^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1) // SPDX-License-Identifier: GPL-2.0-or-later
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3) * Budget Fair Queueing (BFQ) I/O scheduler.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5) * Based on ideas and code from CFQ:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6) * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 7) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 8) * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 9) * Paolo Valente <paolo.valente@unimore.it>
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 10) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 11) * Copyright (C) 2010 Paolo Valente <paolo.valente@unimore.it>
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 12) * Arianna Avanzini <avanzini@google.com>
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 13) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 14) * Copyright (C) 2017 Paolo Valente <paolo.valente@linaro.org>
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 15) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 16) * BFQ is a proportional-share I/O scheduler, with some extra
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 17) * low-latency capabilities. BFQ also supports full hierarchical
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 18) * scheduling through cgroups. Next paragraphs provide an introduction
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 19) * on BFQ inner workings. Details on BFQ benefits, usage and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 20) * limitations can be found in Documentation/block/bfq-iosched.rst.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 21) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 22) * BFQ is a proportional-share storage-I/O scheduling algorithm based
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 23) * on the slice-by-slice service scheme of CFQ. But BFQ assigns
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 24) * budgets, measured in number of sectors, to processes instead of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 25) * time slices. The device is not granted to the in-service process
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 26) * for a given time slice, but until it has exhausted its assigned
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 27) * budget. This change from the time to the service domain enables BFQ
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 28) * to distribute the device throughput among processes as desired,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 29) * without any distortion due to throughput fluctuations, or to device
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 30) * internal queueing. BFQ uses an ad hoc internal scheduler, called
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 31) * B-WF2Q+, to schedule processes according to their budgets. More
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 32) * precisely, BFQ schedules queues associated with processes. Each
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 33) * process/queue is assigned a user-configurable weight, and B-WF2Q+
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 34) * guarantees that each queue receives a fraction of the throughput
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 35) * proportional to its weight. Thanks to the accurate policy of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 36) * B-WF2Q+, BFQ can afford to assign high budgets to I/O-bound
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 37) * processes issuing sequential requests (to boost the throughput),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 38) * and yet guarantee a low latency to interactive and soft real-time
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 39) * applications.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 40) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 41) * In particular, to provide these low-latency guarantees, BFQ
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 42) * explicitly privileges the I/O of two classes of time-sensitive
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 43) * applications: interactive and soft real-time. In more detail, BFQ
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 44) * behaves this way if the low_latency parameter is set (default
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 45) * configuration). This feature enables BFQ to provide applications in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 46) * these classes with a very low latency.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 47) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 48) * To implement this feature, BFQ constantly tries to detect whether
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 49) * the I/O requests in a bfq_queue come from an interactive or a soft
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 50) * real-time application. For brevity, in these cases, the queue is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 51) * said to be interactive or soft real-time. In both cases, BFQ
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 52) * privileges the service of the queue, over that of non-interactive
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 53) * and non-soft-real-time queues. This privileging is performed,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 54) * mainly, by raising the weight of the queue. So, for brevity, we
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 55) * call just weight-raising periods the time periods during which a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 56) * queue is privileged, because deemed interactive or soft real-time.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 57) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 58) * The detection of soft real-time queues/applications is described in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 59) * detail in the comments on the function
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 60) * bfq_bfqq_softrt_next_start. On the other hand, the detection of an
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 61) * interactive queue works as follows: a queue is deemed interactive
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 62) * if it is constantly non empty only for a limited time interval,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 63) * after which it does become empty. The queue may be deemed
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 64) * interactive again (for a limited time), if it restarts being
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 65) * constantly non empty, provided that this happens only after the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 66) * queue has remained empty for a given minimum idle time.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 67) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 68) * By default, BFQ computes automatically the above maximum time
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 69) * interval, i.e., the time interval after which a constantly
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 70) * non-empty queue stops being deemed interactive. Since a queue is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 71) * weight-raised while it is deemed interactive, this maximum time
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 72) * interval happens to coincide with the (maximum) duration of the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 73) * weight-raising for interactive queues.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 74) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 75) * Finally, BFQ also features additional heuristics for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 76) * preserving both a low latency and a high throughput on NCQ-capable,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 77) * rotational or flash-based devices, and to get the job done quickly
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 78) * for applications consisting in many I/O-bound processes.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 79) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 80) * NOTE: if the main or only goal, with a given device, is to achieve
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 81) * the maximum-possible throughput at all times, then do switch off
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 82) * all low-latency heuristics for that device, by setting low_latency
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 83) * to 0.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 84) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 85) * BFQ is described in [1], where also a reference to the initial,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 86) * more theoretical paper on BFQ can be found. The interested reader
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 87) * can find in the latter paper full details on the main algorithm, as
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 88) * well as formulas of the guarantees and formal proofs of all the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 89) * properties. With respect to the version of BFQ presented in these
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 90) * papers, this implementation adds a few more heuristics, such as the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 91) * ones that guarantee a low latency to interactive and soft real-time
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 92) * applications, and a hierarchical extension based on H-WF2Q+.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 93) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 94) * B-WF2Q+ is based on WF2Q+, which is described in [2], together with
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 95) * H-WF2Q+, while the augmented tree used here to implement B-WF2Q+
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 96) * with O(log N) complexity derives from the one introduced with EEVDF
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 97) * in [3].
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 98) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 99) * [1] P. Valente, A. Avanzini, "Evolution of the BFQ Storage I/O
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 100) * Scheduler", Proceedings of the First Workshop on Mobile System
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 101) * Technologies (MST-2015), May 2015.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 102) * http://algogroup.unimore.it/people/paolo/disk_sched/mst-2015.pdf
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 103) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 104) * [2] Jon C.R. Bennett and H. Zhang, "Hierarchical Packet Fair Queueing
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 105) * Algorithms", IEEE/ACM Transactions on Networking, 5(5):675-689,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 106) * Oct 1997.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 107) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 108) * http://www.cs.cmu.edu/~hzhang/papers/TON-97-Oct.ps.gz
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 109) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 110) * [3] I. Stoica and H. Abdel-Wahab, "Earliest Eligible Virtual Deadline
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 111) * First: A Flexible and Accurate Mechanism for Proportional Share
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 112) * Resource Allocation", technical report.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 113) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 114) * http://www.cs.berkeley.edu/~istoica/papers/eevdf-tr-95.pdf
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 115) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 116) #include <linux/module.h>
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 117) #include <linux/slab.h>
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 118) #include <linux/blkdev.h>
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 119) #include <linux/cgroup.h>
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 120) #include <linux/elevator.h>
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 121) #include <linux/ktime.h>
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 122) #include <linux/rbtree.h>
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 123) #include <linux/ioprio.h>
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 124) #include <linux/sbitmap.h>
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 125) #include <linux/delay.h>
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 126) #include <linux/backing-dev.h>
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 127)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 128) #include "blk.h"
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 129) #include "blk-mq.h"
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 130) #include "blk-mq-tag.h"
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 131) #include "blk-mq-sched.h"
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 132) #include "bfq-iosched.h"
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 133) #include "blk-wbt.h"
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 134)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 135) #define BFQ_BFQQ_FNS(name) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 136) void bfq_mark_bfqq_##name(struct bfq_queue *bfqq) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 137) { \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 138) __set_bit(BFQQF_##name, &(bfqq)->flags); \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 139) } \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 140) void bfq_clear_bfqq_##name(struct bfq_queue *bfqq) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 141) { \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 142) __clear_bit(BFQQF_##name, &(bfqq)->flags); \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 143) } \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 144) int bfq_bfqq_##name(const struct bfq_queue *bfqq) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 145) { \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 146) return test_bit(BFQQF_##name, &(bfqq)->flags); \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 147) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 148)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 149) BFQ_BFQQ_FNS(just_created);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 150) BFQ_BFQQ_FNS(busy);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 151) BFQ_BFQQ_FNS(wait_request);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 152) BFQ_BFQQ_FNS(non_blocking_wait_rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 153) BFQ_BFQQ_FNS(fifo_expire);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 154) BFQ_BFQQ_FNS(has_short_ttime);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 155) BFQ_BFQQ_FNS(sync);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 156) BFQ_BFQQ_FNS(IO_bound);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 157) BFQ_BFQQ_FNS(in_large_burst);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 158) BFQ_BFQQ_FNS(coop);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 159) BFQ_BFQQ_FNS(split_coop);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 160) BFQ_BFQQ_FNS(softrt_update);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 161) BFQ_BFQQ_FNS(has_waker);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 162) #undef BFQ_BFQQ_FNS \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 163)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 164) /* Expiration time of sync (0) and async (1) requests, in ns. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 165) static const u64 bfq_fifo_expire[2] = { NSEC_PER_SEC / 4, NSEC_PER_SEC / 8 };
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 166)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 167) /* Maximum backwards seek (magic number lifted from CFQ), in KiB. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 168) static const int bfq_back_max = 16 * 1024;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 169)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 170) /* Penalty of a backwards seek, in number of sectors. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 171) static const int bfq_back_penalty = 2;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 172)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 173) /* Idling period duration, in ns. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 174) static u64 bfq_slice_idle = NSEC_PER_SEC / 125;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 175)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 176) /* Minimum number of assigned budgets for which stats are safe to compute. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 177) static const int bfq_stats_min_budgets = 194;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 178)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 179) /* Default maximum budget values, in sectors and number of requests. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 180) static const int bfq_default_max_budget = 16 * 1024;
^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) * When a sync request is dispatched, the queue that contains that
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 184) * request, and all the ancestor entities of that queue, are charged
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 185) * with the number of sectors of the request. In contrast, if the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 186) * request is async, then the queue and its ancestor entities are
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 187) * charged with the number of sectors of the request, multiplied by
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 188) * the factor below. This throttles the bandwidth for async I/O,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 189) * w.r.t. to sync I/O, and it is done to counter the tendency of async
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 190) * writes to steal I/O throughput to reads.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 191) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 192) * The current value of this parameter is the result of a tuning with
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 193) * several hardware and software configurations. We tried to find the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 194) * lowest value for which writes do not cause noticeable problems to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 195) * reads. In fact, the lower this parameter, the stabler I/O control,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 196) * in the following respect. The lower this parameter is, the less
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 197) * the bandwidth enjoyed by a group decreases
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 198) * - when the group does writes, w.r.t. to when it does reads;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 199) * - when other groups do reads, w.r.t. to when they do writes.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 200) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 201) static const int bfq_async_charge_factor = 3;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 202)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 203) /* Default timeout values, in jiffies, approximating CFQ defaults. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 204) const int bfq_timeout = HZ / 8;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 205)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 206) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 207) * Time limit for merging (see comments in bfq_setup_cooperator). Set
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 208) * to the slowest value that, in our tests, proved to be effective in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 209) * removing false positives, while not causing true positives to miss
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 210) * queue merging.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 211) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 212) * As can be deduced from the low time limit below, queue merging, if
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 213) * successful, happens at the very beginning of the I/O of the involved
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 214) * cooperating processes, as a consequence of the arrival of the very
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 215) * first requests from each cooperator. After that, there is very
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 216) * little chance to find cooperators.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 217) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 218) static const unsigned long bfq_merge_time_limit = HZ/10;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 219)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 220) static struct kmem_cache *bfq_pool;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 221)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 222) /* Below this threshold (in ns), we consider thinktime immediate. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 223) #define BFQ_MIN_TT (2 * NSEC_PER_MSEC)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 224)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 225) /* hw_tag detection: parallel requests threshold and min samples needed. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 226) #define BFQ_HW_QUEUE_THRESHOLD 3
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 227) #define BFQ_HW_QUEUE_SAMPLES 32
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 228)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 229) #define BFQQ_SEEK_THR (sector_t)(8 * 100)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 230) #define BFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 231) #define BFQ_RQ_SEEKY(bfqd, last_pos, rq) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 232) (get_sdist(last_pos, rq) > \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 233) BFQQ_SEEK_THR && \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 234) (!blk_queue_nonrot(bfqd->queue) || \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 235) blk_rq_sectors(rq) < BFQQ_SECT_THR_NONROT))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 236) #define BFQQ_CLOSE_THR (sector_t)(8 * 1024)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 237) #define BFQQ_SEEKY(bfqq) (hweight32(bfqq->seek_history) > 19)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 238) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 239) * Sync random I/O is likely to be confused with soft real-time I/O,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 240) * because it is characterized by limited throughput and apparently
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 241) * isochronous arrival pattern. To avoid false positives, queues
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 242) * containing only random (seeky) I/O are prevented from being tagged
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 243) * as soft real-time.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 244) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 245) #define BFQQ_TOTALLY_SEEKY(bfqq) (bfqq->seek_history == -1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 246)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 247) /* Min number of samples required to perform peak-rate update */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 248) #define BFQ_RATE_MIN_SAMPLES 32
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 249) /* Min observation time interval required to perform a peak-rate update (ns) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 250) #define BFQ_RATE_MIN_INTERVAL (300*NSEC_PER_MSEC)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 251) /* Target observation time interval for a peak-rate update (ns) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 252) #define BFQ_RATE_REF_INTERVAL NSEC_PER_SEC
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 253)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 254) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 255) * Shift used for peak-rate fixed precision calculations.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 256) * With
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 257) * - the current shift: 16 positions
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 258) * - the current type used to store rate: u32
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 259) * - the current unit of measure for rate: [sectors/usec], or, more precisely,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 260) * [(sectors/usec) / 2^BFQ_RATE_SHIFT] to take into account the shift,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 261) * the range of rates that can be stored is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 262) * [1 / 2^BFQ_RATE_SHIFT, 2^(32 - BFQ_RATE_SHIFT)] sectors/usec =
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 263) * [1 / 2^16, 2^16] sectors/usec = [15e-6, 65536] sectors/usec =
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 264) * [15, 65G] sectors/sec
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 265) * Which, assuming a sector size of 512B, corresponds to a range of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 266) * [7.5K, 33T] B/sec
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 267) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 268) #define BFQ_RATE_SHIFT 16
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 269)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 270) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 271) * When configured for computing the duration of the weight-raising
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 272) * for interactive queues automatically (see the comments at the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 273) * beginning of this file), BFQ does it using the following formula:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 274) * duration = (ref_rate / r) * ref_wr_duration,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 275) * where r is the peak rate of the device, and ref_rate and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 276) * ref_wr_duration are two reference parameters. In particular,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 277) * ref_rate is the peak rate of the reference storage device (see
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 278) * below), and ref_wr_duration is about the maximum time needed, with
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 279) * BFQ and while reading two files in parallel, to load typical large
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 280) * applications on the reference device (see the comments on
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 281) * max_service_from_wr below, for more details on how ref_wr_duration
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 282) * is obtained). In practice, the slower/faster the device at hand
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 283) * is, the more/less it takes to load applications with respect to the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 284) * reference device. Accordingly, the longer/shorter BFQ grants
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 285) * weight raising to interactive applications.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 286) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 287) * BFQ uses two different reference pairs (ref_rate, ref_wr_duration),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 288) * depending on whether the device is rotational or non-rotational.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 289) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 290) * In the following definitions, ref_rate[0] and ref_wr_duration[0]
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 291) * are the reference values for a rotational device, whereas
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 292) * ref_rate[1] and ref_wr_duration[1] are the reference values for a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 293) * non-rotational device. The reference rates are not the actual peak
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 294) * rates of the devices used as a reference, but slightly lower
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 295) * values. The reason for using slightly lower values is that the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 296) * peak-rate estimator tends to yield slightly lower values than the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 297) * actual peak rate (it can yield the actual peak rate only if there
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 298) * is only one process doing I/O, and the process does sequential
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 299) * I/O).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 300) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 301) * The reference peak rates are measured in sectors/usec, left-shifted
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 302) * by BFQ_RATE_SHIFT.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 303) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 304) static int ref_rate[2] = {14000, 33000};
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 305) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 306) * To improve readability, a conversion function is used to initialize
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 307) * the following array, which entails that the array can be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 308) * initialized only in a function.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 309) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 310) static int ref_wr_duration[2];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 311)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 312) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 313) * BFQ uses the above-detailed, time-based weight-raising mechanism to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 314) * privilege interactive tasks. This mechanism is vulnerable to the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 315) * following false positives: I/O-bound applications that will go on
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 316) * doing I/O for much longer than the duration of weight
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 317) * raising. These applications have basically no benefit from being
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 318) * weight-raised at the beginning of their I/O. On the opposite end,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 319) * while being weight-raised, these applications
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 320) * a) unjustly steal throughput to applications that may actually need
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 321) * low latency;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 322) * b) make BFQ uselessly perform device idling; device idling results
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 323) * in loss of device throughput with most flash-based storage, and may
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 324) * increase latencies when used purposelessly.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 325) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 326) * BFQ tries to reduce these problems, by adopting the following
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 327) * countermeasure. To introduce this countermeasure, we need first to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 328) * finish explaining how the duration of weight-raising for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 329) * interactive tasks is computed.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 330) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 331) * For a bfq_queue deemed as interactive, the duration of weight
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 332) * raising is dynamically adjusted, as a function of the estimated
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 333) * peak rate of the device, so as to be equal to the time needed to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 334) * execute the 'largest' interactive task we benchmarked so far. By
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 335) * largest task, we mean the task for which each involved process has
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 336) * to do more I/O than for any of the other tasks we benchmarked. This
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 337) * reference interactive task is the start-up of LibreOffice Writer,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 338) * and in this task each process/bfq_queue needs to have at most ~110K
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 339) * sectors transferred.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 340) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 341) * This last piece of information enables BFQ to reduce the actual
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 342) * duration of weight-raising for at least one class of I/O-bound
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 343) * applications: those doing sequential or quasi-sequential I/O. An
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 344) * example is file copy. In fact, once started, the main I/O-bound
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 345) * processes of these applications usually consume the above 110K
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 346) * sectors in much less time than the processes of an application that
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 347) * is starting, because these I/O-bound processes will greedily devote
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 348) * almost all their CPU cycles only to their target,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 349) * throughput-friendly I/O operations. This is even more true if BFQ
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 350) * happens to be underestimating the device peak rate, and thus
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 351) * overestimating the duration of weight raising. But, according to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 352) * our measurements, once transferred 110K sectors, these processes
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 353) * have no right to be weight-raised any longer.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 354) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 355) * Basing on the last consideration, BFQ ends weight-raising for a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 356) * bfq_queue if the latter happens to have received an amount of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 357) * service at least equal to the following constant. The constant is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 358) * set to slightly more than 110K, to have a minimum safety margin.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 359) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 360) * This early ending of weight-raising reduces the amount of time
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 361) * during which interactive false positives cause the two problems
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 362) * described at the beginning of these comments.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 363) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 364) static const unsigned long max_service_from_wr = 120000;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 365)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 366) #define RQ_BIC(rq) icq_to_bic((rq)->elv.priv[0])
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 367) #define RQ_BFQQ(rq) ((rq)->elv.priv[1])
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 368)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 369) struct bfq_queue *bic_to_bfqq(struct bfq_io_cq *bic, bool is_sync)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 370) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 371) return bic->bfqq[is_sync];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 372) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 373)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 374) void bic_set_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq, bool is_sync)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 375) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 376) bic->bfqq[is_sync] = bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 377) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 378)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 379) struct bfq_data *bic_to_bfqd(struct bfq_io_cq *bic)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 380) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 381) return bic->icq.q->elevator->elevator_data;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 382) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 383)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 384) /**
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 385) * icq_to_bic - convert iocontext queue structure to bfq_io_cq.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 386) * @icq: the iocontext queue.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 387) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 388) static struct bfq_io_cq *icq_to_bic(struct io_cq *icq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 389) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 390) /* bic->icq is the first member, %NULL will convert to %NULL */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 391) return container_of(icq, struct bfq_io_cq, icq);
^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) * bfq_bic_lookup - search into @ioc a bic associated to @bfqd.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 396) * @bfqd: the lookup key.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 397) * @ioc: the io_context of the process doing I/O.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 398) * @q: the request queue.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 399) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 400) static struct bfq_io_cq *bfq_bic_lookup(struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 401) struct io_context *ioc,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 402) struct request_queue *q)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 403) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 404) if (ioc) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 405) unsigned long flags;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 406) struct bfq_io_cq *icq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 407)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 408) spin_lock_irqsave(&q->queue_lock, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 409) icq = icq_to_bic(ioc_lookup_icq(ioc, q));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 410) spin_unlock_irqrestore(&q->queue_lock, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 411)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 412) return icq;
^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) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 416) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 417)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 418) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 419) * Scheduler run of queue, if there are requests pending and no one in the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 420) * driver that will restart queueing.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 421) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 422) void bfq_schedule_dispatch(struct bfq_data *bfqd)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 423) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 424) if (bfqd->queued != 0) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 425) bfq_log(bfqd, "schedule dispatch");
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 426) blk_mq_run_hw_queues(bfqd->queue, true);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 427) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 428) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 429)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 430) #define bfq_class_idle(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 431)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 432) #define bfq_sample_valid(samples) ((samples) > 80)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 433)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 434) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 435) * Lifted from AS - choose which of rq1 and rq2 that is best served now.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 436) * We choose the request that is closer to the head right now. Distance
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 437) * behind the head is penalized and only allowed to a certain extent.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 438) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 439) static struct request *bfq_choose_req(struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 440) struct request *rq1,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 441) struct request *rq2,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 442) sector_t last)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 443) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 444) sector_t s1, s2, d1 = 0, d2 = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 445) unsigned long back_max;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 446) #define BFQ_RQ1_WRAP 0x01 /* request 1 wraps */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 447) #define BFQ_RQ2_WRAP 0x02 /* request 2 wraps */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 448) unsigned int wrap = 0; /* bit mask: requests behind the disk head? */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 449)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 450) if (!rq1 || rq1 == rq2)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 451) return rq2;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 452) if (!rq2)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 453) return rq1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 454)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 455) if (rq_is_sync(rq1) && !rq_is_sync(rq2))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 456) return rq1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 457) else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 458) return rq2;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 459) if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 460) return rq1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 461) else if ((rq2->cmd_flags & REQ_META) && !(rq1->cmd_flags & REQ_META))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 462) return rq2;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 463)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 464) s1 = blk_rq_pos(rq1);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 465) s2 = blk_rq_pos(rq2);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 466)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 467) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 468) * By definition, 1KiB is 2 sectors.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 469) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 470) back_max = bfqd->bfq_back_max * 2;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 471)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 472) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 473) * Strict one way elevator _except_ in the case where we allow
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 474) * short backward seeks which are biased as twice the cost of a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 475) * similar forward seek.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 476) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 477) if (s1 >= last)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 478) d1 = s1 - last;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 479) else if (s1 + back_max >= last)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 480) d1 = (last - s1) * bfqd->bfq_back_penalty;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 481) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 482) wrap |= BFQ_RQ1_WRAP;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 483)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 484) if (s2 >= last)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 485) d2 = s2 - last;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 486) else if (s2 + back_max >= last)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 487) d2 = (last - s2) * bfqd->bfq_back_penalty;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 488) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 489) wrap |= BFQ_RQ2_WRAP;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 490)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 491) /* Found required data */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 492)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 493) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 494) * By doing switch() on the bit mask "wrap" we avoid having to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 495) * check two variables for all permutations: --> faster!
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 496) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 497) switch (wrap) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 498) case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 499) if (d1 < d2)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 500) return rq1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 501) else if (d2 < d1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 502) return rq2;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 503)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 504) if (s1 >= s2)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 505) return rq1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 506) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 507) return rq2;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 508)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 509) case BFQ_RQ2_WRAP:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 510) return rq1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 511) case BFQ_RQ1_WRAP:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 512) return rq2;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 513) case BFQ_RQ1_WRAP|BFQ_RQ2_WRAP: /* both rqs wrapped */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 514) default:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 515) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 516) * Since both rqs are wrapped,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 517) * start with the one that's further behind head
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 518) * (--> only *one* back seek required),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 519) * since back seek takes more time than forward.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 520) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 521) if (s1 <= s2)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 522) return rq1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 523) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 524) return rq2;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 525) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 526) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 527)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 528) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 529) * Async I/O can easily starve sync I/O (both sync reads and sync
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 530) * writes), by consuming all tags. Similarly, storms of sync writes,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 531) * such as those that sync(2) may trigger, can starve sync reads.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 532) * Limit depths of async I/O and sync writes so as to counter both
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 533) * problems.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 534) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 535) static void bfq_limit_depth(unsigned int op, struct blk_mq_alloc_data *data)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 536) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 537) struct bfq_data *bfqd = data->q->elevator->elevator_data;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 538)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 539) if (op_is_sync(op) && !op_is_write(op))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 540) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 541)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 542) data->shallow_depth =
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 543) bfqd->word_depths[!!bfqd->wr_busy_queues][op_is_sync(op)];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 544)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 545) bfq_log(bfqd, "[%s] wr_busy %d sync %d depth %u",
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 546) __func__, bfqd->wr_busy_queues, op_is_sync(op),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 547) data->shallow_depth);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 548) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 549)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 550) static struct bfq_queue *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 551) bfq_rq_pos_tree_lookup(struct bfq_data *bfqd, struct rb_root *root,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 552) sector_t sector, struct rb_node **ret_parent,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 553) struct rb_node ***rb_link)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 554) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 555) struct rb_node **p, *parent;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 556) struct bfq_queue *bfqq = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 557)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 558) parent = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 559) p = &root->rb_node;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 560) while (*p) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 561) struct rb_node **n;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 562)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 563) parent = *p;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 564) bfqq = rb_entry(parent, struct bfq_queue, pos_node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 565)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 566) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 567) * Sort strictly based on sector. Smallest to the left,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 568) * largest to the right.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 569) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 570) if (sector > blk_rq_pos(bfqq->next_rq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 571) n = &(*p)->rb_right;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 572) else if (sector < blk_rq_pos(bfqq->next_rq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 573) n = &(*p)->rb_left;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 574) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 575) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 576) p = n;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 577) bfqq = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 578) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 579)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 580) *ret_parent = parent;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 581) if (rb_link)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 582) *rb_link = p;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 583)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 584) bfq_log(bfqd, "rq_pos_tree_lookup %llu: returning %d",
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 585) (unsigned long long)sector,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 586) bfqq ? bfqq->pid : 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 587)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 588) return bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 589) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 590)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 591) static bool bfq_too_late_for_merging(struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 592) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 593) return bfqq->service_from_backlogged > 0 &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 594) time_is_before_jiffies(bfqq->first_IO_time +
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 595) bfq_merge_time_limit);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 596) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 597)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 598) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 599) * The following function is not marked as __cold because it is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 600) * actually cold, but for the same performance goal described in the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 601) * comments on the likely() at the beginning of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 602) * bfq_setup_cooperator(). Unexpectedly, to reach an even lower
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 603) * execution time for the case where this function is not invoked, we
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 604) * had to add an unlikely() in each involved if().
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 605) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 606) void __cold
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 607) bfq_pos_tree_add_move(struct bfq_data *bfqd, struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 608) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 609) struct rb_node **p, *parent;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 610) struct bfq_queue *__bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 611)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 612) if (bfqq->pos_root) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 613) rb_erase(&bfqq->pos_node, bfqq->pos_root);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 614) bfqq->pos_root = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 615) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 616)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 617) /* oom_bfqq does not participate in queue merging */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 618) if (bfqq == &bfqd->oom_bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 619) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 620)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 621) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 622) * bfqq cannot be merged any longer (see comments in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 623) * bfq_setup_cooperator): no point in adding bfqq into the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 624) * position tree.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 625) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 626) if (bfq_too_late_for_merging(bfqq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 627) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 628)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 629) if (bfq_class_idle(bfqq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 630) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 631) if (!bfqq->next_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 632) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 633)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 634) bfqq->pos_root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 635) __bfqq = bfq_rq_pos_tree_lookup(bfqd, bfqq->pos_root,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 636) blk_rq_pos(bfqq->next_rq), &parent, &p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 637) if (!__bfqq) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 638) rb_link_node(&bfqq->pos_node, parent, p);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 639) rb_insert_color(&bfqq->pos_node, bfqq->pos_root);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 640) } else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 641) bfqq->pos_root = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 642) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 643)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 644) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 645) * The following function returns false either if every active queue
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 646) * must receive the same share of the throughput (symmetric scenario),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 647) * or, as a special case, if bfqq must receive a share of the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 648) * throughput lower than or equal to the share that every other active
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 649) * queue must receive. If bfqq does sync I/O, then these are the only
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 650) * two cases where bfqq happens to be guaranteed its share of the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 651) * throughput even if I/O dispatching is not plugged when bfqq remains
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 652) * temporarily empty (for more details, see the comments in the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 653) * function bfq_better_to_idle()). For this reason, the return value
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 654) * of this function is used to check whether I/O-dispatch plugging can
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 655) * be avoided.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 656) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 657) * The above first case (symmetric scenario) occurs when:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 658) * 1) all active queues have the same weight,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 659) * 2) all active queues belong to the same I/O-priority class,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 660) * 3) all active groups at the same level in the groups tree have the same
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 661) * weight,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 662) * 4) all active groups at the same level in the groups tree have the same
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 663) * number of children.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 664) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 665) * Unfortunately, keeping the necessary state for evaluating exactly
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 666) * the last two symmetry sub-conditions above would be quite complex
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 667) * and time consuming. Therefore this function evaluates, instead,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 668) * only the following stronger three sub-conditions, for which it is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 669) * much easier to maintain the needed state:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 670) * 1) all active queues have the same weight,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 671) * 2) all active queues belong to the same I/O-priority class,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 672) * 3) there are no active groups.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 673) * In particular, the last condition is always true if hierarchical
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 674) * support or the cgroups interface are not enabled, thus no state
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 675) * needs to be maintained in this case.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 676) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 677) static bool bfq_asymmetric_scenario(struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 678) struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 679) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 680) bool smallest_weight = bfqq &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 681) bfqq->weight_counter &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 682) bfqq->weight_counter ==
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 683) container_of(
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 684) rb_first_cached(&bfqd->queue_weights_tree),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 685) struct bfq_weight_counter,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 686) weights_node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 687)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 688) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 689) * For queue weights to differ, queue_weights_tree must contain
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 690) * at least two nodes.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 691) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 692) bool varied_queue_weights = !smallest_weight &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 693) !RB_EMPTY_ROOT(&bfqd->queue_weights_tree.rb_root) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 694) (bfqd->queue_weights_tree.rb_root.rb_node->rb_left ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 695) bfqd->queue_weights_tree.rb_root.rb_node->rb_right);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 696)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 697) bool multiple_classes_busy =
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 698) (bfqd->busy_queues[0] && bfqd->busy_queues[1]) ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 699) (bfqd->busy_queues[0] && bfqd->busy_queues[2]) ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 700) (bfqd->busy_queues[1] && bfqd->busy_queues[2]);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 701)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 702) return varied_queue_weights || multiple_classes_busy
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 703) #ifdef CONFIG_BFQ_GROUP_IOSCHED
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 704) || bfqd->num_groups_with_pending_reqs > 0
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 705) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 706) ;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 707) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 708)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 709) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 710) * If the weight-counter tree passed as input contains no counter for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 711) * the weight of the input queue, then add that counter; otherwise just
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 712) * increment the existing counter.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 713) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 714) * Note that weight-counter trees contain few nodes in mostly symmetric
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 715) * scenarios. For example, if all queues have the same weight, then the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 716) * weight-counter tree for the queues may contain at most one node.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 717) * This holds even if low_latency is on, because weight-raised queues
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 718) * are not inserted in the tree.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 719) * In most scenarios, the rate at which nodes are created/destroyed
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 720) * should be low too.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 721) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 722) void bfq_weights_tree_add(struct bfq_data *bfqd, struct bfq_queue *bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 723) struct rb_root_cached *root)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 724) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 725) struct bfq_entity *entity = &bfqq->entity;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 726) struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 727) bool leftmost = true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 728)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 729) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 730) * Do not insert if the queue is already associated with a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 731) * counter, which happens if:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 732) * 1) a request arrival has caused the queue to become both
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 733) * non-weight-raised, and hence change its weight, and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 734) * backlogged; in this respect, each of the two events
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 735) * causes an invocation of this function,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 736) * 2) this is the invocation of this function caused by the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 737) * second event. This second invocation is actually useless,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 738) * and we handle this fact by exiting immediately. More
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 739) * efficient or clearer solutions might possibly be adopted.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 740) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 741) if (bfqq->weight_counter)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 742) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 743)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 744) while (*new) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 745) struct bfq_weight_counter *__counter = container_of(*new,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 746) struct bfq_weight_counter,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 747) weights_node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 748) parent = *new;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 749)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 750) if (entity->weight == __counter->weight) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 751) bfqq->weight_counter = __counter;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 752) goto inc_counter;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 753) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 754) if (entity->weight < __counter->weight)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 755) new = &((*new)->rb_left);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 756) else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 757) new = &((*new)->rb_right);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 758) leftmost = false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 759) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 760) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 761)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 762) bfqq->weight_counter = kzalloc(sizeof(struct bfq_weight_counter),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 763) GFP_ATOMIC);
^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) * In the unlucky event of an allocation failure, we just
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 767) * exit. This will cause the weight of queue to not be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 768) * considered in bfq_asymmetric_scenario, which, in its turn,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 769) * causes the scenario to be deemed wrongly symmetric in case
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 770) * bfqq's weight would have been the only weight making the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 771) * scenario asymmetric. On the bright side, no unbalance will
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 772) * however occur when bfqq becomes inactive again (the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 773) * invocation of this function is triggered by an activation
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 774) * of queue). In fact, bfq_weights_tree_remove does nothing
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 775) * if !bfqq->weight_counter.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 776) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 777) if (unlikely(!bfqq->weight_counter))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 778) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 779)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 780) bfqq->weight_counter->weight = entity->weight;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 781) rb_link_node(&bfqq->weight_counter->weights_node, parent, new);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 782) rb_insert_color_cached(&bfqq->weight_counter->weights_node, root,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 783) leftmost);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 784)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 785) inc_counter:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 786) bfqq->weight_counter->num_active++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 787) bfqq->ref++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 788) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 789)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 790) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 791) * Decrement the weight counter associated with the queue, and, if the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 792) * counter reaches 0, remove the counter from the tree.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 793) * See the comments to the function bfq_weights_tree_add() for considerations
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 794) * about overhead.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 795) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 796) void __bfq_weights_tree_remove(struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 797) struct bfq_queue *bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 798) struct rb_root_cached *root)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 799) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 800) if (!bfqq->weight_counter)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 801) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 802)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 803) bfqq->weight_counter->num_active--;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 804) if (bfqq->weight_counter->num_active > 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 805) goto reset_entity_pointer;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 806)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 807) rb_erase_cached(&bfqq->weight_counter->weights_node, root);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 808) kfree(bfqq->weight_counter);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 809)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 810) reset_entity_pointer:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 811) bfqq->weight_counter = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 812) bfq_put_queue(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 813) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 814)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 815) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 816) * Invoke __bfq_weights_tree_remove on bfqq and decrement the number
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 817) * of active groups for each queue's inactive parent entity.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 818) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 819) void bfq_weights_tree_remove(struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 820) struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 821) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 822) struct bfq_entity *entity = bfqq->entity.parent;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 823)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 824) for_each_entity(entity) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 825) struct bfq_sched_data *sd = entity->my_sched_data;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 826)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 827) if (sd->next_in_service || sd->in_service_entity) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 828) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 829) * entity is still active, because either
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 830) * next_in_service or in_service_entity is not
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 831) * NULL (see the comments on the definition of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 832) * next_in_service for details on why
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 833) * in_service_entity must be checked too).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 834) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 835) * As a consequence, its parent entities are
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 836) * active as well, and thus this loop must
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 837) * stop here.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 838) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 839) break;
^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) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 843) * The decrement of num_groups_with_pending_reqs is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 844) * not performed immediately upon the deactivation of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 845) * entity, but it is delayed to when it also happens
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 846) * that the first leaf descendant bfqq of entity gets
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 847) * all its pending requests completed. The following
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 848) * instructions perform this delayed decrement, if
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 849) * needed. See the comments on
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 850) * num_groups_with_pending_reqs for details.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 851) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 852) if (entity->in_groups_with_pending_reqs) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 853) entity->in_groups_with_pending_reqs = false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 854) bfqd->num_groups_with_pending_reqs--;
^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)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 858) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 859) * Next function is invoked last, because it causes bfqq to be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 860) * freed if the following holds: bfqq is not in service and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 861) * has no dispatched request. DO NOT use bfqq after the next
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 862) * function invocation.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 863) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 864) __bfq_weights_tree_remove(bfqd, bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 865) &bfqd->queue_weights_tree);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 866) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 867)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 868) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 869) * Return expired entry, or NULL to just start from scratch in rbtree.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 870) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 871) static struct request *bfq_check_fifo(struct bfq_queue *bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 872) struct request *last)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 873) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 874) struct request *rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 875)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 876) if (bfq_bfqq_fifo_expire(bfqq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 877) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 878)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 879) bfq_mark_bfqq_fifo_expire(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 880)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 881) rq = rq_entry_fifo(bfqq->fifo.next);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 882)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 883) if (rq == last || ktime_get_ns() < rq->fifo_time)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 884) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 885)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 886) bfq_log_bfqq(bfqq->bfqd, bfqq, "check_fifo: returned %p", rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 887) return rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 888) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 889)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 890) static struct request *bfq_find_next_rq(struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 891) struct bfq_queue *bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 892) struct request *last)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 893) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 894) struct rb_node *rbnext = rb_next(&last->rb_node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 895) struct rb_node *rbprev = rb_prev(&last->rb_node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 896) struct request *next, *prev = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 897)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 898) /* Follow expired path, else get first next available. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 899) next = bfq_check_fifo(bfqq, last);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 900) if (next)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 901) return next;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 902)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 903) if (rbprev)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 904) prev = rb_entry_rq(rbprev);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 905)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 906) if (rbnext)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 907) next = rb_entry_rq(rbnext);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 908) else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 909) rbnext = rb_first(&bfqq->sort_list);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 910) if (rbnext && rbnext != &last->rb_node)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 911) next = rb_entry_rq(rbnext);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 912) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 913)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 914) return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 915) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 916)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 917) /* see the definition of bfq_async_charge_factor for details */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 918) static unsigned long bfq_serv_to_charge(struct request *rq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 919) struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 920) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 921) if (bfq_bfqq_sync(bfqq) || bfqq->wr_coeff > 1 ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 922) bfq_asymmetric_scenario(bfqq->bfqd, bfqq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 923) return blk_rq_sectors(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 924)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 925) return blk_rq_sectors(rq) * bfq_async_charge_factor;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 926) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 927)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 928) /**
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 929) * bfq_updated_next_req - update the queue after a new next_rq selection.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 930) * @bfqd: the device data the queue belongs to.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 931) * @bfqq: the queue to update.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 932) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 933) * If the first request of a queue changes we make sure that the queue
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 934) * has enough budget to serve at least its first request (if the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 935) * request has grown). We do this because if the queue has not enough
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 936) * budget for its first request, it has to go through two dispatch
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 937) * rounds to actually get it dispatched.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 938) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 939) static void bfq_updated_next_req(struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 940) struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 941) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 942) struct bfq_entity *entity = &bfqq->entity;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 943) struct request *next_rq = bfqq->next_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 944) unsigned long new_budget;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 945)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 946) if (!next_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 947) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 948)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 949) if (bfqq == bfqd->in_service_queue)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 950) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 951) * In order not to break guarantees, budgets cannot be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 952) * changed after an entity has been selected.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 953) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 954) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 955)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 956) new_budget = max_t(unsigned long,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 957) max_t(unsigned long, bfqq->max_budget,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 958) bfq_serv_to_charge(next_rq, bfqq)),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 959) entity->service);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 960) if (entity->budget != new_budget) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 961) entity->budget = new_budget;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 962) bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu",
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 963) new_budget);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 964) bfq_requeue_bfqq(bfqd, bfqq, false);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 965) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 966) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 967)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 968) static unsigned int bfq_wr_duration(struct bfq_data *bfqd)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 969) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 970) u64 dur;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 971)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 972) if (bfqd->bfq_wr_max_time > 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 973) return bfqd->bfq_wr_max_time;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 974)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 975) dur = bfqd->rate_dur_prod;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 976) do_div(dur, bfqd->peak_rate);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 977)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 978) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 979) * Limit duration between 3 and 25 seconds. The upper limit
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 980) * has been conservatively set after the following worst case:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 981) * on a QEMU/KVM virtual machine
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 982) * - running in a slow PC
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 983) * - with a virtual disk stacked on a slow low-end 5400rpm HDD
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 984) * - serving a heavy I/O workload, such as the sequential reading
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 985) * of several files
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 986) * mplayer took 23 seconds to start, if constantly weight-raised.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 987) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 988) * As for higher values than that accommodating the above bad
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 989) * scenario, tests show that higher values would often yield
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 990) * the opposite of the desired result, i.e., would worsen
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 991) * responsiveness by allowing non-interactive applications to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 992) * preserve weight raising for too long.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 993) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 994) * On the other end, lower values than 3 seconds make it
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 995) * difficult for most interactive tasks to complete their jobs
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 996) * before weight-raising finishes.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 997) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 998) return clamp_val(dur, msecs_to_jiffies(3000), msecs_to_jiffies(25000));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 999) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1000)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1001) /* switch back from soft real-time to interactive weight raising */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1002) static void switch_back_to_interactive_wr(struct bfq_queue *bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1003) struct bfq_data *bfqd)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1004) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1005) bfqq->wr_coeff = bfqd->bfq_wr_coeff;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1006) bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1007) bfqq->last_wr_start_finish = bfqq->wr_start_at_switch_to_srt;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1008) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1009)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1010) static void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1011) bfq_bfqq_resume_state(struct bfq_queue *bfqq, struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1012) struct bfq_io_cq *bic, bool bfq_already_existing)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1013) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1014) unsigned int old_wr_coeff = bfqq->wr_coeff;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1015) bool busy = bfq_already_existing && bfq_bfqq_busy(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1016)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1017) if (bic->saved_has_short_ttime)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1018) bfq_mark_bfqq_has_short_ttime(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1019) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1020) bfq_clear_bfqq_has_short_ttime(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1021)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1022) if (bic->saved_IO_bound)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1023) bfq_mark_bfqq_IO_bound(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1024) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1025) bfq_clear_bfqq_IO_bound(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1026)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1027) bfqq->entity.new_weight = bic->saved_weight;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1028) bfqq->ttime = bic->saved_ttime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1029) bfqq->wr_coeff = bic->saved_wr_coeff;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1030) bfqq->wr_start_at_switch_to_srt = bic->saved_wr_start_at_switch_to_srt;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1031) bfqq->last_wr_start_finish = bic->saved_last_wr_start_finish;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1032) bfqq->wr_cur_max_time = bic->saved_wr_cur_max_time;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1033)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1034) if (bfqq->wr_coeff > 1 && (bfq_bfqq_in_large_burst(bfqq) ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1035) time_is_before_jiffies(bfqq->last_wr_start_finish +
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1036) bfqq->wr_cur_max_time))) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1037) if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1038) !bfq_bfqq_in_large_burst(bfqq) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1039) time_is_after_eq_jiffies(bfqq->wr_start_at_switch_to_srt +
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1040) bfq_wr_duration(bfqd))) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1041) switch_back_to_interactive_wr(bfqq, bfqd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1042) } else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1043) bfqq->wr_coeff = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1044) bfq_log_bfqq(bfqq->bfqd, bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1045) "resume state: switching off wr");
^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)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1049) /* make sure weight will be updated, however we got here */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1050) bfqq->entity.prio_changed = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1051)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1052) if (likely(!busy))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1053) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1054)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1055) if (old_wr_coeff == 1 && bfqq->wr_coeff > 1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1056) bfqd->wr_busy_queues++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1057) else if (old_wr_coeff > 1 && bfqq->wr_coeff == 1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1058) bfqd->wr_busy_queues--;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1059) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1060)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1061) static int bfqq_process_refs(struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1062) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1063) return bfqq->ref - bfqq->allocated - bfqq->entity.on_st_or_in_serv -
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1064) (bfqq->weight_counter != NULL);
^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) /* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1068) static void bfq_reset_burst_list(struct bfq_data *bfqd, struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1069) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1070) struct bfq_queue *item;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1071) struct hlist_node *n;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1072)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1073) hlist_for_each_entry_safe(item, n, &bfqd->burst_list, burst_list_node)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1074) hlist_del_init(&item->burst_list_node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1075)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1076) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1077) * Start the creation of a new burst list only if there is no
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1078) * active queue. See comments on the conditional invocation of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1079) * bfq_handle_burst().
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1080) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1081) if (bfq_tot_busy_queues(bfqd) == 0) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1082) hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1083) bfqd->burst_size = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1084) } else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1085) bfqd->burst_size = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1086)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1087) bfqd->burst_parent_entity = bfqq->entity.parent;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1088) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1089)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1090) /* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1091) static void bfq_add_to_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1092) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1093) /* Increment burst size to take into account also bfqq */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1094) bfqd->burst_size++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1095)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1096) if (bfqd->burst_size == bfqd->bfq_large_burst_thresh) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1097) struct bfq_queue *pos, *bfqq_item;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1098) struct hlist_node *n;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1099)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1100) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1101) * Enough queues have been activated shortly after each
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1102) * other to consider this burst as large.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1103) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1104) bfqd->large_burst = true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1105)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1106) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1107) * We can now mark all queues in the burst list as
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1108) * belonging to a large burst.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1109) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1110) hlist_for_each_entry(bfqq_item, &bfqd->burst_list,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1111) burst_list_node)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1112) bfq_mark_bfqq_in_large_burst(bfqq_item);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1113) bfq_mark_bfqq_in_large_burst(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1114)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1115) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1116) * From now on, and until the current burst finishes, any
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1117) * new queue being activated shortly after the last queue
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1118) * was inserted in the burst can be immediately marked as
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1119) * belonging to a large burst. So the burst list is not
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1120) * needed any more. Remove it.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1121) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1122) hlist_for_each_entry_safe(pos, n, &bfqd->burst_list,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1123) burst_list_node)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1124) hlist_del_init(&pos->burst_list_node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1125) } else /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1126) * Burst not yet large: add bfqq to the burst list. Do
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1127) * not increment the ref counter for bfqq, because bfqq
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1128) * is removed from the burst list before freeing bfqq
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1129) * in put_queue.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1130) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1131) hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1132) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1133)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1134) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1135) * If many queues belonging to the same group happen to be created
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1136) * shortly after each other, then the processes associated with these
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1137) * queues have typically a common goal. In particular, bursts of queue
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1138) * creations are usually caused by services or applications that spawn
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1139) * many parallel threads/processes. Examples are systemd during boot,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1140) * or git grep. To help these processes get their job done as soon as
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1141) * possible, it is usually better to not grant either weight-raising
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1142) * or device idling to their queues, unless these queues must be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1143) * protected from the I/O flowing through other active queues.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1144) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1145) * In this comment we describe, firstly, the reasons why this fact
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1146) * holds, and, secondly, the next function, which implements the main
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1147) * steps needed to properly mark these queues so that they can then be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1148) * treated in a different way.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1149) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1150) * The above services or applications benefit mostly from a high
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1151) * throughput: the quicker the requests of the activated queues are
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1152) * cumulatively served, the sooner the target job of these queues gets
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1153) * completed. As a consequence, weight-raising any of these queues,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1154) * which also implies idling the device for it, is almost always
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1155) * counterproductive, unless there are other active queues to isolate
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1156) * these new queues from. If there no other active queues, then
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1157) * weight-raising these new queues just lowers throughput in most
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1158) * cases.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1159) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1160) * On the other hand, a burst of queue creations may be caused also by
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1161) * the start of an application that does not consist of a lot of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1162) * parallel I/O-bound threads. In fact, with a complex application,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1163) * several short processes may need to be executed to start-up the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1164) * application. In this respect, to start an application as quickly as
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1165) * possible, the best thing to do is in any case to privilege the I/O
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1166) * related to the application with respect to all other
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1167) * I/O. Therefore, the best strategy to start as quickly as possible
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1168) * an application that causes a burst of queue creations is to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1169) * weight-raise all the queues created during the burst. This is the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1170) * exact opposite of the best strategy for the other type of bursts.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1171) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1172) * In the end, to take the best action for each of the two cases, the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1173) * two types of bursts need to be distinguished. Fortunately, this
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1174) * seems relatively easy, by looking at the sizes of the bursts. In
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1175) * particular, we found a threshold such that only bursts with a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1176) * larger size than that threshold are apparently caused by
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1177) * services or commands such as systemd or git grep. For brevity,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1178) * hereafter we call just 'large' these bursts. BFQ *does not*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1179) * weight-raise queues whose creation occurs in a large burst. In
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1180) * addition, for each of these queues BFQ performs or does not perform
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1181) * idling depending on which choice boosts the throughput more. The
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1182) * exact choice depends on the device and request pattern at
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1183) * hand.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1184) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1185) * Unfortunately, false positives may occur while an interactive task
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1186) * is starting (e.g., an application is being started). The
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1187) * consequence is that the queues associated with the task do not
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1188) * enjoy weight raising as expected. Fortunately these false positives
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1189) * are very rare. They typically occur if some service happens to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1190) * start doing I/O exactly when the interactive task starts.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1191) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1192) * Turning back to the next function, it is invoked only if there are
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1193) * no active queues (apart from active queues that would belong to the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1194) * same, possible burst bfqq would belong to), and it implements all
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1195) * the steps needed to detect the occurrence of a large burst and to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1196) * properly mark all the queues belonging to it (so that they can then
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1197) * be treated in a different way). This goal is achieved by
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1198) * maintaining a "burst list" that holds, temporarily, the queues that
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1199) * belong to the burst in progress. The list is then used to mark
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1200) * these queues as belonging to a large burst if the burst does become
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1201) * large. The main steps are the following.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1202) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1203) * . when the very first queue is created, the queue is inserted into the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1204) * list (as it could be the first queue in a possible burst)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1205) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1206) * . if the current burst has not yet become large, and a queue Q that does
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1207) * not yet belong to the burst is activated shortly after the last time
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1208) * at which a new queue entered the burst list, then the function appends
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1209) * Q to the burst list
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1210) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1211) * . if, as a consequence of the previous step, the burst size reaches
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1212) * the large-burst threshold, then
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1213) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1214) * . all the queues in the burst list are marked as belonging to a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1215) * large burst
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1216) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1217) * . the burst list is deleted; in fact, the burst list already served
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1218) * its purpose (keeping temporarily track of the queues in a burst,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1219) * so as to be able to mark them as belonging to a large burst in the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1220) * previous sub-step), and now is not needed any more
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1221) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1222) * . the device enters a large-burst mode
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1223) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1224) * . if a queue Q that does not belong to the burst is created while
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1225) * the device is in large-burst mode and shortly after the last time
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1226) * at which a queue either entered the burst list or was marked as
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1227) * belonging to the current large burst, then Q is immediately marked
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1228) * as belonging to a large burst.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1229) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1230) * . if a queue Q that does not belong to the burst is created a while
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1231) * later, i.e., not shortly after, than the last time at which a queue
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1232) * either entered the burst list or was marked as belonging to the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1233) * current large burst, then the current burst is deemed as finished and:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1234) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1235) * . the large-burst mode is reset if set
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1236) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1237) * . the burst list is emptied
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1238) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1239) * . Q is inserted in the burst list, as Q may be the first queue
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1240) * in a possible new burst (then the burst list contains just Q
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1241) * after this step).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1242) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1243) static void bfq_handle_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1244) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1245) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1246) * If bfqq is already in the burst list or is part of a large
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1247) * burst, or finally has just been split, then there is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1248) * nothing else to do.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1249) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1250) if (!hlist_unhashed(&bfqq->burst_list_node) ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1251) bfq_bfqq_in_large_burst(bfqq) ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1252) time_is_after_eq_jiffies(bfqq->split_time +
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1253) msecs_to_jiffies(10)))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1254) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1255)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1256) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1257) * If bfqq's creation happens late enough, or bfqq belongs to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1258) * a different group than the burst group, then the current
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1259) * burst is finished, and related data structures must be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1260) * reset.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1261) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1262) * In this respect, consider the special case where bfqq is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1263) * the very first queue created after BFQ is selected for this
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1264) * device. In this case, last_ins_in_burst and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1265) * burst_parent_entity are not yet significant when we get
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1266) * here. But it is easy to verify that, whether or not the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1267) * following condition is true, bfqq will end up being
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1268) * inserted into the burst list. In particular the list will
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1269) * happen to contain only bfqq. And this is exactly what has
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1270) * to happen, as bfqq may be the first queue of the first
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1271) * burst.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1272) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1273) if (time_is_before_jiffies(bfqd->last_ins_in_burst +
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1274) bfqd->bfq_burst_interval) ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1275) bfqq->entity.parent != bfqd->burst_parent_entity) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1276) bfqd->large_burst = false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1277) bfq_reset_burst_list(bfqd, bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1278) goto end;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1279) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1280)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1281) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1282) * If we get here, then bfqq is being activated shortly after the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1283) * last queue. So, if the current burst is also large, we can mark
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1284) * bfqq as belonging to this large burst immediately.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1285) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1286) if (bfqd->large_burst) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1287) bfq_mark_bfqq_in_large_burst(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1288) goto end;
^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) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1292) * If we get here, then a large-burst state has not yet been
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1293) * reached, but bfqq is being activated shortly after the last
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1294) * queue. Then we add bfqq to the burst.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1295) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1296) bfq_add_to_burst(bfqd, bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1297) end:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1298) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1299) * At this point, bfqq either has been added to the current
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1300) * burst or has caused the current burst to terminate and a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1301) * possible new burst to start. In particular, in the second
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1302) * case, bfqq has become the first queue in the possible new
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1303) * burst. In both cases last_ins_in_burst needs to be moved
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1304) * forward.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1305) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1306) bfqd->last_ins_in_burst = jiffies;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1307) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1308)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1309) static int bfq_bfqq_budget_left(struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1310) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1311) struct bfq_entity *entity = &bfqq->entity;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1312)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1313) return entity->budget - entity->service;
^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) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1317) * If enough samples have been computed, return the current max budget
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1318) * stored in bfqd, which is dynamically updated according to the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1319) * estimated disk peak rate; otherwise return the default max budget
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1320) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1321) static int bfq_max_budget(struct bfq_data *bfqd)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1322) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1323) if (bfqd->budgets_assigned < bfq_stats_min_budgets)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1324) return bfq_default_max_budget;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1325) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1326) return bfqd->bfq_max_budget;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1327) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1328)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1329) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1330) * Return min budget, which is a fraction of the current or default
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1331) * max budget (trying with 1/32)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1332) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1333) static int bfq_min_budget(struct bfq_data *bfqd)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1334) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1335) if (bfqd->budgets_assigned < bfq_stats_min_budgets)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1336) return bfq_default_max_budget / 32;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1337) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1338) return bfqd->bfq_max_budget / 32;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1339) }
^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) * The next function, invoked after the input queue bfqq switches from
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1343) * idle to busy, updates the budget of bfqq. The function also tells
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1344) * whether the in-service queue should be expired, by returning
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1345) * true. The purpose of expiring the in-service queue is to give bfqq
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1346) * the chance to possibly preempt the in-service queue, and the reason
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1347) * for preempting the in-service queue is to achieve one of the two
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1348) * goals below.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1349) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1350) * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1351) * expired because it has remained idle. In particular, bfqq may have
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1352) * expired for one of the following two reasons:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1353) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1354) * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1355) * and did not make it to issue a new request before its last
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1356) * request was served;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1357) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1358) * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1359) * a new request before the expiration of the idling-time.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1360) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1361) * Even if bfqq has expired for one of the above reasons, the process
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1362) * associated with the queue may be however issuing requests greedily,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1363) * and thus be sensitive to the bandwidth it receives (bfqq may have
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1364) * remained idle for other reasons: CPU high load, bfqq not enjoying
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1365) * idling, I/O throttling somewhere in the path from the process to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1366) * the I/O scheduler, ...). But if, after every expiration for one of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1367) * the above two reasons, bfqq has to wait for the service of at least
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1368) * one full budget of another queue before being served again, then
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1369) * bfqq is likely to get a much lower bandwidth or resource time than
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1370) * its reserved ones. To address this issue, two countermeasures need
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1371) * to be taken.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1372) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1373) * First, the budget and the timestamps of bfqq need to be updated in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1374) * a special way on bfqq reactivation: they need to be updated as if
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1375) * bfqq did not remain idle and did not expire. In fact, if they are
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1376) * computed as if bfqq expired and remained idle until reactivation,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1377) * then the process associated with bfqq is treated as if, instead of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1378) * being greedy, it stopped issuing requests when bfqq remained idle,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1379) * and restarts issuing requests only on this reactivation. In other
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1380) * words, the scheduler does not help the process recover the "service
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1381) * hole" between bfqq expiration and reactivation. As a consequence,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1382) * the process receives a lower bandwidth than its reserved one. In
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1383) * contrast, to recover this hole, the budget must be updated as if
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1384) * bfqq was not expired at all before this reactivation, i.e., it must
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1385) * be set to the value of the remaining budget when bfqq was
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1386) * expired. Along the same line, timestamps need to be assigned the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1387) * value they had the last time bfqq was selected for service, i.e.,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1388) * before last expiration. Thus timestamps need to be back-shifted
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1389) * with respect to their normal computation (see [1] for more details
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1390) * on this tricky aspect).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1391) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1392) * Secondly, to allow the process to recover the hole, the in-service
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1393) * queue must be expired too, to give bfqq the chance to preempt it
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1394) * immediately. In fact, if bfqq has to wait for a full budget of the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1395) * in-service queue to be completed, then it may become impossible to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1396) * let the process recover the hole, even if the back-shifted
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1397) * timestamps of bfqq are lower than those of the in-service queue. If
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1398) * this happens for most or all of the holes, then the process may not
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1399) * receive its reserved bandwidth. In this respect, it is worth noting
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1400) * that, being the service of outstanding requests unpreemptible, a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1401) * little fraction of the holes may however be unrecoverable, thereby
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1402) * causing a little loss of bandwidth.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1403) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1404) * The last important point is detecting whether bfqq does need this
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1405) * bandwidth recovery. In this respect, the next function deems the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1406) * process associated with bfqq greedy, and thus allows it to recover
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1407) * the hole, if: 1) the process is waiting for the arrival of a new
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1408) * request (which implies that bfqq expired for one of the above two
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1409) * reasons), and 2) such a request has arrived soon. The first
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1410) * condition is controlled through the flag non_blocking_wait_rq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1411) * while the second through the flag arrived_in_time. If both
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1412) * conditions hold, then the function computes the budget in the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1413) * above-described special way, and signals that the in-service queue
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1414) * should be expired. Timestamp back-shifting is done later in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1415) * __bfq_activate_entity.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1416) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1417) * 2. Reduce latency. Even if timestamps are not backshifted to let
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1418) * the process associated with bfqq recover a service hole, bfqq may
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1419) * however happen to have, after being (re)activated, a lower finish
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1420) * timestamp than the in-service queue. That is, the next budget of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1421) * bfqq may have to be completed before the one of the in-service
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1422) * queue. If this is the case, then preempting the in-service queue
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1423) * allows this goal to be achieved, apart from the unpreemptible,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1424) * outstanding requests mentioned above.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1425) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1426) * Unfortunately, regardless of which of the above two goals one wants
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1427) * to achieve, service trees need first to be updated to know whether
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1428) * the in-service queue must be preempted. To have service trees
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1429) * correctly updated, the in-service queue must be expired and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1430) * rescheduled, and bfqq must be scheduled too. This is one of the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1431) * most costly operations (in future versions, the scheduling
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1432) * mechanism may be re-designed in such a way to make it possible to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1433) * know whether preemption is needed without needing to update service
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1434) * trees). In addition, queue preemptions almost always cause random
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1435) * I/O, which may in turn cause loss of throughput. Finally, there may
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1436) * even be no in-service queue when the next function is invoked (so,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1437) * no queue to compare timestamps with). Because of these facts, the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1438) * next function adopts the following simple scheme to avoid costly
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1439) * operations, too frequent preemptions and too many dependencies on
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1440) * the state of the scheduler: it requests the expiration of the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1441) * in-service queue (unconditionally) only for queues that need to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1442) * recover a hole. Then it delegates to other parts of the code the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1443) * responsibility of handling the above case 2.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1444) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1445) static bool bfq_bfqq_update_budg_for_activation(struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1446) struct bfq_queue *bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1447) bool arrived_in_time)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1448) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1449) struct bfq_entity *entity = &bfqq->entity;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1450)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1451) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1452) * In the next compound condition, we check also whether there
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1453) * is some budget left, because otherwise there is no point in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1454) * trying to go on serving bfqq with this same budget: bfqq
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1455) * would be expired immediately after being selected for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1456) * service. This would only cause useless overhead.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1457) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1458) if (bfq_bfqq_non_blocking_wait_rq(bfqq) && arrived_in_time &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1459) bfq_bfqq_budget_left(bfqq) > 0) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1460) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1461) * We do not clear the flag non_blocking_wait_rq here, as
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1462) * the latter is used in bfq_activate_bfqq to signal
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1463) * that timestamps need to be back-shifted (and is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1464) * cleared right after).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1465) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1466)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1467) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1468) * In next assignment we rely on that either
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1469) * entity->service or entity->budget are not updated
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1470) * on expiration if bfqq is empty (see
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1471) * __bfq_bfqq_recalc_budget). Thus both quantities
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1472) * remain unchanged after such an expiration, and the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1473) * following statement therefore assigns to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1474) * entity->budget the remaining budget on such an
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1475) * expiration.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1476) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1477) entity->budget = min_t(unsigned long,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1478) bfq_bfqq_budget_left(bfqq),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1479) bfqq->max_budget);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1480)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1481) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1482) * At this point, we have used entity->service to get
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1483) * the budget left (needed for updating
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1484) * entity->budget). Thus we finally can, and have to,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1485) * reset entity->service. The latter must be reset
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1486) * because bfqq would otherwise be charged again for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1487) * the service it has received during its previous
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1488) * service slot(s).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1489) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1490) entity->service = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1491)
^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) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1496) * We can finally complete expiration, by setting service to 0.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1497) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1498) entity->service = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1499) entity->budget = max_t(unsigned long, bfqq->max_budget,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1500) bfq_serv_to_charge(bfqq->next_rq, bfqq));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1501) bfq_clear_bfqq_non_blocking_wait_rq(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1502) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1503) }
^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) * Return the farthest past time instant according to jiffies
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1507) * macros.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1508) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1509) static unsigned long bfq_smallest_from_now(void)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1510) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1511) return jiffies - MAX_JIFFY_OFFSET;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1512) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1513)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1514) static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1515) struct bfq_queue *bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1516) unsigned int old_wr_coeff,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1517) bool wr_or_deserves_wr,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1518) bool interactive,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1519) bool in_burst,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1520) bool soft_rt)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1521) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1522) if (old_wr_coeff == 1 && wr_or_deserves_wr) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1523) /* start a weight-raising period */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1524) if (interactive) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1525) bfqq->service_from_wr = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1526) bfqq->wr_coeff = bfqd->bfq_wr_coeff;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1527) bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1528) } else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1529) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1530) * No interactive weight raising in progress
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1531) * here: assign minus infinity to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1532) * wr_start_at_switch_to_srt, to make sure
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1533) * that, at the end of the soft-real-time
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1534) * weight raising periods that is starting
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1535) * now, no interactive weight-raising period
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1536) * may be wrongly considered as still in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1537) * progress (and thus actually started by
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1538) * mistake).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1539) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1540) bfqq->wr_start_at_switch_to_srt =
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1541) bfq_smallest_from_now();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1542) bfqq->wr_coeff = bfqd->bfq_wr_coeff *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1543) BFQ_SOFTRT_WEIGHT_FACTOR;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1544) bfqq->wr_cur_max_time =
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1545) bfqd->bfq_wr_rt_max_time;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1546) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1547)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1548) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1549) * If needed, further reduce budget to make sure it is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1550) * close to bfqq's backlog, so as to reduce the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1551) * scheduling-error component due to a too large
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1552) * budget. Do not care about throughput consequences,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1553) * but only about latency. Finally, do not assign a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1554) * too small budget either, to avoid increasing
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1555) * latency by causing too frequent expirations.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1556) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1557) bfqq->entity.budget = min_t(unsigned long,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1558) bfqq->entity.budget,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1559) 2 * bfq_min_budget(bfqd));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1560) } else if (old_wr_coeff > 1) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1561) if (interactive) { /* update wr coeff and duration */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1562) bfqq->wr_coeff = bfqd->bfq_wr_coeff;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1563) bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1564) } else if (in_burst)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1565) bfqq->wr_coeff = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1566) else if (soft_rt) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1567) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1568) * The application is now or still meeting the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1569) * requirements for being deemed soft rt. We
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1570) * can then correctly and safely (re)charge
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1571) * the weight-raising duration for the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1572) * application with the weight-raising
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1573) * duration for soft rt applications.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1574) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1575) * In particular, doing this recharge now, i.e.,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1576) * before the weight-raising period for the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1577) * application finishes, reduces the probability
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1578) * of the following negative scenario:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1579) * 1) the weight of a soft rt application is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1580) * raised at startup (as for any newly
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1581) * created application),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1582) * 2) since the application is not interactive,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1583) * at a certain time weight-raising is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1584) * stopped for the application,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1585) * 3) at that time the application happens to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1586) * still have pending requests, and hence
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1587) * is destined to not have a chance to be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1588) * deemed soft rt before these requests are
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1589) * completed (see the comments to the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1590) * function bfq_bfqq_softrt_next_start()
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1591) * for details on soft rt detection),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1592) * 4) these pending requests experience a high
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1593) * latency because the application is not
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1594) * weight-raised while they are pending.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1595) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1596) if (bfqq->wr_cur_max_time !=
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1597) bfqd->bfq_wr_rt_max_time) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1598) bfqq->wr_start_at_switch_to_srt =
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1599) bfqq->last_wr_start_finish;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1600)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1601) bfqq->wr_cur_max_time =
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1602) bfqd->bfq_wr_rt_max_time;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1603) bfqq->wr_coeff = bfqd->bfq_wr_coeff *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1604) BFQ_SOFTRT_WEIGHT_FACTOR;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1605) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1606) bfqq->last_wr_start_finish = jiffies;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1607) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1608) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1609) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1610)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1611) static bool bfq_bfqq_idle_for_long_time(struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1612) struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1613) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1614) return bfqq->dispatched == 0 &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1615) time_is_before_jiffies(
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1616) bfqq->budget_timeout +
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1617) bfqd->bfq_wr_min_idle_time);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1618) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1619)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1620)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1621) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1622) * Return true if bfqq is in a higher priority class, or has a higher
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1623) * weight than the in-service queue.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1624) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1625) static bool bfq_bfqq_higher_class_or_weight(struct bfq_queue *bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1626) struct bfq_queue *in_serv_bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1627) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1628) int bfqq_weight, in_serv_weight;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1629)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1630) if (bfqq->ioprio_class < in_serv_bfqq->ioprio_class)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1631) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1632)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1633) if (in_serv_bfqq->entity.parent == bfqq->entity.parent) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1634) bfqq_weight = bfqq->entity.weight;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1635) in_serv_weight = in_serv_bfqq->entity.weight;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1636) } else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1637) if (bfqq->entity.parent)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1638) bfqq_weight = bfqq->entity.parent->weight;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1639) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1640) bfqq_weight = bfqq->entity.weight;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1641) if (in_serv_bfqq->entity.parent)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1642) in_serv_weight = in_serv_bfqq->entity.parent->weight;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1643) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1644) in_serv_weight = in_serv_bfqq->entity.weight;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1645) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1646)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1647) return bfqq_weight > in_serv_weight;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1648) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1649)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1650) static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1651) struct bfq_queue *bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1652) int old_wr_coeff,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1653) struct request *rq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1654) bool *interactive)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1655) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1656) bool soft_rt, in_burst, wr_or_deserves_wr,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1657) bfqq_wants_to_preempt,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1658) idle_for_long_time = bfq_bfqq_idle_for_long_time(bfqd, bfqq),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1659) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1660) * See the comments on
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1661) * bfq_bfqq_update_budg_for_activation for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1662) * details on the usage of the next variable.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1663) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1664) arrived_in_time = ktime_get_ns() <=
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1665) bfqq->ttime.last_end_request +
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1666) bfqd->bfq_slice_idle * 3;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1667)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1668)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1669) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1670) * bfqq deserves to be weight-raised if:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1671) * - it is sync,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1672) * - it does not belong to a large burst,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1673) * - it has been idle for enough time or is soft real-time,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1674) * - is linked to a bfq_io_cq (it is not shared in any sense).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1675) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1676) in_burst = bfq_bfqq_in_large_burst(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1677) soft_rt = bfqd->bfq_wr_max_softrt_rate > 0 &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1678) !BFQQ_TOTALLY_SEEKY(bfqq) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1679) !in_burst &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1680) time_is_before_jiffies(bfqq->soft_rt_next_start) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1681) bfqq->dispatched == 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1682) *interactive = !in_burst && idle_for_long_time;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1683) wr_or_deserves_wr = bfqd->low_latency &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1684) (bfqq->wr_coeff > 1 ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1685) (bfq_bfqq_sync(bfqq) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1686) bfqq->bic && (*interactive || soft_rt)));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1687)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1688) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1689) * Using the last flag, update budget and check whether bfqq
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1690) * may want to preempt the in-service queue.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1691) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1692) bfqq_wants_to_preempt =
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1693) bfq_bfqq_update_budg_for_activation(bfqd, bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1694) arrived_in_time);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1695)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1696) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1697) * If bfqq happened to be activated in a burst, but has been
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1698) * idle for much more than an interactive queue, then we
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1699) * assume that, in the overall I/O initiated in the burst, the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1700) * I/O associated with bfqq is finished. So bfqq does not need
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1701) * to be treated as a queue belonging to a burst
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1702) * anymore. Accordingly, we reset bfqq's in_large_burst flag
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1703) * if set, and remove bfqq from the burst list if it's
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1704) * there. We do not decrement burst_size, because the fact
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1705) * that bfqq does not need to belong to the burst list any
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1706) * more does not invalidate the fact that bfqq was created in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1707) * a burst.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1708) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1709) if (likely(!bfq_bfqq_just_created(bfqq)) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1710) idle_for_long_time &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1711) time_is_before_jiffies(
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1712) bfqq->budget_timeout +
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1713) msecs_to_jiffies(10000))) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1714) hlist_del_init(&bfqq->burst_list_node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1715) bfq_clear_bfqq_in_large_burst(bfqq);
^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) bfq_clear_bfqq_just_created(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1719)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1720)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1721) if (!bfq_bfqq_IO_bound(bfqq)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1722) if (arrived_in_time) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1723) bfqq->requests_within_timer++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1724) if (bfqq->requests_within_timer >=
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1725) bfqd->bfq_requests_within_timer)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1726) bfq_mark_bfqq_IO_bound(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1727) } else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1728) bfqq->requests_within_timer = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1729) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1730)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1731) if (bfqd->low_latency) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1732) if (unlikely(time_is_after_jiffies(bfqq->split_time)))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1733) /* wraparound */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1734) bfqq->split_time =
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1735) jiffies - bfqd->bfq_wr_min_idle_time - 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1736)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1737) if (time_is_before_jiffies(bfqq->split_time +
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1738) bfqd->bfq_wr_min_idle_time)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1739) bfq_update_bfqq_wr_on_rq_arrival(bfqd, bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1740) old_wr_coeff,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1741) wr_or_deserves_wr,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1742) *interactive,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1743) in_burst,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1744) soft_rt);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1745)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1746) if (old_wr_coeff != bfqq->wr_coeff)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1747) bfqq->entity.prio_changed = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1748) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1749) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1750)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1751) bfqq->last_idle_bklogged = jiffies;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1752) bfqq->service_from_backlogged = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1753) bfq_clear_bfqq_softrt_update(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1754)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1755) bfq_add_bfqq_busy(bfqd, bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1756)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1757) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1758) * Expire in-service queue only if preemption may be needed
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1759) * for guarantees. In particular, we care only about two
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1760) * cases. The first is that bfqq has to recover a service
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1761) * hole, as explained in the comments on
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1762) * bfq_bfqq_update_budg_for_activation(), i.e., that
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1763) * bfqq_wants_to_preempt is true. However, if bfqq does not
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1764) * carry time-critical I/O, then bfqq's bandwidth is less
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1765) * important than that of queues that carry time-critical I/O.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1766) * So, as a further constraint, we consider this case only if
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1767) * bfqq is at least as weight-raised, i.e., at least as time
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1768) * critical, as the in-service queue.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1769) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1770) * The second case is that bfqq is in a higher priority class,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1771) * or has a higher weight than the in-service queue. If this
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1772) * condition does not hold, we don't care because, even if
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1773) * bfqq does not start to be served immediately, the resulting
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1774) * delay for bfqq's I/O is however lower or much lower than
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1775) * the ideal completion time to be guaranteed to bfqq's I/O.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1776) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1777) * In both cases, preemption is needed only if, according to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1778) * the timestamps of both bfqq and of the in-service queue,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1779) * bfqq actually is the next queue to serve. So, to reduce
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1780) * useless preemptions, the return value of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1781) * next_queue_may_preempt() is considered in the next compound
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1782) * condition too. Yet next_queue_may_preempt() just checks a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1783) * simple, necessary condition for bfqq to be the next queue
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1784) * to serve. In fact, to evaluate a sufficient condition, the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1785) * timestamps of the in-service queue would need to be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1786) * updated, and this operation is quite costly (see the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1787) * comments on bfq_bfqq_update_budg_for_activation()).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1788) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1789) if (bfqd->in_service_queue &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1790) ((bfqq_wants_to_preempt &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1791) bfqq->wr_coeff >= bfqd->in_service_queue->wr_coeff) ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1792) bfq_bfqq_higher_class_or_weight(bfqq, bfqd->in_service_queue)) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1793) next_queue_may_preempt(bfqd))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1794) bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1795) false, BFQQE_PREEMPTED);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1796) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1797)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1798) static void bfq_reset_inject_limit(struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1799) struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1800) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1801) /* invalidate baseline total service time */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1802) bfqq->last_serv_time_ns = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1803)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1804) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1805) * Reset pointer in case we are waiting for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1806) * some request completion.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1807) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1808) bfqd->waited_rq = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1809)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1810) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1811) * If bfqq has a short think time, then start by setting the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1812) * inject limit to 0 prudentially, because the service time of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1813) * an injected I/O request may be higher than the think time
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1814) * of bfqq, and therefore, if one request was injected when
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1815) * bfqq remains empty, this injected request might delay the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1816) * service of the next I/O request for bfqq significantly. In
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1817) * case bfqq can actually tolerate some injection, then the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1818) * adaptive update will however raise the limit soon. This
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1819) * lucky circumstance holds exactly because bfqq has a short
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1820) * think time, and thus, after remaining empty, is likely to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1821) * get new I/O enqueued---and then completed---before being
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1822) * expired. This is the very pattern that gives the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1823) * limit-update algorithm the chance to measure the effect of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1824) * injection on request service times, and then to update the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1825) * limit accordingly.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1826) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1827) * However, in the following special case, the inject limit is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1828) * left to 1 even if the think time is short: bfqq's I/O is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1829) * synchronized with that of some other queue, i.e., bfqq may
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1830) * receive new I/O only after the I/O of the other queue is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1831) * completed. Keeping the inject limit to 1 allows the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1832) * blocking I/O to be served while bfqq is in service. And
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1833) * this is very convenient both for bfqq and for overall
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1834) * throughput, as explained in detail in the comments in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1835) * bfq_update_has_short_ttime().
