2 * Budget Fair Queueing (BFQ) I/O scheduler.
4 * Based on ideas and code from CFQ:
5 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
7 * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
8 * Paolo Valente <paolo.valente@unimore.it>
10 * Copyright (C) 2010 Paolo Valente <paolo.valente@unimore.it>
11 * Arianna Avanzini <avanzini@google.com>
13 * Copyright (C) 2017 Paolo Valente <paolo.valente@linaro.org>
15 * This program is free software; you can redistribute it and/or
16 * modify it under the terms of the GNU General Public License as
17 * published by the Free Software Foundation; either version 2 of the
18 * License, or (at your option) any later version.
20 * This program is distributed in the hope that it will be useful,
21 * but WITHOUT ANY WARRANTY; without even the implied warranty of
22 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
23 * General Public License for more details.
25 * BFQ is a proportional-share I/O scheduler, with some extra
26 * low-latency capabilities. BFQ also supports full hierarchical
27 * scheduling through cgroups. Next paragraphs provide an introduction
28 * on BFQ inner workings. Details on BFQ benefits, usage and
29 * limitations can be found in Documentation/block/bfq-iosched.txt.
31 * BFQ is a proportional-share storage-I/O scheduling algorithm based
32 * on the slice-by-slice service scheme of CFQ. But BFQ assigns
33 * budgets, measured in number of sectors, to processes instead of
34 * time slices. The device is not granted to the in-service process
35 * for a given time slice, but until it has exhausted its assigned
36 * budget. This change from the time to the service domain enables BFQ
37 * to distribute the device throughput among processes as desired,
38 * without any distortion due to throughput fluctuations, or to device
39 * internal queueing. BFQ uses an ad hoc internal scheduler, called
40 * B-WF2Q+, to schedule processes according to their budgets. More
41 * precisely, BFQ schedules queues associated with processes. Each
42 * process/queue is assigned a user-configurable weight, and B-WF2Q+
43 * guarantees that each queue receives a fraction of the throughput
44 * proportional to its weight. Thanks to the accurate policy of
45 * B-WF2Q+, BFQ can afford to assign high budgets to I/O-bound
46 * processes issuing sequential requests (to boost the throughput),
47 * and yet guarantee a low latency to interactive and soft real-time
50 * In particular, to provide these low-latency guarantees, BFQ
51 * explicitly privileges the I/O of two classes of time-sensitive
52 * applications: interactive and soft real-time. In more detail, BFQ
53 * behaves this way if the low_latency parameter is set (default
54 * configuration). This feature enables BFQ to provide applications in
55 * these classes with a very low latency.
57 * To implement this feature, BFQ constantly tries to detect whether
58 * the I/O requests in a bfq_queue come from an interactive or a soft
59 * real-time application. For brevity, in these cases, the queue is
60 * said to be interactive or soft real-time. In both cases, BFQ
61 * privileges the service of the queue, over that of non-interactive
62 * and non-soft-real-time queues. This privileging is performed,
63 * mainly, by raising the weight of the queue. So, for brevity, we
64 * call just weight-raising periods the time periods during which a
65 * queue is privileged, because deemed interactive or soft real-time.
67 * The detection of soft real-time queues/applications is described in
68 * detail in the comments on the function
69 * bfq_bfqq_softrt_next_start. On the other hand, the detection of an
70 * interactive queue works as follows: a queue is deemed interactive
71 * if it is constantly non empty only for a limited time interval,
72 * after which it does become empty. The queue may be deemed
73 * interactive again (for a limited time), if it restarts being
74 * constantly non empty, provided that this happens only after the
75 * queue has remained empty for a given minimum idle time.
77 * By default, BFQ computes automatically the above maximum time
78 * interval, i.e., the time interval after which a constantly
79 * non-empty queue stops being deemed interactive. Since a queue is
80 * weight-raised while it is deemed interactive, this maximum time
81 * interval happens to coincide with the (maximum) duration of the
82 * weight-raising for interactive queues.
84 * Finally, BFQ also features additional heuristics for
85 * preserving both a low latency and a high throughput on NCQ-capable,
86 * rotational or flash-based devices, and to get the job done quickly
87 * for applications consisting in many I/O-bound processes.
89 * NOTE: if the main or only goal, with a given device, is to achieve
90 * the maximum-possible throughput at all times, then do switch off
91 * all low-latency heuristics for that device, by setting low_latency
94 * BFQ is described in [1], where also a reference to the initial,
95 * more theoretical paper on BFQ can be found. The interested reader
96 * can find in the latter paper full details on the main algorithm, as
97 * well as formulas of the guarantees and formal proofs of all the
98 * properties. With respect to the version of BFQ presented in these
99 * papers, this implementation adds a few more heuristics, such as the
100 * ones that guarantee a low latency to interactive and soft real-time
101 * applications, and a hierarchical extension based on H-WF2Q+.
103 * B-WF2Q+ is based on WF2Q+, which is described in [2], together with
104 * H-WF2Q+, while the augmented tree used here to implement B-WF2Q+
105 * with O(log N) complexity derives from the one introduced with EEVDF
108 * [1] P. Valente, A. Avanzini, "Evolution of the BFQ Storage I/O
109 * Scheduler", Proceedings of the First Workshop on Mobile System
110 * Technologies (MST-2015), May 2015.
111 * http://algogroup.unimore.it/people/paolo/disk_sched/mst-2015.pdf
113 * [2] Jon C.R. Bennett and H. Zhang, "Hierarchical Packet Fair Queueing
114 * Algorithms", IEEE/ACM Transactions on Networking, 5(5):675-689,
117 * http://www.cs.cmu.edu/~hzhang/papers/TON-97-Oct.ps.gz
119 * [3] I. Stoica and H. Abdel-Wahab, "Earliest Eligible Virtual Deadline
120 * First: A Flexible and Accurate Mechanism for Proportional Share
121 * Resource Allocation", technical report.
123 * http://www.cs.berkeley.edu/~istoica/papers/eevdf-tr-95.pdf
125 #include <linux/module.h>
126 #include <linux/slab.h>
127 #include <linux/blkdev.h>
128 #include <linux/cgroup.h>
129 #include <linux/elevator.h>
130 #include <linux/ktime.h>
131 #include <linux/rbtree.h>
132 #include <linux/ioprio.h>
133 #include <linux/sbitmap.h>
134 #include <linux/delay.h>
138 #include "blk-mq-tag.h"
139 #include "blk-mq-sched.h"
140 #include "bfq-iosched.h"
143 #define BFQ_BFQQ_FNS(name) \
144 void bfq_mark_bfqq_##name(struct bfq_queue *bfqq) \
146 __set_bit(BFQQF_##name, &(bfqq)->flags); \
148 void bfq_clear_bfqq_##name(struct bfq_queue *bfqq) \
150 __clear_bit(BFQQF_##name, &(bfqq)->flags); \
152 int bfq_bfqq_##name(const struct bfq_queue *bfqq) \
154 return test_bit(BFQQF_##name, &(bfqq)->flags); \
157 BFQ_BFQQ_FNS(just_created);
159 BFQ_BFQQ_FNS(wait_request);
160 BFQ_BFQQ_FNS(non_blocking_wait_rq);
161 BFQ_BFQQ_FNS(fifo_expire);
162 BFQ_BFQQ_FNS(has_short_ttime);
164 BFQ_BFQQ_FNS(IO_bound);
165 BFQ_BFQQ_FNS(in_large_burst);
167 BFQ_BFQQ_FNS(split_coop);
168 BFQ_BFQQ_FNS(softrt_update);
169 #undef BFQ_BFQQ_FNS \
171 /* Expiration time of sync (0) and async (1) requests, in ns. */
172 static const u64 bfq_fifo_expire[2] = { NSEC_PER_SEC / 4, NSEC_PER_SEC / 8 };
174 /* Maximum backwards seek (magic number lifted from CFQ), in KiB. */
175 static const int bfq_back_max = 16 * 1024;
177 /* Penalty of a backwards seek, in number of sectors. */
178 static const int bfq_back_penalty = 2;
180 /* Idling period duration, in ns. */
181 static u64 bfq_slice_idle = NSEC_PER_SEC / 125;
183 /* Minimum number of assigned budgets for which stats are safe to compute. */
184 static const int bfq_stats_min_budgets = 194;
186 /* Default maximum budget values, in sectors and number of requests. */
187 static const int bfq_default_max_budget = 16 * 1024;
190 * When a sync request is dispatched, the queue that contains that
191 * request, and all the ancestor entities of that queue, are charged
192 * with the number of sectors of the request. In constrast, if the
193 * request is async, then the queue and its ancestor entities are
194 * charged with the number of sectors of the request, multiplied by
195 * the factor below. This throttles the bandwidth for async I/O,
196 * w.r.t. to sync I/O, and it is done to counter the tendency of async
197 * writes to steal I/O throughput to reads.
199 * The current value of this parameter is the result of a tuning with
200 * several hardware and software configurations. We tried to find the
201 * lowest value for which writes do not cause noticeable problems to
202 * reads. In fact, the lower this parameter, the stabler I/O control,
203 * in the following respect. The lower this parameter is, the less
204 * the bandwidth enjoyed by a group decreases
205 * - when the group does writes, w.r.t. to when it does reads;
206 * - when other groups do reads, w.r.t. to when they do writes.
208 static const int bfq_async_charge_factor = 3;
210 /* Default timeout values, in jiffies, approximating CFQ defaults. */
211 const int bfq_timeout = HZ / 8;
214 * Time limit for merging (see comments in bfq_setup_cooperator). Set
215 * to the slowest value that, in our tests, proved to be effective in
216 * removing false positives, while not causing true positives to miss
219 * As can be deduced from the low time limit below, queue merging, if
220 * successful, happens at the very beggining of the I/O of the involved
221 * cooperating processes, as a consequence of the arrival of the very
222 * first requests from each cooperator. After that, there is very
223 * little chance to find cooperators.
225 static const unsigned long bfq_merge_time_limit = HZ/10;
227 static struct kmem_cache *bfq_pool;
229 /* Below this threshold (in ns), we consider thinktime immediate. */
230 #define BFQ_MIN_TT (2 * NSEC_PER_MSEC)
232 /* hw_tag detection: parallel requests threshold and min samples needed. */
233 #define BFQ_HW_QUEUE_THRESHOLD 4
234 #define BFQ_HW_QUEUE_SAMPLES 32
236 #define BFQQ_SEEK_THR (sector_t)(8 * 100)
237 #define BFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
238 #define BFQ_RQ_SEEKY(bfqd, last_pos, rq) \
239 (get_sdist(last_pos, rq) > \
241 (!blk_queue_nonrot(bfqd->queue) || \
242 blk_rq_sectors(rq) < BFQQ_SECT_THR_NONROT))
243 #define BFQQ_CLOSE_THR (sector_t)(8 * 1024)
244 #define BFQQ_SEEKY(bfqq) (hweight32(bfqq->seek_history) > 19)
246 /* Min number of samples required to perform peak-rate update */
247 #define BFQ_RATE_MIN_SAMPLES 32
248 /* Min observation time interval required to perform a peak-rate update (ns) */
249 #define BFQ_RATE_MIN_INTERVAL (300*NSEC_PER_MSEC)
250 /* Target observation time interval for a peak-rate update (ns) */
251 #define BFQ_RATE_REF_INTERVAL NSEC_PER_SEC
254 * Shift used for peak-rate fixed precision calculations.
256 * - the current shift: 16 positions
257 * - the current type used to store rate: u32
258 * - the current unit of measure for rate: [sectors/usec], or, more precisely,
259 * [(sectors/usec) / 2^BFQ_RATE_SHIFT] to take into account the shift,
260 * the range of rates that can be stored is
261 * [1 / 2^BFQ_RATE_SHIFT, 2^(32 - BFQ_RATE_SHIFT)] sectors/usec =
262 * [1 / 2^16, 2^16] sectors/usec = [15e-6, 65536] sectors/usec =
263 * [15, 65G] sectors/sec
264 * Which, assuming a sector size of 512B, corresponds to a range of
267 #define BFQ_RATE_SHIFT 16
270 * When configured for computing the duration of the weight-raising
271 * for interactive queues automatically (see the comments at the
272 * beginning of this file), BFQ does it using the following formula:
273 * duration = (ref_rate / r) * ref_wr_duration,
274 * where r is the peak rate of the device, and ref_rate and
275 * ref_wr_duration are two reference parameters. In particular,
276 * ref_rate is the peak rate of the reference storage device (see
277 * below), and ref_wr_duration is about the maximum time needed, with
278 * BFQ and while reading two files in parallel, to load typical large
279 * applications on the reference device (see the comments on
280 * max_service_from_wr below, for more details on how ref_wr_duration
281 * is obtained). In practice, the slower/faster the device at hand
282 * is, the more/less it takes to load applications with respect to the
283 * reference device. Accordingly, the longer/shorter BFQ grants
284 * weight raising to interactive applications.
286 * BFQ uses two different reference pairs (ref_rate, ref_wr_duration),
287 * depending on whether the device is rotational or non-rotational.
289 * In the following definitions, ref_rate[0] and ref_wr_duration[0]
290 * are the reference values for a rotational device, whereas
291 * ref_rate[1] and ref_wr_duration[1] are the reference values for a
292 * non-rotational device. The reference rates are not the actual peak
293 * rates of the devices used as a reference, but slightly lower
294 * values. The reason for using slightly lower values is that the
295 * peak-rate estimator tends to yield slightly lower values than the
296 * actual peak rate (it can yield the actual peak rate only if there
297 * is only one process doing I/O, and the process does sequential
300 * The reference peak rates are measured in sectors/usec, left-shifted
303 static int ref_rate[2] = {14000, 33000};
305 * To improve readability, a conversion function is used to initialize
306 * the following array, which entails that the array can be
307 * initialized only in a function.
309 static int ref_wr_duration[2];
312 * BFQ uses the above-detailed, time-based weight-raising mechanism to
313 * privilege interactive tasks. This mechanism is vulnerable to the
314 * following false positives: I/O-bound applications that will go on
315 * doing I/O for much longer than the duration of weight
316 * raising. These applications have basically no benefit from being
317 * weight-raised at the beginning of their I/O. On the opposite end,
318 * while being weight-raised, these applications
319 * a) unjustly steal throughput to applications that may actually need
321 * b) make BFQ uselessly perform device idling; device idling results
322 * in loss of device throughput with most flash-based storage, and may
323 * increase latencies when used purposelessly.
325 * BFQ tries to reduce these problems, by adopting the following
326 * countermeasure. To introduce this countermeasure, we need first to
327 * finish explaining how the duration of weight-raising for
328 * interactive tasks is computed.
330 * For a bfq_queue deemed as interactive, the duration of weight
331 * raising is dynamically adjusted, as a function of the estimated
332 * peak rate of the device, so as to be equal to the time needed to
333 * execute the 'largest' interactive task we benchmarked so far. By
334 * largest task, we mean the task for which each involved process has
335 * to do more I/O than for any of the other tasks we benchmarked. This
336 * reference interactive task is the start-up of LibreOffice Writer,
337 * and in this task each process/bfq_queue needs to have at most ~110K
338 * sectors transferred.
340 * This last piece of information enables BFQ to reduce the actual
341 * duration of weight-raising for at least one class of I/O-bound
342 * applications: those doing sequential or quasi-sequential I/O. An
343 * example is file copy. In fact, once started, the main I/O-bound
344 * processes of these applications usually consume the above 110K
345 * sectors in much less time than the processes of an application that
346 * is starting, because these I/O-bound processes will greedily devote
347 * almost all their CPU cycles only to their target,
348 * throughput-friendly I/O operations. This is even more true if BFQ
349 * happens to be underestimating the device peak rate, and thus
350 * overestimating the duration of weight raising. But, according to
351 * our measurements, once transferred 110K sectors, these processes
352 * have no right to be weight-raised any longer.
354 * Basing on the last consideration, BFQ ends weight-raising for a
355 * bfq_queue if the latter happens to have received an amount of
356 * service at least equal to the following constant. The constant is
357 * set to slightly more than 110K, to have a minimum safety margin.
359 * This early ending of weight-raising reduces the amount of time
360 * during which interactive false positives cause the two problems
361 * described at the beginning of these comments.
363 static const unsigned long max_service_from_wr = 120000;
365 #define RQ_BIC(rq) icq_to_bic((rq)->elv.priv[0])
366 #define RQ_BFQQ(rq) ((rq)->elv.priv[1])
368 struct bfq_queue *bic_to_bfqq(struct bfq_io_cq *bic, bool is_sync)
370 return bic->bfqq[is_sync];
373 void bic_set_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq, bool is_sync)
375 bic->bfqq[is_sync] = bfqq;
378 struct bfq_data *bic_to_bfqd(struct bfq_io_cq *bic)
380 return bic->icq.q->elevator->elevator_data;
384 * icq_to_bic - convert iocontext queue structure to bfq_io_cq.
385 * @icq: the iocontext queue.
387 static struct bfq_io_cq *icq_to_bic(struct io_cq *icq)
389 /* bic->icq is the first member, %NULL will convert to %NULL */
390 return container_of(icq, struct bfq_io_cq, icq);
394 * bfq_bic_lookup - search into @ioc a bic associated to @bfqd.
395 * @bfqd: the lookup key.
396 * @ioc: the io_context of the process doing I/O.
397 * @q: the request queue.
399 static struct bfq_io_cq *bfq_bic_lookup(struct bfq_data *bfqd,
400 struct io_context *ioc,
401 struct request_queue *q)
405 struct bfq_io_cq *icq;
407 spin_lock_irqsave(&q->queue_lock, flags);
408 icq = icq_to_bic(ioc_lookup_icq(ioc, q));
409 spin_unlock_irqrestore(&q->queue_lock, flags);
418 * Scheduler run of queue, if there are requests pending and no one in the
419 * driver that will restart queueing.
421 void bfq_schedule_dispatch(struct bfq_data *bfqd)
423 if (bfqd->queued != 0) {
424 bfq_log(bfqd, "schedule dispatch");
425 blk_mq_run_hw_queues(bfqd->queue, true);
429 #define bfq_class_idle(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
430 #define bfq_class_rt(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_RT)
432 #define bfq_sample_valid(samples) ((samples) > 80)
435 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
436 * We choose the request that is closesr to the head right now. Distance
437 * behind the head is penalized and only allowed to a certain extent.
439 static struct request *bfq_choose_req(struct bfq_data *bfqd,
444 sector_t s1, s2, d1 = 0, d2 = 0;
445 unsigned long back_max;
446 #define BFQ_RQ1_WRAP 0x01 /* request 1 wraps */
447 #define BFQ_RQ2_WRAP 0x02 /* request 2 wraps */
448 unsigned int wrap = 0; /* bit mask: requests behind the disk head? */
450 if (!rq1 || rq1 == rq2)
455 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
457 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
459 if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
461 else if ((rq2->cmd_flags & REQ_META) && !(rq1->cmd_flags & REQ_META))
464 s1 = blk_rq_pos(rq1);
465 s2 = blk_rq_pos(rq2);
468 * By definition, 1KiB is 2 sectors.
470 back_max = bfqd->bfq_back_max * 2;
473 * Strict one way elevator _except_ in the case where we allow
474 * short backward seeks which are biased as twice the cost of a
475 * similar forward seek.
479 else if (s1 + back_max >= last)
480 d1 = (last - s1) * bfqd->bfq_back_penalty;
482 wrap |= BFQ_RQ1_WRAP;
486 else if (s2 + back_max >= last)
487 d2 = (last - s2) * bfqd->bfq_back_penalty;
489 wrap |= BFQ_RQ2_WRAP;
491 /* Found required data */
494 * By doing switch() on the bit mask "wrap" we avoid having to
495 * check two variables for all permutations: --> faster!
498 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
513 case BFQ_RQ1_WRAP|BFQ_RQ2_WRAP: /* both rqs wrapped */
516 * Since both rqs are wrapped,
517 * start with the one that's further behind head
518 * (--> only *one* back seek required),
519 * since back seek takes more time than forward.
529 * Async I/O can easily starve sync I/O (both sync reads and sync
530 * writes), by consuming all tags. Similarly, storms of sync writes,
531 * such as those that sync(2) may trigger, can starve sync reads.
532 * Limit depths of async I/O and sync writes so as to counter both
535 static void bfq_limit_depth(unsigned int op, struct blk_mq_alloc_data *data)
537 struct bfq_data *bfqd = data->q->elevator->elevator_data;
539 if (op_is_sync(op) && !op_is_write(op))
542 data->shallow_depth =
543 bfqd->word_depths[!!bfqd->wr_busy_queues][op_is_sync(op)];
545 bfq_log(bfqd, "[%s] wr_busy %d sync %d depth %u",
546 __func__, bfqd->wr_busy_queues, op_is_sync(op),
547 data->shallow_depth);
550 static struct bfq_queue *
551 bfq_rq_pos_tree_lookup(struct bfq_data *bfqd, struct rb_root *root,
552 sector_t sector, struct rb_node **ret_parent,
553 struct rb_node ***rb_link)
555 struct rb_node **p, *parent;
556 struct bfq_queue *bfqq = NULL;
564 bfqq = rb_entry(parent, struct bfq_queue, pos_node);
567 * Sort strictly based on sector. Smallest to the left,
568 * largest to the right.
570 if (sector > blk_rq_pos(bfqq->next_rq))
572 else if (sector < blk_rq_pos(bfqq->next_rq))
580 *ret_parent = parent;
584 bfq_log(bfqd, "rq_pos_tree_lookup %llu: returning %d",
585 (unsigned long long)sector,
586 bfqq ? bfqq->pid : 0);
591 static bool bfq_too_late_for_merging(struct bfq_queue *bfqq)
593 return bfqq->service_from_backlogged > 0 &&
594 time_is_before_jiffies(bfqq->first_IO_time +
595 bfq_merge_time_limit);
598 void bfq_pos_tree_add_move(struct bfq_data *bfqd, struct bfq_queue *bfqq)
600 struct rb_node **p, *parent;
601 struct bfq_queue *__bfqq;
603 if (bfqq->pos_root) {
604 rb_erase(&bfqq->pos_node, bfqq->pos_root);
605 bfqq->pos_root = NULL;
609 * bfqq cannot be merged any longer (see comments in
610 * bfq_setup_cooperator): no point in adding bfqq into the
613 if (bfq_too_late_for_merging(bfqq))
616 if (bfq_class_idle(bfqq))
621 bfqq->pos_root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree;
622 __bfqq = bfq_rq_pos_tree_lookup(bfqd, bfqq->pos_root,
623 blk_rq_pos(bfqq->next_rq), &parent, &p);
625 rb_link_node(&bfqq->pos_node, parent, p);
626 rb_insert_color(&bfqq->pos_node, bfqq->pos_root);
628 bfqq->pos_root = NULL;
632 * The following function returns true if every queue must receive the
633 * same share of the throughput (this condition is used when deciding
634 * whether idling may be disabled, see the comments in the function
635 * bfq_better_to_idle()).
637 * Such a scenario occurs when:
638 * 1) all active queues have the same weight,
639 * 2) all active queues belong to the same I/O-priority class,
640 * 3) all active groups at the same level in the groups tree have the same
642 * 4) all active groups at the same level in the groups tree have the same
643 * number of children.
645 * Unfortunately, keeping the necessary state for evaluating exactly
646 * the last two symmetry sub-conditions above would be quite complex
647 * and time consuming. Therefore this function evaluates, instead,
648 * only the following stronger three sub-conditions, for which it is
649 * much easier to maintain the needed state:
650 * 1) all active queues have the same weight,
651 * 2) all active queues belong to the same I/O-priority class,
652 * 3) there are no active groups.
653 * In particular, the last condition is always true if hierarchical
654 * support or the cgroups interface are not enabled, thus no state
655 * needs to be maintained in this case.
657 static bool bfq_symmetric_scenario(struct bfq_data *bfqd)
660 * For queue weights to differ, queue_weights_tree must contain
661 * at least two nodes.
663 bool varied_queue_weights = !RB_EMPTY_ROOT(&bfqd->queue_weights_tree) &&
664 (bfqd->queue_weights_tree.rb_node->rb_left ||
665 bfqd->queue_weights_tree.rb_node->rb_right);
667 bool multiple_classes_busy =
668 (bfqd->busy_queues[0] && bfqd->busy_queues[1]) ||
669 (bfqd->busy_queues[0] && bfqd->busy_queues[2]) ||
670 (bfqd->busy_queues[1] && bfqd->busy_queues[2]);
673 * For queue weights to differ, queue_weights_tree must contain
674 * at least two nodes.
676 return !(varied_queue_weights || multiple_classes_busy
677 #ifdef BFQ_GROUP_IOSCHED_ENABLED
678 || bfqd->num_groups_with_pending_reqs > 0
684 * If the weight-counter tree passed as input contains no counter for
685 * the weight of the input queue, then add that counter; otherwise just
686 * increment the existing counter.
688 * Note that weight-counter trees contain few nodes in mostly symmetric
689 * scenarios. For example, if all queues have the same weight, then the
690 * weight-counter tree for the queues may contain at most one node.
691 * This holds even if low_latency is on, because weight-raised queues
692 * are not inserted in the tree.
693 * In most scenarios, the rate at which nodes are created/destroyed
696 void bfq_weights_tree_add(struct bfq_data *bfqd, struct bfq_queue *bfqq,
697 struct rb_root *root)
699 struct bfq_entity *entity = &bfqq->entity;
700 struct rb_node **new = &(root->rb_node), *parent = NULL;
703 * Do not insert if the queue is already associated with a
704 * counter, which happens if:
705 * 1) a request arrival has caused the queue to become both
706 * non-weight-raised, and hence change its weight, and
707 * backlogged; in this respect, each of the two events
708 * causes an invocation of this function,
709 * 2) this is the invocation of this function caused by the
710 * second event. This second invocation is actually useless,
711 * and we handle this fact by exiting immediately. More
712 * efficient or clearer solutions might possibly be adopted.
714 if (bfqq->weight_counter)
718 struct bfq_weight_counter *__counter = container_of(*new,
719 struct bfq_weight_counter,
723 if (entity->weight == __counter->weight) {
724 bfqq->weight_counter = __counter;
727 if (entity->weight < __counter->weight)
728 new = &((*new)->rb_left);
730 new = &((*new)->rb_right);
733 bfqq->weight_counter = kzalloc(sizeof(struct bfq_weight_counter),
737 * In the unlucky event of an allocation failure, we just
738 * exit. This will cause the weight of queue to not be
739 * considered in bfq_symmetric_scenario, which, in its turn,
740 * causes the scenario to be deemed wrongly symmetric in case
741 * bfqq's weight would have been the only weight making the
742 * scenario asymmetric. On the bright side, no unbalance will
743 * however occur when bfqq becomes inactive again (the
744 * invocation of this function is triggered by an activation
745 * of queue). In fact, bfq_weights_tree_remove does nothing
746 * if !bfqq->weight_counter.
748 if (unlikely(!bfqq->weight_counter))
751 bfqq->weight_counter->weight = entity->weight;
752 rb_link_node(&bfqq->weight_counter->weights_node, parent, new);
753 rb_insert_color(&bfqq->weight_counter->weights_node, root);
756 bfqq->weight_counter->num_active++;
761 * Decrement the weight counter associated with the queue, and, if the
762 * counter reaches 0, remove the counter from the tree.
763 * See the comments to the function bfq_weights_tree_add() for considerations
766 void __bfq_weights_tree_remove(struct bfq_data *bfqd,
767 struct bfq_queue *bfqq,
768 struct rb_root *root)
770 if (!bfqq->weight_counter)
773 bfqq->weight_counter->num_active--;
774 if (bfqq->weight_counter->num_active > 0)
775 goto reset_entity_pointer;
777 rb_erase(&bfqq->weight_counter->weights_node, root);
778 kfree(bfqq->weight_counter);
780 reset_entity_pointer:
781 bfqq->weight_counter = NULL;
786 * Invoke __bfq_weights_tree_remove on bfqq and decrement the number
787 * of active groups for each queue's inactive parent entity.
789 void bfq_weights_tree_remove(struct bfq_data *bfqd,
790 struct bfq_queue *bfqq)
792 struct bfq_entity *entity = bfqq->entity.parent;
794 for_each_entity(entity) {
795 struct bfq_sched_data *sd = entity->my_sched_data;
797 if (sd->next_in_service || sd->in_service_entity) {
799 * entity is still active, because either
800 * next_in_service or in_service_entity is not
801 * NULL (see the comments on the definition of
802 * next_in_service for details on why
803 * in_service_entity must be checked too).
805 * As a consequence, its parent entities are
806 * active as well, and thus this loop must
813 * The decrement of num_groups_with_pending_reqs is
814 * not performed immediately upon the deactivation of
815 * entity, but it is delayed to when it also happens
816 * that the first leaf descendant bfqq of entity gets
817 * all its pending requests completed. The following
818 * instructions perform this delayed decrement, if
819 * needed. See the comments on
820 * num_groups_with_pending_reqs for details.
822 if (entity->in_groups_with_pending_reqs) {
823 entity->in_groups_with_pending_reqs = false;
824 bfqd->num_groups_with_pending_reqs--;
829 * Next function is invoked last, because it causes bfqq to be
830 * freed if the following holds: bfqq is not in service and
831 * has no dispatched request. DO NOT use bfqq after the next
832 * function invocation.
834 __bfq_weights_tree_remove(bfqd, bfqq,
835 &bfqd->queue_weights_tree);
839 * Return expired entry, or NULL to just start from scratch in rbtree.
841 static struct request *bfq_check_fifo(struct bfq_queue *bfqq,
842 struct request *last)
846 if (bfq_bfqq_fifo_expire(bfqq))
849 bfq_mark_bfqq_fifo_expire(bfqq);
851 rq = rq_entry_fifo(bfqq->fifo.next);
853 if (rq == last || ktime_get_ns() < rq->fifo_time)
856 bfq_log_bfqq(bfqq->bfqd, bfqq, "check_fifo: returned %p", rq);
860 static struct request *bfq_find_next_rq(struct bfq_data *bfqd,
861 struct bfq_queue *bfqq,
862 struct request *last)
864 struct rb_node *rbnext = rb_next(&last->rb_node);
865 struct rb_node *rbprev = rb_prev(&last->rb_node);
866 struct request *next, *prev = NULL;
868 /* Follow expired path, else get first next available. */
869 next = bfq_check_fifo(bfqq, last);
874 prev = rb_entry_rq(rbprev);
877 next = rb_entry_rq(rbnext);
879 rbnext = rb_first(&bfqq->sort_list);
880 if (rbnext && rbnext != &last->rb_node)
881 next = rb_entry_rq(rbnext);
884 return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last));
887 /* see the definition of bfq_async_charge_factor for details */
888 static unsigned long bfq_serv_to_charge(struct request *rq,
889 struct bfq_queue *bfqq)
891 if (bfq_bfqq_sync(bfqq) || bfqq->wr_coeff > 1)
892 return blk_rq_sectors(rq);
894 return blk_rq_sectors(rq) * bfq_async_charge_factor;
898 * bfq_updated_next_req - update the queue after a new next_rq selection.
899 * @bfqd: the device data the queue belongs to.
900 * @bfqq: the queue to update.
902 * If the first request of a queue changes we make sure that the queue
903 * has enough budget to serve at least its first request (if the
904 * request has grown). We do this because if the queue has not enough
905 * budget for its first request, it has to go through two dispatch
906 * rounds to actually get it dispatched.
908 static void bfq_updated_next_req(struct bfq_data *bfqd,
909 struct bfq_queue *bfqq)
911 struct bfq_entity *entity = &bfqq->entity;
912 struct request *next_rq = bfqq->next_rq;
913 unsigned long new_budget;
918 if (bfqq == bfqd->in_service_queue)
920 * In order not to break guarantees, budgets cannot be
921 * changed after an entity has been selected.
925 new_budget = max_t(unsigned long,
926 max_t(unsigned long, bfqq->max_budget,
927 bfq_serv_to_charge(next_rq, bfqq)),
929 if (entity->budget != new_budget) {
930 entity->budget = new_budget;
931 bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu",
933 bfq_requeue_bfqq(bfqd, bfqq, false);
937 static unsigned int bfq_wr_duration(struct bfq_data *bfqd)
941 if (bfqd->bfq_wr_max_time > 0)
942 return bfqd->bfq_wr_max_time;
944 dur = bfqd->rate_dur_prod;
945 do_div(dur, bfqd->peak_rate);
948 * Limit duration between 3 and 25 seconds. The upper limit
949 * has been conservatively set after the following worst case:
950 * on a QEMU/KVM virtual machine
951 * - running in a slow PC
952 * - with a virtual disk stacked on a slow low-end 5400rpm HDD
953 * - serving a heavy I/O workload, such as the sequential reading
955 * mplayer took 23 seconds to start, if constantly weight-raised.
957 * As for higher values than that accomodating the above bad
958 * scenario, tests show that higher values would often yield
959 * the opposite of the desired result, i.e., would worsen
960 * responsiveness by allowing non-interactive applications to
961 * preserve weight raising for too long.
963 * On the other end, lower values than 3 seconds make it
964 * difficult for most interactive tasks to complete their jobs
965 * before weight-raising finishes.
967 return clamp_val(dur, msecs_to_jiffies(3000), msecs_to_jiffies(25000));
970 /* switch back from soft real-time to interactive weight raising */
971 static void switch_back_to_interactive_wr(struct bfq_queue *bfqq,
972 struct bfq_data *bfqd)
974 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
975 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
976 bfqq->last_wr_start_finish = bfqq->wr_start_at_switch_to_srt;
980 bfq_bfqq_resume_state(struct bfq_queue *bfqq, struct bfq_data *bfqd,
981 struct bfq_io_cq *bic, bool bfq_already_existing)
983 unsigned int old_wr_coeff = bfqq->wr_coeff;
984 bool busy = bfq_already_existing && bfq_bfqq_busy(bfqq);
986 if (bic->saved_has_short_ttime)
987 bfq_mark_bfqq_has_short_ttime(bfqq);
989 bfq_clear_bfqq_has_short_ttime(bfqq);
991 if (bic->saved_IO_bound)
992 bfq_mark_bfqq_IO_bound(bfqq);
994 bfq_clear_bfqq_IO_bound(bfqq);
996 bfqq->ttime = bic->saved_ttime;
997 bfqq->wr_coeff = bic->saved_wr_coeff;
998 bfqq->wr_start_at_switch_to_srt = bic->saved_wr_start_at_switch_to_srt;
999 bfqq->last_wr_start_finish = bic->saved_last_wr_start_finish;
1000 bfqq->wr_cur_max_time = bic->saved_wr_cur_max_time;
1002 if (bfqq->wr_coeff > 1 && (bfq_bfqq_in_large_burst(bfqq) ||
1003 time_is_before_jiffies(bfqq->last_wr_start_finish +
1004 bfqq->wr_cur_max_time))) {
1005 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
1006 !bfq_bfqq_in_large_burst(bfqq) &&
1007 time_is_after_eq_jiffies(bfqq->wr_start_at_switch_to_srt +
1008 bfq_wr_duration(bfqd))) {
1009 switch_back_to_interactive_wr(bfqq, bfqd);
1012 bfq_log_bfqq(bfqq->bfqd, bfqq,
1013 "resume state: switching off wr");
1017 /* make sure weight will be updated, however we got here */
1018 bfqq->entity.prio_changed = 1;
1023 if (old_wr_coeff == 1 && bfqq->wr_coeff > 1)
1024 bfqd->wr_busy_queues++;
1025 else if (old_wr_coeff > 1 && bfqq->wr_coeff == 1)
1026 bfqd->wr_busy_queues--;
1029 static int bfqq_process_refs(struct bfq_queue *bfqq)
1031 return bfqq->ref - bfqq->allocated - bfqq->entity.on_st -
1032 (bfqq->weight_counter != NULL);
1035 /* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */
1036 static void bfq_reset_burst_list(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1038 struct bfq_queue *item;
1039 struct hlist_node *n;
1041 hlist_for_each_entry_safe(item, n, &bfqd->burst_list, burst_list_node)
1042 hlist_del_init(&item->burst_list_node);
1043 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1044 bfqd->burst_size = 1;
1045 bfqd->burst_parent_entity = bfqq->entity.parent;
1048 /* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */
1049 static void bfq_add_to_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1051 /* Increment burst size to take into account also bfqq */
1054 if (bfqd->burst_size == bfqd->bfq_large_burst_thresh) {
1055 struct bfq_queue *pos, *bfqq_item;
1056 struct hlist_node *n;
1059 * Enough queues have been activated shortly after each
1060 * other to consider this burst as large.
1062 bfqd->large_burst = true;
1065 * We can now mark all queues in the burst list as
1066 * belonging to a large burst.
1068 hlist_for_each_entry(bfqq_item, &bfqd->burst_list,
1070 bfq_mark_bfqq_in_large_burst(bfqq_item);
1071 bfq_mark_bfqq_in_large_burst(bfqq);
1074 * From now on, and until the current burst finishes, any
1075 * new queue being activated shortly after the last queue
1076 * was inserted in the burst can be immediately marked as
1077 * belonging to a large burst. So the burst list is not
1078 * needed any more. Remove it.
1080 hlist_for_each_entry_safe(pos, n, &bfqd->burst_list,
1082 hlist_del_init(&pos->burst_list_node);
1084 * Burst not yet large: add bfqq to the burst list. Do
1085 * not increment the ref counter for bfqq, because bfqq
1086 * is removed from the burst list before freeing bfqq
1089 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1093 * If many queues belonging to the same group happen to be created
1094 * shortly after each other, then the processes associated with these
1095 * queues have typically a common goal. In particular, bursts of queue
1096 * creations are usually caused by services or applications that spawn
1097 * many parallel threads/processes. Examples are systemd during boot,
1098 * or git grep. To help these processes get their job done as soon as
1099 * possible, it is usually better to not grant either weight-raising
1100 * or device idling to their queues.
1102 * In this comment we describe, firstly, the reasons why this fact
1103 * holds, and, secondly, the next function, which implements the main
1104 * steps needed to properly mark these queues so that they can then be
1105 * treated in a different way.
1107 * The above services or applications benefit mostly from a high
1108 * throughput: the quicker the requests of the activated queues are
1109 * cumulatively served, the sooner the target job of these queues gets
1110 * completed. As a consequence, weight-raising any of these queues,
1111 * which also implies idling the device for it, is almost always
1112 * counterproductive. In most cases it just lowers throughput.
1114 * On the other hand, a burst of queue creations may be caused also by
1115 * the start of an application that does not consist of a lot of
1116 * parallel I/O-bound threads. In fact, with a complex application,
1117 * several short processes may need to be executed to start-up the
1118 * application. In this respect, to start an application as quickly as
1119 * possible, the best thing to do is in any case to privilege the I/O
1120 * related to the application with respect to all other
1121 * I/O. Therefore, the best strategy to start as quickly as possible
1122 * an application that causes a burst of queue creations is to
1123 * weight-raise all the queues created during the burst. This is the
1124 * exact opposite of the best strategy for the other type of bursts.
1126 * In the end, to take the best action for each of the two cases, the
1127 * two types of bursts need to be distinguished. Fortunately, this
1128 * seems relatively easy, by looking at the sizes of the bursts. In
1129 * particular, we found a threshold such that only bursts with a
1130 * larger size than that threshold are apparently caused by
1131 * services or commands such as systemd or git grep. For brevity,
1132 * hereafter we call just 'large' these bursts. BFQ *does not*
1133 * weight-raise queues whose creation occurs in a large burst. In
1134 * addition, for each of these queues BFQ performs or does not perform
1135 * idling depending on which choice boosts the throughput more. The
1136 * exact choice depends on the device and request pattern at
1139 * Unfortunately, false positives may occur while an interactive task
1140 * is starting (e.g., an application is being started). The
1141 * consequence is that the queues associated with the task do not
1142 * enjoy weight raising as expected. Fortunately these false positives
1143 * are very rare. They typically occur if some service happens to
1144 * start doing I/O exactly when the interactive task starts.
1146 * Turning back to the next function, it implements all the steps
1147 * needed to detect the occurrence of a large burst and to properly
1148 * mark all the queues belonging to it (so that they can then be
1149 * treated in a different way). This goal is achieved by maintaining a
1150 * "burst list" that holds, temporarily, the queues that belong to the
1151 * burst in progress. The list is then used to mark these queues as
1152 * belonging to a large burst if the burst does become large. The main
1153 * steps are the following.
1155 * . when the very first queue is created, the queue is inserted into the
1156 * list (as it could be the first queue in a possible burst)
1158 * . if the current burst has not yet become large, and a queue Q that does
1159 * not yet belong to the burst is activated shortly after the last time
1160 * at which a new queue entered the burst list, then the function appends
1161 * Q to the burst list
1163 * . if, as a consequence of the previous step, the burst size reaches
1164 * the large-burst threshold, then
1166 * . all the queues in the burst list are marked as belonging to a
1169 * . the burst list is deleted; in fact, the burst list already served
1170 * its purpose (keeping temporarily track of the queues in a burst,
1171 * so as to be able to mark them as belonging to a large burst in the
1172 * previous sub-step), and now is not needed any more
1174 * . the device enters a large-burst mode
1176 * . if a queue Q that does not belong to the burst is created while
1177 * the device is in large-burst mode and shortly after the last time
1178 * at which a queue either entered the burst list or was marked as
1179 * belonging to the current large burst, then Q is immediately marked
1180 * as belonging to a large burst.
1182 * . if a queue Q that does not belong to the burst is created a while
1183 * later, i.e., not shortly after, than the last time at which a queue
1184 * either entered the burst list or was marked as belonging to the
1185 * current large burst, then the current burst is deemed as finished and:
1187 * . the large-burst mode is reset if set
1189 * . the burst list is emptied
1191 * . Q is inserted in the burst list, as Q may be the first queue
1192 * in a possible new burst (then the burst list contains just Q
1195 static void bfq_handle_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1198 * If bfqq is already in the burst list or is part of a large
1199 * burst, or finally has just been split, then there is
1200 * nothing else to do.
1202 if (!hlist_unhashed(&bfqq->burst_list_node) ||
1203 bfq_bfqq_in_large_burst(bfqq) ||
1204 time_is_after_eq_jiffies(bfqq->split_time +
1205 msecs_to_jiffies(10)))
1209 * If bfqq's creation happens late enough, or bfqq belongs to
1210 * a different group than the burst group, then the current
1211 * burst is finished, and related data structures must be
1214 * In this respect, consider the special case where bfqq is
1215 * the very first queue created after BFQ is selected for this
1216 * device. In this case, last_ins_in_burst and
1217 * burst_parent_entity are not yet significant when we get
1218 * here. But it is easy to verify that, whether or not the
1219 * following condition is true, bfqq will end up being
1220 * inserted into the burst list. In particular the list will
1221 * happen to contain only bfqq. And this is exactly what has
1222 * to happen, as bfqq may be the first queue of the first
1225 if (time_is_before_jiffies(bfqd->last_ins_in_burst +
1226 bfqd->bfq_burst_interval) ||
1227 bfqq->entity.parent != bfqd->burst_parent_entity) {
1228 bfqd->large_burst = false;
1229 bfq_reset_burst_list(bfqd, bfqq);
1234 * If we get here, then bfqq is being activated shortly after the
1235 * last queue. So, if the current burst is also large, we can mark
1236 * bfqq as belonging to this large burst immediately.
1238 if (bfqd->large_burst) {
1239 bfq_mark_bfqq_in_large_burst(bfqq);
1244 * If we get here, then a large-burst state has not yet been
1245 * reached, but bfqq is being activated shortly after the last
1246 * queue. Then we add bfqq to the burst.
1248 bfq_add_to_burst(bfqd, bfqq);
1251 * At this point, bfqq either has been added to the current
1252 * burst or has caused the current burst to terminate and a
1253 * possible new burst to start. In particular, in the second
1254 * case, bfqq has become the first queue in the possible new
1255 * burst. In both cases last_ins_in_burst needs to be moved
1258 bfqd->last_ins_in_burst = jiffies;
1261 static int bfq_bfqq_budget_left(struct bfq_queue *bfqq)
1263 struct bfq_entity *entity = &bfqq->entity;
1265 return entity->budget - entity->service;
1269 * If enough samples have been computed, return the current max budget
1270 * stored in bfqd, which is dynamically updated according to the
1271 * estimated disk peak rate; otherwise return the default max budget
1273 static int bfq_max_budget(struct bfq_data *bfqd)
1275 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1276 return bfq_default_max_budget;
1278 return bfqd->bfq_max_budget;
1282 * Return min budget, which is a fraction of the current or default
1283 * max budget (trying with 1/32)
1285 static int bfq_min_budget(struct bfq_data *bfqd)
1287 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1288 return bfq_default_max_budget / 32;
1290 return bfqd->bfq_max_budget / 32;
1294 * The next function, invoked after the input queue bfqq switches from
1295 * idle to busy, updates the budget of bfqq. The function also tells
1296 * whether the in-service queue should be expired, by returning
1297 * true. The purpose of expiring the in-service queue is to give bfqq
1298 * the chance to possibly preempt the in-service queue, and the reason
1299 * for preempting the in-service queue is to achieve one of the two
1302 * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has
1303 * expired because it has remained idle. In particular, bfqq may have
1304 * expired for one of the following two reasons:
1306 * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling
1307 * and did not make it to issue a new request before its last
1308 * request was served;
1310 * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue
1311 * a new request before the expiration of the idling-time.
1313 * Even if bfqq has expired for one of the above reasons, the process
1314 * associated with the queue may be however issuing requests greedily,
1315 * and thus be sensitive to the bandwidth it receives (bfqq may have
1316 * remained idle for other reasons: CPU high load, bfqq not enjoying
1317 * idling, I/O throttling somewhere in the path from the process to
1318 * the I/O scheduler, ...). But if, after every expiration for one of
1319 * the above two reasons, bfqq has to wait for the service of at least
1320 * one full budget of another queue before being served again, then
1321 * bfqq is likely to get a much lower bandwidth or resource time than
1322 * its reserved ones. To address this issue, two countermeasures need
1325 * First, the budget and the timestamps of bfqq need to be updated in
1326 * a special way on bfqq reactivation: they need to be updated as if
1327 * bfqq did not remain idle and did not expire. In fact, if they are
1328 * computed as if bfqq expired and remained idle until reactivation,
1329 * then the process associated with bfqq is treated as if, instead of
1330 * being greedy, it stopped issuing requests when bfqq remained idle,
1331 * and restarts issuing requests only on this reactivation. In other
1332 * words, the scheduler does not help the process recover the "service
1333 * hole" between bfqq expiration and reactivation. As a consequence,
1334 * the process receives a lower bandwidth than its reserved one. In
1335 * contrast, to recover this hole, the budget must be updated as if
1336 * bfqq was not expired at all before this reactivation, i.e., it must
1337 * be set to the value of the remaining budget when bfqq was
1338 * expired. Along the same line, timestamps need to be assigned the
1339 * value they had the last time bfqq was selected for service, i.e.,
1340 * before last expiration. Thus timestamps need to be back-shifted
1341 * with respect to their normal computation (see [1] for more details
1342 * on this tricky aspect).
1344 * Secondly, to allow the process to recover the hole, the in-service
1345 * queue must be expired too, to give bfqq the chance to preempt it
1346 * immediately. In fact, if bfqq has to wait for a full budget of the
1347 * in-service queue to be completed, then it may become impossible to
1348 * let the process recover the hole, even if the back-shifted
1349 * timestamps of bfqq are lower than those of the in-service queue. If
1350 * this happens for most or all of the holes, then the process may not
1351 * receive its reserved bandwidth. In this respect, it is worth noting
1352 * that, being the service of outstanding requests unpreemptible, a
1353 * little fraction of the holes may however be unrecoverable, thereby
1354 * causing a little loss of bandwidth.
1356 * The last important point is detecting whether bfqq does need this
1357 * bandwidth recovery. In this respect, the next function deems the
1358 * process associated with bfqq greedy, and thus allows it to recover
1359 * the hole, if: 1) the process is waiting for the arrival of a new
1360 * request (which implies that bfqq expired for one of the above two
1361 * reasons), and 2) such a request has arrived soon. The first
1362 * condition is controlled through the flag non_blocking_wait_rq,
1363 * while the second through the flag arrived_in_time. If both
1364 * conditions hold, then the function computes the budget in the
1365 * above-described special way, and signals that the in-service queue
1366 * should be expired. Timestamp back-shifting is done later in
1367 * __bfq_activate_entity.
1369 * 2. Reduce latency. Even if timestamps are not backshifted to let
1370 * the process associated with bfqq recover a service hole, bfqq may
1371 * however happen to have, after being (re)activated, a lower finish
1372 * timestamp than the in-service queue. That is, the next budget of
1373 * bfqq may have to be completed before the one of the in-service
1374 * queue. If this is the case, then preempting the in-service queue
1375 * allows this goal to be achieved, apart from the unpreemptible,
1376 * outstanding requests mentioned above.
1378 * Unfortunately, regardless of which of the above two goals one wants
1379 * to achieve, service trees need first to be updated to know whether
1380 * the in-service queue must be preempted. To have service trees
1381 * correctly updated, the in-service queue must be expired and
1382 * rescheduled, and bfqq must be scheduled too. This is one of the
1383 * most costly operations (in future versions, the scheduling
1384 * mechanism may be re-designed in such a way to make it possible to
1385 * know whether preemption is needed without needing to update service
1386 * trees). In addition, queue preemptions almost always cause random
1387 * I/O, and thus loss of throughput. Because of these facts, the next
1388 * function adopts the following simple scheme to avoid both costly
1389 * operations and too frequent preemptions: it requests the expiration
1390 * of the in-service queue (unconditionally) only for queues that need
1391 * to recover a hole, or that either are weight-raised or deserve to
1394 static bool bfq_bfqq_update_budg_for_activation(struct bfq_data *bfqd,
1395 struct bfq_queue *bfqq,
1396 bool arrived_in_time,
1397 bool wr_or_deserves_wr)
1399 struct bfq_entity *entity = &bfqq->entity;
1402 * In the next compound condition, we check also whether there
1403 * is some budget left, because otherwise there is no point in
1404 * trying to go on serving bfqq with this same budget: bfqq
1405 * would be expired immediately after being selected for
1406 * service. This would only cause useless overhead.
1408 if (bfq_bfqq_non_blocking_wait_rq(bfqq) && arrived_in_time &&
1409 bfq_bfqq_budget_left(bfqq) > 0) {
1411 * We do not clear the flag non_blocking_wait_rq here, as
1412 * the latter is used in bfq_activate_bfqq to signal
1413 * that timestamps need to be back-shifted (and is
1414 * cleared right after).
1418 * In next assignment we rely on that either
1419 * entity->service or entity->budget are not updated
1420 * on expiration if bfqq is empty (see
1421 * __bfq_bfqq_recalc_budget). Thus both quantities
1422 * remain unchanged after such an expiration, and the
1423 * following statement therefore assigns to
1424 * entity->budget the remaining budget on such an
1427 entity->budget = min_t(unsigned long,
1428 bfq_bfqq_budget_left(bfqq),
1432 * At this point, we have used entity->service to get
1433 * the budget left (needed for updating
1434 * entity->budget). Thus we finally can, and have to,
1435 * reset entity->service. The latter must be reset
1436 * because bfqq would otherwise be charged again for
1437 * the service it has received during its previous
1440 entity->service = 0;
1446 * We can finally complete expiration, by setting service to 0.
1448 entity->service = 0;
1449 entity->budget = max_t(unsigned long, bfqq->max_budget,
1450 bfq_serv_to_charge(bfqq->next_rq, bfqq));
1451 bfq_clear_bfqq_non_blocking_wait_rq(bfqq);
1452 return wr_or_deserves_wr;
1456 * Return the farthest past time instant according to jiffies
1459 static unsigned long bfq_smallest_from_now(void)
1461 return jiffies - MAX_JIFFY_OFFSET;
1464 static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data *bfqd,
1465 struct bfq_queue *bfqq,
1466 unsigned int old_wr_coeff,
1467 bool wr_or_deserves_wr,
1472 if (old_wr_coeff == 1 && wr_or_deserves_wr) {
1473 /* start a weight-raising period */
1475 bfqq->service_from_wr = 0;
1476 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1477 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1480 * No interactive weight raising in progress
1481 * here: assign minus infinity to
1482 * wr_start_at_switch_to_srt, to make sure
1483 * that, at the end of the soft-real-time
1484 * weight raising periods that is starting
1485 * now, no interactive weight-raising period
1486 * may be wrongly considered as still in
1487 * progress (and thus actually started by
1490 bfqq->wr_start_at_switch_to_srt =
1491 bfq_smallest_from_now();
1492 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1493 BFQ_SOFTRT_WEIGHT_FACTOR;
1494 bfqq->wr_cur_max_time =
1495 bfqd->bfq_wr_rt_max_time;
1499 * If needed, further reduce budget to make sure it is
1500 * close to bfqq's backlog, so as to reduce the
1501 * scheduling-error component due to a too large
1502 * budget. Do not care about throughput consequences,
1503 * but only about latency. Finally, do not assign a
1504 * too small budget either, to avoid increasing
1505 * latency by causing too frequent expirations.
1507 bfqq->entity.budget = min_t(unsigned long,
1508 bfqq->entity.budget,
1509 2 * bfq_min_budget(bfqd));
1510 } else if (old_wr_coeff > 1) {
1511 if (interactive) { /* update wr coeff and duration */
1512 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1513 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1514 } else if (in_burst)
1518 * The application is now or still meeting the
1519 * requirements for being deemed soft rt. We
1520 * can then correctly and safely (re)charge
1521 * the weight-raising duration for the
1522 * application with the weight-raising
1523 * duration for soft rt applications.
1525 * In particular, doing this recharge now, i.e.,
1526 * before the weight-raising period for the
1527 * application finishes, reduces the probability
1528 * of the following negative scenario:
1529 * 1) the weight of a soft rt application is
1530 * raised at startup (as for any newly
1531 * created application),
1532 * 2) since the application is not interactive,
1533 * at a certain time weight-raising is
1534 * stopped for the application,
1535 * 3) at that time the application happens to
1536 * still have pending requests, and hence
1537 * is destined to not have a chance to be
1538 * deemed soft rt before these requests are
1539 * completed (see the comments to the
1540 * function bfq_bfqq_softrt_next_start()
1541 * for details on soft rt detection),
1542 * 4) these pending requests experience a high
1543 * latency because the application is not
1544 * weight-raised while they are pending.
1546 if (bfqq->wr_cur_max_time !=
1547 bfqd->bfq_wr_rt_max_time) {
1548 bfqq->wr_start_at_switch_to_srt =
1549 bfqq->last_wr_start_finish;
1551 bfqq->wr_cur_max_time =
1552 bfqd->bfq_wr_rt_max_time;
1553 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1554 BFQ_SOFTRT_WEIGHT_FACTOR;
1556 bfqq->last_wr_start_finish = jiffies;
1561 static bool bfq_bfqq_idle_for_long_time(struct bfq_data *bfqd,
1562 struct bfq_queue *bfqq)
1564 return bfqq->dispatched == 0 &&
1565 time_is_before_jiffies(
1566 bfqq->budget_timeout +
1567 bfqd->bfq_wr_min_idle_time);
1570 static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data *bfqd,
1571 struct bfq_queue *bfqq,
1576 bool soft_rt, in_burst, wr_or_deserves_wr,
1577 bfqq_wants_to_preempt,
1578 idle_for_long_time = bfq_bfqq_idle_for_long_time(bfqd, bfqq),
1580 * See the comments on
1581 * bfq_bfqq_update_budg_for_activation for
1582 * details on the usage of the next variable.
1584 arrived_in_time = ktime_get_ns() <=
1585 bfqq->ttime.last_end_request +
1586 bfqd->bfq_slice_idle * 3;
1590 * bfqq deserves to be weight-raised if:
1592 * - it does not belong to a large burst,
1593 * - it has been idle for enough time or is soft real-time,
1594 * - is linked to a bfq_io_cq (it is not shared in any sense).
1596 in_burst = bfq_bfqq_in_large_burst(bfqq);
1597 soft_rt = bfqd->bfq_wr_max_softrt_rate > 0 &&
1599 time_is_before_jiffies(bfqq->soft_rt_next_start) &&
1600 bfqq->dispatched == 0;
1601 *interactive = !in_burst && idle_for_long_time;
1602 wr_or_deserves_wr = bfqd->low_latency &&
1603 (bfqq->wr_coeff > 1 ||
1604 (bfq_bfqq_sync(bfqq) &&
1605 bfqq->bic && (*interactive || soft_rt)));
1608 * Using the last flag, update budget and check whether bfqq
1609 * may want to preempt the in-service queue.
1611 bfqq_wants_to_preempt =
1612 bfq_bfqq_update_budg_for_activation(bfqd, bfqq,
1617 * If bfqq happened to be activated in a burst, but has been
1618 * idle for much more than an interactive queue, then we
1619 * assume that, in the overall I/O initiated in the burst, the
1620 * I/O associated with bfqq is finished. So bfqq does not need
1621 * to be treated as a queue belonging to a burst
1622 * anymore. Accordingly, we reset bfqq's in_large_burst flag
1623 * if set, and remove bfqq from the burst list if it's
1624 * there. We do not decrement burst_size, because the fact
1625 * that bfqq does not need to belong to the burst list any
1626 * more does not invalidate the fact that bfqq was created in
1629 if (likely(!bfq_bfqq_just_created(bfqq)) &&
1630 idle_for_long_time &&
1631 time_is_before_jiffies(
1632 bfqq->budget_timeout +
1633 msecs_to_jiffies(10000))) {
1634 hlist_del_init(&bfqq->burst_list_node);
1635 bfq_clear_bfqq_in_large_burst(bfqq);
1638 bfq_clear_bfqq_just_created(bfqq);
1641 if (!bfq_bfqq_IO_bound(bfqq)) {
1642 if (arrived_in_time) {
1643 bfqq->requests_within_timer++;
1644 if (bfqq->requests_within_timer >=
1645 bfqd->bfq_requests_within_timer)
1646 bfq_mark_bfqq_IO_bound(bfqq);
1648 bfqq->requests_within_timer = 0;
1651 if (bfqd->low_latency) {
1652 if (unlikely(time_is_after_jiffies(bfqq->split_time)))
1655 jiffies - bfqd->bfq_wr_min_idle_time - 1;
1657 if (time_is_before_jiffies(bfqq->split_time +
1658 bfqd->bfq_wr_min_idle_time)) {
1659 bfq_update_bfqq_wr_on_rq_arrival(bfqd, bfqq,
1666 if (old_wr_coeff != bfqq->wr_coeff)
1667 bfqq->entity.prio_changed = 1;
1671 bfqq->last_idle_bklogged = jiffies;
1672 bfqq->service_from_backlogged = 0;
1673 bfq_clear_bfqq_softrt_update(bfqq);
1675 bfq_add_bfqq_busy(bfqd, bfqq);
1678 * Expire in-service queue only if preemption may be needed
1679 * for guarantees. In this respect, the function
1680 * next_queue_may_preempt just checks a simple, necessary
1681 * condition, and not a sufficient condition based on
1682 * timestamps. In fact, for the latter condition to be
1683 * evaluated, timestamps would need first to be updated, and
1684 * this operation is quite costly (see the comments on the
1685 * function bfq_bfqq_update_budg_for_activation).
1687 if (bfqd->in_service_queue && bfqq_wants_to_preempt &&
1688 bfqd->in_service_queue->wr_coeff < bfqq->wr_coeff &&
1689 next_queue_may_preempt(bfqd))
1690 bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
1691 false, BFQQE_PREEMPTED);
1694 static void bfq_add_request(struct request *rq)
1696 struct bfq_queue *bfqq = RQ_BFQQ(rq);
1697 struct bfq_data *bfqd = bfqq->bfqd;
1698 struct request *next_rq, *prev;
1699 unsigned int old_wr_coeff = bfqq->wr_coeff;
1700 bool interactive = false;
1702 bfq_log_bfqq(bfqd, bfqq, "add_request %d", rq_is_sync(rq));
1703 bfqq->queued[rq_is_sync(rq)]++;
1706 elv_rb_add(&bfqq->sort_list, rq);
1709 * Check if this request is a better next-serve candidate.
1711 prev = bfqq->next_rq;
1712 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position);
1713 bfqq->next_rq = next_rq;
1716 * Adjust priority tree position, if next_rq changes.
1718 if (prev != bfqq->next_rq)
1719 bfq_pos_tree_add_move(bfqd, bfqq);
1721 if (!bfq_bfqq_busy(bfqq)) /* switching to busy ... */
1722 bfq_bfqq_handle_idle_busy_switch(bfqd, bfqq, old_wr_coeff,
1725 if (bfqd->low_latency && old_wr_coeff == 1 && !rq_is_sync(rq) &&
1726 time_is_before_jiffies(
1727 bfqq->last_wr_start_finish +
1728 bfqd->bfq_wr_min_inter_arr_async)) {
1729 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1730 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1732 bfqd->wr_busy_queues++;
1733 bfqq->entity.prio_changed = 1;
1735 if (prev != bfqq->next_rq)
1736 bfq_updated_next_req(bfqd, bfqq);
1740 * Assign jiffies to last_wr_start_finish in the following
1743 * . if bfqq is not going to be weight-raised, because, for
1744 * non weight-raised queues, last_wr_start_finish stores the
1745 * arrival time of the last request; as of now, this piece
1746 * of information is used only for deciding whether to
1747 * weight-raise async queues
1749 * . if bfqq is not weight-raised, because, if bfqq is now
1750 * switching to weight-raised, then last_wr_start_finish
1751 * stores the time when weight-raising starts
1753 * . if bfqq is interactive, because, regardless of whether
1754 * bfqq is currently weight-raised, the weight-raising
1755 * period must start or restart (this case is considered
1756 * separately because it is not detected by the above
1757 * conditions, if bfqq is already weight-raised)
1759 * last_wr_start_finish has to be updated also if bfqq is soft
1760 * real-time, because the weight-raising period is constantly
1761 * restarted on idle-to-busy transitions for these queues, but
1762 * this is already done in bfq_bfqq_handle_idle_busy_switch if
1765 if (bfqd->low_latency &&
1766 (old_wr_coeff == 1 || bfqq->wr_coeff == 1 || interactive))
1767 bfqq->last_wr_start_finish = jiffies;
1770 static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd,
1772 struct request_queue *q)
1774 struct bfq_queue *bfqq = bfqd->bio_bfqq;
1778 return elv_rb_find(&bfqq->sort_list, bio_end_sector(bio));
1783 static sector_t get_sdist(sector_t last_pos, struct request *rq)
1786 return abs(blk_rq_pos(rq) - last_pos);
1791 #if 0 /* Still not clear if we can do without next two functions */
1792 static void bfq_activate_request(struct request_queue *q, struct request *rq)
1794 struct bfq_data *bfqd = q->elevator->elevator_data;
1796 bfqd->rq_in_driver++;
1799 static void bfq_deactivate_request(struct request_queue *q, struct request *rq)
1801 struct bfq_data *bfqd = q->elevator->elevator_data;
1803 bfqd->rq_in_driver--;
1807 static void bfq_remove_request(struct request_queue *q,
1810 struct bfq_queue *bfqq = RQ_BFQQ(rq);
1811 struct bfq_data *bfqd = bfqq->bfqd;
1812 const int sync = rq_is_sync(rq);
1814 if (bfqq->next_rq == rq) {
1815 bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq);
1816 bfq_updated_next_req(bfqd, bfqq);
1819 if (rq->queuelist.prev != &rq->queuelist)
1820 list_del_init(&rq->queuelist);
1821 bfqq->queued[sync]--;
1823 elv_rb_del(&bfqq->sort_list, rq);
1825 elv_rqhash_del(q, rq);
1826 if (q->last_merge == rq)
1827 q->last_merge = NULL;
1829 if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
1830 bfqq->next_rq = NULL;
1832 if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->in_service_queue) {
1833 bfq_del_bfqq_busy(bfqd, bfqq, false);
1835 * bfqq emptied. In normal operation, when
1836 * bfqq is empty, bfqq->entity.service and
1837 * bfqq->entity.budget must contain,
1838 * respectively, the service received and the
1839 * budget used last time bfqq emptied. These
1840 * facts do not hold in this case, as at least
1841 * this last removal occurred while bfqq is
1842 * not in service. To avoid inconsistencies,
1843 * reset both bfqq->entity.service and
1844 * bfqq->entity.budget, if bfqq has still a
1845 * process that may issue I/O requests to it.
1847 bfqq->entity.budget = bfqq->entity.service = 0;
1851 * Remove queue from request-position tree as it is empty.
1853 if (bfqq->pos_root) {
1854 rb_erase(&bfqq->pos_node, bfqq->pos_root);
1855 bfqq->pos_root = NULL;
1858 bfq_pos_tree_add_move(bfqd, bfqq);
1861 if (rq->cmd_flags & REQ_META)
1862 bfqq->meta_pending--;
1866 static bool bfq_bio_merge(struct blk_mq_hw_ctx *hctx, struct bio *bio)
1868 struct request_queue *q = hctx->queue;
1869 struct bfq_data *bfqd = q->elevator->elevator_data;
1870 struct request *free = NULL;
1872 * bfq_bic_lookup grabs the queue_lock: invoke it now and
1873 * store its return value for later use, to avoid nesting
1874 * queue_lock inside the bfqd->lock. We assume that the bic
1875 * returned by bfq_bic_lookup does not go away before
1876 * bfqd->lock is taken.
1878 struct bfq_io_cq *bic = bfq_bic_lookup(bfqd, current->io_context, q);
1881 spin_lock_irq(&bfqd->lock);
1884 bfqd->bio_bfqq = bic_to_bfqq(bic, op_is_sync(bio->bi_opf));
1886 bfqd->bio_bfqq = NULL;
1887 bfqd->bio_bic = bic;
1889 ret = blk_mq_sched_try_merge(q, bio, &free);
1892 blk_mq_free_request(free);
1893 spin_unlock_irq(&bfqd->lock);
1898 static int bfq_request_merge(struct request_queue *q, struct request **req,
1901 struct bfq_data *bfqd = q->elevator->elevator_data;
1902 struct request *__rq;
1904 __rq = bfq_find_rq_fmerge(bfqd, bio, q);
1905 if (__rq && elv_bio_merge_ok(__rq, bio)) {
1907 return ELEVATOR_FRONT_MERGE;
1910 return ELEVATOR_NO_MERGE;
1913 static struct bfq_queue *bfq_init_rq(struct request *rq);
1915 static void bfq_request_merged(struct request_queue *q, struct request *req,
1916 enum elv_merge type)
1918 if (type == ELEVATOR_FRONT_MERGE &&
1919 rb_prev(&req->rb_node) &&
1921 blk_rq_pos(container_of(rb_prev(&req->rb_node),
1922 struct request, rb_node))) {
1923 struct bfq_queue *bfqq = bfq_init_rq(req);
1924 struct bfq_data *bfqd = bfqq->bfqd;
1925 struct request *prev, *next_rq;
1927 /* Reposition request in its sort_list */
1928 elv_rb_del(&bfqq->sort_list, req);
1929 elv_rb_add(&bfqq->sort_list, req);
1931 /* Choose next request to be served for bfqq */
1932 prev = bfqq->next_rq;
1933 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, req,
1934 bfqd->last_position);
1935 bfqq->next_rq = next_rq;
1937 * If next_rq changes, update both the queue's budget to
1938 * fit the new request and the queue's position in its
1941 if (prev != bfqq->next_rq) {
1942 bfq_updated_next_req(bfqd, bfqq);
1943 bfq_pos_tree_add_move(bfqd, bfqq);
1949 * This function is called to notify the scheduler that the requests
1950 * rq and 'next' have been merged, with 'next' going away. BFQ
1951 * exploits this hook to address the following issue: if 'next' has a
1952 * fifo_time lower that rq, then the fifo_time of rq must be set to
1953 * the value of 'next', to not forget the greater age of 'next'.
1955 * NOTE: in this function we assume that rq is in a bfq_queue, basing
1956 * on that rq is picked from the hash table q->elevator->hash, which,
1957 * in its turn, is filled only with I/O requests present in
1958 * bfq_queues, while BFQ is in use for the request queue q. In fact,
1959 * the function that fills this hash table (elv_rqhash_add) is called
1960 * only by bfq_insert_request.
1962 static void bfq_requests_merged(struct request_queue *q, struct request *rq,
1963 struct request *next)
1965 struct bfq_queue *bfqq = bfq_init_rq(rq),
1966 *next_bfqq = bfq_init_rq(next);
1969 * If next and rq belong to the same bfq_queue and next is older
1970 * than rq, then reposition rq in the fifo (by substituting next
1971 * with rq). Otherwise, if next and rq belong to different
1972 * bfq_queues, never reposition rq: in fact, we would have to
1973 * reposition it with respect to next's position in its own fifo,
1974 * which would most certainly be too expensive with respect to
1977 if (bfqq == next_bfqq &&
1978 !list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1979 next->fifo_time < rq->fifo_time) {
1980 list_del_init(&rq->queuelist);
1981 list_replace_init(&next->queuelist, &rq->queuelist);
1982 rq->fifo_time = next->fifo_time;
1985 if (bfqq->next_rq == next)
1988 bfqg_stats_update_io_merged(bfqq_group(bfqq), next->cmd_flags);
1991 /* Must be called with bfqq != NULL */
1992 static void bfq_bfqq_end_wr(struct bfq_queue *bfqq)
1994 if (bfq_bfqq_busy(bfqq))
1995 bfqq->bfqd->wr_busy_queues--;
1997 bfqq->wr_cur_max_time = 0;
1998 bfqq->last_wr_start_finish = jiffies;
2000 * Trigger a weight change on the next invocation of
2001 * __bfq_entity_update_weight_prio.
2003 bfqq->entity.prio_changed = 1;
2006 void bfq_end_wr_async_queues(struct bfq_data *bfqd,
2007 struct bfq_group *bfqg)
2011 for (i = 0; i < 2; i++)
2012 for (j = 0; j < IOPRIO_BE_NR; j++)
2013 if (bfqg->async_bfqq[i][j])
2014 bfq_bfqq_end_wr(bfqg->async_bfqq[i][j]);
2015 if (bfqg->async_idle_bfqq)
2016 bfq_bfqq_end_wr(bfqg->async_idle_bfqq);
2019 static void bfq_end_wr(struct bfq_data *bfqd)
2021 struct bfq_queue *bfqq;
2023 spin_lock_irq(&bfqd->lock);
2025 list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
2026 bfq_bfqq_end_wr(bfqq);
2027 list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list)
2028 bfq_bfqq_end_wr(bfqq);
2029 bfq_end_wr_async(bfqd);
2031 spin_unlock_irq(&bfqd->lock);
2034 static sector_t bfq_io_struct_pos(void *io_struct, bool request)
2037 return blk_rq_pos(io_struct);
2039 return ((struct bio *)io_struct)->bi_iter.bi_sector;
2042 static int bfq_rq_close_to_sector(void *io_struct, bool request,
2045 return abs(bfq_io_struct_pos(io_struct, request) - sector) <=
2049 static struct bfq_queue *bfqq_find_close(struct bfq_data *bfqd,
2050 struct bfq_queue *bfqq,
2053 struct rb_root *root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree;
2054 struct rb_node *parent, *node;
2055 struct bfq_queue *__bfqq;
2057 if (RB_EMPTY_ROOT(root))
2061 * First, if we find a request starting at the end of the last
2062 * request, choose it.
2064 __bfqq = bfq_rq_pos_tree_lookup(bfqd, root, sector, &parent, NULL);
2069 * If the exact sector wasn't found, the parent of the NULL leaf
2070 * will contain the closest sector (rq_pos_tree sorted by
2071 * next_request position).
2073 __bfqq = rb_entry(parent, struct bfq_queue, pos_node);
2074 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2077 if (blk_rq_pos(__bfqq->next_rq) < sector)
2078 node = rb_next(&__bfqq->pos_node);
2080 node = rb_prev(&__bfqq->pos_node);
2084 __bfqq = rb_entry(node, struct bfq_queue, pos_node);
2085 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2091 static struct bfq_queue *bfq_find_close_cooperator(struct bfq_data *bfqd,
2092 struct bfq_queue *cur_bfqq,
2095 struct bfq_queue *bfqq;
2098 * We shall notice if some of the queues are cooperating,
2099 * e.g., working closely on the same area of the device. In
2100 * that case, we can group them together and: 1) don't waste
2101 * time idling, and 2) serve the union of their requests in
2102 * the best possible order for throughput.
2104 bfqq = bfqq_find_close(bfqd, cur_bfqq, sector);
2105 if (!bfqq || bfqq == cur_bfqq)
2111 static struct bfq_queue *
2112 bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2114 int process_refs, new_process_refs;
2115 struct bfq_queue *__bfqq;
2118 * If there are no process references on the new_bfqq, then it is
2119 * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain
2120 * may have dropped their last reference (not just their last process
2123 if (!bfqq_process_refs(new_bfqq))
2126 /* Avoid a circular list and skip interim queue merges. */
2127 while ((__bfqq = new_bfqq->new_bfqq)) {
2133 process_refs = bfqq_process_refs(bfqq);
2134 new_process_refs = bfqq_process_refs(new_bfqq);
2136 * If the process for the bfqq has gone away, there is no
2137 * sense in merging the queues.
2139 if (process_refs == 0 || new_process_refs == 0)
2142 bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d",
2146 * Merging is just a redirection: the requests of the process
2147 * owning one of the two queues are redirected to the other queue.
2148 * The latter queue, in its turn, is set as shared if this is the
2149 * first time that the requests of some process are redirected to
2152 * We redirect bfqq to new_bfqq and not the opposite, because
2153 * we are in the context of the process owning bfqq, thus we
2154 * have the io_cq of this process. So we can immediately
2155 * configure this io_cq to redirect the requests of the
2156 * process to new_bfqq. In contrast, the io_cq of new_bfqq is
2157 * not available any more (new_bfqq->bic == NULL).
2159 * Anyway, even in case new_bfqq coincides with the in-service
2160 * queue, redirecting requests the in-service queue is the
2161 * best option, as we feed the in-service queue with new
2162 * requests close to the last request served and, by doing so,
2163 * are likely to increase the throughput.
2165 bfqq->new_bfqq = new_bfqq;
2166 new_bfqq->ref += process_refs;
2170 static bool bfq_may_be_close_cooperator(struct bfq_queue *bfqq,
2171 struct bfq_queue *new_bfqq)
2173 if (bfq_too_late_for_merging(new_bfqq))
2176 if (bfq_class_idle(bfqq) || bfq_class_idle(new_bfqq) ||
2177 (bfqq->ioprio_class != new_bfqq->ioprio_class))
2181 * If either of the queues has already been detected as seeky,
2182 * then merging it with the other queue is unlikely to lead to
2185 if (BFQQ_SEEKY(bfqq) || BFQQ_SEEKY(new_bfqq))
2189 * Interleaved I/O is known to be done by (some) applications
2190 * only for reads, so it does not make sense to merge async
2193 if (!bfq_bfqq_sync(bfqq) || !bfq_bfqq_sync(new_bfqq))
2200 * Attempt to schedule a merge of bfqq with the currently in-service
2201 * queue or with a close queue among the scheduled queues. Return
2202 * NULL if no merge was scheduled, a pointer to the shared bfq_queue
2203 * structure otherwise.
2205 * The OOM queue is not allowed to participate to cooperation: in fact, since
2206 * the requests temporarily redirected to the OOM queue could be redirected
2207 * again to dedicated queues at any time, the state needed to correctly
2208 * handle merging with the OOM queue would be quite complex and expensive
2209 * to maintain. Besides, in such a critical condition as an out of memory,
2210 * the benefits of queue merging may be little relevant, or even negligible.
2212 * WARNING: queue merging may impair fairness among non-weight raised
2213 * queues, for at least two reasons: 1) the original weight of a
2214 * merged queue may change during the merged state, 2) even being the
2215 * weight the same, a merged queue may be bloated with many more
2216 * requests than the ones produced by its originally-associated
2219 static struct bfq_queue *
2220 bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2221 void *io_struct, bool request)
2223 struct bfq_queue *in_service_bfqq, *new_bfqq;
2226 * Prevent bfqq from being merged if it has been created too
2227 * long ago. The idea is that true cooperating processes, and
2228 * thus their associated bfq_queues, are supposed to be
2229 * created shortly after each other. This is the case, e.g.,
2230 * for KVM/QEMU and dump I/O threads. Basing on this
2231 * assumption, the following filtering greatly reduces the
2232 * probability that two non-cooperating processes, which just
2233 * happen to do close I/O for some short time interval, have
2234 * their queues merged by mistake.
2236 if (bfq_too_late_for_merging(bfqq))
2240 return bfqq->new_bfqq;
2242 if (!io_struct || unlikely(bfqq == &bfqd->oom_bfqq))
2245 /* If there is only one backlogged queue, don't search. */
2246 if (bfq_tot_busy_queues(bfqd) == 1)
2249 in_service_bfqq = bfqd->in_service_queue;
2251 if (in_service_bfqq && in_service_bfqq != bfqq &&
2252 likely(in_service_bfqq != &bfqd->oom_bfqq) &&
2253 bfq_rq_close_to_sector(io_struct, request, bfqd->last_position) &&
2254 bfqq->entity.parent == in_service_bfqq->entity.parent &&
2255 bfq_may_be_close_cooperator(bfqq, in_service_bfqq)) {
2256 new_bfqq = bfq_setup_merge(bfqq, in_service_bfqq);
2261 * Check whether there is a cooperator among currently scheduled
2262 * queues. The only thing we need is that the bio/request is not
2263 * NULL, as we need it to establish whether a cooperator exists.
2265 new_bfqq = bfq_find_close_cooperator(bfqd, bfqq,
2266 bfq_io_struct_pos(io_struct, request));
2268 if (new_bfqq && likely(new_bfqq != &bfqd->oom_bfqq) &&
2269 bfq_may_be_close_cooperator(bfqq, new_bfqq))
2270 return bfq_setup_merge(bfqq, new_bfqq);
2275 static void bfq_bfqq_save_state(struct bfq_queue *bfqq)
2277 struct bfq_io_cq *bic = bfqq->bic;
2280 * If !bfqq->bic, the queue is already shared or its requests
2281 * have already been redirected to a shared queue; both idle window
2282 * and weight raising state have already been saved. Do nothing.
2287 bic->saved_ttime = bfqq->ttime;
2288 bic->saved_has_short_ttime = bfq_bfqq_has_short_ttime(bfqq);
2289 bic->saved_IO_bound = bfq_bfqq_IO_bound(bfqq);
2290 bic->saved_in_large_burst = bfq_bfqq_in_large_burst(bfqq);
2291 bic->was_in_burst_list = !hlist_unhashed(&bfqq->burst_list_node);
2292 if (unlikely(bfq_bfqq_just_created(bfqq) &&
2293 !bfq_bfqq_in_large_burst(bfqq) &&
2294 bfqq->bfqd->low_latency)) {
2296 * bfqq being merged right after being created: bfqq
2297 * would have deserved interactive weight raising, but
2298 * did not make it to be set in a weight-raised state,
2299 * because of this early merge. Store directly the
2300 * weight-raising state that would have been assigned
2301 * to bfqq, so that to avoid that bfqq unjustly fails
2302 * to enjoy weight raising if split soon.
2304 bic->saved_wr_coeff = bfqq->bfqd->bfq_wr_coeff;
2305 bic->saved_wr_cur_max_time = bfq_wr_duration(bfqq->bfqd);
2306 bic->saved_last_wr_start_finish = jiffies;
2308 bic->saved_wr_coeff = bfqq->wr_coeff;
2309 bic->saved_wr_start_at_switch_to_srt =
2310 bfqq->wr_start_at_switch_to_srt;
2311 bic->saved_last_wr_start_finish = bfqq->last_wr_start_finish;
2312 bic->saved_wr_cur_max_time = bfqq->wr_cur_max_time;
2317 bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic,
2318 struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2320 bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu",
2321 (unsigned long)new_bfqq->pid);
2322 /* Save weight raising and idle window of the merged queues */
2323 bfq_bfqq_save_state(bfqq);
2324 bfq_bfqq_save_state(new_bfqq);
2325 if (bfq_bfqq_IO_bound(bfqq))
2326 bfq_mark_bfqq_IO_bound(new_bfqq);
2327 bfq_clear_bfqq_IO_bound(bfqq);
2330 * If bfqq is weight-raised, then let new_bfqq inherit
2331 * weight-raising. To reduce false positives, neglect the case
2332 * where bfqq has just been created, but has not yet made it
2333 * to be weight-raised (which may happen because EQM may merge
2334 * bfqq even before bfq_add_request is executed for the first
2335 * time for bfqq). Handling this case would however be very
2336 * easy, thanks to the flag just_created.
2338 if (new_bfqq->wr_coeff == 1 && bfqq->wr_coeff > 1) {
2339 new_bfqq->wr_coeff = bfqq->wr_coeff;
2340 new_bfqq->wr_cur_max_time = bfqq->wr_cur_max_time;
2341 new_bfqq->last_wr_start_finish = bfqq->last_wr_start_finish;
2342 new_bfqq->wr_start_at_switch_to_srt =
2343 bfqq->wr_start_at_switch_to_srt;
2344 if (bfq_bfqq_busy(new_bfqq))
2345 bfqd->wr_busy_queues++;
2346 new_bfqq->entity.prio_changed = 1;
2349 if (bfqq->wr_coeff > 1) { /* bfqq has given its wr to new_bfqq */
2351 bfqq->entity.prio_changed = 1;
2352 if (bfq_bfqq_busy(bfqq))
2353 bfqd->wr_busy_queues--;
2356 bfq_log_bfqq(bfqd, new_bfqq, "merge_bfqqs: wr_busy %d",
2357 bfqd->wr_busy_queues);
2360 * Merge queues (that is, let bic redirect its requests to new_bfqq)
2362 bic_set_bfqq(bic, new_bfqq, 1);
2363 bfq_mark_bfqq_coop(new_bfqq);
2365 * new_bfqq now belongs to at least two bics (it is a shared queue):
2366 * set new_bfqq->bic to NULL. bfqq either:
2367 * - does not belong to any bic any more, and hence bfqq->bic must
2368 * be set to NULL, or
2369 * - is a queue whose owning bics have already been redirected to a
2370 * different queue, hence the queue is destined to not belong to
2371 * any bic soon and bfqq->bic is already NULL (therefore the next
2372 * assignment causes no harm).
2374 new_bfqq->bic = NULL;
2376 /* release process reference to bfqq */
2377 bfq_put_queue(bfqq);
2380 static bool bfq_allow_bio_merge(struct request_queue *q, struct request *rq,
2383 struct bfq_data *bfqd = q->elevator->elevator_data;
2384 bool is_sync = op_is_sync(bio->bi_opf);
2385 struct bfq_queue *bfqq = bfqd->bio_bfqq, *new_bfqq;
2388 * Disallow merge of a sync bio into an async request.
2390 if (is_sync && !rq_is_sync(rq))
2394 * Lookup the bfqq that this bio will be queued with. Allow
2395 * merge only if rq is queued there.
2401 * We take advantage of this function to perform an early merge
2402 * of the queues of possible cooperating processes.
2404 new_bfqq = bfq_setup_cooperator(bfqd, bfqq, bio, false);
2407 * bic still points to bfqq, then it has not yet been
2408 * redirected to some other bfq_queue, and a queue
2409 * merge beween bfqq and new_bfqq can be safely
2410 * fulfillled, i.e., bic can be redirected to new_bfqq
2411 * and bfqq can be put.
2413 bfq_merge_bfqqs(bfqd, bfqd->bio_bic, bfqq,
2416 * If we get here, bio will be queued into new_queue,
2417 * so use new_bfqq to decide whether bio and rq can be
2423 * Change also bqfd->bio_bfqq, as
2424 * bfqd->bio_bic now points to new_bfqq, and
2425 * this function may be invoked again (and then may
2426 * use again bqfd->bio_bfqq).
2428 bfqd->bio_bfqq = bfqq;
2431 return bfqq == RQ_BFQQ(rq);
2435 * Set the maximum time for the in-service queue to consume its
2436 * budget. This prevents seeky processes from lowering the throughput.
2437 * In practice, a time-slice service scheme is used with seeky
2440 static void bfq_set_budget_timeout(struct bfq_data *bfqd,
2441 struct bfq_queue *bfqq)
2443 unsigned int timeout_coeff;
2445 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time)
2448 timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight;
2450 bfqd->last_budget_start = ktime_get();
2452 bfqq->budget_timeout = jiffies +
2453 bfqd->bfq_timeout * timeout_coeff;
2456 static void __bfq_set_in_service_queue(struct bfq_data *bfqd,
2457 struct bfq_queue *bfqq)
2460 bfq_clear_bfqq_fifo_expire(bfqq);
2462 bfqd->budgets_assigned = (bfqd->budgets_assigned * 7 + 256) / 8;
2464 if (time_is_before_jiffies(bfqq->last_wr_start_finish) &&
2465 bfqq->wr_coeff > 1 &&
2466 bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
2467 time_is_before_jiffies(bfqq->budget_timeout)) {
2469 * For soft real-time queues, move the start
2470 * of the weight-raising period forward by the
2471 * time the queue has not received any
2472 * service. Otherwise, a relatively long
2473 * service delay is likely to cause the
2474 * weight-raising period of the queue to end,
2475 * because of the short duration of the
2476 * weight-raising period of a soft real-time
2477 * queue. It is worth noting that this move
2478 * is not so dangerous for the other queues,
2479 * because soft real-time queues are not
2482 * To not add a further variable, we use the
2483 * overloaded field budget_timeout to
2484 * determine for how long the queue has not
2485 * received service, i.e., how much time has
2486 * elapsed since the queue expired. However,
2487 * this is a little imprecise, because
2488 * budget_timeout is set to jiffies if bfqq
2489 * not only expires, but also remains with no
2492 if (time_after(bfqq->budget_timeout,
2493 bfqq->last_wr_start_finish))
2494 bfqq->last_wr_start_finish +=
2495 jiffies - bfqq->budget_timeout;
2497 bfqq->last_wr_start_finish = jiffies;
2500 bfq_set_budget_timeout(bfqd, bfqq);
2501 bfq_log_bfqq(bfqd, bfqq,
2502 "set_in_service_queue, cur-budget = %d",
2503 bfqq->entity.budget);
2506 bfqd->in_service_queue = bfqq;
2510 * Get and set a new queue for service.
2512 static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd)
2514 struct bfq_queue *bfqq = bfq_get_next_queue(bfqd);
2516 __bfq_set_in_service_queue(bfqd, bfqq);
2520 static void bfq_arm_slice_timer(struct bfq_data *bfqd)
2522 struct bfq_queue *bfqq = bfqd->in_service_queue;
2525 bfq_mark_bfqq_wait_request(bfqq);
2528 * We don't want to idle for seeks, but we do want to allow
2529 * fair distribution of slice time for a process doing back-to-back
2530 * seeks. So allow a little bit of time for him to submit a new rq.
2532 sl = bfqd->bfq_slice_idle;
2534 * Unless the queue is being weight-raised or the scenario is
2535 * asymmetric, grant only minimum idle time if the queue
2536 * is seeky. A long idling is preserved for a weight-raised
2537 * queue, or, more in general, in an asymmetric scenario,
2538 * because a long idling is needed for guaranteeing to a queue
2539 * its reserved share of the throughput (in particular, it is
2540 * needed if the queue has a higher weight than some other
2543 if (BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 &&
2544 bfq_symmetric_scenario(bfqd))
2545 sl = min_t(u64, sl, BFQ_MIN_TT);
2547 bfqd->last_idling_start = ktime_get();
2548 hrtimer_start(&bfqd->idle_slice_timer, ns_to_ktime(sl),
2550 bfqg_stats_set_start_idle_time(bfqq_group(bfqq));
2554 * In autotuning mode, max_budget is dynamically recomputed as the
2555 * amount of sectors transferred in timeout at the estimated peak
2556 * rate. This enables BFQ to utilize a full timeslice with a full
2557 * budget, even if the in-service queue is served at peak rate. And
2558 * this maximises throughput with sequential workloads.
2560 static unsigned long bfq_calc_max_budget(struct bfq_data *bfqd)
2562 return (u64)bfqd->peak_rate * USEC_PER_MSEC *
2563 jiffies_to_msecs(bfqd->bfq_timeout)>>BFQ_RATE_SHIFT;
2567 * Update parameters related to throughput and responsiveness, as a
2568 * function of the estimated peak rate. See comments on
2569 * bfq_calc_max_budget(), and on the ref_wr_duration array.
2571 static void update_thr_responsiveness_params(struct bfq_data *bfqd)
2573 if (bfqd->bfq_user_max_budget == 0) {
2574 bfqd->bfq_max_budget =
2575 bfq_calc_max_budget(bfqd);
2576 bfq_log(bfqd, "new max_budget = %d", bfqd->bfq_max_budget);
2580 static void bfq_reset_rate_computation(struct bfq_data *bfqd,
2583 if (rq != NULL) { /* new rq dispatch now, reset accordingly */
2584 bfqd->last_dispatch = bfqd->first_dispatch = ktime_get_ns();
2585 bfqd->peak_rate_samples = 1;
2586 bfqd->sequential_samples = 0;
2587 bfqd->tot_sectors_dispatched = bfqd->last_rq_max_size =
2589 } else /* no new rq dispatched, just reset the number of samples */
2590 bfqd->peak_rate_samples = 0; /* full re-init on next disp. */
2593 "reset_rate_computation at end, sample %u/%u tot_sects %llu",
2594 bfqd->peak_rate_samples, bfqd->sequential_samples,
2595 bfqd->tot_sectors_dispatched);
2598 static void bfq_update_rate_reset(struct bfq_data *bfqd, struct request *rq)
2600 u32 rate, weight, divisor;
2603 * For the convergence property to hold (see comments on
2604 * bfq_update_peak_rate()) and for the assessment to be
2605 * reliable, a minimum number of samples must be present, and
2606 * a minimum amount of time must have elapsed. If not so, do
2607 * not compute new rate. Just reset parameters, to get ready
2608 * for a new evaluation attempt.
2610 if (bfqd->peak_rate_samples < BFQ_RATE_MIN_SAMPLES ||
2611 bfqd->delta_from_first < BFQ_RATE_MIN_INTERVAL)
2612 goto reset_computation;
2615 * If a new request completion has occurred after last
2616 * dispatch, then, to approximate the rate at which requests
2617 * have been served by the device, it is more precise to
2618 * extend the observation interval to the last completion.
2620 bfqd->delta_from_first =
2621 max_t(u64, bfqd->delta_from_first,
2622 bfqd->last_completion - bfqd->first_dispatch);
2625 * Rate computed in sects/usec, and not sects/nsec, for
2628 rate = div64_ul(bfqd->tot_sectors_dispatched<<BFQ_RATE_SHIFT,
2629 div_u64(bfqd->delta_from_first, NSEC_PER_USEC));
2632 * Peak rate not updated if:
2633 * - the percentage of sequential dispatches is below 3/4 of the
2634 * total, and rate is below the current estimated peak rate
2635 * - rate is unreasonably high (> 20M sectors/sec)
2637 if ((bfqd->sequential_samples < (3 * bfqd->peak_rate_samples)>>2 &&
2638 rate <= bfqd->peak_rate) ||
2639 rate > 20<<BFQ_RATE_SHIFT)
2640 goto reset_computation;
2643 * We have to update the peak rate, at last! To this purpose,
2644 * we use a low-pass filter. We compute the smoothing constant
2645 * of the filter as a function of the 'weight' of the new
2648 * As can be seen in next formulas, we define this weight as a
2649 * quantity proportional to how sequential the workload is,
2650 * and to how long the observation time interval is.
2652 * The weight runs from 0 to 8. The maximum value of the
2653 * weight, 8, yields the minimum value for the smoothing
2654 * constant. At this minimum value for the smoothing constant,
2655 * the measured rate contributes for half of the next value of
2656 * the estimated peak rate.
2658 * So, the first step is to compute the weight as a function
2659 * of how sequential the workload is. Note that the weight
2660 * cannot reach 9, because bfqd->sequential_samples cannot
2661 * become equal to bfqd->peak_rate_samples, which, in its
2662 * turn, holds true because bfqd->sequential_samples is not
2663 * incremented for the first sample.
2665 weight = (9 * bfqd->sequential_samples) / bfqd->peak_rate_samples;
2668 * Second step: further refine the weight as a function of the
2669 * duration of the observation interval.
2671 weight = min_t(u32, 8,
2672 div_u64(weight * bfqd->delta_from_first,
2673 BFQ_RATE_REF_INTERVAL));
2676 * Divisor ranging from 10, for minimum weight, to 2, for
2679 divisor = 10 - weight;
2682 * Finally, update peak rate:
2684 * peak_rate = peak_rate * (divisor-1) / divisor + rate / divisor
2686 bfqd->peak_rate *= divisor-1;
2687 bfqd->peak_rate /= divisor;
2688 rate /= divisor; /* smoothing constant alpha = 1/divisor */
2690 bfqd->peak_rate += rate;
2693 * For a very slow device, bfqd->peak_rate can reach 0 (see
2694 * the minimum representable values reported in the comments
2695 * on BFQ_RATE_SHIFT). Push to 1 if this happens, to avoid
2696 * divisions by zero where bfqd->peak_rate is used as a
2699 bfqd->peak_rate = max_t(u32, 1, bfqd->peak_rate);
2701 update_thr_responsiveness_params(bfqd);
2704 bfq_reset_rate_computation(bfqd, rq);
2708 * Update the read/write peak rate (the main quantity used for
2709 * auto-tuning, see update_thr_responsiveness_params()).
2711 * It is not trivial to estimate the peak rate (correctly): because of
2712 * the presence of sw and hw queues between the scheduler and the
2713 * device components that finally serve I/O requests, it is hard to
2714 * say exactly when a given dispatched request is served inside the
2715 * device, and for how long. As a consequence, it is hard to know
2716 * precisely at what rate a given set of requests is actually served
2719 * On the opposite end, the dispatch time of any request is trivially
2720 * available, and, from this piece of information, the "dispatch rate"
2721 * of requests can be immediately computed. So, the idea in the next
2722 * function is to use what is known, namely request dispatch times
2723 * (plus, when useful, request completion times), to estimate what is
2724 * unknown, namely in-device request service rate.
2726 * The main issue is that, because of the above facts, the rate at
2727 * which a certain set of requests is dispatched over a certain time
2728 * interval can vary greatly with respect to the rate at which the
2729 * same requests are then served. But, since the size of any
2730 * intermediate queue is limited, and the service scheme is lossless
2731 * (no request is silently dropped), the following obvious convergence
2732 * property holds: the number of requests dispatched MUST become
2733 * closer and closer to the number of requests completed as the
2734 * observation interval grows. This is the key property used in
2735 * the next function to estimate the peak service rate as a function
2736 * of the observed dispatch rate. The function assumes to be invoked
2737 * on every request dispatch.
2739 static void bfq_update_peak_rate(struct bfq_data *bfqd, struct request *rq)
2741 u64 now_ns = ktime_get_ns();
2743 if (bfqd->peak_rate_samples == 0) { /* first dispatch */
2744 bfq_log(bfqd, "update_peak_rate: goto reset, samples %d",
2745 bfqd->peak_rate_samples);
2746 bfq_reset_rate_computation(bfqd, rq);
2747 goto update_last_values; /* will add one sample */
2751 * Device idle for very long: the observation interval lasting
2752 * up to this dispatch cannot be a valid observation interval
2753 * for computing a new peak rate (similarly to the late-
2754 * completion event in bfq_completed_request()). Go to
2755 * update_rate_and_reset to have the following three steps
2757 * - close the observation interval at the last (previous)
2758 * request dispatch or completion
2759 * - compute rate, if possible, for that observation interval
2760 * - start a new observation interval with this dispatch
2762 if (now_ns - bfqd->last_dispatch > 100*NSEC_PER_MSEC &&
2763 bfqd->rq_in_driver == 0)
2764 goto update_rate_and_reset;
2766 /* Update sampling information */
2767 bfqd->peak_rate_samples++;
2769 if ((bfqd->rq_in_driver > 0 ||
2770 now_ns - bfqd->last_completion < BFQ_MIN_TT)
2771 && !BFQ_RQ_SEEKY(bfqd, bfqd->last_position, rq))
2772 bfqd->sequential_samples++;
2774 bfqd->tot_sectors_dispatched += blk_rq_sectors(rq);
2776 /* Reset max observed rq size every 32 dispatches */
2777 if (likely(bfqd->peak_rate_samples % 32))
2778 bfqd->last_rq_max_size =
2779 max_t(u32, blk_rq_sectors(rq), bfqd->last_rq_max_size);
2781 bfqd->last_rq_max_size = blk_rq_sectors(rq);
2783 bfqd->delta_from_first = now_ns - bfqd->first_dispatch;
2785 /* Target observation interval not yet reached, go on sampling */
2786 if (bfqd->delta_from_first < BFQ_RATE_REF_INTERVAL)
2787 goto update_last_values;
2789 update_rate_and_reset:
2790 bfq_update_rate_reset(bfqd, rq);
2792 bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2793 bfqd->last_dispatch = now_ns;
2797 * Remove request from internal lists.
2799 static void bfq_dispatch_remove(struct request_queue *q, struct request *rq)
2801 struct bfq_queue *bfqq = RQ_BFQQ(rq);
2804 * For consistency, the next instruction should have been
2805 * executed after removing the request from the queue and
2806 * dispatching it. We execute instead this instruction before
2807 * bfq_remove_request() (and hence introduce a temporary
2808 * inconsistency), for efficiency. In fact, should this
2809 * dispatch occur for a non in-service bfqq, this anticipated
2810 * increment prevents two counters related to bfqq->dispatched
2811 * from risking to be, first, uselessly decremented, and then
2812 * incremented again when the (new) value of bfqq->dispatched
2813 * happens to be taken into account.
2816 bfq_update_peak_rate(q->elevator->elevator_data, rq);
2818 bfq_remove_request(q, rq);
2821 static void __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq)
2824 * If this bfqq is shared between multiple processes, check
2825 * to make sure that those processes are still issuing I/Os
2826 * within the mean seek distance. If not, it may be time to
2827 * break the queues apart again.
2829 if (bfq_bfqq_coop(bfqq) && BFQQ_SEEKY(bfqq))
2830 bfq_mark_bfqq_split_coop(bfqq);
2832 if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
2833 if (bfqq->dispatched == 0)
2835 * Overloading budget_timeout field to store
2836 * the time at which the queue remains with no
2837 * backlog and no outstanding request; used by
2838 * the weight-raising mechanism.
2840 bfqq->budget_timeout = jiffies;
2842 bfq_del_bfqq_busy(bfqd, bfqq, true);
2844 bfq_requeue_bfqq(bfqd, bfqq, true);
2846 * Resort priority tree of potential close cooperators.
2848 bfq_pos_tree_add_move(bfqd, bfqq);
2852 * All in-service entities must have been properly deactivated
2853 * or requeued before executing the next function, which
2854 * resets all in-service entites as no more in service.
2856 __bfq_bfqd_reset_in_service(bfqd);
2860 * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior.
2861 * @bfqd: device data.
2862 * @bfqq: queue to update.
2863 * @reason: reason for expiration.
2865 * Handle the feedback on @bfqq budget at queue expiration.
2866 * See the body for detailed comments.
2868 static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd,
2869 struct bfq_queue *bfqq,
2870 enum bfqq_expiration reason)
2872 struct request *next_rq;
2873 int budget, min_budget;
2875 min_budget = bfq_min_budget(bfqd);
2877 if (bfqq->wr_coeff == 1)
2878 budget = bfqq->max_budget;
2880 * Use a constant, low budget for weight-raised queues,
2881 * to help achieve a low latency. Keep it slightly higher
2882 * than the minimum possible budget, to cause a little
2883 * bit fewer expirations.
2885 budget = 2 * min_budget;
2887 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %d, budg left %d",
2888 bfqq->entity.budget, bfq_bfqq_budget_left(bfqq));
2889 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %d, min budg %d",
2890 budget, bfq_min_budget(bfqd));
2891 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d",
2892 bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->in_service_queue));
2894 if (bfq_bfqq_sync(bfqq) && bfqq->wr_coeff == 1) {
2897 * Caveat: in all the following cases we trade latency
2900 case BFQQE_TOO_IDLE:
2902 * This is the only case where we may reduce
2903 * the budget: if there is no request of the
2904 * process still waiting for completion, then
2905 * we assume (tentatively) that the timer has
2906 * expired because the batch of requests of
2907 * the process could have been served with a
2908 * smaller budget. Hence, betting that
2909 * process will behave in the same way when it
2910 * becomes backlogged again, we reduce its
2911 * next budget. As long as we guess right,
2912 * this budget cut reduces the latency
2913 * experienced by the process.
2915 * However, if there are still outstanding
2916 * requests, then the process may have not yet
2917 * issued its next request just because it is
2918 * still waiting for the completion of some of
2919 * the still outstanding ones. So in this
2920 * subcase we do not reduce its budget, on the
2921 * contrary we increase it to possibly boost
2922 * the throughput, as discussed in the
2923 * comments to the BUDGET_TIMEOUT case.
2925 if (bfqq->dispatched > 0) /* still outstanding reqs */
2926 budget = min(budget * 2, bfqd->bfq_max_budget);
2928 if (budget > 5 * min_budget)
2929 budget -= 4 * min_budget;
2931 budget = min_budget;
2934 case BFQQE_BUDGET_TIMEOUT:
2936 * We double the budget here because it gives
2937 * the chance to boost the throughput if this
2938 * is not a seeky process (and has bumped into
2939 * this timeout because of, e.g., ZBR).
2941 budget = min(budget * 2, bfqd->bfq_max_budget);
2943 case BFQQE_BUDGET_EXHAUSTED:
2945 * The process still has backlog, and did not
2946 * let either the budget timeout or the disk
2947 * idling timeout expire. Hence it is not
2948 * seeky, has a short thinktime and may be
2949 * happy with a higher budget too. So
2950 * definitely increase the budget of this good
2951 * candidate to boost the disk throughput.
2953 budget = min(budget * 4, bfqd->bfq_max_budget);
2955 case BFQQE_NO_MORE_REQUESTS:
2957 * For queues that expire for this reason, it
2958 * is particularly important to keep the
2959 * budget close to the actual service they
2960 * need. Doing so reduces the timestamp
2961 * misalignment problem described in the
2962 * comments in the body of
2963 * __bfq_activate_entity. In fact, suppose
2964 * that a queue systematically expires for
2965 * BFQQE_NO_MORE_REQUESTS and presents a
2966 * new request in time to enjoy timestamp
2967 * back-shifting. The larger the budget of the
2968 * queue is with respect to the service the
2969 * queue actually requests in each service
2970 * slot, the more times the queue can be
2971 * reactivated with the same virtual finish
2972 * time. It follows that, even if this finish
2973 * time is pushed to the system virtual time
2974 * to reduce the consequent timestamp
2975 * misalignment, the queue unjustly enjoys for
2976 * many re-activations a lower finish time
2977 * than all newly activated queues.
2979 * The service needed by bfqq is measured
2980 * quite precisely by bfqq->entity.service.
2981 * Since bfqq does not enjoy device idling,
2982 * bfqq->entity.service is equal to the number
2983 * of sectors that the process associated with
2984 * bfqq requested to read/write before waiting
2985 * for request completions, or blocking for
2988 budget = max_t(int, bfqq->entity.service, min_budget);
2993 } else if (!bfq_bfqq_sync(bfqq)) {
2995 * Async queues get always the maximum possible
2996 * budget, as for them we do not care about latency
2997 * (in addition, their ability to dispatch is limited
2998 * by the charging factor).
3000 budget = bfqd->bfq_max_budget;
3003 bfqq->max_budget = budget;
3005 if (bfqd->budgets_assigned >= bfq_stats_min_budgets &&
3006 !bfqd->bfq_user_max_budget)
3007 bfqq->max_budget = min(bfqq->max_budget, bfqd->bfq_max_budget);
3010 * If there is still backlog, then assign a new budget, making
3011 * sure that it is large enough for the next request. Since
3012 * the finish time of bfqq must be kept in sync with the
3013 * budget, be sure to call __bfq_bfqq_expire() *after* this
3016 * If there is no backlog, then no need to update the budget;
3017 * it will be updated on the arrival of a new request.
3019 next_rq = bfqq->next_rq;
3021 bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget,
3022 bfq_serv_to_charge(next_rq, bfqq));
3024 bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %d",
3025 next_rq ? blk_rq_sectors(next_rq) : 0,
3026 bfqq->entity.budget);
3030 * Return true if the process associated with bfqq is "slow". The slow
3031 * flag is used, in addition to the budget timeout, to reduce the
3032 * amount of service provided to seeky processes, and thus reduce
3033 * their chances to lower the throughput. More details in the comments
3034 * on the function bfq_bfqq_expire().
3036 * An important observation is in order: as discussed in the comments
3037 * on the function bfq_update_peak_rate(), with devices with internal
3038 * queues, it is hard if ever possible to know when and for how long
3039 * an I/O request is processed by the device (apart from the trivial
3040 * I/O pattern where a new request is dispatched only after the
3041 * previous one has been completed). This makes it hard to evaluate
3042 * the real rate at which the I/O requests of each bfq_queue are
3043 * served. In fact, for an I/O scheduler like BFQ, serving a
3044 * bfq_queue means just dispatching its requests during its service
3045 * slot (i.e., until the budget of the queue is exhausted, or the
3046 * queue remains idle, or, finally, a timeout fires). But, during the
3047 * service slot of a bfq_queue, around 100 ms at most, the device may
3048 * be even still processing requests of bfq_queues served in previous
3049 * service slots. On the opposite end, the requests of the in-service
3050 * bfq_queue may be completed after the service slot of the queue
3053 * Anyway, unless more sophisticated solutions are used
3054 * (where possible), the sum of the sizes of the requests dispatched
3055 * during the service slot of a bfq_queue is probably the only
3056 * approximation available for the service received by the bfq_queue
3057 * during its service slot. And this sum is the quantity used in this
3058 * function to evaluate the I/O speed of a process.
3060 static bool bfq_bfqq_is_slow(struct bfq_data *bfqd, struct bfq_queue *bfqq,
3061 bool compensate, enum bfqq_expiration reason,
3062 unsigned long *delta_ms)
3064 ktime_t delta_ktime;
3066 bool slow = BFQQ_SEEKY(bfqq); /* if delta too short, use seekyness */
3068 if (!bfq_bfqq_sync(bfqq))
3072 delta_ktime = bfqd->last_idling_start;
3074 delta_ktime = ktime_get();
3075 delta_ktime = ktime_sub(delta_ktime, bfqd->last_budget_start);
3076 delta_usecs = ktime_to_us(delta_ktime);
3078 /* don't use too short time intervals */
3079 if (delta_usecs < 1000) {
3080 if (blk_queue_nonrot(bfqd->queue))
3082 * give same worst-case guarantees as idling
3085 *delta_ms = BFQ_MIN_TT / NSEC_PER_MSEC;
3086 else /* charge at least one seek */
3087 *delta_ms = bfq_slice_idle / NSEC_PER_MSEC;
3092 *delta_ms = delta_usecs / USEC_PER_MSEC;
3095 * Use only long (> 20ms) intervals to filter out excessive
3096 * spikes in service rate estimation.
3098 if (delta_usecs > 20000) {
3100 * Caveat for rotational devices: processes doing I/O
3101 * in the slower disk zones tend to be slow(er) even
3102 * if not seeky. In this respect, the estimated peak
3103 * rate is likely to be an average over the disk
3104 * surface. Accordingly, to not be too harsh with
3105 * unlucky processes, a process is deemed slow only if
3106 * its rate has been lower than half of the estimated
3109 slow = bfqq->entity.service < bfqd->bfq_max_budget / 2;
3112 bfq_log_bfqq(bfqd, bfqq, "bfq_bfqq_is_slow: slow %d", slow);
3118 * To be deemed as soft real-time, an application must meet two
3119 * requirements. First, the application must not require an average
3120 * bandwidth higher than the approximate bandwidth required to playback or
3121 * record a compressed high-definition video.
3122 * The next function is invoked on the completion of the last request of a
3123 * batch, to compute the next-start time instant, soft_rt_next_start, such
3124 * that, if the next request of the application does not arrive before
3125 * soft_rt_next_start, then the above requirement on the bandwidth is met.
3127 * The second requirement is that the request pattern of the application is
3128 * isochronous, i.e., that, after issuing a request or a batch of requests,
3129 * the application stops issuing new requests until all its pending requests
3130 * have been completed. After that, the application may issue a new batch,
3132 * For this reason the next function is invoked to compute
3133 * soft_rt_next_start only for applications that meet this requirement,
3134 * whereas soft_rt_next_start is set to infinity for applications that do
3137 * Unfortunately, even a greedy (i.e., I/O-bound) application may
3138 * happen to meet, occasionally or systematically, both the above
3139 * bandwidth and isochrony requirements. This may happen at least in
3140 * the following circumstances. First, if the CPU load is high. The
3141 * application may stop issuing requests while the CPUs are busy
3142 * serving other processes, then restart, then stop again for a while,
3143 * and so on. The other circumstances are related to the storage
3144 * device: the storage device is highly loaded or reaches a low-enough
3145 * throughput with the I/O of the application (e.g., because the I/O
3146 * is random and/or the device is slow). In all these cases, the
3147 * I/O of the application may be simply slowed down enough to meet
3148 * the bandwidth and isochrony requirements. To reduce the probability
3149 * that greedy applications are deemed as soft real-time in these
3150 * corner cases, a further rule is used in the computation of
3151 * soft_rt_next_start: the return value of this function is forced to
3152 * be higher than the maximum between the following two quantities.
3154 * (a) Current time plus: (1) the maximum time for which the arrival
3155 * of a request is waited for when a sync queue becomes idle,
3156 * namely bfqd->bfq_slice_idle, and (2) a few extra jiffies. We
3157 * postpone for a moment the reason for adding a few extra
3158 * jiffies; we get back to it after next item (b). Lower-bounding
3159 * the return value of this function with the current time plus
3160 * bfqd->bfq_slice_idle tends to filter out greedy applications,
3161 * because the latter issue their next request as soon as possible
3162 * after the last one has been completed. In contrast, a soft
3163 * real-time application spends some time processing data, after a
3164 * batch of its requests has been completed.
3166 * (b) Current value of bfqq->soft_rt_next_start. As pointed out
3167 * above, greedy applications may happen to meet both the
3168 * bandwidth and isochrony requirements under heavy CPU or
3169 * storage-device load. In more detail, in these scenarios, these
3170 * applications happen, only for limited time periods, to do I/O
3171 * slowly enough to meet all the requirements described so far,
3172 * including the filtering in above item (a). These slow-speed
3173 * time intervals are usually interspersed between other time
3174 * intervals during which these applications do I/O at a very high
3175 * speed. Fortunately, exactly because of the high speed of the
3176 * I/O in the high-speed intervals, the values returned by this
3177 * function happen to be so high, near the end of any such
3178 * high-speed interval, to be likely to fall *after* the end of
3179 * the low-speed time interval that follows. These high values are
3180 * stored in bfqq->soft_rt_next_start after each invocation of
3181 * this function. As a consequence, if the last value of
3182 * bfqq->soft_rt_next_start is constantly used to lower-bound the
3183 * next value that this function may return, then, from the very
3184 * beginning of a low-speed interval, bfqq->soft_rt_next_start is
3185 * likely to be constantly kept so high that any I/O request
3186 * issued during the low-speed interval is considered as arriving
3187 * to soon for the application to be deemed as soft
3188 * real-time. Then, in the high-speed interval that follows, the
3189 * application will not be deemed as soft real-time, just because
3190 * it will do I/O at a high speed. And so on.
3192 * Getting back to the filtering in item (a), in the following two
3193 * cases this filtering might be easily passed by a greedy
3194 * application, if the reference quantity was just
3195 * bfqd->bfq_slice_idle:
3196 * 1) HZ is so low that the duration of a jiffy is comparable to or
3197 * higher than bfqd->bfq_slice_idle. This happens, e.g., on slow
3198 * devices with HZ=100. The time granularity may be so coarse
3199 * that the approximation, in jiffies, of bfqd->bfq_slice_idle
3200 * is rather lower than the exact value.
3201 * 2) jiffies, instead of increasing at a constant rate, may stop increasing
3202 * for a while, then suddenly 'jump' by several units to recover the lost
3203 * increments. This seems to happen, e.g., inside virtual machines.
3204 * To address this issue, in the filtering in (a) we do not use as a
3205 * reference time interval just bfqd->bfq_slice_idle, but
3206 * bfqd->bfq_slice_idle plus a few jiffies. In particular, we add the
3207 * minimum number of jiffies for which the filter seems to be quite
3208 * precise also in embedded systems and KVM/QEMU virtual machines.
3210 static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data *bfqd,
3211 struct bfq_queue *bfqq)
3213 return max3(bfqq->soft_rt_next_start,
3214 bfqq->last_idle_bklogged +
3215 HZ * bfqq->service_from_backlogged /
3216 bfqd->bfq_wr_max_softrt_rate,
3217 jiffies + nsecs_to_jiffies(bfqq->bfqd->bfq_slice_idle) + 4);
3220 static bool bfq_bfqq_injectable(struct bfq_queue *bfqq)
3222 return BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 &&
3223 blk_queue_nonrot(bfqq->bfqd->queue) &&
3228 * bfq_bfqq_expire - expire a queue.
3229 * @bfqd: device owning the queue.
3230 * @bfqq: the queue to expire.
3231 * @compensate: if true, compensate for the time spent idling.
3232 * @reason: the reason causing the expiration.
3234 * If the process associated with bfqq does slow I/O (e.g., because it
3235 * issues random requests), we charge bfqq with the time it has been
3236 * in service instead of the service it has received (see
3237 * bfq_bfqq_charge_time for details on how this goal is achieved). As
3238 * a consequence, bfqq will typically get higher timestamps upon
3239 * reactivation, and hence it will be rescheduled as if it had
3240 * received more service than what it has actually received. In the
3241 * end, bfqq receives less service in proportion to how slowly its
3242 * associated process consumes its budgets (and hence how seriously it
3243 * tends to lower the throughput). In addition, this time-charging
3244 * strategy guarantees time fairness among slow processes. In
3245 * contrast, if the process associated with bfqq is not slow, we
3246 * charge bfqq exactly with the service it has received.
3248 * Charging time to the first type of queues and the exact service to
3249 * the other has the effect of using the WF2Q+ policy to schedule the
3250 * former on a timeslice basis, without violating service domain
3251 * guarantees among the latter.
3253 void bfq_bfqq_expire(struct bfq_data *bfqd,
3254 struct bfq_queue *bfqq,
3256 enum bfqq_expiration reason)
3259 unsigned long delta = 0;
3260 struct bfq_entity *entity = &bfqq->entity;
3264 * Check whether the process is slow (see bfq_bfqq_is_slow).
3266 slow = bfq_bfqq_is_slow(bfqd, bfqq, compensate, reason, &delta);
3269 * As above explained, charge slow (typically seeky) and
3270 * timed-out queues with the time and not the service
3271 * received, to favor sequential workloads.
3273 * Processes doing I/O in the slower disk zones will tend to
3274 * be slow(er) even if not seeky. Therefore, since the
3275 * estimated peak rate is actually an average over the disk
3276 * surface, these processes may timeout just for bad luck. To
3277 * avoid punishing them, do not charge time to processes that
3278 * succeeded in consuming at least 2/3 of their budget. This
3279 * allows BFQ to preserve enough elasticity to still perform
3280 * bandwidth, and not time, distribution with little unlucky
3281 * or quasi-sequential processes.
3283 if (bfqq->wr_coeff == 1 &&
3285 (reason == BFQQE_BUDGET_TIMEOUT &&
3286 bfq_bfqq_budget_left(bfqq) >= entity->budget / 3)))
3287 bfq_bfqq_charge_time(bfqd, bfqq, delta);
3289 if (reason == BFQQE_TOO_IDLE &&
3290 entity->service <= 2 * entity->budget / 10)
3291 bfq_clear_bfqq_IO_bound(bfqq);
3293 if (bfqd->low_latency && bfqq->wr_coeff == 1)
3294 bfqq->last_wr_start_finish = jiffies;
3296 if (bfqd->low_latency && bfqd->bfq_wr_max_softrt_rate > 0 &&
3297 RB_EMPTY_ROOT(&bfqq->sort_list)) {
3299 * If we get here, and there are no outstanding
3300 * requests, then the request pattern is isochronous
3301 * (see the comments on the function
3302 * bfq_bfqq_softrt_next_start()). Thus we can compute
3303 * soft_rt_next_start. And we do it, unless bfqq is in
3304 * interactive weight raising. We do not do it in the
3305 * latter subcase, for the following reason. bfqq may
3306 * be conveying the I/O needed to load a soft
3307 * real-time application. Such an application will
3308 * actually exhibit a soft real-time I/O pattern after
3309 * it finally starts doing its job. But, if
3310 * soft_rt_next_start is computed here for an
3311 * interactive bfqq, and bfqq had received a lot of
3312 * service before remaining with no outstanding
3313 * request (likely to happen on a fast device), then
3314 * soft_rt_next_start would be assigned such a high
3315 * value that, for a very long time, bfqq would be
3316 * prevented from being possibly considered as soft
3319 * If, instead, the queue still has outstanding
3320 * requests, then we have to wait for the completion
3321 * of all the outstanding requests to discover whether
3322 * the request pattern is actually isochronous.
3324 if (bfqq->dispatched == 0 &&
3325 bfqq->wr_coeff != bfqd->bfq_wr_coeff)
3326 bfqq->soft_rt_next_start =
3327 bfq_bfqq_softrt_next_start(bfqd, bfqq);
3328 else if (bfqq->dispatched > 0) {
3330 * Schedule an update of soft_rt_next_start to when
3331 * the task may be discovered to be isochronous.
3333 bfq_mark_bfqq_softrt_update(bfqq);
3337 bfq_log_bfqq(bfqd, bfqq,
3338 "expire (%d, slow %d, num_disp %d, short_ttime %d)", reason,
3339 slow, bfqq->dispatched, bfq_bfqq_has_short_ttime(bfqq));
3342 * Increase, decrease or leave budget unchanged according to
3345 __bfq_bfqq_recalc_budget(bfqd, bfqq, reason);
3347 __bfq_bfqq_expire(bfqd, bfqq);
3349 if (ref == 1) /* bfqq is gone, no more actions on it */
3352 bfqq->injected_service = 0;
3354 /* mark bfqq as waiting a request only if a bic still points to it */
3355 if (!bfq_bfqq_busy(bfqq) &&
3356 reason != BFQQE_BUDGET_TIMEOUT &&
3357 reason != BFQQE_BUDGET_EXHAUSTED) {
3358 bfq_mark_bfqq_non_blocking_wait_rq(bfqq);
3360 * Not setting service to 0, because, if the next rq
3361 * arrives in time, the queue will go on receiving
3362 * service with this same budget (as if it never expired)
3365 entity->service = 0;
3368 * Reset the received-service counter for every parent entity.
3369 * Differently from what happens with bfqq->entity.service,
3370 * the resetting of this counter never needs to be postponed
3371 * for parent entities. In fact, in case bfqq may have a
3372 * chance to go on being served using the last, partially
3373 * consumed budget, bfqq->entity.service needs to be kept,
3374 * because if bfqq then actually goes on being served using
3375 * the same budget, the last value of bfqq->entity.service is
3376 * needed to properly decrement bfqq->entity.budget by the
3377 * portion already consumed. In contrast, it is not necessary
3378 * to keep entity->service for parent entities too, because
3379 * the bubble up of the new value of bfqq->entity.budget will
3380 * make sure that the budgets of parent entities are correct,
3381 * even in case bfqq and thus parent entities go on receiving
3382 * service with the same budget.
3384 entity = entity->parent;
3385 for_each_entity(entity)
3386 entity->service = 0;
3390 * Budget timeout is not implemented through a dedicated timer, but
3391 * just checked on request arrivals and completions, as well as on
3392 * idle timer expirations.
3394 static bool bfq_bfqq_budget_timeout(struct bfq_queue *bfqq)
3396 return time_is_before_eq_jiffies(bfqq->budget_timeout);
3400 * If we expire a queue that is actively waiting (i.e., with the
3401 * device idled) for the arrival of a new request, then we may incur
3402 * the timestamp misalignment problem described in the body of the
3403 * function __bfq_activate_entity. Hence we return true only if this
3404 * condition does not hold, or if the queue is slow enough to deserve
3405 * only to be kicked off for preserving a high throughput.
3407 static bool bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq)
3409 bfq_log_bfqq(bfqq->bfqd, bfqq,
3410 "may_budget_timeout: wait_request %d left %d timeout %d",
3411 bfq_bfqq_wait_request(bfqq),
3412 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3,
3413 bfq_bfqq_budget_timeout(bfqq));
3415 return (!bfq_bfqq_wait_request(bfqq) ||
3416 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3)
3418 bfq_bfqq_budget_timeout(bfqq);
3421 static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd,
3422 struct bfq_queue *bfqq)
3424 bool rot_without_queueing =
3425 !blk_queue_nonrot(bfqd->queue) && !bfqd->hw_tag,
3426 bfqq_sequential_and_IO_bound,
3429 bfqq_sequential_and_IO_bound = !BFQQ_SEEKY(bfqq) &&
3430 bfq_bfqq_IO_bound(bfqq) && bfq_bfqq_has_short_ttime(bfqq);
3433 * The next variable takes into account the cases where idling
3434 * boosts the throughput.
3436 * The value of the variable is computed considering, first, that
3437 * idling is virtually always beneficial for the throughput if:
3438 * (a) the device is not NCQ-capable and rotational, or
3439 * (b) regardless of the presence of NCQ, the device is rotational and
3440 * the request pattern for bfqq is I/O-bound and sequential, or
3441 * (c) regardless of whether it is rotational, the device is
3442 * not NCQ-capable and the request pattern for bfqq is
3443 * I/O-bound and sequential.
3445 * Secondly, and in contrast to the above item (b), idling an
3446 * NCQ-capable flash-based device would not boost the
3447 * throughput even with sequential I/O; rather it would lower
3448 * the throughput in proportion to how fast the device
3449 * is. Accordingly, the next variable is true if any of the
3450 * above conditions (a), (b) or (c) is true, and, in
3451 * particular, happens to be false if bfqd is an NCQ-capable
3452 * flash-based device.
3454 idling_boosts_thr = rot_without_queueing ||
3455 ((!blk_queue_nonrot(bfqd->queue) || !bfqd->hw_tag) &&
3456 bfqq_sequential_and_IO_bound);
3459 * The return value of this function is equal to that of
3460 * idling_boosts_thr, unless a special case holds. In this
3461 * special case, described below, idling may cause problems to
3462 * weight-raised queues.
3464 * When the request pool is saturated (e.g., in the presence
3465 * of write hogs), if the processes associated with
3466 * non-weight-raised queues ask for requests at a lower rate,
3467 * then processes associated with weight-raised queues have a
3468 * higher probability to get a request from the pool
3469 * immediately (or at least soon) when they need one. Thus
3470 * they have a higher probability to actually get a fraction
3471 * of the device throughput proportional to their high
3472 * weight. This is especially true with NCQ-capable drives,
3473 * which enqueue several requests in advance, and further
3474 * reorder internally-queued requests.
3476 * For this reason, we force to false the return value if
3477 * there are weight-raised busy queues. In this case, and if
3478 * bfqq is not weight-raised, this guarantees that the device
3479 * is not idled for bfqq (if, instead, bfqq is weight-raised,
3480 * then idling will be guaranteed by another variable, see
3481 * below). Combined with the timestamping rules of BFQ (see
3482 * [1] for details), this behavior causes bfqq, and hence any
3483 * sync non-weight-raised queue, to get a lower number of
3484 * requests served, and thus to ask for a lower number of
3485 * requests from the request pool, before the busy
3486 * weight-raised queues get served again. This often mitigates
3487 * starvation problems in the presence of heavy write
3488 * workloads and NCQ, thereby guaranteeing a higher
3489 * application and system responsiveness in these hostile
3492 return idling_boosts_thr &&
3493 bfqd->wr_busy_queues == 0;
3497 * There is a case where idling must be performed not for
3498 * throughput concerns, but to preserve service guarantees.
3500 * To introduce this case, we can note that allowing the drive
3501 * to enqueue more than one request at a time, and hence
3502 * delegating de facto final scheduling decisions to the
3503 * drive's internal scheduler, entails loss of control on the
3504 * actual request service order. In particular, the critical
3505 * situation is when requests from different processes happen
3506 * to be present, at the same time, in the internal queue(s)
3507 * of the drive. In such a situation, the drive, by deciding
3508 * the service order of the internally-queued requests, does
3509 * determine also the actual throughput distribution among
3510 * these processes. But the drive typically has no notion or
3511 * concern about per-process throughput distribution, and
3512 * makes its decisions only on a per-request basis. Therefore,
3513 * the service distribution enforced by the drive's internal
3514 * scheduler is likely to coincide with the desired
3515 * device-throughput distribution only in a completely
3516 * symmetric scenario where:
3517 * (i) each of these processes must get the same throughput as
3519 * (ii) the I/O of each process has the same properties, in
3520 * terms of locality (sequential or random), direction
3521 * (reads or writes), request sizes, greediness
3522 * (from I/O-bound to sporadic), and so on.
3523 * In fact, in such a scenario, the drive tends to treat
3524 * the requests of each of these processes in about the same
3525 * way as the requests of the others, and thus to provide
3526 * each of these processes with about the same throughput
3527 * (which is exactly the desired throughput distribution). In
3528 * contrast, in any asymmetric scenario, device idling is
3529 * certainly needed to guarantee that bfqq receives its
3530 * assigned fraction of the device throughput (see [1] for
3532 * The problem is that idling may significantly reduce
3533 * throughput with certain combinations of types of I/O and
3534 * devices. An important example is sync random I/O, on flash
3535 * storage with command queueing. So, unless bfqq falls in the
3536 * above cases where idling also boosts throughput, it would
3537 * be important to check conditions (i) and (ii) accurately,
3538 * so as to avoid idling when not strictly needed for service
3541 * Unfortunately, it is extremely difficult to thoroughly
3542 * check condition (ii). And, in case there are active groups,
3543 * it becomes very difficult to check condition (i) too. In
3544 * fact, if there are active groups, then, for condition (i)
3545 * to become false, it is enough that an active group contains
3546 * more active processes or sub-groups than some other active
3547 * group. More precisely, for condition (i) to hold because of
3548 * such a group, it is not even necessary that the group is
3549 * (still) active: it is sufficient that, even if the group
3550 * has become inactive, some of its descendant processes still
3551 * have some request already dispatched but still waiting for
3552 * completion. In fact, requests have still to be guaranteed
3553 * their share of the throughput even after being
3554 * dispatched. In this respect, it is easy to show that, if a
3555 * group frequently becomes inactive while still having
3556 * in-flight requests, and if, when this happens, the group is
3557 * not considered in the calculation of whether the scenario
3558 * is asymmetric, then the group may fail to be guaranteed its
3559 * fair share of the throughput (basically because idling may
3560 * not be performed for the descendant processes of the group,
3561 * but it had to be). We address this issue with the
3562 * following bi-modal behavior, implemented in the function
3563 * bfq_symmetric_scenario().
3565 * If there are groups with requests waiting for completion
3566 * (as commented above, some of these groups may even be
3567 * already inactive), then the scenario is tagged as
3568 * asymmetric, conservatively, without checking any of the
3569 * conditions (i) and (ii). So the device is idled for bfqq.
3570 * This behavior matches also the fact that groups are created
3571 * exactly if controlling I/O is a primary concern (to
3572 * preserve bandwidth and latency guarantees).
3574 * On the opposite end, if there are no groups with requests
3575 * waiting for completion, then only condition (i) is actually
3576 * controlled, i.e., provided that condition (i) holds, idling
3577 * is not performed, regardless of whether condition (ii)
3578 * holds. In other words, only if condition (i) does not hold,
3579 * then idling is allowed, and the device tends to be
3580 * prevented from queueing many requests, possibly of several
3581 * processes. Since there are no groups with requests waiting
3582 * for completion, then, to control condition (i) it is enough
3583 * to check just whether all the queues with requests waiting
3584 * for completion also have the same weight.
3586 * Not checking condition (ii) evidently exposes bfqq to the
3587 * risk of getting less throughput than its fair share.
3588 * However, for queues with the same weight, a further
3589 * mechanism, preemption, mitigates or even eliminates this
3590 * problem. And it does so without consequences on overall
3591 * throughput. This mechanism and its benefits are explained
3592 * in the next three paragraphs.
3594 * Even if a queue, say Q, is expired when it remains idle, Q
3595 * can still preempt the new in-service queue if the next
3596 * request of Q arrives soon (see the comments on
3597 * bfq_bfqq_update_budg_for_activation). If all queues and
3598 * groups have the same weight, this form of preemption,
3599 * combined with the hole-recovery heuristic described in the
3600 * comments on function bfq_bfqq_update_budg_for_activation,
3601 * are enough to preserve a correct bandwidth distribution in
3602 * the mid term, even without idling. In fact, even if not
3603 * idling allows the internal queues of the device to contain
3604 * many requests, and thus to reorder requests, we can rather
3605 * safely assume that the internal scheduler still preserves a
3606 * minimum of mid-term fairness.
3608 * More precisely, this preemption-based, idleless approach
3609 * provides fairness in terms of IOPS, and not sectors per
3610 * second. This can be seen with a simple example. Suppose
3611 * that there are two queues with the same weight, but that
3612 * the first queue receives requests of 8 sectors, while the
3613 * second queue receives requests of 1024 sectors. In
3614 * addition, suppose that each of the two queues contains at
3615 * most one request at a time, which implies that each queue
3616 * always remains idle after it is served. Finally, after
3617 * remaining idle, each queue receives very quickly a new
3618 * request. It follows that the two queues are served
3619 * alternatively, preempting each other if needed. This
3620 * implies that, although both queues have the same weight,
3621 * the queue with large requests receives a service that is
3622 * 1024/8 times as high as the service received by the other
3625 * The motivation for using preemption instead of idling (for
3626 * queues with the same weight) is that, by not idling,
3627 * service guarantees are preserved (completely or at least in
3628 * part) without minimally sacrificing throughput. And, if
3629 * there is no active group, then the primary expectation for
3630 * this device is probably a high throughput.
3632 * We are now left only with explaining the additional
3633 * compound condition that is checked below for deciding
3634 * whether the scenario is asymmetric. To explain this
3635 * compound condition, we need to add that the function
3636 * bfq_symmetric_scenario checks the weights of only
3637 * non-weight-raised queues, for efficiency reasons (see
3638 * comments on bfq_weights_tree_add()). Then the fact that
3639 * bfqq is weight-raised is checked explicitly here. More
3640 * precisely, the compound condition below takes into account
3641 * also the fact that, even if bfqq is being weight-raised,
3642 * the scenario is still symmetric if all queues with requests
3643 * waiting for completion happen to be
3644 * weight-raised. Actually, we should be even more precise
3645 * here, and differentiate between interactive weight raising
3646 * and soft real-time weight raising.
3648 * As a side note, it is worth considering that the above
3649 * device-idling countermeasures may however fail in the
3650 * following unlucky scenario: if idling is (correctly)
3651 * disabled in a time period during which all symmetry
3652 * sub-conditions hold, and hence the device is allowed to
3653 * enqueue many requests, but at some later point in time some
3654 * sub-condition stops to hold, then it may become impossible
3655 * to let requests be served in the desired order until all
3656 * the requests already queued in the device have been served.
3658 static bool idling_needed_for_service_guarantees(struct bfq_data *bfqd,
3659 struct bfq_queue *bfqq)
3661 return (bfqq->wr_coeff > 1 &&
3662 bfqd->wr_busy_queues <
3663 bfq_tot_busy_queues(bfqd)) ||
3664 !bfq_symmetric_scenario(bfqd);
3668 * For a queue that becomes empty, device idling is allowed only if
3669 * this function returns true for that queue. As a consequence, since
3670 * device idling plays a critical role for both throughput boosting
3671 * and service guarantees, the return value of this function plays a
3672 * critical role as well.
3674 * In a nutshell, this function returns true only if idling is
3675 * beneficial for throughput or, even if detrimental for throughput,
3676 * idling is however necessary to preserve service guarantees (low
3677 * latency, desired throughput distribution, ...). In particular, on
3678 * NCQ-capable devices, this function tries to return false, so as to
3679 * help keep the drives' internal queues full, whenever this helps the
3680 * device boost the throughput without causing any service-guarantee
3683 * Most of the issues taken into account to get the return value of
3684 * this function are not trivial. We discuss these issues in the two
3685 * functions providing the main pieces of information needed by this
3688 static bool bfq_better_to_idle(struct bfq_queue *bfqq)
3690 struct bfq_data *bfqd = bfqq->bfqd;
3691 bool idling_boosts_thr_with_no_issue, idling_needed_for_service_guar;
3693 if (unlikely(bfqd->strict_guarantees))
3697 * Idling is performed only if slice_idle > 0. In addition, we
3700 * (b) bfqq is in the idle io prio class: in this case we do
3701 * not idle because we want to minimize the bandwidth that
3702 * queues in this class can steal to higher-priority queues
3704 if (bfqd->bfq_slice_idle == 0 || !bfq_bfqq_sync(bfqq) ||
3705 bfq_class_idle(bfqq))
3708 idling_boosts_thr_with_no_issue =
3709 idling_boosts_thr_without_issues(bfqd, bfqq);
3711 idling_needed_for_service_guar =
3712 idling_needed_for_service_guarantees(bfqd, bfqq);
3715 * We have now the two components we need to compute the
3716 * return value of the function, which is true only if idling
3717 * either boosts the throughput (without issues), or is
3718 * necessary to preserve service guarantees.
3720 return idling_boosts_thr_with_no_issue ||
3721 idling_needed_for_service_guar;
3725 * If the in-service queue is empty but the function bfq_better_to_idle
3726 * returns true, then:
3727 * 1) the queue must remain in service and cannot be expired, and
3728 * 2) the device must be idled to wait for the possible arrival of a new
3729 * request for the queue.
3730 * See the comments on the function bfq_better_to_idle for the reasons
3731 * why performing device idling is the best choice to boost the throughput
3732 * and preserve service guarantees when bfq_better_to_idle itself
3735 static bool bfq_bfqq_must_idle(struct bfq_queue *bfqq)
3737 return RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_better_to_idle(bfqq);
3740 static struct bfq_queue *bfq_choose_bfqq_for_injection(struct bfq_data *bfqd)
3742 struct bfq_queue *bfqq;
3745 * A linear search; but, with a high probability, very few
3746 * steps are needed to find a candidate queue, i.e., a queue
3747 * with enough budget left for its next request. In fact:
3748 * - BFQ dynamically updates the budget of every queue so as
3749 * to accommodate the expected backlog of the queue;
3750 * - if a queue gets all its requests dispatched as injected
3751 * service, then the queue is removed from the active list
3752 * (and re-added only if it gets new requests, but with
3753 * enough budget for its new backlog).
3755 list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
3756 if (!RB_EMPTY_ROOT(&bfqq->sort_list) &&
3757 bfq_serv_to_charge(bfqq->next_rq, bfqq) <=
3758 bfq_bfqq_budget_left(bfqq))
3765 * Select a queue for service. If we have a current queue in service,
3766 * check whether to continue servicing it, or retrieve and set a new one.
3768 static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd)
3770 struct bfq_queue *bfqq;
3771 struct request *next_rq;
3772 enum bfqq_expiration reason = BFQQE_BUDGET_TIMEOUT;
3774 bfqq = bfqd->in_service_queue;
3778 bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue");
3781 * Do not expire bfqq for budget timeout if bfqq may be about
3782 * to enjoy device idling. The reason why, in this case, we
3783 * prevent bfqq from expiring is the same as in the comments
3784 * on the case where bfq_bfqq_must_idle() returns true, in
3785 * bfq_completed_request().
3787 if (bfq_may_expire_for_budg_timeout(bfqq) &&
3788 !bfq_bfqq_must_idle(bfqq))
3793 * This loop is rarely executed more than once. Even when it
3794 * happens, it is much more convenient to re-execute this loop
3795 * than to return NULL and trigger a new dispatch to get a
3798 next_rq = bfqq->next_rq;
3800 * If bfqq has requests queued and it has enough budget left to
3801 * serve them, keep the queue, otherwise expire it.
3804 if (bfq_serv_to_charge(next_rq, bfqq) >
3805 bfq_bfqq_budget_left(bfqq)) {
3807 * Expire the queue for budget exhaustion,
3808 * which makes sure that the next budget is
3809 * enough to serve the next request, even if
3810 * it comes from the fifo expired path.
3812 reason = BFQQE_BUDGET_EXHAUSTED;
3816 * The idle timer may be pending because we may
3817 * not disable disk idling even when a new request
3820 if (bfq_bfqq_wait_request(bfqq)) {
3822 * If we get here: 1) at least a new request
3823 * has arrived but we have not disabled the
3824 * timer because the request was too small,
3825 * 2) then the block layer has unplugged
3826 * the device, causing the dispatch to be
3829 * Since the device is unplugged, now the
3830 * requests are probably large enough to
3831 * provide a reasonable throughput.
3832 * So we disable idling.
3834 bfq_clear_bfqq_wait_request(bfqq);
3835 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
3842 * No requests pending. However, if the in-service queue is idling
3843 * for a new request, or has requests waiting for a completion and
3844 * may idle after their completion, then keep it anyway.
3846 * Yet, to boost throughput, inject service from other queues if
3849 if (bfq_bfqq_wait_request(bfqq) ||
3850 (bfqq->dispatched != 0 && bfq_better_to_idle(bfqq))) {
3851 if (bfq_bfqq_injectable(bfqq) &&
3852 bfqq->injected_service * bfqq->inject_coeff <
3853 bfqq->entity.service * 10)
3854 bfqq = bfq_choose_bfqq_for_injection(bfqd);
3861 reason = BFQQE_NO_MORE_REQUESTS;
3863 bfq_bfqq_expire(bfqd, bfqq, false, reason);
3865 bfqq = bfq_set_in_service_queue(bfqd);
3867 bfq_log_bfqq(bfqd, bfqq, "select_queue: checking new queue");
3872 bfq_log_bfqq(bfqd, bfqq, "select_queue: returned this queue");
3874 bfq_log(bfqd, "select_queue: no queue returned");
3879 static void bfq_update_wr_data(struct bfq_data *bfqd, struct bfq_queue *bfqq)
3881 struct bfq_entity *entity = &bfqq->entity;
3883 if (bfqq->wr_coeff > 1) { /* queue is being weight-raised */
3884 bfq_log_bfqq(bfqd, bfqq,
3885 "raising period dur %u/%u msec, old coeff %u, w %d(%d)",
3886 jiffies_to_msecs(jiffies - bfqq->last_wr_start_finish),
3887 jiffies_to_msecs(bfqq->wr_cur_max_time),
3889 bfqq->entity.weight, bfqq->entity.orig_weight);
3891 if (entity->prio_changed)
3892 bfq_log_bfqq(bfqd, bfqq, "WARN: pending prio change");
3895 * If the queue was activated in a burst, or too much
3896 * time has elapsed from the beginning of this
3897 * weight-raising period, then end weight raising.
3899 if (bfq_bfqq_in_large_burst(bfqq))
3900 bfq_bfqq_end_wr(bfqq);
3901 else if (time_is_before_jiffies(bfqq->last_wr_start_finish +
3902 bfqq->wr_cur_max_time)) {
3903 if (bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time ||
3904 time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
3905 bfq_wr_duration(bfqd)))
3906 bfq_bfqq_end_wr(bfqq);
3908 switch_back_to_interactive_wr(bfqq, bfqd);
3909 bfqq->entity.prio_changed = 1;
3912 if (bfqq->wr_coeff > 1 &&
3913 bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time &&
3914 bfqq->service_from_wr > max_service_from_wr) {
3915 /* see comments on max_service_from_wr */
3916 bfq_bfqq_end_wr(bfqq);
3920 * To improve latency (for this or other queues), immediately
3921 * update weight both if it must be raised and if it must be
3922 * lowered. Since, entity may be on some active tree here, and
3923 * might have a pending change of its ioprio class, invoke
3924 * next function with the last parameter unset (see the
3925 * comments on the function).
3927 if ((entity->weight > entity->orig_weight) != (bfqq->wr_coeff > 1))
3928 __bfq_entity_update_weight_prio(bfq_entity_service_tree(entity),
3933 * Dispatch next request from bfqq.
3935 static struct request *bfq_dispatch_rq_from_bfqq(struct bfq_data *bfqd,
3936 struct bfq_queue *bfqq)
3938 struct request *rq = bfqq->next_rq;
3939 unsigned long service_to_charge;
3941 service_to_charge = bfq_serv_to_charge(rq, bfqq);
3943 bfq_bfqq_served(bfqq, service_to_charge);
3945 bfq_dispatch_remove(bfqd->queue, rq);
3947 if (bfqq != bfqd->in_service_queue) {
3948 if (likely(bfqd->in_service_queue))
3949 bfqd->in_service_queue->injected_service +=
3950 bfq_serv_to_charge(rq, bfqq);
3956 * If weight raising has to terminate for bfqq, then next
3957 * function causes an immediate update of bfqq's weight,
3958 * without waiting for next activation. As a consequence, on
3959 * expiration, bfqq will be timestamped as if has never been
3960 * weight-raised during this service slot, even if it has
3961 * received part or even most of the service as a
3962 * weight-raised queue. This inflates bfqq's timestamps, which
3963 * is beneficial, as bfqq is then more willing to leave the
3964 * device immediately to possible other weight-raised queues.
3966 bfq_update_wr_data(bfqd, bfqq);
3969 * Expire bfqq, pretending that its budget expired, if bfqq
3970 * belongs to CLASS_IDLE and other queues are waiting for
3973 if (!(bfq_tot_busy_queues(bfqd) > 1 && bfq_class_idle(bfqq)))
3976 bfq_bfqq_expire(bfqd, bfqq, false, BFQQE_BUDGET_EXHAUSTED);
3982 static bool bfq_has_work(struct blk_mq_hw_ctx *hctx)
3984 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
3987 * Avoiding lock: a race on bfqd->busy_queues should cause at
3988 * most a call to dispatch for nothing
3990 return !list_empty_careful(&bfqd->dispatch) ||
3991 bfq_tot_busy_queues(bfqd) > 0;
3994 static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
3996 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
3997 struct request *rq = NULL;
3998 struct bfq_queue *bfqq = NULL;
4000 if (!list_empty(&bfqd->dispatch)) {
4001 rq = list_first_entry(&bfqd->dispatch, struct request,
4003 list_del_init(&rq->queuelist);
4009 * Increment counters here, because this
4010 * dispatch does not follow the standard
4011 * dispatch flow (where counters are
4016 goto inc_in_driver_start_rq;
4020 * We exploit the bfq_finish_requeue_request hook to
4021 * decrement rq_in_driver, but
4022 * bfq_finish_requeue_request will not be invoked on
4023 * this request. So, to avoid unbalance, just start
4024 * this request, without incrementing rq_in_driver. As
4025 * a negative consequence, rq_in_driver is deceptively
4026 * lower than it should be while this request is in
4027 * service. This may cause bfq_schedule_dispatch to be
4028 * invoked uselessly.
4030 * As for implementing an exact solution, the
4031 * bfq_finish_requeue_request hook, if defined, is
4032 * probably invoked also on this request. So, by
4033 * exploiting this hook, we could 1) increment
4034 * rq_in_driver here, and 2) decrement it in
4035 * bfq_finish_requeue_request. Such a solution would
4036 * let the value of the counter be always accurate,
4037 * but it would entail using an extra interface
4038 * function. This cost seems higher than the benefit,
4039 * being the frequency of non-elevator-private
4040 * requests very low.
4045 bfq_log(bfqd, "dispatch requests: %d busy queues",
4046 bfq_tot_busy_queues(bfqd));
4048 if (bfq_tot_busy_queues(bfqd) == 0)
4052 * Force device to serve one request at a time if
4053 * strict_guarantees is true. Forcing this service scheme is
4054 * currently the ONLY way to guarantee that the request
4055 * service order enforced by the scheduler is respected by a
4056 * queueing device. Otherwise the device is free even to make
4057 * some unlucky request wait for as long as the device
4060 * Of course, serving one request at at time may cause loss of
4063 if (bfqd->strict_guarantees && bfqd->rq_in_driver > 0)
4066 bfqq = bfq_select_queue(bfqd);
4070 rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq);
4073 inc_in_driver_start_rq:
4074 bfqd->rq_in_driver++;
4076 rq->rq_flags |= RQF_STARTED;
4082 #if defined(CONFIG_BFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
4083 static void bfq_update_dispatch_stats(struct request_queue *q,
4085 struct bfq_queue *in_serv_queue,
4086 bool idle_timer_disabled)
4088 struct bfq_queue *bfqq = rq ? RQ_BFQQ(rq) : NULL;
4090 if (!idle_timer_disabled && !bfqq)
4094 * rq and bfqq are guaranteed to exist until this function
4095 * ends, for the following reasons. First, rq can be
4096 * dispatched to the device, and then can be completed and
4097 * freed, only after this function ends. Second, rq cannot be
4098 * merged (and thus freed because of a merge) any longer,
4099 * because it has already started. Thus rq cannot be freed
4100 * before this function ends, and, since rq has a reference to
4101 * bfqq, the same guarantee holds for bfqq too.
4103 * In addition, the following queue lock guarantees that
4104 * bfqq_group(bfqq) exists as well.
4106 spin_lock_irq(&q->queue_lock);
4107 if (idle_timer_disabled)
4109 * Since the idle timer has been disabled,
4110 * in_serv_queue contained some request when
4111 * __bfq_dispatch_request was invoked above, which
4112 * implies that rq was picked exactly from
4113 * in_serv_queue. Thus in_serv_queue == bfqq, and is
4114 * therefore guaranteed to exist because of the above
4117 bfqg_stats_update_idle_time(bfqq_group(in_serv_queue));
4119 struct bfq_group *bfqg = bfqq_group(bfqq);
4121 bfqg_stats_update_avg_queue_size(bfqg);
4122 bfqg_stats_set_start_empty_time(bfqg);
4123 bfqg_stats_update_io_remove(bfqg, rq->cmd_flags);
4125 spin_unlock_irq(&q->queue_lock);
4128 static inline void bfq_update_dispatch_stats(struct request_queue *q,
4130 struct bfq_queue *in_serv_queue,
4131 bool idle_timer_disabled) {}
4134 static struct request *bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
4136 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
4138 struct bfq_queue *in_serv_queue;
4139 bool waiting_rq, idle_timer_disabled;
4141 spin_lock_irq(&bfqd->lock);
4143 in_serv_queue = bfqd->in_service_queue;
4144 waiting_rq = in_serv_queue && bfq_bfqq_wait_request(in_serv_queue);
4146 rq = __bfq_dispatch_request(hctx);
4148 idle_timer_disabled =
4149 waiting_rq && !bfq_bfqq_wait_request(in_serv_queue);
4151 spin_unlock_irq(&bfqd->lock);
4153 bfq_update_dispatch_stats(hctx->queue, rq, in_serv_queue,
4154 idle_timer_disabled);
4160 * Task holds one reference to the queue, dropped when task exits. Each rq
4161 * in-flight on this queue also holds a reference, dropped when rq is freed.
4163 * Scheduler lock must be held here. Recall not to use bfqq after calling
4164 * this function on it.
4166 void bfq_put_queue(struct bfq_queue *bfqq)
4168 #ifdef CONFIG_BFQ_GROUP_IOSCHED
4169 struct bfq_group *bfqg = bfqq_group(bfqq);
4173 bfq_log_bfqq(bfqq->bfqd, bfqq, "put_queue: %p %d",
4180 if (!hlist_unhashed(&bfqq->burst_list_node)) {
4181 hlist_del_init(&bfqq->burst_list_node);
4183 * Decrement also burst size after the removal, if the
4184 * process associated with bfqq is exiting, and thus
4185 * does not contribute to the burst any longer. This
4186 * decrement helps filter out false positives of large
4187 * bursts, when some short-lived process (often due to
4188 * the execution of commands by some service) happens
4189 * to start and exit while a complex application is
4190 * starting, and thus spawning several processes that
4191 * do I/O (and that *must not* be treated as a large
4192 * burst, see comments on bfq_handle_burst).
4194 * In particular, the decrement is performed only if:
4195 * 1) bfqq is not a merged queue, because, if it is,
4196 * then this free of bfqq is not triggered by the exit
4197 * of the process bfqq is associated with, but exactly
4198 * by the fact that bfqq has just been merged.
4199 * 2) burst_size is greater than 0, to handle
4200 * unbalanced decrements. Unbalanced decrements may
4201 * happen in te following case: bfqq is inserted into
4202 * the current burst list--without incrementing
4203 * bust_size--because of a split, but the current
4204 * burst list is not the burst list bfqq belonged to
4205 * (see comments on the case of a split in
4208 if (bfqq->bic && bfqq->bfqd->burst_size > 0)
4209 bfqq->bfqd->burst_size--;
4212 kmem_cache_free(bfq_pool, bfqq);
4213 #ifdef CONFIG_BFQ_GROUP_IOSCHED
4214 bfqg_and_blkg_put(bfqg);
4218 static void bfq_put_cooperator(struct bfq_queue *bfqq)
4220 struct bfq_queue *__bfqq, *next;
4223 * If this queue was scheduled to merge with another queue, be
4224 * sure to drop the reference taken on that queue (and others in
4225 * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs.
4227 __bfqq = bfqq->new_bfqq;
4231 next = __bfqq->new_bfqq;
4232 bfq_put_queue(__bfqq);
4237 static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq)
4239 if (bfqq == bfqd->in_service_queue) {
4240 __bfq_bfqq_expire(bfqd, bfqq);
4241 bfq_schedule_dispatch(bfqd);
4244 bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, bfqq->ref);
4246 bfq_put_cooperator(bfqq);
4248 bfq_put_queue(bfqq); /* release process reference */
4251 static void bfq_exit_icq_bfqq(struct bfq_io_cq *bic, bool is_sync)
4253 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
4254 struct bfq_data *bfqd;
4257 bfqd = bfqq->bfqd; /* NULL if scheduler already exited */
4260 unsigned long flags;
4262 spin_lock_irqsave(&bfqd->lock, flags);
4263 bfq_exit_bfqq(bfqd, bfqq);
4264 bic_set_bfqq(bic, NULL, is_sync);
4265 spin_unlock_irqrestore(&bfqd->lock, flags);
4269 static void bfq_exit_icq(struct io_cq *icq)
4271 struct bfq_io_cq *bic = icq_to_bic(icq);
4273 bfq_exit_icq_bfqq(bic, true);
4274 bfq_exit_icq_bfqq(bic, false);
4278 * Update the entity prio values; note that the new values will not
4279 * be used until the next (re)activation.
4282 bfq_set_next_ioprio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic)
4284 struct task_struct *tsk = current;
4286 struct bfq_data *bfqd = bfqq->bfqd;
4291 ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
4292 switch (ioprio_class) {
4294 dev_err(bfqq->bfqd->queue->backing_dev_info->dev,
4295 "bfq: bad prio class %d\n", ioprio_class);
4297 case IOPRIO_CLASS_NONE:
4299 * No prio set, inherit CPU scheduling settings.
4301 bfqq->new_ioprio = task_nice_ioprio(tsk);
4302 bfqq->new_ioprio_class = task_nice_ioclass(tsk);
4304 case IOPRIO_CLASS_RT:
4305 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
4306 bfqq->new_ioprio_class = IOPRIO_CLASS_RT;
4308 case IOPRIO_CLASS_BE:
4309 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
4310 bfqq->new_ioprio_class = IOPRIO_CLASS_BE;
4312 case IOPRIO_CLASS_IDLE:
4313 bfqq->new_ioprio_class = IOPRIO_CLASS_IDLE;
4314 bfqq->new_ioprio = 7;
4318 if (bfqq->new_ioprio >= IOPRIO_BE_NR) {
4319 pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n",
4321 bfqq->new_ioprio = IOPRIO_BE_NR;
4324 bfqq->entity.new_weight = bfq_ioprio_to_weight(bfqq->new_ioprio);
4325 bfqq->entity.prio_changed = 1;
4328 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
4329 struct bio *bio, bool is_sync,
4330 struct bfq_io_cq *bic);
4332 static void bfq_check_ioprio_change(struct bfq_io_cq *bic, struct bio *bio)
4334 struct bfq_data *bfqd = bic_to_bfqd(bic);
4335 struct bfq_queue *bfqq;
4336 int ioprio = bic->icq.ioc->ioprio;
4339 * This condition may trigger on a newly created bic, be sure to
4340 * drop the lock before returning.
4342 if (unlikely(!bfqd) || likely(bic->ioprio == ioprio))
4345 bic->ioprio = ioprio;
4347 bfqq = bic_to_bfqq(bic, false);
4349 /* release process reference on this queue */
4350 bfq_put_queue(bfqq);
4351 bfqq = bfq_get_queue(bfqd, bio, BLK_RW_ASYNC, bic);
4352 bic_set_bfqq(bic, bfqq, false);
4355 bfqq = bic_to_bfqq(bic, true);
4357 bfq_set_next_ioprio_data(bfqq, bic);
4360 static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4361 struct bfq_io_cq *bic, pid_t pid, int is_sync)
4363 RB_CLEAR_NODE(&bfqq->entity.rb_node);
4364 INIT_LIST_HEAD(&bfqq->fifo);
4365 INIT_HLIST_NODE(&bfqq->burst_list_node);
4371 bfq_set_next_ioprio_data(bfqq, bic);
4375 * No need to mark as has_short_ttime if in
4376 * idle_class, because no device idling is performed
4377 * for queues in idle class
4379 if (!bfq_class_idle(bfqq))
4380 /* tentatively mark as has_short_ttime */
4381 bfq_mark_bfqq_has_short_ttime(bfqq);
4382 bfq_mark_bfqq_sync(bfqq);
4383 bfq_mark_bfqq_just_created(bfqq);
4385 * Aggressively inject a lot of service: up to 90%.
4386 * This coefficient remains constant during bfqq life,
4387 * but this behavior might be changed, after enough
4388 * testing and tuning.
4390 bfqq->inject_coeff = 1;
4392 bfq_clear_bfqq_sync(bfqq);
4394 /* set end request to minus infinity from now */
4395 bfqq->ttime.last_end_request = ktime_get_ns() + 1;
4397 bfq_mark_bfqq_IO_bound(bfqq);
4401 /* Tentative initial value to trade off between thr and lat */
4402 bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3;
4403 bfqq->budget_timeout = bfq_smallest_from_now();
4406 bfqq->last_wr_start_finish = jiffies;
4407 bfqq->wr_start_at_switch_to_srt = bfq_smallest_from_now();
4408 bfqq->split_time = bfq_smallest_from_now();
4411 * To not forget the possibly high bandwidth consumed by a
4412 * process/queue in the recent past,
4413 * bfq_bfqq_softrt_next_start() returns a value at least equal
4414 * to the current value of bfqq->soft_rt_next_start (see
4415 * comments on bfq_bfqq_softrt_next_start). Set
4416 * soft_rt_next_start to now, to mean that bfqq has consumed
4417 * no bandwidth so far.
4419 bfqq->soft_rt_next_start = jiffies;
4421 /* first request is almost certainly seeky */
4422 bfqq->seek_history = 1;
4425 static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd,
4426 struct bfq_group *bfqg,
4427 int ioprio_class, int ioprio)
4429 switch (ioprio_class) {
4430 case IOPRIO_CLASS_RT:
4431 return &bfqg->async_bfqq[0][ioprio];
4432 case IOPRIO_CLASS_NONE:
4433 ioprio = IOPRIO_NORM;
4435 case IOPRIO_CLASS_BE:
4436 return &bfqg->async_bfqq[1][ioprio];
4437 case IOPRIO_CLASS_IDLE:
4438 return &bfqg->async_idle_bfqq;
4444 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
4445 struct bio *bio, bool is_sync,
4446 struct bfq_io_cq *bic)
4448 const int ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
4449 const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
4450 struct bfq_queue **async_bfqq = NULL;
4451 struct bfq_queue *bfqq;
4452 struct bfq_group *bfqg;
4456 bfqg = bfq_find_set_group(bfqd, __bio_blkcg(bio));
4458 bfqq = &bfqd->oom_bfqq;
4463 async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class,
4470 bfqq = kmem_cache_alloc_node(bfq_pool,
4471 GFP_NOWAIT | __GFP_ZERO | __GFP_NOWARN,
4475 bfq_init_bfqq(bfqd, bfqq, bic, current->pid,
4477 bfq_init_entity(&bfqq->entity, bfqg);
4478 bfq_log_bfqq(bfqd, bfqq, "allocated");
4480 bfqq = &bfqd->oom_bfqq;
4481 bfq_log_bfqq(bfqd, bfqq, "using oom bfqq");
4486 * Pin the queue now that it's allocated, scheduler exit will
4491 * Extra group reference, w.r.t. sync
4492 * queue. This extra reference is removed
4493 * only if bfqq->bfqg disappears, to
4494 * guarantee that this queue is not freed
4495 * until its group goes away.
4497 bfq_log_bfqq(bfqd, bfqq, "get_queue, bfqq not in async: %p, %d",
4503 bfqq->ref++; /* get a process reference to this queue */
4504 bfq_log_bfqq(bfqd, bfqq, "get_queue, at end: %p, %d", bfqq, bfqq->ref);
4509 static void bfq_update_io_thinktime(struct bfq_data *bfqd,
4510 struct bfq_queue *bfqq)
4512 struct bfq_ttime *ttime = &bfqq->ttime;
4513 u64 elapsed = ktime_get_ns() - bfqq->ttime.last_end_request;
4515 elapsed = min_t(u64, elapsed, 2ULL * bfqd->bfq_slice_idle);
4517 ttime->ttime_samples = (7*bfqq->ttime.ttime_samples + 256) / 8;
4518 ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed, 8);
4519 ttime->ttime_mean = div64_ul(ttime->ttime_total + 128,
4520 ttime->ttime_samples);
4524 bfq_update_io_seektime(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4527 bfqq->seek_history <<= 1;
4528 bfqq->seek_history |= BFQ_RQ_SEEKY(bfqd, bfqq->last_request_pos, rq);
4531 static void bfq_update_has_short_ttime(struct bfq_data *bfqd,
4532 struct bfq_queue *bfqq,
4533 struct bfq_io_cq *bic)
4535 bool has_short_ttime = true;
4538 * No need to update has_short_ttime if bfqq is async or in
4539 * idle io prio class, or if bfq_slice_idle is zero, because
4540 * no device idling is performed for bfqq in this case.
4542 if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq) ||
4543 bfqd->bfq_slice_idle == 0)
4546 /* Idle window just restored, statistics are meaningless. */
4547 if (time_is_after_eq_jiffies(bfqq->split_time +
4548 bfqd->bfq_wr_min_idle_time))
4551 /* Think time is infinite if no process is linked to
4552 * bfqq. Otherwise check average think time to
4553 * decide whether to mark as has_short_ttime
4555 if (atomic_read(&bic->icq.ioc->active_ref) == 0 ||
4556 (bfq_sample_valid(bfqq->ttime.ttime_samples) &&
4557 bfqq->ttime.ttime_mean > bfqd->bfq_slice_idle))
4558 has_short_ttime = false;
4560 bfq_log_bfqq(bfqd, bfqq, "update_has_short_ttime: has_short_ttime %d",
4563 if (has_short_ttime)
4564 bfq_mark_bfqq_has_short_ttime(bfqq);
4566 bfq_clear_bfqq_has_short_ttime(bfqq);
4570 * Called when a new fs request (rq) is added to bfqq. Check if there's
4571 * something we should do about it.
4573 static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4576 struct bfq_io_cq *bic = RQ_BIC(rq);
4578 if (rq->cmd_flags & REQ_META)
4579 bfqq->meta_pending++;
4581 bfq_update_io_thinktime(bfqd, bfqq);
4582 bfq_update_has_short_ttime(bfqd, bfqq, bic);
4583 bfq_update_io_seektime(bfqd, bfqq, rq);
4585 bfq_log_bfqq(bfqd, bfqq,
4586 "rq_enqueued: has_short_ttime=%d (seeky %d)",
4587 bfq_bfqq_has_short_ttime(bfqq), BFQQ_SEEKY(bfqq));
4589 bfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
4591 if (bfqq == bfqd->in_service_queue && bfq_bfqq_wait_request(bfqq)) {
4592 bool small_req = bfqq->queued[rq_is_sync(rq)] == 1 &&
4593 blk_rq_sectors(rq) < 32;
4594 bool budget_timeout = bfq_bfqq_budget_timeout(bfqq);
4597 * There is just this request queued: if
4598 * - the request is small, and
4599 * - we are idling to boost throughput, and
4600 * - the queue is not to be expired,
4603 * In this way, if the device is being idled to wait
4604 * for a new request from the in-service queue, we
4605 * avoid unplugging the device and committing the
4606 * device to serve just a small request. In contrast
4607 * we wait for the block layer to decide when to
4608 * unplug the device: hopefully, new requests will be
4609 * merged to this one quickly, then the device will be
4610 * unplugged and larger requests will be dispatched.
4612 if (small_req && idling_boosts_thr_without_issues(bfqd, bfqq) &&
4617 * A large enough request arrived, or idling is being
4618 * performed to preserve service guarantees, or
4619 * finally the queue is to be expired: in all these
4620 * cases disk idling is to be stopped, so clear
4621 * wait_request flag and reset timer.
4623 bfq_clear_bfqq_wait_request(bfqq);
4624 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
4627 * The queue is not empty, because a new request just
4628 * arrived. Hence we can safely expire the queue, in
4629 * case of budget timeout, without risking that the
4630 * timestamps of the queue are not updated correctly.
4631 * See [1] for more details.
4634 bfq_bfqq_expire(bfqd, bfqq, false,
4635 BFQQE_BUDGET_TIMEOUT);
4639 /* returns true if it causes the idle timer to be disabled */
4640 static bool __bfq_insert_request(struct bfq_data *bfqd, struct request *rq)
4642 struct bfq_queue *bfqq = RQ_BFQQ(rq),
4643 *new_bfqq = bfq_setup_cooperator(bfqd, bfqq, rq, true);
4644 bool waiting, idle_timer_disabled = false;
4648 * Release the request's reference to the old bfqq
4649 * and make sure one is taken to the shared queue.
4651 new_bfqq->allocated++;
4655 * If the bic associated with the process
4656 * issuing this request still points to bfqq
4657 * (and thus has not been already redirected
4658 * to new_bfqq or even some other bfq_queue),
4659 * then complete the merge and redirect it to
4662 if (bic_to_bfqq(RQ_BIC(rq), 1) == bfqq)
4663 bfq_merge_bfqqs(bfqd, RQ_BIC(rq),
4666 bfq_clear_bfqq_just_created(bfqq);
4668 * rq is about to be enqueued into new_bfqq,
4669 * release rq reference on bfqq
4671 bfq_put_queue(bfqq);
4672 rq->elv.priv[1] = new_bfqq;
4676 waiting = bfqq && bfq_bfqq_wait_request(bfqq);
4677 bfq_add_request(rq);
4678 idle_timer_disabled = waiting && !bfq_bfqq_wait_request(bfqq);
4680 rq->fifo_time = ktime_get_ns() + bfqd->bfq_fifo_expire[rq_is_sync(rq)];
4681 list_add_tail(&rq->queuelist, &bfqq->fifo);
4683 bfq_rq_enqueued(bfqd, bfqq, rq);
4685 return idle_timer_disabled;
4688 #if defined(CONFIG_BFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
4689 static void bfq_update_insert_stats(struct request_queue *q,
4690 struct bfq_queue *bfqq,
4691 bool idle_timer_disabled,
4692 unsigned int cmd_flags)
4698 * bfqq still exists, because it can disappear only after
4699 * either it is merged with another queue, or the process it
4700 * is associated with exits. But both actions must be taken by
4701 * the same process currently executing this flow of
4704 * In addition, the following queue lock guarantees that
4705 * bfqq_group(bfqq) exists as well.
4707 spin_lock_irq(&q->queue_lock);
4708 bfqg_stats_update_io_add(bfqq_group(bfqq), bfqq, cmd_flags);
4709 if (idle_timer_disabled)
4710 bfqg_stats_update_idle_time(bfqq_group(bfqq));
4711 spin_unlock_irq(&q->queue_lock);
4714 static inline void bfq_update_insert_stats(struct request_queue *q,
4715 struct bfq_queue *bfqq,
4716 bool idle_timer_disabled,
4717 unsigned int cmd_flags) {}
4720 static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
4723 struct request_queue *q = hctx->queue;
4724 struct bfq_data *bfqd = q->elevator->elevator_data;
4725 struct bfq_queue *bfqq;
4726 bool idle_timer_disabled = false;
4727 unsigned int cmd_flags;
4729 spin_lock_irq(&bfqd->lock);
4730 if (blk_mq_sched_try_insert_merge(q, rq)) {
4731 spin_unlock_irq(&bfqd->lock);
4735 spin_unlock_irq(&bfqd->lock);
4737 blk_mq_sched_request_inserted(rq);
4739 spin_lock_irq(&bfqd->lock);
4740 bfqq = bfq_init_rq(rq);
4741 if (at_head || blk_rq_is_passthrough(rq)) {
4743 list_add(&rq->queuelist, &bfqd->dispatch);
4745 list_add_tail(&rq->queuelist, &bfqd->dispatch);
4746 } else { /* bfqq is assumed to be non null here */
4747 idle_timer_disabled = __bfq_insert_request(bfqd, rq);
4749 * Update bfqq, because, if a queue merge has occurred
4750 * in __bfq_insert_request, then rq has been
4751 * redirected into a new queue.
4755 if (rq_mergeable(rq)) {
4756 elv_rqhash_add(q, rq);
4763 * Cache cmd_flags before releasing scheduler lock, because rq
4764 * may disappear afterwards (for example, because of a request
4767 cmd_flags = rq->cmd_flags;
4769 spin_unlock_irq(&bfqd->lock);
4771 bfq_update_insert_stats(q, bfqq, idle_timer_disabled,
4775 static void bfq_insert_requests(struct blk_mq_hw_ctx *hctx,
4776 struct list_head *list, bool at_head)
4778 while (!list_empty(list)) {
4781 rq = list_first_entry(list, struct request, queuelist);
4782 list_del_init(&rq->queuelist);
4783 bfq_insert_request(hctx, rq, at_head);
4787 static void bfq_update_hw_tag(struct bfq_data *bfqd)
4789 bfqd->max_rq_in_driver = max_t(int, bfqd->max_rq_in_driver,
4790 bfqd->rq_in_driver);
4792 if (bfqd->hw_tag == 1)
4796 * This sample is valid if the number of outstanding requests
4797 * is large enough to allow a queueing behavior. Note that the
4798 * sum is not exact, as it's not taking into account deactivated
4801 if (bfqd->rq_in_driver + bfqd->queued < BFQ_HW_QUEUE_THRESHOLD)
4804 if (bfqd->hw_tag_samples++ < BFQ_HW_QUEUE_SAMPLES)
4807 bfqd->hw_tag = bfqd->max_rq_in_driver > BFQ_HW_QUEUE_THRESHOLD;
4808 bfqd->max_rq_in_driver = 0;
4809 bfqd->hw_tag_samples = 0;
4812 static void bfq_completed_request(struct bfq_queue *bfqq, struct bfq_data *bfqd)
4817 bfq_update_hw_tag(bfqd);
4819 bfqd->rq_in_driver--;
4822 if (!bfqq->dispatched && !bfq_bfqq_busy(bfqq)) {
4824 * Set budget_timeout (which we overload to store the
4825 * time at which the queue remains with no backlog and
4826 * no outstanding request; used by the weight-raising
4829 bfqq->budget_timeout = jiffies;
4831 bfq_weights_tree_remove(bfqd, bfqq);
4834 now_ns = ktime_get_ns();
4836 bfqq->ttime.last_end_request = now_ns;
4839 * Using us instead of ns, to get a reasonable precision in
4840 * computing rate in next check.
4842 delta_us = div_u64(now_ns - bfqd->last_completion, NSEC_PER_USEC);
4845 * If the request took rather long to complete, and, according
4846 * to the maximum request size recorded, this completion latency
4847 * implies that the request was certainly served at a very low
4848 * rate (less than 1M sectors/sec), then the whole observation
4849 * interval that lasts up to this time instant cannot be a
4850 * valid time interval for computing a new peak rate. Invoke
4851 * bfq_update_rate_reset to have the following three steps
4853 * - close the observation interval at the last (previous)
4854 * request dispatch or completion
4855 * - compute rate, if possible, for that observation interval
4856 * - reset to zero samples, which will trigger a proper
4857 * re-initialization of the observation interval on next
4860 if (delta_us > BFQ_MIN_TT/NSEC_PER_USEC &&
4861 (bfqd->last_rq_max_size<<BFQ_RATE_SHIFT)/delta_us <
4862 1UL<<(BFQ_RATE_SHIFT - 10))
4863 bfq_update_rate_reset(bfqd, NULL);
4864 bfqd->last_completion = now_ns;
4867 * If we are waiting to discover whether the request pattern
4868 * of the task associated with the queue is actually
4869 * isochronous, and both requisites for this condition to hold
4870 * are now satisfied, then compute soft_rt_next_start (see the
4871 * comments on the function bfq_bfqq_softrt_next_start()). We
4872 * do not compute soft_rt_next_start if bfqq is in interactive
4873 * weight raising (see the comments in bfq_bfqq_expire() for
4874 * an explanation). We schedule this delayed update when bfqq
4875 * expires, if it still has in-flight requests.
4877 if (bfq_bfqq_softrt_update(bfqq) && bfqq->dispatched == 0 &&
4878 RB_EMPTY_ROOT(&bfqq->sort_list) &&
4879 bfqq->wr_coeff != bfqd->bfq_wr_coeff)
4880 bfqq->soft_rt_next_start =
4881 bfq_bfqq_softrt_next_start(bfqd, bfqq);
4884 * If this is the in-service queue, check if it needs to be expired,
4885 * or if we want to idle in case it has no pending requests.
4887 if (bfqd->in_service_queue == bfqq) {
4888 if (bfq_bfqq_must_idle(bfqq)) {
4889 if (bfqq->dispatched == 0)
4890 bfq_arm_slice_timer(bfqd);
4892 * If we get here, we do not expire bfqq, even
4893 * if bfqq was in budget timeout or had no
4894 * more requests (as controlled in the next
4895 * conditional instructions). The reason for
4896 * not expiring bfqq is as follows.
4898 * Here bfqq->dispatched > 0 holds, but
4899 * bfq_bfqq_must_idle() returned true. This
4900 * implies that, even if no request arrives
4901 * for bfqq before bfqq->dispatched reaches 0,
4902 * bfqq will, however, not be expired on the
4903 * completion event that causes bfqq->dispatch
4904 * to reach zero. In contrast, on this event,
4905 * bfqq will start enjoying device idling
4906 * (I/O-dispatch plugging).
4908 * But, if we expired bfqq here, bfqq would
4909 * not have the chance to enjoy device idling
4910 * when bfqq->dispatched finally reaches
4911 * zero. This would expose bfqq to violation
4912 * of its reserved service guarantees.
4915 } else if (bfq_may_expire_for_budg_timeout(bfqq))
4916 bfq_bfqq_expire(bfqd, bfqq, false,
4917 BFQQE_BUDGET_TIMEOUT);
4918 else if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
4919 (bfqq->dispatched == 0 ||
4920 !bfq_better_to_idle(bfqq)))
4921 bfq_bfqq_expire(bfqd, bfqq, false,
4922 BFQQE_NO_MORE_REQUESTS);
4925 if (!bfqd->rq_in_driver)
4926 bfq_schedule_dispatch(bfqd);
4929 static void bfq_finish_requeue_request_body(struct bfq_queue *bfqq)
4933 bfq_put_queue(bfqq);
4937 * Handle either a requeue or a finish for rq. The things to do are
4938 * the same in both cases: all references to rq are to be dropped. In
4939 * particular, rq is considered completed from the point of view of
4942 static void bfq_finish_requeue_request(struct request *rq)
4944 struct bfq_queue *bfqq = RQ_BFQQ(rq);
4945 struct bfq_data *bfqd;
4948 * Requeue and finish hooks are invoked in blk-mq without
4949 * checking whether the involved request is actually still
4950 * referenced in the scheduler. To handle this fact, the
4951 * following two checks make this function exit in case of
4952 * spurious invocations, for which there is nothing to do.
4954 * First, check whether rq has nothing to do with an elevator.
4956 if (unlikely(!(rq->rq_flags & RQF_ELVPRIV)))
4960 * rq either is not associated with any icq, or is an already
4961 * requeued request that has not (yet) been re-inserted into
4964 if (!rq->elv.icq || !bfqq)
4969 if (rq->rq_flags & RQF_STARTED)
4970 bfqg_stats_update_completion(bfqq_group(bfqq),
4972 rq->io_start_time_ns,
4975 if (likely(rq->rq_flags & RQF_STARTED)) {
4976 unsigned long flags;
4978 spin_lock_irqsave(&bfqd->lock, flags);
4980 bfq_completed_request(bfqq, bfqd);
4981 bfq_finish_requeue_request_body(bfqq);
4983 spin_unlock_irqrestore(&bfqd->lock, flags);
4986 * Request rq may be still/already in the scheduler,
4987 * in which case we need to remove it (this should
4988 * never happen in case of requeue). And we cannot
4989 * defer such a check and removal, to avoid
4990 * inconsistencies in the time interval from the end
4991 * of this function to the start of the deferred work.
4992 * This situation seems to occur only in process
4993 * context, as a consequence of a merge. In the
4994 * current version of the code, this implies that the
4998 if (!RB_EMPTY_NODE(&rq->rb_node)) {
4999 bfq_remove_request(rq->q, rq);
5000 bfqg_stats_update_io_remove(bfqq_group(bfqq),
5003 bfq_finish_requeue_request_body(bfqq);
5007 * Reset private fields. In case of a requeue, this allows
5008 * this function to correctly do nothing if it is spuriously
5009 * invoked again on this same request (see the check at the
5010 * beginning of the function). Probably, a better general
5011 * design would be to prevent blk-mq from invoking the requeue
5012 * or finish hooks of an elevator, for a request that is not
5013 * referred by that elevator.
5015 * Resetting the following fields would break the
5016 * request-insertion logic if rq is re-inserted into a bfq
5017 * internal queue, without a re-preparation. Here we assume
5018 * that re-insertions of requeued requests, without
5019 * re-preparation, can happen only for pass_through or at_head
5020 * requests (which are not re-inserted into bfq internal
5023 rq->elv.priv[0] = NULL;
5024 rq->elv.priv[1] = NULL;
5028 * Returns NULL if a new bfqq should be allocated, or the old bfqq if this
5029 * was the last process referring to that bfqq.
5031 static struct bfq_queue *
5032 bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq)
5034 bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue");
5036 if (bfqq_process_refs(bfqq) == 1) {
5037 bfqq->pid = current->pid;
5038 bfq_clear_bfqq_coop(bfqq);
5039 bfq_clear_bfqq_split_coop(bfqq);
5043 bic_set_bfqq(bic, NULL, 1);
5045 bfq_put_cooperator(bfqq);
5047 bfq_put_queue(bfqq);
5051 static struct bfq_queue *bfq_get_bfqq_handle_split(struct bfq_data *bfqd,
5052 struct bfq_io_cq *bic,
5054 bool split, bool is_sync,
5057 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
5059 if (likely(bfqq && bfqq != &bfqd->oom_bfqq))
5066 bfq_put_queue(bfqq);
5067 bfqq = bfq_get_queue(bfqd, bio, is_sync, bic);
5069 bic_set_bfqq(bic, bfqq, is_sync);
5070 if (split && is_sync) {
5071 if ((bic->was_in_burst_list && bfqd->large_burst) ||
5072 bic->saved_in_large_burst)
5073 bfq_mark_bfqq_in_large_burst(bfqq);
5075 bfq_clear_bfqq_in_large_burst(bfqq);
5076 if (bic->was_in_burst_list)
5078 * If bfqq was in the current
5079 * burst list before being
5080 * merged, then we have to add
5081 * it back. And we do not need
5082 * to increase burst_size, as
5083 * we did not decrement
5084 * burst_size when we removed
5085 * bfqq from the burst list as
5086 * a consequence of a merge
5088 * bfq_put_queue). In this
5089 * respect, it would be rather
5090 * costly to know whether the
5091 * current burst list is still
5092 * the same burst list from
5093 * which bfqq was removed on
5094 * the merge. To avoid this
5095 * cost, if bfqq was in a
5096 * burst list, then we add
5097 * bfqq to the current burst
5098 * list without any further
5099 * check. This can cause
5100 * inappropriate insertions,
5101 * but rarely enough to not
5102 * harm the detection of large
5103 * bursts significantly.
5105 hlist_add_head(&bfqq->burst_list_node,
5108 bfqq->split_time = jiffies;
5115 * Only reset private fields. The actual request preparation will be
5116 * performed by bfq_init_rq, when rq is either inserted or merged. See
5117 * comments on bfq_init_rq for the reason behind this delayed
5120 static void bfq_prepare_request(struct request *rq, struct bio *bio)
5123 * Regardless of whether we have an icq attached, we have to
5124 * clear the scheduler pointers, as they might point to
5125 * previously allocated bic/bfqq structs.
5127 rq->elv.priv[0] = rq->elv.priv[1] = NULL;
5131 * If needed, init rq, allocate bfq data structures associated with
5132 * rq, and increment reference counters in the destination bfq_queue
5133 * for rq. Return the destination bfq_queue for rq, or NULL is rq is
5134 * not associated with any bfq_queue.
5136 * This function is invoked by the functions that perform rq insertion
5137 * or merging. One may have expected the above preparation operations
5138 * to be performed in bfq_prepare_request, and not delayed to when rq
5139 * is inserted or merged. The rationale behind this delayed
5140 * preparation is that, after the prepare_request hook is invoked for
5141 * rq, rq may still be transformed into a request with no icq, i.e., a
5142 * request not associated with any queue. No bfq hook is invoked to
5143 * signal this tranformation. As a consequence, should these
5144 * preparation operations be performed when the prepare_request hook
5145 * is invoked, and should rq be transformed one moment later, bfq
5146 * would end up in an inconsistent state, because it would have
5147 * incremented some queue counters for an rq destined to
5148 * transformation, without any chance to correctly lower these
5149 * counters back. In contrast, no transformation can still happen for
5150 * rq after rq has been inserted or merged. So, it is safe to execute
5151 * these preparation operations when rq is finally inserted or merged.
5153 static struct bfq_queue *bfq_init_rq(struct request *rq)
5155 struct request_queue *q = rq->q;
5156 struct bio *bio = rq->bio;
5157 struct bfq_data *bfqd = q->elevator->elevator_data;
5158 struct bfq_io_cq *bic;
5159 const int is_sync = rq_is_sync(rq);
5160 struct bfq_queue *bfqq;
5161 bool new_queue = false;
5162 bool bfqq_already_existing = false, split = false;
5164 if (unlikely(!rq->elv.icq))
5168 * Assuming that elv.priv[1] is set only if everything is set
5169 * for this rq. This holds true, because this function is
5170 * invoked only for insertion or merging, and, after such
5171 * events, a request cannot be manipulated any longer before
5172 * being removed from bfq.
5174 if (rq->elv.priv[1])
5175 return rq->elv.priv[1];
5177 bic = icq_to_bic(rq->elv.icq);
5179 bfq_check_ioprio_change(bic, bio);
5181 bfq_bic_update_cgroup(bic, bio);
5183 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio, false, is_sync,
5186 if (likely(!new_queue)) {
5187 /* If the queue was seeky for too long, break it apart. */
5188 if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq)) {
5189 bfq_log_bfqq(bfqd, bfqq, "breaking apart bfqq");
5191 /* Update bic before losing reference to bfqq */
5192 if (bfq_bfqq_in_large_burst(bfqq))
5193 bic->saved_in_large_burst = true;
5195 bfqq = bfq_split_bfqq(bic, bfqq);
5199 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio,
5203 bfqq_already_existing = true;
5209 bfq_log_bfqq(bfqd, bfqq, "get_request %p: bfqq %p, %d",
5210 rq, bfqq, bfqq->ref);
5212 rq->elv.priv[0] = bic;
5213 rq->elv.priv[1] = bfqq;
5216 * If a bfq_queue has only one process reference, it is owned
5217 * by only this bic: we can then set bfqq->bic = bic. in
5218 * addition, if the queue has also just been split, we have to
5221 if (likely(bfqq != &bfqd->oom_bfqq) && bfqq_process_refs(bfqq) == 1) {
5225 * The queue has just been split from a shared
5226 * queue: restore the idle window and the
5227 * possible weight raising period.
5229 bfq_bfqq_resume_state(bfqq, bfqd, bic,
5230 bfqq_already_existing);
5234 if (unlikely(bfq_bfqq_just_created(bfqq)))
5235 bfq_handle_burst(bfqd, bfqq);
5240 static void bfq_idle_slice_timer_body(struct bfq_queue *bfqq)
5242 struct bfq_data *bfqd = bfqq->bfqd;
5243 enum bfqq_expiration reason;
5244 unsigned long flags;
5246 spin_lock_irqsave(&bfqd->lock, flags);
5247 bfq_clear_bfqq_wait_request(bfqq);
5249 if (bfqq != bfqd->in_service_queue) {
5250 spin_unlock_irqrestore(&bfqd->lock, flags);
5254 if (bfq_bfqq_budget_timeout(bfqq))
5256 * Also here the queue can be safely expired
5257 * for budget timeout without wasting
5260 reason = BFQQE_BUDGET_TIMEOUT;
5261 else if (bfqq->queued[0] == 0 && bfqq->queued[1] == 0)
5263 * The queue may not be empty upon timer expiration,
5264 * because we may not disable the timer when the
5265 * first request of the in-service queue arrives
5266 * during disk idling.
5268 reason = BFQQE_TOO_IDLE;
5270 goto schedule_dispatch;
5272 bfq_bfqq_expire(bfqd, bfqq, true, reason);
5275 spin_unlock_irqrestore(&bfqd->lock, flags);
5276 bfq_schedule_dispatch(bfqd);
5280 * Handler of the expiration of the timer running if the in-service queue
5281 * is idling inside its time slice.
5283 static enum hrtimer_restart bfq_idle_slice_timer(struct hrtimer *timer)
5285 struct bfq_data *bfqd = container_of(timer, struct bfq_data,
5287 struct bfq_queue *bfqq = bfqd->in_service_queue;
5290 * Theoretical race here: the in-service queue can be NULL or
5291 * different from the queue that was idling if a new request
5292 * arrives for the current queue and there is a full dispatch
5293 * cycle that changes the in-service queue. This can hardly
5294 * happen, but in the worst case we just expire a queue too
5298 bfq_idle_slice_timer_body(bfqq);
5300 return HRTIMER_NORESTART;
5303 static void __bfq_put_async_bfqq(struct bfq_data *bfqd,
5304 struct bfq_queue **bfqq_ptr)
5306 struct bfq_queue *bfqq = *bfqq_ptr;
5308 bfq_log(bfqd, "put_async_bfqq: %p", bfqq);
5310 bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
5312 bfq_log_bfqq(bfqd, bfqq, "put_async_bfqq: putting %p, %d",
5314 bfq_put_queue(bfqq);
5320 * Release all the bfqg references to its async queues. If we are
5321 * deallocating the group these queues may still contain requests, so
5322 * we reparent them to the root cgroup (i.e., the only one that will
5323 * exist for sure until all the requests on a device are gone).
5325 void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg)
5329 for (i = 0; i < 2; i++)
5330 for (j = 0; j < IOPRIO_BE_NR; j++)
5331 __bfq_put_async_bfqq(bfqd, &bfqg->async_bfqq[i][j]);
5333 __bfq_put_async_bfqq(bfqd, &bfqg->async_idle_bfqq);
5337 * See the comments on bfq_limit_depth for the purpose of
5338 * the depths set in the function. Return minimum shallow depth we'll use.
5340 static unsigned int bfq_update_depths(struct bfq_data *bfqd,
5341 struct sbitmap_queue *bt)
5343 unsigned int i, j, min_shallow = UINT_MAX;
5346 * In-word depths if no bfq_queue is being weight-raised:
5347 * leaving 25% of tags only for sync reads.
5349 * In next formulas, right-shift the value
5350 * (1U<<bt->sb.shift), instead of computing directly
5351 * (1U<<(bt->sb.shift - something)), to be robust against
5352 * any possible value of bt->sb.shift, without having to
5353 * limit 'something'.
5355 /* no more than 50% of tags for async I/O */
5356 bfqd->word_depths[0][0] = max((1U << bt->sb.shift) >> 1, 1U);
5358 * no more than 75% of tags for sync writes (25% extra tags
5359 * w.r.t. async I/O, to prevent async I/O from starving sync
5362 bfqd->word_depths[0][1] = max(((1U << bt->sb.shift) * 3) >> 2, 1U);
5365 * In-word depths in case some bfq_queue is being weight-
5366 * raised: leaving ~63% of tags for sync reads. This is the
5367 * highest percentage for which, in our tests, application
5368 * start-up times didn't suffer from any regression due to tag
5371 /* no more than ~18% of tags for async I/O */
5372 bfqd->word_depths[1][0] = max(((1U << bt->sb.shift) * 3) >> 4, 1U);
5373 /* no more than ~37% of tags for sync writes (~20% extra tags) */
5374 bfqd->word_depths[1][1] = max(((1U << bt->sb.shift) * 6) >> 4, 1U);
5376 for (i = 0; i < 2; i++)
5377 for (j = 0; j < 2; j++)
5378 min_shallow = min(min_shallow, bfqd->word_depths[i][j]);
5383 static int bfq_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int index)
5385 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5386 struct blk_mq_tags *tags = hctx->sched_tags;
5387 unsigned int min_shallow;
5389 min_shallow = bfq_update_depths(bfqd, &tags->bitmap_tags);
5390 sbitmap_queue_min_shallow_depth(&tags->bitmap_tags, min_shallow);
5394 static void bfq_exit_queue(struct elevator_queue *e)
5396 struct bfq_data *bfqd = e->elevator_data;
5397 struct bfq_queue *bfqq, *n;
5399 hrtimer_cancel(&bfqd->idle_slice_timer);
5401 spin_lock_irq(&bfqd->lock);
5402 list_for_each_entry_safe(bfqq, n, &bfqd->idle_list, bfqq_list)
5403 bfq_deactivate_bfqq(bfqd, bfqq, false, false);
5404 spin_unlock_irq(&bfqd->lock);
5406 hrtimer_cancel(&bfqd->idle_slice_timer);
5408 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5409 /* release oom-queue reference to root group */
5410 bfqg_and_blkg_put(bfqd->root_group);
5412 blkcg_deactivate_policy(bfqd->queue, &blkcg_policy_bfq);
5414 spin_lock_irq(&bfqd->lock);
5415 bfq_put_async_queues(bfqd, bfqd->root_group);
5416 kfree(bfqd->root_group);
5417 spin_unlock_irq(&bfqd->lock);
5423 static void bfq_init_root_group(struct bfq_group *root_group,
5424 struct bfq_data *bfqd)
5428 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5429 root_group->entity.parent = NULL;
5430 root_group->my_entity = NULL;
5431 root_group->bfqd = bfqd;
5433 root_group->rq_pos_tree = RB_ROOT;
5434 for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
5435 root_group->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
5436 root_group->sched_data.bfq_class_idle_last_service = jiffies;
5439 static int bfq_init_queue(struct request_queue *q, struct elevator_type *e)
5441 struct bfq_data *bfqd;
5442 struct elevator_queue *eq;
5444 eq = elevator_alloc(q, e);
5448 bfqd = kzalloc_node(sizeof(*bfqd), GFP_KERNEL, q->node);
5450 kobject_put(&eq->kobj);
5453 eq->elevator_data = bfqd;
5455 spin_lock_irq(&q->queue_lock);
5457 spin_unlock_irq(&q->queue_lock);
5460 * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues.
5461 * Grab a permanent reference to it, so that the normal code flow
5462 * will not attempt to free it.
5464 bfq_init_bfqq(bfqd, &bfqd->oom_bfqq, NULL, 1, 0);
5465 bfqd->oom_bfqq.ref++;
5466 bfqd->oom_bfqq.new_ioprio = BFQ_DEFAULT_QUEUE_IOPRIO;
5467 bfqd->oom_bfqq.new_ioprio_class = IOPRIO_CLASS_BE;
5468 bfqd->oom_bfqq.entity.new_weight =
5469 bfq_ioprio_to_weight(bfqd->oom_bfqq.new_ioprio);
5471 /* oom_bfqq does not participate to bursts */
5472 bfq_clear_bfqq_just_created(&bfqd->oom_bfqq);
5475 * Trigger weight initialization, according to ioprio, at the
5476 * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio
5477 * class won't be changed any more.
5479 bfqd->oom_bfqq.entity.prio_changed = 1;
5483 INIT_LIST_HEAD(&bfqd->dispatch);
5485 hrtimer_init(&bfqd->idle_slice_timer, CLOCK_MONOTONIC,
5487 bfqd->idle_slice_timer.function = bfq_idle_slice_timer;
5489 bfqd->queue_weights_tree = RB_ROOT;
5490 bfqd->num_groups_with_pending_reqs = 0;
5492 INIT_LIST_HEAD(&bfqd->active_list);
5493 INIT_LIST_HEAD(&bfqd->idle_list);
5494 INIT_HLIST_HEAD(&bfqd->burst_list);
5498 bfqd->bfq_max_budget = bfq_default_max_budget;
5500 bfqd->bfq_fifo_expire[0] = bfq_fifo_expire[0];
5501 bfqd->bfq_fifo_expire[1] = bfq_fifo_expire[1];
5502 bfqd->bfq_back_max = bfq_back_max;
5503 bfqd->bfq_back_penalty = bfq_back_penalty;
5504 bfqd->bfq_slice_idle = bfq_slice_idle;
5505 bfqd->bfq_timeout = bfq_timeout;
5507 bfqd->bfq_requests_within_timer = 120;
5509 bfqd->bfq_large_burst_thresh = 8;
5510 bfqd->bfq_burst_interval = msecs_to_jiffies(180);
5512 bfqd->low_latency = true;
5515 * Trade-off between responsiveness and fairness.
5517 bfqd->bfq_wr_coeff = 30;
5518 bfqd->bfq_wr_rt_max_time = msecs_to_jiffies(300);
5519 bfqd->bfq_wr_max_time = 0;
5520 bfqd->bfq_wr_min_idle_time = msecs_to_jiffies(2000);
5521 bfqd->bfq_wr_min_inter_arr_async = msecs_to_jiffies(500);
5522 bfqd->bfq_wr_max_softrt_rate = 7000; /*
5523 * Approximate rate required
5524 * to playback or record a
5525 * high-definition compressed
5528 bfqd->wr_busy_queues = 0;
5531 * Begin by assuming, optimistically, that the device peak
5532 * rate is equal to 2/3 of the highest reference rate.
5534 bfqd->rate_dur_prod = ref_rate[blk_queue_nonrot(bfqd->queue)] *
5535 ref_wr_duration[blk_queue_nonrot(bfqd->queue)];
5536 bfqd->peak_rate = ref_rate[blk_queue_nonrot(bfqd->queue)] * 2 / 3;
5538 spin_lock_init(&bfqd->lock);
5541 * The invocation of the next bfq_create_group_hierarchy
5542 * function is the head of a chain of function calls
5543 * (bfq_create_group_hierarchy->blkcg_activate_policy->
5544 * blk_mq_freeze_queue) that may lead to the invocation of the
5545 * has_work hook function. For this reason,
5546 * bfq_create_group_hierarchy is invoked only after all
5547 * scheduler data has been initialized, apart from the fields
5548 * that can be initialized only after invoking
5549 * bfq_create_group_hierarchy. This, in particular, enables
5550 * has_work to correctly return false. Of course, to avoid
5551 * other inconsistencies, the blk-mq stack must then refrain
5552 * from invoking further scheduler hooks before this init
5553 * function is finished.
5555 bfqd->root_group = bfq_create_group_hierarchy(bfqd, q->node);
5556 if (!bfqd->root_group)
5558 bfq_init_root_group(bfqd->root_group, bfqd);
5559 bfq_init_entity(&bfqd->oom_bfqq.entity, bfqd->root_group);
5561 wbt_disable_default(q);
5566 kobject_put(&eq->kobj);
5570 static void bfq_slab_kill(void)
5572 kmem_cache_destroy(bfq_pool);
5575 static int __init bfq_slab_setup(void)
5577 bfq_pool = KMEM_CACHE(bfq_queue, 0);
5583 static ssize_t bfq_var_show(unsigned int var, char *page)
5585 return sprintf(page, "%u\n", var);
5588 static int bfq_var_store(unsigned long *var, const char *page)
5590 unsigned long new_val;
5591 int ret = kstrtoul(page, 10, &new_val);
5599 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
5600 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
5602 struct bfq_data *bfqd = e->elevator_data; \
5603 u64 __data = __VAR; \
5605 __data = jiffies_to_msecs(__data); \
5606 else if (__CONV == 2) \
5607 __data = div_u64(__data, NSEC_PER_MSEC); \
5608 return bfq_var_show(__data, (page)); \
5610 SHOW_FUNCTION(bfq_fifo_expire_sync_show, bfqd->bfq_fifo_expire[1], 2);
5611 SHOW_FUNCTION(bfq_fifo_expire_async_show, bfqd->bfq_fifo_expire[0], 2);
5612 SHOW_FUNCTION(bfq_back_seek_max_show, bfqd->bfq_back_max, 0);
5613 SHOW_FUNCTION(bfq_back_seek_penalty_show, bfqd->bfq_back_penalty, 0);
5614 SHOW_FUNCTION(bfq_slice_idle_show, bfqd->bfq_slice_idle, 2);
5615 SHOW_FUNCTION(bfq_max_budget_show, bfqd->bfq_user_max_budget, 0);
5616 SHOW_FUNCTION(bfq_timeout_sync_show, bfqd->bfq_timeout, 1);
5617 SHOW_FUNCTION(bfq_strict_guarantees_show, bfqd->strict_guarantees, 0);
5618 SHOW_FUNCTION(bfq_low_latency_show, bfqd->low_latency, 0);
5619 #undef SHOW_FUNCTION
5621 #define USEC_SHOW_FUNCTION(__FUNC, __VAR) \
5622 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
5624 struct bfq_data *bfqd = e->elevator_data; \
5625 u64 __data = __VAR; \
5626 __data = div_u64(__data, NSEC_PER_USEC); \
5627 return bfq_var_show(__data, (page)); \
5629 USEC_SHOW_FUNCTION(bfq_slice_idle_us_show, bfqd->bfq_slice_idle);
5630 #undef USEC_SHOW_FUNCTION
5632 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
5634 __FUNC(struct elevator_queue *e, const char *page, size_t count) \
5636 struct bfq_data *bfqd = e->elevator_data; \
5637 unsigned long __data, __min = (MIN), __max = (MAX); \
5640 ret = bfq_var_store(&__data, (page)); \
5643 if (__data < __min) \
5645 else if (__data > __max) \
5648 *(__PTR) = msecs_to_jiffies(__data); \
5649 else if (__CONV == 2) \
5650 *(__PTR) = (u64)__data * NSEC_PER_MSEC; \
5652 *(__PTR) = __data; \
5655 STORE_FUNCTION(bfq_fifo_expire_sync_store, &bfqd->bfq_fifo_expire[1], 1,
5657 STORE_FUNCTION(bfq_fifo_expire_async_store, &bfqd->bfq_fifo_expire[0], 1,
5659 STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0);
5660 STORE_FUNCTION(bfq_back_seek_penalty_store, &bfqd->bfq_back_penalty, 1,
5662 STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 2);
5663 #undef STORE_FUNCTION
5665 #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
5666 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\
5668 struct bfq_data *bfqd = e->elevator_data; \
5669 unsigned long __data, __min = (MIN), __max = (MAX); \
5672 ret = bfq_var_store(&__data, (page)); \
5675 if (__data < __min) \
5677 else if (__data > __max) \
5679 *(__PTR) = (u64)__data * NSEC_PER_USEC; \
5682 USEC_STORE_FUNCTION(bfq_slice_idle_us_store, &bfqd->bfq_slice_idle, 0,
5684 #undef USEC_STORE_FUNCTION
5686 static ssize_t bfq_max_budget_store(struct elevator_queue *e,
5687 const char *page, size_t count)
5689 struct bfq_data *bfqd = e->elevator_data;
5690 unsigned long __data;
5693 ret = bfq_var_store(&__data, (page));
5698 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
5700 if (__data > INT_MAX)
5702 bfqd->bfq_max_budget = __data;
5705 bfqd->bfq_user_max_budget = __data;
5711 * Leaving this name to preserve name compatibility with cfq
5712 * parameters, but this timeout is used for both sync and async.
5714 static ssize_t bfq_timeout_sync_store(struct elevator_queue *e,
5715 const char *page, size_t count)
5717 struct bfq_data *bfqd = e->elevator_data;
5718 unsigned long __data;
5721 ret = bfq_var_store(&__data, (page));
5727 else if (__data > INT_MAX)
5730 bfqd->bfq_timeout = msecs_to_jiffies(__data);
5731 if (bfqd->bfq_user_max_budget == 0)
5732 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
5737 static ssize_t bfq_strict_guarantees_store(struct elevator_queue *e,
5738 const char *page, size_t count)
5740 struct bfq_data *bfqd = e->elevator_data;
5741 unsigned long __data;
5744 ret = bfq_var_store(&__data, (page));
5750 if (!bfqd->strict_guarantees && __data == 1
5751 && bfqd->bfq_slice_idle < 8 * NSEC_PER_MSEC)
5752 bfqd->bfq_slice_idle = 8 * NSEC_PER_MSEC;
5754 bfqd->strict_guarantees = __data;
5759 static ssize_t bfq_low_latency_store(struct elevator_queue *e,
5760 const char *page, size_t count)
5762 struct bfq_data *bfqd = e->elevator_data;
5763 unsigned long __data;
5766 ret = bfq_var_store(&__data, (page));
5772 if (__data == 0 && bfqd->low_latency != 0)
5774 bfqd->low_latency = __data;
5779 #define BFQ_ATTR(name) \
5780 __ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store)
5782 static struct elv_fs_entry bfq_attrs[] = {
5783 BFQ_ATTR(fifo_expire_sync),
5784 BFQ_ATTR(fifo_expire_async),
5785 BFQ_ATTR(back_seek_max),
5786 BFQ_ATTR(back_seek_penalty),
5787 BFQ_ATTR(slice_idle),
5788 BFQ_ATTR(slice_idle_us),
5789 BFQ_ATTR(max_budget),
5790 BFQ_ATTR(timeout_sync),
5791 BFQ_ATTR(strict_guarantees),
5792 BFQ_ATTR(low_latency),
5796 static struct elevator_type iosched_bfq_mq = {
5798 .limit_depth = bfq_limit_depth,
5799 .prepare_request = bfq_prepare_request,
5800 .requeue_request = bfq_finish_requeue_request,
5801 .finish_request = bfq_finish_requeue_request,
5802 .exit_icq = bfq_exit_icq,
5803 .insert_requests = bfq_insert_requests,
5804 .dispatch_request = bfq_dispatch_request,
5805 .next_request = elv_rb_latter_request,
5806 .former_request = elv_rb_former_request,
5807 .allow_merge = bfq_allow_bio_merge,
5808 .bio_merge = bfq_bio_merge,
5809 .request_merge = bfq_request_merge,
5810 .requests_merged = bfq_requests_merged,
5811 .request_merged = bfq_request_merged,
5812 .has_work = bfq_has_work,
5813 .init_hctx = bfq_init_hctx,
5814 .init_sched = bfq_init_queue,
5815 .exit_sched = bfq_exit_queue,
5818 .icq_size = sizeof(struct bfq_io_cq),
5819 .icq_align = __alignof__(struct bfq_io_cq),
5820 .elevator_attrs = bfq_attrs,
5821 .elevator_name = "bfq",
5822 .elevator_owner = THIS_MODULE,
5824 MODULE_ALIAS("bfq-iosched");
5826 static int __init bfq_init(void)
5830 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5831 ret = blkcg_policy_register(&blkcg_policy_bfq);
5837 if (bfq_slab_setup())
5841 * Times to load large popular applications for the typical
5842 * systems installed on the reference devices (see the
5843 * comments before the definition of the next
5844 * array). Actually, we use slightly lower values, as the
5845 * estimated peak rate tends to be smaller than the actual
5846 * peak rate. The reason for this last fact is that estimates
5847 * are computed over much shorter time intervals than the long
5848 * intervals typically used for benchmarking. Why? First, to
5849 * adapt more quickly to variations. Second, because an I/O
5850 * scheduler cannot rely on a peak-rate-evaluation workload to
5851 * be run for a long time.
5853 ref_wr_duration[0] = msecs_to_jiffies(7000); /* actually 8 sec */
5854 ref_wr_duration[1] = msecs_to_jiffies(2500); /* actually 3 sec */
5856 ret = elv_register(&iosched_bfq_mq);
5865 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5866 blkcg_policy_unregister(&blkcg_policy_bfq);
5871 static void __exit bfq_exit(void)
5873 elv_unregister(&iosched_bfq_mq);
5874 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5875 blkcg_policy_unregister(&blkcg_policy_bfq);
5880 module_init(bfq_init);
5881 module_exit(bfq_exit);
5883 MODULE_AUTHOR("Paolo Valente");
5884 MODULE_LICENSE("GPL");
5885 MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler");