_to_fair(struct rq *rq, struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11198) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11199) attach_task_cfs_rq(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11200)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11201) if (task_on_rq_queued(p)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11202) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11203) * We were most likely switched from sched_rt, so
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11204) * kick off the schedule if running, otherwise just see
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11205) * if we can still preempt the current task.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11206) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11207) if (rq->curr == p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11208) resched_curr(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11209) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11210) check_preempt_curr(rq, p, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11211) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11212) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11213)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11214) /* Account for a task changing its policy or group.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11215) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11216) * This routine is mostly called to set cfs_rq->curr field when a task
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11217) * migrates between groups/classes.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11218) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11219) static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11220) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11221) struct sched_entity *se = &p->se;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11222)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11223) #ifdef CONFIG_SMP
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11224) if (task_on_rq_queued(p)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11225) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11226) * Move the next running task to the front of the list, so our
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11227) * cfs_tasks list becomes MRU one.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11228) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11229) list_move(&se->group_node, &rq->cfs_tasks);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11230) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11231) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11232)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11233) for_each_sched_entity(se) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11234) struct cfs_rq *cfs_rq = cfs_rq_of(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11235)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11236) set_next_entity(cfs_rq, se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11237) /* ensure bandwidth has been allocated on our new cfs_rq */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11238) account_cfs_rq_runtime(cfs_rq, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11239) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11240) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11241)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11242) void init_cfs_rq(struct cfs_rq *cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11243) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11244) cfs_rq->tasks_timeline = RB_ROOT_CACHED;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11245) cfs_rq->min_vruntime = (u64)(-(1LL << 20));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11246) #ifndef CONFIG_64BIT
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11247) cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11248) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11249) #ifdef CONFIG_SMP
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11250) raw_spin_lock_init(&cfs_rq->removed.lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11251) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11252) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11253)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11254) #ifdef CONFIG_FAIR_GROUP_SCHED
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11255) static void task_set_group_fair(struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11256) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11257) struct sched_entity *se = &p->se;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11258)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11259) set_task_rq(p, task_cpu(p));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11260) se->depth = se->parent ? se->parent->depth + 1 : 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11261) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11262)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11263) static void task_move_group_fair(struct task_struct *p)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11264) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11265) detach_task_cfs_rq(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11266) set_task_rq(p, task_cpu(p));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11267)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11268) #ifdef CONFIG_SMP
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11269) /* Tell se's cfs_rq has been changed -- migrated */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11270) p->se.avg.last_update_time = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11271) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11272) attach_task_cfs_rq(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11273) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11274)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11275) static void task_change_group_fair(struct task_struct *p, int type)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11276) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11277) switch (type) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11278) case TASK_SET_GROUP:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11279) task_set_group_fair(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11280) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11281)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11282) case TASK_MOVE_GROUP:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11283) task_move_group_fair(p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11284) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11285) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11286) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11287)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11288) void free_fair_sched_group(struct task_group *tg)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11289) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11290) int i;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11291)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11292) destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11293)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11294) for_each_possible_cpu(i) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11295) if (tg->cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11296) kfree(tg->cfs_rq[i]);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11297) if (tg->se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11298) kfree(tg->se[i]);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11299) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11300)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11301) kfree(tg->cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11302) kfree(tg->se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11303) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11304)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11305) int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11306) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11307) struct sched_entity *se;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11308) struct cfs_rq *cfs_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11309) int i;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11310)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11311) tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11312) if (!tg->cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11313) goto err;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11314) tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11315) if (!tg->se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11316) goto err;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11317)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11318) tg->shares = NICE_0_LOAD;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11319)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11320) init_cfs_bandwidth(tg_cfs_bandwidth(tg));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11321)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11322) for_each_possible_cpu(i) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11323) cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11324) GFP_KERNEL, cpu_to_node(i));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11325) if (!cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11326) goto err;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11327)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11328) se = kzalloc_node(sizeof(struct sched_entity),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11329) GFP_KERNEL, cpu_to_node(i));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11330) if (!se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11331) goto err_free_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11332)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11333) init_cfs_rq(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11334) init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11335) init_entity_runnable_average(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11336) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11337)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11338) return 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11339)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11340) err_free_rq:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11341) kfree(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11342) err:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11343) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11344) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11345)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11346) void online_fair_sched_group(struct task_group *tg)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11347) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11348) struct sched_entity *se;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11349) struct rq_flags rf;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11350) struct rq *rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11351) int i;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11352)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11353) for_each_possible_cpu(i) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11354) rq = cpu_rq(i);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11355) se = tg->se[i];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11356) rq_lock_irq(rq, &rf);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11357) update_rq_clock(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11358) attach_entity_cfs_rq(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11359) sync_throttle(tg, i);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11360) rq_unlock_irq(rq, &rf);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11361) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11362) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11363)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11364) void unregister_fair_sched_group(struct task_group *tg)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11365) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11366) unsigned long flags;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11367) struct rq *rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11368) int cpu;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11369)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11370) for_each_possible_cpu(cpu) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11371) if (tg->se[cpu])
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11372) remove_entity_load_avg(tg->se[cpu]);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11373)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11374) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11375) * Only empty task groups can be destroyed; so we can speculatively
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11376) * check on_list without danger of it being re-added.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11377) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11378) if (!tg->cfs_rq[cpu]->on_list)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11379) continue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11380)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11381) rq = cpu_rq(cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11382)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11383) raw_spin_lock_irqsave(&rq->lock, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11384) list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11385) raw_spin_unlock_irqrestore(&rq->lock, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11386) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11387) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11388)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11389) void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11390) struct sched_entity *se, int cpu,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11391) struct sched_entity *parent)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11392) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11393) struct rq *rq = cpu_rq(cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11394)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11395) cfs_rq->tg = tg;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11396) cfs_rq->rq = rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11397) init_cfs_rq_runtime(cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11398)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11399) tg->cfs_rq[cpu] = cfs_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11400) tg->se[cpu] = se;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11401)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11402) /* se could be NULL for root_task_group */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11403) if (!se)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11404) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11405)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11406) if (!parent) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11407) se->cfs_rq = &rq->cfs;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11408) se->depth = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11409) } else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11410) se->cfs_rq = parent->my_q;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11411) se->depth = parent->depth + 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11412) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11413)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11414) se->my_q = cfs_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11415) /* guarantee group entities always have weight */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11416) update_load_set(&se->load, NICE_0_LOAD);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11417) se->parent = parent;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11418) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11419)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11420) static DEFINE_MUTEX(shares_mutex);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11421)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11422) int sched_group_set_shares(struct task_group *tg, unsigned long shares)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11423) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11424) int i;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11425)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11426) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11427) * We can't change the weight of the root cgroup.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11428) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11429) if (!tg->se[0])
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11430) return -EINVAL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11431)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11432) shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11433)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11434) mutex_lock(&shares_mutex);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11435) if (tg->shares == shares)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11436) goto done;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11437)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11438) tg->shares = shares;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11439) for_each_possible_cpu(i) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11440) struct rq *rq = cpu_rq(i);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11441) struct sched_entity *se = tg->se[i];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11442) struct rq_flags rf;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11443)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11444) /* Propagate contribution to hierarchy */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11445) rq_lock_irqsave(rq, &rf);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11446) update_rq_clock(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11447) for_each_sched_entity(se) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11448) update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11449) update_cfs_group(se);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11450) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11451) rq_unlock_irqrestore(rq, &rf);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11452) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11453)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11454) done:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11455) mutex_unlock(&shares_mutex);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11456) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11457) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11458) #else /* CONFIG_FAIR_GROUP_SCHED */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11459)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11460) void free_fair_sched_group(struct task_group *tg) { }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11461)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11462) int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11463) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11464) return 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11465) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11466)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11467) void online_fair_sched_group(struct task_group *tg) { }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11468)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11469) void unregister_fair_sched_group(struct task_group *tg) { }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11470)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11471) #endif /* CONFIG_FAIR_GROUP_SCHED */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11472)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11473)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11474) static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11475) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11476) struct sched_entity *se = &task->se;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11477) unsigned int rr_interval = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11478)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11479) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11480) * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11481) * idle runqueue:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11482) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11483) if (rq->cfs.load.weight)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11484) rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11485)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11486) return rr_interval;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11487) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11488)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11489) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11490) * All the scheduling class methods:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11491) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11492) const struct sched_class fair_sched_class
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11493) __section("__fair_sched_class") = {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11494) .enqueue_task = enqueue_task_fair,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11495) .dequeue_task = dequeue_task_fair,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11496) .yield_task = yield_task_fair,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11497) .yield_to_task = yield_to_task_fair,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11498)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11499) .check_preempt_curr = check_preempt_wakeup,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11500)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11501) .pick_next_task = __pick_next_task_fair,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11502) .put_prev_task = put_prev_task_fair,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11503) .set_next_task = set_next_task_fair,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11504)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11505) #ifdef CONFIG_SMP
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11506) .balance = balance_fair,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11507) .select_task_rq = select_task_rq_fair,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11508) .migrate_task_rq = migrate_task_rq_fair,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11509)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11510) .rq_online = rq_online_fair,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11511) .rq_offline = rq_offline_fair,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11512)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11513) .task_dead = task_dead_fair,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11514) .set_cpus_allowed = set_cpus_allowed_common,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11515) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11516)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11517) .task_tick = task_tick_fair,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11518) .task_fork = task_fork_fair,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11519)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11520) .prio_changed = prio_changed_fair,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11521) .switched_from = switched_from_fair,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11522) .switched_to = switched_to_fair,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11523)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11524) .get_rr_interval = get_rr_interval_fair,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11525)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11526) .update_curr = update_curr_fair,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11527)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11528) #ifdef CONFIG_FAIR_GROUP_SCHED
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11529) .task_change_group = task_change_group_fair,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11530) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11531)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11532) #ifdef CONFIG_UCLAMP_TASK
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11533) .uclamp_enabled = 1,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11534) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11535) };
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11536)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11537) #ifdef CONFIG_SCHED_DEBUG
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11538) void print_cfs_stats(struct seq_file *m, int cpu)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11539) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11540) struct cfs_rq *cfs_rq, *pos;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11541)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11542) rcu_read_lock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11543) for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11544) print_cfs_rq(m, cpu, cfs_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11545) rcu_read_unlock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11546) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11547)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11548) #ifdef CONFIG_NUMA_BALANCING
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11549) void show_numa_stats(struct task_struct *p, struct seq_file *m)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11550) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11551) int node;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11552) unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11553) struct numa_group *ng;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11554)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11555) rcu_read_lock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11556) ng = rcu_dereference(p->numa_group);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11557) for_each_online_node(node) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11558) if (p->numa_faults) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11559) tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11560) tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11561) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11562) if (ng) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11563) gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11564) gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11565) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11566) print_numa_stats(m, node, tsf, tpf, gsf, gpf);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11567) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11568) rcu_read_unlock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11569) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11570) #endif /* CONFIG_NUMA_BALANCING */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11571) #endif /* CONFIG_SCHED_DEBUG */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11572)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11573) __init void init_sched_fair_class(void)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11574) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11575) #ifdef CONFIG_SMP
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11576) open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11577)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11578) #ifdef CONFIG_NO_HZ_COMMON
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11579) nohz.next_balance = jiffies;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11580) nohz.next_blocked = jiffies;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11581) zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11582) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11583) #endif /* SMP */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11584)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11585) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11586)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11587) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11588) * Helper functions to facilitate extracting info from tracepoints.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11589) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11590)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11591) const struct sched_avg *sched_trace_cfs_rq_avg(struct cfs_rq *cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11592) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11593) #ifdef CONFIG_SMP
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11594) return cfs_rq ? &cfs_rq->avg : NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11595) #else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11596) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11597) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11598) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11599) EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_avg);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11600)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11601) char *sched_trace_cfs_rq_path(struct cfs_rq *cfs_rq, char *str, int len)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11602) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11603) if (!cfs_rq) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11604) if (str)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11605) strlcpy(str, "(null)", len);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11606) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11607) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11608) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11609)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11610) cfs_rq_tg_path(cfs_rq, str, len);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11611) return str;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11612) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11613) EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_path);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11614)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11615) int sched_trace_cfs_rq_cpu(struct cfs_rq *cfs_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11616) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11617) return cfs_rq ? cpu_of(rq_of(cfs_rq)) : -1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11618) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11619) EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11620)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11621) const struct sched_avg *sched_trace_rq_avg_rt(struct rq *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11622) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11623) #ifdef CONFIG_SMP
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11624) return rq ? &rq->avg_rt : NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11625) #else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11626) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11627) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11628) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11629) EXPORT_SYMBOL_GPL(sched_trace_rq_avg_rt);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11630)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11631) const struct sched_avg *sched_trace_rq_avg_dl(struct rq *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11632) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11633) #ifdef CONFIG_SMP
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11634) return rq ? &rq->avg_dl : NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11635) #else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11636) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11637) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11638) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11639) EXPORT_SYMBOL_GPL(sched_trace_rq_avg_dl);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11640)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11641) const struct sched_avg *sched_trace_rq_avg_irq(struct rq *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11642) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11643) #if defined(CONFIG_SMP) && defined(CONFIG_HAVE_SCHED_AVG_IRQ)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11644) return rq ? &rq->avg_irq : NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11645) #else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11646) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11647) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11648) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11649) EXPORT_SYMBOL_GPL(sched_trace_rq_avg_irq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11650)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11651) int sched_trace_rq_cpu(struct rq *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11652) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11653) return rq ? cpu_of(rq) : -1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11654) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11655) EXPORT_SYMBOL_GPL(sched_trace_rq_cpu);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11656)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11657) int sched_trace_rq_cpu_capacity(struct rq *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11658) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11659) return rq ?
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11660) #ifdef CONFIG_SMP
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11661) rq->cpu_capacity
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11662) #else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11663) SCHED_CAPACITY_SCALE
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11664) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11665) : -1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11666) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11667) EXPORT_SYMBOL_GPL(sched_trace_rq_cpu_capacity);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11668)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11669) const struct cpumask *sched_trace_rd_span(struct root_domain *rd)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11670) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11671) #ifdef CONFIG_SMP
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11672) return rd ? rd->span : NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11673) #else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11674) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11675) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11676) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11677) EXPORT_SYMBOL_GPL(sched_trace_rd_span);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11678)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11679) int sched_trace_rq_nr_running(struct rq *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11680) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11681) return rq ? rq->nr_running : -1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11682) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11683) EXPORT_SYMBOL_GPL(sched_trace_rq_nr_running);
Orange Pi5 kernel
Deprecated Linux kernel 5.10.110 for OrangePi 5/5B/5+ boards
3 Commits
0 Branches
0 Tags