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1836) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1837) * On the opposite end, if bfqq has a long think time, then
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1838) * start directly by 1, because:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1839) * a) on the bright side, keeping at most one request in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1840) * service in the drive is unlikely to cause any harm to the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1841) * latency of bfqq's requests, as the service time of a single
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1842) * request is likely to be lower than the think time of bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1843) * b) on the downside, after becoming empty, bfqq is likely to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1844) * expire before getting its next request. With this request
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1845) * arrival pattern, it is very hard to sample total service
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1846) * times and update the inject limit accordingly (see comments
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1847) * on bfq_update_inject_limit()). So the limit is likely to be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1848) * never, or at least seldom, updated. As a consequence, by
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1849) * setting the limit to 1, we avoid that no injection ever
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1850) * occurs with bfqq. On the downside, this proactive step
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1851) * further reduces chances to actually compute the baseline
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1852) * total service time. Thus it reduces chances to execute the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1853) * limit-update algorithm and possibly raise the limit to more
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1854) * than 1.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1855) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1856) if (bfq_bfqq_has_short_ttime(bfqq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1857) bfqq->inject_limit = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1858) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1859) bfqq->inject_limit = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1860)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1861) bfqq->decrease_time_jif = jiffies;
^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) static void bfq_add_request(struct request *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1865) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1866) struct bfq_queue *bfqq = RQ_BFQQ(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1867) struct bfq_data *bfqd = bfqq->bfqd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1868) struct request *next_rq, *prev;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1869) unsigned int old_wr_coeff = bfqq->wr_coeff;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1870) bool interactive = false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1871)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1872) bfq_log_bfqq(bfqd, bfqq, "add_request %d", rq_is_sync(rq));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1873) bfqq->queued[rq_is_sync(rq)]++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1874) bfqd->queued++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1875)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1876) if (RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_bfqq_sync(bfqq)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1877) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1878) * Detect whether bfqq's I/O seems synchronized with
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1879) * that of some other queue, i.e., whether bfqq, after
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1880) * remaining empty, happens to receive new I/O only
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1881) * right after some I/O request of the other queue has
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1882) * been completed. We call waker queue the other
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1883) * queue, and we assume, for simplicity, that bfqq may
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1884) * have at most one waker queue.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1885) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1886) * A remarkable throughput boost can be reached by
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1887) * unconditionally injecting the I/O of the waker
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1888) * queue, every time a new bfq_dispatch_request
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1889) * happens to be invoked while I/O is being plugged
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1890) * for bfqq. In addition to boosting throughput, this
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1891) * unblocks bfqq's I/O, thereby improving bandwidth
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1892) * and latency for bfqq. Note that these same results
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1893) * may be achieved with the general injection
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1894) * mechanism, but less effectively. For details on
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1895) * this aspect, see the comments on the choice of the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1896) * queue for injection in bfq_select_queue().
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1897) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1898) * Turning back to the detection of a waker queue, a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1899) * queue Q is deemed as a waker queue for bfqq if, for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1900) * two consecutive times, bfqq happens to become non
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1901) * empty right after a request of Q has been
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1902) * completed. In particular, on the first time, Q is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1903) * tentatively set as a candidate waker queue, while
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1904) * on the second time, the flag
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1905) * bfq_bfqq_has_waker(bfqq) is set to confirm that Q
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1906) * is a waker queue for bfqq. These detection steps
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1907) * are performed only if bfqq has a long think time,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1908) * so as to make it more likely that bfqq's I/O is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1909) * actually being blocked by a synchronization. This
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1910) * last filter, plus the above two-times requirement,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1911) * make false positives less likely.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1912) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1913) * NOTE
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1914) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1915) * The sooner a waker queue is detected, the sooner
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1916) * throughput can be boosted by injecting I/O from the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1917) * waker queue. Fortunately, detection is likely to be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1918) * actually fast, for the following reasons. While
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1919) * blocked by synchronization, bfqq has a long think
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1920) * time. This implies that bfqq's inject limit is at
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1921) * least equal to 1 (see the comments in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1922) * bfq_update_inject_limit()). So, thanks to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1923) * injection, the waker queue is likely to be served
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1924) * during the very first I/O-plugging time interval
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1925) * for bfqq. This triggers the first step of the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1926) * detection mechanism. Thanks again to injection, the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1927) * candidate waker queue is then likely to be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1928) * confirmed no later than during the next
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1929) * I/O-plugging interval for bfqq.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1930) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1931) if (bfqd->last_completed_rq_bfqq &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1932) !bfq_bfqq_has_short_ttime(bfqq) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1933) ktime_get_ns() - bfqd->last_completion <
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1934) 200 * NSEC_PER_USEC) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1935) if (bfqd->last_completed_rq_bfqq != bfqq &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1936) bfqd->last_completed_rq_bfqq !=
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1937) bfqq->waker_bfqq) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1938) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1939) * First synchronization detected with
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1940) * a candidate waker queue, or with a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1941) * different candidate waker queue
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1942) * from the current one.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1943) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1944) bfqq->waker_bfqq = bfqd->last_completed_rq_bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1945)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1946) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1947) * If the waker queue disappears, then
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1948) * bfqq->waker_bfqq must be reset. To
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1949) * this goal, we maintain in each
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1950) * waker queue a list, woken_list, of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1951) * all the queues that reference the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1952) * waker queue through their
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1953) * waker_bfqq pointer. When the waker
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1954) * queue exits, the waker_bfqq pointer
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1955) * of all the queues in the woken_list
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1956) * is reset.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1957) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1958) * In addition, if bfqq is already in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1959) * the woken_list of a waker queue,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1960) * then, before being inserted into
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1961) * the woken_list of a new waker
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1962) * queue, bfqq must be removed from
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1963) * the woken_list of the old waker
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1964) * queue.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1965) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1966) if (!hlist_unhashed(&bfqq->woken_list_node))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1967) hlist_del_init(&bfqq->woken_list_node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1968) hlist_add_head(&bfqq->woken_list_node,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1969) &bfqd->last_completed_rq_bfqq->woken_list);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1970)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1971) bfq_clear_bfqq_has_waker(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1972) } else if (bfqd->last_completed_rq_bfqq ==
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1973) bfqq->waker_bfqq &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1974) !bfq_bfqq_has_waker(bfqq)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1975) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1976) * synchronization with waker_bfqq
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1977) * seen for the second time
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1978) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1979) bfq_mark_bfqq_has_waker(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1980) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1981) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1982)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1983) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1984) * Periodically reset inject limit, to make sure that
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1985) * the latter eventually drops in case workload
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1986) * changes, see step (3) in the comments on
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1987) * bfq_update_inject_limit().
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1988) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1989) if (time_is_before_eq_jiffies(bfqq->decrease_time_jif +
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1990) msecs_to_jiffies(1000)))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1991) bfq_reset_inject_limit(bfqd, bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1992)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1993) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1994) * The following conditions must hold to setup a new
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1995) * sampling of total service time, and then a new
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1996) * update of the inject limit:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1997) * - bfqq is in service, because the total service
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1998) * time is evaluated only for the I/O requests of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 1999) * the queues in service;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2000) * - this is the right occasion to compute or to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2001) * lower the baseline total service time, because
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2002) * there are actually no requests in the drive,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2003) * or
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2004) * the baseline total service time is available, and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2005) * this is the right occasion to compute the other
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2006) * quantity needed to update the inject limit, i.e.,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2007) * the total service time caused by the amount of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2008) * injection allowed by the current value of the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2009) * limit. It is the right occasion because injection
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2010) * has actually been performed during the service
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2011) * hole, and there are still in-flight requests,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2012) * which are very likely to be exactly the injected
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2013) * requests, or part of them;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2014) * - the minimum interval for sampling the total
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2015) * service time and updating the inject limit has
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2016) * elapsed.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2017) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2018) if (bfqq == bfqd->in_service_queue &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2019) (bfqd->rq_in_driver == 0 ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2020) (bfqq->last_serv_time_ns > 0 &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2021) bfqd->rqs_injected && bfqd->rq_in_driver > 0)) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2022) time_is_before_eq_jiffies(bfqq->decrease_time_jif +
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2023) msecs_to_jiffies(10))) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2024) bfqd->last_empty_occupied_ns = ktime_get_ns();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2025) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2026) * Start the state machine for measuring the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2027) * total service time of rq: setting
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2028) * wait_dispatch will cause bfqd->waited_rq to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2029) * be set when rq will be dispatched.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2030) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2031) bfqd->wait_dispatch = true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2032) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2033) * If there is no I/O in service in the drive,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2034) * then possible injection occurred before the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2035) * arrival of rq will not affect the total
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2036) * service time of rq. So the injection limit
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2037) * must not be updated as a function of such
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2038) * total service time, unless new injection
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2039) * occurs before rq is completed. To have the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2040) * injection limit updated only in the latter
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2041) * case, reset rqs_injected here (rqs_injected
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2042) * will be set in case injection is performed
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2043) * on bfqq before rq is completed).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2044) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2045) if (bfqd->rq_in_driver == 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2046) bfqd->rqs_injected = false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2047) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2048) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2049)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2050) elv_rb_add(&bfqq->sort_list, rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2051)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2052) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2053) * Check if this request is a better next-serve candidate.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2054) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2055) prev = bfqq->next_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2056) next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2057) bfqq->next_rq = next_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2058)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2059) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2060) * Adjust priority tree position, if next_rq changes.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2061) * See comments on bfq_pos_tree_add_move() for the unlikely().
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2062) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2063) if (unlikely(!bfqd->nonrot_with_queueing && prev != bfqq->next_rq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2064) bfq_pos_tree_add_move(bfqd, bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2065)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2066) if (!bfq_bfqq_busy(bfqq)) /* switching to busy ... */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2067) bfq_bfqq_handle_idle_busy_switch(bfqd, bfqq, old_wr_coeff,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2068) rq, &interactive);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2069) else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2070) if (bfqd->low_latency && old_wr_coeff == 1 && !rq_is_sync(rq) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2071) time_is_before_jiffies(
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2072) bfqq->last_wr_start_finish +
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2073) bfqd->bfq_wr_min_inter_arr_async)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2074) bfqq->wr_coeff = bfqd->bfq_wr_coeff;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2075) bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2076)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2077) bfqd->wr_busy_queues++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2078) bfqq->entity.prio_changed = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2079) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2080) if (prev != bfqq->next_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2081) bfq_updated_next_req(bfqd, bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2082) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2083)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2084) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2085) * Assign jiffies to last_wr_start_finish in the following
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2086) * cases:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2087) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2088) * . if bfqq is not going to be weight-raised, because, for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2089) * non weight-raised queues, last_wr_start_finish stores the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2090) * arrival time of the last request; as of now, this piece
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2091) * of information is used only for deciding whether to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2092) * weight-raise async queues
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2093) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2094) * . if bfqq is not weight-raised, because, if bfqq is now
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2095) * switching to weight-raised, then last_wr_start_finish
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2096) * stores the time when weight-raising starts
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2097) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2098) * . if bfqq is interactive, because, regardless of whether
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2099) * bfqq is currently weight-raised, the weight-raising
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2100) * period must start or restart (this case is considered
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2101) * separately because it is not detected by the above
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2102) * conditions, if bfqq is already weight-raised)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2103) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2104) * last_wr_start_finish has to be updated also if bfqq is soft
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2105) * real-time, because the weight-raising period is constantly
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2106) * restarted on idle-to-busy transitions for these queues, but
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2107) * this is already done in bfq_bfqq_handle_idle_busy_switch if
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2108) * needed.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2109) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2110) if (bfqd->low_latency &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2111) (old_wr_coeff == 1 || bfqq->wr_coeff == 1 || interactive))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2112) bfqq->last_wr_start_finish = jiffies;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2113) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2114)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2115) static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2116) struct bio *bio,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2117) struct request_queue *q)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2118) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2119) struct bfq_queue *bfqq = bfqd->bio_bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2120)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2121)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2122) if (bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2123) return elv_rb_find(&bfqq->sort_list, bio_end_sector(bio));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2124)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2125) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2126) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2127)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2128) static sector_t get_sdist(sector_t last_pos, struct request *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2129) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2130) if (last_pos)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2131) return abs(blk_rq_pos(rq) - last_pos);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2132)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2133) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2134) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2135)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2136) #if 0 /* Still not clear if we can do without next two functions */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2137) static void bfq_activate_request(struct request_queue *q, struct request *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2138) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2139) struct bfq_data *bfqd = q->elevator->elevator_data;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2140)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2141) bfqd->rq_in_driver++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2142) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2143)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2144) static void bfq_deactivate_request(struct request_queue *q, struct request *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2145) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2146) struct bfq_data *bfqd = q->elevator->elevator_data;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2147)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2148) bfqd->rq_in_driver--;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2149) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2150) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2151)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2152) static void bfq_remove_request(struct request_queue *q,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2153) struct request *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2154) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2155) struct bfq_queue *bfqq = RQ_BFQQ(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2156) struct bfq_data *bfqd = bfqq->bfqd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2157) const int sync = rq_is_sync(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2158)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2159) if (bfqq->next_rq == rq) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2160) bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2161) bfq_updated_next_req(bfqd, bfqq);
^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) if (rq->queuelist.prev != &rq->queuelist)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2165) list_del_init(&rq->queuelist);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2166) bfqq->queued[sync]--;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2167) bfqd->queued--;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2168) elv_rb_del(&bfqq->sort_list, rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2169)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2170) elv_rqhash_del(q, rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2171) if (q->last_merge == rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2172) q->last_merge = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2173)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2174) if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2175) bfqq->next_rq = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2176)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2177) if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->in_service_queue) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2178) bfq_del_bfqq_busy(bfqd, bfqq, false);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2179) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2180) * bfqq emptied. In normal operation, when
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2181) * bfqq is empty, bfqq->entity.service and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2182) * bfqq->entity.budget must contain,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2183) * respectively, the service received and the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2184) * budget used last time bfqq emptied. These
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2185) * facts do not hold in this case, as at least
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2186) * this last removal occurred while bfqq is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2187) * not in service. To avoid inconsistencies,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2188) * reset both bfqq->entity.service and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2189) * bfqq->entity.budget, if bfqq has still a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2190) * process that may issue I/O requests to it.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2191) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2192) bfqq->entity.budget = bfqq->entity.service = 0;
^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) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2196) * Remove queue from request-position tree as it is empty.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2197) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2198) if (bfqq->pos_root) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2199) rb_erase(&bfqq->pos_node, bfqq->pos_root);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2200) bfqq->pos_root = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2201) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2202) } else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2203) /* see comments on bfq_pos_tree_add_move() for the unlikely() */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2204) if (unlikely(!bfqd->nonrot_with_queueing))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2205) bfq_pos_tree_add_move(bfqd, bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2206) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2207)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2208) if (rq->cmd_flags & REQ_META)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2209) bfqq->meta_pending--;
^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)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2213) static bool bfq_bio_merge(struct request_queue *q, struct bio *bio,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2214) unsigned int nr_segs)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2215) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2216) struct bfq_data *bfqd = q->elevator->elevator_data;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2217) struct request *free = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2218) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2219) * bfq_bic_lookup grabs the queue_lock: invoke it now and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2220) * store its return value for later use, to avoid nesting
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2221) * queue_lock inside the bfqd->lock. We assume that the bic
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2222) * returned by bfq_bic_lookup does not go away before
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2223) * bfqd->lock is taken.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2224) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2225) struct bfq_io_cq *bic = bfq_bic_lookup(bfqd, current->io_context, q);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2226) bool ret;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2227)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2228) spin_lock_irq(&bfqd->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2229)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2230) if (bic)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2231) bfqd->bio_bfqq = bic_to_bfqq(bic, op_is_sync(bio->bi_opf));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2232) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2233) bfqd->bio_bfqq = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2234) bfqd->bio_bic = bic;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2235)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2236) ret = blk_mq_sched_try_merge(q, bio, nr_segs, &free);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2237)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2238) if (free)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2239) blk_mq_free_request(free);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2240) spin_unlock_irq(&bfqd->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2241)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2242) return ret;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2243) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2244)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2245) static int bfq_request_merge(struct request_queue *q, struct request **req,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2246) struct bio *bio)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2247) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2248) struct bfq_data *bfqd = q->elevator->elevator_data;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2249) struct request *__rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2250)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2251) __rq = bfq_find_rq_fmerge(bfqd, bio, q);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2252) if (__rq && elv_bio_merge_ok(__rq, bio)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2253) *req = __rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2254)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2255) if (blk_discard_mergable(__rq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2256) return ELEVATOR_DISCARD_MERGE;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2257) return ELEVATOR_FRONT_MERGE;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2258) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2259)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2260) return ELEVATOR_NO_MERGE;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2261) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2262)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2263) static struct bfq_queue *bfq_init_rq(struct request *rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2264)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2265) static void bfq_request_merged(struct request_queue *q, struct request *req,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2266) enum elv_merge type)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2267) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2268) if (type == ELEVATOR_FRONT_MERGE &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2269) rb_prev(&req->rb_node) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2270) blk_rq_pos(req) <
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2271) blk_rq_pos(container_of(rb_prev(&req->rb_node),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2272) struct request, rb_node))) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2273) struct bfq_queue *bfqq = bfq_init_rq(req);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2274) struct bfq_data *bfqd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2275) struct request *prev, *next_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2276)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2277) if (!bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2278) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2279)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2280) bfqd = bfqq->bfqd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2281)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2282) /* Reposition request in its sort_list */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2283) elv_rb_del(&bfqq->sort_list, req);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2284) elv_rb_add(&bfqq->sort_list, req);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2285)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2286) /* Choose next request to be served for bfqq */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2287) prev = bfqq->next_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2288) next_rq = bfq_choose_req(bfqd, bfqq->next_rq, req,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2289) bfqd->last_position);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2290) bfqq->next_rq = next_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2291) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2292) * If next_rq changes, update both the queue's budget to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2293) * fit the new request and the queue's position in its
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2294) * rq_pos_tree.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2295) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2296) if (prev != bfqq->next_rq) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2297) bfq_updated_next_req(bfqd, bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2298) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2299) * See comments on bfq_pos_tree_add_move() for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2300) * the unlikely().
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2301) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2302) if (unlikely(!bfqd->nonrot_with_queueing))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2303) bfq_pos_tree_add_move(bfqd, bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2304) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2305) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2306) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2307)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2308) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2309) * This function is called to notify the scheduler that the requests
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2310) * rq and 'next' have been merged, with 'next' going away. BFQ
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2311) * exploits this hook to address the following issue: if 'next' has a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2312) * fifo_time lower that rq, then the fifo_time of rq must be set to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2313) * the value of 'next', to not forget the greater age of 'next'.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2314) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2315) * NOTE: in this function we assume that rq is in a bfq_queue, basing
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2316) * on that rq is picked from the hash table q->elevator->hash, which,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2317) * in its turn, is filled only with I/O requests present in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2318) * bfq_queues, while BFQ is in use for the request queue q. In fact,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2319) * the function that fills this hash table (elv_rqhash_add) is called
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2320) * only by bfq_insert_request.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2321) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2322) static void bfq_requests_merged(struct request_queue *q, struct request *rq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2323) struct request *next)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2324) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2325) struct bfq_queue *bfqq = bfq_init_rq(rq),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2326) *next_bfqq = bfq_init_rq(next);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2327)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2328) if (!bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2329) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2330)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2331) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2332) * If next and rq belong to the same bfq_queue and next is older
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2333) * than rq, then reposition rq in the fifo (by substituting next
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2334) * with rq). Otherwise, if next and rq belong to different
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2335) * bfq_queues, never reposition rq: in fact, we would have to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2336) * reposition it with respect to next's position in its own fifo,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2337) * which would most certainly be too expensive with respect to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2338) * the benefits.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2339) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2340) if (bfqq == next_bfqq &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2341) !list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2342) next->fifo_time < rq->fifo_time) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2343) list_del_init(&rq->queuelist);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2344) list_replace_init(&next->queuelist, &rq->queuelist);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2345) rq->fifo_time = next->fifo_time;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2346) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2347)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2348) if (bfqq->next_rq == next)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2349) bfqq->next_rq = rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2350)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2351) bfqg_stats_update_io_merged(bfqq_group(bfqq), next->cmd_flags);
^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) /* Must be called with bfqq != NULL */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2355) static void bfq_bfqq_end_wr(struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2356) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2357) if (bfq_bfqq_busy(bfqq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2358) bfqq->bfqd->wr_busy_queues--;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2359) bfqq->wr_coeff = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2360) bfqq->wr_cur_max_time = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2361) bfqq->last_wr_start_finish = jiffies;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2362) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2363) * Trigger a weight change on the next invocation of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2364) * __bfq_entity_update_weight_prio.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2365) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2366) bfqq->entity.prio_changed = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2367) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2368)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2369) void bfq_end_wr_async_queues(struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2370) struct bfq_group *bfqg)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2371) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2372) int i, j;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2373)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2374) for (i = 0; i < 2; i++)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2375) for (j = 0; j < IOPRIO_BE_NR; j++)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2376) if (bfqg->async_bfqq[i][j])
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2377) bfq_bfqq_end_wr(bfqg->async_bfqq[i][j]);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2378) if (bfqg->async_idle_bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2379) bfq_bfqq_end_wr(bfqg->async_idle_bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2380) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2381)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2382) static void bfq_end_wr(struct bfq_data *bfqd)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2383) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2384) struct bfq_queue *bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2385)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2386) spin_lock_irq(&bfqd->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2387)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2388) list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2389) bfq_bfqq_end_wr(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2390) list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2391) bfq_bfqq_end_wr(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2392) bfq_end_wr_async(bfqd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2393)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2394) spin_unlock_irq(&bfqd->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2395) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2396)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2397) static sector_t bfq_io_struct_pos(void *io_struct, bool request)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2398) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2399) if (request)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2400) return blk_rq_pos(io_struct);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2401) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2402) return ((struct bio *)io_struct)->bi_iter.bi_sector;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2403) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2404)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2405) static int bfq_rq_close_to_sector(void *io_struct, bool request,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2406) sector_t sector)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2407) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2408) return abs(bfq_io_struct_pos(io_struct, request) - sector) <=
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2409) BFQQ_CLOSE_THR;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2410) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2411)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2412) static struct bfq_queue *bfqq_find_close(struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2413) struct bfq_queue *bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2414) sector_t sector)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2415) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2416) struct rb_root *root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2417) struct rb_node *parent, *node;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2418) struct bfq_queue *__bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2419)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2420) if (RB_EMPTY_ROOT(root))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2421) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2422)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2423) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2424) * First, if we find a request starting at the end of the last
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2425) * request, choose it.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2426) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2427) __bfqq = bfq_rq_pos_tree_lookup(bfqd, root, sector, &parent, NULL);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2428) if (__bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2429) return __bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2430)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2431) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2432) * If the exact sector wasn't found, the parent of the NULL leaf
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2433) * will contain the closest sector (rq_pos_tree sorted by
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2434) * next_request position).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2435) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2436) __bfqq = rb_entry(parent, struct bfq_queue, pos_node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2437) if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2438) return __bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2439)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2440) if (blk_rq_pos(__bfqq->next_rq) < sector)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2441) node = rb_next(&__bfqq->pos_node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2442) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2443) node = rb_prev(&__bfqq->pos_node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2444) if (!node)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2445) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2446)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2447) __bfqq = rb_entry(node, struct bfq_queue, pos_node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2448) if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2449) return __bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2450)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2451) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2452) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2453)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2454) static struct bfq_queue *bfq_find_close_cooperator(struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2455) struct bfq_queue *cur_bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2456) sector_t sector)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2457) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2458) struct bfq_queue *bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2459)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2460) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2461) * We shall notice if some of the queues are cooperating,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2462) * e.g., working closely on the same area of the device. In
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2463) * that case, we can group them together and: 1) don't waste
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2464) * time idling, and 2) serve the union of their requests in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2465) * the best possible order for throughput.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2466) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2467) bfqq = bfqq_find_close(bfqd, cur_bfqq, sector);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2468) if (!bfqq || bfqq == cur_bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2469) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2470)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2471) return bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2472) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2473)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2474) static struct bfq_queue *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2475) bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2476) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2477) int process_refs, new_process_refs;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2478) struct bfq_queue *__bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2479)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2480) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2481) * If there are no process references on the new_bfqq, then it is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2482) * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2483) * may have dropped their last reference (not just their last process
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2484) * reference).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2485) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2486) if (!bfqq_process_refs(new_bfqq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2487) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2488)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2489) /* Avoid a circular list and skip interim queue merges. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2490) while ((__bfqq = new_bfqq->new_bfqq)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2491) if (__bfqq == bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2492) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2493) new_bfqq = __bfqq;
^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) process_refs = bfqq_process_refs(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2497) new_process_refs = bfqq_process_refs(new_bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2498) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2499) * If the process for the bfqq has gone away, there is no
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2500) * sense in merging the queues.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2501) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2502) if (process_refs == 0 || new_process_refs == 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2503) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2504)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2505) bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d",
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2506) new_bfqq->pid);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2507)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2508) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2509) * Merging is just a redirection: the requests of the process
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2510) * owning one of the two queues are redirected to the other queue.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2511) * The latter queue, in its turn, is set as shared if this is the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2512) * first time that the requests of some process are redirected to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2513) * it.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2514) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2515) * We redirect bfqq to new_bfqq and not the opposite, because
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2516) * we are in the context of the process owning bfqq, thus we
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2517) * have the io_cq of this process. So we can immediately
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2518) * configure this io_cq to redirect the requests of the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2519) * process to new_bfqq. In contrast, the io_cq of new_bfqq is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2520) * not available any more (new_bfqq->bic == NULL).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2521) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2522) * Anyway, even in case new_bfqq coincides with the in-service
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2523) * queue, redirecting requests the in-service queue is the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2524) * best option, as we feed the in-service queue with new
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2525) * requests close to the last request served and, by doing so,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2526) * are likely to increase the throughput.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2527) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2528) bfqq->new_bfqq = new_bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2529) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2530) * The above assignment schedules the following redirections:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2531) * each time some I/O for bfqq arrives, the process that
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2532) * generated that I/O is disassociated from bfqq and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2533) * associated with new_bfqq. Here we increases new_bfqq->ref
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2534) * in advance, adding the number of processes that are
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2535) * expected to be associated with new_bfqq as they happen to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2536) * issue I/O.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2537) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2538) new_bfqq->ref += process_refs;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2539) return new_bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2540) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2541)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2542) static bool bfq_may_be_close_cooperator(struct bfq_queue *bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2543) struct bfq_queue *new_bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2544) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2545) if (bfq_too_late_for_merging(new_bfqq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2546) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2547)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2548) if (bfq_class_idle(bfqq) || bfq_class_idle(new_bfqq) ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2549) (bfqq->ioprio_class != new_bfqq->ioprio_class))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2550) return false;
^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) * If either of the queues has already been detected as seeky,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2554) * then merging it with the other queue is unlikely to lead to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2555) * sequential I/O.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2556) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2557) if (BFQQ_SEEKY(bfqq) || BFQQ_SEEKY(new_bfqq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2558) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2559)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2560) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2561) * Interleaved I/O is known to be done by (some) applications
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2562) * only for reads, so it does not make sense to merge async
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2563) * queues.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2564) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2565) if (!bfq_bfqq_sync(bfqq) || !bfq_bfqq_sync(new_bfqq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2566) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2567)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2568) return true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2569) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2570)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2571) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2572) * Attempt to schedule a merge of bfqq with the currently in-service
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2573) * queue or with a close queue among the scheduled queues. Return
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2574) * NULL if no merge was scheduled, a pointer to the shared bfq_queue
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2575) * structure otherwise.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2576) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2577) * The OOM queue is not allowed to participate to cooperation: in fact, since
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2578) * the requests temporarily redirected to the OOM queue could be redirected
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2579) * again to dedicated queues at any time, the state needed to correctly
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2580) * handle merging with the OOM queue would be quite complex and expensive
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2581) * to maintain. Besides, in such a critical condition as an out of memory,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2582) * the benefits of queue merging may be little relevant, or even negligible.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2583) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2584) * WARNING: queue merging may impair fairness among non-weight raised
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2585) * queues, for at least two reasons: 1) the original weight of a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2586) * merged queue may change during the merged state, 2) even being the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2587) * weight the same, a merged queue may be bloated with many more
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2588) * requests than the ones produced by its originally-associated
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2589) * process.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2590) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2591) static struct bfq_queue *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2592) bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2593) void *io_struct, bool request)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2594) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2595) struct bfq_queue *in_service_bfqq, *new_bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2596)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2597) /* if a merge has already been setup, then proceed with that first */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2598) if (bfqq->new_bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2599) return bfqq->new_bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2600)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2601) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2602) * Do not perform queue merging if the device is non
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2603) * rotational and performs internal queueing. In fact, such a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2604) * device reaches a high speed through internal parallelism
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2605) * and pipelining. This means that, to reach a high
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2606) * throughput, it must have many requests enqueued at the same
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2607) * time. But, in this configuration, the internal scheduling
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2608) * algorithm of the device does exactly the job of queue
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2609) * merging: it reorders requests so as to obtain as much as
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2610) * possible a sequential I/O pattern. As a consequence, with
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2611) * the workload generated by processes doing interleaved I/O,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2612) * the throughput reached by the device is likely to be the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2613) * same, with and without queue merging.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2614) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2615) * Disabling merging also provides a remarkable benefit in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2616) * terms of throughput. Merging tends to make many workloads
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2617) * artificially more uneven, because of shared queues
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2618) * remaining non empty for incomparably more time than
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2619) * non-merged queues. This may accentuate workload
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2620) * asymmetries. For example, if one of the queues in a set of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2621) * merged queues has a higher weight than a normal queue, then
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2622) * the shared queue may inherit such a high weight and, by
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2623) * staying almost always active, may force BFQ to perform I/O
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2624) * plugging most of the time. This evidently makes it harder
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2625) * for BFQ to let the device reach a high throughput.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2626) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2627) * Finally, the likely() macro below is not used because one
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2628) * of the two branches is more likely than the other, but to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2629) * have the code path after the following if() executed as
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2630) * fast as possible for the case of a non rotational device
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2631) * with queueing. We want it because this is the fastest kind
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2632) * of device. On the opposite end, the likely() may lengthen
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2633) * the execution time of BFQ for the case of slower devices
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2634) * (rotational or at least without queueing). But in this case
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2635) * the execution time of BFQ matters very little, if not at
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2636) * all.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2637) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2638) if (likely(bfqd->nonrot_with_queueing))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2639) return NULL;
^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) * Prevent bfqq from being merged if it has been created too
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2643) * long ago. The idea is that true cooperating processes, and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2644) * thus their associated bfq_queues, are supposed to be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2645) * created shortly after each other. This is the case, e.g.,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2646) * for KVM/QEMU and dump I/O threads. Basing on this
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2647) * assumption, the following filtering greatly reduces the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2648) * probability that two non-cooperating processes, which just
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2649) * happen to do close I/O for some short time interval, have
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2650) * their queues merged by mistake.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2651) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2652) if (bfq_too_late_for_merging(bfqq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2653) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2654)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2655) if (!io_struct || unlikely(bfqq == &bfqd->oom_bfqq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2656) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2657)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2658) /* If there is only one backlogged queue, don't search. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2659) if (bfq_tot_busy_queues(bfqd) == 1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2660) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2661)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2662) in_service_bfqq = bfqd->in_service_queue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2663)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2664) if (in_service_bfqq && in_service_bfqq != bfqq &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2665) likely(in_service_bfqq != &bfqd->oom_bfqq) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2666) bfq_rq_close_to_sector(io_struct, request,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2667) bfqd->in_serv_last_pos) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2668) bfqq->entity.parent == in_service_bfqq->entity.parent &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2669) bfq_may_be_close_cooperator(bfqq, in_service_bfqq)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2670) new_bfqq = bfq_setup_merge(bfqq, in_service_bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2671) if (new_bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2672) return new_bfqq;
^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) * Check whether there is a cooperator among currently scheduled
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2676) * queues. The only thing we need is that the bio/request is not
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2677) * NULL, as we need it to establish whether a cooperator exists.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2678) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2679) new_bfqq = bfq_find_close_cooperator(bfqd, bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2680) bfq_io_struct_pos(io_struct, request));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2681)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2682) if (new_bfqq && likely(new_bfqq != &bfqd->oom_bfqq) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2683) bfq_may_be_close_cooperator(bfqq, new_bfqq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2684) return bfq_setup_merge(bfqq, new_bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2685)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2686) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2687) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2688)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2689) static void bfq_bfqq_save_state(struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2690) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2691) struct bfq_io_cq *bic = bfqq->bic;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2692)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2693) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2694) * If !bfqq->bic, the queue is already shared or its requests
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2695) * have already been redirected to a shared queue; both idle window
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2696) * and weight raising state have already been saved. Do nothing.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2697) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2698) if (!bic)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2699) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2700)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2701) bic->saved_weight = bfqq->entity.orig_weight;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2702) bic->saved_ttime = bfqq->ttime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2703) bic->saved_has_short_ttime = bfq_bfqq_has_short_ttime(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2704) bic->saved_IO_bound = bfq_bfqq_IO_bound(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2705) bic->saved_in_large_burst = bfq_bfqq_in_large_burst(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2706) bic->was_in_burst_list = !hlist_unhashed(&bfqq->burst_list_node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2707) if (unlikely(bfq_bfqq_just_created(bfqq) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2708) !bfq_bfqq_in_large_burst(bfqq) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2709) bfqq->bfqd->low_latency)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2710) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2711) * bfqq being merged right after being created: bfqq
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2712) * would have deserved interactive weight raising, but
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2713) * did not make it to be set in a weight-raised state,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2714) * because of this early merge. Store directly the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2715) * weight-raising state that would have been assigned
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2716) * to bfqq, so that to avoid that bfqq unjustly fails
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2717) * to enjoy weight raising if split soon.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2718) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2719) bic->saved_wr_coeff = bfqq->bfqd->bfq_wr_coeff;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2720) bic->saved_wr_start_at_switch_to_srt = bfq_smallest_from_now();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2721) bic->saved_wr_cur_max_time = bfq_wr_duration(bfqq->bfqd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2722) bic->saved_last_wr_start_finish = jiffies;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2723) } else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2724) bic->saved_wr_coeff = bfqq->wr_coeff;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2725) bic->saved_wr_start_at_switch_to_srt =
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2726) bfqq->wr_start_at_switch_to_srt;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2727) bic->saved_last_wr_start_finish = bfqq->last_wr_start_finish;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2728) bic->saved_wr_cur_max_time = bfqq->wr_cur_max_time;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2729) }
^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) void bfq_release_process_ref(struct bfq_data *bfqd, struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2733) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2734) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2735) * To prevent bfqq's service guarantees from being violated,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2736) * bfqq may be left busy, i.e., queued for service, even if
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2737) * empty (see comments in __bfq_bfqq_expire() for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2738) * details). But, if no process will send requests to bfqq any
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2739) * longer, then there is no point in keeping bfqq queued for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2740) * service. In addition, keeping bfqq queued for service, but
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2741) * with no process ref any longer, may have caused bfqq to be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2742) * freed when dequeued from service. But this is assumed to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2743) * never happen.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2744) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2745) if (bfq_bfqq_busy(bfqq) && RB_EMPTY_ROOT(&bfqq->sort_list) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2746) bfqq != bfqd->in_service_queue)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2747) bfq_del_bfqq_busy(bfqd, bfqq, false);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2748)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2749) bfq_put_queue(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2750) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2751)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2752) static void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2753) bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2754) struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2755) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2756) bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu",
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2757) (unsigned long)new_bfqq->pid);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2758) /* Save weight raising and idle window of the merged queues */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2759) bfq_bfqq_save_state(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2760) bfq_bfqq_save_state(new_bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2761) if (bfq_bfqq_IO_bound(bfqq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2762) bfq_mark_bfqq_IO_bound(new_bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2763) bfq_clear_bfqq_IO_bound(bfqq);
^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) * If bfqq is weight-raised, then let new_bfqq inherit
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2767) * weight-raising. To reduce false positives, neglect the case
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2768) * where bfqq has just been created, but has not yet made it
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2769) * to be weight-raised (which may happen because EQM may merge
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2770) * bfqq even before bfq_add_request is executed for the first
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2771) * time for bfqq). Handling this case would however be very
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2772) * easy, thanks to the flag just_created.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2773) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2774) if (new_bfqq->wr_coeff == 1 && bfqq->wr_coeff > 1) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2775) new_bfqq->wr_coeff = bfqq->wr_coeff;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2776) new_bfqq->wr_cur_max_time = bfqq->wr_cur_max_time;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2777) new_bfqq->last_wr_start_finish = bfqq->last_wr_start_finish;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2778) new_bfqq->wr_start_at_switch_to_srt =
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2779) bfqq->wr_start_at_switch_to_srt;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2780) if (bfq_bfqq_busy(new_bfqq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2781) bfqd->wr_busy_queues++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2782) new_bfqq->entity.prio_changed = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2783) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2784)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2785) if (bfqq->wr_coeff > 1) { /* bfqq has given its wr to new_bfqq */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2786) bfqq->wr_coeff = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2787) bfqq->entity.prio_changed = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2788) if (bfq_bfqq_busy(bfqq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2789) bfqd->wr_busy_queues--;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2790) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2791)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2792) bfq_log_bfqq(bfqd, new_bfqq, "merge_bfqqs: wr_busy %d",
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2793) bfqd->wr_busy_queues);
^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) * Merge queues (that is, let bic redirect its requests to new_bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2797) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2798) bic_set_bfqq(bic, new_bfqq, 1);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2799) bfq_mark_bfqq_coop(new_bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2800) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2801) * new_bfqq now belongs to at least two bics (it is a shared queue):
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2802) * set new_bfqq->bic to NULL. bfqq either:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2803) * - does not belong to any bic any more, and hence bfqq->bic must
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2804) * be set to NULL, or
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2805) * - is a queue whose owning bics have already been redirected to a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2806) * different queue, hence the queue is destined to not belong to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2807) * any bic soon and bfqq->bic is already NULL (therefore the next
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2808) * assignment causes no harm).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2809) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2810) new_bfqq->bic = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2811) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2812) * If the queue is shared, the pid is the pid of one of the associated
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2813) * processes. Which pid depends on the exact sequence of merge events
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2814) * the queue underwent. So printing such a pid is useless and confusing
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2815) * because it reports a random pid between those of the associated
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2816) * processes.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2817) * We mark such a queue with a pid -1, and then print SHARED instead of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2818) * a pid in logging messages.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2819) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2820) new_bfqq->pid = -1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2821) bfqq->bic = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2822) bfq_release_process_ref(bfqd, bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2823) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2824)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2825) static bool bfq_allow_bio_merge(struct request_queue *q, struct request *rq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2826) struct bio *bio)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2827) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2828) struct bfq_data *bfqd = q->elevator->elevator_data;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2829) bool is_sync = op_is_sync(bio->bi_opf);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2830) struct bfq_queue *bfqq = bfqd->bio_bfqq, *new_bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2831)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2832) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2833) * Disallow merge of a sync bio into an async request.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2834) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2835) if (is_sync && !rq_is_sync(rq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2836) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2837)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2838) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2839) * Lookup the bfqq that this bio will be queued with. Allow
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2840) * merge only if rq is queued there.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2841) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2842) if (!bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2843) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2844)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2845) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2846) * We take advantage of this function to perform an early merge
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2847) * of the queues of possible cooperating processes.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2848) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2849) new_bfqq = bfq_setup_cooperator(bfqd, bfqq, bio, false);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2850) if (new_bfqq) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2851) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2852) * bic still points to bfqq, then it has not yet been
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2853) * redirected to some other bfq_queue, and a queue
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2854) * merge between bfqq and new_bfqq can be safely
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2855) * fulfilled, i.e., bic can be redirected to new_bfqq
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2856) * and bfqq can be put.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2857) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2858) bfq_merge_bfqqs(bfqd, bfqd->bio_bic, bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2859) new_bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2860) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2861) * If we get here, bio will be queued into new_queue,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2862) * so use new_bfqq to decide whether bio and rq can be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2863) * merged.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2864) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2865) bfqq = new_bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2866)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2867) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2868) * Change also bqfd->bio_bfqq, as
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2869) * bfqd->bio_bic now points to new_bfqq, and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2870) * this function may be invoked again (and then may
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2871) * use again bqfd->bio_bfqq).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2872) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2873) bfqd->bio_bfqq = bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2874) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2875)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2876) return bfqq == RQ_BFQQ(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2877) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2878)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2879) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2880) * Set the maximum time for the in-service queue to consume its
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2881) * budget. This prevents seeky processes from lowering the throughput.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2882) * In practice, a time-slice service scheme is used with seeky
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2883) * processes.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2884) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2885) static void bfq_set_budget_timeout(struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2886) struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2887) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2888) unsigned int timeout_coeff;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2889)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2890) if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2891) timeout_coeff = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2892) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2893) timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2894)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2895) bfqd->last_budget_start = ktime_get();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2896)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2897) bfqq->budget_timeout = jiffies +
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2898) bfqd->bfq_timeout * timeout_coeff;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2899) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2900)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2901) static void __bfq_set_in_service_queue(struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2902) struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2903) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2904) if (bfqq) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2905) bfq_clear_bfqq_fifo_expire(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2906)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2907) bfqd->budgets_assigned = (bfqd->budgets_assigned * 7 + 256) / 8;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2908)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2909) if (time_is_before_jiffies(bfqq->last_wr_start_finish) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2910) bfqq->wr_coeff > 1 &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2911) bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2912) time_is_before_jiffies(bfqq->budget_timeout)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2913) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2914) * For soft real-time queues, move the start
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2915) * of the weight-raising period forward by the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2916) * time the queue has not received any
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2917) * service. Otherwise, a relatively long
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2918) * service delay is likely to cause the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2919) * weight-raising period of the queue to end,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2920) * because of the short duration of the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2921) * weight-raising period of a soft real-time
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2922) * queue. It is worth noting that this move
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2923) * is not so dangerous for the other queues,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2924) * because soft real-time queues are not
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2925) * greedy.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2926) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2927) * To not add a further variable, we use the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2928) * overloaded field budget_timeout to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2929) * determine for how long the queue has not
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2930) * received service, i.e., how much time has
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2931) * elapsed since the queue expired. However,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2932) * this is a little imprecise, because
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2933) * budget_timeout is set to jiffies if bfqq
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2934) * not only expires, but also remains with no
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2935) * request.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2936) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2937) if (time_after(bfqq->budget_timeout,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2938) bfqq->last_wr_start_finish))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2939) bfqq->last_wr_start_finish +=
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2940) jiffies - bfqq->budget_timeout;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2941) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2942) bfqq->last_wr_start_finish = jiffies;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2943) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2944)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2945) bfq_set_budget_timeout(bfqd, bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2946) bfq_log_bfqq(bfqd, bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2947) "set_in_service_queue, cur-budget = %d",
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2948) bfqq->entity.budget);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2949) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2950)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2951) bfqd->in_service_queue = bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2952) bfqd->in_serv_last_pos = 0;
^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) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2956) * Get and set a new queue for service.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2957) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2958) static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2959) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2960) struct bfq_queue *bfqq = bfq_get_next_queue(bfqd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2961)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2962) __bfq_set_in_service_queue(bfqd, bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2963) return bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2964) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2965)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2966) static void bfq_arm_slice_timer(struct bfq_data *bfqd)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2967) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2968) struct bfq_queue *bfqq = bfqd->in_service_queue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2969) u32 sl;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2970)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2971) bfq_mark_bfqq_wait_request(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2972)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2973) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2974) * We don't want to idle for seeks, but we do want to allow
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2975) * fair distribution of slice time for a process doing back-to-back
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2976) * seeks. So allow a little bit of time for him to submit a new rq.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2977) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2978) sl = bfqd->bfq_slice_idle;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2979) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2980) * Unless the queue is being weight-raised or the scenario is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2981) * asymmetric, grant only minimum idle time if the queue
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2982) * is seeky. A long idling is preserved for a weight-raised
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2983) * queue, or, more in general, in an asymmetric scenario,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2984) * because a long idling is needed for guaranteeing to a queue
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2985) * its reserved share of the throughput (in particular, it is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2986) * needed if the queue has a higher weight than some other
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2987) * queue).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2988) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2989) if (BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2990) !bfq_asymmetric_scenario(bfqd, bfqq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2991) sl = min_t(u64, sl, BFQ_MIN_TT);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2992) else if (bfqq->wr_coeff > 1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2993) sl = max_t(u32, sl, 20ULL * NSEC_PER_MSEC);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2994)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2995) bfqd->last_idling_start = ktime_get();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2996) bfqd->last_idling_start_jiffies = jiffies;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2997)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2998) hrtimer_start(&bfqd->idle_slice_timer, ns_to_ktime(sl),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 2999) HRTIMER_MODE_REL);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3000) bfqg_stats_set_start_idle_time(bfqq_group(bfqq));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3001) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3002)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3003) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3004) * In autotuning mode, max_budget is dynamically recomputed as the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3005) * amount of sectors transferred in timeout at the estimated peak
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3006) * rate. This enables BFQ to utilize a full timeslice with a full
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3007) * budget, even if the in-service queue is served at peak rate. And
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3008) * this maximises throughput with sequential workloads.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3009) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3010) static unsigned long bfq_calc_max_budget(struct bfq_data *bfqd)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3011) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3012) return (u64)bfqd->peak_rate * USEC_PER_MSEC *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3013) jiffies_to_msecs(bfqd->bfq_timeout)>>BFQ_RATE_SHIFT;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3014) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3015)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3016) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3017) * Update parameters related to throughput and responsiveness, as a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3018) * function of the estimated peak rate. See comments on
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3019) * bfq_calc_max_budget(), and on the ref_wr_duration array.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3020) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3021) static void update_thr_responsiveness_params(struct bfq_data *bfqd)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3022) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3023) if (bfqd->bfq_user_max_budget == 0) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3024) bfqd->bfq_max_budget =
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3025) bfq_calc_max_budget(bfqd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3026) bfq_log(bfqd, "new max_budget = %d", bfqd->bfq_max_budget);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3027) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3028) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3029)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3030) static void bfq_reset_rate_computation(struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3031) struct request *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3032) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3033) if (rq != NULL) { /* new rq dispatch now, reset accordingly */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3034) bfqd->last_dispatch = bfqd->first_dispatch = ktime_get_ns();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3035) bfqd->peak_rate_samples = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3036) bfqd->sequential_samples = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3037) bfqd->tot_sectors_dispatched = bfqd->last_rq_max_size =
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3038) blk_rq_sectors(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3039) } else /* no new rq dispatched, just reset the number of samples */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3040) bfqd->peak_rate_samples = 0; /* full re-init on next disp. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3041)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3042) bfq_log(bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3043) "reset_rate_computation at end, sample %u/%u tot_sects %llu",
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3044) bfqd->peak_rate_samples, bfqd->sequential_samples,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3045) bfqd->tot_sectors_dispatched);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3046) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3047)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3048) static void bfq_update_rate_reset(struct bfq_data *bfqd, struct request *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3049) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3050) u32 rate, weight, divisor;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3051)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3052) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3053) * For the convergence property to hold (see comments on
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3054) * bfq_update_peak_rate()) and for the assessment to be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3055) * reliable, a minimum number of samples must be present, and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3056) * a minimum amount of time must have elapsed. If not so, do
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3057) * not compute new rate. Just reset parameters, to get ready
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3058) * for a new evaluation attempt.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3059) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3060) if (bfqd->peak_rate_samples < BFQ_RATE_MIN_SAMPLES ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3061) bfqd->delta_from_first < BFQ_RATE_MIN_INTERVAL)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3062) goto reset_computation;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3063)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3064) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3065) * If a new request completion has occurred after last
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3066) * dispatch, then, to approximate the rate at which requests
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3067) * have been served by the device, it is more precise to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3068) * extend the observation interval to the last completion.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3069) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3070) bfqd->delta_from_first =
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3071) max_t(u64, bfqd->delta_from_first,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3072) bfqd->last_completion - bfqd->first_dispatch);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3073)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3074) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3075) * Rate computed in sects/usec, and not sects/nsec, for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3076) * precision issues.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3077) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3078) rate = div64_ul(bfqd->tot_sectors_dispatched<<BFQ_RATE_SHIFT,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3079) div_u64(bfqd->delta_from_first, NSEC_PER_USEC));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3080)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3081) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3082) * Peak rate not updated if:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3083) * - the percentage of sequential dispatches is below 3/4 of the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3084) * total, and rate is below the current estimated peak rate
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3085) * - rate is unreasonably high (> 20M sectors/sec)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3086) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3087) if ((bfqd->sequential_samples < (3 * bfqd->peak_rate_samples)>>2 &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3088) rate <= bfqd->peak_rate) ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3089) rate > 20<<BFQ_RATE_SHIFT)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3090) goto reset_computation;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3091)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3092) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3093) * We have to update the peak rate, at last! To this purpose,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3094) * we use a low-pass filter. We compute the smoothing constant
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3095) * of the filter as a function of the 'weight' of the new
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3096) * measured rate.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3097) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3098) * As can be seen in next formulas, we define this weight as a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3099) * quantity proportional to how sequential the workload is,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3100) * and to how long the observation time interval is.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3101) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3102) * The weight runs from 0 to 8. The maximum value of the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3103) * weight, 8, yields the minimum value for the smoothing
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3104) * constant. At this minimum value for the smoothing constant,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3105) * the measured rate contributes for half of the next value of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3106) * the estimated peak rate.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3107) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3108) * So, the first step is to compute the weight as a function
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3109) * of how sequential the workload is. Note that the weight
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3110) * cannot reach 9, because bfqd->sequential_samples cannot
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3111) * become equal to bfqd->peak_rate_samples, which, in its
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3112) * turn, holds true because bfqd->sequential_samples is not
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3113) * incremented for the first sample.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3114) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3115) weight = (9 * bfqd->sequential_samples) / bfqd->peak_rate_samples;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3116)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3117) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3118) * Second step: further refine the weight as a function of the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3119) * duration of the observation interval.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3120) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3121) weight = min_t(u32, 8,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3122) div_u64(weight * bfqd->delta_from_first,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3123) BFQ_RATE_REF_INTERVAL));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3124)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3125) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3126) * Divisor ranging from 10, for minimum weight, to 2, for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3127) * maximum weight.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3128) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3129) divisor = 10 - weight;
^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) * Finally, update peak rate:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3133) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3134) * peak_rate = peak_rate * (divisor-1) / divisor + rate / divisor
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3135) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3136) bfqd->peak_rate *= divisor-1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3137) bfqd->peak_rate /= divisor;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3138) rate /= divisor; /* smoothing constant alpha = 1/divisor */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3139)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3140) bfqd->peak_rate += rate;
^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) * For a very slow device, bfqd->peak_rate can reach 0 (see
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3144) * the minimum representable values reported in the comments
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3145) * on BFQ_RATE_SHIFT). Push to 1 if this happens, to avoid
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3146) * divisions by zero where bfqd->peak_rate is used as a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3147) * divisor.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3148) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3149) bfqd->peak_rate = max_t(u32, 1, bfqd->peak_rate);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3150)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3151) update_thr_responsiveness_params(bfqd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3152)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3153) reset_computation:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3154) bfq_reset_rate_computation(bfqd, rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3155) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3156)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3157) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3158) * Update the read/write peak rate (the main quantity used for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3159) * auto-tuning, see update_thr_responsiveness_params()).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3160) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3161) * It is not trivial to estimate the peak rate (correctly): because of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3162) * the presence of sw and hw queues between the scheduler and the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3163) * device components that finally serve I/O requests, it is hard to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3164) * say exactly when a given dispatched request is served inside the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3165) * device, and for how long. As a consequence, it is hard to know
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3166) * precisely at what rate a given set of requests is actually served
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3167) * by the device.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3168) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3169) * On the opposite end, the dispatch time of any request is trivially
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3170) * available, and, from this piece of information, the "dispatch rate"
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3171) * of requests can be immediately computed. So, the idea in the next
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3172) * function is to use what is known, namely request dispatch times
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3173) * (plus, when useful, request completion times), to estimate what is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3174) * unknown, namely in-device request service rate.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3175) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3176) * The main issue is that, because of the above facts, the rate at
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3177) * which a certain set of requests is dispatched over a certain time
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3178) * interval can vary greatly with respect to the rate at which the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3179) * same requests are then served. But, since the size of any
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3180) * intermediate queue is limited, and the service scheme is lossless
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3181) * (no request is silently dropped), the following obvious convergence
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3182) * property holds: the number of requests dispatched MUST become
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3183) * closer and closer to the number of requests completed as the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3184) * observation interval grows. This is the key property used in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3185) * the next function to estimate the peak service rate as a function
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3186) * of the observed dispatch rate. The function assumes to be invoked
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3187) * on every request dispatch.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3188) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3189) static void bfq_update_peak_rate(struct bfq_data *bfqd, struct request *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3190) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3191) u64 now_ns = ktime_get_ns();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3192)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3193) if (bfqd->peak_rate_samples == 0) { /* first dispatch */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3194) bfq_log(bfqd, "update_peak_rate: goto reset, samples %d",
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3195) bfqd->peak_rate_samples);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3196) bfq_reset_rate_computation(bfqd, rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3197) goto update_last_values; /* will add one sample */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3198) }
^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) * Device idle for very long: the observation interval lasting
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3202) * up to this dispatch cannot be a valid observation interval
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3203) * for computing a new peak rate (similarly to the late-
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3204) * completion event in bfq_completed_request()). Go to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3205) * update_rate_and_reset to have the following three steps
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3206) * taken:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3207) * - close the observation interval at the last (previous)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3208) * request dispatch or completion
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3209) * - compute rate, if possible, for that observation interval
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3210) * - start a new observation interval with this dispatch
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3211) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3212) if (now_ns - bfqd->last_dispatch > 100*NSEC_PER_MSEC &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3213) bfqd->rq_in_driver == 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3214) goto update_rate_and_reset;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3215)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3216) /* Update sampling information */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3217) bfqd->peak_rate_samples++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3218)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3219) if ((bfqd->rq_in_driver > 0 ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3220) now_ns - bfqd->last_completion < BFQ_MIN_TT)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3221) && !BFQ_RQ_SEEKY(bfqd, bfqd->last_position, rq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3222) bfqd->sequential_samples++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3223)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3224) bfqd->tot_sectors_dispatched += blk_rq_sectors(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3225)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3226) /* Reset max observed rq size every 32 dispatches */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3227) if (likely(bfqd->peak_rate_samples % 32))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3228) bfqd->last_rq_max_size =
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3229) max_t(u32, blk_rq_sectors(rq), bfqd->last_rq_max_size);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3230) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3231) bfqd->last_rq_max_size = blk_rq_sectors(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3232)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3233) bfqd->delta_from_first = now_ns - bfqd->first_dispatch;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3234)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3235) /* Target observation interval not yet reached, go on sampling */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3236) if (bfqd->delta_from_first < BFQ_RATE_REF_INTERVAL)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3237) goto update_last_values;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3238)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3239) update_rate_and_reset:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3240) bfq_update_rate_reset(bfqd, rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3241) update_last_values:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3242) bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3243) if (RQ_BFQQ(rq) == bfqd->in_service_queue)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3244) bfqd->in_serv_last_pos = bfqd->last_position;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3245) bfqd->last_dispatch = now_ns;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3246) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3247)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3248) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3249) * Remove request from internal lists.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3250) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3251) static void bfq_dispatch_remove(struct request_queue *q, struct request *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3252) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3253) struct bfq_queue *bfqq = RQ_BFQQ(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) * For consistency, the next instruction should have been
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3257) * executed after removing the request from the queue and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3258) * dispatching it. We execute instead this instruction before
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3259) * bfq_remove_request() (and hence introduce a temporary
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3260) * inconsistency), for efficiency. In fact, should this
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3261) * dispatch occur for a non in-service bfqq, this anticipated
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3262) * increment prevents two counters related to bfqq->dispatched
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3263) * from risking to be, first, uselessly decremented, and then
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3264) * incremented again when the (new) value of bfqq->dispatched
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3265) * happens to be taken into account.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3266) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3267) bfqq->dispatched++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3268) bfq_update_peak_rate(q->elevator->elevator_data, rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3269)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3270) bfq_remove_request(q, rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3271) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3272)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3273) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3274) * There is a case where idling does not have to be performed for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3275) * throughput concerns, but to preserve the throughput share of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3276) * the process associated with bfqq.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3277) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3278) * To introduce this case, we can note that allowing the drive
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3279) * to enqueue more than one request at a time, and hence
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3280) * delegating de facto final scheduling decisions to the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3281) * drive's internal scheduler, entails loss of control on the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3282) * actual request service order. In particular, the critical
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3283) * situation is when requests from different processes happen
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3284) * to be present, at the same time, in the internal queue(s)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3285) * of the drive. In such a situation, the drive, by deciding
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3286) * the service order of the internally-queued requests, does
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3287) * determine also the actual throughput distribution among
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3288) * these processes. But the drive typically has no notion or
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3289) * concern about per-process throughput distribution, and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3290) * makes its decisions only on a per-request basis. Therefore,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3291) * the service distribution enforced by the drive's internal
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3292) * scheduler is likely to coincide with the desired throughput
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3293) * distribution only in a completely symmetric, or favorably
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3294) * skewed scenario where:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3295) * (i-a) each of these processes must get the same throughput as
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3296) * the others,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3297) * (i-b) in case (i-a) does not hold, it holds that the process
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3298) * associated with bfqq must receive a lower or equal
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3299) * throughput than any of the other processes;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3300) * (ii) the I/O of each process has the same properties, in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3301) * terms of locality (sequential or random), direction
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3302) * (reads or writes), request sizes, greediness
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3303) * (from I/O-bound to sporadic), and so on;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3304)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3305) * In fact, in such a scenario, the drive tends to treat the requests
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3306) * of each process in about the same way as the requests of the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3307) * others, and thus to provide each of these processes with about the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3308) * same throughput. This is exactly the desired throughput
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3309) * distribution if (i-a) holds, or, if (i-b) holds instead, this is an
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3310) * even more convenient distribution for (the process associated with)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3311) * bfqq.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3312) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3313) * In contrast, in any asymmetric or unfavorable scenario, device
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3314) * idling (I/O-dispatch plugging) is certainly needed to guarantee
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3315) * that bfqq receives its assigned fraction of the device throughput
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3316) * (see [1] for details).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3317) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3318) * The problem is that idling may significantly reduce throughput with
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3319) * certain combinations of types of I/O and devices. An important
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3320) * example is sync random I/O on flash storage with command
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3321) * queueing. So, unless bfqq falls in cases where idling also boosts
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3322) * throughput, it is important to check conditions (i-a), i(-b) and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3323) * (ii) accurately, so as to avoid idling when not strictly needed for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3324) * service guarantees.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3325) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3326) * Unfortunately, it is extremely difficult to thoroughly check
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3327) * condition (ii). And, in case there are active groups, it becomes
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3328) * very difficult to check conditions (i-a) and (i-b) too. In fact,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3329) * if there are active groups, then, for conditions (i-a) or (i-b) to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3330) * become false 'indirectly', it is enough that an active group
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3331) * contains more active processes or sub-groups than some other active
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3332) * group. More precisely, for conditions (i-a) or (i-b) to become
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3333) * false because of such a group, it is not even necessary that the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3334) * group is (still) active: it is sufficient that, even if the group
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3335) * has become inactive, some of its descendant processes still have
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3336) * some request already dispatched but still waiting for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3337) * completion. In fact, requests have still to be guaranteed their
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3338) * share of the throughput even after being dispatched. In this
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3339) * respect, it is easy to show that, if a group frequently becomes
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3340) * inactive while still having in-flight requests, and if, when this
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3341) * happens, the group is not considered in the calculation of whether
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3342) * the scenario is asymmetric, then the group may fail to be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3343) * guaranteed its fair share of the throughput (basically because
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3344) * idling may not be performed for the descendant processes of the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3345) * group, but it had to be). We address this issue with the following
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3346) * bi-modal behavior, implemented in the function
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3347) * bfq_asymmetric_scenario().
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3348) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3349) * If there are groups with requests waiting for completion
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3350) * (as commented above, some of these groups may even be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3351) * already inactive), then the scenario is tagged as
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3352) * asymmetric, conservatively, without checking any of the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3353) * conditions (i-a), (i-b) or (ii). So the device is idled for bfqq.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3354) * This behavior matches also the fact that groups are created
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3355) * exactly if controlling I/O is a primary concern (to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3356) * preserve bandwidth and latency guarantees).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3357) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3358) * On the opposite end, if there are no groups with requests waiting
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3359) * for completion, then only conditions (i-a) and (i-b) are actually
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3360) * controlled, i.e., provided that conditions (i-a) or (i-b) holds,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3361) * idling is not performed, regardless of whether condition (ii)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3362) * holds. In other words, only if conditions (i-a) and (i-b) do not
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3363) * hold, then idling is allowed, and the device tends to be prevented
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3364) * from queueing many requests, possibly of several processes. Since
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3365) * there are no groups with requests waiting for completion, then, to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3366) * control conditions (i-a) and (i-b) it is enough to check just
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3367) * whether all the queues with requests waiting for completion also
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3368) * have the same weight.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3369) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3370) * Not checking condition (ii) evidently exposes bfqq to the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3371) * risk of getting less throughput than its fair share.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3372) * However, for queues with the same weight, a further
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3373) * mechanism, preemption, mitigates or even eliminates this
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3374) * problem. And it does so without consequences on overall
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3375) * throughput. This mechanism and its benefits are explained
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3376) * in the next three paragraphs.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3377) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3378) * Even if a queue, say Q, is expired when it remains idle, Q
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3379) * can still preempt the new in-service queue if the next
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3380) * request of Q arrives soon (see the comments on
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3381) * bfq_bfqq_update_budg_for_activation). If all queues and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3382) * groups have the same weight, this form of preemption,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3383) * combined with the hole-recovery heuristic described in the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3384) * comments on function bfq_bfqq_update_budg_for_activation,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3385) * are enough to preserve a correct bandwidth distribution in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3386) * the mid term, even without idling. In fact, even if not
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3387) * idling allows the internal queues of the device to contain
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3388) * many requests, and thus to reorder requests, we can rather
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3389) * safely assume that the internal scheduler still preserves a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3390) * minimum of mid-term fairness.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3391) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3392) * More precisely, this preemption-based, idleless approach
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3393) * provides fairness in terms of IOPS, and not sectors per
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3394) * second. This can be seen with a simple example. Suppose
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3395) * that there are two queues with the same weight, but that
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3396) * the first queue receives requests of 8 sectors, while the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3397) * second queue receives requests of 1024 sectors. In
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3398) * addition, suppose that each of the two queues contains at
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3399) * most one request at a time, which implies that each queue
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3400) * always remains idle after it is served. Finally, after
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3401) * remaining idle, each queue receives very quickly a new
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3402) * request. It follows that the two queues are served
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3403) * alternatively, preempting each other if needed. This
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3404) * implies that, although both queues have the same weight,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3405) * the queue with large requests receives a service that is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3406) * 1024/8 times as high as the service received by the other
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3407) * queue.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3408) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3409) * The motivation for using preemption instead of idling (for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3410) * queues with the same weight) is that, by not idling,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3411) * service guarantees are preserved (completely or at least in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3412) * part) without minimally sacrificing throughput. And, if
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3413) * there is no active group, then the primary expectation for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3414) * this device is probably a high throughput.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3415) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3416) * We are now left only with explaining the two sub-conditions in the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3417) * additional compound condition that is checked below for deciding
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3418) * whether the scenario is asymmetric. To explain the first
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3419) * sub-condition, we need to add that the function
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3420) * bfq_asymmetric_scenario checks the weights of only
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3421) * non-weight-raised queues, for efficiency reasons (see comments on
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3422) * bfq_weights_tree_add()). Then the fact that bfqq is weight-raised
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3423) * is checked explicitly here. More precisely, the compound condition
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3424) * below takes into account also the fact that, even if bfqq is being
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3425) * weight-raised, the scenario is still symmetric if all queues with
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3426) * requests waiting for completion happen to be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3427) * weight-raised. Actually, we should be even more precise here, and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3428) * differentiate between interactive weight raising and soft real-time
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3429) * weight raising.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3430) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3431) * The second sub-condition checked in the compound condition is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3432) * whether there is a fair amount of already in-flight I/O not
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3433) * belonging to bfqq. If so, I/O dispatching is to be plugged, for the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3434) * following reason. The drive may decide to serve in-flight
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3435) * non-bfqq's I/O requests before bfqq's ones, thereby delaying the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3436) * arrival of new I/O requests for bfqq (recall that bfqq is sync). If
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3437) * I/O-dispatching is not plugged, then, while bfqq remains empty, a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3438) * basically uncontrolled amount of I/O from other queues may be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3439) * dispatched too, possibly causing the service of bfqq's I/O to be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3440) * delayed even longer in the drive. This problem gets more and more
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3441) * serious as the speed and the queue depth of the drive grow,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3442) * because, as these two quantities grow, the probability to find no
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3443) * queue busy but many requests in flight grows too. By contrast,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3444) * plugging I/O dispatching minimizes the delay induced by already
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3445) * in-flight I/O, and enables bfqq to recover the bandwidth it may
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3446) * lose because of this delay.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3447) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3448) * As a side note, it is worth considering that the above
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3449) * device-idling countermeasures may however fail in the following
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3450) * unlucky scenario: if I/O-dispatch plugging is (correctly) disabled
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3451) * in a time period during which all symmetry sub-conditions hold, and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3452) * therefore the device is allowed to enqueue many requests, but at
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3453) * some later point in time some sub-condition stops to hold, then it
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3454) * may become impossible to make requests be served in the desired
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3455) * order until all the requests already queued in the device have been
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3456) * served. The last sub-condition commented above somewhat mitigates
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3457) * this problem for weight-raised queues.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3458) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3459) static bool idling_needed_for_service_guarantees(struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3460) struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3461) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3462) /* No point in idling for bfqq if it won't get requests any longer */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3463) if (unlikely(!bfqq_process_refs(bfqq)))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3464) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3465)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3466) return (bfqq->wr_coeff > 1 &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3467) (bfqd->wr_busy_queues <
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3468) bfq_tot_busy_queues(bfqd) ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3469) bfqd->rq_in_driver >=
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3470) bfqq->dispatched + 4)) ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3471) bfq_asymmetric_scenario(bfqd, bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3472) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3473)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3474) static bool __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3475) enum bfqq_expiration reason)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3476) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3477) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3478) * If this bfqq is shared between multiple processes, check
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3479) * to make sure that those processes are still issuing I/Os
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3480) * within the mean seek distance. If not, it may be time to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3481) * break the queues apart again.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3482) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3483) if (bfq_bfqq_coop(bfqq) && BFQQ_SEEKY(bfqq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3484) bfq_mark_bfqq_split_coop(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3485)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3486) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3487) * Consider queues with a higher finish virtual time than
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3488) * bfqq. If idling_needed_for_service_guarantees(bfqq) returns
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3489) * true, then bfqq's bandwidth would be violated if an
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3490) * uncontrolled amount of I/O from these queues were
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3491) * dispatched while bfqq is waiting for its new I/O to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3492) * arrive. This is exactly what may happen if this is a forced
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3493) * expiration caused by a preemption attempt, and if bfqq is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3494) * not re-scheduled. To prevent this from happening, re-queue
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3495) * bfqq if it needs I/O-dispatch plugging, even if it is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3496) * empty. By doing so, bfqq is granted to be served before the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3497) * above queues (provided that bfqq is of course eligible).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3498) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3499) if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3500) !(reason == BFQQE_PREEMPTED &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3501) idling_needed_for_service_guarantees(bfqd, bfqq))) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3502) if (bfqq->dispatched == 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3503) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3504) * Overloading budget_timeout field to store
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3505) * the time at which the queue remains with no
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3506) * backlog and no outstanding request; used by
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3507) * the weight-raising mechanism.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3508) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3509) bfqq->budget_timeout = jiffies;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3510)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3511) bfq_del_bfqq_busy(bfqd, bfqq, true);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3512) } else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3513) bfq_requeue_bfqq(bfqd, bfqq, true);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3514) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3515) * Resort priority tree of potential close cooperators.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3516) * See comments on bfq_pos_tree_add_move() for the unlikely().
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3517) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3518) if (unlikely(!bfqd->nonrot_with_queueing &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3519) !RB_EMPTY_ROOT(&bfqq->sort_list)))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3520) bfq_pos_tree_add_move(bfqd, bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3521) }
^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) * All in-service entities must have been properly deactivated
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3525) * or requeued before executing the next function, which
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3526) * resets all in-service entities as no more in service. This
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3527) * may cause bfqq to be freed. If this happens, the next
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3528) * function returns true.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3529) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3530) return __bfq_bfqd_reset_in_service(bfqd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3531) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3532)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3533) /**
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3534) * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3535) * @bfqd: device data.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3536) * @bfqq: queue to update.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3537) * @reason: reason for expiration.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3538) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3539) * Handle the feedback on @bfqq budget at queue expiration.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3540) * See the body for detailed comments.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3541) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3542) static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3543) struct bfq_queue *bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3544) enum bfqq_expiration reason)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3545) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3546) struct request *next_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3547) int budget, min_budget;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3548)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3549) min_budget = bfq_min_budget(bfqd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3550)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3551) if (bfqq->wr_coeff == 1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3552) budget = bfqq->max_budget;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3553) else /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3554) * Use a constant, low budget for weight-raised queues,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3555) * to help achieve a low latency. Keep it slightly higher
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3556) * than the minimum possible budget, to cause a little
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3557) * bit fewer expirations.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3558) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3559) budget = 2 * min_budget;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3560)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3561) bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %d, budg left %d",
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3562) bfqq->entity.budget, bfq_bfqq_budget_left(bfqq));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3563) bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %d, min budg %d",
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3564) budget, bfq_min_budget(bfqd));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3565) bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d",
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3566) bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->in_service_queue));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3567)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3568) if (bfq_bfqq_sync(bfqq) && bfqq->wr_coeff == 1) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3569) switch (reason) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3570) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3571) * Caveat: in all the following cases we trade latency
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3572) * for throughput.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3573) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3574) case BFQQE_TOO_IDLE:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3575) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3576) * This is the only case where we may reduce
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3577) * the budget: if there is no request of the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3578) * process still waiting for completion, then
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3579) * we assume (tentatively) that the timer has
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3580) * expired because the batch of requests of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3581) * the process could have been served with a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3582) * smaller budget. Hence, betting that
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3583) * process will behave in the same way when it
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3584) * becomes backlogged again, we reduce its
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3585) * next budget. As long as we guess right,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3586) * this budget cut reduces the latency
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3587) * experienced by the process.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3588) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3589) * However, if there are still outstanding
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3590) * requests, then the process may have not yet
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3591) * issued its next request just because it is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3592) * still waiting for the completion of some of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3593) * the still outstanding ones. So in this
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3594) * subcase we do not reduce its budget, on the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3595) * contrary we increase it to possibly boost
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3596) * the throughput, as discussed in the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3597) * comments to the BUDGET_TIMEOUT case.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3598) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3599) if (bfqq->dispatched > 0) /* still outstanding reqs */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3600) budget = min(budget * 2, bfqd->bfq_max_budget);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3601) else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3602) if (budget > 5 * min_budget)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3603) budget -= 4 * min_budget;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3604) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3605) budget = min_budget;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3606) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3607) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3608) case BFQQE_BUDGET_TIMEOUT:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3609) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3610) * We double the budget here because it gives
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3611) * the chance to boost the throughput if this
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3612) * is not a seeky process (and has bumped into
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3613) * this timeout because of, e.g., ZBR).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3614) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3615) budget = min(budget * 2, bfqd->bfq_max_budget);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3616) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3617) case BFQQE_BUDGET_EXHAUSTED:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3618) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3619) * The process still has backlog, and did not
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3620) * let either the budget timeout or the disk
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3621) * idling timeout expire. Hence it is not
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3622) * seeky, has a short thinktime and may be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3623) * happy with a higher budget too. So
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3624) * definitely increase the budget of this good
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3625) * candidate to boost the disk throughput.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3626) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3627) budget = min(budget * 4, bfqd->bfq_max_budget);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3628) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3629) case BFQQE_NO_MORE_REQUESTS:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3630) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3631) * For queues that expire for this reason, it
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3632) * is particularly important to keep the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3633) * budget close to the actual service they
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3634) * need. Doing so reduces the timestamp
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3635) * misalignment problem described in the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3636) * comments in the body of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3637) * __bfq_activate_entity. In fact, suppose
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3638) * that a queue systematically expires for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3639) * BFQQE_NO_MORE_REQUESTS and presents a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3640) * new request in time to enjoy timestamp
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3641) * back-shifting. The larger the budget of the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3642) * queue is with respect to the service the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3643) * queue actually requests in each service
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3644) * slot, the more times the queue can be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3645) * reactivated with the same virtual finish
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3646) * time. It follows that, even if this finish
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3647) * time is pushed to the system virtual time
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3648) * to reduce the consequent timestamp
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3649) * misalignment, the queue unjustly enjoys for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3650) * many re-activations a lower finish time
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3651) * than all newly activated queues.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3652) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3653) * The service needed by bfqq is measured
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3654) * quite precisely by bfqq->entity.service.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3655) * Since bfqq does not enjoy device idling,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3656) * bfqq->entity.service is equal to the number
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3657) * of sectors that the process associated with
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3658) * bfqq requested to read/write before waiting
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3659) * for request completions, or blocking for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3660) * other reasons.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3661) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3662) budget = max_t(int, bfqq->entity.service, min_budget);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3663) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3664) default:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3665) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3666) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3667) } else if (!bfq_bfqq_sync(bfqq)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3668) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3669) * Async queues get always the maximum possible
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3670) * budget, as for them we do not care about latency
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3671) * (in addition, their ability to dispatch is limited
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3672) * by the charging factor).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3673) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3674) budget = bfqd->bfq_max_budget;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3675) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3676)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3677) bfqq->max_budget = budget;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3678)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3679) if (bfqd->budgets_assigned >= bfq_stats_min_budgets &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3680) !bfqd->bfq_user_max_budget)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3681) bfqq->max_budget = min(bfqq->max_budget, bfqd->bfq_max_budget);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3682)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3683) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3684) * If there is still backlog, then assign a new budget, making
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3685) * sure that it is large enough for the next request. Since
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3686) * the finish time of bfqq must be kept in sync with the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3687) * budget, be sure to call __bfq_bfqq_expire() *after* this
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3688) * update.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3689) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3690) * If there is no backlog, then no need to update the budget;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3691) * it will be updated on the arrival of a new request.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3692) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3693) next_rq = bfqq->next_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3694) if (next_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3695) bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3696) bfq_serv_to_charge(next_rq, bfqq));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3697)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3698) bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %d",
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3699) next_rq ? blk_rq_sectors(next_rq) : 0,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3700) bfqq->entity.budget);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3701) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3702)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3703) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3704) * Return true if the process associated with bfqq is "slow". The slow
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3705) * flag is used, in addition to the budget timeout, to reduce the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3706) * amount of service provided to seeky processes, and thus reduce
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3707) * their chances to lower the throughput. More details in the comments
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3708) * on the function bfq_bfqq_expire().
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3709) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3710) * An important observation is in order: as discussed in the comments
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3711) * on the function bfq_update_peak_rate(), with devices with internal
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3712) * queues, it is hard if ever possible to know when and for how long
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3713) * an I/O request is processed by the device (apart from the trivial
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3714) * I/O pattern where a new request is dispatched only after the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3715) * previous one has been completed). This makes it hard to evaluate
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3716) * the real rate at which the I/O requests of each bfq_queue are
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3717) * served. In fact, for an I/O scheduler like BFQ, serving a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3718) * bfq_queue means just dispatching its requests during its service
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3719) * slot (i.e., until the budget of the queue is exhausted, or the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3720) * queue remains idle, or, finally, a timeout fires). But, during the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3721) * service slot of a bfq_queue, around 100 ms at most, the device may
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3722) * be even still processing requests of bfq_queues served in previous
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3723) * service slots. On the opposite end, the requests of the in-service
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3724) * bfq_queue may be completed after the service slot of the queue
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3725) * finishes.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3726) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3727) * Anyway, unless more sophisticated solutions are used
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3728) * (where possible), the sum of the sizes of the requests dispatched
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3729) * during the service slot of a bfq_queue is probably the only
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3730) * approximation available for the service received by the bfq_queue
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3731) * during its service slot. And this sum is the quantity used in this
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3732) * function to evaluate the I/O speed of a process.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3733) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3734) static bool bfq_bfqq_is_slow(struct bfq_data *bfqd, struct bfq_queue *bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3735) bool compensate, enum bfqq_expiration reason,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3736) unsigned long *delta_ms)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3737) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3738) ktime_t delta_ktime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3739) u32 delta_usecs;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3740) bool slow = BFQQ_SEEKY(bfqq); /* if delta too short, use seekyness */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3741)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3742) if (!bfq_bfqq_sync(bfqq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3743) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3744)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3745) if (compensate)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3746) delta_ktime = bfqd->last_idling_start;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3747) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3748) delta_ktime = ktime_get();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3749) delta_ktime = ktime_sub(delta_ktime, bfqd->last_budget_start);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3750) delta_usecs = ktime_to_us(delta_ktime);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3751)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3752) /* don't use too short time intervals */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3753) if (delta_usecs < 1000) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3754) if (blk_queue_nonrot(bfqd->queue))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3755) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3756) * give same worst-case guarantees as idling
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3757) * for seeky
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3758) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3759) *delta_ms = BFQ_MIN_TT / NSEC_PER_MSEC;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3760) else /* charge at least one seek */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3761) *delta_ms = bfq_slice_idle / NSEC_PER_MSEC;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3762)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3763) return slow;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3764) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3765)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3766) *delta_ms = delta_usecs / USEC_PER_MSEC;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3767)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3768) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3769) * Use only long (> 20ms) intervals to filter out excessive
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3770) * spikes in service rate estimation.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3771) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3772) if (delta_usecs > 20000) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3773) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3774) * Caveat for rotational devices: processes doing I/O
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3775) * in the slower disk zones tend to be slow(er) even
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3776) * if not seeky. In this respect, the estimated peak
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3777) * rate is likely to be an average over the disk
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3778) * surface. Accordingly, to not be too harsh with
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3779) * unlucky processes, a process is deemed slow only if
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3780) * its rate has been lower than half of the estimated
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3781) * peak rate.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3782) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3783) slow = bfqq->entity.service < bfqd->bfq_max_budget / 2;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3784) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3785)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3786) bfq_log_bfqq(bfqd, bfqq, "bfq_bfqq_is_slow: slow %d", slow);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3787)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3788) return slow;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3789) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3790)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3791) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3792) * To be deemed as soft real-time, an application must meet two
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3793) * requirements. First, the application must not require an average
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3794) * bandwidth higher than the approximate bandwidth required to playback or
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3795) * record a compressed high-definition video.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3796) * The next function is invoked on the completion of the last request of a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3797) * batch, to compute the next-start time instant, soft_rt_next_start, such
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3798) * that, if the next request of the application does not arrive before
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3799) * soft_rt_next_start, then the above requirement on the bandwidth is met.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3800) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3801) * The second requirement is that the request pattern of the application is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3802) * isochronous, i.e., that, after issuing a request or a batch of requests,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3803) * the application stops issuing new requests until all its pending requests
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3804) * have been completed. After that, the application may issue a new batch,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3805) * and so on.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3806) * For this reason the next function is invoked to compute
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3807) * soft_rt_next_start only for applications that meet this requirement,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3808) * whereas soft_rt_next_start is set to infinity for applications that do
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3809) * not.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3810) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3811) * Unfortunately, even a greedy (i.e., I/O-bound) application may
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3812) * happen to meet, occasionally or systematically, both the above
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3813) * bandwidth and isochrony requirements. This may happen at least in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3814) * the following circumstances. First, if the CPU load is high. The
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3815) * application may stop issuing requests while the CPUs are busy
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3816) * serving other processes, then restart, then stop again for a while,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3817) * and so on. The other circumstances are related to the storage
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3818) * device: the storage device is highly loaded or reaches a low-enough
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3819) * throughput with the I/O of the application (e.g., because the I/O
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3820) * is random and/or the device is slow). In all these cases, the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3821) * I/O of the application may be simply slowed down enough to meet
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3822) * the bandwidth and isochrony requirements. To reduce the probability
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3823) * that greedy applications are deemed as soft real-time in these
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3824) * corner cases, a further rule is used in the computation of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3825) * soft_rt_next_start: the return value of this function is forced to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3826) * be higher than the maximum between the following two quantities.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3827) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3828) * (a) Current time plus: (1) the maximum time for which the arrival
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3829) * of a request is waited for when a sync queue becomes idle,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3830) * namely bfqd->bfq_slice_idle, and (2) a few extra jiffies. We
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3831) * postpone for a moment the reason for adding a few extra
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3832) * jiffies; we get back to it after next item (b). Lower-bounding
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3833) * the return value of this function with the current time plus
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3834) * bfqd->bfq_slice_idle tends to filter out greedy applications,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3835) * because the latter issue their next request as soon as possible
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3836) * after the last one has been completed. In contrast, a soft
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3837) * real-time application spends some time processing data, after a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3838) * batch of its requests has been completed.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3839) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3840) * (b) Current value of bfqq->soft_rt_next_start. As pointed out
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3841) * above, greedy applications may happen to meet both the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3842) * bandwidth and isochrony requirements under heavy CPU or
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3843) * storage-device load. In more detail, in these scenarios, these
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3844) * applications happen, only for limited time periods, to do I/O
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3845) * slowly enough to meet all the requirements described so far,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3846) * including the filtering in above item (a). These slow-speed
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3847) * time intervals are usually interspersed between other time
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3848) * intervals during which these applications do I/O at a very high
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3849) * speed. Fortunately, exactly because of the high speed of the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3850) * I/O in the high-speed intervals, the values returned by this
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3851) * function happen to be so high, near the end of any such
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3852) * high-speed interval, to be likely to fall *after* the end of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3853) * the low-speed time interval that follows. These high values are
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3854) * stored in bfqq->soft_rt_next_start after each invocation of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3855) * this function. As a consequence, if the last value of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3856) * bfqq->soft_rt_next_start is constantly used to lower-bound the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3857) * next value that this function may return, then, from the very
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3858) * beginning of a low-speed interval, bfqq->soft_rt_next_start is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3859) * likely to be constantly kept so high that any I/O request
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3860) * issued during the low-speed interval is considered as arriving
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3861) * to soon for the application to be deemed as soft
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3862) * real-time. Then, in the high-speed interval that follows, the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3863) * application will not be deemed as soft real-time, just because
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3864) * it will do I/O at a high speed. And so on.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3865) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3866) * Getting back to the filtering in item (a), in the following two
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3867) * cases this filtering might be easily passed by a greedy
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3868) * application, if the reference quantity was just
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3869) * bfqd->bfq_slice_idle:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3870) * 1) HZ is so low that the duration of a jiffy is comparable to or
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3871) * higher than bfqd->bfq_slice_idle. This happens, e.g., on slow
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3872) * devices with HZ=100. The time granularity may be so coarse
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3873) * that the approximation, in jiffies, of bfqd->bfq_slice_idle
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3874) * is rather lower than the exact value.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3875) * 2) jiffies, instead of increasing at a constant rate, may stop increasing
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3876) * for a while, then suddenly 'jump' by several units to recover the lost
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3877) * increments. This seems to happen, e.g., inside virtual machines.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3878) * To address this issue, in the filtering in (a) we do not use as a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3879) * reference time interval just bfqd->bfq_slice_idle, but
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3880) * bfqd->bfq_slice_idle plus a few jiffies. In particular, we add the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3881) * minimum number of jiffies for which the filter seems to be quite
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3882) * precise also in embedded systems and KVM/QEMU virtual machines.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3883) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3884) static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3885) struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3886) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3887) return max3(bfqq->soft_rt_next_start,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3888) bfqq->last_idle_bklogged +
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3889) HZ * bfqq->service_from_backlogged /
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3890) bfqd->bfq_wr_max_softrt_rate,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3891) jiffies + nsecs_to_jiffies(bfqq->bfqd->bfq_slice_idle) + 4);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3892) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3893)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3894) /**
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3895) * bfq_bfqq_expire - expire a queue.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3896) * @bfqd: device owning the queue.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3897) * @bfqq: the queue to expire.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3898) * @compensate: if true, compensate for the time spent idling.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3899) * @reason: the reason causing the expiration.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3900) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3901) * If the process associated with bfqq does slow I/O (e.g., because it
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3902) * issues random requests), we charge bfqq with the time it has been
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3903) * in service instead of the service it has received (see
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3904) * bfq_bfqq_charge_time for details on how this goal is achieved). As
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3905) * a consequence, bfqq will typically get higher timestamps upon
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3906) * reactivation, and hence it will be rescheduled as if it had
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3907) * received more service than what it has actually received. In the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3908) * end, bfqq receives less service in proportion to how slowly its
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3909) * associated process consumes its budgets (and hence how seriously it
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3910) * tends to lower the throughput). In addition, this time-charging
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3911) * strategy guarantees time fairness among slow processes. In
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3912) * contrast, if the process associated with bfqq is not slow, we
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3913) * charge bfqq exactly with the service it has received.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3914) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3915) * Charging time to the first type of queues and the exact service to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3916) * the other has the effect of using the WF2Q+ policy to schedule the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3917) * former on a timeslice basis, without violating service domain
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3918) * guarantees among the latter.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3919) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3920) void bfq_bfqq_expire(struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3921) struct bfq_queue *bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3922) bool compensate,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3923) enum bfqq_expiration reason)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3924) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3925) bool slow;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3926) unsigned long delta = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3927) struct bfq_entity *entity = &bfqq->entity;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3928)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3929) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3930) * Check whether the process is slow (see bfq_bfqq_is_slow).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3931) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3932) slow = bfq_bfqq_is_slow(bfqd, bfqq, compensate, reason, &delta);
^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) * As above explained, charge slow (typically seeky) and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3936) * timed-out queues with the time and not the service
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3937) * received, to favor sequential workloads.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3938) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3939) * Processes doing I/O in the slower disk zones will tend to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3940) * be slow(er) even if not seeky. Therefore, since the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3941) * estimated peak rate is actually an average over the disk
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3942) * surface, these processes may timeout just for bad luck. To
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3943) * avoid punishing them, do not charge time to processes that
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3944) * succeeded in consuming at least 2/3 of their budget. This
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3945) * allows BFQ to preserve enough elasticity to still perform
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3946) * bandwidth, and not time, distribution with little unlucky
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3947) * or quasi-sequential processes.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3948) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3949) if (bfqq->wr_coeff == 1 &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3950) (slow ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3951) (reason == BFQQE_BUDGET_TIMEOUT &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3952) bfq_bfqq_budget_left(bfqq) >= entity->budget / 3)))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3953) bfq_bfqq_charge_time(bfqd, bfqq, delta);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3954)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3955) if (reason == BFQQE_TOO_IDLE &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3956) entity->service <= 2 * entity->budget / 10)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3957) bfq_clear_bfqq_IO_bound(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3958)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3959) if (bfqd->low_latency && bfqq->wr_coeff == 1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3960) bfqq->last_wr_start_finish = jiffies;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3961)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3962) if (bfqd->low_latency && bfqd->bfq_wr_max_softrt_rate > 0 &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3963) RB_EMPTY_ROOT(&bfqq->sort_list)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3964) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3965) * If we get here, and there are no outstanding
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3966) * requests, then the request pattern is isochronous
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3967) * (see the comments on the function
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3968) * bfq_bfqq_softrt_next_start()). Thus we can compute
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3969) * soft_rt_next_start. And we do it, unless bfqq is in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3970) * interactive weight raising. We do not do it in the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3971) * latter subcase, for the following reason. bfqq may
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3972) * be conveying the I/O needed to load a soft
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3973) * real-time application. Such an application will
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3974) * actually exhibit a soft real-time I/O pattern after
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3975) * it finally starts doing its job. But, if
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3976) * soft_rt_next_start is computed here for an
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3977) * interactive bfqq, and bfqq had received a lot of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3978) * service before remaining with no outstanding
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3979) * request (likely to happen on a fast device), then
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3980) * soft_rt_next_start would be assigned such a high
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3981) * value that, for a very long time, bfqq would be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3982) * prevented from being possibly considered as soft
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3983) * real time.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3984) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3985) * If, instead, the queue still has outstanding
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3986) * requests, then we have to wait for the completion
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3987) * of all the outstanding requests to discover whether
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3988) * the request pattern is actually isochronous.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3989) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3990) if (bfqq->dispatched == 0 &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3991) bfqq->wr_coeff != bfqd->bfq_wr_coeff)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3992) bfqq->soft_rt_next_start =
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3993) bfq_bfqq_softrt_next_start(bfqd, bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3994) else if (bfqq->dispatched > 0) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3995) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3996) * Schedule an update of soft_rt_next_start to when
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3997) * the task may be discovered to be isochronous.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3998) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 3999) bfq_mark_bfqq_softrt_update(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4000) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4001) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4002)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4003) bfq_log_bfqq(bfqd, bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4004) "expire (%d, slow %d, num_disp %d, short_ttime %d)", reason,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4005) slow, bfqq->dispatched, bfq_bfqq_has_short_ttime(bfqq));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4006)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4007) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4008) * bfqq expired, so no total service time needs to be computed
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4009) * any longer: reset state machine for measuring total service
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4010) * times.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4011) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4012) bfqd->rqs_injected = bfqd->wait_dispatch = false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4013) bfqd->waited_rq = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4014)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4015) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4016) * Increase, decrease or leave budget unchanged according to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4017) * reason.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4018) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4019) __bfq_bfqq_recalc_budget(bfqd, bfqq, reason);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4020) if (__bfq_bfqq_expire(bfqd, bfqq, reason))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4021) /* bfqq is gone, no more actions on it */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4022) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4023)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4024) /* mark bfqq as waiting a request only if a bic still points to it */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4025) if (!bfq_bfqq_busy(bfqq) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4026) reason != BFQQE_BUDGET_TIMEOUT &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4027) reason != BFQQE_BUDGET_EXHAUSTED) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4028) bfq_mark_bfqq_non_blocking_wait_rq(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4029) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4030) * Not setting service to 0, because, if the next rq
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4031) * arrives in time, the queue will go on receiving
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4032) * service with this same budget (as if it never expired)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4033) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4034) } else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4035) entity->service = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4036)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4037) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4038) * Reset the received-service counter for every parent entity.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4039) * Differently from what happens with bfqq->entity.service,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4040) * the resetting of this counter never needs to be postponed
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4041) * for parent entities. In fact, in case bfqq may have a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4042) * chance to go on being served using the last, partially
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4043) * consumed budget, bfqq->entity.service needs to be kept,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4044) * because if bfqq then actually goes on being served using
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4045) * the same budget, the last value of bfqq->entity.service is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4046) * needed to properly decrement bfqq->entity.budget by the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4047) * portion already consumed. In contrast, it is not necessary
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4048) * to keep entity->service for parent entities too, because
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4049) * the bubble up of the new value of bfqq->entity.budget will
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4050) * make sure that the budgets of parent entities are correct,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4051) * even in case bfqq and thus parent entities go on receiving
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4052) * service with the same budget.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4053) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4054) entity = entity->parent;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4055) for_each_entity(entity)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4056) entity->service = 0;
^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) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4060) * Budget timeout is not implemented through a dedicated timer, but
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4061) * just checked on request arrivals and completions, as well as on
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4062) * idle timer expirations.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4063) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4064) static bool bfq_bfqq_budget_timeout(struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4065) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4066) return time_is_before_eq_jiffies(bfqq->budget_timeout);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4067) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4068)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4069) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4070) * If we expire a queue that is actively waiting (i.e., with the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4071) * device idled) for the arrival of a new request, then we may incur
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4072) * the timestamp misalignment problem described in the body of the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4073) * function __bfq_activate_entity. Hence we return true only if this
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4074) * condition does not hold, or if the queue is slow enough to deserve
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4075) * only to be kicked off for preserving a high throughput.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4076) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4077) static bool bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4078) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4079) bfq_log_bfqq(bfqq->bfqd, bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4080) "may_budget_timeout: wait_request %d left %d timeout %d",
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4081) bfq_bfqq_wait_request(bfqq),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4082) bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4083) bfq_bfqq_budget_timeout(bfqq));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4084)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4085) return (!bfq_bfqq_wait_request(bfqq) ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4086) bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4087) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4088) bfq_bfqq_budget_timeout(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4089) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4090)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4091) static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4092) struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4093) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4094) bool rot_without_queueing =
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4095) !blk_queue_nonrot(bfqd->queue) && !bfqd->hw_tag,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4096) bfqq_sequential_and_IO_bound,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4097) idling_boosts_thr;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4098)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4099) /* No point in idling for bfqq if it won't get requests any longer */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4100) if (unlikely(!bfqq_process_refs(bfqq)))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4101) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4102)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4103) bfqq_sequential_and_IO_bound = !BFQQ_SEEKY(bfqq) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4104) bfq_bfqq_IO_bound(bfqq) && bfq_bfqq_has_short_ttime(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4105)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4106) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4107) * The next variable takes into account the cases where idling
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4108) * boosts the throughput.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4109) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4110) * The value of the variable is computed considering, first, that
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4111) * idling is virtually always beneficial for the throughput if:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4112) * (a) the device is not NCQ-capable and rotational, or
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4113) * (b) regardless of the presence of NCQ, the device is rotational and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4114) * the request pattern for bfqq is I/O-bound and sequential, or
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4115) * (c) regardless of whether it is rotational, the device is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4116) * not NCQ-capable and the request pattern for bfqq is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4117) * I/O-bound and sequential.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4118) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4119) * Secondly, and in contrast to the above item (b), idling an
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4120) * NCQ-capable flash-based device would not boost the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4121) * throughput even with sequential I/O; rather it would lower
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4122) * the throughput in proportion to how fast the device
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4123) * is. Accordingly, the next variable is true if any of the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4124) * above conditions (a), (b) or (c) is true, and, in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4125) * particular, happens to be false if bfqd is an NCQ-capable
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4126) * flash-based device.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4127) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4128) idling_boosts_thr = rot_without_queueing ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4129) ((!blk_queue_nonrot(bfqd->queue) || !bfqd->hw_tag) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4130) bfqq_sequential_and_IO_bound);
^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) * The return value of this function is equal to that of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4134) * idling_boosts_thr, unless a special case holds. In this
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4135) * special case, described below, idling may cause problems to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4136) * weight-raised queues.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4137) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4138) * When the request pool is saturated (e.g., in the presence
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4139) * of write hogs), if the processes associated with
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4140) * non-weight-raised queues ask for requests at a lower rate,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4141) * then processes associated with weight-raised queues have a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4142) * higher probability to get a request from the pool
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4143) * immediately (or at least soon) when they need one. Thus
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4144) * they have a higher probability to actually get a fraction
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4145) * of the device throughput proportional to their high
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4146) * weight. This is especially true with NCQ-capable drives,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4147) * which enqueue several requests in advance, and further
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4148) * reorder internally-queued requests.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4149) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4150) * For this reason, we force to false the return value if
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4151) * there are weight-raised busy queues. In this case, and if
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4152) * bfqq is not weight-raised, this guarantees that the device
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4153) * is not idled for bfqq (if, instead, bfqq is weight-raised,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4154) * then idling will be guaranteed by another variable, see
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4155) * below). Combined with the timestamping rules of BFQ (see
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4156) * [1] for details), this behavior causes bfqq, and hence any
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4157) * sync non-weight-raised queue, to get a lower number of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4158) * requests served, and thus to ask for a lower number of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4159) * requests from the request pool, before the busy
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4160) * weight-raised queues get served again. This often mitigates
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4161) * starvation problems in the presence of heavy write
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4162) * workloads and NCQ, thereby guaranteeing a higher
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4163) * application and system responsiveness in these hostile
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4164) * scenarios.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4165) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4166) return idling_boosts_thr &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4167) bfqd->wr_busy_queues == 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4168) }
^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) * For a queue that becomes empty, device idling is allowed only if
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4172) * this function returns true for that queue. As a consequence, since
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4173) * device idling plays a critical role for both throughput boosting
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4174) * and service guarantees, the return value of this function plays a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4175) * critical role as well.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4176) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4177) * In a nutshell, this function returns true only if idling is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4178) * beneficial for throughput or, even if detrimental for throughput,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4179) * idling is however necessary to preserve service guarantees (low
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4180) * latency, desired throughput distribution, ...). In particular, on
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4181) * NCQ-capable devices, this function tries to return false, so as to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4182) * help keep the drives' internal queues full, whenever this helps the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4183) * device boost the throughput without causing any service-guarantee
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4184) * issue.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4185) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4186) * Most of the issues taken into account to get the return value of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4187) * this function are not trivial. We discuss these issues in the two
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4188) * functions providing the main pieces of information needed by this
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4189) * function.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4190) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4191) static bool bfq_better_to_idle(struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4192) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4193) struct bfq_data *bfqd = bfqq->bfqd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4194) bool idling_boosts_thr_with_no_issue, idling_needed_for_service_guar;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4195)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4196) /* No point in idling for bfqq if it won't get requests any longer */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4197) if (unlikely(!bfqq_process_refs(bfqq)))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4198) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4199)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4200) if (unlikely(bfqd->strict_guarantees))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4201) return true;
^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) * Idling is performed only if slice_idle > 0. In addition, we
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4205) * do not idle if
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4206) * (a) bfqq is async
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4207) * (b) bfqq is in the idle io prio class: in this case we do
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4208) * not idle because we want to minimize the bandwidth that
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4209) * queues in this class can steal to higher-priority queues
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4210) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4211) if (bfqd->bfq_slice_idle == 0 || !bfq_bfqq_sync(bfqq) ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4212) bfq_class_idle(bfqq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4213) return false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4214)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4215) idling_boosts_thr_with_no_issue =
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4216) idling_boosts_thr_without_issues(bfqd, bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4217)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4218) idling_needed_for_service_guar =
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4219) idling_needed_for_service_guarantees(bfqd, bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4220)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4221) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4222) * We have now the two components we need to compute the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4223) * return value of the function, which is true only if idling
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4224) * either boosts the throughput (without issues), or is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4225) * necessary to preserve service guarantees.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4226) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4227) return idling_boosts_thr_with_no_issue ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4228) idling_needed_for_service_guar;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4229) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4230)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4231) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4232) * If the in-service queue is empty but the function bfq_better_to_idle
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4233) * returns true, then:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4234) * 1) the queue must remain in service and cannot be expired, and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4235) * 2) the device must be idled to wait for the possible arrival of a new
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4236) * request for the queue.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4237) * See the comments on the function bfq_better_to_idle for the reasons
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4238) * why performing device idling is the best choice to boost the throughput
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4239) * and preserve service guarantees when bfq_better_to_idle itself
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4240) * returns true.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4241) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4242) static bool bfq_bfqq_must_idle(struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4243) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4244) return RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_better_to_idle(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4245) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4246)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4247) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4248) * This function chooses the queue from which to pick the next extra
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4249) * I/O request to inject, if it finds a compatible queue. See the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4250) * comments on bfq_update_inject_limit() for details on the injection
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4251) * mechanism, and for the definitions of the quantities mentioned
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4252) * below.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4253) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4254) static struct bfq_queue *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4255) bfq_choose_bfqq_for_injection(struct bfq_data *bfqd)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4256) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4257) struct bfq_queue *bfqq, *in_serv_bfqq = bfqd->in_service_queue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4258) unsigned int limit = in_serv_bfqq->inject_limit;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4259) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4260) * If
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4261) * - bfqq is not weight-raised and therefore does not carry
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4262) * time-critical I/O,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4263) * or
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4264) * - regardless of whether bfqq is weight-raised, bfqq has
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4265) * however a long think time, during which it can absorb the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4266) * effect of an appropriate number of extra I/O requests
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4267) * from other queues (see bfq_update_inject_limit for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4268) * details on the computation of this number);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4269) * then injection can be performed without restrictions.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4270) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4271) bool in_serv_always_inject = in_serv_bfqq->wr_coeff == 1 ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4272) !bfq_bfqq_has_short_ttime(in_serv_bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4273)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4274) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4275) * If
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4276) * - the baseline total service time could not be sampled yet,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4277) * so the inject limit happens to be still 0, and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4278) * - a lot of time has elapsed since the plugging of I/O
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4279) * dispatching started, so drive speed is being wasted
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4280) * significantly;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4281) * then temporarily raise inject limit to one request.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4282) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4283) if (limit == 0 && in_serv_bfqq->last_serv_time_ns == 0 &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4284) bfq_bfqq_wait_request(in_serv_bfqq) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4285) time_is_before_eq_jiffies(bfqd->last_idling_start_jiffies +
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4286) bfqd->bfq_slice_idle)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4287) )
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4288) limit = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4289)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4290) if (bfqd->rq_in_driver >= limit)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4291) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4292)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4293) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4294) * Linear search of the source queue for injection; but, with
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4295) * a high probability, very few steps are needed to find a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4296) * candidate queue, i.e., a queue with enough budget left for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4297) * its next request. In fact:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4298) * - BFQ dynamically updates the budget of every queue so as
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4299) * to accommodate the expected backlog of the queue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4300) * - if a queue gets all its requests dispatched as injected
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4301) * service, then the queue is removed from the active list
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4302) * (and re-added only if it gets new requests, but then it
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4303) * is assigned again enough budget for its new backlog).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4304) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4305) list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4306) if (!RB_EMPTY_ROOT(&bfqq->sort_list) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4307) (in_serv_always_inject || bfqq->wr_coeff > 1) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4308) bfq_serv_to_charge(bfqq->next_rq, bfqq) <=
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4309) bfq_bfqq_budget_left(bfqq)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4310) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4311) * Allow for only one large in-flight request
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4312) * on non-rotational devices, for the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4313) * following reason. On non-rotationl drives,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4314) * large requests take much longer than
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4315) * smaller requests to be served. In addition,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4316) * the drive prefers to serve large requests
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4317) * w.r.t. to small ones, if it can choose. So,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4318) * having more than one large requests queued
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4319) * in the drive may easily make the next first
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4320) * request of the in-service queue wait for so
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4321) * long to break bfqq's service guarantees. On
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4322) * the bright side, large requests let the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4323) * drive reach a very high throughput, even if
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4324) * there is only one in-flight large request
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4325) * at a time.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4326) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4327) if (blk_queue_nonrot(bfqd->queue) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4328) blk_rq_sectors(bfqq->next_rq) >=
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4329) BFQQ_SECT_THR_NONROT)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4330) limit = min_t(unsigned int, 1, limit);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4331) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4332) limit = in_serv_bfqq->inject_limit;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4333)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4334) if (bfqd->rq_in_driver < limit) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4335) bfqd->rqs_injected = true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4336) return bfqq;
^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)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4340) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4341) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4342)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4343) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4344) * Select a queue for service. If we have a current queue in service,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4345) * check whether to continue servicing it, or retrieve and set a new one.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4346) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4347) static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4348) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4349) struct bfq_queue *bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4350) struct request *next_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4351) enum bfqq_expiration reason = BFQQE_BUDGET_TIMEOUT;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4352)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4353) bfqq = bfqd->in_service_queue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4354) if (!bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4355) goto new_queue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4356)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4357) bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue");
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4358)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4359) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4360) * Do not expire bfqq for budget timeout if bfqq may be about
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4361) * to enjoy device idling. The reason why, in this case, we
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4362) * prevent bfqq from expiring is the same as in the comments
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4363) * on the case where bfq_bfqq_must_idle() returns true, in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4364) * bfq_completed_request().
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4365) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4366) if (bfq_may_expire_for_budg_timeout(bfqq) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4367) !bfq_bfqq_must_idle(bfqq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4368) goto expire;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4369)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4370) check_queue:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4371) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4372) * This loop is rarely executed more than once. Even when it
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4373) * happens, it is much more convenient to re-execute this loop
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4374) * than to return NULL and trigger a new dispatch to get a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4375) * request served.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4376) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4377) next_rq = bfqq->next_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4378) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4379) * If bfqq has requests queued and it has enough budget left to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4380) * serve them, keep the queue, otherwise expire it.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4381) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4382) if (next_rq) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4383) if (bfq_serv_to_charge(next_rq, bfqq) >
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4384) bfq_bfqq_budget_left(bfqq)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4385) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4386) * Expire the queue for budget exhaustion,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4387) * which makes sure that the next budget is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4388) * enough to serve the next request, even if
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4389) * it comes from the fifo expired path.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4390) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4391) reason = BFQQE_BUDGET_EXHAUSTED;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4392) goto expire;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4393) } else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4394) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4395) * The idle timer may be pending because we may
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4396) * not disable disk idling even when a new request
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4397) * arrives.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4398) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4399) if (bfq_bfqq_wait_request(bfqq)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4400) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4401) * If we get here: 1) at least a new request
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4402) * has arrived but we have not disabled the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4403) * timer because the request was too small,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4404) * 2) then the block layer has unplugged
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4405) * the device, causing the dispatch to be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4406) * invoked.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4407) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4408) * Since the device is unplugged, now the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4409) * requests are probably large enough to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4410) * provide a reasonable throughput.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4411) * So we disable idling.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4412) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4413) bfq_clear_bfqq_wait_request(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4414) hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4415) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4416) goto keep_queue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4417) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4418) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4419)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4420) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4421) * No requests pending. However, if the in-service queue is idling
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4422) * for a new request, or has requests waiting for a completion and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4423) * may idle after their completion, then keep it anyway.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4424) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4425) * Yet, inject service from other queues if it boosts
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4426) * throughput and is possible.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4427) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4428) if (bfq_bfqq_wait_request(bfqq) ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4429) (bfqq->dispatched != 0 && bfq_better_to_idle(bfqq))) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4430) struct bfq_queue *async_bfqq =
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4431) bfqq->bic && bfqq->bic->bfqq[0] &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4432) bfq_bfqq_busy(bfqq->bic->bfqq[0]) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4433) bfqq->bic->bfqq[0]->next_rq ?
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4434) bfqq->bic->bfqq[0] : NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4435)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4436) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4437) * The next three mutually-exclusive ifs decide
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4438) * whether to try injection, and choose the queue to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4439) * pick an I/O request from.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4440) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4441) * The first if checks whether the process associated
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4442) * with bfqq has also async I/O pending. If so, it
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4443) * injects such I/O unconditionally. Injecting async
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4444) * I/O from the same process can cause no harm to the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4445) * process. On the contrary, it can only increase
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4446) * bandwidth and reduce latency for the process.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4447) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4448) * The second if checks whether there happens to be a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4449) * non-empty waker queue for bfqq, i.e., a queue whose
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4450) * I/O needs to be completed for bfqq to receive new
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4451) * I/O. This happens, e.g., if bfqq is associated with
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4452) * a process that does some sync. A sync generates
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4453) * extra blocking I/O, which must be completed before
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4454) * the process associated with bfqq can go on with its
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4455) * I/O. If the I/O of the waker queue is not served,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4456) * then bfqq remains empty, and no I/O is dispatched,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4457) * until the idle timeout fires for bfqq. This is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4458) * likely to result in lower bandwidth and higher
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4459) * latencies for bfqq, and in a severe loss of total
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4460) * throughput. The best action to take is therefore to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4461) * serve the waker queue as soon as possible. So do it
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4462) * (without relying on the third alternative below for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4463) * eventually serving waker_bfqq's I/O; see the last
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4464) * paragraph for further details). This systematic
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4465) * injection of I/O from the waker queue does not
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4466) * cause any delay to bfqq's I/O. On the contrary,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4467) * next bfqq's I/O is brought forward dramatically,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4468) * for it is not blocked for milliseconds.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4469) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4470) * The third if checks whether bfqq is a queue for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4471) * which it is better to avoid injection. It is so if
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4472) * bfqq delivers more throughput when served without
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4473) * any further I/O from other queues in the middle, or
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4474) * if the service times of bfqq's I/O requests both
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4475) * count more than overall throughput, and may be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4476) * easily increased by injection (this happens if bfqq
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4477) * has a short think time). If none of these
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4478) * conditions holds, then a candidate queue for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4479) * injection is looked for through
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4480) * bfq_choose_bfqq_for_injection(). Note that the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4481) * latter may return NULL (for example if the inject
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4482) * limit for bfqq is currently 0).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4483) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4484) * NOTE: motivation for the second alternative
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4485) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4486) * Thanks to the way the inject limit is updated in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4487) * bfq_update_has_short_ttime(), it is rather likely
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4488) * that, if I/O is being plugged for bfqq and the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4489) * waker queue has pending I/O requests that are
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4490) * blocking bfqq's I/O, then the third alternative
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4491) * above lets the waker queue get served before the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4492) * I/O-plugging timeout fires. So one may deem the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4493) * second alternative superfluous. It is not, because
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4494) * the third alternative may be way less effective in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4495) * case of a synchronization. For two main
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4496) * reasons. First, throughput may be low because the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4497) * inject limit may be too low to guarantee the same
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4498) * amount of injected I/O, from the waker queue or
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4499) * other queues, that the second alternative
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4500) * guarantees (the second alternative unconditionally
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4501) * injects a pending I/O request of the waker queue
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4502) * for each bfq_dispatch_request()). Second, with the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4503) * third alternative, the duration of the plugging,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4504) * i.e., the time before bfqq finally receives new I/O,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4505) * may not be minimized, because the waker queue may
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4506) * happen to be served only after other queues.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4507) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4508) if (async_bfqq &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4509) icq_to_bic(async_bfqq->next_rq->elv.icq) == bfqq->bic &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4510) bfq_serv_to_charge(async_bfqq->next_rq, async_bfqq) <=
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4511) bfq_bfqq_budget_left(async_bfqq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4512) bfqq = bfqq->bic->bfqq[0];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4513) else if (bfq_bfqq_has_waker(bfqq) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4514) bfq_bfqq_busy(bfqq->waker_bfqq) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4515) bfqq->waker_bfqq->next_rq &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4516) bfq_serv_to_charge(bfqq->waker_bfqq->next_rq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4517) bfqq->waker_bfqq) <=
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4518) bfq_bfqq_budget_left(bfqq->waker_bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4519) )
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4520) bfqq = bfqq->waker_bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4521) else if (!idling_boosts_thr_without_issues(bfqd, bfqq) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4522) (bfqq->wr_coeff == 1 || bfqd->wr_busy_queues > 1 ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4523) !bfq_bfqq_has_short_ttime(bfqq)))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4524) bfqq = bfq_choose_bfqq_for_injection(bfqd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4525) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4526) bfqq = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4527)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4528) goto keep_queue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4529) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4530)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4531) reason = BFQQE_NO_MORE_REQUESTS;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4532) expire:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4533) bfq_bfqq_expire(bfqd, bfqq, false, reason);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4534) new_queue:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4535) bfqq = bfq_set_in_service_queue(bfqd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4536) if (bfqq) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4537) bfq_log_bfqq(bfqd, bfqq, "select_queue: checking new queue");
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4538) goto check_queue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4539) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4540) keep_queue:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4541) if (bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4542) bfq_log_bfqq(bfqd, bfqq, "select_queue: returned this queue");
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4543) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4544) bfq_log(bfqd, "select_queue: no queue returned");
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4545)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4546) return bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4547) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4548)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4549) static void bfq_update_wr_data(struct bfq_data *bfqd, struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4550) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4551) struct bfq_entity *entity = &bfqq->entity;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4552)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4553) if (bfqq->wr_coeff > 1) { /* queue is being weight-raised */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4554) bfq_log_bfqq(bfqd, bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4555) "raising period dur %u/%u msec, old coeff %u, w %d(%d)",
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4556) jiffies_to_msecs(jiffies - bfqq->last_wr_start_finish),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4557) jiffies_to_msecs(bfqq->wr_cur_max_time),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4558) bfqq->wr_coeff,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4559) bfqq->entity.weight, bfqq->entity.orig_weight);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4560)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4561) if (entity->prio_changed)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4562) bfq_log_bfqq(bfqd, bfqq, "WARN: pending prio change");
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4563)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4564) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4565) * If the queue was activated in a burst, or too much
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4566) * time has elapsed from the beginning of this
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4567) * weight-raising period, then end weight raising.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4568) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4569) if (bfq_bfqq_in_large_burst(bfqq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4570) bfq_bfqq_end_wr(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4571) else if (time_is_before_jiffies(bfqq->last_wr_start_finish +
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4572) bfqq->wr_cur_max_time)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4573) if (bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4574) time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4575) bfq_wr_duration(bfqd)))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4576) bfq_bfqq_end_wr(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4577) else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4578) switch_back_to_interactive_wr(bfqq, bfqd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4579) bfqq->entity.prio_changed = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4580) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4581) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4582) if (bfqq->wr_coeff > 1 &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4583) bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4584) bfqq->service_from_wr > max_service_from_wr) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4585) /* see comments on max_service_from_wr */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4586) bfq_bfqq_end_wr(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4587) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4588) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4589) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4590) * To improve latency (for this or other queues), immediately
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4591) * update weight both if it must be raised and if it must be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4592) * lowered. Since, entity may be on some active tree here, and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4593) * might have a pending change of its ioprio class, invoke
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4594) * next function with the last parameter unset (see the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4595) * comments on the function).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4596) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4597) if ((entity->weight > entity->orig_weight) != (bfqq->wr_coeff > 1))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4598) __bfq_entity_update_weight_prio(bfq_entity_service_tree(entity),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4599) entity, false);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4600) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4601)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4602) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4603) * Dispatch next request from bfqq.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4604) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4605) static struct request *bfq_dispatch_rq_from_bfqq(struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4606) struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4607) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4608) struct request *rq = bfqq->next_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4609) unsigned long service_to_charge;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4610)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4611) service_to_charge = bfq_serv_to_charge(rq, bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4612)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4613) bfq_bfqq_served(bfqq, service_to_charge);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4614)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4615) if (bfqq == bfqd->in_service_queue && bfqd->wait_dispatch) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4616) bfqd->wait_dispatch = false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4617) bfqd->waited_rq = rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4618) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4619)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4620) bfq_dispatch_remove(bfqd->queue, rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4621)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4622) if (bfqq != bfqd->in_service_queue)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4623) goto return_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4624)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4625) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4626) * If weight raising has to terminate for bfqq, then next
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4627) * function causes an immediate update of bfqq's weight,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4628) * without waiting for next activation. As a consequence, on
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4629) * expiration, bfqq will be timestamped as if has never been
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4630) * weight-raised during this service slot, even if it has
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4631) * received part or even most of the service as a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4632) * weight-raised queue. This inflates bfqq's timestamps, which
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4633) * is beneficial, as bfqq is then more willing to leave the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4634) * device immediately to possible other weight-raised queues.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4635) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4636) bfq_update_wr_data(bfqd, bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4637)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4638) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4639) * Expire bfqq, pretending that its budget expired, if bfqq
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4640) * belongs to CLASS_IDLE and other queues are waiting for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4641) * service.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4642) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4643) if (!(bfq_tot_busy_queues(bfqd) > 1 && bfq_class_idle(bfqq)))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4644) goto return_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4645)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4646) bfq_bfqq_expire(bfqd, bfqq, false, BFQQE_BUDGET_EXHAUSTED);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4647)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4648) return_rq:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4649) return rq;
^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) static bool bfq_has_work(struct blk_mq_hw_ctx *hctx)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4653) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4654) struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
^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) * Avoiding lock: a race on bfqd->busy_queues should cause at
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4658) * most a call to dispatch for nothing
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4659) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4660) return !list_empty_careful(&bfqd->dispatch) ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4661) bfq_tot_busy_queues(bfqd) > 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4662) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4663)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4664) static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4665) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4666) struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4667) struct request *rq = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4668) struct bfq_queue *bfqq = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4669)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4670) if (!list_empty(&bfqd->dispatch)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4671) rq = list_first_entry(&bfqd->dispatch, struct request,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4672) queuelist);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4673) list_del_init(&rq->queuelist);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4674)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4675) bfqq = RQ_BFQQ(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4676)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4677) if (bfqq) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4678) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4679) * Increment counters here, because this
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4680) * dispatch does not follow the standard
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4681) * dispatch flow (where counters are
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4682) * incremented)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4683) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4684) bfqq->dispatched++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4685)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4686) goto inc_in_driver_start_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4687) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4688)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4689) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4690) * We exploit the bfq_finish_requeue_request hook to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4691) * decrement rq_in_driver, but
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4692) * bfq_finish_requeue_request will not be invoked on
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4693) * this request. So, to avoid unbalance, just start
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4694) * this request, without incrementing rq_in_driver. As
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4695) * a negative consequence, rq_in_driver is deceptively
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4696) * lower than it should be while this request is in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4697) * service. This may cause bfq_schedule_dispatch to be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4698) * invoked uselessly.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4699) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4700) * As for implementing an exact solution, the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4701) * bfq_finish_requeue_request hook, if defined, is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4702) * probably invoked also on this request. So, by
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4703) * exploiting this hook, we could 1) increment
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4704) * rq_in_driver here, and 2) decrement it in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4705) * bfq_finish_requeue_request. Such a solution would
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4706) * let the value of the counter be always accurate,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4707) * but it would entail using an extra interface
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4708) * function. This cost seems higher than the benefit,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4709) * being the frequency of non-elevator-private
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4710) * requests very low.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4711) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4712) goto start_rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4713) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4714)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4715) bfq_log(bfqd, "dispatch requests: %d busy queues",
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4716) bfq_tot_busy_queues(bfqd));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4717)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4718) if (bfq_tot_busy_queues(bfqd) == 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4719) goto exit;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4720)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4721) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4722) * Force device to serve one request at a time if
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4723) * strict_guarantees is true. Forcing this service scheme is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4724) * currently the ONLY way to guarantee that the request
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4725) * service order enforced by the scheduler is respected by a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4726) * queueing device. Otherwise the device is free even to make
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4727) * some unlucky request wait for as long as the device
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4728) * wishes.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4729) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4730) * Of course, serving one request at a time may cause loss of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4731) * throughput.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4732) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4733) if (bfqd->strict_guarantees && bfqd->rq_in_driver > 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4734) goto exit;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4735)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4736) bfqq = bfq_select_queue(bfqd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4737) if (!bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4738) goto exit;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4739)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4740) rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4741)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4742) if (rq) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4743) inc_in_driver_start_rq:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4744) bfqd->rq_in_driver++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4745) start_rq:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4746) rq->rq_flags |= RQF_STARTED;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4747) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4748) exit:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4749) return rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4750) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4751)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4752) #ifdef CONFIG_BFQ_CGROUP_DEBUG
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4753) static void bfq_update_dispatch_stats(struct request_queue *q,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4754) struct request *rq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4755) struct bfq_queue *in_serv_queue,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4756) bool idle_timer_disabled)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4757) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4758) struct bfq_queue *bfqq = rq ? RQ_BFQQ(rq) : NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4759)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4760) if (!idle_timer_disabled && !bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4761) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4762)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4763) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4764) * rq and bfqq are guaranteed to exist until this function
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4765) * ends, for the following reasons. First, rq can be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4766) * dispatched to the device, and then can be completed and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4767) * freed, only after this function ends. Second, rq cannot be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4768) * merged (and thus freed because of a merge) any longer,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4769) * because it has already started. Thus rq cannot be freed
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4770) * before this function ends, and, since rq has a reference to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4771) * bfqq, the same guarantee holds for bfqq too.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4772) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4773) * In addition, the following queue lock guarantees that
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4774) * bfqq_group(bfqq) exists as well.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4775) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4776) spin_lock_irq(&q->queue_lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4777) if (idle_timer_disabled)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4778) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4779) * Since the idle timer has been disabled,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4780) * in_serv_queue contained some request when
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4781) * __bfq_dispatch_request was invoked above, which
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4782) * implies that rq was picked exactly from
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4783) * in_serv_queue. Thus in_serv_queue == bfqq, and is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4784) * therefore guaranteed to exist because of the above
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4785) * arguments.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4786) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4787) bfqg_stats_update_idle_time(bfqq_group(in_serv_queue));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4788) if (bfqq) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4789) struct bfq_group *bfqg = bfqq_group(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4790)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4791) bfqg_stats_update_avg_queue_size(bfqg);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4792) bfqg_stats_set_start_empty_time(bfqg);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4793) bfqg_stats_update_io_remove(bfqg, rq->cmd_flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4794) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4795) spin_unlock_irq(&q->queue_lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4796) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4797) #else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4798) static inline void bfq_update_dispatch_stats(struct request_queue *q,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4799) struct request *rq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4800) struct bfq_queue *in_serv_queue,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4801) bool idle_timer_disabled) {}
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4802) #endif /* CONFIG_BFQ_CGROUP_DEBUG */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4803)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4804) static struct request *bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4805) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4806) struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4807) struct request *rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4808) struct bfq_queue *in_serv_queue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4809) bool waiting_rq, idle_timer_disabled = false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4810)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4811) spin_lock_irq(&bfqd->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4812)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4813) in_serv_queue = bfqd->in_service_queue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4814) waiting_rq = in_serv_queue && bfq_bfqq_wait_request(in_serv_queue);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4815)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4816) rq = __bfq_dispatch_request(hctx);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4817) if (in_serv_queue == bfqd->in_service_queue) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4818) idle_timer_disabled =
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4819) waiting_rq && !bfq_bfqq_wait_request(in_serv_queue);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4820) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4821)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4822) spin_unlock_irq(&bfqd->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4823) bfq_update_dispatch_stats(hctx->queue, rq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4824) idle_timer_disabled ? in_serv_queue : NULL,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4825) idle_timer_disabled);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4826)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4827) return rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4828) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4829)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4830) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4831) * Task holds one reference to the queue, dropped when task exits. Each rq
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4832) * in-flight on this queue also holds a reference, dropped when rq is freed.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4833) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4834) * Scheduler lock must be held here. Recall not to use bfqq after calling
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4835) * this function on it.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4836) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4837) void bfq_put_queue(struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4838) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4839) struct bfq_queue *item;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4840) struct hlist_node *n;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4841) struct bfq_group *bfqg = bfqq_group(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4842)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4843) if (bfqq->bfqd)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4844) bfq_log_bfqq(bfqq->bfqd, bfqq, "put_queue: %p %d",
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4845) bfqq, bfqq->ref);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4846)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4847) bfqq->ref--;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4848) if (bfqq->ref)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4849) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4850)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4851) if (!hlist_unhashed(&bfqq->burst_list_node)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4852) hlist_del_init(&bfqq->burst_list_node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4853) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4854) * Decrement also burst size after the removal, if the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4855) * process associated with bfqq is exiting, and thus
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4856) * does not contribute to the burst any longer. This
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4857) * decrement helps filter out false positives of large
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4858) * bursts, when some short-lived process (often due to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4859) * the execution of commands by some service) happens
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4860) * to start and exit while a complex application is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4861) * starting, and thus spawning several processes that
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4862) * do I/O (and that *must not* be treated as a large
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4863) * burst, see comments on bfq_handle_burst).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4864) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4865) * In particular, the decrement is performed only if:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4866) * 1) bfqq is not a merged queue, because, if it is,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4867) * then this free of bfqq is not triggered by the exit
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4868) * of the process bfqq is associated with, but exactly
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4869) * by the fact that bfqq has just been merged.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4870) * 2) burst_size is greater than 0, to handle
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4871) * unbalanced decrements. Unbalanced decrements may
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4872) * happen in te following case: bfqq is inserted into
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4873) * the current burst list--without incrementing
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4874) * bust_size--because of a split, but the current
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4875) * burst list is not the burst list bfqq belonged to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4876) * (see comments on the case of a split in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4877) * bfq_set_request).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4878) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4879) if (bfqq->bic && bfqq->bfqd->burst_size > 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4880) bfqq->bfqd->burst_size--;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4881) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4882)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4883) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4884) * bfqq does not exist any longer, so it cannot be woken by
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4885) * any other queue, and cannot wake any other queue. Then bfqq
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4886) * must be removed from the woken list of its possible waker
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4887) * queue, and all queues in the woken list of bfqq must stop
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4888) * having a waker queue. Strictly speaking, these updates
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4889) * should be performed when bfqq remains with no I/O source
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4890) * attached to it, which happens before bfqq gets freed. In
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4891) * particular, this happens when the last process associated
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4892) * with bfqq exits or gets associated with a different
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4893) * queue. However, both events lead to bfqq being freed soon,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4894) * and dangling references would come out only after bfqq gets
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4895) * freed. So these updates are done here, as a simple and safe
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4896) * way to handle all cases.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4897) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4898) /* remove bfqq from woken list */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4899) if (!hlist_unhashed(&bfqq->woken_list_node))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4900) hlist_del_init(&bfqq->woken_list_node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4901)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4902) /* reset waker for all queues in woken list */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4903) hlist_for_each_entry_safe(item, n, &bfqq->woken_list,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4904) woken_list_node) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4905) item->waker_bfqq = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4906) bfq_clear_bfqq_has_waker(item);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4907) hlist_del_init(&item->woken_list_node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4908) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4909)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4910) if (bfqq->bfqd && bfqq->bfqd->last_completed_rq_bfqq == bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4911) bfqq->bfqd->last_completed_rq_bfqq = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4912)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4913) kmem_cache_free(bfq_pool, bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4914) bfqg_and_blkg_put(bfqg);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4915) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4916)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4917) static void bfq_put_cooperator(struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4918) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4919) struct bfq_queue *__bfqq, *next;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4920)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4921) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4922) * If this queue was scheduled to merge with another queue, be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4923) * sure to drop the reference taken on that queue (and others in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4924) * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4925) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4926) __bfqq = bfqq->new_bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4927) while (__bfqq) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4928) if (__bfqq == bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4929) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4930) next = __bfqq->new_bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4931) bfq_put_queue(__bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4932) __bfqq = next;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4933) }
^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) static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4937) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4938) if (bfqq == bfqd->in_service_queue) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4939) __bfq_bfqq_expire(bfqd, bfqq, BFQQE_BUDGET_TIMEOUT);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4940) bfq_schedule_dispatch(bfqd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4941) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4942)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4943) bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, bfqq->ref);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4944)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4945) bfq_put_cooperator(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4946)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4947) bfq_release_process_ref(bfqd, bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4948) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4949)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4950) static void bfq_exit_icq_bfqq(struct bfq_io_cq *bic, bool is_sync)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4951) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4952) struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4953) struct bfq_data *bfqd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4954)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4955) if (bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4956) bfqd = bfqq->bfqd; /* NULL if scheduler already exited */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4957)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4958) if (bfqq && bfqd) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4959) unsigned long flags;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4960)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4961) spin_lock_irqsave(&bfqd->lock, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4962) bfqq->bic = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4963) bfq_exit_bfqq(bfqd, bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4964) bic_set_bfqq(bic, NULL, is_sync);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4965) spin_unlock_irqrestore(&bfqd->lock, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4966) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4967) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4968)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4969) static void bfq_exit_icq(struct io_cq *icq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4970) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4971) struct bfq_io_cq *bic = icq_to_bic(icq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4972)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4973) bfq_exit_icq_bfqq(bic, true);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4974) bfq_exit_icq_bfqq(bic, false);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4975) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4976)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4977) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4978) * Update the entity prio values; note that the new values will not
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4979) * be used until the next (re)activation.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4980) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4981) static void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4982) bfq_set_next_ioprio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4983) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4984) struct task_struct *tsk = current;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4985) int ioprio_class;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4986) struct bfq_data *bfqd = bfqq->bfqd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4987)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4988) if (!bfqd)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4989) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4990)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4991) ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4992) switch (ioprio_class) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4993) default:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4994) pr_err("bdi %s: bfq: bad prio class %d\n",
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4995) bdi_dev_name(bfqq->bfqd->queue->backing_dev_info),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4996) ioprio_class);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4997) fallthrough;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4998) case IOPRIO_CLASS_NONE:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 4999) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5000) * No prio set, inherit CPU scheduling settings.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5001) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5002) bfqq->new_ioprio = task_nice_ioprio(tsk);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5003) bfqq->new_ioprio_class = task_nice_ioclass(tsk);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5004) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5005) case IOPRIO_CLASS_RT:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5006) bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5007) bfqq->new_ioprio_class = IOPRIO_CLASS_RT;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5008) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5009) case IOPRIO_CLASS_BE:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5010) bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5011) bfqq->new_ioprio_class = IOPRIO_CLASS_BE;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5012) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5013) case IOPRIO_CLASS_IDLE:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5014) bfqq->new_ioprio_class = IOPRIO_CLASS_IDLE;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5015) bfqq->new_ioprio = 7;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5016) break;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5017) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5018)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5019) if (bfqq->new_ioprio >= IOPRIO_BE_NR) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5020) pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n",
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5021) bfqq->new_ioprio);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5022) bfqq->new_ioprio = IOPRIO_BE_NR - 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5023) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5024)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5025) bfqq->entity.new_weight = bfq_ioprio_to_weight(bfqq->new_ioprio);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5026) bfqq->entity.prio_changed = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5027) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5028)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5029) static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5030) struct bio *bio, bool is_sync,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5031) struct bfq_io_cq *bic);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5032)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5033) static void bfq_check_ioprio_change(struct bfq_io_cq *bic, struct bio *bio)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5034) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5035) struct bfq_data *bfqd = bic_to_bfqd(bic);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5036) struct bfq_queue *bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5037) int ioprio = bic->icq.ioc->ioprio;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5038)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5039) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5040) * This condition may trigger on a newly created bic, be sure to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5041) * drop the lock before returning.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5042) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5043) if (unlikely(!bfqd) || likely(bic->ioprio == ioprio))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5044) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5045)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5046) bic->ioprio = ioprio;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5047)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5048) bfqq = bic_to_bfqq(bic, false);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5049) if (bfqq) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5050) bfq_release_process_ref(bfqd, bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5051) bfqq = bfq_get_queue(bfqd, bio, BLK_RW_ASYNC, bic);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5052) bic_set_bfqq(bic, bfqq, false);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5053) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5054)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5055) bfqq = bic_to_bfqq(bic, true);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5056) if (bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5057) bfq_set_next_ioprio_data(bfqq, bic);
^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) static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5061) struct bfq_io_cq *bic, pid_t pid, int is_sync)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5062) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5063) RB_CLEAR_NODE(&bfqq->entity.rb_node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5064) INIT_LIST_HEAD(&bfqq->fifo);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5065) INIT_HLIST_NODE(&bfqq->burst_list_node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5066) INIT_HLIST_NODE(&bfqq->woken_list_node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5067) INIT_HLIST_HEAD(&bfqq->woken_list);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5068)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5069) bfqq->ref = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5070) bfqq->bfqd = bfqd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5071)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5072) if (bic)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5073) bfq_set_next_ioprio_data(bfqq, bic);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5074)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5075) if (is_sync) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5076) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5077) * No need to mark as has_short_ttime if in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5078) * idle_class, because no device idling is performed
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5079) * for queues in idle class
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5080) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5081) if (!bfq_class_idle(bfqq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5082) /* tentatively mark as has_short_ttime */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5083) bfq_mark_bfqq_has_short_ttime(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5084) bfq_mark_bfqq_sync(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5085) bfq_mark_bfqq_just_created(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5086) } else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5087) bfq_clear_bfqq_sync(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5088)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5089) /* set end request to minus infinity from now */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5090) bfqq->ttime.last_end_request = ktime_get_ns() + 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5091)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5092) bfq_mark_bfqq_IO_bound(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5093)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5094) bfqq->pid = pid;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5095)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5096) /* Tentative initial value to trade off between thr and lat */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5097) bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5098) bfqq->budget_timeout = bfq_smallest_from_now();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5099)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5100) bfqq->wr_coeff = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5101) bfqq->last_wr_start_finish = jiffies;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5102) bfqq->wr_start_at_switch_to_srt = bfq_smallest_from_now();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5103) bfqq->split_time = bfq_smallest_from_now();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5104)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5105) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5106) * To not forget the possibly high bandwidth consumed by a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5107) * process/queue in the recent past,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5108) * bfq_bfqq_softrt_next_start() returns a value at least equal
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5109) * to the current value of bfqq->soft_rt_next_start (see
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5110) * comments on bfq_bfqq_softrt_next_start). Set
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5111) * soft_rt_next_start to now, to mean that bfqq has consumed
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5112) * no bandwidth so far.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5113) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5114) bfqq->soft_rt_next_start = jiffies;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5115)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5116) /* first request is almost certainly seeky */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5117) bfqq->seek_history = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5118) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5119)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5120) static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5121) struct bfq_group *bfqg,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5122) int ioprio_class, int ioprio)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5123) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5124) switch (ioprio_class) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5125) case IOPRIO_CLASS_RT:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5126) return &bfqg->async_bfqq[0][ioprio];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5127) case IOPRIO_CLASS_NONE:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5128) ioprio = IOPRIO_NORM;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5129) fallthrough;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5130) case IOPRIO_CLASS_BE:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5131) return &bfqg->async_bfqq[1][ioprio];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5132) case IOPRIO_CLASS_IDLE:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5133) return &bfqg->async_idle_bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5134) default:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5135) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5136) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5137) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5138)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5139) static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5140) struct bio *bio, bool is_sync,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5141) struct bfq_io_cq *bic)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5142) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5143) const int ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5144) const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5145) struct bfq_queue **async_bfqq = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5146) struct bfq_queue *bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5147) struct bfq_group *bfqg;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5148)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5149) rcu_read_lock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5150)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5151) bfqg = bfq_find_set_group(bfqd, __bio_blkcg(bio));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5152) if (!bfqg) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5153) bfqq = &bfqd->oom_bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5154) goto out;
^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) if (!is_sync) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5158) async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5159) ioprio);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5160) bfqq = *async_bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5161) if (bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5162) goto out;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5163) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5164)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5165) bfqq = kmem_cache_alloc_node(bfq_pool,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5166) GFP_NOWAIT | __GFP_ZERO | __GFP_NOWARN,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5167) bfqd->queue->node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5168)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5169) if (bfqq) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5170) bfq_init_bfqq(bfqd, bfqq, bic, current->pid,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5171) is_sync);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5172) bfq_init_entity(&bfqq->entity, bfqg);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5173) bfq_log_bfqq(bfqd, bfqq, "allocated");
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5174) } else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5175) bfqq = &bfqd->oom_bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5176) bfq_log_bfqq(bfqd, bfqq, "using oom bfqq");
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5177) goto out;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5178) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5179)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5180) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5181) * Pin the queue now that it's allocated, scheduler exit will
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5182) * prune it.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5183) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5184) if (async_bfqq) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5185) bfqq->ref++; /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5186) * Extra group reference, w.r.t. sync
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5187) * queue. This extra reference is removed
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5188) * only if bfqq->bfqg disappears, to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5189) * guarantee that this queue is not freed
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5190) * until its group goes away.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5191) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5192) bfq_log_bfqq(bfqd, bfqq, "get_queue, bfqq not in async: %p, %d",
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5193) bfqq, bfqq->ref);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5194) *async_bfqq = bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5195) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5196)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5197) out:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5198) bfqq->ref++; /* get a process reference to this queue */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5199) bfq_log_bfqq(bfqd, bfqq, "get_queue, at end: %p, %d", bfqq, bfqq->ref);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5200) rcu_read_unlock();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5201) return bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5202) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5203)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5204) static void bfq_update_io_thinktime(struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5205) struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5206) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5207) struct bfq_ttime *ttime = &bfqq->ttime;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5208) u64 elapsed = ktime_get_ns() - bfqq->ttime.last_end_request;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5209)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5210) elapsed = min_t(u64, elapsed, 2ULL * bfqd->bfq_slice_idle);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5211)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5212) ttime->ttime_samples = (7*bfqq->ttime.ttime_samples + 256) / 8;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5213) ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed, 8);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5214) ttime->ttime_mean = div64_ul(ttime->ttime_total + 128,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5215) ttime->ttime_samples);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5216) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5217)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5218) static void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5219) bfq_update_io_seektime(struct bfq_data *bfqd, struct bfq_queue *bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5220) struct request *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5221) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5222) bfqq->seek_history <<= 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5223) bfqq->seek_history |= BFQ_RQ_SEEKY(bfqd, bfqq->last_request_pos, rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5224)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5225) if (bfqq->wr_coeff > 1 &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5226) bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5227) BFQQ_TOTALLY_SEEKY(bfqq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5228) bfq_bfqq_end_wr(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5229) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5230)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5231) static void bfq_update_has_short_ttime(struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5232) struct bfq_queue *bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5233) struct bfq_io_cq *bic)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5234) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5235) bool has_short_ttime = true, state_changed;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5236)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5237) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5238) * No need to update has_short_ttime if bfqq is async or in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5239) * idle io prio class, or if bfq_slice_idle is zero, because
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5240) * no device idling is performed for bfqq in this case.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5241) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5242) if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq) ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5243) bfqd->bfq_slice_idle == 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5244) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5245)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5246) /* Idle window just restored, statistics are meaningless. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5247) if (time_is_after_eq_jiffies(bfqq->split_time +
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5248) bfqd->bfq_wr_min_idle_time))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5249) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5250)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5251) /* Think time is infinite if no process is linked to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5252) * bfqq. Otherwise check average think time to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5253) * decide whether to mark as has_short_ttime
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5254) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5255) if (atomic_read(&bic->icq.ioc->active_ref) == 0 ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5256) (bfq_sample_valid(bfqq->ttime.ttime_samples) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5257) bfqq->ttime.ttime_mean > bfqd->bfq_slice_idle))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5258) has_short_ttime = false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5259)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5260) state_changed = has_short_ttime != bfq_bfqq_has_short_ttime(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5261)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5262) if (has_short_ttime)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5263) bfq_mark_bfqq_has_short_ttime(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5264) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5265) bfq_clear_bfqq_has_short_ttime(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5266)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5267) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5268) * Until the base value for the total service time gets
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5269) * finally computed for bfqq, the inject limit does depend on
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5270) * the think-time state (short|long). In particular, the limit
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5271) * is 0 or 1 if the think time is deemed, respectively, as
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5272) * short or long (details in the comments in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5273) * bfq_update_inject_limit()). Accordingly, the next
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5274) * instructions reset the inject limit if the think-time state
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5275) * has changed and the above base value is still to be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5276) * computed.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5277) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5278) * However, the reset is performed only if more than 100 ms
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5279) * have elapsed since the last update of the inject limit, or
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5280) * (inclusive) if the change is from short to long think
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5281) * time. The reason for this waiting is as follows.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5282) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5283) * bfqq may have a long think time because of a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5284) * synchronization with some other queue, i.e., because the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5285) * I/O of some other queue may need to be completed for bfqq
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5286) * to receive new I/O. Details in the comments on the choice
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5287) * of the queue for injection in bfq_select_queue().
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5288) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5289) * As stressed in those comments, if such a synchronization is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5290) * actually in place, then, without injection on bfqq, the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5291) * blocking I/O cannot happen to served while bfqq is in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5292) * service. As a consequence, if bfqq is granted
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5293) * I/O-dispatch-plugging, then bfqq remains empty, and no I/O
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5294) * is dispatched, until the idle timeout fires. This is likely
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5295) * to result in lower bandwidth and higher latencies for bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5296) * and in a severe loss of total throughput.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5297) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5298) * On the opposite end, a non-zero inject limit may allow the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5299) * I/O that blocks bfqq to be executed soon, and therefore
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5300) * bfqq to receive new I/O soon.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5301) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5302) * But, if the blocking gets actually eliminated, then the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5303) * next think-time sample for bfqq may be very low. This in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5304) * turn may cause bfqq's think time to be deemed
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5305) * short. Without the 100 ms barrier, this new state change
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5306) * would cause the body of the next if to be executed
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5307) * immediately. But this would set to 0 the inject
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5308) * limit. Without injection, the blocking I/O would cause the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5309) * think time of bfqq to become long again, and therefore the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5310) * inject limit to be raised again, and so on. The only effect
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5311) * of such a steady oscillation between the two think-time
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5312) * states would be to prevent effective injection on bfqq.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5313) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5314) * In contrast, if the inject limit is not reset during such a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5315) * long time interval as 100 ms, then the number of short
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5316) * think time samples can grow significantly before the reset
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5317) * is performed. As a consequence, the think time state can
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5318) * become stable before the reset. Therefore there will be no
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5319) * state change when the 100 ms elapse, and no reset of the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5320) * inject limit. The inject limit remains steadily equal to 1
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5321) * both during and after the 100 ms. So injection can be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5322) * performed at all times, and throughput gets boosted.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5323) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5324) * An inject limit equal to 1 is however in conflict, in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5325) * general, with the fact that the think time of bfqq is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5326) * short, because injection may be likely to delay bfqq's I/O
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5327) * (as explained in the comments in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5328) * bfq_update_inject_limit()). But this does not happen in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5329) * this special case, because bfqq's low think time is due to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5330) * an effective handling of a synchronization, through
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5331) * injection. In this special case, bfqq's I/O does not get
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5332) * delayed by injection; on the contrary, bfqq's I/O is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5333) * brought forward, because it is not blocked for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5334) * milliseconds.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5335) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5336) * In addition, serving the blocking I/O much sooner, and much
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5337) * more frequently than once per I/O-plugging timeout, makes
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5338) * it much quicker to detect a waker queue (the concept of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5339) * waker queue is defined in the comments in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5340) * bfq_add_request()). This makes it possible to start sooner
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5341) * to boost throughput more effectively, by injecting the I/O
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5342) * of the waker queue unconditionally on every
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5343) * bfq_dispatch_request().
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5344) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5345) * One last, important benefit of not resetting the inject
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5346) * limit before 100 ms is that, during this time interval, the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5347) * base value for the total service time is likely to get
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5348) * finally computed for bfqq, freeing the inject limit from
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5349) * its relation with the think time.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5350) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5351) if (state_changed && bfqq->last_serv_time_ns == 0 &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5352) (time_is_before_eq_jiffies(bfqq->decrease_time_jif +
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5353) msecs_to_jiffies(100)) ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5354) !has_short_ttime))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5355) bfq_reset_inject_limit(bfqd, bfqq);
^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) * Called when a new fs request (rq) is added to bfqq. Check if there's
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5360) * something we should do about it.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5361) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5362) static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5363) struct request *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5364) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5365) if (rq->cmd_flags & REQ_META)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5366) bfqq->meta_pending++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5367)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5368) bfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5369)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5370) if (bfqq == bfqd->in_service_queue && bfq_bfqq_wait_request(bfqq)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5371) bool small_req = bfqq->queued[rq_is_sync(rq)] == 1 &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5372) blk_rq_sectors(rq) < 32;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5373) bool budget_timeout = bfq_bfqq_budget_timeout(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5374)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5375) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5376) * There is just this request queued: if
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5377) * - the request is small, and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5378) * - we are idling to boost throughput, and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5379) * - the queue is not to be expired,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5380) * then just exit.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5381) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5382) * In this way, if the device is being idled to wait
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5383) * for a new request from the in-service queue, we
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5384) * avoid unplugging the device and committing the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5385) * device to serve just a small request. In contrast
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5386) * we wait for the block layer to decide when to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5387) * unplug the device: hopefully, new requests will be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5388) * merged to this one quickly, then the device will be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5389) * unplugged and larger requests will be dispatched.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5390) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5391) if (small_req && idling_boosts_thr_without_issues(bfqd, bfqq) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5392) !budget_timeout)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5393) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5394)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5395) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5396) * A large enough request arrived, or idling is being
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5397) * performed to preserve service guarantees, or
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5398) * finally the queue is to be expired: in all these
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5399) * cases disk idling is to be stopped, so clear
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5400) * wait_request flag and reset timer.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5401) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5402) bfq_clear_bfqq_wait_request(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5403) hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5404)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5405) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5406) * The queue is not empty, because a new request just
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5407) * arrived. Hence we can safely expire the queue, in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5408) * case of budget timeout, without risking that the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5409) * timestamps of the queue are not updated correctly.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5410) * See [1] for more details.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5411) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5412) if (budget_timeout)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5413) bfq_bfqq_expire(bfqd, bfqq, false,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5414) BFQQE_BUDGET_TIMEOUT);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5415) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5416) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5417)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5418) /* returns true if it causes the idle timer to be disabled */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5419) static bool __bfq_insert_request(struct bfq_data *bfqd, struct request *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5420) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5421) struct bfq_queue *bfqq = RQ_BFQQ(rq),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5422) *new_bfqq = bfq_setup_cooperator(bfqd, bfqq, rq, true);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5423) bool waiting, idle_timer_disabled = false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5424)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5425) if (new_bfqq) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5426) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5427) * Release the request's reference to the old bfqq
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5428) * and make sure one is taken to the shared queue.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5429) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5430) new_bfqq->allocated++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5431) bfqq->allocated--;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5432) new_bfqq->ref++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5433) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5434) * If the bic associated with the process
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5435) * issuing this request still points to bfqq
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5436) * (and thus has not been already redirected
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5437) * to new_bfqq or even some other bfq_queue),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5438) * then complete the merge and redirect it to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5439) * new_bfqq.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5440) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5441) if (bic_to_bfqq(RQ_BIC(rq), 1) == bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5442) bfq_merge_bfqqs(bfqd, RQ_BIC(rq),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5443) bfqq, new_bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5444)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5445) bfq_clear_bfqq_just_created(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5446) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5447) * rq is about to be enqueued into new_bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5448) * release rq reference on bfqq
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5449) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5450) bfq_put_queue(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5451) rq->elv.priv[1] = new_bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5452) bfqq = new_bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5453) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5454)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5455) bfq_update_io_thinktime(bfqd, bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5456) bfq_update_has_short_ttime(bfqd, bfqq, RQ_BIC(rq));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5457) bfq_update_io_seektime(bfqd, bfqq, rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5458)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5459) waiting = bfqq && bfq_bfqq_wait_request(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5460) bfq_add_request(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5461) idle_timer_disabled = waiting && !bfq_bfqq_wait_request(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5462)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5463) rq->fifo_time = ktime_get_ns() + bfqd->bfq_fifo_expire[rq_is_sync(rq)];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5464) list_add_tail(&rq->queuelist, &bfqq->fifo);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5465)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5466) bfq_rq_enqueued(bfqd, bfqq, rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5467)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5468) return idle_timer_disabled;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5469) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5470)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5471) #ifdef CONFIG_BFQ_CGROUP_DEBUG
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5472) static void bfq_update_insert_stats(struct request_queue *q,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5473) struct bfq_queue *bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5474) bool idle_timer_disabled,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5475) unsigned int cmd_flags)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5476) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5477) if (!bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5478) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5479)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5480) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5481) * bfqq still exists, because it can disappear only after
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5482) * either it is merged with another queue, or the process it
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5483) * is associated with exits. But both actions must be taken by
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5484) * the same process currently executing this flow of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5485) * instructions.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5486) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5487) * In addition, the following queue lock guarantees that
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5488) * bfqq_group(bfqq) exists as well.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5489) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5490) spin_lock_irq(&q->queue_lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5491) bfqg_stats_update_io_add(bfqq_group(bfqq), bfqq, cmd_flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5492) if (idle_timer_disabled)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5493) bfqg_stats_update_idle_time(bfqq_group(bfqq));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5494) spin_unlock_irq(&q->queue_lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5495) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5496) #else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5497) static inline void bfq_update_insert_stats(struct request_queue *q,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5498) struct bfq_queue *bfqq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5499) bool idle_timer_disabled,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5500) unsigned int cmd_flags) {}
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5501) #endif /* CONFIG_BFQ_CGROUP_DEBUG */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5502)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5503) static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5504) bool at_head)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5505) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5506) struct request_queue *q = hctx->queue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5507) struct bfq_data *bfqd = q->elevator->elevator_data;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5508) struct bfq_queue *bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5509) bool idle_timer_disabled = false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5510) unsigned int cmd_flags;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5511)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5512) #ifdef CONFIG_BFQ_GROUP_IOSCHED
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5513) if (!cgroup_subsys_on_dfl(io_cgrp_subsys) && rq->bio)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5514) bfqg_stats_update_legacy_io(q, rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5515) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5516) spin_lock_irq(&bfqd->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5517) if (blk_mq_sched_try_insert_merge(q, rq)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5518) spin_unlock_irq(&bfqd->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5519) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5520) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5521)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5522) spin_unlock_irq(&bfqd->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5523)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5524) blk_mq_sched_request_inserted(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5525)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5526) spin_lock_irq(&bfqd->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5527) bfqq = bfq_init_rq(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5528) if (!bfqq || at_head || blk_rq_is_passthrough(rq)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5529) if (at_head)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5530) list_add(&rq->queuelist, &bfqd->dispatch);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5531) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5532) list_add_tail(&rq->queuelist, &bfqd->dispatch);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5533) } else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5534) idle_timer_disabled = __bfq_insert_request(bfqd, rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5535) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5536) * Update bfqq, because, if a queue merge has occurred
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5537) * in __bfq_insert_request, then rq has been
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5538) * redirected into a new queue.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5539) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5540) bfqq = RQ_BFQQ(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5541)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5542) if (rq_mergeable(rq)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5543) elv_rqhash_add(q, rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5544) if (!q->last_merge)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5545) q->last_merge = rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5546) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5547) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5548)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5549) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5550) * Cache cmd_flags before releasing scheduler lock, because rq
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5551) * may disappear afterwards (for example, because of a request
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5552) * merge).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5553) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5554) cmd_flags = rq->cmd_flags;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5555)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5556) spin_unlock_irq(&bfqd->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5557)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5558) bfq_update_insert_stats(q, bfqq, idle_timer_disabled,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5559) cmd_flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5560) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5561)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5562) static void bfq_insert_requests(struct blk_mq_hw_ctx *hctx,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5563) struct list_head *list, bool at_head)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5564) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5565) while (!list_empty(list)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5566) struct request *rq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5567)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5568) rq = list_first_entry(list, struct request, queuelist);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5569) list_del_init(&rq->queuelist);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5570) bfq_insert_request(hctx, rq, at_head);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5571) }
^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) static void bfq_update_hw_tag(struct bfq_data *bfqd)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5575) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5576) struct bfq_queue *bfqq = bfqd->in_service_queue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5577)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5578) bfqd->max_rq_in_driver = max_t(int, bfqd->max_rq_in_driver,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5579) bfqd->rq_in_driver);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5580)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5581) if (bfqd->hw_tag == 1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5582) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5583)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5584) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5585) * This sample is valid if the number of outstanding requests
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5586) * is large enough to allow a queueing behavior. Note that the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5587) * sum is not exact, as it's not taking into account deactivated
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5588) * requests.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5589) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5590) if (bfqd->rq_in_driver + bfqd->queued <= BFQ_HW_QUEUE_THRESHOLD)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5591) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5592)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5593) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5594) * If active queue hasn't enough requests and can idle, bfq might not
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5595) * dispatch sufficient requests to hardware. Don't zero hw_tag in this
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5596) * case
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5597) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5598) if (bfqq && bfq_bfqq_has_short_ttime(bfqq) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5599) bfqq->dispatched + bfqq->queued[0] + bfqq->queued[1] <
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5600) BFQ_HW_QUEUE_THRESHOLD &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5601) bfqd->rq_in_driver < BFQ_HW_QUEUE_THRESHOLD)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5602) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5603)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5604) if (bfqd->hw_tag_samples++ < BFQ_HW_QUEUE_SAMPLES)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5605) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5606)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5607) bfqd->hw_tag = bfqd->max_rq_in_driver > BFQ_HW_QUEUE_THRESHOLD;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5608) bfqd->max_rq_in_driver = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5609) bfqd->hw_tag_samples = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5610)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5611) bfqd->nonrot_with_queueing =
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5612) blk_queue_nonrot(bfqd->queue) && bfqd->hw_tag;
^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) static void bfq_completed_request(struct bfq_queue *bfqq, struct bfq_data *bfqd)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5616) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5617) u64 now_ns;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5618) u32 delta_us;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5619)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5620) bfq_update_hw_tag(bfqd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5621)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5622) bfqd->rq_in_driver--;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5623) bfqq->dispatched--;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5624)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5625) if (!bfqq->dispatched && !bfq_bfqq_busy(bfqq)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5626) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5627) * Set budget_timeout (which we overload to store the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5628) * time at which the queue remains with no backlog and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5629) * no outstanding request; used by the weight-raising
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5630) * mechanism).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5631) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5632) bfqq->budget_timeout = jiffies;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5633)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5634) bfq_weights_tree_remove(bfqd, bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5635) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5636)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5637) now_ns = ktime_get_ns();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5638)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5639) bfqq->ttime.last_end_request = now_ns;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5640)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5641) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5642) * Using us instead of ns, to get a reasonable precision in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5643) * computing rate in next check.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5644) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5645) delta_us = div_u64(now_ns - bfqd->last_completion, NSEC_PER_USEC);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5646)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5647) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5648) * If the request took rather long to complete, and, according
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5649) * to the maximum request size recorded, this completion latency
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5650) * implies that the request was certainly served at a very low
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5651) * rate (less than 1M sectors/sec), then the whole observation
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5652) * interval that lasts up to this time instant cannot be a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5653) * valid time interval for computing a new peak rate. Invoke
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5654) * bfq_update_rate_reset to have the following three steps
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5655) * taken:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5656) * - close the observation interval at the last (previous)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5657) * request dispatch or completion
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5658) * - compute rate, if possible, for that observation interval
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5659) * - reset to zero samples, which will trigger a proper
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5660) * re-initialization of the observation interval on next
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5661) * dispatch
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5662) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5663) if (delta_us > BFQ_MIN_TT/NSEC_PER_USEC &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5664) (bfqd->last_rq_max_size<<BFQ_RATE_SHIFT)/delta_us <
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5665) 1UL<<(BFQ_RATE_SHIFT - 10))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5666) bfq_update_rate_reset(bfqd, NULL);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5667) bfqd->last_completion = now_ns;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5668) bfqd->last_completed_rq_bfqq = bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5669)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5670) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5671) * If we are waiting to discover whether the request pattern
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5672) * of the task associated with the queue is actually
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5673) * isochronous, and both requisites for this condition to hold
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5674) * are now satisfied, then compute soft_rt_next_start (see the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5675) * comments on the function bfq_bfqq_softrt_next_start()). We
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5676) * do not compute soft_rt_next_start if bfqq is in interactive
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5677) * weight raising (see the comments in bfq_bfqq_expire() for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5678) * an explanation). We schedule this delayed update when bfqq
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5679) * expires, if it still has in-flight requests.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5680) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5681) if (bfq_bfqq_softrt_update(bfqq) && bfqq->dispatched == 0 &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5682) RB_EMPTY_ROOT(&bfqq->sort_list) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5683) bfqq->wr_coeff != bfqd->bfq_wr_coeff)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5684) bfqq->soft_rt_next_start =
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5685) bfq_bfqq_softrt_next_start(bfqd, bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5686)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5687) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5688) * If this is the in-service queue, check if it needs to be expired,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5689) * or if we want to idle in case it has no pending requests.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5690) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5691) if (bfqd->in_service_queue == bfqq) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5692) if (bfq_bfqq_must_idle(bfqq)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5693) if (bfqq->dispatched == 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5694) bfq_arm_slice_timer(bfqd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5695) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5696) * If we get here, we do not expire bfqq, even
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5697) * if bfqq was in budget timeout or had no
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5698) * more requests (as controlled in the next
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5699) * conditional instructions). The reason for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5700) * not expiring bfqq is as follows.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5701) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5702) * Here bfqq->dispatched > 0 holds, but
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5703) * bfq_bfqq_must_idle() returned true. This
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5704) * implies that, even if no request arrives
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5705) * for bfqq before bfqq->dispatched reaches 0,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5706) * bfqq will, however, not be expired on the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5707) * completion event that causes bfqq->dispatch
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5708) * to reach zero. In contrast, on this event,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5709) * bfqq will start enjoying device idling
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5710) * (I/O-dispatch plugging).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5711) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5712) * But, if we expired bfqq here, bfqq would
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5713) * not have the chance to enjoy device idling
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5714) * when bfqq->dispatched finally reaches
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5715) * zero. This would expose bfqq to violation
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5716) * of its reserved service guarantees.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5717) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5718) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5719) } else if (bfq_may_expire_for_budg_timeout(bfqq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5720) bfq_bfqq_expire(bfqd, bfqq, false,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5721) BFQQE_BUDGET_TIMEOUT);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5722) else if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5723) (bfqq->dispatched == 0 ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5724) !bfq_better_to_idle(bfqq)))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5725) bfq_bfqq_expire(bfqd, bfqq, false,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5726) BFQQE_NO_MORE_REQUESTS);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5727) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5728)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5729) if (!bfqd->rq_in_driver)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5730) bfq_schedule_dispatch(bfqd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5731) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5732)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5733) static void bfq_finish_requeue_request_body(struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5734) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5735) bfqq->allocated--;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5736)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5737) bfq_put_queue(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5738) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5739)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5740) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5741) * The processes associated with bfqq may happen to generate their
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5742) * cumulative I/O at a lower rate than the rate at which the device
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5743) * could serve the same I/O. This is rather probable, e.g., if only
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5744) * one process is associated with bfqq and the device is an SSD. It
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5745) * results in bfqq becoming often empty while in service. In this
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5746) * respect, if BFQ is allowed to switch to another queue when bfqq
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5747) * remains empty, then the device goes on being fed with I/O requests,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5748) * and the throughput is not affected. In contrast, if BFQ is not
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5749) * allowed to switch to another queue---because bfqq is sync and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5750) * I/O-dispatch needs to be plugged while bfqq is temporarily
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5751) * empty---then, during the service of bfqq, there will be frequent
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5752) * "service holes", i.e., time intervals during which bfqq gets empty
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5753) * and the device can only consume the I/O already queued in its
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5754) * hardware queues. During service holes, the device may even get to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5755) * remaining idle. In the end, during the service of bfqq, the device
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5756) * is driven at a lower speed than the one it can reach with the kind
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5757) * of I/O flowing through bfqq.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5758) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5759) * To counter this loss of throughput, BFQ implements a "request
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5760) * injection mechanism", which tries to fill the above service holes
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5761) * with I/O requests taken from other queues. The hard part in this
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5762) * mechanism is finding the right amount of I/O to inject, so as to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5763) * both boost throughput and not break bfqq's bandwidth and latency
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5764) * guarantees. In this respect, the mechanism maintains a per-queue
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5765) * inject limit, computed as below. While bfqq is empty, the injection
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5766) * mechanism dispatches extra I/O requests only until the total number
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5767) * of I/O requests in flight---i.e., already dispatched but not yet
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5768) * completed---remains lower than this limit.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5769) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5770) * A first definition comes in handy to introduce the algorithm by
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5771) * which the inject limit is computed. We define as first request for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5772) * bfqq, an I/O request for bfqq that arrives while bfqq is in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5773) * service, and causes bfqq to switch from empty to non-empty. The
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5774) * algorithm updates the limit as a function of the effect of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5775) * injection on the service times of only the first requests of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5776) * bfqq. The reason for this restriction is that these are the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5777) * requests whose service time is affected most, because they are the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5778) * first to arrive after injection possibly occurred.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5779) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5780) * To evaluate the effect of injection, the algorithm measures the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5781) * "total service time" of first requests. We define as total service
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5782) * time of an I/O request, the time that elapses since when the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5783) * request is enqueued into bfqq, to when it is completed. This
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5784) * quantity allows the whole effect of injection to be measured. It is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5785) * easy to see why. Suppose that some requests of other queues are
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5786) * actually injected while bfqq is empty, and that a new request R
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5787) * then arrives for bfqq. If the device does start to serve all or
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5788) * part of the injected requests during the service hole, then,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5789) * because of this extra service, it may delay the next invocation of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5790) * the dispatch hook of BFQ. Then, even after R gets eventually
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5791) * dispatched, the device may delay the actual service of R if it is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5792) * still busy serving the extra requests, or if it decides to serve,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5793) * before R, some extra request still present in its queues. As a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5794) * conclusion, the cumulative extra delay caused by injection can be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5795) * easily evaluated by just comparing the total service time of first
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5796) * requests with and without injection.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5797) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5798) * The limit-update algorithm works as follows. On the arrival of a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5799) * first request of bfqq, the algorithm measures the total time of the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5800) * request only if one of the three cases below holds, and, for each
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5801) * case, it updates the limit as described below:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5802) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5803) * (1) If there is no in-flight request. This gives a baseline for the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5804) * total service time of the requests of bfqq. If the baseline has
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5805) * not been computed yet, then, after computing it, the limit is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5806) * set to 1, to start boosting throughput, and to prepare the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5807) * ground for the next case. If the baseline has already been
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5808) * computed, then it is updated, in case it results to be lower
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5809) * than the previous value.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5810) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5811) * (2) If the limit is higher than 0 and there are in-flight
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5812) * requests. By comparing the total service time in this case with
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5813) * the above baseline, it is possible to know at which extent the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5814) * current value of the limit is inflating the total service
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5815) * time. If the inflation is below a certain threshold, then bfqq
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5816) * is assumed to be suffering from no perceivable loss of its
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5817) * service guarantees, and the limit is even tentatively
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5818) * increased. If the inflation is above the threshold, then the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5819) * limit is decreased. Due to the lack of any hysteresis, this
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5820) * logic makes the limit oscillate even in steady workload
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5821) * conditions. Yet we opted for it, because it is fast in reaching
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5822) * the best value for the limit, as a function of the current I/O
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5823) * workload. To reduce oscillations, this step is disabled for a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5824) * short time interval after the limit happens to be decreased.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5825) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5826) * (3) Periodically, after resetting the limit, to make sure that the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5827) * limit eventually drops in case the workload changes. This is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5828) * needed because, after the limit has gone safely up for a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5829) * certain workload, it is impossible to guess whether the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5830) * baseline total service time may have changed, without measuring
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5831) * it again without injection. A more effective version of this
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5832) * step might be to just sample the baseline, by interrupting
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5833) * injection only once, and then to reset/lower the limit only if
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5834) * the total service time with the current limit does happen to be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5835) * too large.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5836) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5837) * More details on each step are provided in the comments on the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5838) * pieces of code that implement these steps: the branch handling the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5839) * transition from empty to non empty in bfq_add_request(), the branch
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5840) * handling injection in bfq_select_queue(), and the function
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5841) * bfq_choose_bfqq_for_injection(). These comments also explain some
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5842) * exceptions, made by the injection mechanism in some special cases.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5843) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5844) static void bfq_update_inject_limit(struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5845) struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5846) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5847) u64 tot_time_ns = ktime_get_ns() - bfqd->last_empty_occupied_ns;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5848) unsigned int old_limit = bfqq->inject_limit;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5849)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5850) if (bfqq->last_serv_time_ns > 0 && bfqd->rqs_injected) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5851) u64 threshold = (bfqq->last_serv_time_ns * 3)>>1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5852)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5853) if (tot_time_ns >= threshold && old_limit > 0) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5854) bfqq->inject_limit--;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5855) bfqq->decrease_time_jif = jiffies;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5856) } else if (tot_time_ns < threshold &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5857) old_limit <= bfqd->max_rq_in_driver)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5858) bfqq->inject_limit++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5859) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5860)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5861) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5862) * Either we still have to compute the base value for the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5863) * total service time, and there seem to be the right
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5864) * conditions to do it, or we can lower the last base value
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5865) * computed.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5866) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5867) * NOTE: (bfqd->rq_in_driver == 1) means that there is no I/O
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5868) * request in flight, because this function is in the code
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5869) * path that handles the completion of a request of bfqq, and,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5870) * in particular, this function is executed before
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5871) * bfqd->rq_in_driver is decremented in such a code path.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5872) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5873) if ((bfqq->last_serv_time_ns == 0 && bfqd->rq_in_driver == 1) ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5874) tot_time_ns < bfqq->last_serv_time_ns) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5875) if (bfqq->last_serv_time_ns == 0) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5876) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5877) * Now we certainly have a base value: make sure we
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5878) * start trying injection.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5879) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5880) bfqq->inject_limit = max_t(unsigned int, 1, old_limit);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5881) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5882) bfqq->last_serv_time_ns = tot_time_ns;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5883) } else if (!bfqd->rqs_injected && bfqd->rq_in_driver == 1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5884) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5885) * No I/O injected and no request still in service in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5886) * the drive: these are the exact conditions for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5887) * computing the base value of the total service time
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5888) * for bfqq. So let's update this value, because it is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5889) * rather variable. For example, it varies if the size
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5890) * or the spatial locality of the I/O requests in bfqq
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5891) * change.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5892) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5893) bfqq->last_serv_time_ns = tot_time_ns;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5894)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5895)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5896) /* update complete, not waiting for any request completion any longer */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5897) bfqd->waited_rq = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5898) bfqd->rqs_injected = false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5899) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5900)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5901) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5902) * Handle either a requeue or a finish for rq. The things to do are
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5903) * the same in both cases: all references to rq are to be dropped. In
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5904) * particular, rq is considered completed from the point of view of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5905) * the scheduler.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5906) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5907) static void bfq_finish_requeue_request(struct request *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5908) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5909) struct bfq_queue *bfqq = RQ_BFQQ(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5910) struct bfq_data *bfqd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5911)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5912) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5913) * rq either is not associated with any icq, or is an already
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5914) * requeued request that has not (yet) been re-inserted into
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5915) * a bfq_queue.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5916) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5917) if (!rq->elv.icq || !bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5918) return;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5919)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5920) bfqd = bfqq->bfqd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5921)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5922) if (rq->rq_flags & RQF_STARTED)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5923) bfqg_stats_update_completion(bfqq_group(bfqq),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5924) rq->start_time_ns,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5925) rq->io_start_time_ns,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5926) rq->cmd_flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5927)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5928) if (likely(rq->rq_flags & RQF_STARTED)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5929) unsigned long flags;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5930)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5931) spin_lock_irqsave(&bfqd->lock, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5932)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5933) if (rq == bfqd->waited_rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5934) bfq_update_inject_limit(bfqd, bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5935)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5936) bfq_completed_request(bfqq, bfqd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5937) bfq_finish_requeue_request_body(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5938)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5939) spin_unlock_irqrestore(&bfqd->lock, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5940) } else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5941) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5942) * Request rq may be still/already in the scheduler,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5943) * in which case we need to remove it (this should
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5944) * never happen in case of requeue). And we cannot
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5945) * defer such a check and removal, to avoid
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5946) * inconsistencies in the time interval from the end
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5947) * of this function to the start of the deferred work.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5948) * This situation seems to occur only in process
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5949) * context, as a consequence of a merge. In the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5950) * current version of the code, this implies that the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5951) * lock is held.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5952) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5953)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5954) if (!RB_EMPTY_NODE(&rq->rb_node)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5955) bfq_remove_request(rq->q, rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5956) bfqg_stats_update_io_remove(bfqq_group(bfqq),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5957) rq->cmd_flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5958) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5959) bfq_finish_requeue_request_body(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5960) }
^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) * Reset private fields. In case of a requeue, this allows
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5964) * this function to correctly do nothing if it is spuriously
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5965) * invoked again on this same request (see the check at the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5966) * beginning of the function). Probably, a better general
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5967) * design would be to prevent blk-mq from invoking the requeue
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5968) * or finish hooks of an elevator, for a request that is not
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5969) * referred by that elevator.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5970) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5971) * Resetting the following fields would break the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5972) * request-insertion logic if rq is re-inserted into a bfq
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5973) * internal queue, without a re-preparation. Here we assume
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5974) * that re-insertions of requeued requests, without
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5975) * re-preparation, can happen only for pass_through or at_head
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5976) * requests (which are not re-inserted into bfq internal
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5977) * queues).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5978) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5979) rq->elv.priv[0] = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5980) rq->elv.priv[1] = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5981) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5982)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5983) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5984) * Removes the association between the current task and bfqq, assuming
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5985) * that bic points to the bfq iocontext of the task.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5986) * Returns NULL if a new bfqq should be allocated, or the old bfqq if this
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5987) * was the last process referring to that bfqq.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5988) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5989) static struct bfq_queue *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5990) bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5991) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5992) bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue");
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5993)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5994) if (bfqq_process_refs(bfqq) == 1) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5995) bfqq->pid = current->pid;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5996) bfq_clear_bfqq_coop(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5997) bfq_clear_bfqq_split_coop(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5998) return bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 5999) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6000)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6001) bic_set_bfqq(bic, NULL, 1);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6002)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6003) bfq_put_cooperator(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6004)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6005) bfq_release_process_ref(bfqq->bfqd, bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6006) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6007) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6008)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6009) static struct bfq_queue *bfq_get_bfqq_handle_split(struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6010) struct bfq_io_cq *bic,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6011) struct bio *bio,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6012) bool split, bool is_sync,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6013) bool *new_queue)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6014) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6015) struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6016)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6017) if (likely(bfqq && bfqq != &bfqd->oom_bfqq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6018) return bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6019)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6020) if (new_queue)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6021) *new_queue = true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6022)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6023) if (bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6024) bfq_put_queue(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6025) bfqq = bfq_get_queue(bfqd, bio, is_sync, bic);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6026)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6027) bic_set_bfqq(bic, bfqq, is_sync);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6028) if (split && is_sync) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6029) if ((bic->was_in_burst_list && bfqd->large_burst) ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6030) bic->saved_in_large_burst)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6031) bfq_mark_bfqq_in_large_burst(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6032) else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6033) bfq_clear_bfqq_in_large_burst(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6034) if (bic->was_in_burst_list)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6035) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6036) * If bfqq was in the current
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6037) * burst list before being
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6038) * merged, then we have to add
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6039) * it back. And we do not need
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6040) * to increase burst_size, as
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6041) * we did not decrement
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6042) * burst_size when we removed
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6043) * bfqq from the burst list as
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6044) * a consequence of a merge
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6045) * (see comments in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6046) * bfq_put_queue). In this
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6047) * respect, it would be rather
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6048) * costly to know whether the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6049) * current burst list is still
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6050) * the same burst list from
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6051) * which bfqq was removed on
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6052) * the merge. To avoid this
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6053) * cost, if bfqq was in a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6054) * burst list, then we add
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6055) * bfqq to the current burst
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6056) * list without any further
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6057) * check. This can cause
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6058) * inappropriate insertions,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6059) * but rarely enough to not
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6060) * harm the detection of large
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6061) * bursts significantly.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6062) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6063) hlist_add_head(&bfqq->burst_list_node,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6064) &bfqd->burst_list);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6065) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6066) bfqq->split_time = jiffies;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6067) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6068)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6069) return bfqq;
^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) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6073) * Only reset private fields. The actual request preparation will be
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6074) * performed by bfq_init_rq, when rq is either inserted or merged. See
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6075) * comments on bfq_init_rq for the reason behind this delayed
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6076) * preparation.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6077) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6078) static void bfq_prepare_request(struct request *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6079) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6080) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6081) * Regardless of whether we have an icq attached, we have to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6082) * clear the scheduler pointers, as they might point to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6083) * previously allocated bic/bfqq structs.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6084) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6085) rq->elv.priv[0] = rq->elv.priv[1] = NULL;
^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) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6089) * If needed, init rq, allocate bfq data structures associated with
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6090) * rq, and increment reference counters in the destination bfq_queue
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6091) * for rq. Return the destination bfq_queue for rq, or NULL is rq is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6092) * not associated with any bfq_queue.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6093) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6094) * This function is invoked by the functions that perform rq insertion
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6095) * or merging. One may have expected the above preparation operations
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6096) * to be performed in bfq_prepare_request, and not delayed to when rq
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6097) * is inserted or merged. The rationale behind this delayed
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6098) * preparation is that, after the prepare_request hook is invoked for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6099) * rq, rq may still be transformed into a request with no icq, i.e., a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6100) * request not associated with any queue. No bfq hook is invoked to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6101) * signal this transformation. As a consequence, should these
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6102) * preparation operations be performed when the prepare_request hook
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6103) * is invoked, and should rq be transformed one moment later, bfq
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6104) * would end up in an inconsistent state, because it would have
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6105) * incremented some queue counters for an rq destined to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6106) * transformation, without any chance to correctly lower these
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6107) * counters back. In contrast, no transformation can still happen for
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6108) * rq after rq has been inserted or merged. So, it is safe to execute
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6109) * these preparation operations when rq is finally inserted or merged.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6110) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6111) static struct bfq_queue *bfq_init_rq(struct request *rq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6112) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6113) struct request_queue *q = rq->q;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6114) struct bio *bio = rq->bio;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6115) struct bfq_data *bfqd = q->elevator->elevator_data;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6116) struct bfq_io_cq *bic;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6117) const int is_sync = rq_is_sync(rq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6118) struct bfq_queue *bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6119) bool new_queue = false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6120) bool bfqq_already_existing = false, split = false;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6121)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6122) if (unlikely(!rq->elv.icq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6123) return NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6124)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6125) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6126) * Assuming that elv.priv[1] is set only if everything is set
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6127) * for this rq. This holds true, because this function is
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6128) * invoked only for insertion or merging, and, after such
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6129) * events, a request cannot be manipulated any longer before
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6130) * being removed from bfq.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6131) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6132) if (rq->elv.priv[1])
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6133) return rq->elv.priv[1];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6134)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6135) bic = icq_to_bic(rq->elv.icq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6136)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6137) bfq_check_ioprio_change(bic, bio);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6138)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6139) bfq_bic_update_cgroup(bic, bio);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6140)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6141) bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio, false, is_sync,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6142) &new_queue);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6143)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6144) if (likely(!new_queue)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6145) /* If the queue was seeky for too long, break it apart. */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6146) if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq)) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6147) bfq_log_bfqq(bfqd, bfqq, "breaking apart bfqq");
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6148)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6149) /* Update bic before losing reference to bfqq */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6150) if (bfq_bfqq_in_large_burst(bfqq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6151) bic->saved_in_large_burst = true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6152)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6153) bfqq = bfq_split_bfqq(bic, bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6154) split = true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6155)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6156) if (!bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6157) bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6158) true, is_sync,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6159) NULL);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6160) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6161) bfqq_already_existing = true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6162) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6163) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6164)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6165) bfqq->allocated++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6166) bfqq->ref++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6167) bfq_log_bfqq(bfqd, bfqq, "get_request %p: bfqq %p, %d",
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6168) rq, bfqq, bfqq->ref);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6169)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6170) rq->elv.priv[0] = bic;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6171) rq->elv.priv[1] = bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6172)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6173) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6174) * If a bfq_queue has only one process reference, it is owned
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6175) * by only this bic: we can then set bfqq->bic = bic. in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6176) * addition, if the queue has also just been split, we have to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6177) * resume its state.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6178) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6179) if (likely(bfqq != &bfqd->oom_bfqq) && bfqq_process_refs(bfqq) == 1) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6180) bfqq->bic = bic;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6181) if (split) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6182) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6183) * The queue has just been split from a shared
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6184) * queue: restore the idle window and the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6185) * possible weight raising period.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6186) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6187) bfq_bfqq_resume_state(bfqq, bfqd, bic,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6188) bfqq_already_existing);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6189) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6190) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6191)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6192) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6193) * Consider bfqq as possibly belonging to a burst of newly
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6194) * created queues only if:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6195) * 1) A burst is actually happening (bfqd->burst_size > 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6196) * or
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6197) * 2) There is no other active queue. In fact, if, in
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6198) * contrast, there are active queues not belonging to the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6199) * possible burst bfqq may belong to, then there is no gain
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6200) * in considering bfqq as belonging to a burst, and
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6201) * therefore in not weight-raising bfqq. See comments on
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6202) * bfq_handle_burst().
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6203) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6204) * This filtering also helps eliminating false positives,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6205) * occurring when bfqq does not belong to an actual large
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6206) * burst, but some background task (e.g., a service) happens
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6207) * to trigger the creation of new queues very close to when
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6208) * bfqq and its possible companion queues are created. See
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6209) * comments on bfq_handle_burst() for further details also on
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6210) * this issue.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6211) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6212) if (unlikely(bfq_bfqq_just_created(bfqq) &&
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6213) (bfqd->burst_size > 0 ||
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6214) bfq_tot_busy_queues(bfqd) == 0)))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6215) bfq_handle_burst(bfqd, bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6216)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6217) return bfqq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6218) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6219)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6220) static void
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6221) bfq_idle_slice_timer_body(struct bfq_data *bfqd, struct bfq_queue *bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6222) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6223) enum bfqq_expiration reason;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6224) unsigned long flags;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6225)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6226) spin_lock_irqsave(&bfqd->lock, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6227)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6228) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6229) * Considering that bfqq may be in race, we should firstly check
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6230) * whether bfqq is in service before doing something on it. If
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6231) * the bfqq in race is not in service, it has already been expired
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6232) * through __bfq_bfqq_expire func and its wait_request flags has
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6233) * been cleared in __bfq_bfqd_reset_in_service func.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6234) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6235) if (bfqq != bfqd->in_service_queue) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6236) spin_unlock_irqrestore(&bfqd->lock, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6237) return;
^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) bfq_clear_bfqq_wait_request(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6241)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6242) if (bfq_bfqq_budget_timeout(bfqq))
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6243) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6244) * Also here the queue can be safely expired
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6245) * for budget timeout without wasting
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6246) * guarantees
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6247) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6248) reason = BFQQE_BUDGET_TIMEOUT;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6249) else if (bfqq->queued[0] == 0 && bfqq->queued[1] == 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6250) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6251) * The queue may not be empty upon timer expiration,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6252) * because we may not disable the timer when the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6253) * first request of the in-service queue arrives
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6254) * during disk idling.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6255) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6256) reason = BFQQE_TOO_IDLE;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6257) else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6258) goto schedule_dispatch;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6259)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6260) bfq_bfqq_expire(bfqd, bfqq, true, reason);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6261)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6262) schedule_dispatch:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6263) spin_unlock_irqrestore(&bfqd->lock, flags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6264) bfq_schedule_dispatch(bfqd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6265) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6266)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6267) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6268) * Handler of the expiration of the timer running if the in-service queue
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6269) * is idling inside its time slice.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6270) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6271) static enum hrtimer_restart bfq_idle_slice_timer(struct hrtimer *timer)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6272) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6273) struct bfq_data *bfqd = container_of(timer, struct bfq_data,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6274) idle_slice_timer);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6275) struct bfq_queue *bfqq = bfqd->in_service_queue;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6276)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6277) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6278) * Theoretical race here: the in-service queue can be NULL or
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6279) * different from the queue that was idling if a new request
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6280) * arrives for the current queue and there is a full dispatch
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6281) * cycle that changes the in-service queue. This can hardly
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6282) * happen, but in the worst case we just expire a queue too
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6283) * early.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6284) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6285) if (bfqq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6286) bfq_idle_slice_timer_body(bfqd, bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6287)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6288) return HRTIMER_NORESTART;
^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 void __bfq_put_async_bfqq(struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6292) struct bfq_queue **bfqq_ptr)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6293) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6294) struct bfq_queue *bfqq = *bfqq_ptr;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6295)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6296) bfq_log(bfqd, "put_async_bfqq: %p", bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6297) if (bfqq) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6298) bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6299)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6300) bfq_log_bfqq(bfqd, bfqq, "put_async_bfqq: putting %p, %d",
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6301) bfqq, bfqq->ref);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6302) bfq_put_queue(bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6303) *bfqq_ptr = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6304) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6305) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6306)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6307) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6308) * Release all the bfqg references to its async queues. If we are
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6309) * deallocating the group these queues may still contain requests, so
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6310) * we reparent them to the root cgroup (i.e., the only one that will
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6311) * exist for sure until all the requests on a device are gone).
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6312) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6313) void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6314) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6315) int i, j;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6316)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6317) for (i = 0; i < 2; i++)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6318) for (j = 0; j < IOPRIO_BE_NR; j++)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6319) __bfq_put_async_bfqq(bfqd, &bfqg->async_bfqq[i][j]);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6320)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6321) __bfq_put_async_bfqq(bfqd, &bfqg->async_idle_bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6322) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6323)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6324) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6325) * See the comments on bfq_limit_depth for the purpose of
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6326) * the depths set in the function. Return minimum shallow depth we'll use.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6327) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6328) static unsigned int bfq_update_depths(struct bfq_data *bfqd,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6329) struct sbitmap_queue *bt)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6330) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6331) unsigned int i, j, min_shallow = UINT_MAX;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6332)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6333) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6334) * In-word depths if no bfq_queue is being weight-raised:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6335) * leaving 25% of tags only for sync reads.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6336) *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6337) * In next formulas, right-shift the value
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6338) * (1U<<bt->sb.shift), instead of computing directly
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6339) * (1U<<(bt->sb.shift - something)), to be robust against
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6340) * any possible value of bt->sb.shift, without having to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6341) * limit 'something'.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6342) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6343) /* no more than 50% of tags for async I/O */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6344) bfqd->word_depths[0][0] = max((1U << bt->sb.shift) >> 1, 1U);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6345) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6346) * no more than 75% of tags for sync writes (25% extra tags
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6347) * w.r.t. async I/O, to prevent async I/O from starving sync
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6348) * writes)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6349) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6350) bfqd->word_depths[0][1] = max(((1U << bt->sb.shift) * 3) >> 2, 1U);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6351)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6352) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6353) * In-word depths in case some bfq_queue is being weight-
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6354) * raised: leaving ~63% of tags for sync reads. This is the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6355) * highest percentage for which, in our tests, application
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6356) * start-up times didn't suffer from any regression due to tag
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6357) * shortage.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6358) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6359) /* no more than ~18% of tags for async I/O */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6360) bfqd->word_depths[1][0] = max(((1U << bt->sb.shift) * 3) >> 4, 1U);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6361) /* no more than ~37% of tags for sync writes (~20% extra tags) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6362) bfqd->word_depths[1][1] = max(((1U << bt->sb.shift) * 6) >> 4, 1U);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6363)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6364) for (i = 0; i < 2; i++)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6365) for (j = 0; j < 2; j++)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6366) min_shallow = min(min_shallow, bfqd->word_depths[i][j]);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6367)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6368) return min_shallow;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6369) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6370)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6371) static void bfq_depth_updated(struct blk_mq_hw_ctx *hctx)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6372) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6373) struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6374) struct blk_mq_tags *tags = hctx->sched_tags;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6375) unsigned int min_shallow;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6376)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6377) min_shallow = bfq_update_depths(bfqd, tags->bitmap_tags);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6378) sbitmap_queue_min_shallow_depth(tags->bitmap_tags, min_shallow);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6379) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6380)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6381) static int bfq_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int index)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6382) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6383) bfq_depth_updated(hctx);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6384) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6385) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6386)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6387) static void bfq_exit_queue(struct elevator_queue *e)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6388) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6389) struct bfq_data *bfqd = e->elevator_data;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6390) struct bfq_queue *bfqq, *n;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6391)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6392) hrtimer_cancel(&bfqd->idle_slice_timer);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6393)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6394) spin_lock_irq(&bfqd->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6395) list_for_each_entry_safe(bfqq, n, &bfqd->idle_list, bfqq_list)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6396) bfq_deactivate_bfqq(bfqd, bfqq, false, false);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6397) spin_unlock_irq(&bfqd->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6398)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6399) hrtimer_cancel(&bfqd->idle_slice_timer);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6400)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6401) /* release oom-queue reference to root group */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6402) bfqg_and_blkg_put(bfqd->root_group);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6403)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6404) #ifdef CONFIG_BFQ_GROUP_IOSCHED
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6405) blkcg_deactivate_policy(bfqd->queue, &blkcg_policy_bfq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6406) #else
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6407) spin_lock_irq(&bfqd->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6408) bfq_put_async_queues(bfqd, bfqd->root_group);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6409) kfree(bfqd->root_group);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6410) spin_unlock_irq(&bfqd->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6411) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6412)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6413) wbt_enable_default(bfqd->queue);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6414)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6415) kfree(bfqd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6416) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6417)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6418) static void bfq_init_root_group(struct bfq_group *root_group,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6419) struct bfq_data *bfqd)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6420) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6421) int i;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6422)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6423) #ifdef CONFIG_BFQ_GROUP_IOSCHED
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6424) root_group->entity.parent = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6425) root_group->my_entity = NULL;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6426) root_group->bfqd = bfqd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6427) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6428) root_group->rq_pos_tree = RB_ROOT;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6429) for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6430) root_group->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6431) root_group->sched_data.bfq_class_idle_last_service = jiffies;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6432) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6433)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6434) static int bfq_init_queue(struct request_queue *q, struct elevator_type *e)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6435) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6436) struct bfq_data *bfqd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6437) struct elevator_queue *eq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6438)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6439) eq = elevator_alloc(q, e);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6440) if (!eq)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6441) return -ENOMEM;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6442)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6443) bfqd = kzalloc_node(sizeof(*bfqd), GFP_KERNEL, q->node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6444) if (!bfqd) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6445) kobject_put(&eq->kobj);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6446) return -ENOMEM;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6447) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6448) eq->elevator_data = bfqd;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6449)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6450) spin_lock_irq(&q->queue_lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6451) q->elevator = eq;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6452) spin_unlock_irq(&q->queue_lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6453)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6454) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6455) * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6456) * Grab a permanent reference to it, so that the normal code flow
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6457) * will not attempt to free it.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6458) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6459) bfq_init_bfqq(bfqd, &bfqd->oom_bfqq, NULL, 1, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6460) bfqd->oom_bfqq.ref++;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6461) bfqd->oom_bfqq.new_ioprio = BFQ_DEFAULT_QUEUE_IOPRIO;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6462) bfqd->oom_bfqq.new_ioprio_class = IOPRIO_CLASS_BE;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6463) bfqd->oom_bfqq.entity.new_weight =
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6464) bfq_ioprio_to_weight(bfqd->oom_bfqq.new_ioprio);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6465)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6466) /* oom_bfqq does not participate to bursts */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6467) bfq_clear_bfqq_just_created(&bfqd->oom_bfqq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6468)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6469) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6470) * Trigger weight initialization, according to ioprio, at the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6471) * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6472) * class won't be changed any more.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6473) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6474) bfqd->oom_bfqq.entity.prio_changed = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6475)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6476) bfqd->queue = q;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6477)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6478) INIT_LIST_HEAD(&bfqd->dispatch);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6479)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6480) hrtimer_init(&bfqd->idle_slice_timer, CLOCK_MONOTONIC,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6481) HRTIMER_MODE_REL);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6482) bfqd->idle_slice_timer.function = bfq_idle_slice_timer;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6483)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6484) bfqd->queue_weights_tree = RB_ROOT_CACHED;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6485) bfqd->num_groups_with_pending_reqs = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6486)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6487) INIT_LIST_HEAD(&bfqd->active_list);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6488) INIT_LIST_HEAD(&bfqd->idle_list);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6489) INIT_HLIST_HEAD(&bfqd->burst_list);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6490)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6491) bfqd->hw_tag = -1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6492) bfqd->nonrot_with_queueing = blk_queue_nonrot(bfqd->queue);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6493)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6494) bfqd->bfq_max_budget = bfq_default_max_budget;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6495)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6496) bfqd->bfq_fifo_expire[0] = bfq_fifo_expire[0];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6497) bfqd->bfq_fifo_expire[1] = bfq_fifo_expire[1];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6498) bfqd->bfq_back_max = bfq_back_max;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6499) bfqd->bfq_back_penalty = bfq_back_penalty;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6500) bfqd->bfq_slice_idle = bfq_slice_idle;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6501) bfqd->bfq_timeout = bfq_timeout;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6502)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6503) bfqd->bfq_requests_within_timer = 120;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6504)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6505) bfqd->bfq_large_burst_thresh = 8;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6506) bfqd->bfq_burst_interval = msecs_to_jiffies(180);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6507)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6508) bfqd->low_latency = true;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6509)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6510) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6511) * Trade-off between responsiveness and fairness.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6512) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6513) bfqd->bfq_wr_coeff = 30;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6514) bfqd->bfq_wr_rt_max_time = msecs_to_jiffies(300);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6515) bfqd->bfq_wr_max_time = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6516) bfqd->bfq_wr_min_idle_time = msecs_to_jiffies(2000);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6517) bfqd->bfq_wr_min_inter_arr_async = msecs_to_jiffies(500);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6518) bfqd->bfq_wr_max_softrt_rate = 7000; /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6519) * Approximate rate required
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6520) * to playback or record a
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6521) * high-definition compressed
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6522) * video.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6523) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6524) bfqd->wr_busy_queues = 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6525)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6526) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6527) * Begin by assuming, optimistically, that the device peak
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6528) * rate is equal to 2/3 of the highest reference rate.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6529) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6530) bfqd->rate_dur_prod = ref_rate[blk_queue_nonrot(bfqd->queue)] *
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6531) ref_wr_duration[blk_queue_nonrot(bfqd->queue)];
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6532) bfqd->peak_rate = ref_rate[blk_queue_nonrot(bfqd->queue)] * 2 / 3;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6533)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6534) spin_lock_init(&bfqd->lock);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6535)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6536) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6537) * The invocation of the next bfq_create_group_hierarchy
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6538) * function is the head of a chain of function calls
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6539) * (bfq_create_group_hierarchy->blkcg_activate_policy->
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6540) * blk_mq_freeze_queue) that may lead to the invocation of the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6541) * has_work hook function. For this reason,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6542) * bfq_create_group_hierarchy is invoked only after all
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6543) * scheduler data has been initialized, apart from the fields
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6544) * that can be initialized only after invoking
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6545) * bfq_create_group_hierarchy. This, in particular, enables
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6546) * has_work to correctly return false. Of course, to avoid
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6547) * other inconsistencies, the blk-mq stack must then refrain
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6548) * from invoking further scheduler hooks before this init
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6549) * function is finished.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6550) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6551) bfqd->root_group = bfq_create_group_hierarchy(bfqd, q->node);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6552) if (!bfqd->root_group)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6553) goto out_free;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6554) bfq_init_root_group(bfqd->root_group, bfqd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6555) bfq_init_entity(&bfqd->oom_bfqq.entity, bfqd->root_group);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6556)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6557) wbt_disable_default(q);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6558) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6559)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6560) out_free:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6561) kfree(bfqd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6562) kobject_put(&eq->kobj);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6563) return -ENOMEM;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6564) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6565)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6566) static void bfq_slab_kill(void)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6567) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6568) kmem_cache_destroy(bfq_pool);
^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) static int __init bfq_slab_setup(void)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6572) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6573) bfq_pool = KMEM_CACHE(bfq_queue, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6574) if (!bfq_pool)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6575) return -ENOMEM;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6576) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6577) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6578)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6579) static ssize_t bfq_var_show(unsigned int var, char *page)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6580) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6581) return sprintf(page, "%u\n", var);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6582) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6583)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6584) static int bfq_var_store(unsigned long *var, const char *page)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6585) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6586) unsigned long new_val;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6587) int ret = kstrtoul(page, 10, &new_val);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6588)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6589) if (ret)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6590) return ret;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6591) *var = new_val;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6592) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6593) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6594)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6595) #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6596) static ssize_t __FUNC(struct elevator_queue *e, char *page) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6597) { \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6598) struct bfq_data *bfqd = e->elevator_data; \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6599) u64 __data = __VAR; \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6600) if (__CONV == 1) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6601) __data = jiffies_to_msecs(__data); \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6602) else if (__CONV == 2) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6603) __data = div_u64(__data, NSEC_PER_MSEC); \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6604) return bfq_var_show(__data, (page)); \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6605) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6606) SHOW_FUNCTION(bfq_fifo_expire_sync_show, bfqd->bfq_fifo_expire[1], 2);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6607) SHOW_FUNCTION(bfq_fifo_expire_async_show, bfqd->bfq_fifo_expire[0], 2);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6608) SHOW_FUNCTION(bfq_back_seek_max_show, bfqd->bfq_back_max, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6609) SHOW_FUNCTION(bfq_back_seek_penalty_show, bfqd->bfq_back_penalty, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6610) SHOW_FUNCTION(bfq_slice_idle_show, bfqd->bfq_slice_idle, 2);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6611) SHOW_FUNCTION(bfq_max_budget_show, bfqd->bfq_user_max_budget, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6612) SHOW_FUNCTION(bfq_timeout_sync_show, bfqd->bfq_timeout, 1);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6613) SHOW_FUNCTION(bfq_strict_guarantees_show, bfqd->strict_guarantees, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6614) SHOW_FUNCTION(bfq_low_latency_show, bfqd->low_latency, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6615) #undef SHOW_FUNCTION
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6616)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6617) #define USEC_SHOW_FUNCTION(__FUNC, __VAR) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6618) static ssize_t __FUNC(struct elevator_queue *e, char *page) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6619) { \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6620) struct bfq_data *bfqd = e->elevator_data; \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6621) u64 __data = __VAR; \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6622) __data = div_u64(__data, NSEC_PER_USEC); \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6623) return bfq_var_show(__data, (page)); \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6624) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6625) USEC_SHOW_FUNCTION(bfq_slice_idle_us_show, bfqd->bfq_slice_idle);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6626) #undef USEC_SHOW_FUNCTION
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6627)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6628) #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6629) static ssize_t \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6630) __FUNC(struct elevator_queue *e, const char *page, size_t count) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6631) { \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6632) struct bfq_data *bfqd = e->elevator_data; \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6633) unsigned long __data, __min = (MIN), __max = (MAX); \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6634) int ret; \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6635) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6636) ret = bfq_var_store(&__data, (page)); \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6637) if (ret) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6638) return ret; \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6639) if (__data < __min) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6640) __data = __min; \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6641) else if (__data > __max) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6642) __data = __max; \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6643) if (__CONV == 1) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6644) *(__PTR) = msecs_to_jiffies(__data); \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6645) else if (__CONV == 2) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6646) *(__PTR) = (u64)__data * NSEC_PER_MSEC; \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6647) else \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6648) *(__PTR) = __data; \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6649) return count; \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6650) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6651) STORE_FUNCTION(bfq_fifo_expire_sync_store, &bfqd->bfq_fifo_expire[1], 1,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6652) INT_MAX, 2);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6653) STORE_FUNCTION(bfq_fifo_expire_async_store, &bfqd->bfq_fifo_expire[0], 1,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6654) INT_MAX, 2);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6655) STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6656) STORE_FUNCTION(bfq_back_seek_penalty_store, &bfqd->bfq_back_penalty, 1,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6657) INT_MAX, 0);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6658) STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 2);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6659) #undef STORE_FUNCTION
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6660)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6661) #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6662) static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6663) { \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6664) struct bfq_data *bfqd = e->elevator_data; \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6665) unsigned long __data, __min = (MIN), __max = (MAX); \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6666) int ret; \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6667) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6668) ret = bfq_var_store(&__data, (page)); \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6669) if (ret) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6670) return ret; \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6671) if (__data < __min) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6672) __data = __min; \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6673) else if (__data > __max) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6674) __data = __max; \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6675) *(__PTR) = (u64)__data * NSEC_PER_USEC; \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6676) return count; \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6677) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6678) USEC_STORE_FUNCTION(bfq_slice_idle_us_store, &bfqd->bfq_slice_idle, 0,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6679) UINT_MAX);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6680) #undef USEC_STORE_FUNCTION
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6681)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6682) static ssize_t bfq_max_budget_store(struct elevator_queue *e,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6683) const char *page, size_t count)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6684) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6685) struct bfq_data *bfqd = e->elevator_data;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6686) unsigned long __data;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6687) int ret;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6688)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6689) ret = bfq_var_store(&__data, (page));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6690) if (ret)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6691) return ret;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6692)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6693) if (__data == 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6694) bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6695) else {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6696) if (__data > INT_MAX)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6697) __data = INT_MAX;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6698) bfqd->bfq_max_budget = __data;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6699) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6700)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6701) bfqd->bfq_user_max_budget = __data;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6702)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6703) return count;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6704) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6705)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6706) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6707) * Leaving this name to preserve name compatibility with cfq
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6708) * parameters, but this timeout is used for both sync and async.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6709) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6710) static ssize_t bfq_timeout_sync_store(struct elevator_queue *e,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6711) const char *page, size_t count)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6712) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6713) struct bfq_data *bfqd = e->elevator_data;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6714) unsigned long __data;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6715) int ret;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6716)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6717) ret = bfq_var_store(&__data, (page));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6718) if (ret)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6719) return ret;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6720)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6721) if (__data < 1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6722) __data = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6723) else if (__data > INT_MAX)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6724) __data = INT_MAX;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6725)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6726) bfqd->bfq_timeout = msecs_to_jiffies(__data);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6727) if (bfqd->bfq_user_max_budget == 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6728) bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6729)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6730) return count;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6731) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6732)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6733) static ssize_t bfq_strict_guarantees_store(struct elevator_queue *e,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6734) const char *page, size_t count)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6735) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6736) struct bfq_data *bfqd = e->elevator_data;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6737) unsigned long __data;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6738) int ret;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6739)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6740) ret = bfq_var_store(&__data, (page));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6741) if (ret)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6742) return ret;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6743)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6744) if (__data > 1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6745) __data = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6746) if (!bfqd->strict_guarantees && __data == 1
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6747) && bfqd->bfq_slice_idle < 8 * NSEC_PER_MSEC)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6748) bfqd->bfq_slice_idle = 8 * NSEC_PER_MSEC;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6749)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6750) bfqd->strict_guarantees = __data;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6751)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6752) return count;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6753) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6754)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6755) static ssize_t bfq_low_latency_store(struct elevator_queue *e,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6756) const char *page, size_t count)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6757) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6758) struct bfq_data *bfqd = e->elevator_data;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6759) unsigned long __data;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6760) int ret;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6761)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6762) ret = bfq_var_store(&__data, (page));
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6763) if (ret)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6764) return ret;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6765)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6766) if (__data > 1)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6767) __data = 1;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6768) if (__data == 0 && bfqd->low_latency != 0)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6769) bfq_end_wr(bfqd);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6770) bfqd->low_latency = __data;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6771)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6772) return count;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6773) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6774)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6775) #define BFQ_ATTR(name) \
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6776) __ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6777)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6778) static struct elv_fs_entry bfq_attrs[] = {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6779) BFQ_ATTR(fifo_expire_sync),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6780) BFQ_ATTR(fifo_expire_async),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6781) BFQ_ATTR(back_seek_max),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6782) BFQ_ATTR(back_seek_penalty),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6783) BFQ_ATTR(slice_idle),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6784) BFQ_ATTR(slice_idle_us),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6785) BFQ_ATTR(max_budget),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6786) BFQ_ATTR(timeout_sync),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6787) BFQ_ATTR(strict_guarantees),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6788) BFQ_ATTR(low_latency),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6789) __ATTR_NULL
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6790) };
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6791)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6792) static struct elevator_type iosched_bfq_mq = {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6793) .ops = {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6794) .limit_depth = bfq_limit_depth,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6795) .prepare_request = bfq_prepare_request,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6796) .requeue_request = bfq_finish_requeue_request,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6797) .finish_request = bfq_finish_requeue_request,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6798) .exit_icq = bfq_exit_icq,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6799) .insert_requests = bfq_insert_requests,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6800) .dispatch_request = bfq_dispatch_request,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6801) .next_request = elv_rb_latter_request,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6802) .former_request = elv_rb_former_request,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6803) .allow_merge = bfq_allow_bio_merge,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6804) .bio_merge = bfq_bio_merge,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6805) .request_merge = bfq_request_merge,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6806) .requests_merged = bfq_requests_merged,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6807) .request_merged = bfq_request_merged,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6808) .has_work = bfq_has_work,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6809) .depth_updated = bfq_depth_updated,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6810) .init_hctx = bfq_init_hctx,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6811) .init_sched = bfq_init_queue,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6812) .exit_sched = bfq_exit_queue,
^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) .icq_size = sizeof(struct bfq_io_cq),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6816) .icq_align = __alignof__(struct bfq_io_cq),
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6817) .elevator_attrs = bfq_attrs,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6818) .elevator_name = "bfq",
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6819) .elevator_owner = THIS_MODULE,
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6820) };
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6821) MODULE_ALIAS("bfq-iosched");
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6822)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6823) static int __init bfq_init(void)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6824) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6825) int ret;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6826)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6827) #ifdef CONFIG_BFQ_GROUP_IOSCHED
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6828) ret = blkcg_policy_register(&blkcg_policy_bfq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6829) if (ret)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6830) return ret;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6831) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6832)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6833) ret = -ENOMEM;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6834) if (bfq_slab_setup())
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6835) goto err_pol_unreg;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6836)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6837) /*
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6838) * Times to load large popular applications for the typical
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6839) * systems installed on the reference devices (see the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6840) * comments before the definition of the next
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6841) * array). Actually, we use slightly lower values, as the
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6842) * estimated peak rate tends to be smaller than the actual
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6843) * peak rate. The reason for this last fact is that estimates
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6844) * are computed over much shorter time intervals than the long
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6845) * intervals typically used for benchmarking. Why? First, to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6846) * adapt more quickly to variations. Second, because an I/O
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6847) * scheduler cannot rely on a peak-rate-evaluation workload to
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6848) * be run for a long time.
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6849) */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6850) ref_wr_duration[0] = msecs_to_jiffies(7000); /* actually 8 sec */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6851) ref_wr_duration[1] = msecs_to_jiffies(2500); /* actually 3 sec */
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6852)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6853) ret = elv_register(&iosched_bfq_mq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6854) if (ret)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6855) goto slab_kill;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6856)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6857) return 0;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6858)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6859) slab_kill:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6860) bfq_slab_kill();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6861) err_pol_unreg:
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6862) #ifdef CONFIG_BFQ_GROUP_IOSCHED
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6863) blkcg_policy_unregister(&blkcg_policy_bfq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6864) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6865) return ret;
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6866) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6867)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6868) static void __exit bfq_exit(void)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6869) {
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6870) elv_unregister(&iosched_bfq_mq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6871) #ifdef CONFIG_BFQ_GROUP_IOSCHED
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6872) blkcg_policy_unregister(&blkcg_policy_bfq);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6873) #endif
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6874) bfq_slab_kill();
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6875) }
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6876)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6877) module_init(bfq_init);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6878) module_exit(bfq_exit);
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6879)
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6880) MODULE_AUTHOR("Paolo Valente");
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6881) MODULE_LICENSE("GPL");
^8f3ce5b39 (kx 2023-10-28 12:00:06 +0300 6882) MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler");
Orange Pi5 kernel
Deprecated Linux kernel 5.10.110 for OrangePi 5/5B/5+ boards
3 Commits
0 Branches
0 Tags