1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Budget Fair Queueing (BFQ) I/O scheduler.
5 * Based on ideas and code from CFQ:
6 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
8 * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
9 * Paolo Valente <paolo.valente@unimore.it>
11 * Copyright (C) 2010 Paolo Valente <paolo.valente@unimore.it>
12 * Arianna Avanzini <avanzini@google.com>
14 * Copyright (C) 2017 Paolo Valente <paolo.valente@linaro.org>
16 * BFQ is a proportional-share I/O scheduler, with some extra
17 * low-latency capabilities. BFQ also supports full hierarchical
18 * scheduling through cgroups. Next paragraphs provide an introduction
19 * on BFQ inner workings. Details on BFQ benefits, usage and
20 * limitations can be found in Documentation/block/bfq-iosched.rst.
22 * BFQ is a proportional-share storage-I/O scheduling algorithm based
23 * on the slice-by-slice service scheme of CFQ. But BFQ assigns
24 * budgets, measured in number of sectors, to processes instead of
25 * time slices. The device is not granted to the in-service process
26 * for a given time slice, but until it has exhausted its assigned
27 * budget. This change from the time to the service domain enables BFQ
28 * to distribute the device throughput among processes as desired,
29 * without any distortion due to throughput fluctuations, or to device
30 * internal queueing. BFQ uses an ad hoc internal scheduler, called
31 * B-WF2Q+, to schedule processes according to their budgets. More
32 * precisely, BFQ schedules queues associated with processes. Each
33 * process/queue is assigned a user-configurable weight, and B-WF2Q+
34 * guarantees that each queue receives a fraction of the throughput
35 * proportional to its weight. Thanks to the accurate policy of
36 * B-WF2Q+, BFQ can afford to assign high budgets to I/O-bound
37 * processes issuing sequential requests (to boost the throughput),
38 * and yet guarantee a low latency to interactive and soft real-time
41 * In particular, to provide these low-latency guarantees, BFQ
42 * explicitly privileges the I/O of two classes of time-sensitive
43 * applications: interactive and soft real-time. In more detail, BFQ
44 * behaves this way if the low_latency parameter is set (default
45 * configuration). This feature enables BFQ to provide applications in
46 * these classes with a very low latency.
48 * To implement this feature, BFQ constantly tries to detect whether
49 * the I/O requests in a bfq_queue come from an interactive or a soft
50 * real-time application. For brevity, in these cases, the queue is
51 * said to be interactive or soft real-time. In both cases, BFQ
52 * privileges the service of the queue, over that of non-interactive
53 * and non-soft-real-time queues. This privileging is performed,
54 * mainly, by raising the weight of the queue. So, for brevity, we
55 * call just weight-raising periods the time periods during which a
56 * queue is privileged, because deemed interactive or soft real-time.
58 * The detection of soft real-time queues/applications is described in
59 * detail in the comments on the function
60 * bfq_bfqq_softrt_next_start. On the other hand, the detection of an
61 * interactive queue works as follows: a queue is deemed interactive
62 * if it is constantly non empty only for a limited time interval,
63 * after which it does become empty. The queue may be deemed
64 * interactive again (for a limited time), if it restarts being
65 * constantly non empty, provided that this happens only after the
66 * queue has remained empty for a given minimum idle time.
68 * By default, BFQ computes automatically the above maximum time
69 * interval, i.e., the time interval after which a constantly
70 * non-empty queue stops being deemed interactive. Since a queue is
71 * weight-raised while it is deemed interactive, this maximum time
72 * interval happens to coincide with the (maximum) duration of the
73 * weight-raising for interactive queues.
75 * Finally, BFQ also features additional heuristics for
76 * preserving both a low latency and a high throughput on NCQ-capable,
77 * rotational or flash-based devices, and to get the job done quickly
78 * for applications consisting in many I/O-bound processes.
80 * NOTE: if the main or only goal, with a given device, is to achieve
81 * the maximum-possible throughput at all times, then do switch off
82 * all low-latency heuristics for that device, by setting low_latency
85 * BFQ is described in [1], where also a reference to the initial,
86 * more theoretical paper on BFQ can be found. The interested reader
87 * can find in the latter paper full details on the main algorithm, as
88 * well as formulas of the guarantees and formal proofs of all the
89 * properties. With respect to the version of BFQ presented in these
90 * papers, this implementation adds a few more heuristics, such as the
91 * ones that guarantee a low latency to interactive and soft real-time
92 * applications, and a hierarchical extension based on H-WF2Q+.
94 * B-WF2Q+ is based on WF2Q+, which is described in [2], together with
95 * H-WF2Q+, while the augmented tree used here to implement B-WF2Q+
96 * with O(log N) complexity derives from the one introduced with EEVDF
99 * [1] P. Valente, A. Avanzini, "Evolution of the BFQ Storage I/O
100 * Scheduler", Proceedings of the First Workshop on Mobile System
101 * Technologies (MST-2015), May 2015.
102 * http://algogroup.unimore.it/people/paolo/disk_sched/mst-2015.pdf
104 * [2] Jon C.R. Bennett and H. Zhang, "Hierarchical Packet Fair Queueing
105 * Algorithms", IEEE/ACM Transactions on Networking, 5(5):675-689,
108 * http://www.cs.cmu.edu/~hzhang/papers/TON-97-Oct.ps.gz
110 * [3] I. Stoica and H. Abdel-Wahab, "Earliest Eligible Virtual Deadline
111 * First: A Flexible and Accurate Mechanism for Proportional Share
112 * Resource Allocation", technical report.
114 * http://www.cs.berkeley.edu/~istoica/papers/eevdf-tr-95.pdf
116 #include <linux/module.h>
117 #include <linux/slab.h>
118 #include <linux/blkdev.h>
119 #include <linux/cgroup.h>
120 #include <linux/ktime.h>
121 #include <linux/rbtree.h>
122 #include <linux/ioprio.h>
123 #include <linux/sbitmap.h>
124 #include <linux/delay.h>
125 #include <linux/backing-dev.h>
127 #include <trace/events/block.h>
129 #include "elevator.h"
132 #include "blk-mq-tag.h"
133 #include "blk-mq-sched.h"
134 #include "bfq-iosched.h"
137 #define BFQ_BFQQ_FNS(name) \
138 void bfq_mark_bfqq_##name(struct bfq_queue *bfqq) \
140 __set_bit(BFQQF_##name, &(bfqq)->flags); \
142 void bfq_clear_bfqq_##name(struct bfq_queue *bfqq) \
144 __clear_bit(BFQQF_##name, &(bfqq)->flags); \
146 int bfq_bfqq_##name(const struct bfq_queue *bfqq) \
148 return test_bit(BFQQF_##name, &(bfqq)->flags); \
151 BFQ_BFQQ_FNS(just_created);
153 BFQ_BFQQ_FNS(wait_request);
154 BFQ_BFQQ_FNS(non_blocking_wait_rq);
155 BFQ_BFQQ_FNS(fifo_expire);
156 BFQ_BFQQ_FNS(has_short_ttime);
158 BFQ_BFQQ_FNS(IO_bound);
159 BFQ_BFQQ_FNS(in_large_burst);
161 BFQ_BFQQ_FNS(split_coop);
162 BFQ_BFQQ_FNS(softrt_update);
163 #undef BFQ_BFQQ_FNS \
165 /* Expiration time of async (0) and sync (1) requests, in ns. */
166 static const u64 bfq_fifo_expire[2] = { NSEC_PER_SEC / 4, NSEC_PER_SEC / 8 };
168 /* Maximum backwards seek (magic number lifted from CFQ), in KiB. */
169 static const int bfq_back_max = 16 * 1024;
171 /* Penalty of a backwards seek, in number of sectors. */
172 static const int bfq_back_penalty = 2;
174 /* Idling period duration, in ns. */
175 static u64 bfq_slice_idle = NSEC_PER_SEC / 125;
177 /* Minimum number of assigned budgets for which stats are safe to compute. */
178 static const int bfq_stats_min_budgets = 194;
180 /* Default maximum budget values, in sectors and number of requests. */
181 static const int bfq_default_max_budget = 16 * 1024;
184 * When a sync request is dispatched, the queue that contains that
185 * request, and all the ancestor entities of that queue, are charged
186 * with the number of sectors of the request. In contrast, if the
187 * request is async, then the queue and its ancestor entities are
188 * charged with the number of sectors of the request, multiplied by
189 * the factor below. This throttles the bandwidth for async I/O,
190 * w.r.t. to sync I/O, and it is done to counter the tendency of async
191 * writes to steal I/O throughput to reads.
193 * The current value of this parameter is the result of a tuning with
194 * several hardware and software configurations. We tried to find the
195 * lowest value for which writes do not cause noticeable problems to
196 * reads. In fact, the lower this parameter, the stabler I/O control,
197 * in the following respect. The lower this parameter is, the less
198 * the bandwidth enjoyed by a group decreases
199 * - when the group does writes, w.r.t. to when it does reads;
200 * - when other groups do reads, w.r.t. to when they do writes.
202 static const int bfq_async_charge_factor = 3;
204 /* Default timeout values, in jiffies, approximating CFQ defaults. */
205 const int bfq_timeout = HZ / 8;
208 * Time limit for merging (see comments in bfq_setup_cooperator). Set
209 * to the slowest value that, in our tests, proved to be effective in
210 * removing false positives, while not causing true positives to miss
213 * As can be deduced from the low time limit below, queue merging, if
214 * successful, happens at the very beginning of the I/O of the involved
215 * cooperating processes, as a consequence of the arrival of the very
216 * first requests from each cooperator. After that, there is very
217 * little chance to find cooperators.
219 static const unsigned long bfq_merge_time_limit = HZ/10;
221 static struct kmem_cache *bfq_pool;
223 /* Below this threshold (in ns), we consider thinktime immediate. */
224 #define BFQ_MIN_TT (2 * NSEC_PER_MSEC)
226 /* hw_tag detection: parallel requests threshold and min samples needed. */
227 #define BFQ_HW_QUEUE_THRESHOLD 3
228 #define BFQ_HW_QUEUE_SAMPLES 32
230 #define BFQQ_SEEK_THR (sector_t)(8 * 100)
231 #define BFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
232 #define BFQ_RQ_SEEKY(bfqd, last_pos, rq) \
233 (get_sdist(last_pos, rq) > \
235 (!blk_queue_nonrot(bfqd->queue) || \
236 blk_rq_sectors(rq) < BFQQ_SECT_THR_NONROT))
237 #define BFQQ_CLOSE_THR (sector_t)(8 * 1024)
238 #define BFQQ_SEEKY(bfqq) (hweight32(bfqq->seek_history) > 19)
240 * Sync random I/O is likely to be confused with soft real-time I/O,
241 * because it is characterized by limited throughput and apparently
242 * isochronous arrival pattern. To avoid false positives, queues
243 * containing only random (seeky) I/O are prevented from being tagged
246 #define BFQQ_TOTALLY_SEEKY(bfqq) (bfqq->seek_history == -1)
248 /* Min number of samples required to perform peak-rate update */
249 #define BFQ_RATE_MIN_SAMPLES 32
250 /* Min observation time interval required to perform a peak-rate update (ns) */
251 #define BFQ_RATE_MIN_INTERVAL (300*NSEC_PER_MSEC)
252 /* Target observation time interval for a peak-rate update (ns) */
253 #define BFQ_RATE_REF_INTERVAL NSEC_PER_SEC
256 * Shift used for peak-rate fixed precision calculations.
258 * - the current shift: 16 positions
259 * - the current type used to store rate: u32
260 * - the current unit of measure for rate: [sectors/usec], or, more precisely,
261 * [(sectors/usec) / 2^BFQ_RATE_SHIFT] to take into account the shift,
262 * the range of rates that can be stored is
263 * [1 / 2^BFQ_RATE_SHIFT, 2^(32 - BFQ_RATE_SHIFT)] sectors/usec =
264 * [1 / 2^16, 2^16] sectors/usec = [15e-6, 65536] sectors/usec =
265 * [15, 65G] sectors/sec
266 * Which, assuming a sector size of 512B, corresponds to a range of
269 #define BFQ_RATE_SHIFT 16
272 * When configured for computing the duration of the weight-raising
273 * for interactive queues automatically (see the comments at the
274 * beginning of this file), BFQ does it using the following formula:
275 * duration = (ref_rate / r) * ref_wr_duration,
276 * where r is the peak rate of the device, and ref_rate and
277 * ref_wr_duration are two reference parameters. In particular,
278 * ref_rate is the peak rate of the reference storage device (see
279 * below), and ref_wr_duration is about the maximum time needed, with
280 * BFQ and while reading two files in parallel, to load typical large
281 * applications on the reference device (see the comments on
282 * max_service_from_wr below, for more details on how ref_wr_duration
283 * is obtained). In practice, the slower/faster the device at hand
284 * is, the more/less it takes to load applications with respect to the
285 * reference device. Accordingly, the longer/shorter BFQ grants
286 * weight raising to interactive applications.
288 * BFQ uses two different reference pairs (ref_rate, ref_wr_duration),
289 * depending on whether the device is rotational or non-rotational.
291 * In the following definitions, ref_rate[0] and ref_wr_duration[0]
292 * are the reference values for a rotational device, whereas
293 * ref_rate[1] and ref_wr_duration[1] are the reference values for a
294 * non-rotational device. The reference rates are not the actual peak
295 * rates of the devices used as a reference, but slightly lower
296 * values. The reason for using slightly lower values is that the
297 * peak-rate estimator tends to yield slightly lower values than the
298 * actual peak rate (it can yield the actual peak rate only if there
299 * is only one process doing I/O, and the process does sequential
302 * The reference peak rates are measured in sectors/usec, left-shifted
305 static int ref_rate[2] = {14000, 33000};
307 * To improve readability, a conversion function is used to initialize
308 * the following array, which entails that the array can be
309 * initialized only in a function.
311 static int ref_wr_duration[2];
314 * BFQ uses the above-detailed, time-based weight-raising mechanism to
315 * privilege interactive tasks. This mechanism is vulnerable to the
316 * following false positives: I/O-bound applications that will go on
317 * doing I/O for much longer than the duration of weight
318 * raising. These applications have basically no benefit from being
319 * weight-raised at the beginning of their I/O. On the opposite end,
320 * while being weight-raised, these applications
321 * a) unjustly steal throughput to applications that may actually need
323 * b) make BFQ uselessly perform device idling; device idling results
324 * in loss of device throughput with most flash-based storage, and may
325 * increase latencies when used purposelessly.
327 * BFQ tries to reduce these problems, by adopting the following
328 * countermeasure. To introduce this countermeasure, we need first to
329 * finish explaining how the duration of weight-raising for
330 * interactive tasks is computed.
332 * For a bfq_queue deemed as interactive, the duration of weight
333 * raising is dynamically adjusted, as a function of the estimated
334 * peak rate of the device, so as to be equal to the time needed to
335 * execute the 'largest' interactive task we benchmarked so far. By
336 * largest task, we mean the task for which each involved process has
337 * to do more I/O than for any of the other tasks we benchmarked. This
338 * reference interactive task is the start-up of LibreOffice Writer,
339 * and in this task each process/bfq_queue needs to have at most ~110K
340 * sectors transferred.
342 * This last piece of information enables BFQ to reduce the actual
343 * duration of weight-raising for at least one class of I/O-bound
344 * applications: those doing sequential or quasi-sequential I/O. An
345 * example is file copy. In fact, once started, the main I/O-bound
346 * processes of these applications usually consume the above 110K
347 * sectors in much less time than the processes of an application that
348 * is starting, because these I/O-bound processes will greedily devote
349 * almost all their CPU cycles only to their target,
350 * throughput-friendly I/O operations. This is even more true if BFQ
351 * happens to be underestimating the device peak rate, and thus
352 * overestimating the duration of weight raising. But, according to
353 * our measurements, once transferred 110K sectors, these processes
354 * have no right to be weight-raised any longer.
356 * Basing on the last consideration, BFQ ends weight-raising for a
357 * bfq_queue if the latter happens to have received an amount of
358 * service at least equal to the following constant. The constant is
359 * set to slightly more than 110K, to have a minimum safety margin.
361 * This early ending of weight-raising reduces the amount of time
362 * during which interactive false positives cause the two problems
363 * described at the beginning of these comments.
365 static const unsigned long max_service_from_wr = 120000;
368 * Maximum time between the creation of two queues, for stable merge
369 * to be activated (in ms)
371 static const unsigned long bfq_activation_stable_merging = 600;
373 * Minimum time to be waited before evaluating delayed stable merge (in ms)
375 static const unsigned long bfq_late_stable_merging = 600;
377 #define RQ_BIC(rq) icq_to_bic((rq)->elv.priv[0])
378 #define RQ_BFQQ(rq) ((rq)->elv.priv[1])
380 struct bfq_queue *bic_to_bfqq(struct bfq_io_cq *bic, bool is_sync)
382 return bic->bfqq[is_sync];
385 static void bfq_put_stable_ref(struct bfq_queue *bfqq);
387 void bic_set_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq, bool is_sync)
390 * If bfqq != NULL, then a non-stable queue merge between
391 * bic->bfqq and bfqq is happening here. This causes troubles
392 * in the following case: bic->bfqq has also been scheduled
393 * for a possible stable merge with bic->stable_merge_bfqq,
394 * and bic->stable_merge_bfqq == bfqq happens to
395 * hold. Troubles occur because bfqq may then undergo a split,
396 * thereby becoming eligible for a stable merge. Yet, if
397 * bic->stable_merge_bfqq points exactly to bfqq, then bfqq
398 * would be stably merged with itself. To avoid this anomaly,
399 * we cancel the stable merge if
400 * bic->stable_merge_bfqq == bfqq.
402 bic->bfqq[is_sync] = bfqq;
404 if (bfqq && bic->stable_merge_bfqq == bfqq) {
406 * Actually, these same instructions are executed also
407 * in bfq_setup_cooperator, in case of abort or actual
408 * execution of a stable merge. We could avoid
409 * repeating these instructions there too, but if we
410 * did so, we would nest even more complexity in this
413 bfq_put_stable_ref(bic->stable_merge_bfqq);
415 bic->stable_merge_bfqq = NULL;
419 struct bfq_data *bic_to_bfqd(struct bfq_io_cq *bic)
421 return bic->icq.q->elevator->elevator_data;
425 * icq_to_bic - convert iocontext queue structure to bfq_io_cq.
426 * @icq: the iocontext queue.
428 static struct bfq_io_cq *icq_to_bic(struct io_cq *icq)
430 /* bic->icq is the first member, %NULL will convert to %NULL */
431 return container_of(icq, struct bfq_io_cq, icq);
435 * bfq_bic_lookup - search into @ioc a bic associated to @bfqd.
436 * @q: the request queue.
438 static struct bfq_io_cq *bfq_bic_lookup(struct request_queue *q)
440 struct bfq_io_cq *icq;
443 if (!current->io_context)
446 spin_lock_irqsave(&q->queue_lock, flags);
447 icq = icq_to_bic(ioc_lookup_icq(q));
448 spin_unlock_irqrestore(&q->queue_lock, flags);
454 * Scheduler run of queue, if there are requests pending and no one in the
455 * driver that will restart queueing.
457 void bfq_schedule_dispatch(struct bfq_data *bfqd)
459 if (bfqd->queued != 0) {
460 bfq_log(bfqd, "schedule dispatch");
461 blk_mq_run_hw_queues(bfqd->queue, true);
465 #define bfq_class_idle(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
467 #define bfq_sample_valid(samples) ((samples) > 80)
470 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
471 * We choose the request that is closer to the head right now. Distance
472 * behind the head is penalized and only allowed to a certain extent.
474 static struct request *bfq_choose_req(struct bfq_data *bfqd,
479 sector_t s1, s2, d1 = 0, d2 = 0;
480 unsigned long back_max;
481 #define BFQ_RQ1_WRAP 0x01 /* request 1 wraps */
482 #define BFQ_RQ2_WRAP 0x02 /* request 2 wraps */
483 unsigned int wrap = 0; /* bit mask: requests behind the disk head? */
485 if (!rq1 || rq1 == rq2)
490 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
492 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
494 if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
496 else if ((rq2->cmd_flags & REQ_META) && !(rq1->cmd_flags & REQ_META))
499 s1 = blk_rq_pos(rq1);
500 s2 = blk_rq_pos(rq2);
503 * By definition, 1KiB is 2 sectors.
505 back_max = bfqd->bfq_back_max * 2;
508 * Strict one way elevator _except_ in the case where we allow
509 * short backward seeks which are biased as twice the cost of a
510 * similar forward seek.
514 else if (s1 + back_max >= last)
515 d1 = (last - s1) * bfqd->bfq_back_penalty;
517 wrap |= BFQ_RQ1_WRAP;
521 else if (s2 + back_max >= last)
522 d2 = (last - s2) * bfqd->bfq_back_penalty;
524 wrap |= BFQ_RQ2_WRAP;
526 /* Found required data */
529 * By doing switch() on the bit mask "wrap" we avoid having to
530 * check two variables for all permutations: --> faster!
533 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
548 case BFQ_RQ1_WRAP|BFQ_RQ2_WRAP: /* both rqs wrapped */
551 * Since both rqs are wrapped,
552 * start with the one that's further behind head
553 * (--> only *one* back seek required),
554 * since back seek takes more time than forward.
563 #define BFQ_LIMIT_INLINE_DEPTH 16
565 #ifdef CONFIG_BFQ_GROUP_IOSCHED
566 static bool bfqq_request_over_limit(struct bfq_queue *bfqq, int limit)
568 struct bfq_data *bfqd = bfqq->bfqd;
569 struct bfq_entity *entity = &bfqq->entity;
570 struct bfq_entity *inline_entities[BFQ_LIMIT_INLINE_DEPTH];
571 struct bfq_entity **entities = inline_entities;
572 int depth, level, alloc_depth = BFQ_LIMIT_INLINE_DEPTH;
573 int class_idx = bfqq->ioprio_class - 1;
574 struct bfq_sched_data *sched_data;
578 if (!entity->on_st_or_in_serv)
582 spin_lock_irq(&bfqd->lock);
583 /* +1 for bfqq entity, root cgroup not included */
584 depth = bfqg_to_blkg(bfqq_group(bfqq))->blkcg->css.cgroup->level + 1;
585 if (depth > alloc_depth) {
586 spin_unlock_irq(&bfqd->lock);
587 if (entities != inline_entities)
589 entities = kmalloc_array(depth, sizeof(*entities), GFP_NOIO);
596 sched_data = entity->sched_data;
597 /* Gather our ancestors as we need to traverse them in reverse order */
599 for_each_entity(entity) {
601 * If at some level entity is not even active, allow request
602 * queueing so that BFQ knows there's work to do and activate
605 if (!entity->on_st_or_in_serv)
607 /* Uh, more parents than cgroup subsystem thinks? */
608 if (WARN_ON_ONCE(level >= depth))
610 entities[level++] = entity;
612 WARN_ON_ONCE(level != depth);
613 for (level--; level >= 0; level--) {
614 entity = entities[level];
616 wsum = bfq_entity_service_tree(entity)->wsum;
620 * For bfqq itself we take into account service trees
621 * of all higher priority classes and multiply their
622 * weights so that low prio queue from higher class
623 * gets more requests than high prio queue from lower
627 for (i = 0; i <= class_idx; i++) {
628 wsum = wsum * IOPRIO_BE_NR +
629 sched_data->service_tree[i].wsum;
632 limit = DIV_ROUND_CLOSEST(limit * entity->weight, wsum);
633 if (entity->allocated >= limit) {
634 bfq_log_bfqq(bfqq->bfqd, bfqq,
635 "too many requests: allocated %d limit %d level %d",
636 entity->allocated, limit, level);
642 spin_unlock_irq(&bfqd->lock);
643 if (entities != inline_entities)
648 static bool bfqq_request_over_limit(struct bfq_queue *bfqq, int limit)
655 * Async I/O can easily starve sync I/O (both sync reads and sync
656 * writes), by consuming all tags. Similarly, storms of sync writes,
657 * such as those that sync(2) may trigger, can starve sync reads.
658 * Limit depths of async I/O and sync writes so as to counter both
661 * Also if a bfq queue or its parent cgroup consume more tags than would be
662 * appropriate for their weight, we trim the available tag depth to 1. This
663 * avoids a situation where one cgroup can starve another cgroup from tags and
664 * thus block service differentiation among cgroups. Note that because the
665 * queue / cgroup already has many requests allocated and queued, this does not
666 * significantly affect service guarantees coming from the BFQ scheduling
669 static void bfq_limit_depth(unsigned int op, struct blk_mq_alloc_data *data)
671 struct bfq_data *bfqd = data->q->elevator->elevator_data;
672 struct bfq_io_cq *bic = bfq_bic_lookup(data->q);
673 struct bfq_queue *bfqq = bic ? bic_to_bfqq(bic, op_is_sync(op)) : NULL;
675 unsigned limit = data->q->nr_requests;
677 /* Sync reads have full depth available */
678 if (op_is_sync(op) && !op_is_write(op)) {
681 depth = bfqd->word_depths[!!bfqd->wr_busy_queues][op_is_sync(op)];
682 limit = (limit * depth) >> bfqd->full_depth_shift;
686 * Does queue (or any parent entity) exceed number of requests that
687 * should be available to it? Heavily limit depth so that it cannot
688 * consume more available requests and thus starve other entities.
690 if (bfqq && bfqq_request_over_limit(bfqq, limit))
693 bfq_log(bfqd, "[%s] wr_busy %d sync %d depth %u",
694 __func__, bfqd->wr_busy_queues, op_is_sync(op), depth);
696 data->shallow_depth = depth;
699 static struct bfq_queue *
700 bfq_rq_pos_tree_lookup(struct bfq_data *bfqd, struct rb_root *root,
701 sector_t sector, struct rb_node **ret_parent,
702 struct rb_node ***rb_link)
704 struct rb_node **p, *parent;
705 struct bfq_queue *bfqq = NULL;
713 bfqq = rb_entry(parent, struct bfq_queue, pos_node);
716 * Sort strictly based on sector. Smallest to the left,
717 * largest to the right.
719 if (sector > blk_rq_pos(bfqq->next_rq))
721 else if (sector < blk_rq_pos(bfqq->next_rq))
729 *ret_parent = parent;
733 bfq_log(bfqd, "rq_pos_tree_lookup %llu: returning %d",
734 (unsigned long long)sector,
735 bfqq ? bfqq->pid : 0);
740 static bool bfq_too_late_for_merging(struct bfq_queue *bfqq)
742 return bfqq->service_from_backlogged > 0 &&
743 time_is_before_jiffies(bfqq->first_IO_time +
744 bfq_merge_time_limit);
748 * The following function is not marked as __cold because it is
749 * actually cold, but for the same performance goal described in the
750 * comments on the likely() at the beginning of
751 * bfq_setup_cooperator(). Unexpectedly, to reach an even lower
752 * execution time for the case where this function is not invoked, we
753 * had to add an unlikely() in each involved if().
756 bfq_pos_tree_add_move(struct bfq_data *bfqd, struct bfq_queue *bfqq)
758 struct rb_node **p, *parent;
759 struct bfq_queue *__bfqq;
761 if (bfqq->pos_root) {
762 rb_erase(&bfqq->pos_node, bfqq->pos_root);
763 bfqq->pos_root = NULL;
766 /* oom_bfqq does not participate in queue merging */
767 if (bfqq == &bfqd->oom_bfqq)
771 * bfqq cannot be merged any longer (see comments in
772 * bfq_setup_cooperator): no point in adding bfqq into the
775 if (bfq_too_late_for_merging(bfqq))
778 if (bfq_class_idle(bfqq))
783 bfqq->pos_root = &bfqq_group(bfqq)->rq_pos_tree;
784 __bfqq = bfq_rq_pos_tree_lookup(bfqd, bfqq->pos_root,
785 blk_rq_pos(bfqq->next_rq), &parent, &p);
787 rb_link_node(&bfqq->pos_node, parent, p);
788 rb_insert_color(&bfqq->pos_node, bfqq->pos_root);
790 bfqq->pos_root = NULL;
794 * The following function returns false either if every active queue
795 * must receive the same share of the throughput (symmetric scenario),
796 * or, as a special case, if bfqq must receive a share of the
797 * throughput lower than or equal to the share that every other active
798 * queue must receive. If bfqq does sync I/O, then these are the only
799 * two cases where bfqq happens to be guaranteed its share of the
800 * throughput even if I/O dispatching is not plugged when bfqq remains
801 * temporarily empty (for more details, see the comments in the
802 * function bfq_better_to_idle()). For this reason, the return value
803 * of this function is used to check whether I/O-dispatch plugging can
806 * The above first case (symmetric scenario) occurs when:
807 * 1) all active queues have the same weight,
808 * 2) all active queues belong to the same I/O-priority class,
809 * 3) all active groups at the same level in the groups tree have the same
811 * 4) all active groups at the same level in the groups tree have the same
812 * number of children.
814 * Unfortunately, keeping the necessary state for evaluating exactly
815 * the last two symmetry sub-conditions above would be quite complex
816 * and time consuming. Therefore this function evaluates, instead,
817 * only the following stronger three sub-conditions, for which it is
818 * much easier to maintain the needed state:
819 * 1) all active queues have the same weight,
820 * 2) all active queues belong to the same I/O-priority class,
821 * 3) there are no active groups.
822 * In particular, the last condition is always true if hierarchical
823 * support or the cgroups interface are not enabled, thus no state
824 * needs to be maintained in this case.
826 static bool bfq_asymmetric_scenario(struct bfq_data *bfqd,
827 struct bfq_queue *bfqq)
829 bool smallest_weight = bfqq &&
830 bfqq->weight_counter &&
831 bfqq->weight_counter ==
833 rb_first_cached(&bfqd->queue_weights_tree),
834 struct bfq_weight_counter,
838 * For queue weights to differ, queue_weights_tree must contain
839 * at least two nodes.
841 bool varied_queue_weights = !smallest_weight &&
842 !RB_EMPTY_ROOT(&bfqd->queue_weights_tree.rb_root) &&
843 (bfqd->queue_weights_tree.rb_root.rb_node->rb_left ||
844 bfqd->queue_weights_tree.rb_root.rb_node->rb_right);
846 bool multiple_classes_busy =
847 (bfqd->busy_queues[0] && bfqd->busy_queues[1]) ||
848 (bfqd->busy_queues[0] && bfqd->busy_queues[2]) ||
849 (bfqd->busy_queues[1] && bfqd->busy_queues[2]);
851 return varied_queue_weights || multiple_classes_busy
852 #ifdef CONFIG_BFQ_GROUP_IOSCHED
853 || bfqd->num_groups_with_pending_reqs > 0
859 * If the weight-counter tree passed as input contains no counter for
860 * the weight of the input queue, then add that counter; otherwise just
861 * increment the existing counter.
863 * Note that weight-counter trees contain few nodes in mostly symmetric
864 * scenarios. For example, if all queues have the same weight, then the
865 * weight-counter tree for the queues may contain at most one node.
866 * This holds even if low_latency is on, because weight-raised queues
867 * are not inserted in the tree.
868 * In most scenarios, the rate at which nodes are created/destroyed
871 void bfq_weights_tree_add(struct bfq_data *bfqd, struct bfq_queue *bfqq,
872 struct rb_root_cached *root)
874 struct bfq_entity *entity = &bfqq->entity;
875 struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
876 bool leftmost = true;
879 * Do not insert if the queue is already associated with a
880 * counter, which happens if:
881 * 1) a request arrival has caused the queue to become both
882 * non-weight-raised, and hence change its weight, and
883 * backlogged; in this respect, each of the two events
884 * causes an invocation of this function,
885 * 2) this is the invocation of this function caused by the
886 * second event. This second invocation is actually useless,
887 * and we handle this fact by exiting immediately. More
888 * efficient or clearer solutions might possibly be adopted.
890 if (bfqq->weight_counter)
894 struct bfq_weight_counter *__counter = container_of(*new,
895 struct bfq_weight_counter,
899 if (entity->weight == __counter->weight) {
900 bfqq->weight_counter = __counter;
903 if (entity->weight < __counter->weight)
904 new = &((*new)->rb_left);
906 new = &((*new)->rb_right);
911 bfqq->weight_counter = kzalloc(sizeof(struct bfq_weight_counter),
915 * In the unlucky event of an allocation failure, we just
916 * exit. This will cause the weight of queue to not be
917 * considered in bfq_asymmetric_scenario, which, in its turn,
918 * causes the scenario to be deemed wrongly symmetric in case
919 * bfqq's weight would have been the only weight making the
920 * scenario asymmetric. On the bright side, no unbalance will
921 * however occur when bfqq becomes inactive again (the
922 * invocation of this function is triggered by an activation
923 * of queue). In fact, bfq_weights_tree_remove does nothing
924 * if !bfqq->weight_counter.
926 if (unlikely(!bfqq->weight_counter))
929 bfqq->weight_counter->weight = entity->weight;
930 rb_link_node(&bfqq->weight_counter->weights_node, parent, new);
931 rb_insert_color_cached(&bfqq->weight_counter->weights_node, root,
935 bfqq->weight_counter->num_active++;
940 * Decrement the weight counter associated with the queue, and, if the
941 * counter reaches 0, remove the counter from the tree.
942 * See the comments to the function bfq_weights_tree_add() for considerations
945 void __bfq_weights_tree_remove(struct bfq_data *bfqd,
946 struct bfq_queue *bfqq,
947 struct rb_root_cached *root)
949 if (!bfqq->weight_counter)
952 bfqq->weight_counter->num_active--;
953 if (bfqq->weight_counter->num_active > 0)
954 goto reset_entity_pointer;
956 rb_erase_cached(&bfqq->weight_counter->weights_node, root);
957 kfree(bfqq->weight_counter);
959 reset_entity_pointer:
960 bfqq->weight_counter = NULL;
965 * Invoke __bfq_weights_tree_remove on bfqq and decrement the number
966 * of active groups for each queue's inactive parent entity.
968 void bfq_weights_tree_remove(struct bfq_data *bfqd,
969 struct bfq_queue *bfqq)
971 struct bfq_entity *entity = bfqq->entity.parent;
973 for_each_entity(entity) {
974 struct bfq_sched_data *sd = entity->my_sched_data;
976 if (sd->next_in_service || sd->in_service_entity) {
978 * entity is still active, because either
979 * next_in_service or in_service_entity is not
980 * NULL (see the comments on the definition of
981 * next_in_service for details on why
982 * in_service_entity must be checked too).
984 * As a consequence, its parent entities are
985 * active as well, and thus this loop must
992 * The decrement of num_groups_with_pending_reqs is
993 * not performed immediately upon the deactivation of
994 * entity, but it is delayed to when it also happens
995 * that the first leaf descendant bfqq of entity gets
996 * all its pending requests completed. The following
997 * instructions perform this delayed decrement, if
998 * needed. See the comments on
999 * num_groups_with_pending_reqs for details.
1001 if (entity->in_groups_with_pending_reqs) {
1002 entity->in_groups_with_pending_reqs = false;
1003 bfqd->num_groups_with_pending_reqs--;
1008 * Next function is invoked last, because it causes bfqq to be
1009 * freed if the following holds: bfqq is not in service and
1010 * has no dispatched request. DO NOT use bfqq after the next
1011 * function invocation.
1013 __bfq_weights_tree_remove(bfqd, bfqq,
1014 &bfqd->queue_weights_tree);
1018 * Return expired entry, or NULL to just start from scratch in rbtree.
1020 static struct request *bfq_check_fifo(struct bfq_queue *bfqq,
1021 struct request *last)
1025 if (bfq_bfqq_fifo_expire(bfqq))
1028 bfq_mark_bfqq_fifo_expire(bfqq);
1030 rq = rq_entry_fifo(bfqq->fifo.next);
1032 if (rq == last || ktime_get_ns() < rq->fifo_time)
1035 bfq_log_bfqq(bfqq->bfqd, bfqq, "check_fifo: returned %p", rq);
1039 static struct request *bfq_find_next_rq(struct bfq_data *bfqd,
1040 struct bfq_queue *bfqq,
1041 struct request *last)
1043 struct rb_node *rbnext = rb_next(&last->rb_node);
1044 struct rb_node *rbprev = rb_prev(&last->rb_node);
1045 struct request *next, *prev = NULL;
1047 /* Follow expired path, else get first next available. */
1048 next = bfq_check_fifo(bfqq, last);
1053 prev = rb_entry_rq(rbprev);
1056 next = rb_entry_rq(rbnext);
1058 rbnext = rb_first(&bfqq->sort_list);
1059 if (rbnext && rbnext != &last->rb_node)
1060 next = rb_entry_rq(rbnext);
1063 return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last));
1066 /* see the definition of bfq_async_charge_factor for details */
1067 static unsigned long bfq_serv_to_charge(struct request *rq,
1068 struct bfq_queue *bfqq)
1070 if (bfq_bfqq_sync(bfqq) || bfqq->wr_coeff > 1 ||
1071 bfq_asymmetric_scenario(bfqq->bfqd, bfqq))
1072 return blk_rq_sectors(rq);
1074 return blk_rq_sectors(rq) * bfq_async_charge_factor;
1078 * bfq_updated_next_req - update the queue after a new next_rq selection.
1079 * @bfqd: the device data the queue belongs to.
1080 * @bfqq: the queue to update.
1082 * If the first request of a queue changes we make sure that the queue
1083 * has enough budget to serve at least its first request (if the
1084 * request has grown). We do this because if the queue has not enough
1085 * budget for its first request, it has to go through two dispatch
1086 * rounds to actually get it dispatched.
1088 static void bfq_updated_next_req(struct bfq_data *bfqd,
1089 struct bfq_queue *bfqq)
1091 struct bfq_entity *entity = &bfqq->entity;
1092 struct request *next_rq = bfqq->next_rq;
1093 unsigned long new_budget;
1098 if (bfqq == bfqd->in_service_queue)
1100 * In order not to break guarantees, budgets cannot be
1101 * changed after an entity has been selected.
1105 new_budget = max_t(unsigned long,
1106 max_t(unsigned long, bfqq->max_budget,
1107 bfq_serv_to_charge(next_rq, bfqq)),
1109 if (entity->budget != new_budget) {
1110 entity->budget = new_budget;
1111 bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu",
1113 bfq_requeue_bfqq(bfqd, bfqq, false);
1117 static unsigned int bfq_wr_duration(struct bfq_data *bfqd)
1121 if (bfqd->bfq_wr_max_time > 0)
1122 return bfqd->bfq_wr_max_time;
1124 dur = bfqd->rate_dur_prod;
1125 do_div(dur, bfqd->peak_rate);
1128 * Limit duration between 3 and 25 seconds. The upper limit
1129 * has been conservatively set after the following worst case:
1130 * on a QEMU/KVM virtual machine
1131 * - running in a slow PC
1132 * - with a virtual disk stacked on a slow low-end 5400rpm HDD
1133 * - serving a heavy I/O workload, such as the sequential reading
1135 * mplayer took 23 seconds to start, if constantly weight-raised.
1137 * As for higher values than that accommodating the above bad
1138 * scenario, tests show that higher values would often yield
1139 * the opposite of the desired result, i.e., would worsen
1140 * responsiveness by allowing non-interactive applications to
1141 * preserve weight raising for too long.
1143 * On the other end, lower values than 3 seconds make it
1144 * difficult for most interactive tasks to complete their jobs
1145 * before weight-raising finishes.
1147 return clamp_val(dur, msecs_to_jiffies(3000), msecs_to_jiffies(25000));
1150 /* switch back from soft real-time to interactive weight raising */
1151 static void switch_back_to_interactive_wr(struct bfq_queue *bfqq,
1152 struct bfq_data *bfqd)
1154 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1155 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1156 bfqq->last_wr_start_finish = bfqq->wr_start_at_switch_to_srt;
1160 bfq_bfqq_resume_state(struct bfq_queue *bfqq, struct bfq_data *bfqd,
1161 struct bfq_io_cq *bic, bool bfq_already_existing)
1163 unsigned int old_wr_coeff = 1;
1164 bool busy = bfq_already_existing && bfq_bfqq_busy(bfqq);
1166 if (bic->saved_has_short_ttime)
1167 bfq_mark_bfqq_has_short_ttime(bfqq);
1169 bfq_clear_bfqq_has_short_ttime(bfqq);
1171 if (bic->saved_IO_bound)
1172 bfq_mark_bfqq_IO_bound(bfqq);
1174 bfq_clear_bfqq_IO_bound(bfqq);
1176 bfqq->last_serv_time_ns = bic->saved_last_serv_time_ns;
1177 bfqq->inject_limit = bic->saved_inject_limit;
1178 bfqq->decrease_time_jif = bic->saved_decrease_time_jif;
1180 bfqq->entity.new_weight = bic->saved_weight;
1181 bfqq->ttime = bic->saved_ttime;
1182 bfqq->io_start_time = bic->saved_io_start_time;
1183 bfqq->tot_idle_time = bic->saved_tot_idle_time;
1185 * Restore weight coefficient only if low_latency is on
1187 if (bfqd->low_latency) {
1188 old_wr_coeff = bfqq->wr_coeff;
1189 bfqq->wr_coeff = bic->saved_wr_coeff;
1191 bfqq->service_from_wr = bic->saved_service_from_wr;
1192 bfqq->wr_start_at_switch_to_srt = bic->saved_wr_start_at_switch_to_srt;
1193 bfqq->last_wr_start_finish = bic->saved_last_wr_start_finish;
1194 bfqq->wr_cur_max_time = bic->saved_wr_cur_max_time;
1196 if (bfqq->wr_coeff > 1 && (bfq_bfqq_in_large_burst(bfqq) ||
1197 time_is_before_jiffies(bfqq->last_wr_start_finish +
1198 bfqq->wr_cur_max_time))) {
1199 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
1200 !bfq_bfqq_in_large_burst(bfqq) &&
1201 time_is_after_eq_jiffies(bfqq->wr_start_at_switch_to_srt +
1202 bfq_wr_duration(bfqd))) {
1203 switch_back_to_interactive_wr(bfqq, bfqd);
1206 bfq_log_bfqq(bfqq->bfqd, bfqq,
1207 "resume state: switching off wr");
1211 /* make sure weight will be updated, however we got here */
1212 bfqq->entity.prio_changed = 1;
1217 if (old_wr_coeff == 1 && bfqq->wr_coeff > 1)
1218 bfqd->wr_busy_queues++;
1219 else if (old_wr_coeff > 1 && bfqq->wr_coeff == 1)
1220 bfqd->wr_busy_queues--;
1223 static int bfqq_process_refs(struct bfq_queue *bfqq)
1225 return bfqq->ref - bfqq->entity.allocated -
1226 bfqq->entity.on_st_or_in_serv -
1227 (bfqq->weight_counter != NULL) - bfqq->stable_ref;
1230 /* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */
1231 static void bfq_reset_burst_list(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1233 struct bfq_queue *item;
1234 struct hlist_node *n;
1236 hlist_for_each_entry_safe(item, n, &bfqd->burst_list, burst_list_node)
1237 hlist_del_init(&item->burst_list_node);
1240 * Start the creation of a new burst list only if there is no
1241 * active queue. See comments on the conditional invocation of
1242 * bfq_handle_burst().
1244 if (bfq_tot_busy_queues(bfqd) == 0) {
1245 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1246 bfqd->burst_size = 1;
1248 bfqd->burst_size = 0;
1250 bfqd->burst_parent_entity = bfqq->entity.parent;
1253 /* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */
1254 static void bfq_add_to_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1256 /* Increment burst size to take into account also bfqq */
1259 if (bfqd->burst_size == bfqd->bfq_large_burst_thresh) {
1260 struct bfq_queue *pos, *bfqq_item;
1261 struct hlist_node *n;
1264 * Enough queues have been activated shortly after each
1265 * other to consider this burst as large.
1267 bfqd->large_burst = true;
1270 * We can now mark all queues in the burst list as
1271 * belonging to a large burst.
1273 hlist_for_each_entry(bfqq_item, &bfqd->burst_list,
1275 bfq_mark_bfqq_in_large_burst(bfqq_item);
1276 bfq_mark_bfqq_in_large_burst(bfqq);
1279 * From now on, and until the current burst finishes, any
1280 * new queue being activated shortly after the last queue
1281 * was inserted in the burst can be immediately marked as
1282 * belonging to a large burst. So the burst list is not
1283 * needed any more. Remove it.
1285 hlist_for_each_entry_safe(pos, n, &bfqd->burst_list,
1287 hlist_del_init(&pos->burst_list_node);
1289 * Burst not yet large: add bfqq to the burst list. Do
1290 * not increment the ref counter for bfqq, because bfqq
1291 * is removed from the burst list before freeing bfqq
1294 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1298 * If many queues belonging to the same group happen to be created
1299 * shortly after each other, then the processes associated with these
1300 * queues have typically a common goal. In particular, bursts of queue
1301 * creations are usually caused by services or applications that spawn
1302 * many parallel threads/processes. Examples are systemd during boot,
1303 * or git grep. To help these processes get their job done as soon as
1304 * possible, it is usually better to not grant either weight-raising
1305 * or device idling to their queues, unless these queues must be
1306 * protected from the I/O flowing through other active queues.
1308 * In this comment we describe, firstly, the reasons why this fact
1309 * holds, and, secondly, the next function, which implements the main
1310 * steps needed to properly mark these queues so that they can then be
1311 * treated in a different way.
1313 * The above services or applications benefit mostly from a high
1314 * throughput: the quicker the requests of the activated queues are
1315 * cumulatively served, the sooner the target job of these queues gets
1316 * completed. As a consequence, weight-raising any of these queues,
1317 * which also implies idling the device for it, is almost always
1318 * counterproductive, unless there are other active queues to isolate
1319 * these new queues from. If there no other active queues, then
1320 * weight-raising these new queues just lowers throughput in most
1323 * On the other hand, a burst of queue creations may be caused also by
1324 * the start of an application that does not consist of a lot of
1325 * parallel I/O-bound threads. In fact, with a complex application,
1326 * several short processes may need to be executed to start-up the
1327 * application. In this respect, to start an application as quickly as
1328 * possible, the best thing to do is in any case to privilege the I/O
1329 * related to the application with respect to all other
1330 * I/O. Therefore, the best strategy to start as quickly as possible
1331 * an application that causes a burst of queue creations is to
1332 * weight-raise all the queues created during the burst. This is the
1333 * exact opposite of the best strategy for the other type of bursts.
1335 * In the end, to take the best action for each of the two cases, the
1336 * two types of bursts need to be distinguished. Fortunately, this
1337 * seems relatively easy, by looking at the sizes of the bursts. In
1338 * particular, we found a threshold such that only bursts with a
1339 * larger size than that threshold are apparently caused by
1340 * services or commands such as systemd or git grep. For brevity,
1341 * hereafter we call just 'large' these bursts. BFQ *does not*
1342 * weight-raise queues whose creation occurs in a large burst. In
1343 * addition, for each of these queues BFQ performs or does not perform
1344 * idling depending on which choice boosts the throughput more. The
1345 * exact choice depends on the device and request pattern at
1348 * Unfortunately, false positives may occur while an interactive task
1349 * is starting (e.g., an application is being started). The
1350 * consequence is that the queues associated with the task do not
1351 * enjoy weight raising as expected. Fortunately these false positives
1352 * are very rare. They typically occur if some service happens to
1353 * start doing I/O exactly when the interactive task starts.
1355 * Turning back to the next function, it is invoked only if there are
1356 * no active queues (apart from active queues that would belong to the
1357 * same, possible burst bfqq would belong to), and it implements all
1358 * the steps needed to detect the occurrence of a large burst and to
1359 * properly mark all the queues belonging to it (so that they can then
1360 * be treated in a different way). This goal is achieved by
1361 * maintaining a "burst list" that holds, temporarily, the queues that
1362 * belong to the burst in progress. The list is then used to mark
1363 * these queues as belonging to a large burst if the burst does become
1364 * large. The main steps are the following.
1366 * . when the very first queue is created, the queue is inserted into the
1367 * list (as it could be the first queue in a possible burst)
1369 * . if the current burst has not yet become large, and a queue Q that does
1370 * not yet belong to the burst is activated shortly after the last time
1371 * at which a new queue entered the burst list, then the function appends
1372 * Q to the burst list
1374 * . if, as a consequence of the previous step, the burst size reaches
1375 * the large-burst threshold, then
1377 * . all the queues in the burst list are marked as belonging to a
1380 * . the burst list is deleted; in fact, the burst list already served
1381 * its purpose (keeping temporarily track of the queues in a burst,
1382 * so as to be able to mark them as belonging to a large burst in the
1383 * previous sub-step), and now is not needed any more
1385 * . the device enters a large-burst mode
1387 * . if a queue Q that does not belong to the burst is created while
1388 * the device is in large-burst mode and shortly after the last time
1389 * at which a queue either entered the burst list or was marked as
1390 * belonging to the current large burst, then Q is immediately marked
1391 * as belonging to a large burst.
1393 * . if a queue Q that does not belong to the burst is created a while
1394 * later, i.e., not shortly after, than the last time at which a queue
1395 * either entered the burst list or was marked as belonging to the
1396 * current large burst, then the current burst is deemed as finished and:
1398 * . the large-burst mode is reset if set
1400 * . the burst list is emptied
1402 * . Q is inserted in the burst list, as Q may be the first queue
1403 * in a possible new burst (then the burst list contains just Q
1406 static void bfq_handle_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1409 * If bfqq is already in the burst list or is part of a large
1410 * burst, or finally has just been split, then there is
1411 * nothing else to do.
1413 if (!hlist_unhashed(&bfqq->burst_list_node) ||
1414 bfq_bfqq_in_large_burst(bfqq) ||
1415 time_is_after_eq_jiffies(bfqq->split_time +
1416 msecs_to_jiffies(10)))
1420 * If bfqq's creation happens late enough, or bfqq belongs to
1421 * a different group than the burst group, then the current
1422 * burst is finished, and related data structures must be
1425 * In this respect, consider the special case where bfqq is
1426 * the very first queue created after BFQ is selected for this
1427 * device. In this case, last_ins_in_burst and
1428 * burst_parent_entity are not yet significant when we get
1429 * here. But it is easy to verify that, whether or not the
1430 * following condition is true, bfqq will end up being
1431 * inserted into the burst list. In particular the list will
1432 * happen to contain only bfqq. And this is exactly what has
1433 * to happen, as bfqq may be the first queue of the first
1436 if (time_is_before_jiffies(bfqd->last_ins_in_burst +
1437 bfqd->bfq_burst_interval) ||
1438 bfqq->entity.parent != bfqd->burst_parent_entity) {
1439 bfqd->large_burst = false;
1440 bfq_reset_burst_list(bfqd, bfqq);
1445 * If we get here, then bfqq is being activated shortly after the
1446 * last queue. So, if the current burst is also large, we can mark
1447 * bfqq as belonging to this large burst immediately.
1449 if (bfqd->large_burst) {
1450 bfq_mark_bfqq_in_large_burst(bfqq);
1455 * If we get here, then a large-burst state has not yet been
1456 * reached, but bfqq is being activated shortly after the last
1457 * queue. Then we add bfqq to the burst.
1459 bfq_add_to_burst(bfqd, bfqq);
1462 * At this point, bfqq either has been added to the current
1463 * burst or has caused the current burst to terminate and a
1464 * possible new burst to start. In particular, in the second
1465 * case, bfqq has become the first queue in the possible new
1466 * burst. In both cases last_ins_in_burst needs to be moved
1469 bfqd->last_ins_in_burst = jiffies;
1472 static int bfq_bfqq_budget_left(struct bfq_queue *bfqq)
1474 struct bfq_entity *entity = &bfqq->entity;
1476 return entity->budget - entity->service;
1480 * If enough samples have been computed, return the current max budget
1481 * stored in bfqd, which is dynamically updated according to the
1482 * estimated disk peak rate; otherwise return the default max budget
1484 static int bfq_max_budget(struct bfq_data *bfqd)
1486 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1487 return bfq_default_max_budget;
1489 return bfqd->bfq_max_budget;
1493 * Return min budget, which is a fraction of the current or default
1494 * max budget (trying with 1/32)
1496 static int bfq_min_budget(struct bfq_data *bfqd)
1498 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1499 return bfq_default_max_budget / 32;
1501 return bfqd->bfq_max_budget / 32;
1505 * The next function, invoked after the input queue bfqq switches from
1506 * idle to busy, updates the budget of bfqq. The function also tells
1507 * whether the in-service queue should be expired, by returning
1508 * true. The purpose of expiring the in-service queue is to give bfqq
1509 * the chance to possibly preempt the in-service queue, and the reason
1510 * for preempting the in-service queue is to achieve one of the two
1513 * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has
1514 * expired because it has remained idle. In particular, bfqq may have
1515 * expired for one of the following two reasons:
1517 * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling
1518 * and did not make it to issue a new request before its last
1519 * request was served;
1521 * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue
1522 * a new request before the expiration of the idling-time.
1524 * Even if bfqq has expired for one of the above reasons, the process
1525 * associated with the queue may be however issuing requests greedily,
1526 * and thus be sensitive to the bandwidth it receives (bfqq may have
1527 * remained idle for other reasons: CPU high load, bfqq not enjoying
1528 * idling, I/O throttling somewhere in the path from the process to
1529 * the I/O scheduler, ...). But if, after every expiration for one of
1530 * the above two reasons, bfqq has to wait for the service of at least
1531 * one full budget of another queue before being served again, then
1532 * bfqq is likely to get a much lower bandwidth or resource time than
1533 * its reserved ones. To address this issue, two countermeasures need
1536 * First, the budget and the timestamps of bfqq need to be updated in
1537 * a special way on bfqq reactivation: they need to be updated as if
1538 * bfqq did not remain idle and did not expire. In fact, if they are
1539 * computed as if bfqq expired and remained idle until reactivation,
1540 * then the process associated with bfqq is treated as if, instead of
1541 * being greedy, it stopped issuing requests when bfqq remained idle,
1542 * and restarts issuing requests only on this reactivation. In other
1543 * words, the scheduler does not help the process recover the "service
1544 * hole" between bfqq expiration and reactivation. As a consequence,
1545 * the process receives a lower bandwidth than its reserved one. In
1546 * contrast, to recover this hole, the budget must be updated as if
1547 * bfqq was not expired at all before this reactivation, i.e., it must
1548 * be set to the value of the remaining budget when bfqq was
1549 * expired. Along the same line, timestamps need to be assigned the
1550 * value they had the last time bfqq was selected for service, i.e.,
1551 * before last expiration. Thus timestamps need to be back-shifted
1552 * with respect to their normal computation (see [1] for more details
1553 * on this tricky aspect).
1555 * Secondly, to allow the process to recover the hole, the in-service
1556 * queue must be expired too, to give bfqq the chance to preempt it
1557 * immediately. In fact, if bfqq has to wait for a full budget of the
1558 * in-service queue to be completed, then it may become impossible to
1559 * let the process recover the hole, even if the back-shifted
1560 * timestamps of bfqq are lower than those of the in-service queue. If
1561 * this happens for most or all of the holes, then the process may not
1562 * receive its reserved bandwidth. In this respect, it is worth noting
1563 * that, being the service of outstanding requests unpreemptible, a
1564 * little fraction of the holes may however be unrecoverable, thereby
1565 * causing a little loss of bandwidth.
1567 * The last important point is detecting whether bfqq does need this
1568 * bandwidth recovery. In this respect, the next function deems the
1569 * process associated with bfqq greedy, and thus allows it to recover
1570 * the hole, if: 1) the process is waiting for the arrival of a new
1571 * request (which implies that bfqq expired for one of the above two
1572 * reasons), and 2) such a request has arrived soon. The first
1573 * condition is controlled through the flag non_blocking_wait_rq,
1574 * while the second through the flag arrived_in_time. If both
1575 * conditions hold, then the function computes the budget in the
1576 * above-described special way, and signals that the in-service queue
1577 * should be expired. Timestamp back-shifting is done later in
1578 * __bfq_activate_entity.
1580 * 2. Reduce latency. Even if timestamps are not backshifted to let
1581 * the process associated with bfqq recover a service hole, bfqq may
1582 * however happen to have, after being (re)activated, a lower finish
1583 * timestamp than the in-service queue. That is, the next budget of
1584 * bfqq may have to be completed before the one of the in-service
1585 * queue. If this is the case, then preempting the in-service queue
1586 * allows this goal to be achieved, apart from the unpreemptible,
1587 * outstanding requests mentioned above.
1589 * Unfortunately, regardless of which of the above two goals one wants
1590 * to achieve, service trees need first to be updated to know whether
1591 * the in-service queue must be preempted. To have service trees
1592 * correctly updated, the in-service queue must be expired and
1593 * rescheduled, and bfqq must be scheduled too. This is one of the
1594 * most costly operations (in future versions, the scheduling
1595 * mechanism may be re-designed in such a way to make it possible to
1596 * know whether preemption is needed without needing to update service
1597 * trees). In addition, queue preemptions almost always cause random
1598 * I/O, which may in turn cause loss of throughput. Finally, there may
1599 * even be no in-service queue when the next function is invoked (so,
1600 * no queue to compare timestamps with). Because of these facts, the
1601 * next function adopts the following simple scheme to avoid costly
1602 * operations, too frequent preemptions and too many dependencies on
1603 * the state of the scheduler: it requests the expiration of the
1604 * in-service queue (unconditionally) only for queues that need to
1605 * recover a hole. Then it delegates to other parts of the code the
1606 * responsibility of handling the above case 2.
1608 static bool bfq_bfqq_update_budg_for_activation(struct bfq_data *bfqd,
1609 struct bfq_queue *bfqq,
1610 bool arrived_in_time)
1612 struct bfq_entity *entity = &bfqq->entity;
1615 * In the next compound condition, we check also whether there
1616 * is some budget left, because otherwise there is no point in
1617 * trying to go on serving bfqq with this same budget: bfqq
1618 * would be expired immediately after being selected for
1619 * service. This would only cause useless overhead.
1621 if (bfq_bfqq_non_blocking_wait_rq(bfqq) && arrived_in_time &&
1622 bfq_bfqq_budget_left(bfqq) > 0) {
1624 * We do not clear the flag non_blocking_wait_rq here, as
1625 * the latter is used in bfq_activate_bfqq to signal
1626 * that timestamps need to be back-shifted (and is
1627 * cleared right after).
1631 * In next assignment we rely on that either
1632 * entity->service or entity->budget are not updated
1633 * on expiration if bfqq is empty (see
1634 * __bfq_bfqq_recalc_budget). Thus both quantities
1635 * remain unchanged after such an expiration, and the
1636 * following statement therefore assigns to
1637 * entity->budget the remaining budget on such an
1640 entity->budget = min_t(unsigned long,
1641 bfq_bfqq_budget_left(bfqq),
1645 * At this point, we have used entity->service to get
1646 * the budget left (needed for updating
1647 * entity->budget). Thus we finally can, and have to,
1648 * reset entity->service. The latter must be reset
1649 * because bfqq would otherwise be charged again for
1650 * the service it has received during its previous
1653 entity->service = 0;
1659 * We can finally complete expiration, by setting service to 0.
1661 entity->service = 0;
1662 entity->budget = max_t(unsigned long, bfqq->max_budget,
1663 bfq_serv_to_charge(bfqq->next_rq, bfqq));
1664 bfq_clear_bfqq_non_blocking_wait_rq(bfqq);
1669 * Return the farthest past time instant according to jiffies
1672 static unsigned long bfq_smallest_from_now(void)
1674 return jiffies - MAX_JIFFY_OFFSET;
1677 static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data *bfqd,
1678 struct bfq_queue *bfqq,
1679 unsigned int old_wr_coeff,
1680 bool wr_or_deserves_wr,
1685 if (old_wr_coeff == 1 && wr_or_deserves_wr) {
1686 /* start a weight-raising period */
1688 bfqq->service_from_wr = 0;
1689 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1690 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1693 * No interactive weight raising in progress
1694 * here: assign minus infinity to
1695 * wr_start_at_switch_to_srt, to make sure
1696 * that, at the end of the soft-real-time
1697 * weight raising periods that is starting
1698 * now, no interactive weight-raising period
1699 * may be wrongly considered as still in
1700 * progress (and thus actually started by
1703 bfqq->wr_start_at_switch_to_srt =
1704 bfq_smallest_from_now();
1705 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1706 BFQ_SOFTRT_WEIGHT_FACTOR;
1707 bfqq->wr_cur_max_time =
1708 bfqd->bfq_wr_rt_max_time;
1712 * If needed, further reduce budget to make sure it is
1713 * close to bfqq's backlog, so as to reduce the
1714 * scheduling-error component due to a too large
1715 * budget. Do not care about throughput consequences,
1716 * but only about latency. Finally, do not assign a
1717 * too small budget either, to avoid increasing
1718 * latency by causing too frequent expirations.
1720 bfqq->entity.budget = min_t(unsigned long,
1721 bfqq->entity.budget,
1722 2 * bfq_min_budget(bfqd));
1723 } else if (old_wr_coeff > 1) {
1724 if (interactive) { /* update wr coeff and duration */
1725 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1726 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1727 } else if (in_burst)
1731 * The application is now or still meeting the
1732 * requirements for being deemed soft rt. We
1733 * can then correctly and safely (re)charge
1734 * the weight-raising duration for the
1735 * application with the weight-raising
1736 * duration for soft rt applications.
1738 * In particular, doing this recharge now, i.e.,
1739 * before the weight-raising period for the
1740 * application finishes, reduces the probability
1741 * of the following negative scenario:
1742 * 1) the weight of a soft rt application is
1743 * raised at startup (as for any newly
1744 * created application),
1745 * 2) since the application is not interactive,
1746 * at a certain time weight-raising is
1747 * stopped for the application,
1748 * 3) at that time the application happens to
1749 * still have pending requests, and hence
1750 * is destined to not have a chance to be
1751 * deemed soft rt before these requests are
1752 * completed (see the comments to the
1753 * function bfq_bfqq_softrt_next_start()
1754 * for details on soft rt detection),
1755 * 4) these pending requests experience a high
1756 * latency because the application is not
1757 * weight-raised while they are pending.
1759 if (bfqq->wr_cur_max_time !=
1760 bfqd->bfq_wr_rt_max_time) {
1761 bfqq->wr_start_at_switch_to_srt =
1762 bfqq->last_wr_start_finish;
1764 bfqq->wr_cur_max_time =
1765 bfqd->bfq_wr_rt_max_time;
1766 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1767 BFQ_SOFTRT_WEIGHT_FACTOR;
1769 bfqq->last_wr_start_finish = jiffies;
1774 static bool bfq_bfqq_idle_for_long_time(struct bfq_data *bfqd,
1775 struct bfq_queue *bfqq)
1777 return bfqq->dispatched == 0 &&
1778 time_is_before_jiffies(
1779 bfqq->budget_timeout +
1780 bfqd->bfq_wr_min_idle_time);
1785 * Return true if bfqq is in a higher priority class, or has a higher
1786 * weight than the in-service queue.
1788 static bool bfq_bfqq_higher_class_or_weight(struct bfq_queue *bfqq,
1789 struct bfq_queue *in_serv_bfqq)
1791 int bfqq_weight, in_serv_weight;
1793 if (bfqq->ioprio_class < in_serv_bfqq->ioprio_class)
1796 if (in_serv_bfqq->entity.parent == bfqq->entity.parent) {
1797 bfqq_weight = bfqq->entity.weight;
1798 in_serv_weight = in_serv_bfqq->entity.weight;
1800 if (bfqq->entity.parent)
1801 bfqq_weight = bfqq->entity.parent->weight;
1803 bfqq_weight = bfqq->entity.weight;
1804 if (in_serv_bfqq->entity.parent)
1805 in_serv_weight = in_serv_bfqq->entity.parent->weight;
1807 in_serv_weight = in_serv_bfqq->entity.weight;
1810 return bfqq_weight > in_serv_weight;
1813 static bool bfq_better_to_idle(struct bfq_queue *bfqq);
1815 static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data *bfqd,
1816 struct bfq_queue *bfqq,
1821 bool soft_rt, in_burst, wr_or_deserves_wr,
1822 bfqq_wants_to_preempt,
1823 idle_for_long_time = bfq_bfqq_idle_for_long_time(bfqd, bfqq),
1825 * See the comments on
1826 * bfq_bfqq_update_budg_for_activation for
1827 * details on the usage of the next variable.
1829 arrived_in_time = ktime_get_ns() <=
1830 bfqq->ttime.last_end_request +
1831 bfqd->bfq_slice_idle * 3;
1835 * bfqq deserves to be weight-raised if:
1837 * - it does not belong to a large burst,
1838 * - it has been idle for enough time or is soft real-time,
1839 * - is linked to a bfq_io_cq (it is not shared in any sense),
1840 * - has a default weight (otherwise we assume the user wanted
1841 * to control its weight explicitly)
1843 in_burst = bfq_bfqq_in_large_burst(bfqq);
1844 soft_rt = bfqd->bfq_wr_max_softrt_rate > 0 &&
1845 !BFQQ_TOTALLY_SEEKY(bfqq) &&
1847 time_is_before_jiffies(bfqq->soft_rt_next_start) &&
1848 bfqq->dispatched == 0 &&
1849 bfqq->entity.new_weight == 40;
1850 *interactive = !in_burst && idle_for_long_time &&
1851 bfqq->entity.new_weight == 40;
1853 * Merged bfq_queues are kept out of weight-raising
1854 * (low-latency) mechanisms. The reason is that these queues
1855 * are usually created for non-interactive and
1856 * non-soft-real-time tasks. Yet this is not the case for
1857 * stably-merged queues. These queues are merged just because
1858 * they are created shortly after each other. So they may
1859 * easily serve the I/O of an interactive or soft-real time
1860 * application, if the application happens to spawn multiple
1861 * processes. So let also stably-merged queued enjoy weight
1864 wr_or_deserves_wr = bfqd->low_latency &&
1865 (bfqq->wr_coeff > 1 ||
1866 (bfq_bfqq_sync(bfqq) &&
1867 (bfqq->bic || RQ_BIC(rq)->stably_merged) &&
1868 (*interactive || soft_rt)));
1871 * Using the last flag, update budget and check whether bfqq
1872 * may want to preempt the in-service queue.
1874 bfqq_wants_to_preempt =
1875 bfq_bfqq_update_budg_for_activation(bfqd, bfqq,
1879 * If bfqq happened to be activated in a burst, but has been
1880 * idle for much more than an interactive queue, then we
1881 * assume that, in the overall I/O initiated in the burst, the
1882 * I/O associated with bfqq is finished. So bfqq does not need
1883 * to be treated as a queue belonging to a burst
1884 * anymore. Accordingly, we reset bfqq's in_large_burst flag
1885 * if set, and remove bfqq from the burst list if it's
1886 * there. We do not decrement burst_size, because the fact
1887 * that bfqq does not need to belong to the burst list any
1888 * more does not invalidate the fact that bfqq was created in
1891 if (likely(!bfq_bfqq_just_created(bfqq)) &&
1892 idle_for_long_time &&
1893 time_is_before_jiffies(
1894 bfqq->budget_timeout +
1895 msecs_to_jiffies(10000))) {
1896 hlist_del_init(&bfqq->burst_list_node);
1897 bfq_clear_bfqq_in_large_burst(bfqq);
1900 bfq_clear_bfqq_just_created(bfqq);
1902 if (bfqd->low_latency) {
1903 if (unlikely(time_is_after_jiffies(bfqq->split_time)))
1906 jiffies - bfqd->bfq_wr_min_idle_time - 1;
1908 if (time_is_before_jiffies(bfqq->split_time +
1909 bfqd->bfq_wr_min_idle_time)) {
1910 bfq_update_bfqq_wr_on_rq_arrival(bfqd, bfqq,
1917 if (old_wr_coeff != bfqq->wr_coeff)
1918 bfqq->entity.prio_changed = 1;
1922 bfqq->last_idle_bklogged = jiffies;
1923 bfqq->service_from_backlogged = 0;
1924 bfq_clear_bfqq_softrt_update(bfqq);
1926 bfq_add_bfqq_busy(bfqd, bfqq);
1929 * Expire in-service queue if preemption may be needed for
1930 * guarantees or throughput. As for guarantees, we care
1931 * explicitly about two cases. The first is that bfqq has to
1932 * recover a service hole, as explained in the comments on
1933 * bfq_bfqq_update_budg_for_activation(), i.e., that
1934 * bfqq_wants_to_preempt is true. However, if bfqq does not
1935 * carry time-critical I/O, then bfqq's bandwidth is less
1936 * important than that of queues that carry time-critical I/O.
1937 * So, as a further constraint, we consider this case only if
1938 * bfqq is at least as weight-raised, i.e., at least as time
1939 * critical, as the in-service queue.
1941 * The second case is that bfqq is in a higher priority class,
1942 * or has a higher weight than the in-service queue. If this
1943 * condition does not hold, we don't care because, even if
1944 * bfqq does not start to be served immediately, the resulting
1945 * delay for bfqq's I/O is however lower or much lower than
1946 * the ideal completion time to be guaranteed to bfqq's I/O.
1948 * In both cases, preemption is needed only if, according to
1949 * the timestamps of both bfqq and of the in-service queue,
1950 * bfqq actually is the next queue to serve. So, to reduce
1951 * useless preemptions, the return value of
1952 * next_queue_may_preempt() is considered in the next compound
1953 * condition too. Yet next_queue_may_preempt() just checks a
1954 * simple, necessary condition for bfqq to be the next queue
1955 * to serve. In fact, to evaluate a sufficient condition, the
1956 * timestamps of the in-service queue would need to be
1957 * updated, and this operation is quite costly (see the
1958 * comments on bfq_bfqq_update_budg_for_activation()).
1960 * As for throughput, we ask bfq_better_to_idle() whether we
1961 * still need to plug I/O dispatching. If bfq_better_to_idle()
1962 * says no, then plugging is not needed any longer, either to
1963 * boost throughput or to perserve service guarantees. Then
1964 * the best option is to stop plugging I/O, as not doing so
1965 * would certainly lower throughput. We may end up in this
1966 * case if: (1) upon a dispatch attempt, we detected that it
1967 * was better to plug I/O dispatch, and to wait for a new
1968 * request to arrive for the currently in-service queue, but
1969 * (2) this switch of bfqq to busy changes the scenario.
1971 if (bfqd->in_service_queue &&
1972 ((bfqq_wants_to_preempt &&
1973 bfqq->wr_coeff >= bfqd->in_service_queue->wr_coeff) ||
1974 bfq_bfqq_higher_class_or_weight(bfqq, bfqd->in_service_queue) ||
1975 !bfq_better_to_idle(bfqd->in_service_queue)) &&
1976 next_queue_may_preempt(bfqd))
1977 bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
1978 false, BFQQE_PREEMPTED);
1981 static void bfq_reset_inject_limit(struct bfq_data *bfqd,
1982 struct bfq_queue *bfqq)
1984 /* invalidate baseline total service time */
1985 bfqq->last_serv_time_ns = 0;
1988 * Reset pointer in case we are waiting for
1989 * some request completion.
1991 bfqd->waited_rq = NULL;
1994 * If bfqq has a short think time, then start by setting the
1995 * inject limit to 0 prudentially, because the service time of
1996 * an injected I/O request may be higher than the think time
1997 * of bfqq, and therefore, if one request was injected when
1998 * bfqq remains empty, this injected request might delay the
1999 * service of the next I/O request for bfqq significantly. In
2000 * case bfqq can actually tolerate some injection, then the
2001 * adaptive update will however raise the limit soon. This
2002 * lucky circumstance holds exactly because bfqq has a short
2003 * think time, and thus, after remaining empty, is likely to
2004 * get new I/O enqueued---and then completed---before being
2005 * expired. This is the very pattern that gives the
2006 * limit-update algorithm the chance to measure the effect of
2007 * injection on request service times, and then to update the
2008 * limit accordingly.
2010 * However, in the following special case, the inject limit is
2011 * left to 1 even if the think time is short: bfqq's I/O is
2012 * synchronized with that of some other queue, i.e., bfqq may
2013 * receive new I/O only after the I/O of the other queue is
2014 * completed. Keeping the inject limit to 1 allows the
2015 * blocking I/O to be served while bfqq is in service. And
2016 * this is very convenient both for bfqq and for overall
2017 * throughput, as explained in detail in the comments in
2018 * bfq_update_has_short_ttime().
2020 * On the opposite end, if bfqq has a long think time, then
2021 * start directly by 1, because:
2022 * a) on the bright side, keeping at most one request in
2023 * service in the drive is unlikely to cause any harm to the
2024 * latency of bfqq's requests, as the service time of a single
2025 * request is likely to be lower than the think time of bfqq;
2026 * b) on the downside, after becoming empty, bfqq is likely to
2027 * expire before getting its next request. With this request
2028 * arrival pattern, it is very hard to sample total service
2029 * times and update the inject limit accordingly (see comments
2030 * on bfq_update_inject_limit()). So the limit is likely to be
2031 * never, or at least seldom, updated. As a consequence, by
2032 * setting the limit to 1, we avoid that no injection ever
2033 * occurs with bfqq. On the downside, this proactive step
2034 * further reduces chances to actually compute the baseline
2035 * total service time. Thus it reduces chances to execute the
2036 * limit-update algorithm and possibly raise the limit to more
2039 if (bfq_bfqq_has_short_ttime(bfqq))
2040 bfqq->inject_limit = 0;
2042 bfqq->inject_limit = 1;
2044 bfqq->decrease_time_jif = jiffies;
2047 static void bfq_update_io_intensity(struct bfq_queue *bfqq, u64 now_ns)
2049 u64 tot_io_time = now_ns - bfqq->io_start_time;
2051 if (RB_EMPTY_ROOT(&bfqq->sort_list) && bfqq->dispatched == 0)
2052 bfqq->tot_idle_time +=
2053 now_ns - bfqq->ttime.last_end_request;
2055 if (unlikely(bfq_bfqq_just_created(bfqq)))
2059 * Must be busy for at least about 80% of the time to be
2060 * considered I/O bound.
2062 if (bfqq->tot_idle_time * 5 > tot_io_time)
2063 bfq_clear_bfqq_IO_bound(bfqq);
2065 bfq_mark_bfqq_IO_bound(bfqq);
2068 * Keep an observation window of at most 200 ms in the past
2071 if (tot_io_time > 200 * NSEC_PER_MSEC) {
2072 bfqq->io_start_time = now_ns - (tot_io_time>>1);
2073 bfqq->tot_idle_time >>= 1;
2078 * Detect whether bfqq's I/O seems synchronized with that of some
2079 * other queue, i.e., whether bfqq, after remaining empty, happens to
2080 * receive new I/O only right after some I/O request of the other
2081 * queue has been completed. We call waker queue the other queue, and
2082 * we assume, for simplicity, that bfqq may have at most one waker
2085 * A remarkable throughput boost can be reached by unconditionally
2086 * injecting the I/O of the waker queue, every time a new
2087 * bfq_dispatch_request happens to be invoked while I/O is being
2088 * plugged for bfqq. In addition to boosting throughput, this
2089 * unblocks bfqq's I/O, thereby improving bandwidth and latency for
2090 * bfqq. Note that these same results may be achieved with the general
2091 * injection mechanism, but less effectively. For details on this
2092 * aspect, see the comments on the choice of the queue for injection
2093 * in bfq_select_queue().
2095 * Turning back to the detection of a waker queue, a queue Q is deemed as a
2096 * waker queue for bfqq if, for three consecutive times, bfqq happens to become
2097 * non empty right after a request of Q has been completed within given
2098 * timeout. In this respect, even if bfqq is empty, we do not check for a waker
2099 * if it still has some in-flight I/O. In fact, in this case bfqq is actually
2100 * still being served by the drive, and may receive new I/O on the completion
2101 * of some of the in-flight requests. In particular, on the first time, Q is
2102 * tentatively set as a candidate waker queue, while on the third consecutive
2103 * time that Q is detected, the field waker_bfqq is set to Q, to confirm that Q
2104 * is a waker queue for bfqq. These detection steps are performed only if bfqq
2105 * has a long think time, so as to make it more likely that bfqq's I/O is
2106 * actually being blocked by a synchronization. This last filter, plus the
2107 * above three-times requirement and time limit for detection, make false
2108 * positives less likely.
2112 * The sooner a waker queue is detected, the sooner throughput can be
2113 * boosted by injecting I/O from the waker queue. Fortunately,
2114 * detection is likely to be actually fast, for the following
2115 * reasons. While blocked by synchronization, bfqq has a long think
2116 * time. This implies that bfqq's inject limit is at least equal to 1
2117 * (see the comments in bfq_update_inject_limit()). So, thanks to
2118 * injection, the waker queue is likely to be served during the very
2119 * first I/O-plugging time interval for bfqq. This triggers the first
2120 * step of the detection mechanism. Thanks again to injection, the
2121 * candidate waker queue is then likely to be confirmed no later than
2122 * during the next I/O-plugging interval for bfqq.
2126 * On queue merging all waker information is lost.
2128 static void bfq_check_waker(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2131 char waker_name[MAX_BFQQ_NAME_LENGTH];
2133 if (!bfqd->last_completed_rq_bfqq ||
2134 bfqd->last_completed_rq_bfqq == bfqq ||
2135 bfq_bfqq_has_short_ttime(bfqq) ||
2136 bfqq->dispatched > 0 ||
2137 now_ns - bfqd->last_completion >= 4 * NSEC_PER_MSEC ||
2138 bfqd->last_completed_rq_bfqq == bfqq->waker_bfqq)
2142 * We reset waker detection logic also if too much time has passed
2143 * since the first detection. If wakeups are rare, pointless idling
2144 * doesn't hurt throughput that much. The condition below makes sure
2145 * we do not uselessly idle blocking waker in more than 1/64 cases.
2147 if (bfqd->last_completed_rq_bfqq !=
2148 bfqq->tentative_waker_bfqq ||
2149 now_ns > bfqq->waker_detection_started +
2150 128 * (u64)bfqd->bfq_slice_idle) {
2152 * First synchronization detected with a
2153 * candidate waker queue, or with a different
2154 * candidate waker queue from the current one.
2156 bfqq->tentative_waker_bfqq =
2157 bfqd->last_completed_rq_bfqq;
2158 bfqq->num_waker_detections = 1;
2159 bfqq->waker_detection_started = now_ns;
2160 bfq_bfqq_name(bfqq->tentative_waker_bfqq, waker_name,
2161 MAX_BFQQ_NAME_LENGTH);
2162 bfq_log_bfqq(bfqd, bfqq, "set tentative waker %s", waker_name);
2163 } else /* Same tentative waker queue detected again */
2164 bfqq->num_waker_detections++;
2166 if (bfqq->num_waker_detections == 3) {
2167 bfqq->waker_bfqq = bfqd->last_completed_rq_bfqq;
2168 bfqq->tentative_waker_bfqq = NULL;
2169 bfq_bfqq_name(bfqq->waker_bfqq, waker_name,
2170 MAX_BFQQ_NAME_LENGTH);
2171 bfq_log_bfqq(bfqd, bfqq, "set waker %s", waker_name);
2174 * If the waker queue disappears, then
2175 * bfqq->waker_bfqq must be reset. To
2176 * this goal, we maintain in each
2177 * waker queue a list, woken_list, of
2178 * all the queues that reference the
2179 * waker queue through their
2180 * waker_bfqq pointer. When the waker
2181 * queue exits, the waker_bfqq pointer
2182 * of all the queues in the woken_list
2185 * In addition, if bfqq is already in
2186 * the woken_list of a waker queue,
2187 * then, before being inserted into
2188 * the woken_list of a new waker
2189 * queue, bfqq must be removed from
2190 * the woken_list of the old waker
2193 if (!hlist_unhashed(&bfqq->woken_list_node))
2194 hlist_del_init(&bfqq->woken_list_node);
2195 hlist_add_head(&bfqq->woken_list_node,
2196 &bfqd->last_completed_rq_bfqq->woken_list);
2200 static void bfq_add_request(struct request *rq)
2202 struct bfq_queue *bfqq = RQ_BFQQ(rq);
2203 struct bfq_data *bfqd = bfqq->bfqd;
2204 struct request *next_rq, *prev;
2205 unsigned int old_wr_coeff = bfqq->wr_coeff;
2206 bool interactive = false;
2207 u64 now_ns = ktime_get_ns();
2209 bfq_log_bfqq(bfqd, bfqq, "add_request %d", rq_is_sync(rq));
2210 bfqq->queued[rq_is_sync(rq)]++;
2213 if (RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_bfqq_sync(bfqq)) {
2214 bfq_check_waker(bfqd, bfqq, now_ns);
2217 * Periodically reset inject limit, to make sure that
2218 * the latter eventually drops in case workload
2219 * changes, see step (3) in the comments on
2220 * bfq_update_inject_limit().
2222 if (time_is_before_eq_jiffies(bfqq->decrease_time_jif +
2223 msecs_to_jiffies(1000)))
2224 bfq_reset_inject_limit(bfqd, bfqq);
2227 * The following conditions must hold to setup a new
2228 * sampling of total service time, and then a new
2229 * update of the inject limit:
2230 * - bfqq is in service, because the total service
2231 * time is evaluated only for the I/O requests of
2232 * the queues in service;
2233 * - this is the right occasion to compute or to
2234 * lower the baseline total service time, because
2235 * there are actually no requests in the drive,
2237 * the baseline total service time is available, and
2238 * this is the right occasion to compute the other
2239 * quantity needed to update the inject limit, i.e.,
2240 * the total service time caused by the amount of
2241 * injection allowed by the current value of the
2242 * limit. It is the right occasion because injection
2243 * has actually been performed during the service
2244 * hole, and there are still in-flight requests,
2245 * which are very likely to be exactly the injected
2246 * requests, or part of them;
2247 * - the minimum interval for sampling the total
2248 * service time and updating the inject limit has
2251 if (bfqq == bfqd->in_service_queue &&
2252 (bfqd->rq_in_driver == 0 ||
2253 (bfqq->last_serv_time_ns > 0 &&
2254 bfqd->rqs_injected && bfqd->rq_in_driver > 0)) &&
2255 time_is_before_eq_jiffies(bfqq->decrease_time_jif +
2256 msecs_to_jiffies(10))) {
2257 bfqd->last_empty_occupied_ns = ktime_get_ns();
2259 * Start the state machine for measuring the
2260 * total service time of rq: setting
2261 * wait_dispatch will cause bfqd->waited_rq to
2262 * be set when rq will be dispatched.
2264 bfqd->wait_dispatch = true;
2266 * If there is no I/O in service in the drive,
2267 * then possible injection occurred before the
2268 * arrival of rq will not affect the total
2269 * service time of rq. So the injection limit
2270 * must not be updated as a function of such
2271 * total service time, unless new injection
2272 * occurs before rq is completed. To have the
2273 * injection limit updated only in the latter
2274 * case, reset rqs_injected here (rqs_injected
2275 * will be set in case injection is performed
2276 * on bfqq before rq is completed).
2278 if (bfqd->rq_in_driver == 0)
2279 bfqd->rqs_injected = false;
2283 if (bfq_bfqq_sync(bfqq))
2284 bfq_update_io_intensity(bfqq, now_ns);
2286 elv_rb_add(&bfqq->sort_list, rq);
2289 * Check if this request is a better next-serve candidate.
2291 prev = bfqq->next_rq;
2292 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position);
2293 bfqq->next_rq = next_rq;
2296 * Adjust priority tree position, if next_rq changes.
2297 * See comments on bfq_pos_tree_add_move() for the unlikely().
2299 if (unlikely(!bfqd->nonrot_with_queueing && prev != bfqq->next_rq))
2300 bfq_pos_tree_add_move(bfqd, bfqq);
2302 if (!bfq_bfqq_busy(bfqq)) /* switching to busy ... */
2303 bfq_bfqq_handle_idle_busy_switch(bfqd, bfqq, old_wr_coeff,
2306 if (bfqd->low_latency && old_wr_coeff == 1 && !rq_is_sync(rq) &&
2307 time_is_before_jiffies(
2308 bfqq->last_wr_start_finish +
2309 bfqd->bfq_wr_min_inter_arr_async)) {
2310 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
2311 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
2313 bfqd->wr_busy_queues++;
2314 bfqq->entity.prio_changed = 1;
2316 if (prev != bfqq->next_rq)
2317 bfq_updated_next_req(bfqd, bfqq);
2321 * Assign jiffies to last_wr_start_finish in the following
2324 * . if bfqq is not going to be weight-raised, because, for
2325 * non weight-raised queues, last_wr_start_finish stores the
2326 * arrival time of the last request; as of now, this piece
2327 * of information is used only for deciding whether to
2328 * weight-raise async queues
2330 * . if bfqq is not weight-raised, because, if bfqq is now
2331 * switching to weight-raised, then last_wr_start_finish
2332 * stores the time when weight-raising starts
2334 * . if bfqq is interactive, because, regardless of whether
2335 * bfqq is currently weight-raised, the weight-raising
2336 * period must start or restart (this case is considered
2337 * separately because it is not detected by the above
2338 * conditions, if bfqq is already weight-raised)
2340 * last_wr_start_finish has to be updated also if bfqq is soft
2341 * real-time, because the weight-raising period is constantly
2342 * restarted on idle-to-busy transitions for these queues, but
2343 * this is already done in bfq_bfqq_handle_idle_busy_switch if
2346 if (bfqd->low_latency &&
2347 (old_wr_coeff == 1 || bfqq->wr_coeff == 1 || interactive))
2348 bfqq->last_wr_start_finish = jiffies;
2351 static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd,
2353 struct request_queue *q)
2355 struct bfq_queue *bfqq = bfqd->bio_bfqq;
2359 return elv_rb_find(&bfqq->sort_list, bio_end_sector(bio));
2364 static sector_t get_sdist(sector_t last_pos, struct request *rq)
2367 return abs(blk_rq_pos(rq) - last_pos);
2372 #if 0 /* Still not clear if we can do without next two functions */
2373 static void bfq_activate_request(struct request_queue *q, struct request *rq)
2375 struct bfq_data *bfqd = q->elevator->elevator_data;
2377 bfqd->rq_in_driver++;
2380 static void bfq_deactivate_request(struct request_queue *q, struct request *rq)
2382 struct bfq_data *bfqd = q->elevator->elevator_data;
2384 bfqd->rq_in_driver--;
2388 static void bfq_remove_request(struct request_queue *q,
2391 struct bfq_queue *bfqq = RQ_BFQQ(rq);
2392 struct bfq_data *bfqd = bfqq->bfqd;
2393 const int sync = rq_is_sync(rq);
2395 if (bfqq->next_rq == rq) {
2396 bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq);
2397 bfq_updated_next_req(bfqd, bfqq);
2400 if (rq->queuelist.prev != &rq->queuelist)
2401 list_del_init(&rq->queuelist);
2402 bfqq->queued[sync]--;
2404 elv_rb_del(&bfqq->sort_list, rq);
2406 elv_rqhash_del(q, rq);
2407 if (q->last_merge == rq)
2408 q->last_merge = NULL;
2410 if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
2411 bfqq->next_rq = NULL;
2413 if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->in_service_queue) {
2414 bfq_del_bfqq_busy(bfqd, bfqq, false);
2416 * bfqq emptied. In normal operation, when
2417 * bfqq is empty, bfqq->entity.service and
2418 * bfqq->entity.budget must contain,
2419 * respectively, the service received and the
2420 * budget used last time bfqq emptied. These
2421 * facts do not hold in this case, as at least
2422 * this last removal occurred while bfqq is
2423 * not in service. To avoid inconsistencies,
2424 * reset both bfqq->entity.service and
2425 * bfqq->entity.budget, if bfqq has still a
2426 * process that may issue I/O requests to it.
2428 bfqq->entity.budget = bfqq->entity.service = 0;
2432 * Remove queue from request-position tree as it is empty.
2434 if (bfqq->pos_root) {
2435 rb_erase(&bfqq->pos_node, bfqq->pos_root);
2436 bfqq->pos_root = NULL;
2439 /* see comments on bfq_pos_tree_add_move() for the unlikely() */
2440 if (unlikely(!bfqd->nonrot_with_queueing))
2441 bfq_pos_tree_add_move(bfqd, bfqq);
2444 if (rq->cmd_flags & REQ_META)
2445 bfqq->meta_pending--;
2449 static bool bfq_bio_merge(struct request_queue *q, struct bio *bio,
2450 unsigned int nr_segs)
2452 struct bfq_data *bfqd = q->elevator->elevator_data;
2453 struct request *free = NULL;
2455 * bfq_bic_lookup grabs the queue_lock: invoke it now and
2456 * store its return value for later use, to avoid nesting
2457 * queue_lock inside the bfqd->lock. We assume that the bic
2458 * returned by bfq_bic_lookup does not go away before
2459 * bfqd->lock is taken.
2461 struct bfq_io_cq *bic = bfq_bic_lookup(q);
2464 spin_lock_irq(&bfqd->lock);
2467 bfqd->bio_bfqq = bic_to_bfqq(bic, op_is_sync(bio->bi_opf));
2469 bfqd->bio_bfqq = NULL;
2470 bfqd->bio_bic = bic;
2472 ret = blk_mq_sched_try_merge(q, bio, nr_segs, &free);
2474 spin_unlock_irq(&bfqd->lock);
2476 blk_mq_free_request(free);
2481 static int bfq_request_merge(struct request_queue *q, struct request **req,
2484 struct bfq_data *bfqd = q->elevator->elevator_data;
2485 struct request *__rq;
2487 __rq = bfq_find_rq_fmerge(bfqd, bio, q);
2488 if (__rq && elv_bio_merge_ok(__rq, bio)) {
2491 if (blk_discard_mergable(__rq))
2492 return ELEVATOR_DISCARD_MERGE;
2493 return ELEVATOR_FRONT_MERGE;
2496 return ELEVATOR_NO_MERGE;
2499 static struct bfq_queue *bfq_init_rq(struct request *rq);
2501 static void bfq_request_merged(struct request_queue *q, struct request *req,
2502 enum elv_merge type)
2504 if (type == ELEVATOR_FRONT_MERGE &&
2505 rb_prev(&req->rb_node) &&
2507 blk_rq_pos(container_of(rb_prev(&req->rb_node),
2508 struct request, rb_node))) {
2509 struct bfq_queue *bfqq = bfq_init_rq(req);
2510 struct bfq_data *bfqd;
2511 struct request *prev, *next_rq;
2518 /* Reposition request in its sort_list */
2519 elv_rb_del(&bfqq->sort_list, req);
2520 elv_rb_add(&bfqq->sort_list, req);
2522 /* Choose next request to be served for bfqq */
2523 prev = bfqq->next_rq;
2524 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, req,
2525 bfqd->last_position);
2526 bfqq->next_rq = next_rq;
2528 * If next_rq changes, update both the queue's budget to
2529 * fit the new request and the queue's position in its
2532 if (prev != bfqq->next_rq) {
2533 bfq_updated_next_req(bfqd, bfqq);
2535 * See comments on bfq_pos_tree_add_move() for
2538 if (unlikely(!bfqd->nonrot_with_queueing))
2539 bfq_pos_tree_add_move(bfqd, bfqq);
2545 * This function is called to notify the scheduler that the requests
2546 * rq and 'next' have been merged, with 'next' going away. BFQ
2547 * exploits this hook to address the following issue: if 'next' has a
2548 * fifo_time lower that rq, then the fifo_time of rq must be set to
2549 * the value of 'next', to not forget the greater age of 'next'.
2551 * NOTE: in this function we assume that rq is in a bfq_queue, basing
2552 * on that rq is picked from the hash table q->elevator->hash, which,
2553 * in its turn, is filled only with I/O requests present in
2554 * bfq_queues, while BFQ is in use for the request queue q. In fact,
2555 * the function that fills this hash table (elv_rqhash_add) is called
2556 * only by bfq_insert_request.
2558 static void bfq_requests_merged(struct request_queue *q, struct request *rq,
2559 struct request *next)
2561 struct bfq_queue *bfqq = bfq_init_rq(rq),
2562 *next_bfqq = bfq_init_rq(next);
2568 * If next and rq belong to the same bfq_queue and next is older
2569 * than rq, then reposition rq in the fifo (by substituting next
2570 * with rq). Otherwise, if next and rq belong to different
2571 * bfq_queues, never reposition rq: in fact, we would have to
2572 * reposition it with respect to next's position in its own fifo,
2573 * which would most certainly be too expensive with respect to
2576 if (bfqq == next_bfqq &&
2577 !list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2578 next->fifo_time < rq->fifo_time) {
2579 list_del_init(&rq->queuelist);
2580 list_replace_init(&next->queuelist, &rq->queuelist);
2581 rq->fifo_time = next->fifo_time;
2584 if (bfqq->next_rq == next)
2587 bfqg_stats_update_io_merged(bfqq_group(bfqq), next->cmd_flags);
2589 /* Merged request may be in the IO scheduler. Remove it. */
2590 if (!RB_EMPTY_NODE(&next->rb_node)) {
2591 bfq_remove_request(next->q, next);
2593 bfqg_stats_update_io_remove(bfqq_group(next_bfqq),
2598 /* Must be called with bfqq != NULL */
2599 static void bfq_bfqq_end_wr(struct bfq_queue *bfqq)
2602 * If bfqq has been enjoying interactive weight-raising, then
2603 * reset soft_rt_next_start. We do it for the following
2604 * reason. bfqq may have been conveying the I/O needed to load
2605 * a soft real-time application. Such an application actually
2606 * exhibits a soft real-time I/O pattern after it finishes
2607 * loading, and finally starts doing its job. But, if bfqq has
2608 * been receiving a lot of bandwidth so far (likely to happen
2609 * on a fast device), then soft_rt_next_start now contains a
2610 * high value that. So, without this reset, bfqq would be
2611 * prevented from being possibly considered as soft_rt for a
2615 if (bfqq->wr_cur_max_time !=
2616 bfqq->bfqd->bfq_wr_rt_max_time)
2617 bfqq->soft_rt_next_start = jiffies;
2619 if (bfq_bfqq_busy(bfqq))
2620 bfqq->bfqd->wr_busy_queues--;
2622 bfqq->wr_cur_max_time = 0;
2623 bfqq->last_wr_start_finish = jiffies;
2625 * Trigger a weight change on the next invocation of
2626 * __bfq_entity_update_weight_prio.
2628 bfqq->entity.prio_changed = 1;
2631 void bfq_end_wr_async_queues(struct bfq_data *bfqd,
2632 struct bfq_group *bfqg)
2636 for (i = 0; i < 2; i++)
2637 for (j = 0; j < IOPRIO_NR_LEVELS; j++)
2638 if (bfqg->async_bfqq[i][j])
2639 bfq_bfqq_end_wr(bfqg->async_bfqq[i][j]);
2640 if (bfqg->async_idle_bfqq)
2641 bfq_bfqq_end_wr(bfqg->async_idle_bfqq);
2644 static void bfq_end_wr(struct bfq_data *bfqd)
2646 struct bfq_queue *bfqq;
2648 spin_lock_irq(&bfqd->lock);
2650 list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
2651 bfq_bfqq_end_wr(bfqq);
2652 list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list)
2653 bfq_bfqq_end_wr(bfqq);
2654 bfq_end_wr_async(bfqd);
2656 spin_unlock_irq(&bfqd->lock);
2659 static sector_t bfq_io_struct_pos(void *io_struct, bool request)
2662 return blk_rq_pos(io_struct);
2664 return ((struct bio *)io_struct)->bi_iter.bi_sector;
2667 static int bfq_rq_close_to_sector(void *io_struct, bool request,
2670 return abs(bfq_io_struct_pos(io_struct, request) - sector) <=
2674 static struct bfq_queue *bfqq_find_close(struct bfq_data *bfqd,
2675 struct bfq_queue *bfqq,
2678 struct rb_root *root = &bfqq_group(bfqq)->rq_pos_tree;
2679 struct rb_node *parent, *node;
2680 struct bfq_queue *__bfqq;
2682 if (RB_EMPTY_ROOT(root))
2686 * First, if we find a request starting at the end of the last
2687 * request, choose it.
2689 __bfqq = bfq_rq_pos_tree_lookup(bfqd, root, sector, &parent, NULL);
2694 * If the exact sector wasn't found, the parent of the NULL leaf
2695 * will contain the closest sector (rq_pos_tree sorted by
2696 * next_request position).
2698 __bfqq = rb_entry(parent, struct bfq_queue, pos_node);
2699 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2702 if (blk_rq_pos(__bfqq->next_rq) < sector)
2703 node = rb_next(&__bfqq->pos_node);
2705 node = rb_prev(&__bfqq->pos_node);
2709 __bfqq = rb_entry(node, struct bfq_queue, pos_node);
2710 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2716 static struct bfq_queue *bfq_find_close_cooperator(struct bfq_data *bfqd,
2717 struct bfq_queue *cur_bfqq,
2720 struct bfq_queue *bfqq;
2723 * We shall notice if some of the queues are cooperating,
2724 * e.g., working closely on the same area of the device. In
2725 * that case, we can group them together and: 1) don't waste
2726 * time idling, and 2) serve the union of their requests in
2727 * the best possible order for throughput.
2729 bfqq = bfqq_find_close(bfqd, cur_bfqq, sector);
2730 if (!bfqq || bfqq == cur_bfqq)
2736 static struct bfq_queue *
2737 bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2739 int process_refs, new_process_refs;
2740 struct bfq_queue *__bfqq;
2743 * If there are no process references on the new_bfqq, then it is
2744 * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain
2745 * may have dropped their last reference (not just their last process
2748 if (!bfqq_process_refs(new_bfqq))
2751 /* Avoid a circular list and skip interim queue merges. */
2752 while ((__bfqq = new_bfqq->new_bfqq)) {
2758 process_refs = bfqq_process_refs(bfqq);
2759 new_process_refs = bfqq_process_refs(new_bfqq);
2761 * If the process for the bfqq has gone away, there is no
2762 * sense in merging the queues.
2764 if (process_refs == 0 || new_process_refs == 0)
2767 bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d",
2771 * Merging is just a redirection: the requests of the process
2772 * owning one of the two queues are redirected to the other queue.
2773 * The latter queue, in its turn, is set as shared if this is the
2774 * first time that the requests of some process are redirected to
2777 * We redirect bfqq to new_bfqq and not the opposite, because
2778 * we are in the context of the process owning bfqq, thus we
2779 * have the io_cq of this process. So we can immediately
2780 * configure this io_cq to redirect the requests of the
2781 * process to new_bfqq. In contrast, the io_cq of new_bfqq is
2782 * not available any more (new_bfqq->bic == NULL).
2784 * Anyway, even in case new_bfqq coincides with the in-service
2785 * queue, redirecting requests the in-service queue is the
2786 * best option, as we feed the in-service queue with new
2787 * requests close to the last request served and, by doing so,
2788 * are likely to increase the throughput.
2790 bfqq->new_bfqq = new_bfqq;
2792 * The above assignment schedules the following redirections:
2793 * each time some I/O for bfqq arrives, the process that
2794 * generated that I/O is disassociated from bfqq and
2795 * associated with new_bfqq. Here we increases new_bfqq->ref
2796 * in advance, adding the number of processes that are
2797 * expected to be associated with new_bfqq as they happen to
2800 new_bfqq->ref += process_refs;
2804 static bool bfq_may_be_close_cooperator(struct bfq_queue *bfqq,
2805 struct bfq_queue *new_bfqq)
2807 if (bfq_too_late_for_merging(new_bfqq))
2810 if (bfq_class_idle(bfqq) || bfq_class_idle(new_bfqq) ||
2811 (bfqq->ioprio_class != new_bfqq->ioprio_class))
2815 * If either of the queues has already been detected as seeky,
2816 * then merging it with the other queue is unlikely to lead to
2819 if (BFQQ_SEEKY(bfqq) || BFQQ_SEEKY(new_bfqq))
2823 * Interleaved I/O is known to be done by (some) applications
2824 * only for reads, so it does not make sense to merge async
2827 if (!bfq_bfqq_sync(bfqq) || !bfq_bfqq_sync(new_bfqq))
2833 static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd,
2834 struct bfq_queue *bfqq);
2837 * Attempt to schedule a merge of bfqq with the currently in-service
2838 * queue or with a close queue among the scheduled queues. Return
2839 * NULL if no merge was scheduled, a pointer to the shared bfq_queue
2840 * structure otherwise.
2842 * The OOM queue is not allowed to participate to cooperation: in fact, since
2843 * the requests temporarily redirected to the OOM queue could be redirected
2844 * again to dedicated queues at any time, the state needed to correctly
2845 * handle merging with the OOM queue would be quite complex and expensive
2846 * to maintain. Besides, in such a critical condition as an out of memory,
2847 * the benefits of queue merging may be little relevant, or even negligible.
2849 * WARNING: queue merging may impair fairness among non-weight raised
2850 * queues, for at least two reasons: 1) the original weight of a
2851 * merged queue may change during the merged state, 2) even being the
2852 * weight the same, a merged queue may be bloated with many more
2853 * requests than the ones produced by its originally-associated
2856 static struct bfq_queue *
2857 bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2858 void *io_struct, bool request, struct bfq_io_cq *bic)
2860 struct bfq_queue *in_service_bfqq, *new_bfqq;
2862 /* if a merge has already been setup, then proceed with that first */
2864 return bfqq->new_bfqq;
2867 * Check delayed stable merge for rotational or non-queueing
2868 * devs. For this branch to be executed, bfqq must not be
2869 * currently merged with some other queue (i.e., bfqq->bic
2870 * must be non null). If we considered also merged queues,
2871 * then we should also check whether bfqq has already been
2872 * merged with bic->stable_merge_bfqq. But this would be
2873 * costly and complicated.
2875 if (unlikely(!bfqd->nonrot_with_queueing)) {
2877 * Make sure also that bfqq is sync, because
2878 * bic->stable_merge_bfqq may point to some queue (for
2879 * stable merging) also if bic is associated with a
2880 * sync queue, but this bfqq is async
2882 if (bfq_bfqq_sync(bfqq) && bic->stable_merge_bfqq &&
2883 !bfq_bfqq_just_created(bfqq) &&
2884 time_is_before_jiffies(bfqq->split_time +
2885 msecs_to_jiffies(bfq_late_stable_merging)) &&
2886 time_is_before_jiffies(bfqq->creation_time +
2887 msecs_to_jiffies(bfq_late_stable_merging))) {
2888 struct bfq_queue *stable_merge_bfqq =
2889 bic->stable_merge_bfqq;
2890 int proc_ref = min(bfqq_process_refs(bfqq),
2891 bfqq_process_refs(stable_merge_bfqq));
2893 /* deschedule stable merge, because done or aborted here */
2894 bfq_put_stable_ref(stable_merge_bfqq);
2896 bic->stable_merge_bfqq = NULL;
2898 if (!idling_boosts_thr_without_issues(bfqd, bfqq) &&
2900 /* next function will take at least one ref */
2901 struct bfq_queue *new_bfqq =
2902 bfq_setup_merge(bfqq, stable_merge_bfqq);
2904 bic->stably_merged = true;
2905 if (new_bfqq && new_bfqq->bic)
2906 new_bfqq->bic->stably_merged = true;
2914 * Do not perform queue merging if the device is non
2915 * rotational and performs internal queueing. In fact, such a
2916 * device reaches a high speed through internal parallelism
2917 * and pipelining. This means that, to reach a high
2918 * throughput, it must have many requests enqueued at the same
2919 * time. But, in this configuration, the internal scheduling
2920 * algorithm of the device does exactly the job of queue
2921 * merging: it reorders requests so as to obtain as much as
2922 * possible a sequential I/O pattern. As a consequence, with
2923 * the workload generated by processes doing interleaved I/O,
2924 * the throughput reached by the device is likely to be the
2925 * same, with and without queue merging.
2927 * Disabling merging also provides a remarkable benefit in
2928 * terms of throughput. Merging tends to make many workloads
2929 * artificially more uneven, because of shared queues
2930 * remaining non empty for incomparably more time than
2931 * non-merged queues. This may accentuate workload
2932 * asymmetries. For example, if one of the queues in a set of
2933 * merged queues has a higher weight than a normal queue, then
2934 * the shared queue may inherit such a high weight and, by
2935 * staying almost always active, may force BFQ to perform I/O
2936 * plugging most of the time. This evidently makes it harder
2937 * for BFQ to let the device reach a high throughput.
2939 * Finally, the likely() macro below is not used because one
2940 * of the two branches is more likely than the other, but to
2941 * have the code path after the following if() executed as
2942 * fast as possible for the case of a non rotational device
2943 * with queueing. We want it because this is the fastest kind
2944 * of device. On the opposite end, the likely() may lengthen
2945 * the execution time of BFQ for the case of slower devices
2946 * (rotational or at least without queueing). But in this case
2947 * the execution time of BFQ matters very little, if not at
2950 if (likely(bfqd->nonrot_with_queueing))
2954 * Prevent bfqq from being merged if it has been created too
2955 * long ago. The idea is that true cooperating processes, and
2956 * thus their associated bfq_queues, are supposed to be
2957 * created shortly after each other. This is the case, e.g.,
2958 * for KVM/QEMU and dump I/O threads. Basing on this
2959 * assumption, the following filtering greatly reduces the
2960 * probability that two non-cooperating processes, which just
2961 * happen to do close I/O for some short time interval, have
2962 * their queues merged by mistake.
2964 if (bfq_too_late_for_merging(bfqq))
2967 if (!io_struct || unlikely(bfqq == &bfqd->oom_bfqq))
2970 /* If there is only one backlogged queue, don't search. */
2971 if (bfq_tot_busy_queues(bfqd) == 1)
2974 in_service_bfqq = bfqd->in_service_queue;
2976 if (in_service_bfqq && in_service_bfqq != bfqq &&
2977 likely(in_service_bfqq != &bfqd->oom_bfqq) &&
2978 bfq_rq_close_to_sector(io_struct, request,
2979 bfqd->in_serv_last_pos) &&
2980 bfqq->entity.parent == in_service_bfqq->entity.parent &&
2981 bfq_may_be_close_cooperator(bfqq, in_service_bfqq)) {
2982 new_bfqq = bfq_setup_merge(bfqq, in_service_bfqq);
2987 * Check whether there is a cooperator among currently scheduled
2988 * queues. The only thing we need is that the bio/request is not
2989 * NULL, as we need it to establish whether a cooperator exists.
2991 new_bfqq = bfq_find_close_cooperator(bfqd, bfqq,
2992 bfq_io_struct_pos(io_struct, request));
2994 if (new_bfqq && likely(new_bfqq != &bfqd->oom_bfqq) &&
2995 bfq_may_be_close_cooperator(bfqq, new_bfqq))
2996 return bfq_setup_merge(bfqq, new_bfqq);
3001 static void bfq_bfqq_save_state(struct bfq_queue *bfqq)
3003 struct bfq_io_cq *bic = bfqq->bic;
3006 * If !bfqq->bic, the queue is already shared or its requests
3007 * have already been redirected to a shared queue; both idle window
3008 * and weight raising state have already been saved. Do nothing.
3013 bic->saved_last_serv_time_ns = bfqq->last_serv_time_ns;
3014 bic->saved_inject_limit = bfqq->inject_limit;
3015 bic->saved_decrease_time_jif = bfqq->decrease_time_jif;
3017 bic->saved_weight = bfqq->entity.orig_weight;
3018 bic->saved_ttime = bfqq->ttime;
3019 bic->saved_has_short_ttime = bfq_bfqq_has_short_ttime(bfqq);
3020 bic->saved_IO_bound = bfq_bfqq_IO_bound(bfqq);
3021 bic->saved_io_start_time = bfqq->io_start_time;
3022 bic->saved_tot_idle_time = bfqq->tot_idle_time;
3023 bic->saved_in_large_burst = bfq_bfqq_in_large_burst(bfqq);
3024 bic->was_in_burst_list = !hlist_unhashed(&bfqq->burst_list_node);
3025 if (unlikely(bfq_bfqq_just_created(bfqq) &&
3026 !bfq_bfqq_in_large_burst(bfqq) &&
3027 bfqq->bfqd->low_latency)) {
3029 * bfqq being merged right after being created: bfqq
3030 * would have deserved interactive weight raising, but
3031 * did not make it to be set in a weight-raised state,
3032 * because of this early merge. Store directly the
3033 * weight-raising state that would have been assigned
3034 * to bfqq, so that to avoid that bfqq unjustly fails
3035 * to enjoy weight raising if split soon.
3037 bic->saved_wr_coeff = bfqq->bfqd->bfq_wr_coeff;
3038 bic->saved_wr_start_at_switch_to_srt = bfq_smallest_from_now();
3039 bic->saved_wr_cur_max_time = bfq_wr_duration(bfqq->bfqd);
3040 bic->saved_last_wr_start_finish = jiffies;
3042 bic->saved_wr_coeff = bfqq->wr_coeff;
3043 bic->saved_wr_start_at_switch_to_srt =
3044 bfqq->wr_start_at_switch_to_srt;
3045 bic->saved_service_from_wr = bfqq->service_from_wr;
3046 bic->saved_last_wr_start_finish = bfqq->last_wr_start_finish;
3047 bic->saved_wr_cur_max_time = bfqq->wr_cur_max_time;
3053 bfq_reassign_last_bfqq(struct bfq_queue *cur_bfqq, struct bfq_queue *new_bfqq)
3055 if (cur_bfqq->entity.parent &&
3056 cur_bfqq->entity.parent->last_bfqq_created == cur_bfqq)
3057 cur_bfqq->entity.parent->last_bfqq_created = new_bfqq;
3058 else if (cur_bfqq->bfqd && cur_bfqq->bfqd->last_bfqq_created == cur_bfqq)
3059 cur_bfqq->bfqd->last_bfqq_created = new_bfqq;
3062 void bfq_release_process_ref(struct bfq_data *bfqd, struct bfq_queue *bfqq)
3065 * To prevent bfqq's service guarantees from being violated,
3066 * bfqq may be left busy, i.e., queued for service, even if
3067 * empty (see comments in __bfq_bfqq_expire() for
3068 * details). But, if no process will send requests to bfqq any
3069 * longer, then there is no point in keeping bfqq queued for
3070 * service. In addition, keeping bfqq queued for service, but
3071 * with no process ref any longer, may have caused bfqq to be
3072 * freed when dequeued from service. But this is assumed to
3075 if (bfq_bfqq_busy(bfqq) && RB_EMPTY_ROOT(&bfqq->sort_list) &&
3076 bfqq != bfqd->in_service_queue)
3077 bfq_del_bfqq_busy(bfqd, bfqq, false);
3079 bfq_reassign_last_bfqq(bfqq, NULL);
3081 bfq_put_queue(bfqq);
3085 bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic,
3086 struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
3088 bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu",
3089 (unsigned long)new_bfqq->pid);
3090 /* Save weight raising and idle window of the merged queues */
3091 bfq_bfqq_save_state(bfqq);
3092 bfq_bfqq_save_state(new_bfqq);
3093 if (bfq_bfqq_IO_bound(bfqq))
3094 bfq_mark_bfqq_IO_bound(new_bfqq);
3095 bfq_clear_bfqq_IO_bound(bfqq);
3098 * The processes associated with bfqq are cooperators of the
3099 * processes associated with new_bfqq. So, if bfqq has a
3100 * waker, then assume that all these processes will be happy
3101 * to let bfqq's waker freely inject I/O when they have no
3104 if (bfqq->waker_bfqq && !new_bfqq->waker_bfqq &&
3105 bfqq->waker_bfqq != new_bfqq) {
3106 new_bfqq->waker_bfqq = bfqq->waker_bfqq;
3107 new_bfqq->tentative_waker_bfqq = NULL;
3110 * If the waker queue disappears, then
3111 * new_bfqq->waker_bfqq must be reset. So insert
3112 * new_bfqq into the woken_list of the waker. See
3113 * bfq_check_waker for details.
3115 hlist_add_head(&new_bfqq->woken_list_node,
3116 &new_bfqq->waker_bfqq->woken_list);
3121 * If bfqq is weight-raised, then let new_bfqq inherit
3122 * weight-raising. To reduce false positives, neglect the case
3123 * where bfqq has just been created, but has not yet made it
3124 * to be weight-raised (which may happen because EQM may merge
3125 * bfqq even before bfq_add_request is executed for the first
3126 * time for bfqq). Handling this case would however be very
3127 * easy, thanks to the flag just_created.
3129 if (new_bfqq->wr_coeff == 1 && bfqq->wr_coeff > 1) {
3130 new_bfqq->wr_coeff = bfqq->wr_coeff;
3131 new_bfqq->wr_cur_max_time = bfqq->wr_cur_max_time;
3132 new_bfqq->last_wr_start_finish = bfqq->last_wr_start_finish;
3133 new_bfqq->wr_start_at_switch_to_srt =
3134 bfqq->wr_start_at_switch_to_srt;
3135 if (bfq_bfqq_busy(new_bfqq))
3136 bfqd->wr_busy_queues++;
3137 new_bfqq->entity.prio_changed = 1;
3140 if (bfqq->wr_coeff > 1) { /* bfqq has given its wr to new_bfqq */
3142 bfqq->entity.prio_changed = 1;
3143 if (bfq_bfqq_busy(bfqq))
3144 bfqd->wr_busy_queues--;
3147 bfq_log_bfqq(bfqd, new_bfqq, "merge_bfqqs: wr_busy %d",
3148 bfqd->wr_busy_queues);
3151 * Merge queues (that is, let bic redirect its requests to new_bfqq)
3153 bic_set_bfqq(bic, new_bfqq, 1);
3154 bfq_mark_bfqq_coop(new_bfqq);
3156 * new_bfqq now belongs to at least two bics (it is a shared queue):
3157 * set new_bfqq->bic to NULL. bfqq either:
3158 * - does not belong to any bic any more, and hence bfqq->bic must
3159 * be set to NULL, or
3160 * - is a queue whose owning bics have already been redirected to a
3161 * different queue, hence the queue is destined to not belong to
3162 * any bic soon and bfqq->bic is already NULL (therefore the next
3163 * assignment causes no harm).
3165 new_bfqq->bic = NULL;
3167 * If the queue is shared, the pid is the pid of one of the associated
3168 * processes. Which pid depends on the exact sequence of merge events
3169 * the queue underwent. So printing such a pid is useless and confusing
3170 * because it reports a random pid between those of the associated
3172 * We mark such a queue with a pid -1, and then print SHARED instead of
3173 * a pid in logging messages.
3178 bfq_reassign_last_bfqq(bfqq, new_bfqq);
3180 bfq_release_process_ref(bfqd, bfqq);
3183 static bool bfq_allow_bio_merge(struct request_queue *q, struct request *rq,
3186 struct bfq_data *bfqd = q->elevator->elevator_data;
3187 bool is_sync = op_is_sync(bio->bi_opf);
3188 struct bfq_queue *bfqq = bfqd->bio_bfqq, *new_bfqq;
3191 * Disallow merge of a sync bio into an async request.
3193 if (is_sync && !rq_is_sync(rq))
3197 * Lookup the bfqq that this bio will be queued with. Allow
3198 * merge only if rq is queued there.
3204 * We take advantage of this function to perform an early merge
3205 * of the queues of possible cooperating processes.
3207 new_bfqq = bfq_setup_cooperator(bfqd, bfqq, bio, false, bfqd->bio_bic);
3210 * bic still points to bfqq, then it has not yet been
3211 * redirected to some other bfq_queue, and a queue
3212 * merge between bfqq and new_bfqq can be safely
3213 * fulfilled, i.e., bic can be redirected to new_bfqq
3214 * and bfqq can be put.
3216 bfq_merge_bfqqs(bfqd, bfqd->bio_bic, bfqq,
3219 * If we get here, bio will be queued into new_queue,
3220 * so use new_bfqq to decide whether bio and rq can be
3226 * Change also bqfd->bio_bfqq, as
3227 * bfqd->bio_bic now points to new_bfqq, and
3228 * this function may be invoked again (and then may
3229 * use again bqfd->bio_bfqq).
3231 bfqd->bio_bfqq = bfqq;
3234 return bfqq == RQ_BFQQ(rq);
3238 * Set the maximum time for the in-service queue to consume its
3239 * budget. This prevents seeky processes from lowering the throughput.
3240 * In practice, a time-slice service scheme is used with seeky
3243 static void bfq_set_budget_timeout(struct bfq_data *bfqd,
3244 struct bfq_queue *bfqq)
3246 unsigned int timeout_coeff;
3248 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time)
3251 timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight;
3253 bfqd->last_budget_start = ktime_get();
3255 bfqq->budget_timeout = jiffies +
3256 bfqd->bfq_timeout * timeout_coeff;
3259 static void __bfq_set_in_service_queue(struct bfq_data *bfqd,
3260 struct bfq_queue *bfqq)
3263 bfq_clear_bfqq_fifo_expire(bfqq);
3265 bfqd->budgets_assigned = (bfqd->budgets_assigned * 7 + 256) / 8;
3267 if (time_is_before_jiffies(bfqq->last_wr_start_finish) &&
3268 bfqq->wr_coeff > 1 &&
3269 bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
3270 time_is_before_jiffies(bfqq->budget_timeout)) {
3272 * For soft real-time queues, move the start
3273 * of the weight-raising period forward by the
3274 * time the queue has not received any
3275 * service. Otherwise, a relatively long
3276 * service delay is likely to cause the
3277 * weight-raising period of the queue to end,
3278 * because of the short duration of the
3279 * weight-raising period of a soft real-time
3280 * queue. It is worth noting that this move
3281 * is not so dangerous for the other queues,
3282 * because soft real-time queues are not
3285 * To not add a further variable, we use the
3286 * overloaded field budget_timeout to
3287 * determine for how long the queue has not
3288 * received service, i.e., how much time has
3289 * elapsed since the queue expired. However,
3290 * this is a little imprecise, because
3291 * budget_timeout is set to jiffies if bfqq
3292 * not only expires, but also remains with no
3295 if (time_after(bfqq->budget_timeout,
3296 bfqq->last_wr_start_finish))
3297 bfqq->last_wr_start_finish +=
3298 jiffies - bfqq->budget_timeout;
3300 bfqq->last_wr_start_finish = jiffies;
3303 bfq_set_budget_timeout(bfqd, bfqq);
3304 bfq_log_bfqq(bfqd, bfqq,
3305 "set_in_service_queue, cur-budget = %d",
3306 bfqq->entity.budget);
3309 bfqd->in_service_queue = bfqq;
3310 bfqd->in_serv_last_pos = 0;
3314 * Get and set a new queue for service.
3316 static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd)
3318 struct bfq_queue *bfqq = bfq_get_next_queue(bfqd);
3320 __bfq_set_in_service_queue(bfqd, bfqq);
3324 static void bfq_arm_slice_timer(struct bfq_data *bfqd)
3326 struct bfq_queue *bfqq = bfqd->in_service_queue;
3329 bfq_mark_bfqq_wait_request(bfqq);
3332 * We don't want to idle for seeks, but we do want to allow
3333 * fair distribution of slice time for a process doing back-to-back
3334 * seeks. So allow a little bit of time for him to submit a new rq.
3336 sl = bfqd->bfq_slice_idle;
3338 * Unless the queue is being weight-raised or the scenario is
3339 * asymmetric, grant only minimum idle time if the queue
3340 * is seeky. A long idling is preserved for a weight-raised
3341 * queue, or, more in general, in an asymmetric scenario,
3342 * because a long idling is needed for guaranteeing to a queue
3343 * its reserved share of the throughput (in particular, it is
3344 * needed if the queue has a higher weight than some other
3347 if (BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 &&
3348 !bfq_asymmetric_scenario(bfqd, bfqq))
3349 sl = min_t(u64, sl, BFQ_MIN_TT);
3350 else if (bfqq->wr_coeff > 1)
3351 sl = max_t(u32, sl, 20ULL * NSEC_PER_MSEC);
3353 bfqd->last_idling_start = ktime_get();
3354 bfqd->last_idling_start_jiffies = jiffies;
3356 hrtimer_start(&bfqd->idle_slice_timer, ns_to_ktime(sl),
3358 bfqg_stats_set_start_idle_time(bfqq_group(bfqq));
3362 * In autotuning mode, max_budget is dynamically recomputed as the
3363 * amount of sectors transferred in timeout at the estimated peak
3364 * rate. This enables BFQ to utilize a full timeslice with a full
3365 * budget, even if the in-service queue is served at peak rate. And
3366 * this maximises throughput with sequential workloads.
3368 static unsigned long bfq_calc_max_budget(struct bfq_data *bfqd)
3370 return (u64)bfqd->peak_rate * USEC_PER_MSEC *
3371 jiffies_to_msecs(bfqd->bfq_timeout)>>BFQ_RATE_SHIFT;
3375 * Update parameters related to throughput and responsiveness, as a
3376 * function of the estimated peak rate. See comments on
3377 * bfq_calc_max_budget(), and on the ref_wr_duration array.
3379 static void update_thr_responsiveness_params(struct bfq_data *bfqd)
3381 if (bfqd->bfq_user_max_budget == 0) {
3382 bfqd->bfq_max_budget =
3383 bfq_calc_max_budget(bfqd);
3384 bfq_log(bfqd, "new max_budget = %d", bfqd->bfq_max_budget);
3388 static void bfq_reset_rate_computation(struct bfq_data *bfqd,
3391 if (rq != NULL) { /* new rq dispatch now, reset accordingly */
3392 bfqd->last_dispatch = bfqd->first_dispatch = ktime_get_ns();
3393 bfqd->peak_rate_samples = 1;
3394 bfqd->sequential_samples = 0;
3395 bfqd->tot_sectors_dispatched = bfqd->last_rq_max_size =
3397 } else /* no new rq dispatched, just reset the number of samples */
3398 bfqd->peak_rate_samples = 0; /* full re-init on next disp. */
3401 "reset_rate_computation at end, sample %u/%u tot_sects %llu",
3402 bfqd->peak_rate_samples, bfqd->sequential_samples,
3403 bfqd->tot_sectors_dispatched);
3406 static void bfq_update_rate_reset(struct bfq_data *bfqd, struct request *rq)
3408 u32 rate, weight, divisor;
3411 * For the convergence property to hold (see comments on
3412 * bfq_update_peak_rate()) and for the assessment to be
3413 * reliable, a minimum number of samples must be present, and
3414 * a minimum amount of time must have elapsed. If not so, do
3415 * not compute new rate. Just reset parameters, to get ready
3416 * for a new evaluation attempt.
3418 if (bfqd->peak_rate_samples < BFQ_RATE_MIN_SAMPLES ||
3419 bfqd->delta_from_first < BFQ_RATE_MIN_INTERVAL)
3420 goto reset_computation;
3423 * If a new request completion has occurred after last
3424 * dispatch, then, to approximate the rate at which requests
3425 * have been served by the device, it is more precise to
3426 * extend the observation interval to the last completion.
3428 bfqd->delta_from_first =
3429 max_t(u64, bfqd->delta_from_first,
3430 bfqd->last_completion - bfqd->first_dispatch);
3433 * Rate computed in sects/usec, and not sects/nsec, for
3436 rate = div64_ul(bfqd->tot_sectors_dispatched<<BFQ_RATE_SHIFT,
3437 div_u64(bfqd->delta_from_first, NSEC_PER_USEC));
3440 * Peak rate not updated if:
3441 * - the percentage of sequential dispatches is below 3/4 of the
3442 * total, and rate is below the current estimated peak rate
3443 * - rate is unreasonably high (> 20M sectors/sec)
3445 if ((bfqd->sequential_samples < (3 * bfqd->peak_rate_samples)>>2 &&
3446 rate <= bfqd->peak_rate) ||
3447 rate > 20<<BFQ_RATE_SHIFT)
3448 goto reset_computation;
3451 * We have to update the peak rate, at last! To this purpose,
3452 * we use a low-pass filter. We compute the smoothing constant
3453 * of the filter as a function of the 'weight' of the new
3456 * As can be seen in next formulas, we define this weight as a
3457 * quantity proportional to how sequential the workload is,
3458 * and to how long the observation time interval is.
3460 * The weight runs from 0 to 8. The maximum value of the
3461 * weight, 8, yields the minimum value for the smoothing
3462 * constant. At this minimum value for the smoothing constant,
3463 * the measured rate contributes for half of the next value of
3464 * the estimated peak rate.
3466 * So, the first step is to compute the weight as a function
3467 * of how sequential the workload is. Note that the weight
3468 * cannot reach 9, because bfqd->sequential_samples cannot
3469 * become equal to bfqd->peak_rate_samples, which, in its
3470 * turn, holds true because bfqd->sequential_samples is not
3471 * incremented for the first sample.
3473 weight = (9 * bfqd->sequential_samples) / bfqd->peak_rate_samples;
3476 * Second step: further refine the weight as a function of the
3477 * duration of the observation interval.
3479 weight = min_t(u32, 8,
3480 div_u64(weight * bfqd->delta_from_first,
3481 BFQ_RATE_REF_INTERVAL));
3484 * Divisor ranging from 10, for minimum weight, to 2, for
3487 divisor = 10 - weight;
3490 * Finally, update peak rate:
3492 * peak_rate = peak_rate * (divisor-1) / divisor + rate / divisor
3494 bfqd->peak_rate *= divisor-1;
3495 bfqd->peak_rate /= divisor;
3496 rate /= divisor; /* smoothing constant alpha = 1/divisor */
3498 bfqd->peak_rate += rate;
3501 * For a very slow device, bfqd->peak_rate can reach 0 (see
3502 * the minimum representable values reported in the comments
3503 * on BFQ_RATE_SHIFT). Push to 1 if this happens, to avoid
3504 * divisions by zero where bfqd->peak_rate is used as a
3507 bfqd->peak_rate = max_t(u32, 1, bfqd->peak_rate);
3509 update_thr_responsiveness_params(bfqd);
3512 bfq_reset_rate_computation(bfqd, rq);
3516 * Update the read/write peak rate (the main quantity used for
3517 * auto-tuning, see update_thr_responsiveness_params()).
3519 * It is not trivial to estimate the peak rate (correctly): because of
3520 * the presence of sw and hw queues between the scheduler and the
3521 * device components that finally serve I/O requests, it is hard to
3522 * say exactly when a given dispatched request is served inside the
3523 * device, and for how long. As a consequence, it is hard to know
3524 * precisely at what rate a given set of requests is actually served
3527 * On the opposite end, the dispatch time of any request is trivially
3528 * available, and, from this piece of information, the "dispatch rate"
3529 * of requests can be immediately computed. So, the idea in the next
3530 * function is to use what is known, namely request dispatch times
3531 * (plus, when useful, request completion times), to estimate what is
3532 * unknown, namely in-device request service rate.
3534 * The main issue is that, because of the above facts, the rate at
3535 * which a certain set of requests is dispatched over a certain time
3536 * interval can vary greatly with respect to the rate at which the
3537 * same requests are then served. But, since the size of any
3538 * intermediate queue is limited, and the service scheme is lossless
3539 * (no request is silently dropped), the following obvious convergence
3540 * property holds: the number of requests dispatched MUST become
3541 * closer and closer to the number of requests completed as the
3542 * observation interval grows. This is the key property used in
3543 * the next function to estimate the peak service rate as a function
3544 * of the observed dispatch rate. The function assumes to be invoked
3545 * on every request dispatch.
3547 static void bfq_update_peak_rate(struct bfq_data *bfqd, struct request *rq)
3549 u64 now_ns = ktime_get_ns();
3551 if (bfqd->peak_rate_samples == 0) { /* first dispatch */
3552 bfq_log(bfqd, "update_peak_rate: goto reset, samples %d",
3553 bfqd->peak_rate_samples);
3554 bfq_reset_rate_computation(bfqd, rq);
3555 goto update_last_values; /* will add one sample */
3559 * Device idle for very long: the observation interval lasting
3560 * up to this dispatch cannot be a valid observation interval
3561 * for computing a new peak rate (similarly to the late-
3562 * completion event in bfq_completed_request()). Go to
3563 * update_rate_and_reset to have the following three steps
3565 * - close the observation interval at the last (previous)
3566 * request dispatch or completion
3567 * - compute rate, if possible, for that observation interval
3568 * - start a new observation interval with this dispatch
3570 if (now_ns - bfqd->last_dispatch > 100*NSEC_PER_MSEC &&
3571 bfqd->rq_in_driver == 0)
3572 goto update_rate_and_reset;
3574 /* Update sampling information */
3575 bfqd->peak_rate_samples++;
3577 if ((bfqd->rq_in_driver > 0 ||
3578 now_ns - bfqd->last_completion < BFQ_MIN_TT)
3579 && !BFQ_RQ_SEEKY(bfqd, bfqd->last_position, rq))
3580 bfqd->sequential_samples++;
3582 bfqd->tot_sectors_dispatched += blk_rq_sectors(rq);
3584 /* Reset max observed rq size every 32 dispatches */
3585 if (likely(bfqd->peak_rate_samples % 32))
3586 bfqd->last_rq_max_size =
3587 max_t(u32, blk_rq_sectors(rq), bfqd->last_rq_max_size);
3589 bfqd->last_rq_max_size = blk_rq_sectors(rq);
3591 bfqd->delta_from_first = now_ns - bfqd->first_dispatch;
3593 /* Target observation interval not yet reached, go on sampling */
3594 if (bfqd->delta_from_first < BFQ_RATE_REF_INTERVAL)
3595 goto update_last_values;
3597 update_rate_and_reset:
3598 bfq_update_rate_reset(bfqd, rq);
3600 bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
3601 if (RQ_BFQQ(rq) == bfqd->in_service_queue)
3602 bfqd->in_serv_last_pos = bfqd->last_position;
3603 bfqd->last_dispatch = now_ns;
3607 * Remove request from internal lists.
3609 static void bfq_dispatch_remove(struct request_queue *q, struct request *rq)
3611 struct bfq_queue *bfqq = RQ_BFQQ(rq);
3614 * For consistency, the next instruction should have been
3615 * executed after removing the request from the queue and
3616 * dispatching it. We execute instead this instruction before
3617 * bfq_remove_request() (and hence introduce a temporary
3618 * inconsistency), for efficiency. In fact, should this
3619 * dispatch occur for a non in-service bfqq, this anticipated
3620 * increment prevents two counters related to bfqq->dispatched
3621 * from risking to be, first, uselessly decremented, and then
3622 * incremented again when the (new) value of bfqq->dispatched
3623 * happens to be taken into account.
3626 bfq_update_peak_rate(q->elevator->elevator_data, rq);
3628 bfq_remove_request(q, rq);
3632 * There is a case where idling does not have to be performed for
3633 * throughput concerns, but to preserve the throughput share of
3634 * the process associated with bfqq.
3636 * To introduce this case, we can note that allowing the drive
3637 * to enqueue more than one request at a time, and hence
3638 * delegating de facto final scheduling decisions to the
3639 * drive's internal scheduler, entails loss of control on the
3640 * actual request service order. In particular, the critical
3641 * situation is when requests from different processes happen
3642 * to be present, at the same time, in the internal queue(s)
3643 * of the drive. In such a situation, the drive, by deciding
3644 * the service order of the internally-queued requests, does
3645 * determine also the actual throughput distribution among
3646 * these processes. But the drive typically has no notion or
3647 * concern about per-process throughput distribution, and
3648 * makes its decisions only on a per-request basis. Therefore,
3649 * the service distribution enforced by the drive's internal
3650 * scheduler is likely to coincide with the desired throughput
3651 * distribution only in a completely symmetric, or favorably
3652 * skewed scenario where:
3653 * (i-a) each of these processes must get the same throughput as
3655 * (i-b) in case (i-a) does not hold, it holds that the process
3656 * associated with bfqq must receive a lower or equal
3657 * throughput than any of the other processes;
3658 * (ii) the I/O of each process has the same properties, in
3659 * terms of locality (sequential or random), direction
3660 * (reads or writes), request sizes, greediness
3661 * (from I/O-bound to sporadic), and so on;
3663 * In fact, in such a scenario, the drive tends to treat the requests
3664 * of each process in about the same way as the requests of the
3665 * others, and thus to provide each of these processes with about the
3666 * same throughput. This is exactly the desired throughput
3667 * distribution if (i-a) holds, or, if (i-b) holds instead, this is an
3668 * even more convenient distribution for (the process associated with)
3671 * In contrast, in any asymmetric or unfavorable scenario, device
3672 * idling (I/O-dispatch plugging) is certainly needed to guarantee
3673 * that bfqq receives its assigned fraction of the device throughput
3674 * (see [1] for details).
3676 * The problem is that idling may significantly reduce throughput with
3677 * certain combinations of types of I/O and devices. An important
3678 * example is sync random I/O on flash storage with command
3679 * queueing. So, unless bfqq falls in cases where idling also boosts
3680 * throughput, it is important to check conditions (i-a), i(-b) and
3681 * (ii) accurately, so as to avoid idling when not strictly needed for
3682 * service guarantees.
3684 * Unfortunately, it is extremely difficult to thoroughly check
3685 * condition (ii). And, in case there are active groups, it becomes
3686 * very difficult to check conditions (i-a) and (i-b) too. In fact,
3687 * if there are active groups, then, for conditions (i-a) or (i-b) to
3688 * become false 'indirectly', it is enough that an active group
3689 * contains more active processes or sub-groups than some other active
3690 * group. More precisely, for conditions (i-a) or (i-b) to become
3691 * false because of such a group, it is not even necessary that the
3692 * group is (still) active: it is sufficient that, even if the group
3693 * has become inactive, some of its descendant processes still have
3694 * some request already dispatched but still waiting for
3695 * completion. In fact, requests have still to be guaranteed their
3696 * share of the throughput even after being dispatched. In this
3697 * respect, it is easy to show that, if a group frequently becomes
3698 * inactive while still having in-flight requests, and if, when this
3699 * happens, the group is not considered in the calculation of whether
3700 * the scenario is asymmetric, then the group may fail to be
3701 * guaranteed its fair share of the throughput (basically because
3702 * idling may not be performed for the descendant processes of the
3703 * group, but it had to be). We address this issue with the following
3704 * bi-modal behavior, implemented in the function
3705 * bfq_asymmetric_scenario().
3707 * If there are groups with requests waiting for completion
3708 * (as commented above, some of these groups may even be
3709 * already inactive), then the scenario is tagged as
3710 * asymmetric, conservatively, without checking any of the
3711 * conditions (i-a), (i-b) or (ii). So the device is idled for bfqq.
3712 * This behavior matches also the fact that groups are created
3713 * exactly if controlling I/O is a primary concern (to
3714 * preserve bandwidth and latency guarantees).
3716 * On the opposite end, if there are no groups with requests waiting
3717 * for completion, then only conditions (i-a) and (i-b) are actually
3718 * controlled, i.e., provided that conditions (i-a) or (i-b) holds,
3719 * idling is not performed, regardless of whether condition (ii)
3720 * holds. In other words, only if conditions (i-a) and (i-b) do not
3721 * hold, then idling is allowed, and the device tends to be prevented
3722 * from queueing many requests, possibly of several processes. Since
3723 * there are no groups with requests waiting for completion, then, to
3724 * control conditions (i-a) and (i-b) it is enough to check just
3725 * whether all the queues with requests waiting for completion also
3726 * have the same weight.
3728 * Not checking condition (ii) evidently exposes bfqq to the
3729 * risk of getting less throughput than its fair share.
3730 * However, for queues with the same weight, a further
3731 * mechanism, preemption, mitigates or even eliminates this
3732 * problem. And it does so without consequences on overall
3733 * throughput. This mechanism and its benefits are explained
3734 * in the next three paragraphs.
3736 * Even if a queue, say Q, is expired when it remains idle, Q
3737 * can still preempt the new in-service queue if the next
3738 * request of Q arrives soon (see the comments on
3739 * bfq_bfqq_update_budg_for_activation). If all queues and
3740 * groups have the same weight, this form of preemption,
3741 * combined with the hole-recovery heuristic described in the
3742 * comments on function bfq_bfqq_update_budg_for_activation,
3743 * are enough to preserve a correct bandwidth distribution in
3744 * the mid term, even without idling. In fact, even if not
3745 * idling allows the internal queues of the device to contain
3746 * many requests, and thus to reorder requests, we can rather
3747 * safely assume that the internal scheduler still preserves a
3748 * minimum of mid-term fairness.
3750 * More precisely, this preemption-based, idleless approach
3751 * provides fairness in terms of IOPS, and not sectors per
3752 * second. This can be seen with a simple example. Suppose
3753 * that there are two queues with the same weight, but that
3754 * the first queue receives requests of 8 sectors, while the
3755 * second queue receives requests of 1024 sectors. In
3756 * addition, suppose that each of the two queues contains at
3757 * most one request at a time, which implies that each queue
3758 * always remains idle after it is served. Finally, after
3759 * remaining idle, each queue receives very quickly a new
3760 * request. It follows that the two queues are served
3761 * alternatively, preempting each other if needed. This
3762 * implies that, although both queues have the same weight,
3763 * the queue with large requests receives a service that is
3764 * 1024/8 times as high as the service received by the other
3767 * The motivation for using preemption instead of idling (for
3768 * queues with the same weight) is that, by not idling,
3769 * service guarantees are preserved (completely or at least in
3770 * part) without minimally sacrificing throughput. And, if
3771 * there is no active group, then the primary expectation for
3772 * this device is probably a high throughput.
3774 * We are now left only with explaining the two sub-conditions in the
3775 * additional compound condition that is checked below for deciding
3776 * whether the scenario is asymmetric. To explain the first
3777 * sub-condition, we need to add that the function
3778 * bfq_asymmetric_scenario checks the weights of only
3779 * non-weight-raised queues, for efficiency reasons (see comments on
3780 * bfq_weights_tree_add()). Then the fact that bfqq is weight-raised
3781 * is checked explicitly here. More precisely, the compound condition
3782 * below takes into account also the fact that, even if bfqq is being
3783 * weight-raised, the scenario is still symmetric if all queues with
3784 * requests waiting for completion happen to be
3785 * weight-raised. Actually, we should be even more precise here, and
3786 * differentiate between interactive weight raising and soft real-time
3789 * The second sub-condition checked in the compound condition is
3790 * whether there is a fair amount of already in-flight I/O not
3791 * belonging to bfqq. If so, I/O dispatching is to be plugged, for the
3792 * following reason. The drive may decide to serve in-flight
3793 * non-bfqq's I/O requests before bfqq's ones, thereby delaying the
3794 * arrival of new I/O requests for bfqq (recall that bfqq is sync). If
3795 * I/O-dispatching is not plugged, then, while bfqq remains empty, a
3796 * basically uncontrolled amount of I/O from other queues may be
3797 * dispatched too, possibly causing the service of bfqq's I/O to be
3798 * delayed even longer in the drive. This problem gets more and more
3799 * serious as the speed and the queue depth of the drive grow,
3800 * because, as these two quantities grow, the probability to find no
3801 * queue busy but many requests in flight grows too. By contrast,
3802 * plugging I/O dispatching minimizes the delay induced by already
3803 * in-flight I/O, and enables bfqq to recover the bandwidth it may
3804 * lose because of this delay.
3806 * As a side note, it is worth considering that the above
3807 * device-idling countermeasures may however fail in the following
3808 * unlucky scenario: if I/O-dispatch plugging is (correctly) disabled
3809 * in a time period during which all symmetry sub-conditions hold, and
3810 * therefore the device is allowed to enqueue many requests, but at
3811 * some later point in time some sub-condition stops to hold, then it
3812 * may become impossible to make requests be served in the desired
3813 * order until all the requests already queued in the device have been
3814 * served. The last sub-condition commented above somewhat mitigates
3815 * this problem for weight-raised queues.
3817 * However, as an additional mitigation for this problem, we preserve
3818 * plugging for a special symmetric case that may suddenly turn into
3819 * asymmetric: the case where only bfqq is busy. In this case, not
3820 * expiring bfqq does not cause any harm to any other queues in terms
3821 * of service guarantees. In contrast, it avoids the following unlucky
3822 * sequence of events: (1) bfqq is expired, (2) a new queue with a
3823 * lower weight than bfqq becomes busy (or more queues), (3) the new
3824 * queue is served until a new request arrives for bfqq, (4) when bfqq
3825 * is finally served, there are so many requests of the new queue in
3826 * the drive that the pending requests for bfqq take a lot of time to
3827 * be served. In particular, event (2) may case even already
3828 * dispatched requests of bfqq to be delayed, inside the drive. So, to
3829 * avoid this series of events, the scenario is preventively declared
3830 * as asymmetric also if bfqq is the only busy queues
3832 static bool idling_needed_for_service_guarantees(struct bfq_data *bfqd,
3833 struct bfq_queue *bfqq)
3835 int tot_busy_queues = bfq_tot_busy_queues(bfqd);
3837 /* No point in idling for bfqq if it won't get requests any longer */
3838 if (unlikely(!bfqq_process_refs(bfqq)))
3841 return (bfqq->wr_coeff > 1 &&
3842 (bfqd->wr_busy_queues <
3844 bfqd->rq_in_driver >=
3845 bfqq->dispatched + 4)) ||
3846 bfq_asymmetric_scenario(bfqd, bfqq) ||
3847 tot_busy_queues == 1;
3850 static bool __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq,
3851 enum bfqq_expiration reason)
3854 * If this bfqq is shared between multiple processes, check
3855 * to make sure that those processes are still issuing I/Os
3856 * within the mean seek distance. If not, it may be time to
3857 * break the queues apart again.
3859 if (bfq_bfqq_coop(bfqq) && BFQQ_SEEKY(bfqq))
3860 bfq_mark_bfqq_split_coop(bfqq);
3863 * Consider queues with a higher finish virtual time than
3864 * bfqq. If idling_needed_for_service_guarantees(bfqq) returns
3865 * true, then bfqq's bandwidth would be violated if an
3866 * uncontrolled amount of I/O from these queues were
3867 * dispatched while bfqq is waiting for its new I/O to
3868 * arrive. This is exactly what may happen if this is a forced
3869 * expiration caused by a preemption attempt, and if bfqq is
3870 * not re-scheduled. To prevent this from happening, re-queue
3871 * bfqq if it needs I/O-dispatch plugging, even if it is
3872 * empty. By doing so, bfqq is granted to be served before the
3873 * above queues (provided that bfqq is of course eligible).
3875 if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
3876 !(reason == BFQQE_PREEMPTED &&
3877 idling_needed_for_service_guarantees(bfqd, bfqq))) {
3878 if (bfqq->dispatched == 0)
3880 * Overloading budget_timeout field to store
3881 * the time at which the queue remains with no
3882 * backlog and no outstanding request; used by
3883 * the weight-raising mechanism.
3885 bfqq->budget_timeout = jiffies;
3887 bfq_del_bfqq_busy(bfqd, bfqq, true);
3889 bfq_requeue_bfqq(bfqd, bfqq, true);
3891 * Resort priority tree of potential close cooperators.
3892 * See comments on bfq_pos_tree_add_move() for the unlikely().
3894 if (unlikely(!bfqd->nonrot_with_queueing &&
3895 !RB_EMPTY_ROOT(&bfqq->sort_list)))
3896 bfq_pos_tree_add_move(bfqd, bfqq);
3900 * All in-service entities must have been properly deactivated
3901 * or requeued before executing the next function, which
3902 * resets all in-service entities as no more in service. This
3903 * may cause bfqq to be freed. If this happens, the next
3904 * function returns true.
3906 return __bfq_bfqd_reset_in_service(bfqd);
3910 * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior.
3911 * @bfqd: device data.
3912 * @bfqq: queue to update.
3913 * @reason: reason for expiration.
3915 * Handle the feedback on @bfqq budget at queue expiration.
3916 * See the body for detailed comments.
3918 static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd,
3919 struct bfq_queue *bfqq,
3920 enum bfqq_expiration reason)
3922 struct request *next_rq;
3923 int budget, min_budget;
3925 min_budget = bfq_min_budget(bfqd);
3927 if (bfqq->wr_coeff == 1)
3928 budget = bfqq->max_budget;
3930 * Use a constant, low budget for weight-raised queues,
3931 * to help achieve a low latency. Keep it slightly higher
3932 * than the minimum possible budget, to cause a little
3933 * bit fewer expirations.
3935 budget = 2 * min_budget;
3937 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %d, budg left %d",
3938 bfqq->entity.budget, bfq_bfqq_budget_left(bfqq));
3939 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %d, min budg %d",
3940 budget, bfq_min_budget(bfqd));
3941 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d",
3942 bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->in_service_queue));
3944 if (bfq_bfqq_sync(bfqq) && bfqq->wr_coeff == 1) {
3947 * Caveat: in all the following cases we trade latency
3950 case BFQQE_TOO_IDLE:
3952 * This is the only case where we may reduce
3953 * the budget: if there is no request of the
3954 * process still waiting for completion, then
3955 * we assume (tentatively) that the timer has
3956 * expired because the batch of requests of
3957 * the process could have been served with a
3958 * smaller budget. Hence, betting that
3959 * process will behave in the same way when it
3960 * becomes backlogged again, we reduce its
3961 * next budget. As long as we guess right,
3962 * this budget cut reduces the latency
3963 * experienced by the process.
3965 * However, if there are still outstanding
3966 * requests, then the process may have not yet
3967 * issued its next request just because it is
3968 * still waiting for the completion of some of
3969 * the still outstanding ones. So in this
3970 * subcase we do not reduce its budget, on the
3971 * contrary we increase it to possibly boost
3972 * the throughput, as discussed in the
3973 * comments to the BUDGET_TIMEOUT case.
3975 if (bfqq->dispatched > 0) /* still outstanding reqs */
3976 budget = min(budget * 2, bfqd->bfq_max_budget);
3978 if (budget > 5 * min_budget)
3979 budget -= 4 * min_budget;
3981 budget = min_budget;
3984 case BFQQE_BUDGET_TIMEOUT:
3986 * We double the budget here because it gives
3987 * the chance to boost the throughput if this
3988 * is not a seeky process (and has bumped into
3989 * this timeout because of, e.g., ZBR).
3991 budget = min(budget * 2, bfqd->bfq_max_budget);
3993 case BFQQE_BUDGET_EXHAUSTED:
3995 * The process still has backlog, and did not
3996 * let either the budget timeout or the disk
3997 * idling timeout expire. Hence it is not
3998 * seeky, has a short thinktime and may be
3999 * happy with a higher budget too. So
4000 * definitely increase the budget of this good
4001 * candidate to boost the disk throughput.
4003 budget = min(budget * 4, bfqd->bfq_max_budget);
4005 case BFQQE_NO_MORE_REQUESTS:
4007 * For queues that expire for this reason, it
4008 * is particularly important to keep the
4009 * budget close to the actual service they
4010 * need. Doing so reduces the timestamp
4011 * misalignment problem described in the
4012 * comments in the body of
4013 * __bfq_activate_entity. In fact, suppose
4014 * that a queue systematically expires for
4015 * BFQQE_NO_MORE_REQUESTS and presents a
4016 * new request in time to enjoy timestamp
4017 * back-shifting. The larger the budget of the
4018 * queue is with respect to the service the
4019 * queue actually requests in each service
4020 * slot, the more times the queue can be
4021 * reactivated with the same virtual finish
4022 * time. It follows that, even if this finish
4023 * time is pushed to the system virtual time
4024 * to reduce the consequent timestamp
4025 * misalignment, the queue unjustly enjoys for
4026 * many re-activations a lower finish time
4027 * than all newly activated queues.
4029 * The service needed by bfqq is measured
4030 * quite precisely by bfqq->entity.service.
4031 * Since bfqq does not enjoy device idling,
4032 * bfqq->entity.service is equal to the number
4033 * of sectors that the process associated with
4034 * bfqq requested to read/write before waiting
4035 * for request completions, or blocking for
4038 budget = max_t(int, bfqq->entity.service, min_budget);
4043 } else if (!bfq_bfqq_sync(bfqq)) {
4045 * Async queues get always the maximum possible
4046 * budget, as for them we do not care about latency
4047 * (in addition, their ability to dispatch is limited
4048 * by the charging factor).
4050 budget = bfqd->bfq_max_budget;
4053 bfqq->max_budget = budget;
4055 if (bfqd->budgets_assigned >= bfq_stats_min_budgets &&
4056 !bfqd->bfq_user_max_budget)
4057 bfqq->max_budget = min(bfqq->max_budget, bfqd->bfq_max_budget);
4060 * If there is still backlog, then assign a new budget, making
4061 * sure that it is large enough for the next request. Since
4062 * the finish time of bfqq must be kept in sync with the
4063 * budget, be sure to call __bfq_bfqq_expire() *after* this
4066 * If there is no backlog, then no need to update the budget;
4067 * it will be updated on the arrival of a new request.
4069 next_rq = bfqq->next_rq;
4071 bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget,
4072 bfq_serv_to_charge(next_rq, bfqq));
4074 bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %d",
4075 next_rq ? blk_rq_sectors(next_rq) : 0,
4076 bfqq->entity.budget);
4080 * Return true if the process associated with bfqq is "slow". The slow
4081 * flag is used, in addition to the budget timeout, to reduce the
4082 * amount of service provided to seeky processes, and thus reduce
4083 * their chances to lower the throughput. More details in the comments
4084 * on the function bfq_bfqq_expire().
4086 * An important observation is in order: as discussed in the comments
4087 * on the function bfq_update_peak_rate(), with devices with internal
4088 * queues, it is hard if ever possible to know when and for how long
4089 * an I/O request is processed by the device (apart from the trivial
4090 * I/O pattern where a new request is dispatched only after the
4091 * previous one has been completed). This makes it hard to evaluate
4092 * the real rate at which the I/O requests of each bfq_queue are
4093 * served. In fact, for an I/O scheduler like BFQ, serving a
4094 * bfq_queue means just dispatching its requests during its service
4095 * slot (i.e., until the budget of the queue is exhausted, or the
4096 * queue remains idle, or, finally, a timeout fires). But, during the
4097 * service slot of a bfq_queue, around 100 ms at most, the device may
4098 * be even still processing requests of bfq_queues served in previous
4099 * service slots. On the opposite end, the requests of the in-service
4100 * bfq_queue may be completed after the service slot of the queue
4103 * Anyway, unless more sophisticated solutions are used
4104 * (where possible), the sum of the sizes of the requests dispatched
4105 * during the service slot of a bfq_queue is probably the only
4106 * approximation available for the service received by the bfq_queue
4107 * during its service slot. And this sum is the quantity used in this
4108 * function to evaluate the I/O speed of a process.
4110 static bool bfq_bfqq_is_slow(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4111 bool compensate, enum bfqq_expiration reason,
4112 unsigned long *delta_ms)
4114 ktime_t delta_ktime;
4116 bool slow = BFQQ_SEEKY(bfqq); /* if delta too short, use seekyness */
4118 if (!bfq_bfqq_sync(bfqq))
4122 delta_ktime = bfqd->last_idling_start;
4124 delta_ktime = ktime_get();
4125 delta_ktime = ktime_sub(delta_ktime, bfqd->last_budget_start);
4126 delta_usecs = ktime_to_us(delta_ktime);
4128 /* don't use too short time intervals */
4129 if (delta_usecs < 1000) {
4130 if (blk_queue_nonrot(bfqd->queue))
4132 * give same worst-case guarantees as idling
4135 *delta_ms = BFQ_MIN_TT / NSEC_PER_MSEC;
4136 else /* charge at least one seek */
4137 *delta_ms = bfq_slice_idle / NSEC_PER_MSEC;
4142 *delta_ms = delta_usecs / USEC_PER_MSEC;
4145 * Use only long (> 20ms) intervals to filter out excessive
4146 * spikes in service rate estimation.
4148 if (delta_usecs > 20000) {
4150 * Caveat for rotational devices: processes doing I/O
4151 * in the slower disk zones tend to be slow(er) even
4152 * if not seeky. In this respect, the estimated peak
4153 * rate is likely to be an average over the disk
4154 * surface. Accordingly, to not be too harsh with
4155 * unlucky processes, a process is deemed slow only if
4156 * its rate has been lower than half of the estimated
4159 slow = bfqq->entity.service < bfqd->bfq_max_budget / 2;
4162 bfq_log_bfqq(bfqd, bfqq, "bfq_bfqq_is_slow: slow %d", slow);
4168 * To be deemed as soft real-time, an application must meet two
4169 * requirements. First, the application must not require an average
4170 * bandwidth higher than the approximate bandwidth required to playback or
4171 * record a compressed high-definition video.
4172 * The next function is invoked on the completion of the last request of a
4173 * batch, to compute the next-start time instant, soft_rt_next_start, such
4174 * that, if the next request of the application does not arrive before
4175 * soft_rt_next_start, then the above requirement on the bandwidth is met.
4177 * The second requirement is that the request pattern of the application is
4178 * isochronous, i.e., that, after issuing a request or a batch of requests,
4179 * the application stops issuing new requests until all its pending requests
4180 * have been completed. After that, the application may issue a new batch,
4182 * For this reason the next function is invoked to compute
4183 * soft_rt_next_start only for applications that meet this requirement,
4184 * whereas soft_rt_next_start is set to infinity for applications that do
4187 * Unfortunately, even a greedy (i.e., I/O-bound) application may
4188 * happen to meet, occasionally or systematically, both the above
4189 * bandwidth and isochrony requirements. This may happen at least in
4190 * the following circumstances. First, if the CPU load is high. The
4191 * application may stop issuing requests while the CPUs are busy
4192 * serving other processes, then restart, then stop again for a while,
4193 * and so on. The other circumstances are related to the storage
4194 * device: the storage device is highly loaded or reaches a low-enough
4195 * throughput with the I/O of the application (e.g., because the I/O
4196 * is random and/or the device is slow). In all these cases, the
4197 * I/O of the application may be simply slowed down enough to meet
4198 * the bandwidth and isochrony requirements. To reduce the probability
4199 * that greedy applications are deemed as soft real-time in these
4200 * corner cases, a further rule is used in the computation of
4201 * soft_rt_next_start: the return value of this function is forced to
4202 * be higher than the maximum between the following two quantities.
4204 * (a) Current time plus: (1) the maximum time for which the arrival
4205 * of a request is waited for when a sync queue becomes idle,
4206 * namely bfqd->bfq_slice_idle, and (2) a few extra jiffies. We
4207 * postpone for a moment the reason for adding a few extra
4208 * jiffies; we get back to it after next item (b). Lower-bounding
4209 * the return value of this function with the current time plus
4210 * bfqd->bfq_slice_idle tends to filter out greedy applications,
4211 * because the latter issue their next request as soon as possible
4212 * after the last one has been completed. In contrast, a soft
4213 * real-time application spends some time processing data, after a
4214 * batch of its requests has been completed.
4216 * (b) Current value of bfqq->soft_rt_next_start. As pointed out
4217 * above, greedy applications may happen to meet both the
4218 * bandwidth and isochrony requirements under heavy CPU or
4219 * storage-device load. In more detail, in these scenarios, these
4220 * applications happen, only for limited time periods, to do I/O
4221 * slowly enough to meet all the requirements described so far,
4222 * including the filtering in above item (a). These slow-speed
4223 * time intervals are usually interspersed between other time
4224 * intervals during which these applications do I/O at a very high
4225 * speed. Fortunately, exactly because of the high speed of the
4226 * I/O in the high-speed intervals, the values returned by this
4227 * function happen to be so high, near the end of any such
4228 * high-speed interval, to be likely to fall *after* the end of
4229 * the low-speed time interval that follows. These high values are
4230 * stored in bfqq->soft_rt_next_start after each invocation of
4231 * this function. As a consequence, if the last value of
4232 * bfqq->soft_rt_next_start is constantly used to lower-bound the
4233 * next value that this function may return, then, from the very
4234 * beginning of a low-speed interval, bfqq->soft_rt_next_start is
4235 * likely to be constantly kept so high that any I/O request
4236 * issued during the low-speed interval is considered as arriving
4237 * to soon for the application to be deemed as soft
4238 * real-time. Then, in the high-speed interval that follows, the
4239 * application will not be deemed as soft real-time, just because
4240 * it will do I/O at a high speed. And so on.
4242 * Getting back to the filtering in item (a), in the following two
4243 * cases this filtering might be easily passed by a greedy
4244 * application, if the reference quantity was just
4245 * bfqd->bfq_slice_idle:
4246 * 1) HZ is so low that the duration of a jiffy is comparable to or
4247 * higher than bfqd->bfq_slice_idle. This happens, e.g., on slow
4248 * devices with HZ=100. The time granularity may be so coarse
4249 * that the approximation, in jiffies, of bfqd->bfq_slice_idle
4250 * is rather lower than the exact value.
4251 * 2) jiffies, instead of increasing at a constant rate, may stop increasing
4252 * for a while, then suddenly 'jump' by several units to recover the lost
4253 * increments. This seems to happen, e.g., inside virtual machines.
4254 * To address this issue, in the filtering in (a) we do not use as a
4255 * reference time interval just bfqd->bfq_slice_idle, but
4256 * bfqd->bfq_slice_idle plus a few jiffies. In particular, we add the
4257 * minimum number of jiffies for which the filter seems to be quite
4258 * precise also in embedded systems and KVM/QEMU virtual machines.
4260 static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data *bfqd,
4261 struct bfq_queue *bfqq)
4263 return max3(bfqq->soft_rt_next_start,
4264 bfqq->last_idle_bklogged +
4265 HZ * bfqq->service_from_backlogged /
4266 bfqd->bfq_wr_max_softrt_rate,
4267 jiffies + nsecs_to_jiffies(bfqq->bfqd->bfq_slice_idle) + 4);
4271 * bfq_bfqq_expire - expire a queue.
4272 * @bfqd: device owning the queue.
4273 * @bfqq: the queue to expire.
4274 * @compensate: if true, compensate for the time spent idling.
4275 * @reason: the reason causing the expiration.
4277 * If the process associated with bfqq does slow I/O (e.g., because it
4278 * issues random requests), we charge bfqq with the time it has been
4279 * in service instead of the service it has received (see
4280 * bfq_bfqq_charge_time for details on how this goal is achieved). As
4281 * a consequence, bfqq will typically get higher timestamps upon
4282 * reactivation, and hence it will be rescheduled as if it had
4283 * received more service than what it has actually received. In the
4284 * end, bfqq receives less service in proportion to how slowly its
4285 * associated process consumes its budgets (and hence how seriously it
4286 * tends to lower the throughput). In addition, this time-charging
4287 * strategy guarantees time fairness among slow processes. In
4288 * contrast, if the process associated with bfqq is not slow, we
4289 * charge bfqq exactly with the service it has received.
4291 * Charging time to the first type of queues and the exact service to
4292 * the other has the effect of using the WF2Q+ policy to schedule the
4293 * former on a timeslice basis, without violating service domain
4294 * guarantees among the latter.
4296 void bfq_bfqq_expire(struct bfq_data *bfqd,
4297 struct bfq_queue *bfqq,
4299 enum bfqq_expiration reason)
4302 unsigned long delta = 0;
4303 struct bfq_entity *entity = &bfqq->entity;
4306 * Check whether the process is slow (see bfq_bfqq_is_slow).
4308 slow = bfq_bfqq_is_slow(bfqd, bfqq, compensate, reason, &delta);
4311 * As above explained, charge slow (typically seeky) and
4312 * timed-out queues with the time and not the service
4313 * received, to favor sequential workloads.
4315 * Processes doing I/O in the slower disk zones will tend to
4316 * be slow(er) even if not seeky. Therefore, since the
4317 * estimated peak rate is actually an average over the disk
4318 * surface, these processes may timeout just for bad luck. To
4319 * avoid punishing them, do not charge time to processes that
4320 * succeeded in consuming at least 2/3 of their budget. This
4321 * allows BFQ to preserve enough elasticity to still perform
4322 * bandwidth, and not time, distribution with little unlucky
4323 * or quasi-sequential processes.
4325 if (bfqq->wr_coeff == 1 &&
4327 (reason == BFQQE_BUDGET_TIMEOUT &&
4328 bfq_bfqq_budget_left(bfqq) >= entity->budget / 3)))
4329 bfq_bfqq_charge_time(bfqd, bfqq, delta);
4331 if (bfqd->low_latency && bfqq->wr_coeff == 1)
4332 bfqq->last_wr_start_finish = jiffies;
4334 if (bfqd->low_latency && bfqd->bfq_wr_max_softrt_rate > 0 &&
4335 RB_EMPTY_ROOT(&bfqq->sort_list)) {
4337 * If we get here, and there are no outstanding
4338 * requests, then the request pattern is isochronous
4339 * (see the comments on the function
4340 * bfq_bfqq_softrt_next_start()). Therefore we can
4341 * compute soft_rt_next_start.
4343 * If, instead, the queue still has outstanding
4344 * requests, then we have to wait for the completion
4345 * of all the outstanding requests to discover whether
4346 * the request pattern is actually isochronous.
4348 if (bfqq->dispatched == 0)
4349 bfqq->soft_rt_next_start =
4350 bfq_bfqq_softrt_next_start(bfqd, bfqq);
4351 else if (bfqq->dispatched > 0) {
4353 * Schedule an update of soft_rt_next_start to when
4354 * the task may be discovered to be isochronous.
4356 bfq_mark_bfqq_softrt_update(bfqq);
4360 bfq_log_bfqq(bfqd, bfqq,
4361 "expire (%d, slow %d, num_disp %d, short_ttime %d)", reason,
4362 slow, bfqq->dispatched, bfq_bfqq_has_short_ttime(bfqq));
4365 * bfqq expired, so no total service time needs to be computed
4366 * any longer: reset state machine for measuring total service
4369 bfqd->rqs_injected = bfqd->wait_dispatch = false;
4370 bfqd->waited_rq = NULL;
4373 * Increase, decrease or leave budget unchanged according to
4376 __bfq_bfqq_recalc_budget(bfqd, bfqq, reason);
4377 if (__bfq_bfqq_expire(bfqd, bfqq, reason))
4378 /* bfqq is gone, no more actions on it */
4381 /* mark bfqq as waiting a request only if a bic still points to it */
4382 if (!bfq_bfqq_busy(bfqq) &&
4383 reason != BFQQE_BUDGET_TIMEOUT &&
4384 reason != BFQQE_BUDGET_EXHAUSTED) {
4385 bfq_mark_bfqq_non_blocking_wait_rq(bfqq);
4387 * Not setting service to 0, because, if the next rq
4388 * arrives in time, the queue will go on receiving
4389 * service with this same budget (as if it never expired)
4392 entity->service = 0;
4395 * Reset the received-service counter for every parent entity.
4396 * Differently from what happens with bfqq->entity.service,
4397 * the resetting of this counter never needs to be postponed
4398 * for parent entities. In fact, in case bfqq may have a
4399 * chance to go on being served using the last, partially
4400 * consumed budget, bfqq->entity.service needs to be kept,
4401 * because if bfqq then actually goes on being served using
4402 * the same budget, the last value of bfqq->entity.service is
4403 * needed to properly decrement bfqq->entity.budget by the
4404 * portion already consumed. In contrast, it is not necessary
4405 * to keep entity->service for parent entities too, because
4406 * the bubble up of the new value of bfqq->entity.budget will
4407 * make sure that the budgets of parent entities are correct,
4408 * even in case bfqq and thus parent entities go on receiving
4409 * service with the same budget.
4411 entity = entity->parent;
4412 for_each_entity(entity)
4413 entity->service = 0;
4417 * Budget timeout is not implemented through a dedicated timer, but
4418 * just checked on request arrivals and completions, as well as on
4419 * idle timer expirations.
4421 static bool bfq_bfqq_budget_timeout(struct bfq_queue *bfqq)
4423 return time_is_before_eq_jiffies(bfqq->budget_timeout);
4427 * If we expire a queue that is actively waiting (i.e., with the
4428 * device idled) for the arrival of a new request, then we may incur
4429 * the timestamp misalignment problem described in the body of the
4430 * function __bfq_activate_entity. Hence we return true only if this
4431 * condition does not hold, or if the queue is slow enough to deserve
4432 * only to be kicked off for preserving a high throughput.
4434 static bool bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq)
4436 bfq_log_bfqq(bfqq->bfqd, bfqq,
4437 "may_budget_timeout: wait_request %d left %d timeout %d",
4438 bfq_bfqq_wait_request(bfqq),
4439 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3,
4440 bfq_bfqq_budget_timeout(bfqq));
4442 return (!bfq_bfqq_wait_request(bfqq) ||
4443 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3)
4445 bfq_bfqq_budget_timeout(bfqq);
4448 static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd,
4449 struct bfq_queue *bfqq)
4451 bool rot_without_queueing =
4452 !blk_queue_nonrot(bfqd->queue) && !bfqd->hw_tag,
4453 bfqq_sequential_and_IO_bound,
4456 /* No point in idling for bfqq if it won't get requests any longer */
4457 if (unlikely(!bfqq_process_refs(bfqq)))
4460 bfqq_sequential_and_IO_bound = !BFQQ_SEEKY(bfqq) &&
4461 bfq_bfqq_IO_bound(bfqq) && bfq_bfqq_has_short_ttime(bfqq);
4464 * The next variable takes into account the cases where idling
4465 * boosts the throughput.
4467 * The value of the variable is computed considering, first, that
4468 * idling is virtually always beneficial for the throughput if:
4469 * (a) the device is not NCQ-capable and rotational, or
4470 * (b) regardless of the presence of NCQ, the device is rotational and
4471 * the request pattern for bfqq is I/O-bound and sequential, or
4472 * (c) regardless of whether it is rotational, the device is
4473 * not NCQ-capable and the request pattern for bfqq is
4474 * I/O-bound and sequential.
4476 * Secondly, and in contrast to the above item (b), idling an
4477 * NCQ-capable flash-based device would not boost the
4478 * throughput even with sequential I/O; rather it would lower
4479 * the throughput in proportion to how fast the device
4480 * is. Accordingly, the next variable is true if any of the
4481 * above conditions (a), (b) or (c) is true, and, in
4482 * particular, happens to be false if bfqd is an NCQ-capable
4483 * flash-based device.
4485 idling_boosts_thr = rot_without_queueing ||
4486 ((!blk_queue_nonrot(bfqd->queue) || !bfqd->hw_tag) &&
4487 bfqq_sequential_and_IO_bound);
4490 * The return value of this function is equal to that of
4491 * idling_boosts_thr, unless a special case holds. In this
4492 * special case, described below, idling may cause problems to
4493 * weight-raised queues.
4495 * When the request pool is saturated (e.g., in the presence
4496 * of write hogs), if the processes associated with
4497 * non-weight-raised queues ask for requests at a lower rate,
4498 * then processes associated with weight-raised queues have a
4499 * higher probability to get a request from the pool
4500 * immediately (or at least soon) when they need one. Thus
4501 * they have a higher probability to actually get a fraction
4502 * of the device throughput proportional to their high
4503 * weight. This is especially true with NCQ-capable drives,
4504 * which enqueue several requests in advance, and further
4505 * reorder internally-queued requests.
4507 * For this reason, we force to false the return value if
4508 * there are weight-raised busy queues. In this case, and if
4509 * bfqq is not weight-raised, this guarantees that the device
4510 * is not idled for bfqq (if, instead, bfqq is weight-raised,
4511 * then idling will be guaranteed by another variable, see
4512 * below). Combined with the timestamping rules of BFQ (see
4513 * [1] for details), this behavior causes bfqq, and hence any
4514 * sync non-weight-raised queue, to get a lower number of
4515 * requests served, and thus to ask for a lower number of
4516 * requests from the request pool, before the busy
4517 * weight-raised queues get served again. This often mitigates
4518 * starvation problems in the presence of heavy write
4519 * workloads and NCQ, thereby guaranteeing a higher
4520 * application and system responsiveness in these hostile
4523 return idling_boosts_thr &&
4524 bfqd->wr_busy_queues == 0;
4528 * For a queue that becomes empty, device idling is allowed only if
4529 * this function returns true for that queue. As a consequence, since
4530 * device idling plays a critical role for both throughput boosting
4531 * and service guarantees, the return value of this function plays a
4532 * critical role as well.
4534 * In a nutshell, this function returns true only if idling is
4535 * beneficial for throughput or, even if detrimental for throughput,
4536 * idling is however necessary to preserve service guarantees (low
4537 * latency, desired throughput distribution, ...). In particular, on
4538 * NCQ-capable devices, this function tries to return false, so as to
4539 * help keep the drives' internal queues full, whenever this helps the
4540 * device boost the throughput without causing any service-guarantee
4543 * Most of the issues taken into account to get the return value of
4544 * this function are not trivial. We discuss these issues in the two
4545 * functions providing the main pieces of information needed by this
4548 static bool bfq_better_to_idle(struct bfq_queue *bfqq)
4550 struct bfq_data *bfqd = bfqq->bfqd;
4551 bool idling_boosts_thr_with_no_issue, idling_needed_for_service_guar;
4553 /* No point in idling for bfqq if it won't get requests any longer */
4554 if (unlikely(!bfqq_process_refs(bfqq)))
4557 if (unlikely(bfqd->strict_guarantees))
4561 * Idling is performed only if slice_idle > 0. In addition, we
4564 * (b) bfqq is in the idle io prio class: in this case we do
4565 * not idle because we want to minimize the bandwidth that
4566 * queues in this class can steal to higher-priority queues
4568 if (bfqd->bfq_slice_idle == 0 || !bfq_bfqq_sync(bfqq) ||
4569 bfq_class_idle(bfqq))
4572 idling_boosts_thr_with_no_issue =
4573 idling_boosts_thr_without_issues(bfqd, bfqq);
4575 idling_needed_for_service_guar =
4576 idling_needed_for_service_guarantees(bfqd, bfqq);
4579 * We have now the two components we need to compute the
4580 * return value of the function, which is true only if idling
4581 * either boosts the throughput (without issues), or is
4582 * necessary to preserve service guarantees.
4584 return idling_boosts_thr_with_no_issue ||
4585 idling_needed_for_service_guar;
4589 * If the in-service queue is empty but the function bfq_better_to_idle
4590 * returns true, then:
4591 * 1) the queue must remain in service and cannot be expired, and
4592 * 2) the device must be idled to wait for the possible arrival of a new
4593 * request for the queue.
4594 * See the comments on the function bfq_better_to_idle for the reasons
4595 * why performing device idling is the best choice to boost the throughput
4596 * and preserve service guarantees when bfq_better_to_idle itself
4599 static bool bfq_bfqq_must_idle(struct bfq_queue *bfqq)
4601 return RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_better_to_idle(bfqq);
4605 * This function chooses the queue from which to pick the next extra
4606 * I/O request to inject, if it finds a compatible queue. See the
4607 * comments on bfq_update_inject_limit() for details on the injection
4608 * mechanism, and for the definitions of the quantities mentioned
4611 static struct bfq_queue *
4612 bfq_choose_bfqq_for_injection(struct bfq_data *bfqd)
4614 struct bfq_queue *bfqq, *in_serv_bfqq = bfqd->in_service_queue;
4615 unsigned int limit = in_serv_bfqq->inject_limit;
4618 * - bfqq is not weight-raised and therefore does not carry
4619 * time-critical I/O,
4621 * - regardless of whether bfqq is weight-raised, bfqq has
4622 * however a long think time, during which it can absorb the
4623 * effect of an appropriate number of extra I/O requests
4624 * from other queues (see bfq_update_inject_limit for
4625 * details on the computation of this number);
4626 * then injection can be performed without restrictions.
4628 bool in_serv_always_inject = in_serv_bfqq->wr_coeff == 1 ||
4629 !bfq_bfqq_has_short_ttime(in_serv_bfqq);
4633 * - the baseline total service time could not be sampled yet,
4634 * so the inject limit happens to be still 0, and
4635 * - a lot of time has elapsed since the plugging of I/O
4636 * dispatching started, so drive speed is being wasted
4638 * then temporarily raise inject limit to one request.
4640 if (limit == 0 && in_serv_bfqq->last_serv_time_ns == 0 &&
4641 bfq_bfqq_wait_request(in_serv_bfqq) &&
4642 time_is_before_eq_jiffies(bfqd->last_idling_start_jiffies +
4643 bfqd->bfq_slice_idle)
4647 if (bfqd->rq_in_driver >= limit)
4651 * Linear search of the source queue for injection; but, with
4652 * a high probability, very few steps are needed to find a
4653 * candidate queue, i.e., a queue with enough budget left for
4654 * its next request. In fact:
4655 * - BFQ dynamically updates the budget of every queue so as
4656 * to accommodate the expected backlog of the queue;
4657 * - if a queue gets all its requests dispatched as injected
4658 * service, then the queue is removed from the active list
4659 * (and re-added only if it gets new requests, but then it
4660 * is assigned again enough budget for its new backlog).
4662 list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
4663 if (!RB_EMPTY_ROOT(&bfqq->sort_list) &&
4664 (in_serv_always_inject || bfqq->wr_coeff > 1) &&
4665 bfq_serv_to_charge(bfqq->next_rq, bfqq) <=
4666 bfq_bfqq_budget_left(bfqq)) {
4668 * Allow for only one large in-flight request
4669 * on non-rotational devices, for the
4670 * following reason. On non-rotationl drives,
4671 * large requests take much longer than
4672 * smaller requests to be served. In addition,
4673 * the drive prefers to serve large requests
4674 * w.r.t. to small ones, if it can choose. So,
4675 * having more than one large requests queued
4676 * in the drive may easily make the next first
4677 * request of the in-service queue wait for so
4678 * long to break bfqq's service guarantees. On
4679 * the bright side, large requests let the
4680 * drive reach a very high throughput, even if
4681 * there is only one in-flight large request
4684 if (blk_queue_nonrot(bfqd->queue) &&
4685 blk_rq_sectors(bfqq->next_rq) >=
4686 BFQQ_SECT_THR_NONROT)
4687 limit = min_t(unsigned int, 1, limit);
4689 limit = in_serv_bfqq->inject_limit;
4691 if (bfqd->rq_in_driver < limit) {
4692 bfqd->rqs_injected = true;
4701 * Select a queue for service. If we have a current queue in service,
4702 * check whether to continue servicing it, or retrieve and set a new one.
4704 static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd)
4706 struct bfq_queue *bfqq;
4707 struct request *next_rq;
4708 enum bfqq_expiration reason = BFQQE_BUDGET_TIMEOUT;
4710 bfqq = bfqd->in_service_queue;
4714 bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue");
4717 * Do not expire bfqq for budget timeout if bfqq may be about
4718 * to enjoy device idling. The reason why, in this case, we
4719 * prevent bfqq from expiring is the same as in the comments
4720 * on the case where bfq_bfqq_must_idle() returns true, in
4721 * bfq_completed_request().
4723 if (bfq_may_expire_for_budg_timeout(bfqq) &&
4724 !bfq_bfqq_must_idle(bfqq))
4729 * This loop is rarely executed more than once. Even when it
4730 * happens, it is much more convenient to re-execute this loop
4731 * than to return NULL and trigger a new dispatch to get a
4734 next_rq = bfqq->next_rq;
4736 * If bfqq has requests queued and it has enough budget left to
4737 * serve them, keep the queue, otherwise expire it.
4740 if (bfq_serv_to_charge(next_rq, bfqq) >
4741 bfq_bfqq_budget_left(bfqq)) {
4743 * Expire the queue for budget exhaustion,
4744 * which makes sure that the next budget is
4745 * enough to serve the next request, even if
4746 * it comes from the fifo expired path.
4748 reason = BFQQE_BUDGET_EXHAUSTED;
4752 * The idle timer may be pending because we may
4753 * not disable disk idling even when a new request
4756 if (bfq_bfqq_wait_request(bfqq)) {
4758 * If we get here: 1) at least a new request
4759 * has arrived but we have not disabled the
4760 * timer because the request was too small,
4761 * 2) then the block layer has unplugged
4762 * the device, causing the dispatch to be
4765 * Since the device is unplugged, now the
4766 * requests are probably large enough to
4767 * provide a reasonable throughput.
4768 * So we disable idling.
4770 bfq_clear_bfqq_wait_request(bfqq);
4771 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
4778 * No requests pending. However, if the in-service queue is idling
4779 * for a new request, or has requests waiting for a completion and
4780 * may idle after their completion, then keep it anyway.
4782 * Yet, inject service from other queues if it boosts
4783 * throughput and is possible.
4785 if (bfq_bfqq_wait_request(bfqq) ||
4786 (bfqq->dispatched != 0 && bfq_better_to_idle(bfqq))) {
4787 struct bfq_queue *async_bfqq =
4788 bfqq->bic && bfqq->bic->bfqq[0] &&
4789 bfq_bfqq_busy(bfqq->bic->bfqq[0]) &&
4790 bfqq->bic->bfqq[0]->next_rq ?
4791 bfqq->bic->bfqq[0] : NULL;
4792 struct bfq_queue *blocked_bfqq =
4793 !hlist_empty(&bfqq->woken_list) ?
4794 container_of(bfqq->woken_list.first,
4800 * The next four mutually-exclusive ifs decide
4801 * whether to try injection, and choose the queue to
4802 * pick an I/O request from.
4804 * The first if checks whether the process associated
4805 * with bfqq has also async I/O pending. If so, it
4806 * injects such I/O unconditionally. Injecting async
4807 * I/O from the same process can cause no harm to the
4808 * process. On the contrary, it can only increase
4809 * bandwidth and reduce latency for the process.
4811 * The second if checks whether there happens to be a
4812 * non-empty waker queue for bfqq, i.e., a queue whose
4813 * I/O needs to be completed for bfqq to receive new
4814 * I/O. This happens, e.g., if bfqq is associated with
4815 * a process that does some sync. A sync generates
4816 * extra blocking I/O, which must be completed before
4817 * the process associated with bfqq can go on with its
4818 * I/O. If the I/O of the waker queue is not served,
4819 * then bfqq remains empty, and no I/O is dispatched,
4820 * until the idle timeout fires for bfqq. This is
4821 * likely to result in lower bandwidth and higher
4822 * latencies for bfqq, and in a severe loss of total
4823 * throughput. The best action to take is therefore to
4824 * serve the waker queue as soon as possible. So do it
4825 * (without relying on the third alternative below for
4826 * eventually serving waker_bfqq's I/O; see the last
4827 * paragraph for further details). This systematic
4828 * injection of I/O from the waker queue does not
4829 * cause any delay to bfqq's I/O. On the contrary,
4830 * next bfqq's I/O is brought forward dramatically,
4831 * for it is not blocked for milliseconds.
4833 * The third if checks whether there is a queue woken
4834 * by bfqq, and currently with pending I/O. Such a
4835 * woken queue does not steal bandwidth from bfqq,
4836 * because it remains soon without I/O if bfqq is not
4837 * served. So there is virtually no risk of loss of
4838 * bandwidth for bfqq if this woken queue has I/O
4839 * dispatched while bfqq is waiting for new I/O.
4841 * The fourth if checks whether bfqq is a queue for
4842 * which it is better to avoid injection. It is so if
4843 * bfqq delivers more throughput when served without
4844 * any further I/O from other queues in the middle, or
4845 * if the service times of bfqq's I/O requests both
4846 * count more than overall throughput, and may be
4847 * easily increased by injection (this happens if bfqq
4848 * has a short think time). If none of these
4849 * conditions holds, then a candidate queue for
4850 * injection is looked for through
4851 * bfq_choose_bfqq_for_injection(). Note that the
4852 * latter may return NULL (for example if the inject
4853 * limit for bfqq is currently 0).
4855 * NOTE: motivation for the second alternative
4857 * Thanks to the way the inject limit is updated in
4858 * bfq_update_has_short_ttime(), it is rather likely
4859 * that, if I/O is being plugged for bfqq and the
4860 * waker queue has pending I/O requests that are
4861 * blocking bfqq's I/O, then the fourth alternative
4862 * above lets the waker queue get served before the
4863 * I/O-plugging timeout fires. So one may deem the
4864 * second alternative superfluous. It is not, because
4865 * the fourth alternative may be way less effective in
4866 * case of a synchronization. For two main
4867 * reasons. First, throughput may be low because the
4868 * inject limit may be too low to guarantee the same
4869 * amount of injected I/O, from the waker queue or
4870 * other queues, that the second alternative
4871 * guarantees (the second alternative unconditionally
4872 * injects a pending I/O request of the waker queue
4873 * for each bfq_dispatch_request()). Second, with the
4874 * fourth alternative, the duration of the plugging,
4875 * i.e., the time before bfqq finally receives new I/O,
4876 * may not be minimized, because the waker queue may
4877 * happen to be served only after other queues.
4880 icq_to_bic(async_bfqq->next_rq->elv.icq) == bfqq->bic &&
4881 bfq_serv_to_charge(async_bfqq->next_rq, async_bfqq) <=
4882 bfq_bfqq_budget_left(async_bfqq))
4883 bfqq = bfqq->bic->bfqq[0];
4884 else if (bfqq->waker_bfqq &&
4885 bfq_bfqq_busy(bfqq->waker_bfqq) &&
4886 bfqq->waker_bfqq->next_rq &&
4887 bfq_serv_to_charge(bfqq->waker_bfqq->next_rq,
4888 bfqq->waker_bfqq) <=
4889 bfq_bfqq_budget_left(bfqq->waker_bfqq)
4891 bfqq = bfqq->waker_bfqq;
4892 else if (blocked_bfqq &&
4893 bfq_bfqq_busy(blocked_bfqq) &&
4894 blocked_bfqq->next_rq &&
4895 bfq_serv_to_charge(blocked_bfqq->next_rq,
4897 bfq_bfqq_budget_left(blocked_bfqq)
4899 bfqq = blocked_bfqq;
4900 else if (!idling_boosts_thr_without_issues(bfqd, bfqq) &&
4901 (bfqq->wr_coeff == 1 || bfqd->wr_busy_queues > 1 ||
4902 !bfq_bfqq_has_short_ttime(bfqq)))
4903 bfqq = bfq_choose_bfqq_for_injection(bfqd);
4910 reason = BFQQE_NO_MORE_REQUESTS;
4912 bfq_bfqq_expire(bfqd, bfqq, false, reason);
4914 bfqq = bfq_set_in_service_queue(bfqd);
4916 bfq_log_bfqq(bfqd, bfqq, "select_queue: checking new queue");
4921 bfq_log_bfqq(bfqd, bfqq, "select_queue: returned this queue");
4923 bfq_log(bfqd, "select_queue: no queue returned");
4928 static void bfq_update_wr_data(struct bfq_data *bfqd, struct bfq_queue *bfqq)
4930 struct bfq_entity *entity = &bfqq->entity;
4932 if (bfqq->wr_coeff > 1) { /* queue is being weight-raised */
4933 bfq_log_bfqq(bfqd, bfqq,
4934 "raising period dur %u/%u msec, old coeff %u, w %d(%d)",
4935 jiffies_to_msecs(jiffies - bfqq->last_wr_start_finish),
4936 jiffies_to_msecs(bfqq->wr_cur_max_time),
4938 bfqq->entity.weight, bfqq->entity.orig_weight);
4940 if (entity->prio_changed)
4941 bfq_log_bfqq(bfqd, bfqq, "WARN: pending prio change");
4944 * If the queue was activated in a burst, or too much
4945 * time has elapsed from the beginning of this
4946 * weight-raising period, then end weight raising.
4948 if (bfq_bfqq_in_large_burst(bfqq))
4949 bfq_bfqq_end_wr(bfqq);
4950 else if (time_is_before_jiffies(bfqq->last_wr_start_finish +
4951 bfqq->wr_cur_max_time)) {
4952 if (bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time ||
4953 time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
4954 bfq_wr_duration(bfqd))) {
4956 * Either in interactive weight
4957 * raising, or in soft_rt weight
4959 * interactive-weight-raising period
4960 * elapsed (so no switch back to
4961 * interactive weight raising).
4963 bfq_bfqq_end_wr(bfqq);
4965 * soft_rt finishing while still in
4966 * interactive period, switch back to
4967 * interactive weight raising
4969 switch_back_to_interactive_wr(bfqq, bfqd);
4970 bfqq->entity.prio_changed = 1;
4973 if (bfqq->wr_coeff > 1 &&
4974 bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time &&
4975 bfqq->service_from_wr > max_service_from_wr) {
4976 /* see comments on max_service_from_wr */
4977 bfq_bfqq_end_wr(bfqq);
4981 * To improve latency (for this or other queues), immediately
4982 * update weight both if it must be raised and if it must be
4983 * lowered. Since, entity may be on some active tree here, and
4984 * might have a pending change of its ioprio class, invoke
4985 * next function with the last parameter unset (see the
4986 * comments on the function).
4988 if ((entity->weight > entity->orig_weight) != (bfqq->wr_coeff > 1))
4989 __bfq_entity_update_weight_prio(bfq_entity_service_tree(entity),
4994 * Dispatch next request from bfqq.
4996 static struct request *bfq_dispatch_rq_from_bfqq(struct bfq_data *bfqd,
4997 struct bfq_queue *bfqq)
4999 struct request *rq = bfqq->next_rq;
5000 unsigned long service_to_charge;
5002 service_to_charge = bfq_serv_to_charge(rq, bfqq);
5004 bfq_bfqq_served(bfqq, service_to_charge);
5006 if (bfqq == bfqd->in_service_queue && bfqd->wait_dispatch) {
5007 bfqd->wait_dispatch = false;
5008 bfqd->waited_rq = rq;
5011 bfq_dispatch_remove(bfqd->queue, rq);
5013 if (bfqq != bfqd->in_service_queue)
5017 * If weight raising has to terminate for bfqq, then next
5018 * function causes an immediate update of bfqq's weight,
5019 * without waiting for next activation. As a consequence, on
5020 * expiration, bfqq will be timestamped as if has never been
5021 * weight-raised during this service slot, even if it has
5022 * received part or even most of the service as a
5023 * weight-raised queue. This inflates bfqq's timestamps, which
5024 * is beneficial, as bfqq is then more willing to leave the
5025 * device immediately to possible other weight-raised queues.
5027 bfq_update_wr_data(bfqd, bfqq);
5030 * Expire bfqq, pretending that its budget expired, if bfqq
5031 * belongs to CLASS_IDLE and other queues are waiting for
5034 if (!(bfq_tot_busy_queues(bfqd) > 1 && bfq_class_idle(bfqq)))
5037 bfq_bfqq_expire(bfqd, bfqq, false, BFQQE_BUDGET_EXHAUSTED);
5043 static bool bfq_has_work(struct blk_mq_hw_ctx *hctx)
5045 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5048 * Avoiding lock: a race on bfqd->busy_queues should cause at
5049 * most a call to dispatch for nothing
5051 return !list_empty_careful(&bfqd->dispatch) ||
5052 bfq_tot_busy_queues(bfqd) > 0;
5055 static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
5057 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5058 struct request *rq = NULL;
5059 struct bfq_queue *bfqq = NULL;
5061 if (!list_empty(&bfqd->dispatch)) {
5062 rq = list_first_entry(&bfqd->dispatch, struct request,
5064 list_del_init(&rq->queuelist);
5070 * Increment counters here, because this
5071 * dispatch does not follow the standard
5072 * dispatch flow (where counters are
5077 goto inc_in_driver_start_rq;
5081 * We exploit the bfq_finish_requeue_request hook to
5082 * decrement rq_in_driver, but
5083 * bfq_finish_requeue_request will not be invoked on
5084 * this request. So, to avoid unbalance, just start
5085 * this request, without incrementing rq_in_driver. As
5086 * a negative consequence, rq_in_driver is deceptively
5087 * lower than it should be while this request is in
5088 * service. This may cause bfq_schedule_dispatch to be
5089 * invoked uselessly.
5091 * As for implementing an exact solution, the
5092 * bfq_finish_requeue_request hook, if defined, is
5093 * probably invoked also on this request. So, by
5094 * exploiting this hook, we could 1) increment
5095 * rq_in_driver here, and 2) decrement it in
5096 * bfq_finish_requeue_request. Such a solution would
5097 * let the value of the counter be always accurate,
5098 * but it would entail using an extra interface
5099 * function. This cost seems higher than the benefit,
5100 * being the frequency of non-elevator-private
5101 * requests very low.
5106 bfq_log(bfqd, "dispatch requests: %d busy queues",
5107 bfq_tot_busy_queues(bfqd));
5109 if (bfq_tot_busy_queues(bfqd) == 0)
5113 * Force device to serve one request at a time if
5114 * strict_guarantees is true. Forcing this service scheme is
5115 * currently the ONLY way to guarantee that the request
5116 * service order enforced by the scheduler is respected by a
5117 * queueing device. Otherwise the device is free even to make
5118 * some unlucky request wait for as long as the device
5121 * Of course, serving one request at a time may cause loss of
5124 if (bfqd->strict_guarantees && bfqd->rq_in_driver > 0)
5127 bfqq = bfq_select_queue(bfqd);
5131 rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq);
5134 inc_in_driver_start_rq:
5135 bfqd->rq_in_driver++;
5137 rq->rq_flags |= RQF_STARTED;
5143 #ifdef CONFIG_BFQ_CGROUP_DEBUG
5144 static void bfq_update_dispatch_stats(struct request_queue *q,
5146 struct bfq_queue *in_serv_queue,
5147 bool idle_timer_disabled)
5149 struct bfq_queue *bfqq = rq ? RQ_BFQQ(rq) : NULL;
5151 if (!idle_timer_disabled && !bfqq)
5155 * rq and bfqq are guaranteed to exist until this function
5156 * ends, for the following reasons. First, rq can be
5157 * dispatched to the device, and then can be completed and
5158 * freed, only after this function ends. Second, rq cannot be
5159 * merged (and thus freed because of a merge) any longer,
5160 * because it has already started. Thus rq cannot be freed
5161 * before this function ends, and, since rq has a reference to
5162 * bfqq, the same guarantee holds for bfqq too.
5164 * In addition, the following queue lock guarantees that
5165 * bfqq_group(bfqq) exists as well.
5167 spin_lock_irq(&q->queue_lock);
5168 if (idle_timer_disabled)
5170 * Since the idle timer has been disabled,
5171 * in_serv_queue contained some request when
5172 * __bfq_dispatch_request was invoked above, which
5173 * implies that rq was picked exactly from
5174 * in_serv_queue. Thus in_serv_queue == bfqq, and is
5175 * therefore guaranteed to exist because of the above
5178 bfqg_stats_update_idle_time(bfqq_group(in_serv_queue));
5180 struct bfq_group *bfqg = bfqq_group(bfqq);
5182 bfqg_stats_update_avg_queue_size(bfqg);
5183 bfqg_stats_set_start_empty_time(bfqg);
5184 bfqg_stats_update_io_remove(bfqg, rq->cmd_flags);
5186 spin_unlock_irq(&q->queue_lock);
5189 static inline void bfq_update_dispatch_stats(struct request_queue *q,
5191 struct bfq_queue *in_serv_queue,
5192 bool idle_timer_disabled) {}
5193 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
5195 static struct request *bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
5197 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5199 struct bfq_queue *in_serv_queue;
5200 bool waiting_rq, idle_timer_disabled = false;
5202 spin_lock_irq(&bfqd->lock);
5204 in_serv_queue = bfqd->in_service_queue;
5205 waiting_rq = in_serv_queue && bfq_bfqq_wait_request(in_serv_queue);
5207 rq = __bfq_dispatch_request(hctx);
5208 if (in_serv_queue == bfqd->in_service_queue) {
5209 idle_timer_disabled =
5210 waiting_rq && !bfq_bfqq_wait_request(in_serv_queue);
5213 spin_unlock_irq(&bfqd->lock);
5214 bfq_update_dispatch_stats(hctx->queue, rq,
5215 idle_timer_disabled ? in_serv_queue : NULL,
5216 idle_timer_disabled);
5222 * Task holds one reference to the queue, dropped when task exits. Each rq
5223 * in-flight on this queue also holds a reference, dropped when rq is freed.
5225 * Scheduler lock must be held here. Recall not to use bfqq after calling
5226 * this function on it.
5228 void bfq_put_queue(struct bfq_queue *bfqq)
5230 struct bfq_queue *item;
5231 struct hlist_node *n;
5232 struct bfq_group *bfqg = bfqq_group(bfqq);
5235 bfq_log_bfqq(bfqq->bfqd, bfqq, "put_queue: %p %d",
5242 if (!hlist_unhashed(&bfqq->burst_list_node)) {
5243 hlist_del_init(&bfqq->burst_list_node);
5245 * Decrement also burst size after the removal, if the
5246 * process associated with bfqq is exiting, and thus
5247 * does not contribute to the burst any longer. This
5248 * decrement helps filter out false positives of large
5249 * bursts, when some short-lived process (often due to
5250 * the execution of commands by some service) happens
5251 * to start and exit while a complex application is
5252 * starting, and thus spawning several processes that
5253 * do I/O (and that *must not* be treated as a large
5254 * burst, see comments on bfq_handle_burst).
5256 * In particular, the decrement is performed only if:
5257 * 1) bfqq is not a merged queue, because, if it is,
5258 * then this free of bfqq is not triggered by the exit
5259 * of the process bfqq is associated with, but exactly
5260 * by the fact that bfqq has just been merged.
5261 * 2) burst_size is greater than 0, to handle
5262 * unbalanced decrements. Unbalanced decrements may
5263 * happen in te following case: bfqq is inserted into
5264 * the current burst list--without incrementing
5265 * bust_size--because of a split, but the current
5266 * burst list is not the burst list bfqq belonged to
5267 * (see comments on the case of a split in
5270 if (bfqq->bic && bfqq->bfqd->burst_size > 0)
5271 bfqq->bfqd->burst_size--;
5275 * bfqq does not exist any longer, so it cannot be woken by
5276 * any other queue, and cannot wake any other queue. Then bfqq
5277 * must be removed from the woken list of its possible waker
5278 * queue, and all queues in the woken list of bfqq must stop
5279 * having a waker queue. Strictly speaking, these updates
5280 * should be performed when bfqq remains with no I/O source
5281 * attached to it, which happens before bfqq gets freed. In
5282 * particular, this happens when the last process associated
5283 * with bfqq exits or gets associated with a different
5284 * queue. However, both events lead to bfqq being freed soon,
5285 * and dangling references would come out only after bfqq gets
5286 * freed. So these updates are done here, as a simple and safe
5287 * way to handle all cases.
5289 /* remove bfqq from woken list */
5290 if (!hlist_unhashed(&bfqq->woken_list_node))
5291 hlist_del_init(&bfqq->woken_list_node);
5293 /* reset waker for all queues in woken list */
5294 hlist_for_each_entry_safe(item, n, &bfqq->woken_list,
5296 item->waker_bfqq = NULL;
5297 hlist_del_init(&item->woken_list_node);
5300 if (bfqq->bfqd && bfqq->bfqd->last_completed_rq_bfqq == bfqq)
5301 bfqq->bfqd->last_completed_rq_bfqq = NULL;
5303 kmem_cache_free(bfq_pool, bfqq);
5304 bfqg_and_blkg_put(bfqg);
5307 static void bfq_put_stable_ref(struct bfq_queue *bfqq)
5310 bfq_put_queue(bfqq);
5313 static void bfq_put_cooperator(struct bfq_queue *bfqq)
5315 struct bfq_queue *__bfqq, *next;
5318 * If this queue was scheduled to merge with another queue, be
5319 * sure to drop the reference taken on that queue (and others in
5320 * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs.
5322 __bfqq = bfqq->new_bfqq;
5326 next = __bfqq->new_bfqq;
5327 bfq_put_queue(__bfqq);
5332 static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq)
5334 if (bfqq == bfqd->in_service_queue) {
5335 __bfq_bfqq_expire(bfqd, bfqq, BFQQE_BUDGET_TIMEOUT);
5336 bfq_schedule_dispatch(bfqd);
5339 bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, bfqq->ref);
5341 bfq_put_cooperator(bfqq);
5343 bfq_release_process_ref(bfqd, bfqq);
5346 static void bfq_exit_icq_bfqq(struct bfq_io_cq *bic, bool is_sync)
5348 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
5349 struct bfq_data *bfqd;
5352 bfqd = bfqq->bfqd; /* NULL if scheduler already exited */
5355 unsigned long flags;
5357 spin_lock_irqsave(&bfqd->lock, flags);
5359 bfq_exit_bfqq(bfqd, bfqq);
5360 bic_set_bfqq(bic, NULL, is_sync);
5361 spin_unlock_irqrestore(&bfqd->lock, flags);
5365 static void bfq_exit_icq(struct io_cq *icq)
5367 struct bfq_io_cq *bic = icq_to_bic(icq);
5369 if (bic->stable_merge_bfqq) {
5370 struct bfq_data *bfqd = bic->stable_merge_bfqq->bfqd;
5373 * bfqd is NULL if scheduler already exited, and in
5374 * that case this is the last time bfqq is accessed.
5377 unsigned long flags;
5379 spin_lock_irqsave(&bfqd->lock, flags);
5380 bfq_put_stable_ref(bic->stable_merge_bfqq);
5381 spin_unlock_irqrestore(&bfqd->lock, flags);
5383 bfq_put_stable_ref(bic->stable_merge_bfqq);
5387 bfq_exit_icq_bfqq(bic, true);
5388 bfq_exit_icq_bfqq(bic, false);
5392 * Update the entity prio values; note that the new values will not
5393 * be used until the next (re)activation.
5396 bfq_set_next_ioprio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic)
5398 struct task_struct *tsk = current;
5400 struct bfq_data *bfqd = bfqq->bfqd;
5405 ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
5406 switch (ioprio_class) {
5408 pr_err("bdi %s: bfq: bad prio class %d\n",
5409 bdi_dev_name(bfqq->bfqd->queue->disk->bdi),
5412 case IOPRIO_CLASS_NONE:
5414 * No prio set, inherit CPU scheduling settings.
5416 bfqq->new_ioprio = task_nice_ioprio(tsk);
5417 bfqq->new_ioprio_class = task_nice_ioclass(tsk);
5419 case IOPRIO_CLASS_RT:
5420 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
5421 bfqq->new_ioprio_class = IOPRIO_CLASS_RT;
5423 case IOPRIO_CLASS_BE:
5424 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
5425 bfqq->new_ioprio_class = IOPRIO_CLASS_BE;
5427 case IOPRIO_CLASS_IDLE:
5428 bfqq->new_ioprio_class = IOPRIO_CLASS_IDLE;
5429 bfqq->new_ioprio = 7;
5433 if (bfqq->new_ioprio >= IOPRIO_NR_LEVELS) {
5434 pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n",
5436 bfqq->new_ioprio = IOPRIO_NR_LEVELS - 1;
5439 bfqq->entity.new_weight = bfq_ioprio_to_weight(bfqq->new_ioprio);
5440 bfq_log_bfqq(bfqd, bfqq, "new_ioprio %d new_weight %d",
5441 bfqq->new_ioprio, bfqq->entity.new_weight);
5442 bfqq->entity.prio_changed = 1;
5445 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
5446 struct bio *bio, bool is_sync,
5447 struct bfq_io_cq *bic,
5450 static void bfq_check_ioprio_change(struct bfq_io_cq *bic, struct bio *bio)
5452 struct bfq_data *bfqd = bic_to_bfqd(bic);
5453 struct bfq_queue *bfqq;
5454 int ioprio = bic->icq.ioc->ioprio;
5457 * This condition may trigger on a newly created bic, be sure to
5458 * drop the lock before returning.
5460 if (unlikely(!bfqd) || likely(bic->ioprio == ioprio))
5463 bic->ioprio = ioprio;
5465 bfqq = bic_to_bfqq(bic, false);
5467 bfq_release_process_ref(bfqd, bfqq);
5468 bfqq = bfq_get_queue(bfqd, bio, false, bic, true);
5469 bic_set_bfqq(bic, bfqq, false);
5472 bfqq = bic_to_bfqq(bic, true);
5474 bfq_set_next_ioprio_data(bfqq, bic);
5477 static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5478 struct bfq_io_cq *bic, pid_t pid, int is_sync)
5480 u64 now_ns = ktime_get_ns();
5482 RB_CLEAR_NODE(&bfqq->entity.rb_node);
5483 INIT_LIST_HEAD(&bfqq->fifo);
5484 INIT_HLIST_NODE(&bfqq->burst_list_node);
5485 INIT_HLIST_NODE(&bfqq->woken_list_node);
5486 INIT_HLIST_HEAD(&bfqq->woken_list);
5492 bfq_set_next_ioprio_data(bfqq, bic);
5496 * No need to mark as has_short_ttime if in
5497 * idle_class, because no device idling is performed
5498 * for queues in idle class
5500 if (!bfq_class_idle(bfqq))
5501 /* tentatively mark as has_short_ttime */
5502 bfq_mark_bfqq_has_short_ttime(bfqq);
5503 bfq_mark_bfqq_sync(bfqq);
5504 bfq_mark_bfqq_just_created(bfqq);
5506 bfq_clear_bfqq_sync(bfqq);
5508 /* set end request to minus infinity from now */
5509 bfqq->ttime.last_end_request = now_ns + 1;
5511 bfqq->creation_time = jiffies;
5513 bfqq->io_start_time = now_ns;
5515 bfq_mark_bfqq_IO_bound(bfqq);
5519 /* Tentative initial value to trade off between thr and lat */
5520 bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3;
5521 bfqq->budget_timeout = bfq_smallest_from_now();
5524 bfqq->last_wr_start_finish = jiffies;
5525 bfqq->wr_start_at_switch_to_srt = bfq_smallest_from_now();
5526 bfqq->split_time = bfq_smallest_from_now();
5529 * To not forget the possibly high bandwidth consumed by a
5530 * process/queue in the recent past,
5531 * bfq_bfqq_softrt_next_start() returns a value at least equal
5532 * to the current value of bfqq->soft_rt_next_start (see
5533 * comments on bfq_bfqq_softrt_next_start). Set
5534 * soft_rt_next_start to now, to mean that bfqq has consumed
5535 * no bandwidth so far.
5537 bfqq->soft_rt_next_start = jiffies;
5539 /* first request is almost certainly seeky */
5540 bfqq->seek_history = 1;
5543 static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd,
5544 struct bfq_group *bfqg,
5545 int ioprio_class, int ioprio)
5547 switch (ioprio_class) {
5548 case IOPRIO_CLASS_RT:
5549 return &bfqg->async_bfqq[0][ioprio];
5550 case IOPRIO_CLASS_NONE:
5551 ioprio = IOPRIO_BE_NORM;
5553 case IOPRIO_CLASS_BE:
5554 return &bfqg->async_bfqq[1][ioprio];
5555 case IOPRIO_CLASS_IDLE:
5556 return &bfqg->async_idle_bfqq;
5562 static struct bfq_queue *
5563 bfq_do_early_stable_merge(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5564 struct bfq_io_cq *bic,
5565 struct bfq_queue *last_bfqq_created)
5567 struct bfq_queue *new_bfqq =
5568 bfq_setup_merge(bfqq, last_bfqq_created);
5574 new_bfqq->bic->stably_merged = true;
5575 bic->stably_merged = true;
5578 * Reusing merge functions. This implies that
5579 * bfqq->bic must be set too, for
5580 * bfq_merge_bfqqs to correctly save bfqq's
5581 * state before killing it.
5584 bfq_merge_bfqqs(bfqd, bic, bfqq, new_bfqq);
5590 * Many throughput-sensitive workloads are made of several parallel
5591 * I/O flows, with all flows generated by the same application, or
5592 * more generically by the same task (e.g., system boot). The most
5593 * counterproductive action with these workloads is plugging I/O
5594 * dispatch when one of the bfq_queues associated with these flows
5595 * remains temporarily empty.
5597 * To avoid this plugging, BFQ has been using a burst-handling
5598 * mechanism for years now. This mechanism has proven effective for
5599 * throughput, and not detrimental for service guarantees. The
5600 * following function pushes this mechanism a little bit further,
5601 * basing on the following two facts.
5603 * First, all the I/O flows of a the same application or task
5604 * contribute to the execution/completion of that common application
5605 * or task. So the performance figures that matter are total
5606 * throughput of the flows and task-wide I/O latency. In particular,
5607 * these flows do not need to be protected from each other, in terms
5608 * of individual bandwidth or latency.
5610 * Second, the above fact holds regardless of the number of flows.
5612 * Putting these two facts together, this commits merges stably the
5613 * bfq_queues associated with these I/O flows, i.e., with the
5614 * processes that generate these IO/ flows, regardless of how many the
5615 * involved processes are.
5617 * To decide whether a set of bfq_queues is actually associated with
5618 * the I/O flows of a common application or task, and to merge these
5619 * queues stably, this function operates as follows: given a bfq_queue,
5620 * say Q2, currently being created, and the last bfq_queue, say Q1,
5621 * created before Q2, Q2 is merged stably with Q1 if
5622 * - very little time has elapsed since when Q1 was created
5623 * - Q2 has the same ioprio as Q1
5624 * - Q2 belongs to the same group as Q1
5626 * Merging bfq_queues also reduces scheduling overhead. A fio test
5627 * with ten random readers on /dev/nullb shows a throughput boost of
5628 * 40%, with a quadcore. Since BFQ's execution time amounts to ~50% of
5629 * the total per-request processing time, the above throughput boost
5630 * implies that BFQ's overhead is reduced by more than 50%.
5632 * This new mechanism most certainly obsoletes the current
5633 * burst-handling heuristics. We keep those heuristics for the moment.
5635 static struct bfq_queue *bfq_do_or_sched_stable_merge(struct bfq_data *bfqd,
5636 struct bfq_queue *bfqq,
5637 struct bfq_io_cq *bic)
5639 struct bfq_queue **source_bfqq = bfqq->entity.parent ?
5640 &bfqq->entity.parent->last_bfqq_created :
5641 &bfqd->last_bfqq_created;
5643 struct bfq_queue *last_bfqq_created = *source_bfqq;
5646 * If last_bfqq_created has not been set yet, then init it. If
5647 * it has been set already, but too long ago, then move it
5648 * forward to bfqq. Finally, move also if bfqq belongs to a
5649 * different group than last_bfqq_created, or if bfqq has a
5650 * different ioprio or ioprio_class. If none of these
5651 * conditions holds true, then try an early stable merge or
5652 * schedule a delayed stable merge.
5654 * A delayed merge is scheduled (instead of performing an
5655 * early merge), in case bfqq might soon prove to be more
5656 * throughput-beneficial if not merged. Currently this is
5657 * possible only if bfqd is rotational with no queueing. For
5658 * such a drive, not merging bfqq is better for throughput if
5659 * bfqq happens to contain sequential I/O. So, we wait a
5660 * little bit for enough I/O to flow through bfqq. After that,
5661 * if such an I/O is sequential, then the merge is
5662 * canceled. Otherwise the merge is finally performed.
5664 if (!last_bfqq_created ||
5665 time_before(last_bfqq_created->creation_time +
5666 msecs_to_jiffies(bfq_activation_stable_merging),
5667 bfqq->creation_time) ||
5668 bfqq->entity.parent != last_bfqq_created->entity.parent ||
5669 bfqq->ioprio != last_bfqq_created->ioprio ||
5670 bfqq->ioprio_class != last_bfqq_created->ioprio_class)
5671 *source_bfqq = bfqq;
5672 else if (time_after_eq(last_bfqq_created->creation_time +
5673 bfqd->bfq_burst_interval,
5674 bfqq->creation_time)) {
5675 if (likely(bfqd->nonrot_with_queueing))
5677 * With this type of drive, leaving
5678 * bfqq alone may provide no
5679 * throughput benefits compared with
5680 * merging bfqq. So merge bfqq now.
5682 bfqq = bfq_do_early_stable_merge(bfqd, bfqq,
5685 else { /* schedule tentative stable merge */
5687 * get reference on last_bfqq_created,
5688 * to prevent it from being freed,
5689 * until we decide whether to merge
5691 last_bfqq_created->ref++;
5693 * need to keep track of stable refs, to
5694 * compute process refs correctly
5696 last_bfqq_created->stable_ref++;
5698 * Record the bfqq to merge to.
5700 bic->stable_merge_bfqq = last_bfqq_created;
5708 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
5709 struct bio *bio, bool is_sync,
5710 struct bfq_io_cq *bic,
5713 const int ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
5714 const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
5715 struct bfq_queue **async_bfqq = NULL;
5716 struct bfq_queue *bfqq;
5717 struct bfq_group *bfqg;
5721 bfqg = bfq_find_set_group(bfqd, __bio_blkcg(bio));
5723 bfqq = &bfqd->oom_bfqq;
5728 async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class,
5735 bfqq = kmem_cache_alloc_node(bfq_pool,
5736 GFP_NOWAIT | __GFP_ZERO | __GFP_NOWARN,
5740 bfq_init_bfqq(bfqd, bfqq, bic, current->pid,
5742 bfq_init_entity(&bfqq->entity, bfqg);
5743 bfq_log_bfqq(bfqd, bfqq, "allocated");
5745 bfqq = &bfqd->oom_bfqq;
5746 bfq_log_bfqq(bfqd, bfqq, "using oom bfqq");
5751 * Pin the queue now that it's allocated, scheduler exit will
5756 * Extra group reference, w.r.t. sync
5757 * queue. This extra reference is removed
5758 * only if bfqq->bfqg disappears, to
5759 * guarantee that this queue is not freed
5760 * until its group goes away.
5762 bfq_log_bfqq(bfqd, bfqq, "get_queue, bfqq not in async: %p, %d",
5768 bfqq->ref++; /* get a process reference to this queue */
5770 if (bfqq != &bfqd->oom_bfqq && is_sync && !respawn)
5771 bfqq = bfq_do_or_sched_stable_merge(bfqd, bfqq, bic);
5777 static void bfq_update_io_thinktime(struct bfq_data *bfqd,
5778 struct bfq_queue *bfqq)
5780 struct bfq_ttime *ttime = &bfqq->ttime;
5784 * We are really interested in how long it takes for the queue to
5785 * become busy when there is no outstanding IO for this queue. So
5786 * ignore cases when the bfq queue has already IO queued.
5788 if (bfqq->dispatched || bfq_bfqq_busy(bfqq))
5790 elapsed = ktime_get_ns() - bfqq->ttime.last_end_request;
5791 elapsed = min_t(u64, elapsed, 2ULL * bfqd->bfq_slice_idle);
5793 ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
5794 ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed, 8);
5795 ttime->ttime_mean = div64_ul(ttime->ttime_total + 128,
5796 ttime->ttime_samples);
5800 bfq_update_io_seektime(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5803 bfqq->seek_history <<= 1;
5804 bfqq->seek_history |= BFQ_RQ_SEEKY(bfqd, bfqq->last_request_pos, rq);
5806 if (bfqq->wr_coeff > 1 &&
5807 bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
5808 BFQQ_TOTALLY_SEEKY(bfqq)) {
5809 if (time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
5810 bfq_wr_duration(bfqd))) {
5812 * In soft_rt weight raising with the
5813 * interactive-weight-raising period
5814 * elapsed (so no switch back to
5815 * interactive weight raising).
5817 bfq_bfqq_end_wr(bfqq);
5819 * stopping soft_rt weight raising
5820 * while still in interactive period,
5821 * switch back to interactive weight
5824 switch_back_to_interactive_wr(bfqq, bfqd);
5825 bfqq->entity.prio_changed = 1;
5830 static void bfq_update_has_short_ttime(struct bfq_data *bfqd,
5831 struct bfq_queue *bfqq,
5832 struct bfq_io_cq *bic)
5834 bool has_short_ttime = true, state_changed;
5837 * No need to update has_short_ttime if bfqq is async or in
5838 * idle io prio class, or if bfq_slice_idle is zero, because
5839 * no device idling is performed for bfqq in this case.
5841 if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq) ||
5842 bfqd->bfq_slice_idle == 0)
5845 /* Idle window just restored, statistics are meaningless. */
5846 if (time_is_after_eq_jiffies(bfqq->split_time +
5847 bfqd->bfq_wr_min_idle_time))
5850 /* Think time is infinite if no process is linked to
5851 * bfqq. Otherwise check average think time to decide whether
5852 * to mark as has_short_ttime. To this goal, compare average
5853 * think time with half the I/O-plugging timeout.
5855 if (atomic_read(&bic->icq.ioc->active_ref) == 0 ||
5856 (bfq_sample_valid(bfqq->ttime.ttime_samples) &&
5857 bfqq->ttime.ttime_mean > bfqd->bfq_slice_idle>>1))
5858 has_short_ttime = false;
5860 state_changed = has_short_ttime != bfq_bfqq_has_short_ttime(bfqq);
5862 if (has_short_ttime)
5863 bfq_mark_bfqq_has_short_ttime(bfqq);
5865 bfq_clear_bfqq_has_short_ttime(bfqq);
5868 * Until the base value for the total service time gets
5869 * finally computed for bfqq, the inject limit does depend on
5870 * the think-time state (short|long). In particular, the limit
5871 * is 0 or 1 if the think time is deemed, respectively, as
5872 * short or long (details in the comments in
5873 * bfq_update_inject_limit()). Accordingly, the next
5874 * instructions reset the inject limit if the think-time state
5875 * has changed and the above base value is still to be
5878 * However, the reset is performed only if more than 100 ms
5879 * have elapsed since the last update of the inject limit, or
5880 * (inclusive) if the change is from short to long think
5881 * time. The reason for this waiting is as follows.
5883 * bfqq may have a long think time because of a
5884 * synchronization with some other queue, i.e., because the
5885 * I/O of some other queue may need to be completed for bfqq
5886 * to receive new I/O. Details in the comments on the choice
5887 * of the queue for injection in bfq_select_queue().
5889 * As stressed in those comments, if such a synchronization is
5890 * actually in place, then, without injection on bfqq, the
5891 * blocking I/O cannot happen to served while bfqq is in
5892 * service. As a consequence, if bfqq is granted
5893 * I/O-dispatch-plugging, then bfqq remains empty, and no I/O
5894 * is dispatched, until the idle timeout fires. This is likely
5895 * to result in lower bandwidth and higher latencies for bfqq,
5896 * and in a severe loss of total throughput.
5898 * On the opposite end, a non-zero inject limit may allow the
5899 * I/O that blocks bfqq to be executed soon, and therefore
5900 * bfqq to receive new I/O soon.
5902 * But, if the blocking gets actually eliminated, then the
5903 * next think-time sample for bfqq may be very low. This in
5904 * turn may cause bfqq's think time to be deemed
5905 * short. Without the 100 ms barrier, this new state change
5906 * would cause the body of the next if to be executed
5907 * immediately. But this would set to 0 the inject
5908 * limit. Without injection, the blocking I/O would cause the
5909 * think time of bfqq to become long again, and therefore the
5910 * inject limit to be raised again, and so on. The only effect
5911 * of such a steady oscillation between the two think-time
5912 * states would be to prevent effective injection on bfqq.
5914 * In contrast, if the inject limit is not reset during such a
5915 * long time interval as 100 ms, then the number of short
5916 * think time samples can grow significantly before the reset
5917 * is performed. As a consequence, the think time state can
5918 * become stable before the reset. Therefore there will be no
5919 * state change when the 100 ms elapse, and no reset of the
5920 * inject limit. The inject limit remains steadily equal to 1
5921 * both during and after the 100 ms. So injection can be
5922 * performed at all times, and throughput gets boosted.
5924 * An inject limit equal to 1 is however in conflict, in
5925 * general, with the fact that the think time of bfqq is
5926 * short, because injection may be likely to delay bfqq's I/O
5927 * (as explained in the comments in
5928 * bfq_update_inject_limit()). But this does not happen in
5929 * this special case, because bfqq's low think time is due to
5930 * an effective handling of a synchronization, through
5931 * injection. In this special case, bfqq's I/O does not get
5932 * delayed by injection; on the contrary, bfqq's I/O is
5933 * brought forward, because it is not blocked for
5936 * In addition, serving the blocking I/O much sooner, and much
5937 * more frequently than once per I/O-plugging timeout, makes
5938 * it much quicker to detect a waker queue (the concept of
5939 * waker queue is defined in the comments in
5940 * bfq_add_request()). This makes it possible to start sooner
5941 * to boost throughput more effectively, by injecting the I/O
5942 * of the waker queue unconditionally on every
5943 * bfq_dispatch_request().
5945 * One last, important benefit of not resetting the inject
5946 * limit before 100 ms is that, during this time interval, the
5947 * base value for the total service time is likely to get
5948 * finally computed for bfqq, freeing the inject limit from
5949 * its relation with the think time.
5951 if (state_changed && bfqq->last_serv_time_ns == 0 &&
5952 (time_is_before_eq_jiffies(bfqq->decrease_time_jif +
5953 msecs_to_jiffies(100)) ||
5955 bfq_reset_inject_limit(bfqd, bfqq);
5959 * Called when a new fs request (rq) is added to bfqq. Check if there's
5960 * something we should do about it.
5962 static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5965 if (rq->cmd_flags & REQ_META)
5966 bfqq->meta_pending++;
5968 bfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
5970 if (bfqq == bfqd->in_service_queue && bfq_bfqq_wait_request(bfqq)) {
5971 bool small_req = bfqq->queued[rq_is_sync(rq)] == 1 &&
5972 blk_rq_sectors(rq) < 32;
5973 bool budget_timeout = bfq_bfqq_budget_timeout(bfqq);
5976 * There is just this request queued: if
5977 * - the request is small, and
5978 * - we are idling to boost throughput, and
5979 * - the queue is not to be expired,
5982 * In this way, if the device is being idled to wait
5983 * for a new request from the in-service queue, we
5984 * avoid unplugging the device and committing the
5985 * device to serve just a small request. In contrast
5986 * we wait for the block layer to decide when to
5987 * unplug the device: hopefully, new requests will be
5988 * merged to this one quickly, then the device will be
5989 * unplugged and larger requests will be dispatched.
5991 if (small_req && idling_boosts_thr_without_issues(bfqd, bfqq) &&
5996 * A large enough request arrived, or idling is being
5997 * performed to preserve service guarantees, or
5998 * finally the queue is to be expired: in all these
5999 * cases disk idling is to be stopped, so clear
6000 * wait_request flag and reset timer.
6002 bfq_clear_bfqq_wait_request(bfqq);
6003 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
6006 * The queue is not empty, because a new request just
6007 * arrived. Hence we can safely expire the queue, in
6008 * case of budget timeout, without risking that the
6009 * timestamps of the queue are not updated correctly.
6010 * See [1] for more details.
6013 bfq_bfqq_expire(bfqd, bfqq, false,
6014 BFQQE_BUDGET_TIMEOUT);
6018 static void bfqq_request_allocated(struct bfq_queue *bfqq)
6020 struct bfq_entity *entity = &bfqq->entity;
6022 for_each_entity(entity)
6023 entity->allocated++;
6026 static void bfqq_request_freed(struct bfq_queue *bfqq)
6028 struct bfq_entity *entity = &bfqq->entity;
6030 for_each_entity(entity)
6031 entity->allocated--;
6034 /* returns true if it causes the idle timer to be disabled */
6035 static bool __bfq_insert_request(struct bfq_data *bfqd, struct request *rq)
6037 struct bfq_queue *bfqq = RQ_BFQQ(rq),
6038 *new_bfqq = bfq_setup_cooperator(bfqd, bfqq, rq, true,
6040 bool waiting, idle_timer_disabled = false;
6044 * Release the request's reference to the old bfqq
6045 * and make sure one is taken to the shared queue.
6047 bfqq_request_allocated(new_bfqq);
6048 bfqq_request_freed(bfqq);
6051 * If the bic associated with the process
6052 * issuing this request still points to bfqq
6053 * (and thus has not been already redirected
6054 * to new_bfqq or even some other bfq_queue),
6055 * then complete the merge and redirect it to
6058 if (bic_to_bfqq(RQ_BIC(rq), 1) == bfqq)
6059 bfq_merge_bfqqs(bfqd, RQ_BIC(rq),
6062 bfq_clear_bfqq_just_created(bfqq);
6064 * rq is about to be enqueued into new_bfqq,
6065 * release rq reference on bfqq
6067 bfq_put_queue(bfqq);
6068 rq->elv.priv[1] = new_bfqq;
6072 bfq_update_io_thinktime(bfqd, bfqq);
6073 bfq_update_has_short_ttime(bfqd, bfqq, RQ_BIC(rq));
6074 bfq_update_io_seektime(bfqd, bfqq, rq);
6076 waiting = bfqq && bfq_bfqq_wait_request(bfqq);
6077 bfq_add_request(rq);
6078 idle_timer_disabled = waiting && !bfq_bfqq_wait_request(bfqq);
6080 rq->fifo_time = ktime_get_ns() + bfqd->bfq_fifo_expire[rq_is_sync(rq)];
6081 list_add_tail(&rq->queuelist, &bfqq->fifo);
6083 bfq_rq_enqueued(bfqd, bfqq, rq);
6085 return idle_timer_disabled;
6088 #ifdef CONFIG_BFQ_CGROUP_DEBUG
6089 static void bfq_update_insert_stats(struct request_queue *q,
6090 struct bfq_queue *bfqq,
6091 bool idle_timer_disabled,
6092 unsigned int cmd_flags)
6098 * bfqq still exists, because it can disappear only after
6099 * either it is merged with another queue, or the process it
6100 * is associated with exits. But both actions must be taken by
6101 * the same process currently executing this flow of
6104 * In addition, the following queue lock guarantees that
6105 * bfqq_group(bfqq) exists as well.
6107 spin_lock_irq(&q->queue_lock);
6108 bfqg_stats_update_io_add(bfqq_group(bfqq), bfqq, cmd_flags);
6109 if (idle_timer_disabled)
6110 bfqg_stats_update_idle_time(bfqq_group(bfqq));
6111 spin_unlock_irq(&q->queue_lock);
6114 static inline void bfq_update_insert_stats(struct request_queue *q,
6115 struct bfq_queue *bfqq,
6116 bool idle_timer_disabled,
6117 unsigned int cmd_flags) {}
6118 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
6120 static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
6123 struct request_queue *q = hctx->queue;
6124 struct bfq_data *bfqd = q->elevator->elevator_data;
6125 struct bfq_queue *bfqq;
6126 bool idle_timer_disabled = false;
6127 unsigned int cmd_flags;
6130 #ifdef CONFIG_BFQ_GROUP_IOSCHED
6131 if (!cgroup_subsys_on_dfl(io_cgrp_subsys) && rq->bio)
6132 bfqg_stats_update_legacy_io(q, rq);
6134 spin_lock_irq(&bfqd->lock);
6135 if (blk_mq_sched_try_insert_merge(q, rq, &free)) {
6136 spin_unlock_irq(&bfqd->lock);
6137 blk_mq_free_requests(&free);
6141 spin_unlock_irq(&bfqd->lock);
6143 trace_block_rq_insert(rq);
6145 spin_lock_irq(&bfqd->lock);
6146 bfqq = bfq_init_rq(rq);
6147 if (!bfqq || at_head) {
6149 list_add(&rq->queuelist, &bfqd->dispatch);
6151 list_add_tail(&rq->queuelist, &bfqd->dispatch);
6153 idle_timer_disabled = __bfq_insert_request(bfqd, rq);
6155 * Update bfqq, because, if a queue merge has occurred
6156 * in __bfq_insert_request, then rq has been
6157 * redirected into a new queue.
6161 if (rq_mergeable(rq)) {
6162 elv_rqhash_add(q, rq);
6169 * Cache cmd_flags before releasing scheduler lock, because rq
6170 * may disappear afterwards (for example, because of a request
6173 cmd_flags = rq->cmd_flags;
6174 spin_unlock_irq(&bfqd->lock);
6176 bfq_update_insert_stats(q, bfqq, idle_timer_disabled,
6180 static void bfq_insert_requests(struct blk_mq_hw_ctx *hctx,
6181 struct list_head *list, bool at_head)
6183 while (!list_empty(list)) {
6186 rq = list_first_entry(list, struct request, queuelist);
6187 list_del_init(&rq->queuelist);
6188 bfq_insert_request(hctx, rq, at_head);
6192 static void bfq_update_hw_tag(struct bfq_data *bfqd)
6194 struct bfq_queue *bfqq = bfqd->in_service_queue;
6196 bfqd->max_rq_in_driver = max_t(int, bfqd->max_rq_in_driver,
6197 bfqd->rq_in_driver);
6199 if (bfqd->hw_tag == 1)
6203 * This sample is valid if the number of outstanding requests
6204 * is large enough to allow a queueing behavior. Note that the
6205 * sum is not exact, as it's not taking into account deactivated
6208 if (bfqd->rq_in_driver + bfqd->queued <= BFQ_HW_QUEUE_THRESHOLD)
6212 * If active queue hasn't enough requests and can idle, bfq might not
6213 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
6216 if (bfqq && bfq_bfqq_has_short_ttime(bfqq) &&
6217 bfqq->dispatched + bfqq->queued[0] + bfqq->queued[1] <
6218 BFQ_HW_QUEUE_THRESHOLD &&
6219 bfqd->rq_in_driver < BFQ_HW_QUEUE_THRESHOLD)
6222 if (bfqd->hw_tag_samples++ < BFQ_HW_QUEUE_SAMPLES)
6225 bfqd->hw_tag = bfqd->max_rq_in_driver > BFQ_HW_QUEUE_THRESHOLD;
6226 bfqd->max_rq_in_driver = 0;
6227 bfqd->hw_tag_samples = 0;
6229 bfqd->nonrot_with_queueing =
6230 blk_queue_nonrot(bfqd->queue) && bfqd->hw_tag;
6233 static void bfq_completed_request(struct bfq_queue *bfqq, struct bfq_data *bfqd)
6238 bfq_update_hw_tag(bfqd);
6240 bfqd->rq_in_driver--;
6243 if (!bfqq->dispatched && !bfq_bfqq_busy(bfqq)) {
6245 * Set budget_timeout (which we overload to store the
6246 * time at which the queue remains with no backlog and
6247 * no outstanding request; used by the weight-raising
6250 bfqq->budget_timeout = jiffies;
6252 bfq_weights_tree_remove(bfqd, bfqq);
6255 now_ns = ktime_get_ns();
6257 bfqq->ttime.last_end_request = now_ns;
6260 * Using us instead of ns, to get a reasonable precision in
6261 * computing rate in next check.
6263 delta_us = div_u64(now_ns - bfqd->last_completion, NSEC_PER_USEC);
6266 * If the request took rather long to complete, and, according
6267 * to the maximum request size recorded, this completion latency
6268 * implies that the request was certainly served at a very low
6269 * rate (less than 1M sectors/sec), then the whole observation
6270 * interval that lasts up to this time instant cannot be a
6271 * valid time interval for computing a new peak rate. Invoke
6272 * bfq_update_rate_reset to have the following three steps
6274 * - close the observation interval at the last (previous)
6275 * request dispatch or completion
6276 * - compute rate, if possible, for that observation interval
6277 * - reset to zero samples, which will trigger a proper
6278 * re-initialization of the observation interval on next
6281 if (delta_us > BFQ_MIN_TT/NSEC_PER_USEC &&
6282 (bfqd->last_rq_max_size<<BFQ_RATE_SHIFT)/delta_us <
6283 1UL<<(BFQ_RATE_SHIFT - 10))
6284 bfq_update_rate_reset(bfqd, NULL);
6285 bfqd->last_completion = now_ns;
6287 * Shared queues are likely to receive I/O at a high
6288 * rate. This may deceptively let them be considered as wakers
6289 * of other queues. But a false waker will unjustly steal
6290 * bandwidth to its supposedly woken queue. So considering
6291 * also shared queues in the waking mechanism may cause more
6292 * control troubles than throughput benefits. Then reset
6293 * last_completed_rq_bfqq if bfqq is a shared queue.
6295 if (!bfq_bfqq_coop(bfqq))
6296 bfqd->last_completed_rq_bfqq = bfqq;
6298 bfqd->last_completed_rq_bfqq = NULL;
6301 * If we are waiting to discover whether the request pattern
6302 * of the task associated with the queue is actually
6303 * isochronous, and both requisites for this condition to hold
6304 * are now satisfied, then compute soft_rt_next_start (see the
6305 * comments on the function bfq_bfqq_softrt_next_start()). We
6306 * do not compute soft_rt_next_start if bfqq is in interactive
6307 * weight raising (see the comments in bfq_bfqq_expire() for
6308 * an explanation). We schedule this delayed update when bfqq
6309 * expires, if it still has in-flight requests.
6311 if (bfq_bfqq_softrt_update(bfqq) && bfqq->dispatched == 0 &&
6312 RB_EMPTY_ROOT(&bfqq->sort_list) &&
6313 bfqq->wr_coeff != bfqd->bfq_wr_coeff)
6314 bfqq->soft_rt_next_start =
6315 bfq_bfqq_softrt_next_start(bfqd, bfqq);
6318 * If this is the in-service queue, check if it needs to be expired,
6319 * or if we want to idle in case it has no pending requests.
6321 if (bfqd->in_service_queue == bfqq) {
6322 if (bfq_bfqq_must_idle(bfqq)) {
6323 if (bfqq->dispatched == 0)
6324 bfq_arm_slice_timer(bfqd);
6326 * If we get here, we do not expire bfqq, even
6327 * if bfqq was in budget timeout or had no
6328 * more requests (as controlled in the next
6329 * conditional instructions). The reason for
6330 * not expiring bfqq is as follows.
6332 * Here bfqq->dispatched > 0 holds, but
6333 * bfq_bfqq_must_idle() returned true. This
6334 * implies that, even if no request arrives
6335 * for bfqq before bfqq->dispatched reaches 0,
6336 * bfqq will, however, not be expired on the
6337 * completion event that causes bfqq->dispatch
6338 * to reach zero. In contrast, on this event,
6339 * bfqq will start enjoying device idling
6340 * (I/O-dispatch plugging).
6342 * But, if we expired bfqq here, bfqq would
6343 * not have the chance to enjoy device idling
6344 * when bfqq->dispatched finally reaches
6345 * zero. This would expose bfqq to violation
6346 * of its reserved service guarantees.
6349 } else if (bfq_may_expire_for_budg_timeout(bfqq))
6350 bfq_bfqq_expire(bfqd, bfqq, false,
6351 BFQQE_BUDGET_TIMEOUT);
6352 else if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
6353 (bfqq->dispatched == 0 ||
6354 !bfq_better_to_idle(bfqq)))
6355 bfq_bfqq_expire(bfqd, bfqq, false,
6356 BFQQE_NO_MORE_REQUESTS);
6359 if (!bfqd->rq_in_driver)
6360 bfq_schedule_dispatch(bfqd);
6363 static void bfq_finish_requeue_request_body(struct bfq_queue *bfqq)
6365 bfqq_request_freed(bfqq);
6366 bfq_put_queue(bfqq);
6370 * The processes associated with bfqq may happen to generate their
6371 * cumulative I/O at a lower rate than the rate at which the device
6372 * could serve the same I/O. This is rather probable, e.g., if only
6373 * one process is associated with bfqq and the device is an SSD. It
6374 * results in bfqq becoming often empty while in service. In this
6375 * respect, if BFQ is allowed to switch to another queue when bfqq
6376 * remains empty, then the device goes on being fed with I/O requests,
6377 * and the throughput is not affected. In contrast, if BFQ is not
6378 * allowed to switch to another queue---because bfqq is sync and
6379 * I/O-dispatch needs to be plugged while bfqq is temporarily
6380 * empty---then, during the service of bfqq, there will be frequent
6381 * "service holes", i.e., time intervals during which bfqq gets empty
6382 * and the device can only consume the I/O already queued in its
6383 * hardware queues. During service holes, the device may even get to
6384 * remaining idle. In the end, during the service of bfqq, the device
6385 * is driven at a lower speed than the one it can reach with the kind
6386 * of I/O flowing through bfqq.
6388 * To counter this loss of throughput, BFQ implements a "request
6389 * injection mechanism", which tries to fill the above service holes
6390 * with I/O requests taken from other queues. The hard part in this
6391 * mechanism is finding the right amount of I/O to inject, so as to
6392 * both boost throughput and not break bfqq's bandwidth and latency
6393 * guarantees. In this respect, the mechanism maintains a per-queue
6394 * inject limit, computed as below. While bfqq is empty, the injection
6395 * mechanism dispatches extra I/O requests only until the total number
6396 * of I/O requests in flight---i.e., already dispatched but not yet
6397 * completed---remains lower than this limit.
6399 * A first definition comes in handy to introduce the algorithm by
6400 * which the inject limit is computed. We define as first request for
6401 * bfqq, an I/O request for bfqq that arrives while bfqq is in
6402 * service, and causes bfqq to switch from empty to non-empty. The
6403 * algorithm updates the limit as a function of the effect of
6404 * injection on the service times of only the first requests of
6405 * bfqq. The reason for this restriction is that these are the
6406 * requests whose service time is affected most, because they are the
6407 * first to arrive after injection possibly occurred.
6409 * To evaluate the effect of injection, the algorithm measures the
6410 * "total service time" of first requests. We define as total service
6411 * time of an I/O request, the time that elapses since when the
6412 * request is enqueued into bfqq, to when it is completed. This
6413 * quantity allows the whole effect of injection to be measured. It is
6414 * easy to see why. Suppose that some requests of other queues are
6415 * actually injected while bfqq is empty, and that a new request R
6416 * then arrives for bfqq. If the device does start to serve all or
6417 * part of the injected requests during the service hole, then,
6418 * because of this extra service, it may delay the next invocation of
6419 * the dispatch hook of BFQ. Then, even after R gets eventually
6420 * dispatched, the device may delay the actual service of R if it is
6421 * still busy serving the extra requests, or if it decides to serve,
6422 * before R, some extra request still present in its queues. As a
6423 * conclusion, the cumulative extra delay caused by injection can be
6424 * easily evaluated by just comparing the total service time of first
6425 * requests with and without injection.
6427 * The limit-update algorithm works as follows. On the arrival of a
6428 * first request of bfqq, the algorithm measures the total time of the
6429 * request only if one of the three cases below holds, and, for each
6430 * case, it updates the limit as described below:
6432 * (1) If there is no in-flight request. This gives a baseline for the
6433 * total service time of the requests of bfqq. If the baseline has
6434 * not been computed yet, then, after computing it, the limit is
6435 * set to 1, to start boosting throughput, and to prepare the
6436 * ground for the next case. If the baseline has already been
6437 * computed, then it is updated, in case it results to be lower
6438 * than the previous value.
6440 * (2) If the limit is higher than 0 and there are in-flight
6441 * requests. By comparing the total service time in this case with
6442 * the above baseline, it is possible to know at which extent the
6443 * current value of the limit is inflating the total service
6444 * time. If the inflation is below a certain threshold, then bfqq
6445 * is assumed to be suffering from no perceivable loss of its
6446 * service guarantees, and the limit is even tentatively
6447 * increased. If the inflation is above the threshold, then the
6448 * limit is decreased. Due to the lack of any hysteresis, this
6449 * logic makes the limit oscillate even in steady workload
6450 * conditions. Yet we opted for it, because it is fast in reaching
6451 * the best value for the limit, as a function of the current I/O
6452 * workload. To reduce oscillations, this step is disabled for a
6453 * short time interval after the limit happens to be decreased.
6455 * (3) Periodically, after resetting the limit, to make sure that the
6456 * limit eventually drops in case the workload changes. This is
6457 * needed because, after the limit has gone safely up for a
6458 * certain workload, it is impossible to guess whether the
6459 * baseline total service time may have changed, without measuring
6460 * it again without injection. A more effective version of this
6461 * step might be to just sample the baseline, by interrupting
6462 * injection only once, and then to reset/lower the limit only if
6463 * the total service time with the current limit does happen to be
6466 * More details on each step are provided in the comments on the
6467 * pieces of code that implement these steps: the branch handling the
6468 * transition from empty to non empty in bfq_add_request(), the branch
6469 * handling injection in bfq_select_queue(), and the function
6470 * bfq_choose_bfqq_for_injection(). These comments also explain some
6471 * exceptions, made by the injection mechanism in some special cases.
6473 static void bfq_update_inject_limit(struct bfq_data *bfqd,
6474 struct bfq_queue *bfqq)
6476 u64 tot_time_ns = ktime_get_ns() - bfqd->last_empty_occupied_ns;
6477 unsigned int old_limit = bfqq->inject_limit;
6479 if (bfqq->last_serv_time_ns > 0 && bfqd->rqs_injected) {
6480 u64 threshold = (bfqq->last_serv_time_ns * 3)>>1;
6482 if (tot_time_ns >= threshold && old_limit > 0) {
6483 bfqq->inject_limit--;
6484 bfqq->decrease_time_jif = jiffies;
6485 } else if (tot_time_ns < threshold &&
6486 old_limit <= bfqd->max_rq_in_driver)
6487 bfqq->inject_limit++;
6491 * Either we still have to compute the base value for the
6492 * total service time, and there seem to be the right
6493 * conditions to do it, or we can lower the last base value
6496 * NOTE: (bfqd->rq_in_driver == 1) means that there is no I/O
6497 * request in flight, because this function is in the code
6498 * path that handles the completion of a request of bfqq, and,
6499 * in particular, this function is executed before
6500 * bfqd->rq_in_driver is decremented in such a code path.
6502 if ((bfqq->last_serv_time_ns == 0 && bfqd->rq_in_driver == 1) ||
6503 tot_time_ns < bfqq->last_serv_time_ns) {
6504 if (bfqq->last_serv_time_ns == 0) {
6506 * Now we certainly have a base value: make sure we
6507 * start trying injection.
6509 bfqq->inject_limit = max_t(unsigned int, 1, old_limit);
6511 bfqq->last_serv_time_ns = tot_time_ns;
6512 } else if (!bfqd->rqs_injected && bfqd->rq_in_driver == 1)
6514 * No I/O injected and no request still in service in
6515 * the drive: these are the exact conditions for
6516 * computing the base value of the total service time
6517 * for bfqq. So let's update this value, because it is
6518 * rather variable. For example, it varies if the size
6519 * or the spatial locality of the I/O requests in bfqq
6522 bfqq->last_serv_time_ns = tot_time_ns;
6525 /* update complete, not waiting for any request completion any longer */
6526 bfqd->waited_rq = NULL;
6527 bfqd->rqs_injected = false;
6531 * Handle either a requeue or a finish for rq. The things to do are
6532 * the same in both cases: all references to rq are to be dropped. In
6533 * particular, rq is considered completed from the point of view of
6536 static void bfq_finish_requeue_request(struct request *rq)
6538 struct bfq_queue *bfqq = RQ_BFQQ(rq);
6539 struct bfq_data *bfqd;
6540 unsigned long flags;
6543 * rq either is not associated with any icq, or is an already
6544 * requeued request that has not (yet) been re-inserted into
6547 if (!rq->elv.icq || !bfqq)
6552 if (rq->rq_flags & RQF_STARTED)
6553 bfqg_stats_update_completion(bfqq_group(bfqq),
6555 rq->io_start_time_ns,
6558 spin_lock_irqsave(&bfqd->lock, flags);
6559 if (likely(rq->rq_flags & RQF_STARTED)) {
6560 if (rq == bfqd->waited_rq)
6561 bfq_update_inject_limit(bfqd, bfqq);
6563 bfq_completed_request(bfqq, bfqd);
6565 bfq_finish_requeue_request_body(bfqq);
6566 spin_unlock_irqrestore(&bfqd->lock, flags);
6569 * Reset private fields. In case of a requeue, this allows
6570 * this function to correctly do nothing if it is spuriously
6571 * invoked again on this same request (see the check at the
6572 * beginning of the function). Probably, a better general
6573 * design would be to prevent blk-mq from invoking the requeue
6574 * or finish hooks of an elevator, for a request that is not
6575 * referred by that elevator.
6577 * Resetting the following fields would break the
6578 * request-insertion logic if rq is re-inserted into a bfq
6579 * internal queue, without a re-preparation. Here we assume
6580 * that re-insertions of requeued requests, without
6581 * re-preparation, can happen only for pass_through or at_head
6582 * requests (which are not re-inserted into bfq internal
6585 rq->elv.priv[0] = NULL;
6586 rq->elv.priv[1] = NULL;
6589 static void bfq_finish_request(struct request *rq)
6591 bfq_finish_requeue_request(rq);
6594 put_io_context(rq->elv.icq->ioc);
6600 * Removes the association between the current task and bfqq, assuming
6601 * that bic points to the bfq iocontext of the task.
6602 * Returns NULL if a new bfqq should be allocated, or the old bfqq if this
6603 * was the last process referring to that bfqq.
6605 static struct bfq_queue *
6606 bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq)
6608 bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue");
6610 if (bfqq_process_refs(bfqq) == 1) {
6611 bfqq->pid = current->pid;
6612 bfq_clear_bfqq_coop(bfqq);
6613 bfq_clear_bfqq_split_coop(bfqq);
6617 bic_set_bfqq(bic, NULL, 1);
6619 bfq_put_cooperator(bfqq);
6621 bfq_release_process_ref(bfqq->bfqd, bfqq);
6625 static struct bfq_queue *bfq_get_bfqq_handle_split(struct bfq_data *bfqd,
6626 struct bfq_io_cq *bic,
6628 bool split, bool is_sync,
6631 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
6633 if (likely(bfqq && bfqq != &bfqd->oom_bfqq))
6640 bfq_put_queue(bfqq);
6641 bfqq = bfq_get_queue(bfqd, bio, is_sync, bic, split);
6643 bic_set_bfqq(bic, bfqq, is_sync);
6644 if (split && is_sync) {
6645 if ((bic->was_in_burst_list && bfqd->large_burst) ||
6646 bic->saved_in_large_burst)
6647 bfq_mark_bfqq_in_large_burst(bfqq);
6649 bfq_clear_bfqq_in_large_burst(bfqq);
6650 if (bic->was_in_burst_list)
6652 * If bfqq was in the current
6653 * burst list before being
6654 * merged, then we have to add
6655 * it back. And we do not need
6656 * to increase burst_size, as
6657 * we did not decrement
6658 * burst_size when we removed
6659 * bfqq from the burst list as
6660 * a consequence of a merge
6662 * bfq_put_queue). In this
6663 * respect, it would be rather
6664 * costly to know whether the
6665 * current burst list is still
6666 * the same burst list from
6667 * which bfqq was removed on
6668 * the merge. To avoid this
6669 * cost, if bfqq was in a
6670 * burst list, then we add
6671 * bfqq to the current burst
6672 * list without any further
6673 * check. This can cause
6674 * inappropriate insertions,
6675 * but rarely enough to not
6676 * harm the detection of large
6677 * bursts significantly.
6679 hlist_add_head(&bfqq->burst_list_node,
6682 bfqq->split_time = jiffies;
6689 * Only reset private fields. The actual request preparation will be
6690 * performed by bfq_init_rq, when rq is either inserted or merged. See
6691 * comments on bfq_init_rq for the reason behind this delayed
6694 static void bfq_prepare_request(struct request *rq)
6696 rq->elv.icq = ioc_find_get_icq(rq->q);
6699 * Regardless of whether we have an icq attached, we have to
6700 * clear the scheduler pointers, as they might point to
6701 * previously allocated bic/bfqq structs.
6703 rq->elv.priv[0] = rq->elv.priv[1] = NULL;
6707 * If needed, init rq, allocate bfq data structures associated with
6708 * rq, and increment reference counters in the destination bfq_queue
6709 * for rq. Return the destination bfq_queue for rq, or NULL is rq is
6710 * not associated with any bfq_queue.
6712 * This function is invoked by the functions that perform rq insertion
6713 * or merging. One may have expected the above preparation operations
6714 * to be performed in bfq_prepare_request, and not delayed to when rq
6715 * is inserted or merged. The rationale behind this delayed
6716 * preparation is that, after the prepare_request hook is invoked for
6717 * rq, rq may still be transformed into a request with no icq, i.e., a
6718 * request not associated with any queue. No bfq hook is invoked to
6719 * signal this transformation. As a consequence, should these
6720 * preparation operations be performed when the prepare_request hook
6721 * is invoked, and should rq be transformed one moment later, bfq
6722 * would end up in an inconsistent state, because it would have
6723 * incremented some queue counters for an rq destined to
6724 * transformation, without any chance to correctly lower these
6725 * counters back. In contrast, no transformation can still happen for
6726 * rq after rq has been inserted or merged. So, it is safe to execute
6727 * these preparation operations when rq is finally inserted or merged.
6729 static struct bfq_queue *bfq_init_rq(struct request *rq)
6731 struct request_queue *q = rq->q;
6732 struct bio *bio = rq->bio;
6733 struct bfq_data *bfqd = q->elevator->elevator_data;
6734 struct bfq_io_cq *bic;
6735 const int is_sync = rq_is_sync(rq);
6736 struct bfq_queue *bfqq;
6737 bool new_queue = false;
6738 bool bfqq_already_existing = false, split = false;
6740 if (unlikely(!rq->elv.icq))
6744 * Assuming that elv.priv[1] is set only if everything is set
6745 * for this rq. This holds true, because this function is
6746 * invoked only for insertion or merging, and, after such
6747 * events, a request cannot be manipulated any longer before
6748 * being removed from bfq.
6750 if (rq->elv.priv[1])
6751 return rq->elv.priv[1];
6753 bic = icq_to_bic(rq->elv.icq);
6755 bfq_check_ioprio_change(bic, bio);
6757 bfq_bic_update_cgroup(bic, bio);
6759 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio, false, is_sync,
6762 if (likely(!new_queue)) {
6763 /* If the queue was seeky for too long, break it apart. */
6764 if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq) &&
6765 !bic->stably_merged) {
6766 struct bfq_queue *old_bfqq = bfqq;
6768 /* Update bic before losing reference to bfqq */
6769 if (bfq_bfqq_in_large_burst(bfqq))
6770 bic->saved_in_large_burst = true;
6772 bfqq = bfq_split_bfqq(bic, bfqq);
6776 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio,
6779 bfqq->waker_bfqq = old_bfqq->waker_bfqq;
6780 bfqq->tentative_waker_bfqq = NULL;
6783 * If the waker queue disappears, then
6784 * new_bfqq->waker_bfqq must be
6785 * reset. So insert new_bfqq into the
6786 * woken_list of the waker. See
6787 * bfq_check_waker for details.
6789 if (bfqq->waker_bfqq)
6790 hlist_add_head(&bfqq->woken_list_node,
6791 &bfqq->waker_bfqq->woken_list);
6793 bfqq_already_existing = true;
6797 bfqq_request_allocated(bfqq);
6799 bfq_log_bfqq(bfqd, bfqq, "get_request %p: bfqq %p, %d",
6800 rq, bfqq, bfqq->ref);
6802 rq->elv.priv[0] = bic;
6803 rq->elv.priv[1] = bfqq;
6806 * If a bfq_queue has only one process reference, it is owned
6807 * by only this bic: we can then set bfqq->bic = bic. in
6808 * addition, if the queue has also just been split, we have to
6811 if (likely(bfqq != &bfqd->oom_bfqq) && bfqq_process_refs(bfqq) == 1) {
6815 * The queue has just been split from a shared
6816 * queue: restore the idle window and the
6817 * possible weight raising period.
6819 bfq_bfqq_resume_state(bfqq, bfqd, bic,
6820 bfqq_already_existing);
6825 * Consider bfqq as possibly belonging to a burst of newly
6826 * created queues only if:
6827 * 1) A burst is actually happening (bfqd->burst_size > 0)
6829 * 2) There is no other active queue. In fact, if, in
6830 * contrast, there are active queues not belonging to the
6831 * possible burst bfqq may belong to, then there is no gain
6832 * in considering bfqq as belonging to a burst, and
6833 * therefore in not weight-raising bfqq. See comments on
6834 * bfq_handle_burst().
6836 * This filtering also helps eliminating false positives,
6837 * occurring when bfqq does not belong to an actual large
6838 * burst, but some background task (e.g., a service) happens
6839 * to trigger the creation of new queues very close to when
6840 * bfqq and its possible companion queues are created. See
6841 * comments on bfq_handle_burst() for further details also on
6844 if (unlikely(bfq_bfqq_just_created(bfqq) &&
6845 (bfqd->burst_size > 0 ||
6846 bfq_tot_busy_queues(bfqd) == 0)))
6847 bfq_handle_burst(bfqd, bfqq);
6853 bfq_idle_slice_timer_body(struct bfq_data *bfqd, struct bfq_queue *bfqq)
6855 enum bfqq_expiration reason;
6856 unsigned long flags;
6858 spin_lock_irqsave(&bfqd->lock, flags);
6861 * Considering that bfqq may be in race, we should firstly check
6862 * whether bfqq is in service before doing something on it. If
6863 * the bfqq in race is not in service, it has already been expired
6864 * through __bfq_bfqq_expire func and its wait_request flags has
6865 * been cleared in __bfq_bfqd_reset_in_service func.
6867 if (bfqq != bfqd->in_service_queue) {
6868 spin_unlock_irqrestore(&bfqd->lock, flags);
6872 bfq_clear_bfqq_wait_request(bfqq);
6874 if (bfq_bfqq_budget_timeout(bfqq))
6876 * Also here the queue can be safely expired
6877 * for budget timeout without wasting
6880 reason = BFQQE_BUDGET_TIMEOUT;
6881 else if (bfqq->queued[0] == 0 && bfqq->queued[1] == 0)
6883 * The queue may not be empty upon timer expiration,
6884 * because we may not disable the timer when the
6885 * first request of the in-service queue arrives
6886 * during disk idling.
6888 reason = BFQQE_TOO_IDLE;
6890 goto schedule_dispatch;
6892 bfq_bfqq_expire(bfqd, bfqq, true, reason);
6895 spin_unlock_irqrestore(&bfqd->lock, flags);
6896 bfq_schedule_dispatch(bfqd);
6900 * Handler of the expiration of the timer running if the in-service queue
6901 * is idling inside its time slice.
6903 static enum hrtimer_restart bfq_idle_slice_timer(struct hrtimer *timer)
6905 struct bfq_data *bfqd = container_of(timer, struct bfq_data,
6907 struct bfq_queue *bfqq = bfqd->in_service_queue;
6910 * Theoretical race here: the in-service queue can be NULL or
6911 * different from the queue that was idling if a new request
6912 * arrives for the current queue and there is a full dispatch
6913 * cycle that changes the in-service queue. This can hardly
6914 * happen, but in the worst case we just expire a queue too
6918 bfq_idle_slice_timer_body(bfqd, bfqq);
6920 return HRTIMER_NORESTART;
6923 static void __bfq_put_async_bfqq(struct bfq_data *bfqd,
6924 struct bfq_queue **bfqq_ptr)
6926 struct bfq_queue *bfqq = *bfqq_ptr;
6928 bfq_log(bfqd, "put_async_bfqq: %p", bfqq);
6930 bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
6932 bfq_log_bfqq(bfqd, bfqq, "put_async_bfqq: putting %p, %d",
6934 bfq_put_queue(bfqq);
6940 * Release all the bfqg references to its async queues. If we are
6941 * deallocating the group these queues may still contain requests, so
6942 * we reparent them to the root cgroup (i.e., the only one that will
6943 * exist for sure until all the requests on a device are gone).
6945 void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg)
6949 for (i = 0; i < 2; i++)
6950 for (j = 0; j < IOPRIO_NR_LEVELS; j++)
6951 __bfq_put_async_bfqq(bfqd, &bfqg->async_bfqq[i][j]);
6953 __bfq_put_async_bfqq(bfqd, &bfqg->async_idle_bfqq);
6957 * See the comments on bfq_limit_depth for the purpose of
6958 * the depths set in the function. Return minimum shallow depth we'll use.
6960 static void bfq_update_depths(struct bfq_data *bfqd, struct sbitmap_queue *bt)
6962 unsigned int depth = 1U << bt->sb.shift;
6964 bfqd->full_depth_shift = bt->sb.shift;
6966 * In-word depths if no bfq_queue is being weight-raised:
6967 * leaving 25% of tags only for sync reads.
6969 * In next formulas, right-shift the value
6970 * (1U<<bt->sb.shift), instead of computing directly
6971 * (1U<<(bt->sb.shift - something)), to be robust against
6972 * any possible value of bt->sb.shift, without having to
6973 * limit 'something'.
6975 /* no more than 50% of tags for async I/O */
6976 bfqd->word_depths[0][0] = max(depth >> 1, 1U);
6978 * no more than 75% of tags for sync writes (25% extra tags
6979 * w.r.t. async I/O, to prevent async I/O from starving sync
6982 bfqd->word_depths[0][1] = max((depth * 3) >> 2, 1U);
6985 * In-word depths in case some bfq_queue is being weight-
6986 * raised: leaving ~63% of tags for sync reads. This is the
6987 * highest percentage for which, in our tests, application
6988 * start-up times didn't suffer from any regression due to tag
6991 /* no more than ~18% of tags for async I/O */
6992 bfqd->word_depths[1][0] = max((depth * 3) >> 4, 1U);
6993 /* no more than ~37% of tags for sync writes (~20% extra tags) */
6994 bfqd->word_depths[1][1] = max((depth * 6) >> 4, 1U);
6997 static void bfq_depth_updated(struct blk_mq_hw_ctx *hctx)
6999 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
7000 struct blk_mq_tags *tags = hctx->sched_tags;
7002 bfq_update_depths(bfqd, &tags->bitmap_tags);
7003 sbitmap_queue_min_shallow_depth(&tags->bitmap_tags, 1);
7006 static int bfq_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int index)
7008 bfq_depth_updated(hctx);
7012 static void bfq_exit_queue(struct elevator_queue *e)
7014 struct bfq_data *bfqd = e->elevator_data;
7015 struct bfq_queue *bfqq, *n;
7017 hrtimer_cancel(&bfqd->idle_slice_timer);
7019 spin_lock_irq(&bfqd->lock);
7020 list_for_each_entry_safe(bfqq, n, &bfqd->idle_list, bfqq_list)
7021 bfq_deactivate_bfqq(bfqd, bfqq, false, false);
7022 spin_unlock_irq(&bfqd->lock);
7024 hrtimer_cancel(&bfqd->idle_slice_timer);
7026 /* release oom-queue reference to root group */
7027 bfqg_and_blkg_put(bfqd->root_group);
7029 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7030 blkcg_deactivate_policy(bfqd->queue, &blkcg_policy_bfq);
7032 spin_lock_irq(&bfqd->lock);
7033 bfq_put_async_queues(bfqd, bfqd->root_group);
7034 kfree(bfqd->root_group);
7035 spin_unlock_irq(&bfqd->lock);
7038 wbt_enable_default(bfqd->queue);
7043 static void bfq_init_root_group(struct bfq_group *root_group,
7044 struct bfq_data *bfqd)
7048 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7049 root_group->entity.parent = NULL;
7050 root_group->my_entity = NULL;
7051 root_group->bfqd = bfqd;
7053 root_group->rq_pos_tree = RB_ROOT;
7054 for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
7055 root_group->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
7056 root_group->sched_data.bfq_class_idle_last_service = jiffies;
7059 static int bfq_init_queue(struct request_queue *q, struct elevator_type *e)
7061 struct bfq_data *bfqd;
7062 struct elevator_queue *eq;
7064 eq = elevator_alloc(q, e);
7068 bfqd = kzalloc_node(sizeof(*bfqd), GFP_KERNEL, q->node);
7070 kobject_put(&eq->kobj);
7073 eq->elevator_data = bfqd;
7075 spin_lock_irq(&q->queue_lock);
7077 spin_unlock_irq(&q->queue_lock);
7080 * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues.
7081 * Grab a permanent reference to it, so that the normal code flow
7082 * will not attempt to free it.
7084 bfq_init_bfqq(bfqd, &bfqd->oom_bfqq, NULL, 1, 0);
7085 bfqd->oom_bfqq.ref++;
7086 bfqd->oom_bfqq.new_ioprio = BFQ_DEFAULT_QUEUE_IOPRIO;
7087 bfqd->oom_bfqq.new_ioprio_class = IOPRIO_CLASS_BE;
7088 bfqd->oom_bfqq.entity.new_weight =
7089 bfq_ioprio_to_weight(bfqd->oom_bfqq.new_ioprio);
7091 /* oom_bfqq does not participate to bursts */
7092 bfq_clear_bfqq_just_created(&bfqd->oom_bfqq);
7095 * Trigger weight initialization, according to ioprio, at the
7096 * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio
7097 * class won't be changed any more.
7099 bfqd->oom_bfqq.entity.prio_changed = 1;
7103 INIT_LIST_HEAD(&bfqd->dispatch);
7105 hrtimer_init(&bfqd->idle_slice_timer, CLOCK_MONOTONIC,
7107 bfqd->idle_slice_timer.function = bfq_idle_slice_timer;
7109 bfqd->queue_weights_tree = RB_ROOT_CACHED;
7110 bfqd->num_groups_with_pending_reqs = 0;
7112 INIT_LIST_HEAD(&bfqd->active_list);
7113 INIT_LIST_HEAD(&bfqd->idle_list);
7114 INIT_HLIST_HEAD(&bfqd->burst_list);
7117 bfqd->nonrot_with_queueing = blk_queue_nonrot(bfqd->queue);
7119 bfqd->bfq_max_budget = bfq_default_max_budget;
7121 bfqd->bfq_fifo_expire[0] = bfq_fifo_expire[0];
7122 bfqd->bfq_fifo_expire[1] = bfq_fifo_expire[1];
7123 bfqd->bfq_back_max = bfq_back_max;
7124 bfqd->bfq_back_penalty = bfq_back_penalty;
7125 bfqd->bfq_slice_idle = bfq_slice_idle;
7126 bfqd->bfq_timeout = bfq_timeout;
7128 bfqd->bfq_large_burst_thresh = 8;
7129 bfqd->bfq_burst_interval = msecs_to_jiffies(180);
7131 bfqd->low_latency = true;
7134 * Trade-off between responsiveness and fairness.
7136 bfqd->bfq_wr_coeff = 30;
7137 bfqd->bfq_wr_rt_max_time = msecs_to_jiffies(300);
7138 bfqd->bfq_wr_max_time = 0;
7139 bfqd->bfq_wr_min_idle_time = msecs_to_jiffies(2000);
7140 bfqd->bfq_wr_min_inter_arr_async = msecs_to_jiffies(500);
7141 bfqd->bfq_wr_max_softrt_rate = 7000; /*
7142 * Approximate rate required
7143 * to playback or record a
7144 * high-definition compressed
7147 bfqd->wr_busy_queues = 0;
7150 * Begin by assuming, optimistically, that the device peak
7151 * rate is equal to 2/3 of the highest reference rate.
7153 bfqd->rate_dur_prod = ref_rate[blk_queue_nonrot(bfqd->queue)] *
7154 ref_wr_duration[blk_queue_nonrot(bfqd->queue)];
7155 bfqd->peak_rate = ref_rate[blk_queue_nonrot(bfqd->queue)] * 2 / 3;
7157 spin_lock_init(&bfqd->lock);
7160 * The invocation of the next bfq_create_group_hierarchy
7161 * function is the head of a chain of function calls
7162 * (bfq_create_group_hierarchy->blkcg_activate_policy->
7163 * blk_mq_freeze_queue) that may lead to the invocation of the
7164 * has_work hook function. For this reason,
7165 * bfq_create_group_hierarchy is invoked only after all
7166 * scheduler data has been initialized, apart from the fields
7167 * that can be initialized only after invoking
7168 * bfq_create_group_hierarchy. This, in particular, enables
7169 * has_work to correctly return false. Of course, to avoid
7170 * other inconsistencies, the blk-mq stack must then refrain
7171 * from invoking further scheduler hooks before this init
7172 * function is finished.
7174 bfqd->root_group = bfq_create_group_hierarchy(bfqd, q->node);
7175 if (!bfqd->root_group)
7177 bfq_init_root_group(bfqd->root_group, bfqd);
7178 bfq_init_entity(&bfqd->oom_bfqq.entity, bfqd->root_group);
7180 wbt_disable_default(q);
7185 kobject_put(&eq->kobj);
7189 static void bfq_slab_kill(void)
7191 kmem_cache_destroy(bfq_pool);
7194 static int __init bfq_slab_setup(void)
7196 bfq_pool = KMEM_CACHE(bfq_queue, 0);
7202 static ssize_t bfq_var_show(unsigned int var, char *page)
7204 return sprintf(page, "%u\n", var);
7207 static int bfq_var_store(unsigned long *var, const char *page)
7209 unsigned long new_val;
7210 int ret = kstrtoul(page, 10, &new_val);
7218 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
7219 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
7221 struct bfq_data *bfqd = e->elevator_data; \
7222 u64 __data = __VAR; \
7224 __data = jiffies_to_msecs(__data); \
7225 else if (__CONV == 2) \
7226 __data = div_u64(__data, NSEC_PER_MSEC); \
7227 return bfq_var_show(__data, (page)); \
7229 SHOW_FUNCTION(bfq_fifo_expire_sync_show, bfqd->bfq_fifo_expire[1], 2);
7230 SHOW_FUNCTION(bfq_fifo_expire_async_show, bfqd->bfq_fifo_expire[0], 2);
7231 SHOW_FUNCTION(bfq_back_seek_max_show, bfqd->bfq_back_max, 0);
7232 SHOW_FUNCTION(bfq_back_seek_penalty_show, bfqd->bfq_back_penalty, 0);
7233 SHOW_FUNCTION(bfq_slice_idle_show, bfqd->bfq_slice_idle, 2);
7234 SHOW_FUNCTION(bfq_max_budget_show, bfqd->bfq_user_max_budget, 0);
7235 SHOW_FUNCTION(bfq_timeout_sync_show, bfqd->bfq_timeout, 1);
7236 SHOW_FUNCTION(bfq_strict_guarantees_show, bfqd->strict_guarantees, 0);
7237 SHOW_FUNCTION(bfq_low_latency_show, bfqd->low_latency, 0);
7238 #undef SHOW_FUNCTION
7240 #define USEC_SHOW_FUNCTION(__FUNC, __VAR) \
7241 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
7243 struct bfq_data *bfqd = e->elevator_data; \
7244 u64 __data = __VAR; \
7245 __data = div_u64(__data, NSEC_PER_USEC); \
7246 return bfq_var_show(__data, (page)); \
7248 USEC_SHOW_FUNCTION(bfq_slice_idle_us_show, bfqd->bfq_slice_idle);
7249 #undef USEC_SHOW_FUNCTION
7251 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
7253 __FUNC(struct elevator_queue *e, const char *page, size_t count) \
7255 struct bfq_data *bfqd = e->elevator_data; \
7256 unsigned long __data, __min = (MIN), __max = (MAX); \
7259 ret = bfq_var_store(&__data, (page)); \
7262 if (__data < __min) \
7264 else if (__data > __max) \
7267 *(__PTR) = msecs_to_jiffies(__data); \
7268 else if (__CONV == 2) \
7269 *(__PTR) = (u64)__data * NSEC_PER_MSEC; \
7271 *(__PTR) = __data; \
7274 STORE_FUNCTION(bfq_fifo_expire_sync_store, &bfqd->bfq_fifo_expire[1], 1,
7276 STORE_FUNCTION(bfq_fifo_expire_async_store, &bfqd->bfq_fifo_expire[0], 1,
7278 STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0);
7279 STORE_FUNCTION(bfq_back_seek_penalty_store, &bfqd->bfq_back_penalty, 1,
7281 STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 2);
7282 #undef STORE_FUNCTION
7284 #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
7285 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\
7287 struct bfq_data *bfqd = e->elevator_data; \
7288 unsigned long __data, __min = (MIN), __max = (MAX); \
7291 ret = bfq_var_store(&__data, (page)); \
7294 if (__data < __min) \
7296 else if (__data > __max) \
7298 *(__PTR) = (u64)__data * NSEC_PER_USEC; \
7301 USEC_STORE_FUNCTION(bfq_slice_idle_us_store, &bfqd->bfq_slice_idle, 0,
7303 #undef USEC_STORE_FUNCTION
7305 static ssize_t bfq_max_budget_store(struct elevator_queue *e,
7306 const char *page, size_t count)
7308 struct bfq_data *bfqd = e->elevator_data;
7309 unsigned long __data;
7312 ret = bfq_var_store(&__data, (page));
7317 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
7319 if (__data > INT_MAX)
7321 bfqd->bfq_max_budget = __data;
7324 bfqd->bfq_user_max_budget = __data;
7330 * Leaving this name to preserve name compatibility with cfq
7331 * parameters, but this timeout is used for both sync and async.
7333 static ssize_t bfq_timeout_sync_store(struct elevator_queue *e,
7334 const char *page, size_t count)
7336 struct bfq_data *bfqd = e->elevator_data;
7337 unsigned long __data;
7340 ret = bfq_var_store(&__data, (page));
7346 else if (__data > INT_MAX)
7349 bfqd->bfq_timeout = msecs_to_jiffies(__data);
7350 if (bfqd->bfq_user_max_budget == 0)
7351 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
7356 static ssize_t bfq_strict_guarantees_store(struct elevator_queue *e,
7357 const char *page, size_t count)
7359 struct bfq_data *bfqd = e->elevator_data;
7360 unsigned long __data;
7363 ret = bfq_var_store(&__data, (page));
7369 if (!bfqd->strict_guarantees && __data == 1
7370 && bfqd->bfq_slice_idle < 8 * NSEC_PER_MSEC)
7371 bfqd->bfq_slice_idle = 8 * NSEC_PER_MSEC;
7373 bfqd->strict_guarantees = __data;
7378 static ssize_t bfq_low_latency_store(struct elevator_queue *e,
7379 const char *page, size_t count)
7381 struct bfq_data *bfqd = e->elevator_data;
7382 unsigned long __data;
7385 ret = bfq_var_store(&__data, (page));
7391 if (__data == 0 && bfqd->low_latency != 0)
7393 bfqd->low_latency = __data;
7398 #define BFQ_ATTR(name) \
7399 __ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store)
7401 static struct elv_fs_entry bfq_attrs[] = {
7402 BFQ_ATTR(fifo_expire_sync),
7403 BFQ_ATTR(fifo_expire_async),
7404 BFQ_ATTR(back_seek_max),
7405 BFQ_ATTR(back_seek_penalty),
7406 BFQ_ATTR(slice_idle),
7407 BFQ_ATTR(slice_idle_us),
7408 BFQ_ATTR(max_budget),
7409 BFQ_ATTR(timeout_sync),
7410 BFQ_ATTR(strict_guarantees),
7411 BFQ_ATTR(low_latency),
7415 static struct elevator_type iosched_bfq_mq = {
7417 .limit_depth = bfq_limit_depth,
7418 .prepare_request = bfq_prepare_request,
7419 .requeue_request = bfq_finish_requeue_request,
7420 .finish_request = bfq_finish_request,
7421 .exit_icq = bfq_exit_icq,
7422 .insert_requests = bfq_insert_requests,
7423 .dispatch_request = bfq_dispatch_request,
7424 .next_request = elv_rb_latter_request,
7425 .former_request = elv_rb_former_request,
7426 .allow_merge = bfq_allow_bio_merge,
7427 .bio_merge = bfq_bio_merge,
7428 .request_merge = bfq_request_merge,
7429 .requests_merged = bfq_requests_merged,
7430 .request_merged = bfq_request_merged,
7431 .has_work = bfq_has_work,
7432 .depth_updated = bfq_depth_updated,
7433 .init_hctx = bfq_init_hctx,
7434 .init_sched = bfq_init_queue,
7435 .exit_sched = bfq_exit_queue,
7438 .icq_size = sizeof(struct bfq_io_cq),
7439 .icq_align = __alignof__(struct bfq_io_cq),
7440 .elevator_attrs = bfq_attrs,
7441 .elevator_name = "bfq",
7442 .elevator_owner = THIS_MODULE,
7444 MODULE_ALIAS("bfq-iosched");
7446 static int __init bfq_init(void)
7450 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7451 ret = blkcg_policy_register(&blkcg_policy_bfq);
7457 if (bfq_slab_setup())
7461 * Times to load large popular applications for the typical
7462 * systems installed on the reference devices (see the
7463 * comments before the definition of the next
7464 * array). Actually, we use slightly lower values, as the
7465 * estimated peak rate tends to be smaller than the actual
7466 * peak rate. The reason for this last fact is that estimates
7467 * are computed over much shorter time intervals than the long
7468 * intervals typically used for benchmarking. Why? First, to
7469 * adapt more quickly to variations. Second, because an I/O
7470 * scheduler cannot rely on a peak-rate-evaluation workload to
7471 * be run for a long time.
7473 ref_wr_duration[0] = msecs_to_jiffies(7000); /* actually 8 sec */
7474 ref_wr_duration[1] = msecs_to_jiffies(2500); /* actually 3 sec */
7476 ret = elv_register(&iosched_bfq_mq);
7485 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7486 blkcg_policy_unregister(&blkcg_policy_bfq);
7491 static void __exit bfq_exit(void)
7493 elv_unregister(&iosched_bfq_mq);
7494 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7495 blkcg_policy_unregister(&blkcg_policy_bfq);
7500 module_init(bfq_init);
7501 module_exit(bfq_exit);
7503 MODULE_AUTHOR("Paolo Valente");
7504 MODULE_LICENSE("GPL");
7505 MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler");