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) ((struct bfq_io_cq *)((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 lockdep_assert_held(&bfqd->lock);
461 if (bfqd->queued != 0) {
462 bfq_log(bfqd, "schedule dispatch");
463 blk_mq_run_hw_queues(bfqd->queue, true);
467 #define bfq_class_idle(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
469 #define bfq_sample_valid(samples) ((samples) > 80)
472 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
473 * We choose the request that is closer to the head right now. Distance
474 * behind the head is penalized and only allowed to a certain extent.
476 static struct request *bfq_choose_req(struct bfq_data *bfqd,
481 sector_t s1, s2, d1 = 0, d2 = 0;
482 unsigned long back_max;
483 #define BFQ_RQ1_WRAP 0x01 /* request 1 wraps */
484 #define BFQ_RQ2_WRAP 0x02 /* request 2 wraps */
485 unsigned int wrap = 0; /* bit mask: requests behind the disk head? */
487 if (!rq1 || rq1 == rq2)
492 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
494 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
496 if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
498 else if ((rq2->cmd_flags & REQ_META) && !(rq1->cmd_flags & REQ_META))
501 s1 = blk_rq_pos(rq1);
502 s2 = blk_rq_pos(rq2);
505 * By definition, 1KiB is 2 sectors.
507 back_max = bfqd->bfq_back_max * 2;
510 * Strict one way elevator _except_ in the case where we allow
511 * short backward seeks which are biased as twice the cost of a
512 * similar forward seek.
516 else if (s1 + back_max >= last)
517 d1 = (last - s1) * bfqd->bfq_back_penalty;
519 wrap |= BFQ_RQ1_WRAP;
523 else if (s2 + back_max >= last)
524 d2 = (last - s2) * bfqd->bfq_back_penalty;
526 wrap |= BFQ_RQ2_WRAP;
528 /* Found required data */
531 * By doing switch() on the bit mask "wrap" we avoid having to
532 * check two variables for all permutations: --> faster!
535 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
550 case BFQ_RQ1_WRAP|BFQ_RQ2_WRAP: /* both rqs wrapped */
553 * Since both rqs are wrapped,
554 * start with the one that's further behind head
555 * (--> only *one* back seek required),
556 * since back seek takes more time than forward.
565 #define BFQ_LIMIT_INLINE_DEPTH 16
567 #ifdef CONFIG_BFQ_GROUP_IOSCHED
568 static bool bfqq_request_over_limit(struct bfq_queue *bfqq, int limit)
570 struct bfq_data *bfqd = bfqq->bfqd;
571 struct bfq_entity *entity = &bfqq->entity;
572 struct bfq_entity *inline_entities[BFQ_LIMIT_INLINE_DEPTH];
573 struct bfq_entity **entities = inline_entities;
574 int depth, level, alloc_depth = BFQ_LIMIT_INLINE_DEPTH;
575 int class_idx = bfqq->ioprio_class - 1;
576 struct bfq_sched_data *sched_data;
580 if (!entity->on_st_or_in_serv)
584 spin_lock_irq(&bfqd->lock);
585 /* +1 for bfqq entity, root cgroup not included */
586 depth = bfqg_to_blkg(bfqq_group(bfqq))->blkcg->css.cgroup->level + 1;
587 if (depth > alloc_depth) {
588 spin_unlock_irq(&bfqd->lock);
589 if (entities != inline_entities)
591 entities = kmalloc_array(depth, sizeof(*entities), GFP_NOIO);
598 sched_data = entity->sched_data;
599 /* Gather our ancestors as we need to traverse them in reverse order */
601 for_each_entity(entity) {
603 * If at some level entity is not even active, allow request
604 * queueing so that BFQ knows there's work to do and activate
607 if (!entity->on_st_or_in_serv)
609 /* Uh, more parents than cgroup subsystem thinks? */
610 if (WARN_ON_ONCE(level >= depth))
612 entities[level++] = entity;
614 WARN_ON_ONCE(level != depth);
615 for (level--; level >= 0; level--) {
616 entity = entities[level];
618 wsum = bfq_entity_service_tree(entity)->wsum;
622 * For bfqq itself we take into account service trees
623 * of all higher priority classes and multiply their
624 * weights so that low prio queue from higher class
625 * gets more requests than high prio queue from lower
629 for (i = 0; i <= class_idx; i++) {
630 wsum = wsum * IOPRIO_BE_NR +
631 sched_data->service_tree[i].wsum;
634 limit = DIV_ROUND_CLOSEST(limit * entity->weight, wsum);
635 if (entity->allocated >= limit) {
636 bfq_log_bfqq(bfqq->bfqd, bfqq,
637 "too many requests: allocated %d limit %d level %d",
638 entity->allocated, limit, level);
644 spin_unlock_irq(&bfqd->lock);
645 if (entities != inline_entities)
650 static bool bfqq_request_over_limit(struct bfq_queue *bfqq, int limit)
657 * Async I/O can easily starve sync I/O (both sync reads and sync
658 * writes), by consuming all tags. Similarly, storms of sync writes,
659 * such as those that sync(2) may trigger, can starve sync reads.
660 * Limit depths of async I/O and sync writes so as to counter both
663 * Also if a bfq queue or its parent cgroup consume more tags than would be
664 * appropriate for their weight, we trim the available tag depth to 1. This
665 * avoids a situation where one cgroup can starve another cgroup from tags and
666 * thus block service differentiation among cgroups. Note that because the
667 * queue / cgroup already has many requests allocated and queued, this does not
668 * significantly affect service guarantees coming from the BFQ scheduling
671 static void bfq_limit_depth(blk_opf_t opf, struct blk_mq_alloc_data *data)
673 struct bfq_data *bfqd = data->q->elevator->elevator_data;
674 struct bfq_io_cq *bic = bfq_bic_lookup(data->q);
675 struct bfq_queue *bfqq = bic ? bic_to_bfqq(bic, op_is_sync(opf)) : NULL;
677 unsigned limit = data->q->nr_requests;
679 /* Sync reads have full depth available */
680 if (op_is_sync(opf) && !op_is_write(opf)) {
683 depth = bfqd->word_depths[!!bfqd->wr_busy_queues][op_is_sync(opf)];
684 limit = (limit * depth) >> bfqd->full_depth_shift;
688 * Does queue (or any parent entity) exceed number of requests that
689 * should be available to it? Heavily limit depth so that it cannot
690 * consume more available requests and thus starve other entities.
692 if (bfqq && bfqq_request_over_limit(bfqq, limit))
695 bfq_log(bfqd, "[%s] wr_busy %d sync %d depth %u",
696 __func__, bfqd->wr_busy_queues, op_is_sync(opf), depth);
698 data->shallow_depth = depth;
701 static struct bfq_queue *
702 bfq_rq_pos_tree_lookup(struct bfq_data *bfqd, struct rb_root *root,
703 sector_t sector, struct rb_node **ret_parent,
704 struct rb_node ***rb_link)
706 struct rb_node **p, *parent;
707 struct bfq_queue *bfqq = NULL;
715 bfqq = rb_entry(parent, struct bfq_queue, pos_node);
718 * Sort strictly based on sector. Smallest to the left,
719 * largest to the right.
721 if (sector > blk_rq_pos(bfqq->next_rq))
723 else if (sector < blk_rq_pos(bfqq->next_rq))
731 *ret_parent = parent;
735 bfq_log(bfqd, "rq_pos_tree_lookup %llu: returning %d",
736 (unsigned long long)sector,
737 bfqq ? bfqq->pid : 0);
742 static bool bfq_too_late_for_merging(struct bfq_queue *bfqq)
744 return bfqq->service_from_backlogged > 0 &&
745 time_is_before_jiffies(bfqq->first_IO_time +
746 bfq_merge_time_limit);
750 * The following function is not marked as __cold because it is
751 * actually cold, but for the same performance goal described in the
752 * comments on the likely() at the beginning of
753 * bfq_setup_cooperator(). Unexpectedly, to reach an even lower
754 * execution time for the case where this function is not invoked, we
755 * had to add an unlikely() in each involved if().
758 bfq_pos_tree_add_move(struct bfq_data *bfqd, struct bfq_queue *bfqq)
760 struct rb_node **p, *parent;
761 struct bfq_queue *__bfqq;
763 if (bfqq->pos_root) {
764 rb_erase(&bfqq->pos_node, bfqq->pos_root);
765 bfqq->pos_root = NULL;
768 /* oom_bfqq does not participate in queue merging */
769 if (bfqq == &bfqd->oom_bfqq)
773 * bfqq cannot be merged any longer (see comments in
774 * bfq_setup_cooperator): no point in adding bfqq into the
777 if (bfq_too_late_for_merging(bfqq))
780 if (bfq_class_idle(bfqq))
785 bfqq->pos_root = &bfqq_group(bfqq)->rq_pos_tree;
786 __bfqq = bfq_rq_pos_tree_lookup(bfqd, bfqq->pos_root,
787 blk_rq_pos(bfqq->next_rq), &parent, &p);
789 rb_link_node(&bfqq->pos_node, parent, p);
790 rb_insert_color(&bfqq->pos_node, bfqq->pos_root);
792 bfqq->pos_root = NULL;
796 * The following function returns false either if every active queue
797 * must receive the same share of the throughput (symmetric scenario),
798 * or, as a special case, if bfqq must receive a share of the
799 * throughput lower than or equal to the share that every other active
800 * queue must receive. If bfqq does sync I/O, then these are the only
801 * two cases where bfqq happens to be guaranteed its share of the
802 * throughput even if I/O dispatching is not plugged when bfqq remains
803 * temporarily empty (for more details, see the comments in the
804 * function bfq_better_to_idle()). For this reason, the return value
805 * of this function is used to check whether I/O-dispatch plugging can
808 * The above first case (symmetric scenario) occurs when:
809 * 1) all active queues have the same weight,
810 * 2) all active queues belong to the same I/O-priority class,
811 * 3) all active groups at the same level in the groups tree have the same
813 * 4) all active groups at the same level in the groups tree have the same
814 * number of children.
816 * Unfortunately, keeping the necessary state for evaluating exactly
817 * the last two symmetry sub-conditions above would be quite complex
818 * and time consuming. Therefore this function evaluates, instead,
819 * only the following stronger three sub-conditions, for which it is
820 * much easier to maintain the needed state:
821 * 1) all active queues have the same weight,
822 * 2) all active queues belong to the same I/O-priority class,
823 * 3) there is at most one active group.
824 * In particular, the last condition is always true if hierarchical
825 * support or the cgroups interface are not enabled, thus no state
826 * needs to be maintained in this case.
828 static bool bfq_asymmetric_scenario(struct bfq_data *bfqd,
829 struct bfq_queue *bfqq)
831 bool smallest_weight = bfqq &&
832 bfqq->weight_counter &&
833 bfqq->weight_counter ==
835 rb_first_cached(&bfqd->queue_weights_tree),
836 struct bfq_weight_counter,
840 * For queue weights to differ, queue_weights_tree must contain
841 * at least two nodes.
843 bool varied_queue_weights = !smallest_weight &&
844 !RB_EMPTY_ROOT(&bfqd->queue_weights_tree.rb_root) &&
845 (bfqd->queue_weights_tree.rb_root.rb_node->rb_left ||
846 bfqd->queue_weights_tree.rb_root.rb_node->rb_right);
848 bool multiple_classes_busy =
849 (bfqd->busy_queues[0] && bfqd->busy_queues[1]) ||
850 (bfqd->busy_queues[0] && bfqd->busy_queues[2]) ||
851 (bfqd->busy_queues[1] && bfqd->busy_queues[2]);
853 return varied_queue_weights || multiple_classes_busy
854 #ifdef CONFIG_BFQ_GROUP_IOSCHED
855 || bfqd->num_groups_with_pending_reqs > 1
861 * If the weight-counter tree passed as input contains no counter for
862 * the weight of the input queue, then add that counter; otherwise just
863 * increment the existing counter.
865 * Note that weight-counter trees contain few nodes in mostly symmetric
866 * scenarios. For example, if all queues have the same weight, then the
867 * weight-counter tree for the queues may contain at most one node.
868 * This holds even if low_latency is on, because weight-raised queues
869 * are not inserted in the tree.
870 * In most scenarios, the rate at which nodes are created/destroyed
873 void bfq_weights_tree_add(struct bfq_queue *bfqq)
875 struct rb_root_cached *root = &bfqq->bfqd->queue_weights_tree;
876 struct bfq_entity *entity = &bfqq->entity;
877 struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
878 bool leftmost = true;
881 * Do not insert if the queue is already associated with a
882 * counter, which happens if:
883 * 1) a request arrival has caused the queue to become both
884 * non-weight-raised, and hence change its weight, and
885 * backlogged; in this respect, each of the two events
886 * causes an invocation of this function,
887 * 2) this is the invocation of this function caused by the
888 * second event. This second invocation is actually useless,
889 * and we handle this fact by exiting immediately. More
890 * efficient or clearer solutions might possibly be adopted.
892 if (bfqq->weight_counter)
896 struct bfq_weight_counter *__counter = container_of(*new,
897 struct bfq_weight_counter,
901 if (entity->weight == __counter->weight) {
902 bfqq->weight_counter = __counter;
905 if (entity->weight < __counter->weight)
906 new = &((*new)->rb_left);
908 new = &((*new)->rb_right);
913 bfqq->weight_counter = kzalloc(sizeof(struct bfq_weight_counter),
917 * In the unlucky event of an allocation failure, we just
918 * exit. This will cause the weight of queue to not be
919 * considered in bfq_asymmetric_scenario, which, in its turn,
920 * causes the scenario to be deemed wrongly symmetric in case
921 * bfqq's weight would have been the only weight making the
922 * scenario asymmetric. On the bright side, no unbalance will
923 * however occur when bfqq becomes inactive again (the
924 * invocation of this function is triggered by an activation
925 * of queue). In fact, bfq_weights_tree_remove does nothing
926 * if !bfqq->weight_counter.
928 if (unlikely(!bfqq->weight_counter))
931 bfqq->weight_counter->weight = entity->weight;
932 rb_link_node(&bfqq->weight_counter->weights_node, parent, new);
933 rb_insert_color_cached(&bfqq->weight_counter->weights_node, root,
937 bfqq->weight_counter->num_active++;
942 * Decrement the weight counter associated with the queue, and, if the
943 * counter reaches 0, remove the counter from the tree.
944 * See the comments to the function bfq_weights_tree_add() for considerations
947 void bfq_weights_tree_remove(struct bfq_queue *bfqq)
949 struct rb_root_cached *root;
951 if (!bfqq->weight_counter)
954 root = &bfqq->bfqd->queue_weights_tree;
955 bfqq->weight_counter->num_active--;
956 if (bfqq->weight_counter->num_active > 0)
957 goto reset_entity_pointer;
959 rb_erase_cached(&bfqq->weight_counter->weights_node, root);
960 kfree(bfqq->weight_counter);
962 reset_entity_pointer:
963 bfqq->weight_counter = NULL;
968 * Return expired entry, or NULL to just start from scratch in rbtree.
970 static struct request *bfq_check_fifo(struct bfq_queue *bfqq,
971 struct request *last)
975 if (bfq_bfqq_fifo_expire(bfqq))
978 bfq_mark_bfqq_fifo_expire(bfqq);
980 rq = rq_entry_fifo(bfqq->fifo.next);
982 if (rq == last || ktime_get_ns() < rq->fifo_time)
985 bfq_log_bfqq(bfqq->bfqd, bfqq, "check_fifo: returned %p", rq);
989 static struct request *bfq_find_next_rq(struct bfq_data *bfqd,
990 struct bfq_queue *bfqq,
991 struct request *last)
993 struct rb_node *rbnext = rb_next(&last->rb_node);
994 struct rb_node *rbprev = rb_prev(&last->rb_node);
995 struct request *next, *prev = NULL;
997 /* Follow expired path, else get first next available. */
998 next = bfq_check_fifo(bfqq, last);
1003 prev = rb_entry_rq(rbprev);
1006 next = rb_entry_rq(rbnext);
1008 rbnext = rb_first(&bfqq->sort_list);
1009 if (rbnext && rbnext != &last->rb_node)
1010 next = rb_entry_rq(rbnext);
1013 return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last));
1016 /* see the definition of bfq_async_charge_factor for details */
1017 static unsigned long bfq_serv_to_charge(struct request *rq,
1018 struct bfq_queue *bfqq)
1020 if (bfq_bfqq_sync(bfqq) || bfqq->wr_coeff > 1 ||
1021 bfq_asymmetric_scenario(bfqq->bfqd, bfqq))
1022 return blk_rq_sectors(rq);
1024 return blk_rq_sectors(rq) * bfq_async_charge_factor;
1028 * bfq_updated_next_req - update the queue after a new next_rq selection.
1029 * @bfqd: the device data the queue belongs to.
1030 * @bfqq: the queue to update.
1032 * If the first request of a queue changes we make sure that the queue
1033 * has enough budget to serve at least its first request (if the
1034 * request has grown). We do this because if the queue has not enough
1035 * budget for its first request, it has to go through two dispatch
1036 * rounds to actually get it dispatched.
1038 static void bfq_updated_next_req(struct bfq_data *bfqd,
1039 struct bfq_queue *bfqq)
1041 struct bfq_entity *entity = &bfqq->entity;
1042 struct request *next_rq = bfqq->next_rq;
1043 unsigned long new_budget;
1048 if (bfqq == bfqd->in_service_queue)
1050 * In order not to break guarantees, budgets cannot be
1051 * changed after an entity has been selected.
1055 new_budget = max_t(unsigned long,
1056 max_t(unsigned long, bfqq->max_budget,
1057 bfq_serv_to_charge(next_rq, bfqq)),
1059 if (entity->budget != new_budget) {
1060 entity->budget = new_budget;
1061 bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu",
1063 bfq_requeue_bfqq(bfqd, bfqq, false);
1067 static unsigned int bfq_wr_duration(struct bfq_data *bfqd)
1071 if (bfqd->bfq_wr_max_time > 0)
1072 return bfqd->bfq_wr_max_time;
1074 dur = bfqd->rate_dur_prod;
1075 do_div(dur, bfqd->peak_rate);
1078 * Limit duration between 3 and 25 seconds. The upper limit
1079 * has been conservatively set after the following worst case:
1080 * on a QEMU/KVM virtual machine
1081 * - running in a slow PC
1082 * - with a virtual disk stacked on a slow low-end 5400rpm HDD
1083 * - serving a heavy I/O workload, such as the sequential reading
1085 * mplayer took 23 seconds to start, if constantly weight-raised.
1087 * As for higher values than that accommodating the above bad
1088 * scenario, tests show that higher values would often yield
1089 * the opposite of the desired result, i.e., would worsen
1090 * responsiveness by allowing non-interactive applications to
1091 * preserve weight raising for too long.
1093 * On the other end, lower values than 3 seconds make it
1094 * difficult for most interactive tasks to complete their jobs
1095 * before weight-raising finishes.
1097 return clamp_val(dur, msecs_to_jiffies(3000), msecs_to_jiffies(25000));
1100 /* switch back from soft real-time to interactive weight raising */
1101 static void switch_back_to_interactive_wr(struct bfq_queue *bfqq,
1102 struct bfq_data *bfqd)
1104 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1105 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1106 bfqq->last_wr_start_finish = bfqq->wr_start_at_switch_to_srt;
1110 bfq_bfqq_resume_state(struct bfq_queue *bfqq, struct bfq_data *bfqd,
1111 struct bfq_io_cq *bic, bool bfq_already_existing)
1113 unsigned int old_wr_coeff = 1;
1114 bool busy = bfq_already_existing && bfq_bfqq_busy(bfqq);
1116 if (bic->saved_has_short_ttime)
1117 bfq_mark_bfqq_has_short_ttime(bfqq);
1119 bfq_clear_bfqq_has_short_ttime(bfqq);
1121 if (bic->saved_IO_bound)
1122 bfq_mark_bfqq_IO_bound(bfqq);
1124 bfq_clear_bfqq_IO_bound(bfqq);
1126 bfqq->last_serv_time_ns = bic->saved_last_serv_time_ns;
1127 bfqq->inject_limit = bic->saved_inject_limit;
1128 bfqq->decrease_time_jif = bic->saved_decrease_time_jif;
1130 bfqq->entity.new_weight = bic->saved_weight;
1131 bfqq->ttime = bic->saved_ttime;
1132 bfqq->io_start_time = bic->saved_io_start_time;
1133 bfqq->tot_idle_time = bic->saved_tot_idle_time;
1135 * Restore weight coefficient only if low_latency is on
1137 if (bfqd->low_latency) {
1138 old_wr_coeff = bfqq->wr_coeff;
1139 bfqq->wr_coeff = bic->saved_wr_coeff;
1141 bfqq->service_from_wr = bic->saved_service_from_wr;
1142 bfqq->wr_start_at_switch_to_srt = bic->saved_wr_start_at_switch_to_srt;
1143 bfqq->last_wr_start_finish = bic->saved_last_wr_start_finish;
1144 bfqq->wr_cur_max_time = bic->saved_wr_cur_max_time;
1146 if (bfqq->wr_coeff > 1 && (bfq_bfqq_in_large_burst(bfqq) ||
1147 time_is_before_jiffies(bfqq->last_wr_start_finish +
1148 bfqq->wr_cur_max_time))) {
1149 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
1150 !bfq_bfqq_in_large_burst(bfqq) &&
1151 time_is_after_eq_jiffies(bfqq->wr_start_at_switch_to_srt +
1152 bfq_wr_duration(bfqd))) {
1153 switch_back_to_interactive_wr(bfqq, bfqd);
1156 bfq_log_bfqq(bfqq->bfqd, bfqq,
1157 "resume state: switching off wr");
1161 /* make sure weight will be updated, however we got here */
1162 bfqq->entity.prio_changed = 1;
1167 if (old_wr_coeff == 1 && bfqq->wr_coeff > 1)
1168 bfqd->wr_busy_queues++;
1169 else if (old_wr_coeff > 1 && bfqq->wr_coeff == 1)
1170 bfqd->wr_busy_queues--;
1173 static int bfqq_process_refs(struct bfq_queue *bfqq)
1175 return bfqq->ref - bfqq->entity.allocated -
1176 bfqq->entity.on_st_or_in_serv -
1177 (bfqq->weight_counter != NULL) - bfqq->stable_ref;
1180 /* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */
1181 static void bfq_reset_burst_list(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1183 struct bfq_queue *item;
1184 struct hlist_node *n;
1186 hlist_for_each_entry_safe(item, n, &bfqd->burst_list, burst_list_node)
1187 hlist_del_init(&item->burst_list_node);
1190 * Start the creation of a new burst list only if there is no
1191 * active queue. See comments on the conditional invocation of
1192 * bfq_handle_burst().
1194 if (bfq_tot_busy_queues(bfqd) == 0) {
1195 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1196 bfqd->burst_size = 1;
1198 bfqd->burst_size = 0;
1200 bfqd->burst_parent_entity = bfqq->entity.parent;
1203 /* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */
1204 static void bfq_add_to_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1206 /* Increment burst size to take into account also bfqq */
1209 if (bfqd->burst_size == bfqd->bfq_large_burst_thresh) {
1210 struct bfq_queue *pos, *bfqq_item;
1211 struct hlist_node *n;
1214 * Enough queues have been activated shortly after each
1215 * other to consider this burst as large.
1217 bfqd->large_burst = true;
1220 * We can now mark all queues in the burst list as
1221 * belonging to a large burst.
1223 hlist_for_each_entry(bfqq_item, &bfqd->burst_list,
1225 bfq_mark_bfqq_in_large_burst(bfqq_item);
1226 bfq_mark_bfqq_in_large_burst(bfqq);
1229 * From now on, and until the current burst finishes, any
1230 * new queue being activated shortly after the last queue
1231 * was inserted in the burst can be immediately marked as
1232 * belonging to a large burst. So the burst list is not
1233 * needed any more. Remove it.
1235 hlist_for_each_entry_safe(pos, n, &bfqd->burst_list,
1237 hlist_del_init(&pos->burst_list_node);
1239 * Burst not yet large: add bfqq to the burst list. Do
1240 * not increment the ref counter for bfqq, because bfqq
1241 * is removed from the burst list before freeing bfqq
1244 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1248 * If many queues belonging to the same group happen to be created
1249 * shortly after each other, then the processes associated with these
1250 * queues have typically a common goal. In particular, bursts of queue
1251 * creations are usually caused by services or applications that spawn
1252 * many parallel threads/processes. Examples are systemd during boot,
1253 * or git grep. To help these processes get their job done as soon as
1254 * possible, it is usually better to not grant either weight-raising
1255 * or device idling to their queues, unless these queues must be
1256 * protected from the I/O flowing through other active queues.
1258 * In this comment we describe, firstly, the reasons why this fact
1259 * holds, and, secondly, the next function, which implements the main
1260 * steps needed to properly mark these queues so that they can then be
1261 * treated in a different way.
1263 * The above services or applications benefit mostly from a high
1264 * throughput: the quicker the requests of the activated queues are
1265 * cumulatively served, the sooner the target job of these queues gets
1266 * completed. As a consequence, weight-raising any of these queues,
1267 * which also implies idling the device for it, is almost always
1268 * counterproductive, unless there are other active queues to isolate
1269 * these new queues from. If there no other active queues, then
1270 * weight-raising these new queues just lowers throughput in most
1273 * On the other hand, a burst of queue creations may be caused also by
1274 * the start of an application that does not consist of a lot of
1275 * parallel I/O-bound threads. In fact, with a complex application,
1276 * several short processes may need to be executed to start-up the
1277 * application. In this respect, to start an application as quickly as
1278 * possible, the best thing to do is in any case to privilege the I/O
1279 * related to the application with respect to all other
1280 * I/O. Therefore, the best strategy to start as quickly as possible
1281 * an application that causes a burst of queue creations is to
1282 * weight-raise all the queues created during the burst. This is the
1283 * exact opposite of the best strategy for the other type of bursts.
1285 * In the end, to take the best action for each of the two cases, the
1286 * two types of bursts need to be distinguished. Fortunately, this
1287 * seems relatively easy, by looking at the sizes of the bursts. In
1288 * particular, we found a threshold such that only bursts with a
1289 * larger size than that threshold are apparently caused by
1290 * services or commands such as systemd or git grep. For brevity,
1291 * hereafter we call just 'large' these bursts. BFQ *does not*
1292 * weight-raise queues whose creation occurs in a large burst. In
1293 * addition, for each of these queues BFQ performs or does not perform
1294 * idling depending on which choice boosts the throughput more. The
1295 * exact choice depends on the device and request pattern at
1298 * Unfortunately, false positives may occur while an interactive task
1299 * is starting (e.g., an application is being started). The
1300 * consequence is that the queues associated with the task do not
1301 * enjoy weight raising as expected. Fortunately these false positives
1302 * are very rare. They typically occur if some service happens to
1303 * start doing I/O exactly when the interactive task starts.
1305 * Turning back to the next function, it is invoked only if there are
1306 * no active queues (apart from active queues that would belong to the
1307 * same, possible burst bfqq would belong to), and it implements all
1308 * the steps needed to detect the occurrence of a large burst and to
1309 * properly mark all the queues belonging to it (so that they can then
1310 * be treated in a different way). This goal is achieved by
1311 * maintaining a "burst list" that holds, temporarily, the queues that
1312 * belong to the burst in progress. The list is then used to mark
1313 * these queues as belonging to a large burst if the burst does become
1314 * large. The main steps are the following.
1316 * . when the very first queue is created, the queue is inserted into the
1317 * list (as it could be the first queue in a possible burst)
1319 * . if the current burst has not yet become large, and a queue Q that does
1320 * not yet belong to the burst is activated shortly after the last time
1321 * at which a new queue entered the burst list, then the function appends
1322 * Q to the burst list
1324 * . if, as a consequence of the previous step, the burst size reaches
1325 * the large-burst threshold, then
1327 * . all the queues in the burst list are marked as belonging to a
1330 * . the burst list is deleted; in fact, the burst list already served
1331 * its purpose (keeping temporarily track of the queues in a burst,
1332 * so as to be able to mark them as belonging to a large burst in the
1333 * previous sub-step), and now is not needed any more
1335 * . the device enters a large-burst mode
1337 * . if a queue Q that does not belong to the burst is created while
1338 * the device is in large-burst mode and shortly after the last time
1339 * at which a queue either entered the burst list or was marked as
1340 * belonging to the current large burst, then Q is immediately marked
1341 * as belonging to a large burst.
1343 * . if a queue Q that does not belong to the burst is created a while
1344 * later, i.e., not shortly after, than the last time at which a queue
1345 * either entered the burst list or was marked as belonging to the
1346 * current large burst, then the current burst is deemed as finished and:
1348 * . the large-burst mode is reset if set
1350 * . the burst list is emptied
1352 * . Q is inserted in the burst list, as Q may be the first queue
1353 * in a possible new burst (then the burst list contains just Q
1356 static void bfq_handle_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1359 * If bfqq is already in the burst list or is part of a large
1360 * burst, or finally has just been split, then there is
1361 * nothing else to do.
1363 if (!hlist_unhashed(&bfqq->burst_list_node) ||
1364 bfq_bfqq_in_large_burst(bfqq) ||
1365 time_is_after_eq_jiffies(bfqq->split_time +
1366 msecs_to_jiffies(10)))
1370 * If bfqq's creation happens late enough, or bfqq belongs to
1371 * a different group than the burst group, then the current
1372 * burst is finished, and related data structures must be
1375 * In this respect, consider the special case where bfqq is
1376 * the very first queue created after BFQ is selected for this
1377 * device. In this case, last_ins_in_burst and
1378 * burst_parent_entity are not yet significant when we get
1379 * here. But it is easy to verify that, whether or not the
1380 * following condition is true, bfqq will end up being
1381 * inserted into the burst list. In particular the list will
1382 * happen to contain only bfqq. And this is exactly what has
1383 * to happen, as bfqq may be the first queue of the first
1386 if (time_is_before_jiffies(bfqd->last_ins_in_burst +
1387 bfqd->bfq_burst_interval) ||
1388 bfqq->entity.parent != bfqd->burst_parent_entity) {
1389 bfqd->large_burst = false;
1390 bfq_reset_burst_list(bfqd, bfqq);
1395 * If we get here, then bfqq is being activated shortly after the
1396 * last queue. So, if the current burst is also large, we can mark
1397 * bfqq as belonging to this large burst immediately.
1399 if (bfqd->large_burst) {
1400 bfq_mark_bfqq_in_large_burst(bfqq);
1405 * If we get here, then a large-burst state has not yet been
1406 * reached, but bfqq is being activated shortly after the last
1407 * queue. Then we add bfqq to the burst.
1409 bfq_add_to_burst(bfqd, bfqq);
1412 * At this point, bfqq either has been added to the current
1413 * burst or has caused the current burst to terminate and a
1414 * possible new burst to start. In particular, in the second
1415 * case, bfqq has become the first queue in the possible new
1416 * burst. In both cases last_ins_in_burst needs to be moved
1419 bfqd->last_ins_in_burst = jiffies;
1422 static int bfq_bfqq_budget_left(struct bfq_queue *bfqq)
1424 struct bfq_entity *entity = &bfqq->entity;
1426 return entity->budget - entity->service;
1430 * If enough samples have been computed, return the current max budget
1431 * stored in bfqd, which is dynamically updated according to the
1432 * estimated disk peak rate; otherwise return the default max budget
1434 static int bfq_max_budget(struct bfq_data *bfqd)
1436 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1437 return bfq_default_max_budget;
1439 return bfqd->bfq_max_budget;
1443 * Return min budget, which is a fraction of the current or default
1444 * max budget (trying with 1/32)
1446 static int bfq_min_budget(struct bfq_data *bfqd)
1448 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1449 return bfq_default_max_budget / 32;
1451 return bfqd->bfq_max_budget / 32;
1455 * The next function, invoked after the input queue bfqq switches from
1456 * idle to busy, updates the budget of bfqq. The function also tells
1457 * whether the in-service queue should be expired, by returning
1458 * true. The purpose of expiring the in-service queue is to give bfqq
1459 * the chance to possibly preempt the in-service queue, and the reason
1460 * for preempting the in-service queue is to achieve one of the two
1463 * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has
1464 * expired because it has remained idle. In particular, bfqq may have
1465 * expired for one of the following two reasons:
1467 * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling
1468 * and did not make it to issue a new request before its last
1469 * request was served;
1471 * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue
1472 * a new request before the expiration of the idling-time.
1474 * Even if bfqq has expired for one of the above reasons, the process
1475 * associated with the queue may be however issuing requests greedily,
1476 * and thus be sensitive to the bandwidth it receives (bfqq may have
1477 * remained idle for other reasons: CPU high load, bfqq not enjoying
1478 * idling, I/O throttling somewhere in the path from the process to
1479 * the I/O scheduler, ...). But if, after every expiration for one of
1480 * the above two reasons, bfqq has to wait for the service of at least
1481 * one full budget of another queue before being served again, then
1482 * bfqq is likely to get a much lower bandwidth or resource time than
1483 * its reserved ones. To address this issue, two countermeasures need
1486 * First, the budget and the timestamps of bfqq need to be updated in
1487 * a special way on bfqq reactivation: they need to be updated as if
1488 * bfqq did not remain idle and did not expire. In fact, if they are
1489 * computed as if bfqq expired and remained idle until reactivation,
1490 * then the process associated with bfqq is treated as if, instead of
1491 * being greedy, it stopped issuing requests when bfqq remained idle,
1492 * and restarts issuing requests only on this reactivation. In other
1493 * words, the scheduler does not help the process recover the "service
1494 * hole" between bfqq expiration and reactivation. As a consequence,
1495 * the process receives a lower bandwidth than its reserved one. In
1496 * contrast, to recover this hole, the budget must be updated as if
1497 * bfqq was not expired at all before this reactivation, i.e., it must
1498 * be set to the value of the remaining budget when bfqq was
1499 * expired. Along the same line, timestamps need to be assigned the
1500 * value they had the last time bfqq was selected for service, i.e.,
1501 * before last expiration. Thus timestamps need to be back-shifted
1502 * with respect to their normal computation (see [1] for more details
1503 * on this tricky aspect).
1505 * Secondly, to allow the process to recover the hole, the in-service
1506 * queue must be expired too, to give bfqq the chance to preempt it
1507 * immediately. In fact, if bfqq has to wait for a full budget of the
1508 * in-service queue to be completed, then it may become impossible to
1509 * let the process recover the hole, even if the back-shifted
1510 * timestamps of bfqq are lower than those of the in-service queue. If
1511 * this happens for most or all of the holes, then the process may not
1512 * receive its reserved bandwidth. In this respect, it is worth noting
1513 * that, being the service of outstanding requests unpreemptible, a
1514 * little fraction of the holes may however be unrecoverable, thereby
1515 * causing a little loss of bandwidth.
1517 * The last important point is detecting whether bfqq does need this
1518 * bandwidth recovery. In this respect, the next function deems the
1519 * process associated with bfqq greedy, and thus allows it to recover
1520 * the hole, if: 1) the process is waiting for the arrival of a new
1521 * request (which implies that bfqq expired for one of the above two
1522 * reasons), and 2) such a request has arrived soon. The first
1523 * condition is controlled through the flag non_blocking_wait_rq,
1524 * while the second through the flag arrived_in_time. If both
1525 * conditions hold, then the function computes the budget in the
1526 * above-described special way, and signals that the in-service queue
1527 * should be expired. Timestamp back-shifting is done later in
1528 * __bfq_activate_entity.
1530 * 2. Reduce latency. Even if timestamps are not backshifted to let
1531 * the process associated with bfqq recover a service hole, bfqq may
1532 * however happen to have, after being (re)activated, a lower finish
1533 * timestamp than the in-service queue. That is, the next budget of
1534 * bfqq may have to be completed before the one of the in-service
1535 * queue. If this is the case, then preempting the in-service queue
1536 * allows this goal to be achieved, apart from the unpreemptible,
1537 * outstanding requests mentioned above.
1539 * Unfortunately, regardless of which of the above two goals one wants
1540 * to achieve, service trees need first to be updated to know whether
1541 * the in-service queue must be preempted. To have service trees
1542 * correctly updated, the in-service queue must be expired and
1543 * rescheduled, and bfqq must be scheduled too. This is one of the
1544 * most costly operations (in future versions, the scheduling
1545 * mechanism may be re-designed in such a way to make it possible to
1546 * know whether preemption is needed without needing to update service
1547 * trees). In addition, queue preemptions almost always cause random
1548 * I/O, which may in turn cause loss of throughput. Finally, there may
1549 * even be no in-service queue when the next function is invoked (so,
1550 * no queue to compare timestamps with). Because of these facts, the
1551 * next function adopts the following simple scheme to avoid costly
1552 * operations, too frequent preemptions and too many dependencies on
1553 * the state of the scheduler: it requests the expiration of the
1554 * in-service queue (unconditionally) only for queues that need to
1555 * recover a hole. Then it delegates to other parts of the code the
1556 * responsibility of handling the above case 2.
1558 static bool bfq_bfqq_update_budg_for_activation(struct bfq_data *bfqd,
1559 struct bfq_queue *bfqq,
1560 bool arrived_in_time)
1562 struct bfq_entity *entity = &bfqq->entity;
1565 * In the next compound condition, we check also whether there
1566 * is some budget left, because otherwise there is no point in
1567 * trying to go on serving bfqq with this same budget: bfqq
1568 * would be expired immediately after being selected for
1569 * service. This would only cause useless overhead.
1571 if (bfq_bfqq_non_blocking_wait_rq(bfqq) && arrived_in_time &&
1572 bfq_bfqq_budget_left(bfqq) > 0) {
1574 * We do not clear the flag non_blocking_wait_rq here, as
1575 * the latter is used in bfq_activate_bfqq to signal
1576 * that timestamps need to be back-shifted (and is
1577 * cleared right after).
1581 * In next assignment we rely on that either
1582 * entity->service or entity->budget are not updated
1583 * on expiration if bfqq is empty (see
1584 * __bfq_bfqq_recalc_budget). Thus both quantities
1585 * remain unchanged after such an expiration, and the
1586 * following statement therefore assigns to
1587 * entity->budget the remaining budget on such an
1590 entity->budget = min_t(unsigned long,
1591 bfq_bfqq_budget_left(bfqq),
1595 * At this point, we have used entity->service to get
1596 * the budget left (needed for updating
1597 * entity->budget). Thus we finally can, and have to,
1598 * reset entity->service. The latter must be reset
1599 * because bfqq would otherwise be charged again for
1600 * the service it has received during its previous
1603 entity->service = 0;
1609 * We can finally complete expiration, by setting service to 0.
1611 entity->service = 0;
1612 entity->budget = max_t(unsigned long, bfqq->max_budget,
1613 bfq_serv_to_charge(bfqq->next_rq, bfqq));
1614 bfq_clear_bfqq_non_blocking_wait_rq(bfqq);
1619 * Return the farthest past time instant according to jiffies
1622 static unsigned long bfq_smallest_from_now(void)
1624 return jiffies - MAX_JIFFY_OFFSET;
1627 static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data *bfqd,
1628 struct bfq_queue *bfqq,
1629 unsigned int old_wr_coeff,
1630 bool wr_or_deserves_wr,
1635 if (old_wr_coeff == 1 && wr_or_deserves_wr) {
1636 /* start a weight-raising period */
1638 bfqq->service_from_wr = 0;
1639 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1640 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1643 * No interactive weight raising in progress
1644 * here: assign minus infinity to
1645 * wr_start_at_switch_to_srt, to make sure
1646 * that, at the end of the soft-real-time
1647 * weight raising periods that is starting
1648 * now, no interactive weight-raising period
1649 * may be wrongly considered as still in
1650 * progress (and thus actually started by
1653 bfqq->wr_start_at_switch_to_srt =
1654 bfq_smallest_from_now();
1655 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1656 BFQ_SOFTRT_WEIGHT_FACTOR;
1657 bfqq->wr_cur_max_time =
1658 bfqd->bfq_wr_rt_max_time;
1662 * If needed, further reduce budget to make sure it is
1663 * close to bfqq's backlog, so as to reduce the
1664 * scheduling-error component due to a too large
1665 * budget. Do not care about throughput consequences,
1666 * but only about latency. Finally, do not assign a
1667 * too small budget either, to avoid increasing
1668 * latency by causing too frequent expirations.
1670 bfqq->entity.budget = min_t(unsigned long,
1671 bfqq->entity.budget,
1672 2 * bfq_min_budget(bfqd));
1673 } else if (old_wr_coeff > 1) {
1674 if (interactive) { /* update wr coeff and duration */
1675 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1676 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1677 } else if (in_burst)
1681 * The application is now or still meeting the
1682 * requirements for being deemed soft rt. We
1683 * can then correctly and safely (re)charge
1684 * the weight-raising duration for the
1685 * application with the weight-raising
1686 * duration for soft rt applications.
1688 * In particular, doing this recharge now, i.e.,
1689 * before the weight-raising period for the
1690 * application finishes, reduces the probability
1691 * of the following negative scenario:
1692 * 1) the weight of a soft rt application is
1693 * raised at startup (as for any newly
1694 * created application),
1695 * 2) since the application is not interactive,
1696 * at a certain time weight-raising is
1697 * stopped for the application,
1698 * 3) at that time the application happens to
1699 * still have pending requests, and hence
1700 * is destined to not have a chance to be
1701 * deemed soft rt before these requests are
1702 * completed (see the comments to the
1703 * function bfq_bfqq_softrt_next_start()
1704 * for details on soft rt detection),
1705 * 4) these pending requests experience a high
1706 * latency because the application is not
1707 * weight-raised while they are pending.
1709 if (bfqq->wr_cur_max_time !=
1710 bfqd->bfq_wr_rt_max_time) {
1711 bfqq->wr_start_at_switch_to_srt =
1712 bfqq->last_wr_start_finish;
1714 bfqq->wr_cur_max_time =
1715 bfqd->bfq_wr_rt_max_time;
1716 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1717 BFQ_SOFTRT_WEIGHT_FACTOR;
1719 bfqq->last_wr_start_finish = jiffies;
1724 static bool bfq_bfqq_idle_for_long_time(struct bfq_data *bfqd,
1725 struct bfq_queue *bfqq)
1727 return bfqq->dispatched == 0 &&
1728 time_is_before_jiffies(
1729 bfqq->budget_timeout +
1730 bfqd->bfq_wr_min_idle_time);
1735 * Return true if bfqq is in a higher priority class, or has a higher
1736 * weight than the in-service queue.
1738 static bool bfq_bfqq_higher_class_or_weight(struct bfq_queue *bfqq,
1739 struct bfq_queue *in_serv_bfqq)
1741 int bfqq_weight, in_serv_weight;
1743 if (bfqq->ioprio_class < in_serv_bfqq->ioprio_class)
1746 if (in_serv_bfqq->entity.parent == bfqq->entity.parent) {
1747 bfqq_weight = bfqq->entity.weight;
1748 in_serv_weight = in_serv_bfqq->entity.weight;
1750 if (bfqq->entity.parent)
1751 bfqq_weight = bfqq->entity.parent->weight;
1753 bfqq_weight = bfqq->entity.weight;
1754 if (in_serv_bfqq->entity.parent)
1755 in_serv_weight = in_serv_bfqq->entity.parent->weight;
1757 in_serv_weight = in_serv_bfqq->entity.weight;
1760 return bfqq_weight > in_serv_weight;
1763 static bool bfq_better_to_idle(struct bfq_queue *bfqq);
1765 static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data *bfqd,
1766 struct bfq_queue *bfqq,
1771 bool soft_rt, in_burst, wr_or_deserves_wr,
1772 bfqq_wants_to_preempt,
1773 idle_for_long_time = bfq_bfqq_idle_for_long_time(bfqd, bfqq),
1775 * See the comments on
1776 * bfq_bfqq_update_budg_for_activation for
1777 * details on the usage of the next variable.
1779 arrived_in_time = ktime_get_ns() <=
1780 bfqq->ttime.last_end_request +
1781 bfqd->bfq_slice_idle * 3;
1785 * bfqq deserves to be weight-raised if:
1787 * - it does not belong to a large burst,
1788 * - it has been idle for enough time or is soft real-time,
1789 * - is linked to a bfq_io_cq (it is not shared in any sense),
1790 * - has a default weight (otherwise we assume the user wanted
1791 * to control its weight explicitly)
1793 in_burst = bfq_bfqq_in_large_burst(bfqq);
1794 soft_rt = bfqd->bfq_wr_max_softrt_rate > 0 &&
1795 !BFQQ_TOTALLY_SEEKY(bfqq) &&
1797 time_is_before_jiffies(bfqq->soft_rt_next_start) &&
1798 bfqq->dispatched == 0 &&
1799 bfqq->entity.new_weight == 40;
1800 *interactive = !in_burst && idle_for_long_time &&
1801 bfqq->entity.new_weight == 40;
1803 * Merged bfq_queues are kept out of weight-raising
1804 * (low-latency) mechanisms. The reason is that these queues
1805 * are usually created for non-interactive and
1806 * non-soft-real-time tasks. Yet this is not the case for
1807 * stably-merged queues. These queues are merged just because
1808 * they are created shortly after each other. So they may
1809 * easily serve the I/O of an interactive or soft-real time
1810 * application, if the application happens to spawn multiple
1811 * processes. So let also stably-merged queued enjoy weight
1814 wr_or_deserves_wr = bfqd->low_latency &&
1815 (bfqq->wr_coeff > 1 ||
1816 (bfq_bfqq_sync(bfqq) &&
1817 (bfqq->bic || RQ_BIC(rq)->stably_merged) &&
1818 (*interactive || soft_rt)));
1821 * Using the last flag, update budget and check whether bfqq
1822 * may want to preempt the in-service queue.
1824 bfqq_wants_to_preempt =
1825 bfq_bfqq_update_budg_for_activation(bfqd, bfqq,
1829 * If bfqq happened to be activated in a burst, but has been
1830 * idle for much more than an interactive queue, then we
1831 * assume that, in the overall I/O initiated in the burst, the
1832 * I/O associated with bfqq is finished. So bfqq does not need
1833 * to be treated as a queue belonging to a burst
1834 * anymore. Accordingly, we reset bfqq's in_large_burst flag
1835 * if set, and remove bfqq from the burst list if it's
1836 * there. We do not decrement burst_size, because the fact
1837 * that bfqq does not need to belong to the burst list any
1838 * more does not invalidate the fact that bfqq was created in
1841 if (likely(!bfq_bfqq_just_created(bfqq)) &&
1842 idle_for_long_time &&
1843 time_is_before_jiffies(
1844 bfqq->budget_timeout +
1845 msecs_to_jiffies(10000))) {
1846 hlist_del_init(&bfqq->burst_list_node);
1847 bfq_clear_bfqq_in_large_burst(bfqq);
1850 bfq_clear_bfqq_just_created(bfqq);
1852 if (bfqd->low_latency) {
1853 if (unlikely(time_is_after_jiffies(bfqq->split_time)))
1856 jiffies - bfqd->bfq_wr_min_idle_time - 1;
1858 if (time_is_before_jiffies(bfqq->split_time +
1859 bfqd->bfq_wr_min_idle_time)) {
1860 bfq_update_bfqq_wr_on_rq_arrival(bfqd, bfqq,
1867 if (old_wr_coeff != bfqq->wr_coeff)
1868 bfqq->entity.prio_changed = 1;
1872 bfqq->last_idle_bklogged = jiffies;
1873 bfqq->service_from_backlogged = 0;
1874 bfq_clear_bfqq_softrt_update(bfqq);
1876 bfq_add_bfqq_busy(bfqq);
1879 * Expire in-service queue if preemption may be needed for
1880 * guarantees or throughput. As for guarantees, we care
1881 * explicitly about two cases. The first is that bfqq has to
1882 * recover a service hole, as explained in the comments on
1883 * bfq_bfqq_update_budg_for_activation(), i.e., that
1884 * bfqq_wants_to_preempt is true. However, if bfqq does not
1885 * carry time-critical I/O, then bfqq's bandwidth is less
1886 * important than that of queues that carry time-critical I/O.
1887 * So, as a further constraint, we consider this case only if
1888 * bfqq is at least as weight-raised, i.e., at least as time
1889 * critical, as the in-service queue.
1891 * The second case is that bfqq is in a higher priority class,
1892 * or has a higher weight than the in-service queue. If this
1893 * condition does not hold, we don't care because, even if
1894 * bfqq does not start to be served immediately, the resulting
1895 * delay for bfqq's I/O is however lower or much lower than
1896 * the ideal completion time to be guaranteed to bfqq's I/O.
1898 * In both cases, preemption is needed only if, according to
1899 * the timestamps of both bfqq and of the in-service queue,
1900 * bfqq actually is the next queue to serve. So, to reduce
1901 * useless preemptions, the return value of
1902 * next_queue_may_preempt() is considered in the next compound
1903 * condition too. Yet next_queue_may_preempt() just checks a
1904 * simple, necessary condition for bfqq to be the next queue
1905 * to serve. In fact, to evaluate a sufficient condition, the
1906 * timestamps of the in-service queue would need to be
1907 * updated, and this operation is quite costly (see the
1908 * comments on bfq_bfqq_update_budg_for_activation()).
1910 * As for throughput, we ask bfq_better_to_idle() whether we
1911 * still need to plug I/O dispatching. If bfq_better_to_idle()
1912 * says no, then plugging is not needed any longer, either to
1913 * boost throughput or to perserve service guarantees. Then
1914 * the best option is to stop plugging I/O, as not doing so
1915 * would certainly lower throughput. We may end up in this
1916 * case if: (1) upon a dispatch attempt, we detected that it
1917 * was better to plug I/O dispatch, and to wait for a new
1918 * request to arrive for the currently in-service queue, but
1919 * (2) this switch of bfqq to busy changes the scenario.
1921 if (bfqd->in_service_queue &&
1922 ((bfqq_wants_to_preempt &&
1923 bfqq->wr_coeff >= bfqd->in_service_queue->wr_coeff) ||
1924 bfq_bfqq_higher_class_or_weight(bfqq, bfqd->in_service_queue) ||
1925 !bfq_better_to_idle(bfqd->in_service_queue)) &&
1926 next_queue_may_preempt(bfqd))
1927 bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
1928 false, BFQQE_PREEMPTED);
1931 static void bfq_reset_inject_limit(struct bfq_data *bfqd,
1932 struct bfq_queue *bfqq)
1934 /* invalidate baseline total service time */
1935 bfqq->last_serv_time_ns = 0;
1938 * Reset pointer in case we are waiting for
1939 * some request completion.
1941 bfqd->waited_rq = NULL;
1944 * If bfqq has a short think time, then start by setting the
1945 * inject limit to 0 prudentially, because the service time of
1946 * an injected I/O request may be higher than the think time
1947 * of bfqq, and therefore, if one request was injected when
1948 * bfqq remains empty, this injected request might delay the
1949 * service of the next I/O request for bfqq significantly. In
1950 * case bfqq can actually tolerate some injection, then the
1951 * adaptive update will however raise the limit soon. This
1952 * lucky circumstance holds exactly because bfqq has a short
1953 * think time, and thus, after remaining empty, is likely to
1954 * get new I/O enqueued---and then completed---before being
1955 * expired. This is the very pattern that gives the
1956 * limit-update algorithm the chance to measure the effect of
1957 * injection on request service times, and then to update the
1958 * limit accordingly.
1960 * However, in the following special case, the inject limit is
1961 * left to 1 even if the think time is short: bfqq's I/O is
1962 * synchronized with that of some other queue, i.e., bfqq may
1963 * receive new I/O only after the I/O of the other queue is
1964 * completed. Keeping the inject limit to 1 allows the
1965 * blocking I/O to be served while bfqq is in service. And
1966 * this is very convenient both for bfqq and for overall
1967 * throughput, as explained in detail in the comments in
1968 * bfq_update_has_short_ttime().
1970 * On the opposite end, if bfqq has a long think time, then
1971 * start directly by 1, because:
1972 * a) on the bright side, keeping at most one request in
1973 * service in the drive is unlikely to cause any harm to the
1974 * latency of bfqq's requests, as the service time of a single
1975 * request is likely to be lower than the think time of bfqq;
1976 * b) on the downside, after becoming empty, bfqq is likely to
1977 * expire before getting its next request. With this request
1978 * arrival pattern, it is very hard to sample total service
1979 * times and update the inject limit accordingly (see comments
1980 * on bfq_update_inject_limit()). So the limit is likely to be
1981 * never, or at least seldom, updated. As a consequence, by
1982 * setting the limit to 1, we avoid that no injection ever
1983 * occurs with bfqq. On the downside, this proactive step
1984 * further reduces chances to actually compute the baseline
1985 * total service time. Thus it reduces chances to execute the
1986 * limit-update algorithm and possibly raise the limit to more
1989 if (bfq_bfqq_has_short_ttime(bfqq))
1990 bfqq->inject_limit = 0;
1992 bfqq->inject_limit = 1;
1994 bfqq->decrease_time_jif = jiffies;
1997 static void bfq_update_io_intensity(struct bfq_queue *bfqq, u64 now_ns)
1999 u64 tot_io_time = now_ns - bfqq->io_start_time;
2001 if (RB_EMPTY_ROOT(&bfqq->sort_list) && bfqq->dispatched == 0)
2002 bfqq->tot_idle_time +=
2003 now_ns - bfqq->ttime.last_end_request;
2005 if (unlikely(bfq_bfqq_just_created(bfqq)))
2009 * Must be busy for at least about 80% of the time to be
2010 * considered I/O bound.
2012 if (bfqq->tot_idle_time * 5 > tot_io_time)
2013 bfq_clear_bfqq_IO_bound(bfqq);
2015 bfq_mark_bfqq_IO_bound(bfqq);
2018 * Keep an observation window of at most 200 ms in the past
2021 if (tot_io_time > 200 * NSEC_PER_MSEC) {
2022 bfqq->io_start_time = now_ns - (tot_io_time>>1);
2023 bfqq->tot_idle_time >>= 1;
2028 * Detect whether bfqq's I/O seems synchronized with that of some
2029 * other queue, i.e., whether bfqq, after remaining empty, happens to
2030 * receive new I/O only right after some I/O request of the other
2031 * queue has been completed. We call waker queue the other queue, and
2032 * we assume, for simplicity, that bfqq may have at most one waker
2035 * A remarkable throughput boost can be reached by unconditionally
2036 * injecting the I/O of the waker queue, every time a new
2037 * bfq_dispatch_request happens to be invoked while I/O is being
2038 * plugged for bfqq. In addition to boosting throughput, this
2039 * unblocks bfqq's I/O, thereby improving bandwidth and latency for
2040 * bfqq. Note that these same results may be achieved with the general
2041 * injection mechanism, but less effectively. For details on this
2042 * aspect, see the comments on the choice of the queue for injection
2043 * in bfq_select_queue().
2045 * Turning back to the detection of a waker queue, a queue Q is deemed as a
2046 * waker queue for bfqq if, for three consecutive times, bfqq happens to become
2047 * non empty right after a request of Q has been completed within given
2048 * timeout. In this respect, even if bfqq is empty, we do not check for a waker
2049 * if it still has some in-flight I/O. In fact, in this case bfqq is actually
2050 * still being served by the drive, and may receive new I/O on the completion
2051 * of some of the in-flight requests. In particular, on the first time, Q is
2052 * tentatively set as a candidate waker queue, while on the third consecutive
2053 * time that Q is detected, the field waker_bfqq is set to Q, to confirm that Q
2054 * is a waker queue for bfqq. These detection steps are performed only if bfqq
2055 * has a long think time, so as to make it more likely that bfqq's I/O is
2056 * actually being blocked by a synchronization. This last filter, plus the
2057 * above three-times requirement and time limit for detection, make false
2058 * positives less likely.
2062 * The sooner a waker queue is detected, the sooner throughput can be
2063 * boosted by injecting I/O from the waker queue. Fortunately,
2064 * detection is likely to be actually fast, for the following
2065 * reasons. While blocked by synchronization, bfqq has a long think
2066 * time. This implies that bfqq's inject limit is at least equal to 1
2067 * (see the comments in bfq_update_inject_limit()). So, thanks to
2068 * injection, the waker queue is likely to be served during the very
2069 * first I/O-plugging time interval for bfqq. This triggers the first
2070 * step of the detection mechanism. Thanks again to injection, the
2071 * candidate waker queue is then likely to be confirmed no later than
2072 * during the next I/O-plugging interval for bfqq.
2076 * On queue merging all waker information is lost.
2078 static void bfq_check_waker(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2081 char waker_name[MAX_BFQQ_NAME_LENGTH];
2083 if (!bfqd->last_completed_rq_bfqq ||
2084 bfqd->last_completed_rq_bfqq == bfqq ||
2085 bfq_bfqq_has_short_ttime(bfqq) ||
2086 now_ns - bfqd->last_completion >= 4 * NSEC_PER_MSEC ||
2087 bfqd->last_completed_rq_bfqq == &bfqd->oom_bfqq ||
2088 bfqq == &bfqd->oom_bfqq)
2092 * We reset waker detection logic also if too much time has passed
2093 * since the first detection. If wakeups are rare, pointless idling
2094 * doesn't hurt throughput that much. The condition below makes sure
2095 * we do not uselessly idle blocking waker in more than 1/64 cases.
2097 if (bfqd->last_completed_rq_bfqq !=
2098 bfqq->tentative_waker_bfqq ||
2099 now_ns > bfqq->waker_detection_started +
2100 128 * (u64)bfqd->bfq_slice_idle) {
2102 * First synchronization detected with a
2103 * candidate waker queue, or with a different
2104 * candidate waker queue from the current one.
2106 bfqq->tentative_waker_bfqq =
2107 bfqd->last_completed_rq_bfqq;
2108 bfqq->num_waker_detections = 1;
2109 bfqq->waker_detection_started = now_ns;
2110 bfq_bfqq_name(bfqq->tentative_waker_bfqq, waker_name,
2111 MAX_BFQQ_NAME_LENGTH);
2112 bfq_log_bfqq(bfqd, bfqq, "set tentative waker %s", waker_name);
2113 } else /* Same tentative waker queue detected again */
2114 bfqq->num_waker_detections++;
2116 if (bfqq->num_waker_detections == 3) {
2117 bfqq->waker_bfqq = bfqd->last_completed_rq_bfqq;
2118 bfqq->tentative_waker_bfqq = NULL;
2119 bfq_bfqq_name(bfqq->waker_bfqq, waker_name,
2120 MAX_BFQQ_NAME_LENGTH);
2121 bfq_log_bfqq(bfqd, bfqq, "set waker %s", waker_name);
2124 * If the waker queue disappears, then
2125 * bfqq->waker_bfqq must be reset. To
2126 * this goal, we maintain in each
2127 * waker queue a list, woken_list, of
2128 * all the queues that reference the
2129 * waker queue through their
2130 * waker_bfqq pointer. When the waker
2131 * queue exits, the waker_bfqq pointer
2132 * of all the queues in the woken_list
2135 * In addition, if bfqq is already in
2136 * the woken_list of a waker queue,
2137 * then, before being inserted into
2138 * the woken_list of a new waker
2139 * queue, bfqq must be removed from
2140 * the woken_list of the old waker
2143 if (!hlist_unhashed(&bfqq->woken_list_node))
2144 hlist_del_init(&bfqq->woken_list_node);
2145 hlist_add_head(&bfqq->woken_list_node,
2146 &bfqd->last_completed_rq_bfqq->woken_list);
2150 static void bfq_add_request(struct request *rq)
2152 struct bfq_queue *bfqq = RQ_BFQQ(rq);
2153 struct bfq_data *bfqd = bfqq->bfqd;
2154 struct request *next_rq, *prev;
2155 unsigned int old_wr_coeff = bfqq->wr_coeff;
2156 bool interactive = false;
2157 u64 now_ns = ktime_get_ns();
2159 bfq_log_bfqq(bfqd, bfqq, "add_request %d", rq_is_sync(rq));
2160 bfqq->queued[rq_is_sync(rq)]++;
2162 * Updating of 'bfqd->queued' is protected by 'bfqd->lock', however, it
2163 * may be read without holding the lock in bfq_has_work().
2165 WRITE_ONCE(bfqd->queued, bfqd->queued + 1);
2167 if (bfq_bfqq_sync(bfqq) && RQ_BIC(rq)->requests <= 1) {
2168 bfq_check_waker(bfqd, bfqq, now_ns);
2171 * Periodically reset inject limit, to make sure that
2172 * the latter eventually drops in case workload
2173 * changes, see step (3) in the comments on
2174 * bfq_update_inject_limit().
2176 if (time_is_before_eq_jiffies(bfqq->decrease_time_jif +
2177 msecs_to_jiffies(1000)))
2178 bfq_reset_inject_limit(bfqd, bfqq);
2181 * The following conditions must hold to setup a new
2182 * sampling of total service time, and then a new
2183 * update of the inject limit:
2184 * - bfqq is in service, because the total service
2185 * time is evaluated only for the I/O requests of
2186 * the queues in service;
2187 * - this is the right occasion to compute or to
2188 * lower the baseline total service time, because
2189 * there are actually no requests in the drive,
2191 * the baseline total service time is available, and
2192 * this is the right occasion to compute the other
2193 * quantity needed to update the inject limit, i.e.,
2194 * the total service time caused by the amount of
2195 * injection allowed by the current value of the
2196 * limit. It is the right occasion because injection
2197 * has actually been performed during the service
2198 * hole, and there are still in-flight requests,
2199 * which are very likely to be exactly the injected
2200 * requests, or part of them;
2201 * - the minimum interval for sampling the total
2202 * service time and updating the inject limit has
2205 if (bfqq == bfqd->in_service_queue &&
2206 (bfqd->rq_in_driver == 0 ||
2207 (bfqq->last_serv_time_ns > 0 &&
2208 bfqd->rqs_injected && bfqd->rq_in_driver > 0)) &&
2209 time_is_before_eq_jiffies(bfqq->decrease_time_jif +
2210 msecs_to_jiffies(10))) {
2211 bfqd->last_empty_occupied_ns = ktime_get_ns();
2213 * Start the state machine for measuring the
2214 * total service time of rq: setting
2215 * wait_dispatch will cause bfqd->waited_rq to
2216 * be set when rq will be dispatched.
2218 bfqd->wait_dispatch = true;
2220 * If there is no I/O in service in the drive,
2221 * then possible injection occurred before the
2222 * arrival of rq will not affect the total
2223 * service time of rq. So the injection limit
2224 * must not be updated as a function of such
2225 * total service time, unless new injection
2226 * occurs before rq is completed. To have the
2227 * injection limit updated only in the latter
2228 * case, reset rqs_injected here (rqs_injected
2229 * will be set in case injection is performed
2230 * on bfqq before rq is completed).
2232 if (bfqd->rq_in_driver == 0)
2233 bfqd->rqs_injected = false;
2237 if (bfq_bfqq_sync(bfqq))
2238 bfq_update_io_intensity(bfqq, now_ns);
2240 elv_rb_add(&bfqq->sort_list, rq);
2243 * Check if this request is a better next-serve candidate.
2245 prev = bfqq->next_rq;
2246 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position);
2247 bfqq->next_rq = next_rq;
2250 * Adjust priority tree position, if next_rq changes.
2251 * See comments on bfq_pos_tree_add_move() for the unlikely().
2253 if (unlikely(!bfqd->nonrot_with_queueing && prev != bfqq->next_rq))
2254 bfq_pos_tree_add_move(bfqd, bfqq);
2256 if (!bfq_bfqq_busy(bfqq)) /* switching to busy ... */
2257 bfq_bfqq_handle_idle_busy_switch(bfqd, bfqq, old_wr_coeff,
2260 if (bfqd->low_latency && old_wr_coeff == 1 && !rq_is_sync(rq) &&
2261 time_is_before_jiffies(
2262 bfqq->last_wr_start_finish +
2263 bfqd->bfq_wr_min_inter_arr_async)) {
2264 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
2265 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
2267 bfqd->wr_busy_queues++;
2268 bfqq->entity.prio_changed = 1;
2270 if (prev != bfqq->next_rq)
2271 bfq_updated_next_req(bfqd, bfqq);
2275 * Assign jiffies to last_wr_start_finish in the following
2278 * . if bfqq is not going to be weight-raised, because, for
2279 * non weight-raised queues, last_wr_start_finish stores the
2280 * arrival time of the last request; as of now, this piece
2281 * of information is used only for deciding whether to
2282 * weight-raise async queues
2284 * . if bfqq is not weight-raised, because, if bfqq is now
2285 * switching to weight-raised, then last_wr_start_finish
2286 * stores the time when weight-raising starts
2288 * . if bfqq is interactive, because, regardless of whether
2289 * bfqq is currently weight-raised, the weight-raising
2290 * period must start or restart (this case is considered
2291 * separately because it is not detected by the above
2292 * conditions, if bfqq is already weight-raised)
2294 * last_wr_start_finish has to be updated also if bfqq is soft
2295 * real-time, because the weight-raising period is constantly
2296 * restarted on idle-to-busy transitions for these queues, but
2297 * this is already done in bfq_bfqq_handle_idle_busy_switch if
2300 if (bfqd->low_latency &&
2301 (old_wr_coeff == 1 || bfqq->wr_coeff == 1 || interactive))
2302 bfqq->last_wr_start_finish = jiffies;
2305 static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd,
2307 struct request_queue *q)
2309 struct bfq_queue *bfqq = bfqd->bio_bfqq;
2313 return elv_rb_find(&bfqq->sort_list, bio_end_sector(bio));
2318 static sector_t get_sdist(sector_t last_pos, struct request *rq)
2321 return abs(blk_rq_pos(rq) - last_pos);
2326 static void bfq_remove_request(struct request_queue *q,
2329 struct bfq_queue *bfqq = RQ_BFQQ(rq);
2330 struct bfq_data *bfqd = bfqq->bfqd;
2331 const int sync = rq_is_sync(rq);
2333 if (bfqq->next_rq == rq) {
2334 bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq);
2335 bfq_updated_next_req(bfqd, bfqq);
2338 if (rq->queuelist.prev != &rq->queuelist)
2339 list_del_init(&rq->queuelist);
2340 bfqq->queued[sync]--;
2342 * Updating of 'bfqd->queued' is protected by 'bfqd->lock', however, it
2343 * may be read without holding the lock in bfq_has_work().
2345 WRITE_ONCE(bfqd->queued, bfqd->queued - 1);
2346 elv_rb_del(&bfqq->sort_list, rq);
2348 elv_rqhash_del(q, rq);
2349 if (q->last_merge == rq)
2350 q->last_merge = NULL;
2352 if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
2353 bfqq->next_rq = NULL;
2355 if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->in_service_queue) {
2356 bfq_del_bfqq_busy(bfqq, false);
2358 * bfqq emptied. In normal operation, when
2359 * bfqq is empty, bfqq->entity.service and
2360 * bfqq->entity.budget must contain,
2361 * respectively, the service received and the
2362 * budget used last time bfqq emptied. These
2363 * facts do not hold in this case, as at least
2364 * this last removal occurred while bfqq is
2365 * not in service. To avoid inconsistencies,
2366 * reset both bfqq->entity.service and
2367 * bfqq->entity.budget, if bfqq has still a
2368 * process that may issue I/O requests to it.
2370 bfqq->entity.budget = bfqq->entity.service = 0;
2374 * Remove queue from request-position tree as it is empty.
2376 if (bfqq->pos_root) {
2377 rb_erase(&bfqq->pos_node, bfqq->pos_root);
2378 bfqq->pos_root = NULL;
2381 /* see comments on bfq_pos_tree_add_move() for the unlikely() */
2382 if (unlikely(!bfqd->nonrot_with_queueing))
2383 bfq_pos_tree_add_move(bfqd, bfqq);
2386 if (rq->cmd_flags & REQ_META)
2387 bfqq->meta_pending--;
2391 static bool bfq_bio_merge(struct request_queue *q, struct bio *bio,
2392 unsigned int nr_segs)
2394 struct bfq_data *bfqd = q->elevator->elevator_data;
2395 struct request *free = NULL;
2397 * bfq_bic_lookup grabs the queue_lock: invoke it now and
2398 * store its return value for later use, to avoid nesting
2399 * queue_lock inside the bfqd->lock. We assume that the bic
2400 * returned by bfq_bic_lookup does not go away before
2401 * bfqd->lock is taken.
2403 struct bfq_io_cq *bic = bfq_bic_lookup(q);
2406 spin_lock_irq(&bfqd->lock);
2410 * Make sure cgroup info is uptodate for current process before
2411 * considering the merge.
2413 bfq_bic_update_cgroup(bic, bio);
2415 bfqd->bio_bfqq = bic_to_bfqq(bic, op_is_sync(bio->bi_opf));
2417 bfqd->bio_bfqq = NULL;
2419 bfqd->bio_bic = bic;
2421 ret = blk_mq_sched_try_merge(q, bio, nr_segs, &free);
2423 spin_unlock_irq(&bfqd->lock);
2425 blk_mq_free_request(free);
2430 static int bfq_request_merge(struct request_queue *q, struct request **req,
2433 struct bfq_data *bfqd = q->elevator->elevator_data;
2434 struct request *__rq;
2436 __rq = bfq_find_rq_fmerge(bfqd, bio, q);
2437 if (__rq && elv_bio_merge_ok(__rq, bio)) {
2440 if (blk_discard_mergable(__rq))
2441 return ELEVATOR_DISCARD_MERGE;
2442 return ELEVATOR_FRONT_MERGE;
2445 return ELEVATOR_NO_MERGE;
2448 static void bfq_request_merged(struct request_queue *q, struct request *req,
2449 enum elv_merge type)
2451 if (type == ELEVATOR_FRONT_MERGE &&
2452 rb_prev(&req->rb_node) &&
2454 blk_rq_pos(container_of(rb_prev(&req->rb_node),
2455 struct request, rb_node))) {
2456 struct bfq_queue *bfqq = RQ_BFQQ(req);
2457 struct bfq_data *bfqd;
2458 struct request *prev, *next_rq;
2465 /* Reposition request in its sort_list */
2466 elv_rb_del(&bfqq->sort_list, req);
2467 elv_rb_add(&bfqq->sort_list, req);
2469 /* Choose next request to be served for bfqq */
2470 prev = bfqq->next_rq;
2471 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, req,
2472 bfqd->last_position);
2473 bfqq->next_rq = next_rq;
2475 * If next_rq changes, update both the queue's budget to
2476 * fit the new request and the queue's position in its
2479 if (prev != bfqq->next_rq) {
2480 bfq_updated_next_req(bfqd, bfqq);
2482 * See comments on bfq_pos_tree_add_move() for
2485 if (unlikely(!bfqd->nonrot_with_queueing))
2486 bfq_pos_tree_add_move(bfqd, bfqq);
2492 * This function is called to notify the scheduler that the requests
2493 * rq and 'next' have been merged, with 'next' going away. BFQ
2494 * exploits this hook to address the following issue: if 'next' has a
2495 * fifo_time lower that rq, then the fifo_time of rq must be set to
2496 * the value of 'next', to not forget the greater age of 'next'.
2498 * NOTE: in this function we assume that rq is in a bfq_queue, basing
2499 * on that rq is picked from the hash table q->elevator->hash, which,
2500 * in its turn, is filled only with I/O requests present in
2501 * bfq_queues, while BFQ is in use for the request queue q. In fact,
2502 * the function that fills this hash table (elv_rqhash_add) is called
2503 * only by bfq_insert_request.
2505 static void bfq_requests_merged(struct request_queue *q, struct request *rq,
2506 struct request *next)
2508 struct bfq_queue *bfqq = RQ_BFQQ(rq),
2509 *next_bfqq = RQ_BFQQ(next);
2515 * If next and rq belong to the same bfq_queue and next is older
2516 * than rq, then reposition rq in the fifo (by substituting next
2517 * with rq). Otherwise, if next and rq belong to different
2518 * bfq_queues, never reposition rq: in fact, we would have to
2519 * reposition it with respect to next's position in its own fifo,
2520 * which would most certainly be too expensive with respect to
2523 if (bfqq == next_bfqq &&
2524 !list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2525 next->fifo_time < rq->fifo_time) {
2526 list_del_init(&rq->queuelist);
2527 list_replace_init(&next->queuelist, &rq->queuelist);
2528 rq->fifo_time = next->fifo_time;
2531 if (bfqq->next_rq == next)
2534 bfqg_stats_update_io_merged(bfqq_group(bfqq), next->cmd_flags);
2536 /* Merged request may be in the IO scheduler. Remove it. */
2537 if (!RB_EMPTY_NODE(&next->rb_node)) {
2538 bfq_remove_request(next->q, next);
2540 bfqg_stats_update_io_remove(bfqq_group(next_bfqq),
2545 /* Must be called with bfqq != NULL */
2546 static void bfq_bfqq_end_wr(struct bfq_queue *bfqq)
2549 * If bfqq has been enjoying interactive weight-raising, then
2550 * reset soft_rt_next_start. We do it for the following
2551 * reason. bfqq may have been conveying the I/O needed to load
2552 * a soft real-time application. Such an application actually
2553 * exhibits a soft real-time I/O pattern after it finishes
2554 * loading, and finally starts doing its job. But, if bfqq has
2555 * been receiving a lot of bandwidth so far (likely to happen
2556 * on a fast device), then soft_rt_next_start now contains a
2557 * high value that. So, without this reset, bfqq would be
2558 * prevented from being possibly considered as soft_rt for a
2562 if (bfqq->wr_cur_max_time !=
2563 bfqq->bfqd->bfq_wr_rt_max_time)
2564 bfqq->soft_rt_next_start = jiffies;
2566 if (bfq_bfqq_busy(bfqq))
2567 bfqq->bfqd->wr_busy_queues--;
2569 bfqq->wr_cur_max_time = 0;
2570 bfqq->last_wr_start_finish = jiffies;
2572 * Trigger a weight change on the next invocation of
2573 * __bfq_entity_update_weight_prio.
2575 bfqq->entity.prio_changed = 1;
2578 void bfq_end_wr_async_queues(struct bfq_data *bfqd,
2579 struct bfq_group *bfqg)
2583 for (i = 0; i < 2; i++)
2584 for (j = 0; j < IOPRIO_NR_LEVELS; j++)
2585 if (bfqg->async_bfqq[i][j])
2586 bfq_bfqq_end_wr(bfqg->async_bfqq[i][j]);
2587 if (bfqg->async_idle_bfqq)
2588 bfq_bfqq_end_wr(bfqg->async_idle_bfqq);
2591 static void bfq_end_wr(struct bfq_data *bfqd)
2593 struct bfq_queue *bfqq;
2595 spin_lock_irq(&bfqd->lock);
2597 list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
2598 bfq_bfqq_end_wr(bfqq);
2599 list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list)
2600 bfq_bfqq_end_wr(bfqq);
2601 bfq_end_wr_async(bfqd);
2603 spin_unlock_irq(&bfqd->lock);
2606 static sector_t bfq_io_struct_pos(void *io_struct, bool request)
2609 return blk_rq_pos(io_struct);
2611 return ((struct bio *)io_struct)->bi_iter.bi_sector;
2614 static int bfq_rq_close_to_sector(void *io_struct, bool request,
2617 return abs(bfq_io_struct_pos(io_struct, request) - sector) <=
2621 static struct bfq_queue *bfqq_find_close(struct bfq_data *bfqd,
2622 struct bfq_queue *bfqq,
2625 struct rb_root *root = &bfqq_group(bfqq)->rq_pos_tree;
2626 struct rb_node *parent, *node;
2627 struct bfq_queue *__bfqq;
2629 if (RB_EMPTY_ROOT(root))
2633 * First, if we find a request starting at the end of the last
2634 * request, choose it.
2636 __bfqq = bfq_rq_pos_tree_lookup(bfqd, root, sector, &parent, NULL);
2641 * If the exact sector wasn't found, the parent of the NULL leaf
2642 * will contain the closest sector (rq_pos_tree sorted by
2643 * next_request position).
2645 __bfqq = rb_entry(parent, struct bfq_queue, pos_node);
2646 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2649 if (blk_rq_pos(__bfqq->next_rq) < sector)
2650 node = rb_next(&__bfqq->pos_node);
2652 node = rb_prev(&__bfqq->pos_node);
2656 __bfqq = rb_entry(node, struct bfq_queue, pos_node);
2657 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2663 static struct bfq_queue *bfq_find_close_cooperator(struct bfq_data *bfqd,
2664 struct bfq_queue *cur_bfqq,
2667 struct bfq_queue *bfqq;
2670 * We shall notice if some of the queues are cooperating,
2671 * e.g., working closely on the same area of the device. In
2672 * that case, we can group them together and: 1) don't waste
2673 * time idling, and 2) serve the union of their requests in
2674 * the best possible order for throughput.
2676 bfqq = bfqq_find_close(bfqd, cur_bfqq, sector);
2677 if (!bfqq || bfqq == cur_bfqq)
2683 static struct bfq_queue *
2684 bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2686 int process_refs, new_process_refs;
2687 struct bfq_queue *__bfqq;
2690 * If there are no process references on the new_bfqq, then it is
2691 * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain
2692 * may have dropped their last reference (not just their last process
2695 if (!bfqq_process_refs(new_bfqq))
2698 /* Avoid a circular list and skip interim queue merges. */
2699 while ((__bfqq = new_bfqq->new_bfqq)) {
2705 process_refs = bfqq_process_refs(bfqq);
2706 new_process_refs = bfqq_process_refs(new_bfqq);
2708 * If the process for the bfqq has gone away, there is no
2709 * sense in merging the queues.
2711 if (process_refs == 0 || new_process_refs == 0)
2715 * Make sure merged queues belong to the same parent. Parents could
2716 * have changed since the time we decided the two queues are suitable
2719 if (new_bfqq->entity.parent != bfqq->entity.parent)
2722 bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d",
2726 * Merging is just a redirection: the requests of the process
2727 * owning one of the two queues are redirected to the other queue.
2728 * The latter queue, in its turn, is set as shared if this is the
2729 * first time that the requests of some process are redirected to
2732 * We redirect bfqq to new_bfqq and not the opposite, because
2733 * we are in the context of the process owning bfqq, thus we
2734 * have the io_cq of this process. So we can immediately
2735 * configure this io_cq to redirect the requests of the
2736 * process to new_bfqq. In contrast, the io_cq of new_bfqq is
2737 * not available any more (new_bfqq->bic == NULL).
2739 * Anyway, even in case new_bfqq coincides with the in-service
2740 * queue, redirecting requests the in-service queue is the
2741 * best option, as we feed the in-service queue with new
2742 * requests close to the last request served and, by doing so,
2743 * are likely to increase the throughput.
2745 bfqq->new_bfqq = new_bfqq;
2747 * The above assignment schedules the following redirections:
2748 * each time some I/O for bfqq arrives, the process that
2749 * generated that I/O is disassociated from bfqq and
2750 * associated with new_bfqq. Here we increases new_bfqq->ref
2751 * in advance, adding the number of processes that are
2752 * expected to be associated with new_bfqq as they happen to
2755 new_bfqq->ref += process_refs;
2759 static bool bfq_may_be_close_cooperator(struct bfq_queue *bfqq,
2760 struct bfq_queue *new_bfqq)
2762 if (bfq_too_late_for_merging(new_bfqq))
2765 if (bfq_class_idle(bfqq) || bfq_class_idle(new_bfqq) ||
2766 (bfqq->ioprio_class != new_bfqq->ioprio_class))
2770 * If either of the queues has already been detected as seeky,
2771 * then merging it with the other queue is unlikely to lead to
2774 if (BFQQ_SEEKY(bfqq) || BFQQ_SEEKY(new_bfqq))
2778 * Interleaved I/O is known to be done by (some) applications
2779 * only for reads, so it does not make sense to merge async
2782 if (!bfq_bfqq_sync(bfqq) || !bfq_bfqq_sync(new_bfqq))
2788 static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd,
2789 struct bfq_queue *bfqq);
2792 * Attempt to schedule a merge of bfqq with the currently in-service
2793 * queue or with a close queue among the scheduled queues. Return
2794 * NULL if no merge was scheduled, a pointer to the shared bfq_queue
2795 * structure otherwise.
2797 * The OOM queue is not allowed to participate to cooperation: in fact, since
2798 * the requests temporarily redirected to the OOM queue could be redirected
2799 * again to dedicated queues at any time, the state needed to correctly
2800 * handle merging with the OOM queue would be quite complex and expensive
2801 * to maintain. Besides, in such a critical condition as an out of memory,
2802 * the benefits of queue merging may be little relevant, or even negligible.
2804 * WARNING: queue merging may impair fairness among non-weight raised
2805 * queues, for at least two reasons: 1) the original weight of a
2806 * merged queue may change during the merged state, 2) even being the
2807 * weight the same, a merged queue may be bloated with many more
2808 * requests than the ones produced by its originally-associated
2811 static struct bfq_queue *
2812 bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2813 void *io_struct, bool request, struct bfq_io_cq *bic)
2815 struct bfq_queue *in_service_bfqq, *new_bfqq;
2817 /* if a merge has already been setup, then proceed with that first */
2819 return bfqq->new_bfqq;
2822 * Check delayed stable merge for rotational or non-queueing
2823 * devs. For this branch to be executed, bfqq must not be
2824 * currently merged with some other queue (i.e., bfqq->bic
2825 * must be non null). If we considered also merged queues,
2826 * then we should also check whether bfqq has already been
2827 * merged with bic->stable_merge_bfqq. But this would be
2828 * costly and complicated.
2830 if (unlikely(!bfqd->nonrot_with_queueing)) {
2832 * Make sure also that bfqq is sync, because
2833 * bic->stable_merge_bfqq may point to some queue (for
2834 * stable merging) also if bic is associated with a
2835 * sync queue, but this bfqq is async
2837 if (bfq_bfqq_sync(bfqq) && bic->stable_merge_bfqq &&
2838 !bfq_bfqq_just_created(bfqq) &&
2839 time_is_before_jiffies(bfqq->split_time +
2840 msecs_to_jiffies(bfq_late_stable_merging)) &&
2841 time_is_before_jiffies(bfqq->creation_time +
2842 msecs_to_jiffies(bfq_late_stable_merging))) {
2843 struct bfq_queue *stable_merge_bfqq =
2844 bic->stable_merge_bfqq;
2845 int proc_ref = min(bfqq_process_refs(bfqq),
2846 bfqq_process_refs(stable_merge_bfqq));
2848 /* deschedule stable merge, because done or aborted here */
2849 bfq_put_stable_ref(stable_merge_bfqq);
2851 bic->stable_merge_bfqq = NULL;
2853 if (!idling_boosts_thr_without_issues(bfqd, bfqq) &&
2855 /* next function will take at least one ref */
2856 struct bfq_queue *new_bfqq =
2857 bfq_setup_merge(bfqq, stable_merge_bfqq);
2860 bic->stably_merged = true;
2862 new_bfqq->bic->stably_merged =
2872 * Do not perform queue merging if the device is non
2873 * rotational and performs internal queueing. In fact, such a
2874 * device reaches a high speed through internal parallelism
2875 * and pipelining. This means that, to reach a high
2876 * throughput, it must have many requests enqueued at the same
2877 * time. But, in this configuration, the internal scheduling
2878 * algorithm of the device does exactly the job of queue
2879 * merging: it reorders requests so as to obtain as much as
2880 * possible a sequential I/O pattern. As a consequence, with
2881 * the workload generated by processes doing interleaved I/O,
2882 * the throughput reached by the device is likely to be the
2883 * same, with and without queue merging.
2885 * Disabling merging also provides a remarkable benefit in
2886 * terms of throughput. Merging tends to make many workloads
2887 * artificially more uneven, because of shared queues
2888 * remaining non empty for incomparably more time than
2889 * non-merged queues. This may accentuate workload
2890 * asymmetries. For example, if one of the queues in a set of
2891 * merged queues has a higher weight than a normal queue, then
2892 * the shared queue may inherit such a high weight and, by
2893 * staying almost always active, may force BFQ to perform I/O
2894 * plugging most of the time. This evidently makes it harder
2895 * for BFQ to let the device reach a high throughput.
2897 * Finally, the likely() macro below is not used because one
2898 * of the two branches is more likely than the other, but to
2899 * have the code path after the following if() executed as
2900 * fast as possible for the case of a non rotational device
2901 * with queueing. We want it because this is the fastest kind
2902 * of device. On the opposite end, the likely() may lengthen
2903 * the execution time of BFQ for the case of slower devices
2904 * (rotational or at least without queueing). But in this case
2905 * the execution time of BFQ matters very little, if not at
2908 if (likely(bfqd->nonrot_with_queueing))
2912 * Prevent bfqq from being merged if it has been created too
2913 * long ago. The idea is that true cooperating processes, and
2914 * thus their associated bfq_queues, are supposed to be
2915 * created shortly after each other. This is the case, e.g.,
2916 * for KVM/QEMU and dump I/O threads. Basing on this
2917 * assumption, the following filtering greatly reduces the
2918 * probability that two non-cooperating processes, which just
2919 * happen to do close I/O for some short time interval, have
2920 * their queues merged by mistake.
2922 if (bfq_too_late_for_merging(bfqq))
2925 if (!io_struct || unlikely(bfqq == &bfqd->oom_bfqq))
2928 /* If there is only one backlogged queue, don't search. */
2929 if (bfq_tot_busy_queues(bfqd) == 1)
2932 in_service_bfqq = bfqd->in_service_queue;
2934 if (in_service_bfqq && in_service_bfqq != bfqq &&
2935 likely(in_service_bfqq != &bfqd->oom_bfqq) &&
2936 bfq_rq_close_to_sector(io_struct, request,
2937 bfqd->in_serv_last_pos) &&
2938 bfqq->entity.parent == in_service_bfqq->entity.parent &&
2939 bfq_may_be_close_cooperator(bfqq, in_service_bfqq)) {
2940 new_bfqq = bfq_setup_merge(bfqq, in_service_bfqq);
2945 * Check whether there is a cooperator among currently scheduled
2946 * queues. The only thing we need is that the bio/request is not
2947 * NULL, as we need it to establish whether a cooperator exists.
2949 new_bfqq = bfq_find_close_cooperator(bfqd, bfqq,
2950 bfq_io_struct_pos(io_struct, request));
2952 if (new_bfqq && likely(new_bfqq != &bfqd->oom_bfqq) &&
2953 bfq_may_be_close_cooperator(bfqq, new_bfqq))
2954 return bfq_setup_merge(bfqq, new_bfqq);
2959 static void bfq_bfqq_save_state(struct bfq_queue *bfqq)
2961 struct bfq_io_cq *bic = bfqq->bic;
2964 * If !bfqq->bic, the queue is already shared or its requests
2965 * have already been redirected to a shared queue; both idle window
2966 * and weight raising state have already been saved. Do nothing.
2971 bic->saved_last_serv_time_ns = bfqq->last_serv_time_ns;
2972 bic->saved_inject_limit = bfqq->inject_limit;
2973 bic->saved_decrease_time_jif = bfqq->decrease_time_jif;
2975 bic->saved_weight = bfqq->entity.orig_weight;
2976 bic->saved_ttime = bfqq->ttime;
2977 bic->saved_has_short_ttime = bfq_bfqq_has_short_ttime(bfqq);
2978 bic->saved_IO_bound = bfq_bfqq_IO_bound(bfqq);
2979 bic->saved_io_start_time = bfqq->io_start_time;
2980 bic->saved_tot_idle_time = bfqq->tot_idle_time;
2981 bic->saved_in_large_burst = bfq_bfqq_in_large_burst(bfqq);
2982 bic->was_in_burst_list = !hlist_unhashed(&bfqq->burst_list_node);
2983 if (unlikely(bfq_bfqq_just_created(bfqq) &&
2984 !bfq_bfqq_in_large_burst(bfqq) &&
2985 bfqq->bfqd->low_latency)) {
2987 * bfqq being merged right after being created: bfqq
2988 * would have deserved interactive weight raising, but
2989 * did not make it to be set in a weight-raised state,
2990 * because of this early merge. Store directly the
2991 * weight-raising state that would have been assigned
2992 * to bfqq, so that to avoid that bfqq unjustly fails
2993 * to enjoy weight raising if split soon.
2995 bic->saved_wr_coeff = bfqq->bfqd->bfq_wr_coeff;
2996 bic->saved_wr_start_at_switch_to_srt = bfq_smallest_from_now();
2997 bic->saved_wr_cur_max_time = bfq_wr_duration(bfqq->bfqd);
2998 bic->saved_last_wr_start_finish = jiffies;
3000 bic->saved_wr_coeff = bfqq->wr_coeff;
3001 bic->saved_wr_start_at_switch_to_srt =
3002 bfqq->wr_start_at_switch_to_srt;
3003 bic->saved_service_from_wr = bfqq->service_from_wr;
3004 bic->saved_last_wr_start_finish = bfqq->last_wr_start_finish;
3005 bic->saved_wr_cur_max_time = bfqq->wr_cur_max_time;
3011 bfq_reassign_last_bfqq(struct bfq_queue *cur_bfqq, struct bfq_queue *new_bfqq)
3013 if (cur_bfqq->entity.parent &&
3014 cur_bfqq->entity.parent->last_bfqq_created == cur_bfqq)
3015 cur_bfqq->entity.parent->last_bfqq_created = new_bfqq;
3016 else if (cur_bfqq->bfqd && cur_bfqq->bfqd->last_bfqq_created == cur_bfqq)
3017 cur_bfqq->bfqd->last_bfqq_created = new_bfqq;
3020 void bfq_release_process_ref(struct bfq_data *bfqd, struct bfq_queue *bfqq)
3023 * To prevent bfqq's service guarantees from being violated,
3024 * bfqq may be left busy, i.e., queued for service, even if
3025 * empty (see comments in __bfq_bfqq_expire() for
3026 * details). But, if no process will send requests to bfqq any
3027 * longer, then there is no point in keeping bfqq queued for
3028 * service. In addition, keeping bfqq queued for service, but
3029 * with no process ref any longer, may have caused bfqq to be
3030 * freed when dequeued from service. But this is assumed to
3033 if (bfq_bfqq_busy(bfqq) && RB_EMPTY_ROOT(&bfqq->sort_list) &&
3034 bfqq != bfqd->in_service_queue)
3035 bfq_del_bfqq_busy(bfqq, false);
3037 bfq_reassign_last_bfqq(bfqq, NULL);
3039 bfq_put_queue(bfqq);
3043 bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic,
3044 struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
3046 bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu",
3047 (unsigned long)new_bfqq->pid);
3048 /* Save weight raising and idle window of the merged queues */
3049 bfq_bfqq_save_state(bfqq);
3050 bfq_bfqq_save_state(new_bfqq);
3051 if (bfq_bfqq_IO_bound(bfqq))
3052 bfq_mark_bfqq_IO_bound(new_bfqq);
3053 bfq_clear_bfqq_IO_bound(bfqq);
3056 * The processes associated with bfqq are cooperators of the
3057 * processes associated with new_bfqq. So, if bfqq has a
3058 * waker, then assume that all these processes will be happy
3059 * to let bfqq's waker freely inject I/O when they have no
3062 if (bfqq->waker_bfqq && !new_bfqq->waker_bfqq &&
3063 bfqq->waker_bfqq != new_bfqq) {
3064 new_bfqq->waker_bfqq = bfqq->waker_bfqq;
3065 new_bfqq->tentative_waker_bfqq = NULL;
3068 * If the waker queue disappears, then
3069 * new_bfqq->waker_bfqq must be reset. So insert
3070 * new_bfqq into the woken_list of the waker. See
3071 * bfq_check_waker for details.
3073 hlist_add_head(&new_bfqq->woken_list_node,
3074 &new_bfqq->waker_bfqq->woken_list);
3079 * If bfqq is weight-raised, then let new_bfqq inherit
3080 * weight-raising. To reduce false positives, neglect the case
3081 * where bfqq has just been created, but has not yet made it
3082 * to be weight-raised (which may happen because EQM may merge
3083 * bfqq even before bfq_add_request is executed for the first
3084 * time for bfqq). Handling this case would however be very
3085 * easy, thanks to the flag just_created.
3087 if (new_bfqq->wr_coeff == 1 && bfqq->wr_coeff > 1) {
3088 new_bfqq->wr_coeff = bfqq->wr_coeff;
3089 new_bfqq->wr_cur_max_time = bfqq->wr_cur_max_time;
3090 new_bfqq->last_wr_start_finish = bfqq->last_wr_start_finish;
3091 new_bfqq->wr_start_at_switch_to_srt =
3092 bfqq->wr_start_at_switch_to_srt;
3093 if (bfq_bfqq_busy(new_bfqq))
3094 bfqd->wr_busy_queues++;
3095 new_bfqq->entity.prio_changed = 1;
3098 if (bfqq->wr_coeff > 1) { /* bfqq has given its wr to new_bfqq */
3100 bfqq->entity.prio_changed = 1;
3101 if (bfq_bfqq_busy(bfqq))
3102 bfqd->wr_busy_queues--;
3105 bfq_log_bfqq(bfqd, new_bfqq, "merge_bfqqs: wr_busy %d",
3106 bfqd->wr_busy_queues);
3109 * Merge queues (that is, let bic redirect its requests to new_bfqq)
3111 bic_set_bfqq(bic, new_bfqq, 1);
3112 bfq_mark_bfqq_coop(new_bfqq);
3114 * new_bfqq now belongs to at least two bics (it is a shared queue):
3115 * set new_bfqq->bic to NULL. bfqq either:
3116 * - does not belong to any bic any more, and hence bfqq->bic must
3117 * be set to NULL, or
3118 * - is a queue whose owning bics have already been redirected to a
3119 * different queue, hence the queue is destined to not belong to
3120 * any bic soon and bfqq->bic is already NULL (therefore the next
3121 * assignment causes no harm).
3123 new_bfqq->bic = NULL;
3125 * If the queue is shared, the pid is the pid of one of the associated
3126 * processes. Which pid depends on the exact sequence of merge events
3127 * the queue underwent. So printing such a pid is useless and confusing
3128 * because it reports a random pid between those of the associated
3130 * We mark such a queue with a pid -1, and then print SHARED instead of
3131 * a pid in logging messages.
3136 bfq_reassign_last_bfqq(bfqq, new_bfqq);
3138 bfq_release_process_ref(bfqd, bfqq);
3141 static bool bfq_allow_bio_merge(struct request_queue *q, struct request *rq,
3144 struct bfq_data *bfqd = q->elevator->elevator_data;
3145 bool is_sync = op_is_sync(bio->bi_opf);
3146 struct bfq_queue *bfqq = bfqd->bio_bfqq, *new_bfqq;
3149 * Disallow merge of a sync bio into an async request.
3151 if (is_sync && !rq_is_sync(rq))
3155 * Lookup the bfqq that this bio will be queued with. Allow
3156 * merge only if rq is queued there.
3162 * We take advantage of this function to perform an early merge
3163 * of the queues of possible cooperating processes.
3165 new_bfqq = bfq_setup_cooperator(bfqd, bfqq, bio, false, bfqd->bio_bic);
3168 * bic still points to bfqq, then it has not yet been
3169 * redirected to some other bfq_queue, and a queue
3170 * merge between bfqq and new_bfqq can be safely
3171 * fulfilled, i.e., bic can be redirected to new_bfqq
3172 * and bfqq can be put.
3174 bfq_merge_bfqqs(bfqd, bfqd->bio_bic, bfqq,
3177 * If we get here, bio will be queued into new_queue,
3178 * so use new_bfqq to decide whether bio and rq can be
3184 * Change also bqfd->bio_bfqq, as
3185 * bfqd->bio_bic now points to new_bfqq, and
3186 * this function may be invoked again (and then may
3187 * use again bqfd->bio_bfqq).
3189 bfqd->bio_bfqq = bfqq;
3192 return bfqq == RQ_BFQQ(rq);
3196 * Set the maximum time for the in-service queue to consume its
3197 * budget. This prevents seeky processes from lowering the throughput.
3198 * In practice, a time-slice service scheme is used with seeky
3201 static void bfq_set_budget_timeout(struct bfq_data *bfqd,
3202 struct bfq_queue *bfqq)
3204 unsigned int timeout_coeff;
3206 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time)
3209 timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight;
3211 bfqd->last_budget_start = ktime_get();
3213 bfqq->budget_timeout = jiffies +
3214 bfqd->bfq_timeout * timeout_coeff;
3217 static void __bfq_set_in_service_queue(struct bfq_data *bfqd,
3218 struct bfq_queue *bfqq)
3221 bfq_clear_bfqq_fifo_expire(bfqq);
3223 bfqd->budgets_assigned = (bfqd->budgets_assigned * 7 + 256) / 8;
3225 if (time_is_before_jiffies(bfqq->last_wr_start_finish) &&
3226 bfqq->wr_coeff > 1 &&
3227 bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
3228 time_is_before_jiffies(bfqq->budget_timeout)) {
3230 * For soft real-time queues, move the start
3231 * of the weight-raising period forward by the
3232 * time the queue has not received any
3233 * service. Otherwise, a relatively long
3234 * service delay is likely to cause the
3235 * weight-raising period of the queue to end,
3236 * because of the short duration of the
3237 * weight-raising period of a soft real-time
3238 * queue. It is worth noting that this move
3239 * is not so dangerous for the other queues,
3240 * because soft real-time queues are not
3243 * To not add a further variable, we use the
3244 * overloaded field budget_timeout to
3245 * determine for how long the queue has not
3246 * received service, i.e., how much time has
3247 * elapsed since the queue expired. However,
3248 * this is a little imprecise, because
3249 * budget_timeout is set to jiffies if bfqq
3250 * not only expires, but also remains with no
3253 if (time_after(bfqq->budget_timeout,
3254 bfqq->last_wr_start_finish))
3255 bfqq->last_wr_start_finish +=
3256 jiffies - bfqq->budget_timeout;
3258 bfqq->last_wr_start_finish = jiffies;
3261 bfq_set_budget_timeout(bfqd, bfqq);
3262 bfq_log_bfqq(bfqd, bfqq,
3263 "set_in_service_queue, cur-budget = %d",
3264 bfqq->entity.budget);
3267 bfqd->in_service_queue = bfqq;
3268 bfqd->in_serv_last_pos = 0;
3272 * Get and set a new queue for service.
3274 static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd)
3276 struct bfq_queue *bfqq = bfq_get_next_queue(bfqd);
3278 __bfq_set_in_service_queue(bfqd, bfqq);
3282 static void bfq_arm_slice_timer(struct bfq_data *bfqd)
3284 struct bfq_queue *bfqq = bfqd->in_service_queue;
3287 bfq_mark_bfqq_wait_request(bfqq);
3290 * We don't want to idle for seeks, but we do want to allow
3291 * fair distribution of slice time for a process doing back-to-back
3292 * seeks. So allow a little bit of time for him to submit a new rq.
3294 sl = bfqd->bfq_slice_idle;
3296 * Unless the queue is being weight-raised or the scenario is
3297 * asymmetric, grant only minimum idle time if the queue
3298 * is seeky. A long idling is preserved for a weight-raised
3299 * queue, or, more in general, in an asymmetric scenario,
3300 * because a long idling is needed for guaranteeing to a queue
3301 * its reserved share of the throughput (in particular, it is
3302 * needed if the queue has a higher weight than some other
3305 if (BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 &&
3306 !bfq_asymmetric_scenario(bfqd, bfqq))
3307 sl = min_t(u64, sl, BFQ_MIN_TT);
3308 else if (bfqq->wr_coeff > 1)
3309 sl = max_t(u32, sl, 20ULL * NSEC_PER_MSEC);
3311 bfqd->last_idling_start = ktime_get();
3312 bfqd->last_idling_start_jiffies = jiffies;
3314 hrtimer_start(&bfqd->idle_slice_timer, ns_to_ktime(sl),
3316 bfqg_stats_set_start_idle_time(bfqq_group(bfqq));
3320 * In autotuning mode, max_budget is dynamically recomputed as the
3321 * amount of sectors transferred in timeout at the estimated peak
3322 * rate. This enables BFQ to utilize a full timeslice with a full
3323 * budget, even if the in-service queue is served at peak rate. And
3324 * this maximises throughput with sequential workloads.
3326 static unsigned long bfq_calc_max_budget(struct bfq_data *bfqd)
3328 return (u64)bfqd->peak_rate * USEC_PER_MSEC *
3329 jiffies_to_msecs(bfqd->bfq_timeout)>>BFQ_RATE_SHIFT;
3333 * Update parameters related to throughput and responsiveness, as a
3334 * function of the estimated peak rate. See comments on
3335 * bfq_calc_max_budget(), and on the ref_wr_duration array.
3337 static void update_thr_responsiveness_params(struct bfq_data *bfqd)
3339 if (bfqd->bfq_user_max_budget == 0) {
3340 bfqd->bfq_max_budget =
3341 bfq_calc_max_budget(bfqd);
3342 bfq_log(bfqd, "new max_budget = %d", bfqd->bfq_max_budget);
3346 static void bfq_reset_rate_computation(struct bfq_data *bfqd,
3349 if (rq != NULL) { /* new rq dispatch now, reset accordingly */
3350 bfqd->last_dispatch = bfqd->first_dispatch = ktime_get_ns();
3351 bfqd->peak_rate_samples = 1;
3352 bfqd->sequential_samples = 0;
3353 bfqd->tot_sectors_dispatched = bfqd->last_rq_max_size =
3355 } else /* no new rq dispatched, just reset the number of samples */
3356 bfqd->peak_rate_samples = 0; /* full re-init on next disp. */
3359 "reset_rate_computation at end, sample %u/%u tot_sects %llu",
3360 bfqd->peak_rate_samples, bfqd->sequential_samples,
3361 bfqd->tot_sectors_dispatched);
3364 static void bfq_update_rate_reset(struct bfq_data *bfqd, struct request *rq)
3366 u32 rate, weight, divisor;
3369 * For the convergence property to hold (see comments on
3370 * bfq_update_peak_rate()) and for the assessment to be
3371 * reliable, a minimum number of samples must be present, and
3372 * a minimum amount of time must have elapsed. If not so, do
3373 * not compute new rate. Just reset parameters, to get ready
3374 * for a new evaluation attempt.
3376 if (bfqd->peak_rate_samples < BFQ_RATE_MIN_SAMPLES ||
3377 bfqd->delta_from_first < BFQ_RATE_MIN_INTERVAL)
3378 goto reset_computation;
3381 * If a new request completion has occurred after last
3382 * dispatch, then, to approximate the rate at which requests
3383 * have been served by the device, it is more precise to
3384 * extend the observation interval to the last completion.
3386 bfqd->delta_from_first =
3387 max_t(u64, bfqd->delta_from_first,
3388 bfqd->last_completion - bfqd->first_dispatch);
3391 * Rate computed in sects/usec, and not sects/nsec, for
3394 rate = div64_ul(bfqd->tot_sectors_dispatched<<BFQ_RATE_SHIFT,
3395 div_u64(bfqd->delta_from_first, NSEC_PER_USEC));
3398 * Peak rate not updated if:
3399 * - the percentage of sequential dispatches is below 3/4 of the
3400 * total, and rate is below the current estimated peak rate
3401 * - rate is unreasonably high (> 20M sectors/sec)
3403 if ((bfqd->sequential_samples < (3 * bfqd->peak_rate_samples)>>2 &&
3404 rate <= bfqd->peak_rate) ||
3405 rate > 20<<BFQ_RATE_SHIFT)
3406 goto reset_computation;
3409 * We have to update the peak rate, at last! To this purpose,
3410 * we use a low-pass filter. We compute the smoothing constant
3411 * of the filter as a function of the 'weight' of the new
3414 * As can be seen in next formulas, we define this weight as a
3415 * quantity proportional to how sequential the workload is,
3416 * and to how long the observation time interval is.
3418 * The weight runs from 0 to 8. The maximum value of the
3419 * weight, 8, yields the minimum value for the smoothing
3420 * constant. At this minimum value for the smoothing constant,
3421 * the measured rate contributes for half of the next value of
3422 * the estimated peak rate.
3424 * So, the first step is to compute the weight as a function
3425 * of how sequential the workload is. Note that the weight
3426 * cannot reach 9, because bfqd->sequential_samples cannot
3427 * become equal to bfqd->peak_rate_samples, which, in its
3428 * turn, holds true because bfqd->sequential_samples is not
3429 * incremented for the first sample.
3431 weight = (9 * bfqd->sequential_samples) / bfqd->peak_rate_samples;
3434 * Second step: further refine the weight as a function of the
3435 * duration of the observation interval.
3437 weight = min_t(u32, 8,
3438 div_u64(weight * bfqd->delta_from_first,
3439 BFQ_RATE_REF_INTERVAL));
3442 * Divisor ranging from 10, for minimum weight, to 2, for
3445 divisor = 10 - weight;
3448 * Finally, update peak rate:
3450 * peak_rate = peak_rate * (divisor-1) / divisor + rate / divisor
3452 bfqd->peak_rate *= divisor-1;
3453 bfqd->peak_rate /= divisor;
3454 rate /= divisor; /* smoothing constant alpha = 1/divisor */
3456 bfqd->peak_rate += rate;
3459 * For a very slow device, bfqd->peak_rate can reach 0 (see
3460 * the minimum representable values reported in the comments
3461 * on BFQ_RATE_SHIFT). Push to 1 if this happens, to avoid
3462 * divisions by zero where bfqd->peak_rate is used as a
3465 bfqd->peak_rate = max_t(u32, 1, bfqd->peak_rate);
3467 update_thr_responsiveness_params(bfqd);
3470 bfq_reset_rate_computation(bfqd, rq);
3474 * Update the read/write peak rate (the main quantity used for
3475 * auto-tuning, see update_thr_responsiveness_params()).
3477 * It is not trivial to estimate the peak rate (correctly): because of
3478 * the presence of sw and hw queues between the scheduler and the
3479 * device components that finally serve I/O requests, it is hard to
3480 * say exactly when a given dispatched request is served inside the
3481 * device, and for how long. As a consequence, it is hard to know
3482 * precisely at what rate a given set of requests is actually served
3485 * On the opposite end, the dispatch time of any request is trivially
3486 * available, and, from this piece of information, the "dispatch rate"
3487 * of requests can be immediately computed. So, the idea in the next
3488 * function is to use what is known, namely request dispatch times
3489 * (plus, when useful, request completion times), to estimate what is
3490 * unknown, namely in-device request service rate.
3492 * The main issue is that, because of the above facts, the rate at
3493 * which a certain set of requests is dispatched over a certain time
3494 * interval can vary greatly with respect to the rate at which the
3495 * same requests are then served. But, since the size of any
3496 * intermediate queue is limited, and the service scheme is lossless
3497 * (no request is silently dropped), the following obvious convergence
3498 * property holds: the number of requests dispatched MUST become
3499 * closer and closer to the number of requests completed as the
3500 * observation interval grows. This is the key property used in
3501 * the next function to estimate the peak service rate as a function
3502 * of the observed dispatch rate. The function assumes to be invoked
3503 * on every request dispatch.
3505 static void bfq_update_peak_rate(struct bfq_data *bfqd, struct request *rq)
3507 u64 now_ns = ktime_get_ns();
3509 if (bfqd->peak_rate_samples == 0) { /* first dispatch */
3510 bfq_log(bfqd, "update_peak_rate: goto reset, samples %d",
3511 bfqd->peak_rate_samples);
3512 bfq_reset_rate_computation(bfqd, rq);
3513 goto update_last_values; /* will add one sample */
3517 * Device idle for very long: the observation interval lasting
3518 * up to this dispatch cannot be a valid observation interval
3519 * for computing a new peak rate (similarly to the late-
3520 * completion event in bfq_completed_request()). Go to
3521 * update_rate_and_reset to have the following three steps
3523 * - close the observation interval at the last (previous)
3524 * request dispatch or completion
3525 * - compute rate, if possible, for that observation interval
3526 * - start a new observation interval with this dispatch
3528 if (now_ns - bfqd->last_dispatch > 100*NSEC_PER_MSEC &&
3529 bfqd->rq_in_driver == 0)
3530 goto update_rate_and_reset;
3532 /* Update sampling information */
3533 bfqd->peak_rate_samples++;
3535 if ((bfqd->rq_in_driver > 0 ||
3536 now_ns - bfqd->last_completion < BFQ_MIN_TT)
3537 && !BFQ_RQ_SEEKY(bfqd, bfqd->last_position, rq))
3538 bfqd->sequential_samples++;
3540 bfqd->tot_sectors_dispatched += blk_rq_sectors(rq);
3542 /* Reset max observed rq size every 32 dispatches */
3543 if (likely(bfqd->peak_rate_samples % 32))
3544 bfqd->last_rq_max_size =
3545 max_t(u32, blk_rq_sectors(rq), bfqd->last_rq_max_size);
3547 bfqd->last_rq_max_size = blk_rq_sectors(rq);
3549 bfqd->delta_from_first = now_ns - bfqd->first_dispatch;
3551 /* Target observation interval not yet reached, go on sampling */
3552 if (bfqd->delta_from_first < BFQ_RATE_REF_INTERVAL)
3553 goto update_last_values;
3555 update_rate_and_reset:
3556 bfq_update_rate_reset(bfqd, rq);
3558 bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
3559 if (RQ_BFQQ(rq) == bfqd->in_service_queue)
3560 bfqd->in_serv_last_pos = bfqd->last_position;
3561 bfqd->last_dispatch = now_ns;
3565 * Remove request from internal lists.
3567 static void bfq_dispatch_remove(struct request_queue *q, struct request *rq)
3569 struct bfq_queue *bfqq = RQ_BFQQ(rq);
3572 * For consistency, the next instruction should have been
3573 * executed after removing the request from the queue and
3574 * dispatching it. We execute instead this instruction before
3575 * bfq_remove_request() (and hence introduce a temporary
3576 * inconsistency), for efficiency. In fact, should this
3577 * dispatch occur for a non in-service bfqq, this anticipated
3578 * increment prevents two counters related to bfqq->dispatched
3579 * from risking to be, first, uselessly decremented, and then
3580 * incremented again when the (new) value of bfqq->dispatched
3581 * happens to be taken into account.
3584 bfq_update_peak_rate(q->elevator->elevator_data, rq);
3586 bfq_remove_request(q, rq);
3590 * There is a case where idling does not have to be performed for
3591 * throughput concerns, but to preserve the throughput share of
3592 * the process associated with bfqq.
3594 * To introduce this case, we can note that allowing the drive
3595 * to enqueue more than one request at a time, and hence
3596 * delegating de facto final scheduling decisions to the
3597 * drive's internal scheduler, entails loss of control on the
3598 * actual request service order. In particular, the critical
3599 * situation is when requests from different processes happen
3600 * to be present, at the same time, in the internal queue(s)
3601 * of the drive. In such a situation, the drive, by deciding
3602 * the service order of the internally-queued requests, does
3603 * determine also the actual throughput distribution among
3604 * these processes. But the drive typically has no notion or
3605 * concern about per-process throughput distribution, and
3606 * makes its decisions only on a per-request basis. Therefore,
3607 * the service distribution enforced by the drive's internal
3608 * scheduler is likely to coincide with the desired throughput
3609 * distribution only in a completely symmetric, or favorably
3610 * skewed scenario where:
3611 * (i-a) each of these processes must get the same throughput as
3613 * (i-b) in case (i-a) does not hold, it holds that the process
3614 * associated with bfqq must receive a lower or equal
3615 * throughput than any of the other processes;
3616 * (ii) the I/O of each process has the same properties, in
3617 * terms of locality (sequential or random), direction
3618 * (reads or writes), request sizes, greediness
3619 * (from I/O-bound to sporadic), and so on;
3621 * In fact, in such a scenario, the drive tends to treat the requests
3622 * of each process in about the same way as the requests of the
3623 * others, and thus to provide each of these processes with about the
3624 * same throughput. This is exactly the desired throughput
3625 * distribution if (i-a) holds, or, if (i-b) holds instead, this is an
3626 * even more convenient distribution for (the process associated with)
3629 * In contrast, in any asymmetric or unfavorable scenario, device
3630 * idling (I/O-dispatch plugging) is certainly needed to guarantee
3631 * that bfqq receives its assigned fraction of the device throughput
3632 * (see [1] for details).
3634 * The problem is that idling may significantly reduce throughput with
3635 * certain combinations of types of I/O and devices. An important
3636 * example is sync random I/O on flash storage with command
3637 * queueing. So, unless bfqq falls in cases where idling also boosts
3638 * throughput, it is important to check conditions (i-a), i(-b) and
3639 * (ii) accurately, so as to avoid idling when not strictly needed for
3640 * service guarantees.
3642 * Unfortunately, it is extremely difficult to thoroughly check
3643 * condition (ii). And, in case there are active groups, it becomes
3644 * very difficult to check conditions (i-a) and (i-b) too. In fact,
3645 * if there are active groups, then, for conditions (i-a) or (i-b) to
3646 * become false 'indirectly', it is enough that an active group
3647 * contains more active processes or sub-groups than some other active
3648 * group. More precisely, for conditions (i-a) or (i-b) to become
3649 * false because of such a group, it is not even necessary that the
3650 * group is (still) active: it is sufficient that, even if the group
3651 * has become inactive, some of its descendant processes still have
3652 * some request already dispatched but still waiting for
3653 * completion. In fact, requests have still to be guaranteed their
3654 * share of the throughput even after being dispatched. In this
3655 * respect, it is easy to show that, if a group frequently becomes
3656 * inactive while still having in-flight requests, and if, when this
3657 * happens, the group is not considered in the calculation of whether
3658 * the scenario is asymmetric, then the group may fail to be
3659 * guaranteed its fair share of the throughput (basically because
3660 * idling may not be performed for the descendant processes of the
3661 * group, but it had to be). We address this issue with the following
3662 * bi-modal behavior, implemented in the function
3663 * bfq_asymmetric_scenario().
3665 * If there are groups with requests waiting for completion
3666 * (as commented above, some of these groups may even be
3667 * already inactive), then the scenario is tagged as
3668 * asymmetric, conservatively, without checking any of the
3669 * conditions (i-a), (i-b) or (ii). So the device is idled for bfqq.
3670 * This behavior matches also the fact that groups are created
3671 * exactly if controlling I/O is a primary concern (to
3672 * preserve bandwidth and latency guarantees).
3674 * On the opposite end, if there are no groups with requests waiting
3675 * for completion, then only conditions (i-a) and (i-b) are actually
3676 * controlled, i.e., provided that conditions (i-a) or (i-b) holds,
3677 * idling is not performed, regardless of whether condition (ii)
3678 * holds. In other words, only if conditions (i-a) and (i-b) do not
3679 * hold, then idling is allowed, and the device tends to be prevented
3680 * from queueing many requests, possibly of several processes. Since
3681 * there are no groups with requests waiting for completion, then, to
3682 * control conditions (i-a) and (i-b) it is enough to check just
3683 * whether all the queues with requests waiting for completion also
3684 * have the same weight.
3686 * Not checking condition (ii) evidently exposes bfqq to the
3687 * risk of getting less throughput than its fair share.
3688 * However, for queues with the same weight, a further
3689 * mechanism, preemption, mitigates or even eliminates this
3690 * problem. And it does so without consequences on overall
3691 * throughput. This mechanism and its benefits are explained
3692 * in the next three paragraphs.
3694 * Even if a queue, say Q, is expired when it remains idle, Q
3695 * can still preempt the new in-service queue if the next
3696 * request of Q arrives soon (see the comments on
3697 * bfq_bfqq_update_budg_for_activation). If all queues and
3698 * groups have the same weight, this form of preemption,
3699 * combined with the hole-recovery heuristic described in the
3700 * comments on function bfq_bfqq_update_budg_for_activation,
3701 * are enough to preserve a correct bandwidth distribution in
3702 * the mid term, even without idling. In fact, even if not
3703 * idling allows the internal queues of the device to contain
3704 * many requests, and thus to reorder requests, we can rather
3705 * safely assume that the internal scheduler still preserves a
3706 * minimum of mid-term fairness.
3708 * More precisely, this preemption-based, idleless approach
3709 * provides fairness in terms of IOPS, and not sectors per
3710 * second. This can be seen with a simple example. Suppose
3711 * that there are two queues with the same weight, but that
3712 * the first queue receives requests of 8 sectors, while the
3713 * second queue receives requests of 1024 sectors. In
3714 * addition, suppose that each of the two queues contains at
3715 * most one request at a time, which implies that each queue
3716 * always remains idle after it is served. Finally, after
3717 * remaining idle, each queue receives very quickly a new
3718 * request. It follows that the two queues are served
3719 * alternatively, preempting each other if needed. This
3720 * implies that, although both queues have the same weight,
3721 * the queue with large requests receives a service that is
3722 * 1024/8 times as high as the service received by the other
3725 * The motivation for using preemption instead of idling (for
3726 * queues with the same weight) is that, by not idling,
3727 * service guarantees are preserved (completely or at least in
3728 * part) without minimally sacrificing throughput. And, if
3729 * there is no active group, then the primary expectation for
3730 * this device is probably a high throughput.
3732 * We are now left only with explaining the two sub-conditions in the
3733 * additional compound condition that is checked below for deciding
3734 * whether the scenario is asymmetric. To explain the first
3735 * sub-condition, we need to add that the function
3736 * bfq_asymmetric_scenario checks the weights of only
3737 * non-weight-raised queues, for efficiency reasons (see comments on
3738 * bfq_weights_tree_add()). Then the fact that bfqq is weight-raised
3739 * is checked explicitly here. More precisely, the compound condition
3740 * below takes into account also the fact that, even if bfqq is being
3741 * weight-raised, the scenario is still symmetric if all queues with
3742 * requests waiting for completion happen to be
3743 * weight-raised. Actually, we should be even more precise here, and
3744 * differentiate between interactive weight raising and soft real-time
3747 * The second sub-condition checked in the compound condition is
3748 * whether there is a fair amount of already in-flight I/O not
3749 * belonging to bfqq. If so, I/O dispatching is to be plugged, for the
3750 * following reason. The drive may decide to serve in-flight
3751 * non-bfqq's I/O requests before bfqq's ones, thereby delaying the
3752 * arrival of new I/O requests for bfqq (recall that bfqq is sync). If
3753 * I/O-dispatching is not plugged, then, while bfqq remains empty, a
3754 * basically uncontrolled amount of I/O from other queues may be
3755 * dispatched too, possibly causing the service of bfqq's I/O to be
3756 * delayed even longer in the drive. This problem gets more and more
3757 * serious as the speed and the queue depth of the drive grow,
3758 * because, as these two quantities grow, the probability to find no
3759 * queue busy but many requests in flight grows too. By contrast,
3760 * plugging I/O dispatching minimizes the delay induced by already
3761 * in-flight I/O, and enables bfqq to recover the bandwidth it may
3762 * lose because of this delay.
3764 * As a side note, it is worth considering that the above
3765 * device-idling countermeasures may however fail in the following
3766 * unlucky scenario: if I/O-dispatch plugging is (correctly) disabled
3767 * in a time period during which all symmetry sub-conditions hold, and
3768 * therefore the device is allowed to enqueue many requests, but at
3769 * some later point in time some sub-condition stops to hold, then it
3770 * may become impossible to make requests be served in the desired
3771 * order until all the requests already queued in the device have been
3772 * served. The last sub-condition commented above somewhat mitigates
3773 * this problem for weight-raised queues.
3775 * However, as an additional mitigation for this problem, we preserve
3776 * plugging for a special symmetric case that may suddenly turn into
3777 * asymmetric: the case where only bfqq is busy. In this case, not
3778 * expiring bfqq does not cause any harm to any other queues in terms
3779 * of service guarantees. In contrast, it avoids the following unlucky
3780 * sequence of events: (1) bfqq is expired, (2) a new queue with a
3781 * lower weight than bfqq becomes busy (or more queues), (3) the new
3782 * queue is served until a new request arrives for bfqq, (4) when bfqq
3783 * is finally served, there are so many requests of the new queue in
3784 * the drive that the pending requests for bfqq take a lot of time to
3785 * be served. In particular, event (2) may case even already
3786 * dispatched requests of bfqq to be delayed, inside the drive. So, to
3787 * avoid this series of events, the scenario is preventively declared
3788 * as asymmetric also if bfqq is the only busy queues
3790 static bool idling_needed_for_service_guarantees(struct bfq_data *bfqd,
3791 struct bfq_queue *bfqq)
3793 int tot_busy_queues = bfq_tot_busy_queues(bfqd);
3795 /* No point in idling for bfqq if it won't get requests any longer */
3796 if (unlikely(!bfqq_process_refs(bfqq)))
3799 return (bfqq->wr_coeff > 1 &&
3800 (bfqd->wr_busy_queues <
3802 bfqd->rq_in_driver >=
3803 bfqq->dispatched + 4)) ||
3804 bfq_asymmetric_scenario(bfqd, bfqq) ||
3805 tot_busy_queues == 1;
3808 static bool __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq,
3809 enum bfqq_expiration reason)
3812 * If this bfqq is shared between multiple processes, check
3813 * to make sure that those processes are still issuing I/Os
3814 * within the mean seek distance. If not, it may be time to
3815 * break the queues apart again.
3817 if (bfq_bfqq_coop(bfqq) && BFQQ_SEEKY(bfqq))
3818 bfq_mark_bfqq_split_coop(bfqq);
3821 * Consider queues with a higher finish virtual time than
3822 * bfqq. If idling_needed_for_service_guarantees(bfqq) returns
3823 * true, then bfqq's bandwidth would be violated if an
3824 * uncontrolled amount of I/O from these queues were
3825 * dispatched while bfqq is waiting for its new I/O to
3826 * arrive. This is exactly what may happen if this is a forced
3827 * expiration caused by a preemption attempt, and if bfqq is
3828 * not re-scheduled. To prevent this from happening, re-queue
3829 * bfqq if it needs I/O-dispatch plugging, even if it is
3830 * empty. By doing so, bfqq is granted to be served before the
3831 * above queues (provided that bfqq is of course eligible).
3833 if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
3834 !(reason == BFQQE_PREEMPTED &&
3835 idling_needed_for_service_guarantees(bfqd, bfqq))) {
3836 if (bfqq->dispatched == 0)
3838 * Overloading budget_timeout field to store
3839 * the time at which the queue remains with no
3840 * backlog and no outstanding request; used by
3841 * the weight-raising mechanism.
3843 bfqq->budget_timeout = jiffies;
3845 bfq_del_bfqq_busy(bfqq, true);
3847 bfq_requeue_bfqq(bfqd, bfqq, true);
3849 * Resort priority tree of potential close cooperators.
3850 * See comments on bfq_pos_tree_add_move() for the unlikely().
3852 if (unlikely(!bfqd->nonrot_with_queueing &&
3853 !RB_EMPTY_ROOT(&bfqq->sort_list)))
3854 bfq_pos_tree_add_move(bfqd, bfqq);
3858 * All in-service entities must have been properly deactivated
3859 * or requeued before executing the next function, which
3860 * resets all in-service entities as no more in service. This
3861 * may cause bfqq to be freed. If this happens, the next
3862 * function returns true.
3864 return __bfq_bfqd_reset_in_service(bfqd);
3868 * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior.
3869 * @bfqd: device data.
3870 * @bfqq: queue to update.
3871 * @reason: reason for expiration.
3873 * Handle the feedback on @bfqq budget at queue expiration.
3874 * See the body for detailed comments.
3876 static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd,
3877 struct bfq_queue *bfqq,
3878 enum bfqq_expiration reason)
3880 struct request *next_rq;
3881 int budget, min_budget;
3883 min_budget = bfq_min_budget(bfqd);
3885 if (bfqq->wr_coeff == 1)
3886 budget = bfqq->max_budget;
3888 * Use a constant, low budget for weight-raised queues,
3889 * to help achieve a low latency. Keep it slightly higher
3890 * than the minimum possible budget, to cause a little
3891 * bit fewer expirations.
3893 budget = 2 * min_budget;
3895 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %d, budg left %d",
3896 bfqq->entity.budget, bfq_bfqq_budget_left(bfqq));
3897 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %d, min budg %d",
3898 budget, bfq_min_budget(bfqd));
3899 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d",
3900 bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->in_service_queue));
3902 if (bfq_bfqq_sync(bfqq) && bfqq->wr_coeff == 1) {
3905 * Caveat: in all the following cases we trade latency
3908 case BFQQE_TOO_IDLE:
3910 * This is the only case where we may reduce
3911 * the budget: if there is no request of the
3912 * process still waiting for completion, then
3913 * we assume (tentatively) that the timer has
3914 * expired because the batch of requests of
3915 * the process could have been served with a
3916 * smaller budget. Hence, betting that
3917 * process will behave in the same way when it
3918 * becomes backlogged again, we reduce its
3919 * next budget. As long as we guess right,
3920 * this budget cut reduces the latency
3921 * experienced by the process.
3923 * However, if there are still outstanding
3924 * requests, then the process may have not yet
3925 * issued its next request just because it is
3926 * still waiting for the completion of some of
3927 * the still outstanding ones. So in this
3928 * subcase we do not reduce its budget, on the
3929 * contrary we increase it to possibly boost
3930 * the throughput, as discussed in the
3931 * comments to the BUDGET_TIMEOUT case.
3933 if (bfqq->dispatched > 0) /* still outstanding reqs */
3934 budget = min(budget * 2, bfqd->bfq_max_budget);
3936 if (budget > 5 * min_budget)
3937 budget -= 4 * min_budget;
3939 budget = min_budget;
3942 case BFQQE_BUDGET_TIMEOUT:
3944 * We double the budget here because it gives
3945 * the chance to boost the throughput if this
3946 * is not a seeky process (and has bumped into
3947 * this timeout because of, e.g., ZBR).
3949 budget = min(budget * 2, bfqd->bfq_max_budget);
3951 case BFQQE_BUDGET_EXHAUSTED:
3953 * The process still has backlog, and did not
3954 * let either the budget timeout or the disk
3955 * idling timeout expire. Hence it is not
3956 * seeky, has a short thinktime and may be
3957 * happy with a higher budget too. So
3958 * definitely increase the budget of this good
3959 * candidate to boost the disk throughput.
3961 budget = min(budget * 4, bfqd->bfq_max_budget);
3963 case BFQQE_NO_MORE_REQUESTS:
3965 * For queues that expire for this reason, it
3966 * is particularly important to keep the
3967 * budget close to the actual service they
3968 * need. Doing so reduces the timestamp
3969 * misalignment problem described in the
3970 * comments in the body of
3971 * __bfq_activate_entity. In fact, suppose
3972 * that a queue systematically expires for
3973 * BFQQE_NO_MORE_REQUESTS and presents a
3974 * new request in time to enjoy timestamp
3975 * back-shifting. The larger the budget of the
3976 * queue is with respect to the service the
3977 * queue actually requests in each service
3978 * slot, the more times the queue can be
3979 * reactivated with the same virtual finish
3980 * time. It follows that, even if this finish
3981 * time is pushed to the system virtual time
3982 * to reduce the consequent timestamp
3983 * misalignment, the queue unjustly enjoys for
3984 * many re-activations a lower finish time
3985 * than all newly activated queues.
3987 * The service needed by bfqq is measured
3988 * quite precisely by bfqq->entity.service.
3989 * Since bfqq does not enjoy device idling,
3990 * bfqq->entity.service is equal to the number
3991 * of sectors that the process associated with
3992 * bfqq requested to read/write before waiting
3993 * for request completions, or blocking for
3996 budget = max_t(int, bfqq->entity.service, min_budget);
4001 } else if (!bfq_bfqq_sync(bfqq)) {
4003 * Async queues get always the maximum possible
4004 * budget, as for them we do not care about latency
4005 * (in addition, their ability to dispatch is limited
4006 * by the charging factor).
4008 budget = bfqd->bfq_max_budget;
4011 bfqq->max_budget = budget;
4013 if (bfqd->budgets_assigned >= bfq_stats_min_budgets &&
4014 !bfqd->bfq_user_max_budget)
4015 bfqq->max_budget = min(bfqq->max_budget, bfqd->bfq_max_budget);
4018 * If there is still backlog, then assign a new budget, making
4019 * sure that it is large enough for the next request. Since
4020 * the finish time of bfqq must be kept in sync with the
4021 * budget, be sure to call __bfq_bfqq_expire() *after* this
4024 * If there is no backlog, then no need to update the budget;
4025 * it will be updated on the arrival of a new request.
4027 next_rq = bfqq->next_rq;
4029 bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget,
4030 bfq_serv_to_charge(next_rq, bfqq));
4032 bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %d",
4033 next_rq ? blk_rq_sectors(next_rq) : 0,
4034 bfqq->entity.budget);
4038 * Return true if the process associated with bfqq is "slow". The slow
4039 * flag is used, in addition to the budget timeout, to reduce the
4040 * amount of service provided to seeky processes, and thus reduce
4041 * their chances to lower the throughput. More details in the comments
4042 * on the function bfq_bfqq_expire().
4044 * An important observation is in order: as discussed in the comments
4045 * on the function bfq_update_peak_rate(), with devices with internal
4046 * queues, it is hard if ever possible to know when and for how long
4047 * an I/O request is processed by the device (apart from the trivial
4048 * I/O pattern where a new request is dispatched only after the
4049 * previous one has been completed). This makes it hard to evaluate
4050 * the real rate at which the I/O requests of each bfq_queue are
4051 * served. In fact, for an I/O scheduler like BFQ, serving a
4052 * bfq_queue means just dispatching its requests during its service
4053 * slot (i.e., until the budget of the queue is exhausted, or the
4054 * queue remains idle, or, finally, a timeout fires). But, during the
4055 * service slot of a bfq_queue, around 100 ms at most, the device may
4056 * be even still processing requests of bfq_queues served in previous
4057 * service slots. On the opposite end, the requests of the in-service
4058 * bfq_queue may be completed after the service slot of the queue
4061 * Anyway, unless more sophisticated solutions are used
4062 * (where possible), the sum of the sizes of the requests dispatched
4063 * during the service slot of a bfq_queue is probably the only
4064 * approximation available for the service received by the bfq_queue
4065 * during its service slot. And this sum is the quantity used in this
4066 * function to evaluate the I/O speed of a process.
4068 static bool bfq_bfqq_is_slow(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4069 bool compensate, enum bfqq_expiration reason,
4070 unsigned long *delta_ms)
4072 ktime_t delta_ktime;
4074 bool slow = BFQQ_SEEKY(bfqq); /* if delta too short, use seekyness */
4076 if (!bfq_bfqq_sync(bfqq))
4080 delta_ktime = bfqd->last_idling_start;
4082 delta_ktime = ktime_get();
4083 delta_ktime = ktime_sub(delta_ktime, bfqd->last_budget_start);
4084 delta_usecs = ktime_to_us(delta_ktime);
4086 /* don't use too short time intervals */
4087 if (delta_usecs < 1000) {
4088 if (blk_queue_nonrot(bfqd->queue))
4090 * give same worst-case guarantees as idling
4093 *delta_ms = BFQ_MIN_TT / NSEC_PER_MSEC;
4094 else /* charge at least one seek */
4095 *delta_ms = bfq_slice_idle / NSEC_PER_MSEC;
4100 *delta_ms = delta_usecs / USEC_PER_MSEC;
4103 * Use only long (> 20ms) intervals to filter out excessive
4104 * spikes in service rate estimation.
4106 if (delta_usecs > 20000) {
4108 * Caveat for rotational devices: processes doing I/O
4109 * in the slower disk zones tend to be slow(er) even
4110 * if not seeky. In this respect, the estimated peak
4111 * rate is likely to be an average over the disk
4112 * surface. Accordingly, to not be too harsh with
4113 * unlucky processes, a process is deemed slow only if
4114 * its rate has been lower than half of the estimated
4117 slow = bfqq->entity.service < bfqd->bfq_max_budget / 2;
4120 bfq_log_bfqq(bfqd, bfqq, "bfq_bfqq_is_slow: slow %d", slow);
4126 * To be deemed as soft real-time, an application must meet two
4127 * requirements. First, the application must not require an average
4128 * bandwidth higher than the approximate bandwidth required to playback or
4129 * record a compressed high-definition video.
4130 * The next function is invoked on the completion of the last request of a
4131 * batch, to compute the next-start time instant, soft_rt_next_start, such
4132 * that, if the next request of the application does not arrive before
4133 * soft_rt_next_start, then the above requirement on the bandwidth is met.
4135 * The second requirement is that the request pattern of the application is
4136 * isochronous, i.e., that, after issuing a request or a batch of requests,
4137 * the application stops issuing new requests until all its pending requests
4138 * have been completed. After that, the application may issue a new batch,
4140 * For this reason the next function is invoked to compute
4141 * soft_rt_next_start only for applications that meet this requirement,
4142 * whereas soft_rt_next_start is set to infinity for applications that do
4145 * Unfortunately, even a greedy (i.e., I/O-bound) application may
4146 * happen to meet, occasionally or systematically, both the above
4147 * bandwidth and isochrony requirements. This may happen at least in
4148 * the following circumstances. First, if the CPU load is high. The
4149 * application may stop issuing requests while the CPUs are busy
4150 * serving other processes, then restart, then stop again for a while,
4151 * and so on. The other circumstances are related to the storage
4152 * device: the storage device is highly loaded or reaches a low-enough
4153 * throughput with the I/O of the application (e.g., because the I/O
4154 * is random and/or the device is slow). In all these cases, the
4155 * I/O of the application may be simply slowed down enough to meet
4156 * the bandwidth and isochrony requirements. To reduce the probability
4157 * that greedy applications are deemed as soft real-time in these
4158 * corner cases, a further rule is used in the computation of
4159 * soft_rt_next_start: the return value of this function is forced to
4160 * be higher than the maximum between the following two quantities.
4162 * (a) Current time plus: (1) the maximum time for which the arrival
4163 * of a request is waited for when a sync queue becomes idle,
4164 * namely bfqd->bfq_slice_idle, and (2) a few extra jiffies. We
4165 * postpone for a moment the reason for adding a few extra
4166 * jiffies; we get back to it after next item (b). Lower-bounding
4167 * the return value of this function with the current time plus
4168 * bfqd->bfq_slice_idle tends to filter out greedy applications,
4169 * because the latter issue their next request as soon as possible
4170 * after the last one has been completed. In contrast, a soft
4171 * real-time application spends some time processing data, after a
4172 * batch of its requests has been completed.
4174 * (b) Current value of bfqq->soft_rt_next_start. As pointed out
4175 * above, greedy applications may happen to meet both the
4176 * bandwidth and isochrony requirements under heavy CPU or
4177 * storage-device load. In more detail, in these scenarios, these
4178 * applications happen, only for limited time periods, to do I/O
4179 * slowly enough to meet all the requirements described so far,
4180 * including the filtering in above item (a). These slow-speed
4181 * time intervals are usually interspersed between other time
4182 * intervals during which these applications do I/O at a very high
4183 * speed. Fortunately, exactly because of the high speed of the
4184 * I/O in the high-speed intervals, the values returned by this
4185 * function happen to be so high, near the end of any such
4186 * high-speed interval, to be likely to fall *after* the end of
4187 * the low-speed time interval that follows. These high values are
4188 * stored in bfqq->soft_rt_next_start after each invocation of
4189 * this function. As a consequence, if the last value of
4190 * bfqq->soft_rt_next_start is constantly used to lower-bound the
4191 * next value that this function may return, then, from the very
4192 * beginning of a low-speed interval, bfqq->soft_rt_next_start is
4193 * likely to be constantly kept so high that any I/O request
4194 * issued during the low-speed interval is considered as arriving
4195 * to soon for the application to be deemed as soft
4196 * real-time. Then, in the high-speed interval that follows, the
4197 * application will not be deemed as soft real-time, just because
4198 * it will do I/O at a high speed. And so on.
4200 * Getting back to the filtering in item (a), in the following two
4201 * cases this filtering might be easily passed by a greedy
4202 * application, if the reference quantity was just
4203 * bfqd->bfq_slice_idle:
4204 * 1) HZ is so low that the duration of a jiffy is comparable to or
4205 * higher than bfqd->bfq_slice_idle. This happens, e.g., on slow
4206 * devices with HZ=100. The time granularity may be so coarse
4207 * that the approximation, in jiffies, of bfqd->bfq_slice_idle
4208 * is rather lower than the exact value.
4209 * 2) jiffies, instead of increasing at a constant rate, may stop increasing
4210 * for a while, then suddenly 'jump' by several units to recover the lost
4211 * increments. This seems to happen, e.g., inside virtual machines.
4212 * To address this issue, in the filtering in (a) we do not use as a
4213 * reference time interval just bfqd->bfq_slice_idle, but
4214 * bfqd->bfq_slice_idle plus a few jiffies. In particular, we add the
4215 * minimum number of jiffies for which the filter seems to be quite
4216 * precise also in embedded systems and KVM/QEMU virtual machines.
4218 static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data *bfqd,
4219 struct bfq_queue *bfqq)
4221 return max3(bfqq->soft_rt_next_start,
4222 bfqq->last_idle_bklogged +
4223 HZ * bfqq->service_from_backlogged /
4224 bfqd->bfq_wr_max_softrt_rate,
4225 jiffies + nsecs_to_jiffies(bfqq->bfqd->bfq_slice_idle) + 4);
4229 * bfq_bfqq_expire - expire a queue.
4230 * @bfqd: device owning the queue.
4231 * @bfqq: the queue to expire.
4232 * @compensate: if true, compensate for the time spent idling.
4233 * @reason: the reason causing the expiration.
4235 * If the process associated with bfqq does slow I/O (e.g., because it
4236 * issues random requests), we charge bfqq with the time it has been
4237 * in service instead of the service it has received (see
4238 * bfq_bfqq_charge_time for details on how this goal is achieved). As
4239 * a consequence, bfqq will typically get higher timestamps upon
4240 * reactivation, and hence it will be rescheduled as if it had
4241 * received more service than what it has actually received. In the
4242 * end, bfqq receives less service in proportion to how slowly its
4243 * associated process consumes its budgets (and hence how seriously it
4244 * tends to lower the throughput). In addition, this time-charging
4245 * strategy guarantees time fairness among slow processes. In
4246 * contrast, if the process associated with bfqq is not slow, we
4247 * charge bfqq exactly with the service it has received.
4249 * Charging time to the first type of queues and the exact service to
4250 * the other has the effect of using the WF2Q+ policy to schedule the
4251 * former on a timeslice basis, without violating service domain
4252 * guarantees among the latter.
4254 void bfq_bfqq_expire(struct bfq_data *bfqd,
4255 struct bfq_queue *bfqq,
4257 enum bfqq_expiration reason)
4260 unsigned long delta = 0;
4261 struct bfq_entity *entity = &bfqq->entity;
4264 * Check whether the process is slow (see bfq_bfqq_is_slow).
4266 slow = bfq_bfqq_is_slow(bfqd, bfqq, compensate, reason, &delta);
4269 * As above explained, charge slow (typically seeky) and
4270 * timed-out queues with the time and not the service
4271 * received, to favor sequential workloads.
4273 * Processes doing I/O in the slower disk zones will tend to
4274 * be slow(er) even if not seeky. Therefore, since the
4275 * estimated peak rate is actually an average over the disk
4276 * surface, these processes may timeout just for bad luck. To
4277 * avoid punishing them, do not charge time to processes that
4278 * succeeded in consuming at least 2/3 of their budget. This
4279 * allows BFQ to preserve enough elasticity to still perform
4280 * bandwidth, and not time, distribution with little unlucky
4281 * or quasi-sequential processes.
4283 if (bfqq->wr_coeff == 1 &&
4285 (reason == BFQQE_BUDGET_TIMEOUT &&
4286 bfq_bfqq_budget_left(bfqq) >= entity->budget / 3)))
4287 bfq_bfqq_charge_time(bfqd, bfqq, delta);
4289 if (bfqd->low_latency && bfqq->wr_coeff == 1)
4290 bfqq->last_wr_start_finish = jiffies;
4292 if (bfqd->low_latency && bfqd->bfq_wr_max_softrt_rate > 0 &&
4293 RB_EMPTY_ROOT(&bfqq->sort_list)) {
4295 * If we get here, and there are no outstanding
4296 * requests, then the request pattern is isochronous
4297 * (see the comments on the function
4298 * bfq_bfqq_softrt_next_start()). Therefore we can
4299 * compute soft_rt_next_start.
4301 * If, instead, the queue still has outstanding
4302 * requests, then we have to wait for the completion
4303 * of all the outstanding requests to discover whether
4304 * the request pattern is actually isochronous.
4306 if (bfqq->dispatched == 0)
4307 bfqq->soft_rt_next_start =
4308 bfq_bfqq_softrt_next_start(bfqd, bfqq);
4309 else if (bfqq->dispatched > 0) {
4311 * Schedule an update of soft_rt_next_start to when
4312 * the task may be discovered to be isochronous.
4314 bfq_mark_bfqq_softrt_update(bfqq);
4318 bfq_log_bfqq(bfqd, bfqq,
4319 "expire (%d, slow %d, num_disp %d, short_ttime %d)", reason,
4320 slow, bfqq->dispatched, bfq_bfqq_has_short_ttime(bfqq));
4323 * bfqq expired, so no total service time needs to be computed
4324 * any longer: reset state machine for measuring total service
4327 bfqd->rqs_injected = bfqd->wait_dispatch = false;
4328 bfqd->waited_rq = NULL;
4331 * Increase, decrease or leave budget unchanged according to
4334 __bfq_bfqq_recalc_budget(bfqd, bfqq, reason);
4335 if (__bfq_bfqq_expire(bfqd, bfqq, reason))
4336 /* bfqq is gone, no more actions on it */
4339 /* mark bfqq as waiting a request only if a bic still points to it */
4340 if (!bfq_bfqq_busy(bfqq) &&
4341 reason != BFQQE_BUDGET_TIMEOUT &&
4342 reason != BFQQE_BUDGET_EXHAUSTED) {
4343 bfq_mark_bfqq_non_blocking_wait_rq(bfqq);
4345 * Not setting service to 0, because, if the next rq
4346 * arrives in time, the queue will go on receiving
4347 * service with this same budget (as if it never expired)
4350 entity->service = 0;
4353 * Reset the received-service counter for every parent entity.
4354 * Differently from what happens with bfqq->entity.service,
4355 * the resetting of this counter never needs to be postponed
4356 * for parent entities. In fact, in case bfqq may have a
4357 * chance to go on being served using the last, partially
4358 * consumed budget, bfqq->entity.service needs to be kept,
4359 * because if bfqq then actually goes on being served using
4360 * the same budget, the last value of bfqq->entity.service is
4361 * needed to properly decrement bfqq->entity.budget by the
4362 * portion already consumed. In contrast, it is not necessary
4363 * to keep entity->service for parent entities too, because
4364 * the bubble up of the new value of bfqq->entity.budget will
4365 * make sure that the budgets of parent entities are correct,
4366 * even in case bfqq and thus parent entities go on receiving
4367 * service with the same budget.
4369 entity = entity->parent;
4370 for_each_entity(entity)
4371 entity->service = 0;
4375 * Budget timeout is not implemented through a dedicated timer, but
4376 * just checked on request arrivals and completions, as well as on
4377 * idle timer expirations.
4379 static bool bfq_bfqq_budget_timeout(struct bfq_queue *bfqq)
4381 return time_is_before_eq_jiffies(bfqq->budget_timeout);
4385 * If we expire a queue that is actively waiting (i.e., with the
4386 * device idled) for the arrival of a new request, then we may incur
4387 * the timestamp misalignment problem described in the body of the
4388 * function __bfq_activate_entity. Hence we return true only if this
4389 * condition does not hold, or if the queue is slow enough to deserve
4390 * only to be kicked off for preserving a high throughput.
4392 static bool bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq)
4394 bfq_log_bfqq(bfqq->bfqd, bfqq,
4395 "may_budget_timeout: wait_request %d left %d timeout %d",
4396 bfq_bfqq_wait_request(bfqq),
4397 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3,
4398 bfq_bfqq_budget_timeout(bfqq));
4400 return (!bfq_bfqq_wait_request(bfqq) ||
4401 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3)
4403 bfq_bfqq_budget_timeout(bfqq);
4406 static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd,
4407 struct bfq_queue *bfqq)
4409 bool rot_without_queueing =
4410 !blk_queue_nonrot(bfqd->queue) && !bfqd->hw_tag,
4411 bfqq_sequential_and_IO_bound,
4414 /* No point in idling for bfqq if it won't get requests any longer */
4415 if (unlikely(!bfqq_process_refs(bfqq)))
4418 bfqq_sequential_and_IO_bound = !BFQQ_SEEKY(bfqq) &&
4419 bfq_bfqq_IO_bound(bfqq) && bfq_bfqq_has_short_ttime(bfqq);
4422 * The next variable takes into account the cases where idling
4423 * boosts the throughput.
4425 * The value of the variable is computed considering, first, that
4426 * idling is virtually always beneficial for the throughput if:
4427 * (a) the device is not NCQ-capable and rotational, or
4428 * (b) regardless of the presence of NCQ, the device is rotational and
4429 * the request pattern for bfqq is I/O-bound and sequential, or
4430 * (c) regardless of whether it is rotational, the device is
4431 * not NCQ-capable and the request pattern for bfqq is
4432 * I/O-bound and sequential.
4434 * Secondly, and in contrast to the above item (b), idling an
4435 * NCQ-capable flash-based device would not boost the
4436 * throughput even with sequential I/O; rather it would lower
4437 * the throughput in proportion to how fast the device
4438 * is. Accordingly, the next variable is true if any of the
4439 * above conditions (a), (b) or (c) is true, and, in
4440 * particular, happens to be false if bfqd is an NCQ-capable
4441 * flash-based device.
4443 idling_boosts_thr = rot_without_queueing ||
4444 ((!blk_queue_nonrot(bfqd->queue) || !bfqd->hw_tag) &&
4445 bfqq_sequential_and_IO_bound);
4448 * The return value of this function is equal to that of
4449 * idling_boosts_thr, unless a special case holds. In this
4450 * special case, described below, idling may cause problems to
4451 * weight-raised queues.
4453 * When the request pool is saturated (e.g., in the presence
4454 * of write hogs), if the processes associated with
4455 * non-weight-raised queues ask for requests at a lower rate,
4456 * then processes associated with weight-raised queues have a
4457 * higher probability to get a request from the pool
4458 * immediately (or at least soon) when they need one. Thus
4459 * they have a higher probability to actually get a fraction
4460 * of the device throughput proportional to their high
4461 * weight. This is especially true with NCQ-capable drives,
4462 * which enqueue several requests in advance, and further
4463 * reorder internally-queued requests.
4465 * For this reason, we force to false the return value if
4466 * there are weight-raised busy queues. In this case, and if
4467 * bfqq is not weight-raised, this guarantees that the device
4468 * is not idled for bfqq (if, instead, bfqq is weight-raised,
4469 * then idling will be guaranteed by another variable, see
4470 * below). Combined with the timestamping rules of BFQ (see
4471 * [1] for details), this behavior causes bfqq, and hence any
4472 * sync non-weight-raised queue, to get a lower number of
4473 * requests served, and thus to ask for a lower number of
4474 * requests from the request pool, before the busy
4475 * weight-raised queues get served again. This often mitigates
4476 * starvation problems in the presence of heavy write
4477 * workloads and NCQ, thereby guaranteeing a higher
4478 * application and system responsiveness in these hostile
4481 return idling_boosts_thr &&
4482 bfqd->wr_busy_queues == 0;
4486 * For a queue that becomes empty, device idling is allowed only if
4487 * this function returns true for that queue. As a consequence, since
4488 * device idling plays a critical role for both throughput boosting
4489 * and service guarantees, the return value of this function plays a
4490 * critical role as well.
4492 * In a nutshell, this function returns true only if idling is
4493 * beneficial for throughput or, even if detrimental for throughput,
4494 * idling is however necessary to preserve service guarantees (low
4495 * latency, desired throughput distribution, ...). In particular, on
4496 * NCQ-capable devices, this function tries to return false, so as to
4497 * help keep the drives' internal queues full, whenever this helps the
4498 * device boost the throughput without causing any service-guarantee
4501 * Most of the issues taken into account to get the return value of
4502 * this function are not trivial. We discuss these issues in the two
4503 * functions providing the main pieces of information needed by this
4506 static bool bfq_better_to_idle(struct bfq_queue *bfqq)
4508 struct bfq_data *bfqd = bfqq->bfqd;
4509 bool idling_boosts_thr_with_no_issue, idling_needed_for_service_guar;
4511 /* No point in idling for bfqq if it won't get requests any longer */
4512 if (unlikely(!bfqq_process_refs(bfqq)))
4515 if (unlikely(bfqd->strict_guarantees))
4519 * Idling is performed only if slice_idle > 0. In addition, we
4522 * (b) bfqq is in the idle io prio class: in this case we do
4523 * not idle because we want to minimize the bandwidth that
4524 * queues in this class can steal to higher-priority queues
4526 if (bfqd->bfq_slice_idle == 0 || !bfq_bfqq_sync(bfqq) ||
4527 bfq_class_idle(bfqq))
4530 idling_boosts_thr_with_no_issue =
4531 idling_boosts_thr_without_issues(bfqd, bfqq);
4533 idling_needed_for_service_guar =
4534 idling_needed_for_service_guarantees(bfqd, bfqq);
4537 * We have now the two components we need to compute the
4538 * return value of the function, which is true only if idling
4539 * either boosts the throughput (without issues), or is
4540 * necessary to preserve service guarantees.
4542 return idling_boosts_thr_with_no_issue ||
4543 idling_needed_for_service_guar;
4547 * If the in-service queue is empty but the function bfq_better_to_idle
4548 * returns true, then:
4549 * 1) the queue must remain in service and cannot be expired, and
4550 * 2) the device must be idled to wait for the possible arrival of a new
4551 * request for the queue.
4552 * See the comments on the function bfq_better_to_idle for the reasons
4553 * why performing device idling is the best choice to boost the throughput
4554 * and preserve service guarantees when bfq_better_to_idle itself
4557 static bool bfq_bfqq_must_idle(struct bfq_queue *bfqq)
4559 return RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_better_to_idle(bfqq);
4563 * This function chooses the queue from which to pick the next extra
4564 * I/O request to inject, if it finds a compatible queue. See the
4565 * comments on bfq_update_inject_limit() for details on the injection
4566 * mechanism, and for the definitions of the quantities mentioned
4569 static struct bfq_queue *
4570 bfq_choose_bfqq_for_injection(struct bfq_data *bfqd)
4572 struct bfq_queue *bfqq, *in_serv_bfqq = bfqd->in_service_queue;
4573 unsigned int limit = in_serv_bfqq->inject_limit;
4576 * - bfqq is not weight-raised and therefore does not carry
4577 * time-critical I/O,
4579 * - regardless of whether bfqq is weight-raised, bfqq has
4580 * however a long think time, during which it can absorb the
4581 * effect of an appropriate number of extra I/O requests
4582 * from other queues (see bfq_update_inject_limit for
4583 * details on the computation of this number);
4584 * then injection can be performed without restrictions.
4586 bool in_serv_always_inject = in_serv_bfqq->wr_coeff == 1 ||
4587 !bfq_bfqq_has_short_ttime(in_serv_bfqq);
4591 * - the baseline total service time could not be sampled yet,
4592 * so the inject limit happens to be still 0, and
4593 * - a lot of time has elapsed since the plugging of I/O
4594 * dispatching started, so drive speed is being wasted
4596 * then temporarily raise inject limit to one request.
4598 if (limit == 0 && in_serv_bfqq->last_serv_time_ns == 0 &&
4599 bfq_bfqq_wait_request(in_serv_bfqq) &&
4600 time_is_before_eq_jiffies(bfqd->last_idling_start_jiffies +
4601 bfqd->bfq_slice_idle)
4605 if (bfqd->rq_in_driver >= limit)
4609 * Linear search of the source queue for injection; but, with
4610 * a high probability, very few steps are needed to find a
4611 * candidate queue, i.e., a queue with enough budget left for
4612 * its next request. In fact:
4613 * - BFQ dynamically updates the budget of every queue so as
4614 * to accommodate the expected backlog of the queue;
4615 * - if a queue gets all its requests dispatched as injected
4616 * service, then the queue is removed from the active list
4617 * (and re-added only if it gets new requests, but then it
4618 * is assigned again enough budget for its new backlog).
4620 list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
4621 if (!RB_EMPTY_ROOT(&bfqq->sort_list) &&
4622 (in_serv_always_inject || bfqq->wr_coeff > 1) &&
4623 bfq_serv_to_charge(bfqq->next_rq, bfqq) <=
4624 bfq_bfqq_budget_left(bfqq)) {
4626 * Allow for only one large in-flight request
4627 * on non-rotational devices, for the
4628 * following reason. On non-rotationl drives,
4629 * large requests take much longer than
4630 * smaller requests to be served. In addition,
4631 * the drive prefers to serve large requests
4632 * w.r.t. to small ones, if it can choose. So,
4633 * having more than one large requests queued
4634 * in the drive may easily make the next first
4635 * request of the in-service queue wait for so
4636 * long to break bfqq's service guarantees. On
4637 * the bright side, large requests let the
4638 * drive reach a very high throughput, even if
4639 * there is only one in-flight large request
4642 if (blk_queue_nonrot(bfqd->queue) &&
4643 blk_rq_sectors(bfqq->next_rq) >=
4644 BFQQ_SECT_THR_NONROT)
4645 limit = min_t(unsigned int, 1, limit);
4647 limit = in_serv_bfqq->inject_limit;
4649 if (bfqd->rq_in_driver < limit) {
4650 bfqd->rqs_injected = true;
4659 * Select a queue for service. If we have a current queue in service,
4660 * check whether to continue servicing it, or retrieve and set a new one.
4662 static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd)
4664 struct bfq_queue *bfqq;
4665 struct request *next_rq;
4666 enum bfqq_expiration reason = BFQQE_BUDGET_TIMEOUT;
4668 bfqq = bfqd->in_service_queue;
4672 bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue");
4675 * Do not expire bfqq for budget timeout if bfqq may be about
4676 * to enjoy device idling. The reason why, in this case, we
4677 * prevent bfqq from expiring is the same as in the comments
4678 * on the case where bfq_bfqq_must_idle() returns true, in
4679 * bfq_completed_request().
4681 if (bfq_may_expire_for_budg_timeout(bfqq) &&
4682 !bfq_bfqq_must_idle(bfqq))
4687 * This loop is rarely executed more than once. Even when it
4688 * happens, it is much more convenient to re-execute this loop
4689 * than to return NULL and trigger a new dispatch to get a
4692 next_rq = bfqq->next_rq;
4694 * If bfqq has requests queued and it has enough budget left to
4695 * serve them, keep the queue, otherwise expire it.
4698 if (bfq_serv_to_charge(next_rq, bfqq) >
4699 bfq_bfqq_budget_left(bfqq)) {
4701 * Expire the queue for budget exhaustion,
4702 * which makes sure that the next budget is
4703 * enough to serve the next request, even if
4704 * it comes from the fifo expired path.
4706 reason = BFQQE_BUDGET_EXHAUSTED;
4710 * The idle timer may be pending because we may
4711 * not disable disk idling even when a new request
4714 if (bfq_bfqq_wait_request(bfqq)) {
4716 * If we get here: 1) at least a new request
4717 * has arrived but we have not disabled the
4718 * timer because the request was too small,
4719 * 2) then the block layer has unplugged
4720 * the device, causing the dispatch to be
4723 * Since the device is unplugged, now the
4724 * requests are probably large enough to
4725 * provide a reasonable throughput.
4726 * So we disable idling.
4728 bfq_clear_bfqq_wait_request(bfqq);
4729 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
4736 * No requests pending. However, if the in-service queue is idling
4737 * for a new request, or has requests waiting for a completion and
4738 * may idle after their completion, then keep it anyway.
4740 * Yet, inject service from other queues if it boosts
4741 * throughput and is possible.
4743 if (bfq_bfqq_wait_request(bfqq) ||
4744 (bfqq->dispatched != 0 && bfq_better_to_idle(bfqq))) {
4745 struct bfq_queue *async_bfqq =
4746 bfqq->bic && bfqq->bic->bfqq[0] &&
4747 bfq_bfqq_busy(bfqq->bic->bfqq[0]) &&
4748 bfqq->bic->bfqq[0]->next_rq ?
4749 bfqq->bic->bfqq[0] : NULL;
4750 struct bfq_queue *blocked_bfqq =
4751 !hlist_empty(&bfqq->woken_list) ?
4752 container_of(bfqq->woken_list.first,
4758 * The next four mutually-exclusive ifs decide
4759 * whether to try injection, and choose the queue to
4760 * pick an I/O request from.
4762 * The first if checks whether the process associated
4763 * with bfqq has also async I/O pending. If so, it
4764 * injects such I/O unconditionally. Injecting async
4765 * I/O from the same process can cause no harm to the
4766 * process. On the contrary, it can only increase
4767 * bandwidth and reduce latency for the process.
4769 * The second if checks whether there happens to be a
4770 * non-empty waker queue for bfqq, i.e., a queue whose
4771 * I/O needs to be completed for bfqq to receive new
4772 * I/O. This happens, e.g., if bfqq is associated with
4773 * a process that does some sync. A sync generates
4774 * extra blocking I/O, which must be completed before
4775 * the process associated with bfqq can go on with its
4776 * I/O. If the I/O of the waker queue is not served,
4777 * then bfqq remains empty, and no I/O is dispatched,
4778 * until the idle timeout fires for bfqq. This is
4779 * likely to result in lower bandwidth and higher
4780 * latencies for bfqq, and in a severe loss of total
4781 * throughput. The best action to take is therefore to
4782 * serve the waker queue as soon as possible. So do it
4783 * (without relying on the third alternative below for
4784 * eventually serving waker_bfqq's I/O; see the last
4785 * paragraph for further details). This systematic
4786 * injection of I/O from the waker queue does not
4787 * cause any delay to bfqq's I/O. On the contrary,
4788 * next bfqq's I/O is brought forward dramatically,
4789 * for it is not blocked for milliseconds.
4791 * The third if checks whether there is a queue woken
4792 * by bfqq, and currently with pending I/O. Such a
4793 * woken queue does not steal bandwidth from bfqq,
4794 * because it remains soon without I/O if bfqq is not
4795 * served. So there is virtually no risk of loss of
4796 * bandwidth for bfqq if this woken queue has I/O
4797 * dispatched while bfqq is waiting for new I/O.
4799 * The fourth if checks whether bfqq is a queue for
4800 * which it is better to avoid injection. It is so if
4801 * bfqq delivers more throughput when served without
4802 * any further I/O from other queues in the middle, or
4803 * if the service times of bfqq's I/O requests both
4804 * count more than overall throughput, and may be
4805 * easily increased by injection (this happens if bfqq
4806 * has a short think time). If none of these
4807 * conditions holds, then a candidate queue for
4808 * injection is looked for through
4809 * bfq_choose_bfqq_for_injection(). Note that the
4810 * latter may return NULL (for example if the inject
4811 * limit for bfqq is currently 0).
4813 * NOTE: motivation for the second alternative
4815 * Thanks to the way the inject limit is updated in
4816 * bfq_update_has_short_ttime(), it is rather likely
4817 * that, if I/O is being plugged for bfqq and the
4818 * waker queue has pending I/O requests that are
4819 * blocking bfqq's I/O, then the fourth alternative
4820 * above lets the waker queue get served before the
4821 * I/O-plugging timeout fires. So one may deem the
4822 * second alternative superfluous. It is not, because
4823 * the fourth alternative may be way less effective in
4824 * case of a synchronization. For two main
4825 * reasons. First, throughput may be low because the
4826 * inject limit may be too low to guarantee the same
4827 * amount of injected I/O, from the waker queue or
4828 * other queues, that the second alternative
4829 * guarantees (the second alternative unconditionally
4830 * injects a pending I/O request of the waker queue
4831 * for each bfq_dispatch_request()). Second, with the
4832 * fourth alternative, the duration of the plugging,
4833 * i.e., the time before bfqq finally receives new I/O,
4834 * may not be minimized, because the waker queue may
4835 * happen to be served only after other queues.
4838 icq_to_bic(async_bfqq->next_rq->elv.icq) == bfqq->bic &&
4839 bfq_serv_to_charge(async_bfqq->next_rq, async_bfqq) <=
4840 bfq_bfqq_budget_left(async_bfqq))
4841 bfqq = bfqq->bic->bfqq[0];
4842 else if (bfqq->waker_bfqq &&
4843 bfq_bfqq_busy(bfqq->waker_bfqq) &&
4844 bfqq->waker_bfqq->next_rq &&
4845 bfq_serv_to_charge(bfqq->waker_bfqq->next_rq,
4846 bfqq->waker_bfqq) <=
4847 bfq_bfqq_budget_left(bfqq->waker_bfqq)
4849 bfqq = bfqq->waker_bfqq;
4850 else if (blocked_bfqq &&
4851 bfq_bfqq_busy(blocked_bfqq) &&
4852 blocked_bfqq->next_rq &&
4853 bfq_serv_to_charge(blocked_bfqq->next_rq,
4855 bfq_bfqq_budget_left(blocked_bfqq)
4857 bfqq = blocked_bfqq;
4858 else if (!idling_boosts_thr_without_issues(bfqd, bfqq) &&
4859 (bfqq->wr_coeff == 1 || bfqd->wr_busy_queues > 1 ||
4860 !bfq_bfqq_has_short_ttime(bfqq)))
4861 bfqq = bfq_choose_bfqq_for_injection(bfqd);
4868 reason = BFQQE_NO_MORE_REQUESTS;
4870 bfq_bfqq_expire(bfqd, bfqq, false, reason);
4872 bfqq = bfq_set_in_service_queue(bfqd);
4874 bfq_log_bfqq(bfqd, bfqq, "select_queue: checking new queue");
4879 bfq_log_bfqq(bfqd, bfqq, "select_queue: returned this queue");
4881 bfq_log(bfqd, "select_queue: no queue returned");
4886 static void bfq_update_wr_data(struct bfq_data *bfqd, struct bfq_queue *bfqq)
4888 struct bfq_entity *entity = &bfqq->entity;
4890 if (bfqq->wr_coeff > 1) { /* queue is being weight-raised */
4891 bfq_log_bfqq(bfqd, bfqq,
4892 "raising period dur %u/%u msec, old coeff %u, w %d(%d)",
4893 jiffies_to_msecs(jiffies - bfqq->last_wr_start_finish),
4894 jiffies_to_msecs(bfqq->wr_cur_max_time),
4896 bfqq->entity.weight, bfqq->entity.orig_weight);
4898 if (entity->prio_changed)
4899 bfq_log_bfqq(bfqd, bfqq, "WARN: pending prio change");
4902 * If the queue was activated in a burst, or too much
4903 * time has elapsed from the beginning of this
4904 * weight-raising period, then end weight raising.
4906 if (bfq_bfqq_in_large_burst(bfqq))
4907 bfq_bfqq_end_wr(bfqq);
4908 else if (time_is_before_jiffies(bfqq->last_wr_start_finish +
4909 bfqq->wr_cur_max_time)) {
4910 if (bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time ||
4911 time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
4912 bfq_wr_duration(bfqd))) {
4914 * Either in interactive weight
4915 * raising, or in soft_rt weight
4917 * interactive-weight-raising period
4918 * elapsed (so no switch back to
4919 * interactive weight raising).
4921 bfq_bfqq_end_wr(bfqq);
4923 * soft_rt finishing while still in
4924 * interactive period, switch back to
4925 * interactive weight raising
4927 switch_back_to_interactive_wr(bfqq, bfqd);
4928 bfqq->entity.prio_changed = 1;
4931 if (bfqq->wr_coeff > 1 &&
4932 bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time &&
4933 bfqq->service_from_wr > max_service_from_wr) {
4934 /* see comments on max_service_from_wr */
4935 bfq_bfqq_end_wr(bfqq);
4939 * To improve latency (for this or other queues), immediately
4940 * update weight both if it must be raised and if it must be
4941 * lowered. Since, entity may be on some active tree here, and
4942 * might have a pending change of its ioprio class, invoke
4943 * next function with the last parameter unset (see the
4944 * comments on the function).
4946 if ((entity->weight > entity->orig_weight) != (bfqq->wr_coeff > 1))
4947 __bfq_entity_update_weight_prio(bfq_entity_service_tree(entity),
4952 * Dispatch next request from bfqq.
4954 static struct request *bfq_dispatch_rq_from_bfqq(struct bfq_data *bfqd,
4955 struct bfq_queue *bfqq)
4957 struct request *rq = bfqq->next_rq;
4958 unsigned long service_to_charge;
4960 service_to_charge = bfq_serv_to_charge(rq, bfqq);
4962 bfq_bfqq_served(bfqq, service_to_charge);
4964 if (bfqq == bfqd->in_service_queue && bfqd->wait_dispatch) {
4965 bfqd->wait_dispatch = false;
4966 bfqd->waited_rq = rq;
4969 bfq_dispatch_remove(bfqd->queue, rq);
4971 if (bfqq != bfqd->in_service_queue)
4975 * If weight raising has to terminate for bfqq, then next
4976 * function causes an immediate update of bfqq's weight,
4977 * without waiting for next activation. As a consequence, on
4978 * expiration, bfqq will be timestamped as if has never been
4979 * weight-raised during this service slot, even if it has
4980 * received part or even most of the service as a
4981 * weight-raised queue. This inflates bfqq's timestamps, which
4982 * is beneficial, as bfqq is then more willing to leave the
4983 * device immediately to possible other weight-raised queues.
4985 bfq_update_wr_data(bfqd, bfqq);
4988 * Expire bfqq, pretending that its budget expired, if bfqq
4989 * belongs to CLASS_IDLE and other queues are waiting for
4992 if (!(bfq_tot_busy_queues(bfqd) > 1 && bfq_class_idle(bfqq)))
4995 bfq_bfqq_expire(bfqd, bfqq, false, BFQQE_BUDGET_EXHAUSTED);
5001 static bool bfq_has_work(struct blk_mq_hw_ctx *hctx)
5003 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5006 * Avoiding lock: a race on bfqd->queued should cause at
5007 * most a call to dispatch for nothing
5009 return !list_empty_careful(&bfqd->dispatch) ||
5010 READ_ONCE(bfqd->queued);
5013 static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
5015 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5016 struct request *rq = NULL;
5017 struct bfq_queue *bfqq = NULL;
5019 if (!list_empty(&bfqd->dispatch)) {
5020 rq = list_first_entry(&bfqd->dispatch, struct request,
5022 list_del_init(&rq->queuelist);
5028 * Increment counters here, because this
5029 * dispatch does not follow the standard
5030 * dispatch flow (where counters are
5035 goto inc_in_driver_start_rq;
5039 * We exploit the bfq_finish_requeue_request hook to
5040 * decrement rq_in_driver, but
5041 * bfq_finish_requeue_request will not be invoked on
5042 * this request. So, to avoid unbalance, just start
5043 * this request, without incrementing rq_in_driver. As
5044 * a negative consequence, rq_in_driver is deceptively
5045 * lower than it should be while this request is in
5046 * service. This may cause bfq_schedule_dispatch to be
5047 * invoked uselessly.
5049 * As for implementing an exact solution, the
5050 * bfq_finish_requeue_request hook, if defined, is
5051 * probably invoked also on this request. So, by
5052 * exploiting this hook, we could 1) increment
5053 * rq_in_driver here, and 2) decrement it in
5054 * bfq_finish_requeue_request. Such a solution would
5055 * let the value of the counter be always accurate,
5056 * but it would entail using an extra interface
5057 * function. This cost seems higher than the benefit,
5058 * being the frequency of non-elevator-private
5059 * requests very low.
5064 bfq_log(bfqd, "dispatch requests: %d busy queues",
5065 bfq_tot_busy_queues(bfqd));
5067 if (bfq_tot_busy_queues(bfqd) == 0)
5071 * Force device to serve one request at a time if
5072 * strict_guarantees is true. Forcing this service scheme is
5073 * currently the ONLY way to guarantee that the request
5074 * service order enforced by the scheduler is respected by a
5075 * queueing device. Otherwise the device is free even to make
5076 * some unlucky request wait for as long as the device
5079 * Of course, serving one request at a time may cause loss of
5082 if (bfqd->strict_guarantees && bfqd->rq_in_driver > 0)
5085 bfqq = bfq_select_queue(bfqd);
5089 rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq);
5092 inc_in_driver_start_rq:
5093 bfqd->rq_in_driver++;
5095 rq->rq_flags |= RQF_STARTED;
5101 #ifdef CONFIG_BFQ_CGROUP_DEBUG
5102 static void bfq_update_dispatch_stats(struct request_queue *q,
5104 struct bfq_queue *in_serv_queue,
5105 bool idle_timer_disabled)
5107 struct bfq_queue *bfqq = rq ? RQ_BFQQ(rq) : NULL;
5109 if (!idle_timer_disabled && !bfqq)
5113 * rq and bfqq are guaranteed to exist until this function
5114 * ends, for the following reasons. First, rq can be
5115 * dispatched to the device, and then can be completed and
5116 * freed, only after this function ends. Second, rq cannot be
5117 * merged (and thus freed because of a merge) any longer,
5118 * because it has already started. Thus rq cannot be freed
5119 * before this function ends, and, since rq has a reference to
5120 * bfqq, the same guarantee holds for bfqq too.
5122 * In addition, the following queue lock guarantees that
5123 * bfqq_group(bfqq) exists as well.
5125 spin_lock_irq(&q->queue_lock);
5126 if (idle_timer_disabled)
5128 * Since the idle timer has been disabled,
5129 * in_serv_queue contained some request when
5130 * __bfq_dispatch_request was invoked above, which
5131 * implies that rq was picked exactly from
5132 * in_serv_queue. Thus in_serv_queue == bfqq, and is
5133 * therefore guaranteed to exist because of the above
5136 bfqg_stats_update_idle_time(bfqq_group(in_serv_queue));
5138 struct bfq_group *bfqg = bfqq_group(bfqq);
5140 bfqg_stats_update_avg_queue_size(bfqg);
5141 bfqg_stats_set_start_empty_time(bfqg);
5142 bfqg_stats_update_io_remove(bfqg, rq->cmd_flags);
5144 spin_unlock_irq(&q->queue_lock);
5147 static inline void bfq_update_dispatch_stats(struct request_queue *q,
5149 struct bfq_queue *in_serv_queue,
5150 bool idle_timer_disabled) {}
5151 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
5153 static struct request *bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
5155 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5157 struct bfq_queue *in_serv_queue;
5158 bool waiting_rq, idle_timer_disabled = false;
5160 spin_lock_irq(&bfqd->lock);
5162 in_serv_queue = bfqd->in_service_queue;
5163 waiting_rq = in_serv_queue && bfq_bfqq_wait_request(in_serv_queue);
5165 rq = __bfq_dispatch_request(hctx);
5166 if (in_serv_queue == bfqd->in_service_queue) {
5167 idle_timer_disabled =
5168 waiting_rq && !bfq_bfqq_wait_request(in_serv_queue);
5171 spin_unlock_irq(&bfqd->lock);
5172 bfq_update_dispatch_stats(hctx->queue, rq,
5173 idle_timer_disabled ? in_serv_queue : NULL,
5174 idle_timer_disabled);
5180 * Task holds one reference to the queue, dropped when task exits. Each rq
5181 * in-flight on this queue also holds a reference, dropped when rq is freed.
5183 * Scheduler lock must be held here. Recall not to use bfqq after calling
5184 * this function on it.
5186 void bfq_put_queue(struct bfq_queue *bfqq)
5188 struct bfq_queue *item;
5189 struct hlist_node *n;
5190 struct bfq_group *bfqg = bfqq_group(bfqq);
5192 bfq_log_bfqq(bfqq->bfqd, bfqq, "put_queue: %p %d", bfqq, bfqq->ref);
5198 if (!hlist_unhashed(&bfqq->burst_list_node)) {
5199 hlist_del_init(&bfqq->burst_list_node);
5201 * Decrement also burst size after the removal, if the
5202 * process associated with bfqq is exiting, and thus
5203 * does not contribute to the burst any longer. This
5204 * decrement helps filter out false positives of large
5205 * bursts, when some short-lived process (often due to
5206 * the execution of commands by some service) happens
5207 * to start and exit while a complex application is
5208 * starting, and thus spawning several processes that
5209 * do I/O (and that *must not* be treated as a large
5210 * burst, see comments on bfq_handle_burst).
5212 * In particular, the decrement is performed only if:
5213 * 1) bfqq is not a merged queue, because, if it is,
5214 * then this free of bfqq is not triggered by the exit
5215 * of the process bfqq is associated with, but exactly
5216 * by the fact that bfqq has just been merged.
5217 * 2) burst_size is greater than 0, to handle
5218 * unbalanced decrements. Unbalanced decrements may
5219 * happen in te following case: bfqq is inserted into
5220 * the current burst list--without incrementing
5221 * bust_size--because of a split, but the current
5222 * burst list is not the burst list bfqq belonged to
5223 * (see comments on the case of a split in
5226 if (bfqq->bic && bfqq->bfqd->burst_size > 0)
5227 bfqq->bfqd->burst_size--;
5231 * bfqq does not exist any longer, so it cannot be woken by
5232 * any other queue, and cannot wake any other queue. Then bfqq
5233 * must be removed from the woken list of its possible waker
5234 * queue, and all queues in the woken list of bfqq must stop
5235 * having a waker queue. Strictly speaking, these updates
5236 * should be performed when bfqq remains with no I/O source
5237 * attached to it, which happens before bfqq gets freed. In
5238 * particular, this happens when the last process associated
5239 * with bfqq exits or gets associated with a different
5240 * queue. However, both events lead to bfqq being freed soon,
5241 * and dangling references would come out only after bfqq gets
5242 * freed. So these updates are done here, as a simple and safe
5243 * way to handle all cases.
5245 /* remove bfqq from woken list */
5246 if (!hlist_unhashed(&bfqq->woken_list_node))
5247 hlist_del_init(&bfqq->woken_list_node);
5249 /* reset waker for all queues in woken list */
5250 hlist_for_each_entry_safe(item, n, &bfqq->woken_list,
5252 item->waker_bfqq = NULL;
5253 hlist_del_init(&item->woken_list_node);
5256 if (bfqq->bfqd->last_completed_rq_bfqq == bfqq)
5257 bfqq->bfqd->last_completed_rq_bfqq = NULL;
5259 kmem_cache_free(bfq_pool, bfqq);
5260 bfqg_and_blkg_put(bfqg);
5263 static void bfq_put_stable_ref(struct bfq_queue *bfqq)
5266 bfq_put_queue(bfqq);
5269 void bfq_put_cooperator(struct bfq_queue *bfqq)
5271 struct bfq_queue *__bfqq, *next;
5274 * If this queue was scheduled to merge with another queue, be
5275 * sure to drop the reference taken on that queue (and others in
5276 * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs.
5278 __bfqq = bfqq->new_bfqq;
5282 next = __bfqq->new_bfqq;
5283 bfq_put_queue(__bfqq);
5288 static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq)
5290 if (bfqq == bfqd->in_service_queue) {
5291 __bfq_bfqq_expire(bfqd, bfqq, BFQQE_BUDGET_TIMEOUT);
5292 bfq_schedule_dispatch(bfqd);
5295 bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, bfqq->ref);
5297 bfq_put_cooperator(bfqq);
5299 bfq_release_process_ref(bfqd, bfqq);
5302 static void bfq_exit_icq_bfqq(struct bfq_io_cq *bic, bool is_sync)
5304 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
5305 struct bfq_data *bfqd;
5308 bfqd = bfqq->bfqd; /* NULL if scheduler already exited */
5311 unsigned long flags;
5313 spin_lock_irqsave(&bfqd->lock, flags);
5315 bfq_exit_bfqq(bfqd, bfqq);
5316 bic_set_bfqq(bic, NULL, is_sync);
5317 spin_unlock_irqrestore(&bfqd->lock, flags);
5321 static void bfq_exit_icq(struct io_cq *icq)
5323 struct bfq_io_cq *bic = icq_to_bic(icq);
5325 if (bic->stable_merge_bfqq) {
5326 struct bfq_data *bfqd = bic->stable_merge_bfqq->bfqd;
5329 * bfqd is NULL if scheduler already exited, and in
5330 * that case this is the last time bfqq is accessed.
5333 unsigned long flags;
5335 spin_lock_irqsave(&bfqd->lock, flags);
5336 bfq_put_stable_ref(bic->stable_merge_bfqq);
5337 spin_unlock_irqrestore(&bfqd->lock, flags);
5339 bfq_put_stable_ref(bic->stable_merge_bfqq);
5343 bfq_exit_icq_bfqq(bic, true);
5344 bfq_exit_icq_bfqq(bic, false);
5348 * Update the entity prio values; note that the new values will not
5349 * be used until the next (re)activation.
5352 bfq_set_next_ioprio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic)
5354 struct task_struct *tsk = current;
5356 struct bfq_data *bfqd = bfqq->bfqd;
5361 ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
5362 switch (ioprio_class) {
5364 pr_err("bdi %s: bfq: bad prio class %d\n",
5365 bdi_dev_name(bfqq->bfqd->queue->disk->bdi),
5368 case IOPRIO_CLASS_NONE:
5370 * No prio set, inherit CPU scheduling settings.
5372 bfqq->new_ioprio = task_nice_ioprio(tsk);
5373 bfqq->new_ioprio_class = task_nice_ioclass(tsk);
5375 case IOPRIO_CLASS_RT:
5376 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
5377 bfqq->new_ioprio_class = IOPRIO_CLASS_RT;
5379 case IOPRIO_CLASS_BE:
5380 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
5381 bfqq->new_ioprio_class = IOPRIO_CLASS_BE;
5383 case IOPRIO_CLASS_IDLE:
5384 bfqq->new_ioprio_class = IOPRIO_CLASS_IDLE;
5385 bfqq->new_ioprio = 7;
5389 if (bfqq->new_ioprio >= IOPRIO_NR_LEVELS) {
5390 pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n",
5392 bfqq->new_ioprio = IOPRIO_NR_LEVELS - 1;
5395 bfqq->entity.new_weight = bfq_ioprio_to_weight(bfqq->new_ioprio);
5396 bfq_log_bfqq(bfqd, bfqq, "new_ioprio %d new_weight %d",
5397 bfqq->new_ioprio, bfqq->entity.new_weight);
5398 bfqq->entity.prio_changed = 1;
5401 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
5402 struct bio *bio, bool is_sync,
5403 struct bfq_io_cq *bic,
5406 static void bfq_check_ioprio_change(struct bfq_io_cq *bic, struct bio *bio)
5408 struct bfq_data *bfqd = bic_to_bfqd(bic);
5409 struct bfq_queue *bfqq;
5410 int ioprio = bic->icq.ioc->ioprio;
5413 * This condition may trigger on a newly created bic, be sure to
5414 * drop the lock before returning.
5416 if (unlikely(!bfqd) || likely(bic->ioprio == ioprio))
5419 bic->ioprio = ioprio;
5421 bfqq = bic_to_bfqq(bic, false);
5423 bfq_release_process_ref(bfqd, bfqq);
5424 bfqq = bfq_get_queue(bfqd, bio, false, bic, true);
5425 bic_set_bfqq(bic, bfqq, false);
5428 bfqq = bic_to_bfqq(bic, true);
5430 bfq_set_next_ioprio_data(bfqq, bic);
5433 static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5434 struct bfq_io_cq *bic, pid_t pid, int is_sync)
5436 u64 now_ns = ktime_get_ns();
5438 RB_CLEAR_NODE(&bfqq->entity.rb_node);
5439 INIT_LIST_HEAD(&bfqq->fifo);
5440 INIT_HLIST_NODE(&bfqq->burst_list_node);
5441 INIT_HLIST_NODE(&bfqq->woken_list_node);
5442 INIT_HLIST_HEAD(&bfqq->woken_list);
5448 bfq_set_next_ioprio_data(bfqq, bic);
5452 * No need to mark as has_short_ttime if in
5453 * idle_class, because no device idling is performed
5454 * for queues in idle class
5456 if (!bfq_class_idle(bfqq))
5457 /* tentatively mark as has_short_ttime */
5458 bfq_mark_bfqq_has_short_ttime(bfqq);
5459 bfq_mark_bfqq_sync(bfqq);
5460 bfq_mark_bfqq_just_created(bfqq);
5462 bfq_clear_bfqq_sync(bfqq);
5464 /* set end request to minus infinity from now */
5465 bfqq->ttime.last_end_request = now_ns + 1;
5467 bfqq->creation_time = jiffies;
5469 bfqq->io_start_time = now_ns;
5471 bfq_mark_bfqq_IO_bound(bfqq);
5475 /* Tentative initial value to trade off between thr and lat */
5476 bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3;
5477 bfqq->budget_timeout = bfq_smallest_from_now();
5480 bfqq->last_wr_start_finish = jiffies;
5481 bfqq->wr_start_at_switch_to_srt = bfq_smallest_from_now();
5482 bfqq->split_time = bfq_smallest_from_now();
5485 * To not forget the possibly high bandwidth consumed by a
5486 * process/queue in the recent past,
5487 * bfq_bfqq_softrt_next_start() returns a value at least equal
5488 * to the current value of bfqq->soft_rt_next_start (see
5489 * comments on bfq_bfqq_softrt_next_start). Set
5490 * soft_rt_next_start to now, to mean that bfqq has consumed
5491 * no bandwidth so far.
5493 bfqq->soft_rt_next_start = jiffies;
5495 /* first request is almost certainly seeky */
5496 bfqq->seek_history = 1;
5499 static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd,
5500 struct bfq_group *bfqg,
5501 int ioprio_class, int ioprio)
5503 switch (ioprio_class) {
5504 case IOPRIO_CLASS_RT:
5505 return &bfqg->async_bfqq[0][ioprio];
5506 case IOPRIO_CLASS_NONE:
5507 ioprio = IOPRIO_BE_NORM;
5509 case IOPRIO_CLASS_BE:
5510 return &bfqg->async_bfqq[1][ioprio];
5511 case IOPRIO_CLASS_IDLE:
5512 return &bfqg->async_idle_bfqq;
5518 static struct bfq_queue *
5519 bfq_do_early_stable_merge(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5520 struct bfq_io_cq *bic,
5521 struct bfq_queue *last_bfqq_created)
5523 struct bfq_queue *new_bfqq =
5524 bfq_setup_merge(bfqq, last_bfqq_created);
5530 new_bfqq->bic->stably_merged = true;
5531 bic->stably_merged = true;
5534 * Reusing merge functions. This implies that
5535 * bfqq->bic must be set too, for
5536 * bfq_merge_bfqqs to correctly save bfqq's
5537 * state before killing it.
5540 bfq_merge_bfqqs(bfqd, bic, bfqq, new_bfqq);
5546 * Many throughput-sensitive workloads are made of several parallel
5547 * I/O flows, with all flows generated by the same application, or
5548 * more generically by the same task (e.g., system boot). The most
5549 * counterproductive action with these workloads is plugging I/O
5550 * dispatch when one of the bfq_queues associated with these flows
5551 * remains temporarily empty.
5553 * To avoid this plugging, BFQ has been using a burst-handling
5554 * mechanism for years now. This mechanism has proven effective for
5555 * throughput, and not detrimental for service guarantees. The
5556 * following function pushes this mechanism a little bit further,
5557 * basing on the following two facts.
5559 * First, all the I/O flows of a the same application or task
5560 * contribute to the execution/completion of that common application
5561 * or task. So the performance figures that matter are total
5562 * throughput of the flows and task-wide I/O latency. In particular,
5563 * these flows do not need to be protected from each other, in terms
5564 * of individual bandwidth or latency.
5566 * Second, the above fact holds regardless of the number of flows.
5568 * Putting these two facts together, this commits merges stably the
5569 * bfq_queues associated with these I/O flows, i.e., with the
5570 * processes that generate these IO/ flows, regardless of how many the
5571 * involved processes are.
5573 * To decide whether a set of bfq_queues is actually associated with
5574 * the I/O flows of a common application or task, and to merge these
5575 * queues stably, this function operates as follows: given a bfq_queue,
5576 * say Q2, currently being created, and the last bfq_queue, say Q1,
5577 * created before Q2, Q2 is merged stably with Q1 if
5578 * - very little time has elapsed since when Q1 was created
5579 * - Q2 has the same ioprio as Q1
5580 * - Q2 belongs to the same group as Q1
5582 * Merging bfq_queues also reduces scheduling overhead. A fio test
5583 * with ten random readers on /dev/nullb shows a throughput boost of
5584 * 40%, with a quadcore. Since BFQ's execution time amounts to ~50% of
5585 * the total per-request processing time, the above throughput boost
5586 * implies that BFQ's overhead is reduced by more than 50%.
5588 * This new mechanism most certainly obsoletes the current
5589 * burst-handling heuristics. We keep those heuristics for the moment.
5591 static struct bfq_queue *bfq_do_or_sched_stable_merge(struct bfq_data *bfqd,
5592 struct bfq_queue *bfqq,
5593 struct bfq_io_cq *bic)
5595 struct bfq_queue **source_bfqq = bfqq->entity.parent ?
5596 &bfqq->entity.parent->last_bfqq_created :
5597 &bfqd->last_bfqq_created;
5599 struct bfq_queue *last_bfqq_created = *source_bfqq;
5602 * If last_bfqq_created has not been set yet, then init it. If
5603 * it has been set already, but too long ago, then move it
5604 * forward to bfqq. Finally, move also if bfqq belongs to a
5605 * different group than last_bfqq_created, or if bfqq has a
5606 * different ioprio or ioprio_class. If none of these
5607 * conditions holds true, then try an early stable merge or
5608 * schedule a delayed stable merge.
5610 * A delayed merge is scheduled (instead of performing an
5611 * early merge), in case bfqq might soon prove to be more
5612 * throughput-beneficial if not merged. Currently this is
5613 * possible only if bfqd is rotational with no queueing. For
5614 * such a drive, not merging bfqq is better for throughput if
5615 * bfqq happens to contain sequential I/O. So, we wait a
5616 * little bit for enough I/O to flow through bfqq. After that,
5617 * if such an I/O is sequential, then the merge is
5618 * canceled. Otherwise the merge is finally performed.
5620 if (!last_bfqq_created ||
5621 time_before(last_bfqq_created->creation_time +
5622 msecs_to_jiffies(bfq_activation_stable_merging),
5623 bfqq->creation_time) ||
5624 bfqq->entity.parent != last_bfqq_created->entity.parent ||
5625 bfqq->ioprio != last_bfqq_created->ioprio ||
5626 bfqq->ioprio_class != last_bfqq_created->ioprio_class)
5627 *source_bfqq = bfqq;
5628 else if (time_after_eq(last_bfqq_created->creation_time +
5629 bfqd->bfq_burst_interval,
5630 bfqq->creation_time)) {
5631 if (likely(bfqd->nonrot_with_queueing))
5633 * With this type of drive, leaving
5634 * bfqq alone may provide no
5635 * throughput benefits compared with
5636 * merging bfqq. So merge bfqq now.
5638 bfqq = bfq_do_early_stable_merge(bfqd, bfqq,
5641 else { /* schedule tentative stable merge */
5643 * get reference on last_bfqq_created,
5644 * to prevent it from being freed,
5645 * until we decide whether to merge
5647 last_bfqq_created->ref++;
5649 * need to keep track of stable refs, to
5650 * compute process refs correctly
5652 last_bfqq_created->stable_ref++;
5654 * Record the bfqq to merge to.
5656 bic->stable_merge_bfqq = last_bfqq_created;
5664 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
5665 struct bio *bio, bool is_sync,
5666 struct bfq_io_cq *bic,
5669 const int ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
5670 const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
5671 struct bfq_queue **async_bfqq = NULL;
5672 struct bfq_queue *bfqq;
5673 struct bfq_group *bfqg;
5675 bfqg = bfq_bio_bfqg(bfqd, bio);
5677 async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class,
5684 bfqq = kmem_cache_alloc_node(bfq_pool,
5685 GFP_NOWAIT | __GFP_ZERO | __GFP_NOWARN,
5689 bfq_init_bfqq(bfqd, bfqq, bic, current->pid,
5691 bfq_init_entity(&bfqq->entity, bfqg);
5692 bfq_log_bfqq(bfqd, bfqq, "allocated");
5694 bfqq = &bfqd->oom_bfqq;
5695 bfq_log_bfqq(bfqd, bfqq, "using oom bfqq");
5700 * Pin the queue now that it's allocated, scheduler exit will
5705 * Extra group reference, w.r.t. sync
5706 * queue. This extra reference is removed
5707 * only if bfqq->bfqg disappears, to
5708 * guarantee that this queue is not freed
5709 * until its group goes away.
5711 bfq_log_bfqq(bfqd, bfqq, "get_queue, bfqq not in async: %p, %d",
5717 bfqq->ref++; /* get a process reference to this queue */
5719 if (bfqq != &bfqd->oom_bfqq && is_sync && !respawn)
5720 bfqq = bfq_do_or_sched_stable_merge(bfqd, bfqq, bic);
5724 static void bfq_update_io_thinktime(struct bfq_data *bfqd,
5725 struct bfq_queue *bfqq)
5727 struct bfq_ttime *ttime = &bfqq->ttime;
5731 * We are really interested in how long it takes for the queue to
5732 * become busy when there is no outstanding IO for this queue. So
5733 * ignore cases when the bfq queue has already IO queued.
5735 if (bfqq->dispatched || bfq_bfqq_busy(bfqq))
5737 elapsed = ktime_get_ns() - bfqq->ttime.last_end_request;
5738 elapsed = min_t(u64, elapsed, 2ULL * bfqd->bfq_slice_idle);
5740 ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
5741 ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed, 8);
5742 ttime->ttime_mean = div64_ul(ttime->ttime_total + 128,
5743 ttime->ttime_samples);
5747 bfq_update_io_seektime(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5750 bfqq->seek_history <<= 1;
5751 bfqq->seek_history |= BFQ_RQ_SEEKY(bfqd, bfqq->last_request_pos, rq);
5753 if (bfqq->wr_coeff > 1 &&
5754 bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
5755 BFQQ_TOTALLY_SEEKY(bfqq)) {
5756 if (time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
5757 bfq_wr_duration(bfqd))) {
5759 * In soft_rt weight raising with the
5760 * interactive-weight-raising period
5761 * elapsed (so no switch back to
5762 * interactive weight raising).
5764 bfq_bfqq_end_wr(bfqq);
5766 * stopping soft_rt weight raising
5767 * while still in interactive period,
5768 * switch back to interactive weight
5771 switch_back_to_interactive_wr(bfqq, bfqd);
5772 bfqq->entity.prio_changed = 1;
5777 static void bfq_update_has_short_ttime(struct bfq_data *bfqd,
5778 struct bfq_queue *bfqq,
5779 struct bfq_io_cq *bic)
5781 bool has_short_ttime = true, state_changed;
5784 * No need to update has_short_ttime if bfqq is async or in
5785 * idle io prio class, or if bfq_slice_idle is zero, because
5786 * no device idling is performed for bfqq in this case.
5788 if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq) ||
5789 bfqd->bfq_slice_idle == 0)
5792 /* Idle window just restored, statistics are meaningless. */
5793 if (time_is_after_eq_jiffies(bfqq->split_time +
5794 bfqd->bfq_wr_min_idle_time))
5797 /* Think time is infinite if no process is linked to
5798 * bfqq. Otherwise check average think time to decide whether
5799 * to mark as has_short_ttime. To this goal, compare average
5800 * think time with half the I/O-plugging timeout.
5802 if (atomic_read(&bic->icq.ioc->active_ref) == 0 ||
5803 (bfq_sample_valid(bfqq->ttime.ttime_samples) &&
5804 bfqq->ttime.ttime_mean > bfqd->bfq_slice_idle>>1))
5805 has_short_ttime = false;
5807 state_changed = has_short_ttime != bfq_bfqq_has_short_ttime(bfqq);
5809 if (has_short_ttime)
5810 bfq_mark_bfqq_has_short_ttime(bfqq);
5812 bfq_clear_bfqq_has_short_ttime(bfqq);
5815 * Until the base value for the total service time gets
5816 * finally computed for bfqq, the inject limit does depend on
5817 * the think-time state (short|long). In particular, the limit
5818 * is 0 or 1 if the think time is deemed, respectively, as
5819 * short or long (details in the comments in
5820 * bfq_update_inject_limit()). Accordingly, the next
5821 * instructions reset the inject limit if the think-time state
5822 * has changed and the above base value is still to be
5825 * However, the reset is performed only if more than 100 ms
5826 * have elapsed since the last update of the inject limit, or
5827 * (inclusive) if the change is from short to long think
5828 * time. The reason for this waiting is as follows.
5830 * bfqq may have a long think time because of a
5831 * synchronization with some other queue, i.e., because the
5832 * I/O of some other queue may need to be completed for bfqq
5833 * to receive new I/O. Details in the comments on the choice
5834 * of the queue for injection in bfq_select_queue().
5836 * As stressed in those comments, if such a synchronization is
5837 * actually in place, then, without injection on bfqq, the
5838 * blocking I/O cannot happen to served while bfqq is in
5839 * service. As a consequence, if bfqq is granted
5840 * I/O-dispatch-plugging, then bfqq remains empty, and no I/O
5841 * is dispatched, until the idle timeout fires. This is likely
5842 * to result in lower bandwidth and higher latencies for bfqq,
5843 * and in a severe loss of total throughput.
5845 * On the opposite end, a non-zero inject limit may allow the
5846 * I/O that blocks bfqq to be executed soon, and therefore
5847 * bfqq to receive new I/O soon.
5849 * But, if the blocking gets actually eliminated, then the
5850 * next think-time sample for bfqq may be very low. This in
5851 * turn may cause bfqq's think time to be deemed
5852 * short. Without the 100 ms barrier, this new state change
5853 * would cause the body of the next if to be executed
5854 * immediately. But this would set to 0 the inject
5855 * limit. Without injection, the blocking I/O would cause the
5856 * think time of bfqq to become long again, and therefore the
5857 * inject limit to be raised again, and so on. The only effect
5858 * of such a steady oscillation between the two think-time
5859 * states would be to prevent effective injection on bfqq.
5861 * In contrast, if the inject limit is not reset during such a
5862 * long time interval as 100 ms, then the number of short
5863 * think time samples can grow significantly before the reset
5864 * is performed. As a consequence, the think time state can
5865 * become stable before the reset. Therefore there will be no
5866 * state change when the 100 ms elapse, and no reset of the
5867 * inject limit. The inject limit remains steadily equal to 1
5868 * both during and after the 100 ms. So injection can be
5869 * performed at all times, and throughput gets boosted.
5871 * An inject limit equal to 1 is however in conflict, in
5872 * general, with the fact that the think time of bfqq is
5873 * short, because injection may be likely to delay bfqq's I/O
5874 * (as explained in the comments in
5875 * bfq_update_inject_limit()). But this does not happen in
5876 * this special case, because bfqq's low think time is due to
5877 * an effective handling of a synchronization, through
5878 * injection. In this special case, bfqq's I/O does not get
5879 * delayed by injection; on the contrary, bfqq's I/O is
5880 * brought forward, because it is not blocked for
5883 * In addition, serving the blocking I/O much sooner, and much
5884 * more frequently than once per I/O-plugging timeout, makes
5885 * it much quicker to detect a waker queue (the concept of
5886 * waker queue is defined in the comments in
5887 * bfq_add_request()). This makes it possible to start sooner
5888 * to boost throughput more effectively, by injecting the I/O
5889 * of the waker queue unconditionally on every
5890 * bfq_dispatch_request().
5892 * One last, important benefit of not resetting the inject
5893 * limit before 100 ms is that, during this time interval, the
5894 * base value for the total service time is likely to get
5895 * finally computed for bfqq, freeing the inject limit from
5896 * its relation with the think time.
5898 if (state_changed && bfqq->last_serv_time_ns == 0 &&
5899 (time_is_before_eq_jiffies(bfqq->decrease_time_jif +
5900 msecs_to_jiffies(100)) ||
5902 bfq_reset_inject_limit(bfqd, bfqq);
5906 * Called when a new fs request (rq) is added to bfqq. Check if there's
5907 * something we should do about it.
5909 static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5912 if (rq->cmd_flags & REQ_META)
5913 bfqq->meta_pending++;
5915 bfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
5917 if (bfqq == bfqd->in_service_queue && bfq_bfqq_wait_request(bfqq)) {
5918 bool small_req = bfqq->queued[rq_is_sync(rq)] == 1 &&
5919 blk_rq_sectors(rq) < 32;
5920 bool budget_timeout = bfq_bfqq_budget_timeout(bfqq);
5923 * There is just this request queued: if
5924 * - the request is small, and
5925 * - we are idling to boost throughput, and
5926 * - the queue is not to be expired,
5929 * In this way, if the device is being idled to wait
5930 * for a new request from the in-service queue, we
5931 * avoid unplugging the device and committing the
5932 * device to serve just a small request. In contrast
5933 * we wait for the block layer to decide when to
5934 * unplug the device: hopefully, new requests will be
5935 * merged to this one quickly, then the device will be
5936 * unplugged and larger requests will be dispatched.
5938 if (small_req && idling_boosts_thr_without_issues(bfqd, bfqq) &&
5943 * A large enough request arrived, or idling is being
5944 * performed to preserve service guarantees, or
5945 * finally the queue is to be expired: in all these
5946 * cases disk idling is to be stopped, so clear
5947 * wait_request flag and reset timer.
5949 bfq_clear_bfqq_wait_request(bfqq);
5950 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
5953 * The queue is not empty, because a new request just
5954 * arrived. Hence we can safely expire the queue, in
5955 * case of budget timeout, without risking that the
5956 * timestamps of the queue are not updated correctly.
5957 * See [1] for more details.
5960 bfq_bfqq_expire(bfqd, bfqq, false,
5961 BFQQE_BUDGET_TIMEOUT);
5965 static void bfqq_request_allocated(struct bfq_queue *bfqq)
5967 struct bfq_entity *entity = &bfqq->entity;
5969 for_each_entity(entity)
5970 entity->allocated++;
5973 static void bfqq_request_freed(struct bfq_queue *bfqq)
5975 struct bfq_entity *entity = &bfqq->entity;
5977 for_each_entity(entity)
5978 entity->allocated--;
5981 /* returns true if it causes the idle timer to be disabled */
5982 static bool __bfq_insert_request(struct bfq_data *bfqd, struct request *rq)
5984 struct bfq_queue *bfqq = RQ_BFQQ(rq),
5985 *new_bfqq = bfq_setup_cooperator(bfqd, bfqq, rq, true,
5987 bool waiting, idle_timer_disabled = false;
5991 * Release the request's reference to the old bfqq
5992 * and make sure one is taken to the shared queue.
5994 bfqq_request_allocated(new_bfqq);
5995 bfqq_request_freed(bfqq);
5998 * If the bic associated with the process
5999 * issuing this request still points to bfqq
6000 * (and thus has not been already redirected
6001 * to new_bfqq or even some other bfq_queue),
6002 * then complete the merge and redirect it to
6005 if (bic_to_bfqq(RQ_BIC(rq), 1) == bfqq)
6006 bfq_merge_bfqqs(bfqd, RQ_BIC(rq),
6009 bfq_clear_bfqq_just_created(bfqq);
6011 * rq is about to be enqueued into new_bfqq,
6012 * release rq reference on bfqq
6014 bfq_put_queue(bfqq);
6015 rq->elv.priv[1] = new_bfqq;
6019 bfq_update_io_thinktime(bfqd, bfqq);
6020 bfq_update_has_short_ttime(bfqd, bfqq, RQ_BIC(rq));
6021 bfq_update_io_seektime(bfqd, bfqq, rq);
6023 waiting = bfqq && bfq_bfqq_wait_request(bfqq);
6024 bfq_add_request(rq);
6025 idle_timer_disabled = waiting && !bfq_bfqq_wait_request(bfqq);
6027 rq->fifo_time = ktime_get_ns() + bfqd->bfq_fifo_expire[rq_is_sync(rq)];
6028 list_add_tail(&rq->queuelist, &bfqq->fifo);
6030 bfq_rq_enqueued(bfqd, bfqq, rq);
6032 return idle_timer_disabled;
6035 #ifdef CONFIG_BFQ_CGROUP_DEBUG
6036 static void bfq_update_insert_stats(struct request_queue *q,
6037 struct bfq_queue *bfqq,
6038 bool idle_timer_disabled,
6039 blk_opf_t cmd_flags)
6045 * bfqq still exists, because it can disappear only after
6046 * either it is merged with another queue, or the process it
6047 * is associated with exits. But both actions must be taken by
6048 * the same process currently executing this flow of
6051 * In addition, the following queue lock guarantees that
6052 * bfqq_group(bfqq) exists as well.
6054 spin_lock_irq(&q->queue_lock);
6055 bfqg_stats_update_io_add(bfqq_group(bfqq), bfqq, cmd_flags);
6056 if (idle_timer_disabled)
6057 bfqg_stats_update_idle_time(bfqq_group(bfqq));
6058 spin_unlock_irq(&q->queue_lock);
6061 static inline void bfq_update_insert_stats(struct request_queue *q,
6062 struct bfq_queue *bfqq,
6063 bool idle_timer_disabled,
6064 blk_opf_t cmd_flags) {}
6065 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
6067 static struct bfq_queue *bfq_init_rq(struct request *rq);
6069 static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
6072 struct request_queue *q = hctx->queue;
6073 struct bfq_data *bfqd = q->elevator->elevator_data;
6074 struct bfq_queue *bfqq;
6075 bool idle_timer_disabled = false;
6076 blk_opf_t cmd_flags;
6079 #ifdef CONFIG_BFQ_GROUP_IOSCHED
6080 if (!cgroup_subsys_on_dfl(io_cgrp_subsys) && rq->bio)
6081 bfqg_stats_update_legacy_io(q, rq);
6083 spin_lock_irq(&bfqd->lock);
6084 bfqq = bfq_init_rq(rq);
6085 if (blk_mq_sched_try_insert_merge(q, rq, &free)) {
6086 spin_unlock_irq(&bfqd->lock);
6087 blk_mq_free_requests(&free);
6091 trace_block_rq_insert(rq);
6093 if (!bfqq || at_head) {
6095 list_add(&rq->queuelist, &bfqd->dispatch);
6097 list_add_tail(&rq->queuelist, &bfqd->dispatch);
6099 idle_timer_disabled = __bfq_insert_request(bfqd, rq);
6101 * Update bfqq, because, if a queue merge has occurred
6102 * in __bfq_insert_request, then rq has been
6103 * redirected into a new queue.
6107 if (rq_mergeable(rq)) {
6108 elv_rqhash_add(q, rq);
6115 * Cache cmd_flags before releasing scheduler lock, because rq
6116 * may disappear afterwards (for example, because of a request
6119 cmd_flags = rq->cmd_flags;
6120 spin_unlock_irq(&bfqd->lock);
6122 bfq_update_insert_stats(q, bfqq, idle_timer_disabled,
6126 static void bfq_insert_requests(struct blk_mq_hw_ctx *hctx,
6127 struct list_head *list, bool at_head)
6129 while (!list_empty(list)) {
6132 rq = list_first_entry(list, struct request, queuelist);
6133 list_del_init(&rq->queuelist);
6134 bfq_insert_request(hctx, rq, at_head);
6138 static void bfq_update_hw_tag(struct bfq_data *bfqd)
6140 struct bfq_queue *bfqq = bfqd->in_service_queue;
6142 bfqd->max_rq_in_driver = max_t(int, bfqd->max_rq_in_driver,
6143 bfqd->rq_in_driver);
6145 if (bfqd->hw_tag == 1)
6149 * This sample is valid if the number of outstanding requests
6150 * is large enough to allow a queueing behavior. Note that the
6151 * sum is not exact, as it's not taking into account deactivated
6154 if (bfqd->rq_in_driver + bfqd->queued <= BFQ_HW_QUEUE_THRESHOLD)
6158 * If active queue hasn't enough requests and can idle, bfq might not
6159 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
6162 if (bfqq && bfq_bfqq_has_short_ttime(bfqq) &&
6163 bfqq->dispatched + bfqq->queued[0] + bfqq->queued[1] <
6164 BFQ_HW_QUEUE_THRESHOLD &&
6165 bfqd->rq_in_driver < BFQ_HW_QUEUE_THRESHOLD)
6168 if (bfqd->hw_tag_samples++ < BFQ_HW_QUEUE_SAMPLES)
6171 bfqd->hw_tag = bfqd->max_rq_in_driver > BFQ_HW_QUEUE_THRESHOLD;
6172 bfqd->max_rq_in_driver = 0;
6173 bfqd->hw_tag_samples = 0;
6175 bfqd->nonrot_with_queueing =
6176 blk_queue_nonrot(bfqd->queue) && bfqd->hw_tag;
6179 static void bfq_completed_request(struct bfq_queue *bfqq, struct bfq_data *bfqd)
6184 bfq_update_hw_tag(bfqd);
6186 bfqd->rq_in_driver--;
6189 if (!bfqq->dispatched && !bfq_bfqq_busy(bfqq)) {
6191 * Set budget_timeout (which we overload to store the
6192 * time at which the queue remains with no backlog and
6193 * no outstanding request; used by the weight-raising
6196 bfqq->budget_timeout = jiffies;
6198 bfq_del_bfqq_in_groups_with_pending_reqs(bfqq);
6199 bfq_weights_tree_remove(bfqq);
6202 now_ns = ktime_get_ns();
6204 bfqq->ttime.last_end_request = now_ns;
6207 * Using us instead of ns, to get a reasonable precision in
6208 * computing rate in next check.
6210 delta_us = div_u64(now_ns - bfqd->last_completion, NSEC_PER_USEC);
6213 * If the request took rather long to complete, and, according
6214 * to the maximum request size recorded, this completion latency
6215 * implies that the request was certainly served at a very low
6216 * rate (less than 1M sectors/sec), then the whole observation
6217 * interval that lasts up to this time instant cannot be a
6218 * valid time interval for computing a new peak rate. Invoke
6219 * bfq_update_rate_reset to have the following three steps
6221 * - close the observation interval at the last (previous)
6222 * request dispatch or completion
6223 * - compute rate, if possible, for that observation interval
6224 * - reset to zero samples, which will trigger a proper
6225 * re-initialization of the observation interval on next
6228 if (delta_us > BFQ_MIN_TT/NSEC_PER_USEC &&
6229 (bfqd->last_rq_max_size<<BFQ_RATE_SHIFT)/delta_us <
6230 1UL<<(BFQ_RATE_SHIFT - 10))
6231 bfq_update_rate_reset(bfqd, NULL);
6232 bfqd->last_completion = now_ns;
6234 * Shared queues are likely to receive I/O at a high
6235 * rate. This may deceptively let them be considered as wakers
6236 * of other queues. But a false waker will unjustly steal
6237 * bandwidth to its supposedly woken queue. So considering
6238 * also shared queues in the waking mechanism may cause more
6239 * control troubles than throughput benefits. Then reset
6240 * last_completed_rq_bfqq if bfqq is a shared queue.
6242 if (!bfq_bfqq_coop(bfqq))
6243 bfqd->last_completed_rq_bfqq = bfqq;
6245 bfqd->last_completed_rq_bfqq = NULL;
6248 * If we are waiting to discover whether the request pattern
6249 * of the task associated with the queue is actually
6250 * isochronous, and both requisites for this condition to hold
6251 * are now satisfied, then compute soft_rt_next_start (see the
6252 * comments on the function bfq_bfqq_softrt_next_start()). We
6253 * do not compute soft_rt_next_start if bfqq is in interactive
6254 * weight raising (see the comments in bfq_bfqq_expire() for
6255 * an explanation). We schedule this delayed update when bfqq
6256 * expires, if it still has in-flight requests.
6258 if (bfq_bfqq_softrt_update(bfqq) && bfqq->dispatched == 0 &&
6259 RB_EMPTY_ROOT(&bfqq->sort_list) &&
6260 bfqq->wr_coeff != bfqd->bfq_wr_coeff)
6261 bfqq->soft_rt_next_start =
6262 bfq_bfqq_softrt_next_start(bfqd, bfqq);
6265 * If this is the in-service queue, check if it needs to be expired,
6266 * or if we want to idle in case it has no pending requests.
6268 if (bfqd->in_service_queue == bfqq) {
6269 if (bfq_bfqq_must_idle(bfqq)) {
6270 if (bfqq->dispatched == 0)
6271 bfq_arm_slice_timer(bfqd);
6273 * If we get here, we do not expire bfqq, even
6274 * if bfqq was in budget timeout or had no
6275 * more requests (as controlled in the next
6276 * conditional instructions). The reason for
6277 * not expiring bfqq is as follows.
6279 * Here bfqq->dispatched > 0 holds, but
6280 * bfq_bfqq_must_idle() returned true. This
6281 * implies that, even if no request arrives
6282 * for bfqq before bfqq->dispatched reaches 0,
6283 * bfqq will, however, not be expired on the
6284 * completion event that causes bfqq->dispatch
6285 * to reach zero. In contrast, on this event,
6286 * bfqq will start enjoying device idling
6287 * (I/O-dispatch plugging).
6289 * But, if we expired bfqq here, bfqq would
6290 * not have the chance to enjoy device idling
6291 * when bfqq->dispatched finally reaches
6292 * zero. This would expose bfqq to violation
6293 * of its reserved service guarantees.
6296 } else if (bfq_may_expire_for_budg_timeout(bfqq))
6297 bfq_bfqq_expire(bfqd, bfqq, false,
6298 BFQQE_BUDGET_TIMEOUT);
6299 else if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
6300 (bfqq->dispatched == 0 ||
6301 !bfq_better_to_idle(bfqq)))
6302 bfq_bfqq_expire(bfqd, bfqq, false,
6303 BFQQE_NO_MORE_REQUESTS);
6306 if (!bfqd->rq_in_driver)
6307 bfq_schedule_dispatch(bfqd);
6311 * The processes associated with bfqq may happen to generate their
6312 * cumulative I/O at a lower rate than the rate at which the device
6313 * could serve the same I/O. This is rather probable, e.g., if only
6314 * one process is associated with bfqq and the device is an SSD. It
6315 * results in bfqq becoming often empty while in service. In this
6316 * respect, if BFQ is allowed to switch to another queue when bfqq
6317 * remains empty, then the device goes on being fed with I/O requests,
6318 * and the throughput is not affected. In contrast, if BFQ is not
6319 * allowed to switch to another queue---because bfqq is sync and
6320 * I/O-dispatch needs to be plugged while bfqq is temporarily
6321 * empty---then, during the service of bfqq, there will be frequent
6322 * "service holes", i.e., time intervals during which bfqq gets empty
6323 * and the device can only consume the I/O already queued in its
6324 * hardware queues. During service holes, the device may even get to
6325 * remaining idle. In the end, during the service of bfqq, the device
6326 * is driven at a lower speed than the one it can reach with the kind
6327 * of I/O flowing through bfqq.
6329 * To counter this loss of throughput, BFQ implements a "request
6330 * injection mechanism", which tries to fill the above service holes
6331 * with I/O requests taken from other queues. The hard part in this
6332 * mechanism is finding the right amount of I/O to inject, so as to
6333 * both boost throughput and not break bfqq's bandwidth and latency
6334 * guarantees. In this respect, the mechanism maintains a per-queue
6335 * inject limit, computed as below. While bfqq is empty, the injection
6336 * mechanism dispatches extra I/O requests only until the total number
6337 * of I/O requests in flight---i.e., already dispatched but not yet
6338 * completed---remains lower than this limit.
6340 * A first definition comes in handy to introduce the algorithm by
6341 * which the inject limit is computed. We define as first request for
6342 * bfqq, an I/O request for bfqq that arrives while bfqq is in
6343 * service, and causes bfqq to switch from empty to non-empty. The
6344 * algorithm updates the limit as a function of the effect of
6345 * injection on the service times of only the first requests of
6346 * bfqq. The reason for this restriction is that these are the
6347 * requests whose service time is affected most, because they are the
6348 * first to arrive after injection possibly occurred.
6350 * To evaluate the effect of injection, the algorithm measures the
6351 * "total service time" of first requests. We define as total service
6352 * time of an I/O request, the time that elapses since when the
6353 * request is enqueued into bfqq, to when it is completed. This
6354 * quantity allows the whole effect of injection to be measured. It is
6355 * easy to see why. Suppose that some requests of other queues are
6356 * actually injected while bfqq is empty, and that a new request R
6357 * then arrives for bfqq. If the device does start to serve all or
6358 * part of the injected requests during the service hole, then,
6359 * because of this extra service, it may delay the next invocation of
6360 * the dispatch hook of BFQ. Then, even after R gets eventually
6361 * dispatched, the device may delay the actual service of R if it is
6362 * still busy serving the extra requests, or if it decides to serve,
6363 * before R, some extra request still present in its queues. As a
6364 * conclusion, the cumulative extra delay caused by injection can be
6365 * easily evaluated by just comparing the total service time of first
6366 * requests with and without injection.
6368 * The limit-update algorithm works as follows. On the arrival of a
6369 * first request of bfqq, the algorithm measures the total time of the
6370 * request only if one of the three cases below holds, and, for each
6371 * case, it updates the limit as described below:
6373 * (1) If there is no in-flight request. This gives a baseline for the
6374 * total service time of the requests of bfqq. If the baseline has
6375 * not been computed yet, then, after computing it, the limit is
6376 * set to 1, to start boosting throughput, and to prepare the
6377 * ground for the next case. If the baseline has already been
6378 * computed, then it is updated, in case it results to be lower
6379 * than the previous value.
6381 * (2) If the limit is higher than 0 and there are in-flight
6382 * requests. By comparing the total service time in this case with
6383 * the above baseline, it is possible to know at which extent the
6384 * current value of the limit is inflating the total service
6385 * time. If the inflation is below a certain threshold, then bfqq
6386 * is assumed to be suffering from no perceivable loss of its
6387 * service guarantees, and the limit is even tentatively
6388 * increased. If the inflation is above the threshold, then the
6389 * limit is decreased. Due to the lack of any hysteresis, this
6390 * logic makes the limit oscillate even in steady workload
6391 * conditions. Yet we opted for it, because it is fast in reaching
6392 * the best value for the limit, as a function of the current I/O
6393 * workload. To reduce oscillations, this step is disabled for a
6394 * short time interval after the limit happens to be decreased.
6396 * (3) Periodically, after resetting the limit, to make sure that the
6397 * limit eventually drops in case the workload changes. This is
6398 * needed because, after the limit has gone safely up for a
6399 * certain workload, it is impossible to guess whether the
6400 * baseline total service time may have changed, without measuring
6401 * it again without injection. A more effective version of this
6402 * step might be to just sample the baseline, by interrupting
6403 * injection only once, and then to reset/lower the limit only if
6404 * the total service time with the current limit does happen to be
6407 * More details on each step are provided in the comments on the
6408 * pieces of code that implement these steps: the branch handling the
6409 * transition from empty to non empty in bfq_add_request(), the branch
6410 * handling injection in bfq_select_queue(), and the function
6411 * bfq_choose_bfqq_for_injection(). These comments also explain some
6412 * exceptions, made by the injection mechanism in some special cases.
6414 static void bfq_update_inject_limit(struct bfq_data *bfqd,
6415 struct bfq_queue *bfqq)
6417 u64 tot_time_ns = ktime_get_ns() - bfqd->last_empty_occupied_ns;
6418 unsigned int old_limit = bfqq->inject_limit;
6420 if (bfqq->last_serv_time_ns > 0 && bfqd->rqs_injected) {
6421 u64 threshold = (bfqq->last_serv_time_ns * 3)>>1;
6423 if (tot_time_ns >= threshold && old_limit > 0) {
6424 bfqq->inject_limit--;
6425 bfqq->decrease_time_jif = jiffies;
6426 } else if (tot_time_ns < threshold &&
6427 old_limit <= bfqd->max_rq_in_driver)
6428 bfqq->inject_limit++;
6432 * Either we still have to compute the base value for the
6433 * total service time, and there seem to be the right
6434 * conditions to do it, or we can lower the last base value
6437 * NOTE: (bfqd->rq_in_driver == 1) means that there is no I/O
6438 * request in flight, because this function is in the code
6439 * path that handles the completion of a request of bfqq, and,
6440 * in particular, this function is executed before
6441 * bfqd->rq_in_driver is decremented in such a code path.
6443 if ((bfqq->last_serv_time_ns == 0 && bfqd->rq_in_driver == 1) ||
6444 tot_time_ns < bfqq->last_serv_time_ns) {
6445 if (bfqq->last_serv_time_ns == 0) {
6447 * Now we certainly have a base value: make sure we
6448 * start trying injection.
6450 bfqq->inject_limit = max_t(unsigned int, 1, old_limit);
6452 bfqq->last_serv_time_ns = tot_time_ns;
6453 } else if (!bfqd->rqs_injected && bfqd->rq_in_driver == 1)
6455 * No I/O injected and no request still in service in
6456 * the drive: these are the exact conditions for
6457 * computing the base value of the total service time
6458 * for bfqq. So let's update this value, because it is
6459 * rather variable. For example, it varies if the size
6460 * or the spatial locality of the I/O requests in bfqq
6463 bfqq->last_serv_time_ns = tot_time_ns;
6466 /* update complete, not waiting for any request completion any longer */
6467 bfqd->waited_rq = NULL;
6468 bfqd->rqs_injected = false;
6472 * Handle either a requeue or a finish for rq. The things to do are
6473 * the same in both cases: all references to rq are to be dropped. In
6474 * particular, rq is considered completed from the point of view of
6477 static void bfq_finish_requeue_request(struct request *rq)
6479 struct bfq_queue *bfqq = RQ_BFQQ(rq);
6480 struct bfq_data *bfqd;
6481 unsigned long flags;
6484 * rq either is not associated with any icq, or is an already
6485 * requeued request that has not (yet) been re-inserted into
6488 if (!rq->elv.icq || !bfqq)
6493 if (rq->rq_flags & RQF_STARTED)
6494 bfqg_stats_update_completion(bfqq_group(bfqq),
6496 rq->io_start_time_ns,
6499 spin_lock_irqsave(&bfqd->lock, flags);
6500 if (likely(rq->rq_flags & RQF_STARTED)) {
6501 if (rq == bfqd->waited_rq)
6502 bfq_update_inject_limit(bfqd, bfqq);
6504 bfq_completed_request(bfqq, bfqd);
6506 bfqq_request_freed(bfqq);
6507 bfq_put_queue(bfqq);
6508 RQ_BIC(rq)->requests--;
6509 spin_unlock_irqrestore(&bfqd->lock, flags);
6512 * Reset private fields. In case of a requeue, this allows
6513 * this function to correctly do nothing if it is spuriously
6514 * invoked again on this same request (see the check at the
6515 * beginning of the function). Probably, a better general
6516 * design would be to prevent blk-mq from invoking the requeue
6517 * or finish hooks of an elevator, for a request that is not
6518 * referred by that elevator.
6520 * Resetting the following fields would break the
6521 * request-insertion logic if rq is re-inserted into a bfq
6522 * internal queue, without a re-preparation. Here we assume
6523 * that re-insertions of requeued requests, without
6524 * re-preparation, can happen only for pass_through or at_head
6525 * requests (which are not re-inserted into bfq internal
6528 rq->elv.priv[0] = NULL;
6529 rq->elv.priv[1] = NULL;
6532 static void bfq_finish_request(struct request *rq)
6534 bfq_finish_requeue_request(rq);
6537 put_io_context(rq->elv.icq->ioc);
6543 * Removes the association between the current task and bfqq, assuming
6544 * that bic points to the bfq iocontext of the task.
6545 * Returns NULL if a new bfqq should be allocated, or the old bfqq if this
6546 * was the last process referring to that bfqq.
6548 static struct bfq_queue *
6549 bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq)
6551 bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue");
6553 if (bfqq_process_refs(bfqq) == 1) {
6554 bfqq->pid = current->pid;
6555 bfq_clear_bfqq_coop(bfqq);
6556 bfq_clear_bfqq_split_coop(bfqq);
6560 bic_set_bfqq(bic, NULL, 1);
6562 bfq_put_cooperator(bfqq);
6564 bfq_release_process_ref(bfqq->bfqd, bfqq);
6568 static struct bfq_queue *bfq_get_bfqq_handle_split(struct bfq_data *bfqd,
6569 struct bfq_io_cq *bic,
6571 bool split, bool is_sync,
6574 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
6576 if (likely(bfqq && bfqq != &bfqd->oom_bfqq))
6583 bfq_put_queue(bfqq);
6584 bfqq = bfq_get_queue(bfqd, bio, is_sync, bic, split);
6586 bic_set_bfqq(bic, bfqq, is_sync);
6587 if (split && is_sync) {
6588 if ((bic->was_in_burst_list && bfqd->large_burst) ||
6589 bic->saved_in_large_burst)
6590 bfq_mark_bfqq_in_large_burst(bfqq);
6592 bfq_clear_bfqq_in_large_burst(bfqq);
6593 if (bic->was_in_burst_list)
6595 * If bfqq was in the current
6596 * burst list before being
6597 * merged, then we have to add
6598 * it back. And we do not need
6599 * to increase burst_size, as
6600 * we did not decrement
6601 * burst_size when we removed
6602 * bfqq from the burst list as
6603 * a consequence of a merge
6605 * bfq_put_queue). In this
6606 * respect, it would be rather
6607 * costly to know whether the
6608 * current burst list is still
6609 * the same burst list from
6610 * which bfqq was removed on
6611 * the merge. To avoid this
6612 * cost, if bfqq was in a
6613 * burst list, then we add
6614 * bfqq to the current burst
6615 * list without any further
6616 * check. This can cause
6617 * inappropriate insertions,
6618 * but rarely enough to not
6619 * harm the detection of large
6620 * bursts significantly.
6622 hlist_add_head(&bfqq->burst_list_node,
6625 bfqq->split_time = jiffies;
6632 * Only reset private fields. The actual request preparation will be
6633 * performed by bfq_init_rq, when rq is either inserted or merged. See
6634 * comments on bfq_init_rq for the reason behind this delayed
6637 static void bfq_prepare_request(struct request *rq)
6639 rq->elv.icq = ioc_find_get_icq(rq->q);
6642 * Regardless of whether we have an icq attached, we have to
6643 * clear the scheduler pointers, as they might point to
6644 * previously allocated bic/bfqq structs.
6646 rq->elv.priv[0] = rq->elv.priv[1] = NULL;
6650 * If needed, init rq, allocate bfq data structures associated with
6651 * rq, and increment reference counters in the destination bfq_queue
6652 * for rq. Return the destination bfq_queue for rq, or NULL is rq is
6653 * not associated with any bfq_queue.
6655 * This function is invoked by the functions that perform rq insertion
6656 * or merging. One may have expected the above preparation operations
6657 * to be performed in bfq_prepare_request, and not delayed to when rq
6658 * is inserted or merged. The rationale behind this delayed
6659 * preparation is that, after the prepare_request hook is invoked for
6660 * rq, rq may still be transformed into a request with no icq, i.e., a
6661 * request not associated with any queue. No bfq hook is invoked to
6662 * signal this transformation. As a consequence, should these
6663 * preparation operations be performed when the prepare_request hook
6664 * is invoked, and should rq be transformed one moment later, bfq
6665 * would end up in an inconsistent state, because it would have
6666 * incremented some queue counters for an rq destined to
6667 * transformation, without any chance to correctly lower these
6668 * counters back. In contrast, no transformation can still happen for
6669 * rq after rq has been inserted or merged. So, it is safe to execute
6670 * these preparation operations when rq is finally inserted or merged.
6672 static struct bfq_queue *bfq_init_rq(struct request *rq)
6674 struct request_queue *q = rq->q;
6675 struct bio *bio = rq->bio;
6676 struct bfq_data *bfqd = q->elevator->elevator_data;
6677 struct bfq_io_cq *bic;
6678 const int is_sync = rq_is_sync(rq);
6679 struct bfq_queue *bfqq;
6680 bool new_queue = false;
6681 bool bfqq_already_existing = false, split = false;
6683 if (unlikely(!rq->elv.icq))
6687 * Assuming that elv.priv[1] is set only if everything is set
6688 * for this rq. This holds true, because this function is
6689 * invoked only for insertion or merging, and, after such
6690 * events, a request cannot be manipulated any longer before
6691 * being removed from bfq.
6693 if (rq->elv.priv[1])
6694 return rq->elv.priv[1];
6696 bic = icq_to_bic(rq->elv.icq);
6698 bfq_check_ioprio_change(bic, bio);
6700 bfq_bic_update_cgroup(bic, bio);
6702 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio, false, is_sync,
6705 if (likely(!new_queue)) {
6706 /* If the queue was seeky for too long, break it apart. */
6707 if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq) &&
6708 !bic->stably_merged) {
6709 struct bfq_queue *old_bfqq = bfqq;
6711 /* Update bic before losing reference to bfqq */
6712 if (bfq_bfqq_in_large_burst(bfqq))
6713 bic->saved_in_large_burst = true;
6715 bfqq = bfq_split_bfqq(bic, bfqq);
6719 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio,
6722 if (unlikely(bfqq == &bfqd->oom_bfqq))
6723 bfqq_already_existing = true;
6725 bfqq_already_existing = true;
6727 if (!bfqq_already_existing) {
6728 bfqq->waker_bfqq = old_bfqq->waker_bfqq;
6729 bfqq->tentative_waker_bfqq = NULL;
6732 * If the waker queue disappears, then
6733 * new_bfqq->waker_bfqq must be
6734 * reset. So insert new_bfqq into the
6735 * woken_list of the waker. See
6736 * bfq_check_waker for details.
6738 if (bfqq->waker_bfqq)
6739 hlist_add_head(&bfqq->woken_list_node,
6740 &bfqq->waker_bfqq->woken_list);
6745 bfqq_request_allocated(bfqq);
6748 bfq_log_bfqq(bfqd, bfqq, "get_request %p: bfqq %p, %d",
6749 rq, bfqq, bfqq->ref);
6751 rq->elv.priv[0] = bic;
6752 rq->elv.priv[1] = bfqq;
6755 * If a bfq_queue has only one process reference, it is owned
6756 * by only this bic: we can then set bfqq->bic = bic. in
6757 * addition, if the queue has also just been split, we have to
6760 if (likely(bfqq != &bfqd->oom_bfqq) && bfqq_process_refs(bfqq) == 1) {
6764 * The queue has just been split from a shared
6765 * queue: restore the idle window and the
6766 * possible weight raising period.
6768 bfq_bfqq_resume_state(bfqq, bfqd, bic,
6769 bfqq_already_existing);
6774 * Consider bfqq as possibly belonging to a burst of newly
6775 * created queues only if:
6776 * 1) A burst is actually happening (bfqd->burst_size > 0)
6778 * 2) There is no other active queue. In fact, if, in
6779 * contrast, there are active queues not belonging to the
6780 * possible burst bfqq may belong to, then there is no gain
6781 * in considering bfqq as belonging to a burst, and
6782 * therefore in not weight-raising bfqq. See comments on
6783 * bfq_handle_burst().
6785 * This filtering also helps eliminating false positives,
6786 * occurring when bfqq does not belong to an actual large
6787 * burst, but some background task (e.g., a service) happens
6788 * to trigger the creation of new queues very close to when
6789 * bfqq and its possible companion queues are created. See
6790 * comments on bfq_handle_burst() for further details also on
6793 if (unlikely(bfq_bfqq_just_created(bfqq) &&
6794 (bfqd->burst_size > 0 ||
6795 bfq_tot_busy_queues(bfqd) == 0)))
6796 bfq_handle_burst(bfqd, bfqq);
6802 bfq_idle_slice_timer_body(struct bfq_data *bfqd, struct bfq_queue *bfqq)
6804 enum bfqq_expiration reason;
6805 unsigned long flags;
6807 spin_lock_irqsave(&bfqd->lock, flags);
6810 * Considering that bfqq may be in race, we should firstly check
6811 * whether bfqq is in service before doing something on it. If
6812 * the bfqq in race is not in service, it has already been expired
6813 * through __bfq_bfqq_expire func and its wait_request flags has
6814 * been cleared in __bfq_bfqd_reset_in_service func.
6816 if (bfqq != bfqd->in_service_queue) {
6817 spin_unlock_irqrestore(&bfqd->lock, flags);
6821 bfq_clear_bfqq_wait_request(bfqq);
6823 if (bfq_bfqq_budget_timeout(bfqq))
6825 * Also here the queue can be safely expired
6826 * for budget timeout without wasting
6829 reason = BFQQE_BUDGET_TIMEOUT;
6830 else if (bfqq->queued[0] == 0 && bfqq->queued[1] == 0)
6832 * The queue may not be empty upon timer expiration,
6833 * because we may not disable the timer when the
6834 * first request of the in-service queue arrives
6835 * during disk idling.
6837 reason = BFQQE_TOO_IDLE;
6839 goto schedule_dispatch;
6841 bfq_bfqq_expire(bfqd, bfqq, true, reason);
6844 bfq_schedule_dispatch(bfqd);
6845 spin_unlock_irqrestore(&bfqd->lock, flags);
6849 * Handler of the expiration of the timer running if the in-service queue
6850 * is idling inside its time slice.
6852 static enum hrtimer_restart bfq_idle_slice_timer(struct hrtimer *timer)
6854 struct bfq_data *bfqd = container_of(timer, struct bfq_data,
6856 struct bfq_queue *bfqq = bfqd->in_service_queue;
6859 * Theoretical race here: the in-service queue can be NULL or
6860 * different from the queue that was idling if a new request
6861 * arrives for the current queue and there is a full dispatch
6862 * cycle that changes the in-service queue. This can hardly
6863 * happen, but in the worst case we just expire a queue too
6867 bfq_idle_slice_timer_body(bfqd, bfqq);
6869 return HRTIMER_NORESTART;
6872 static void __bfq_put_async_bfqq(struct bfq_data *bfqd,
6873 struct bfq_queue **bfqq_ptr)
6875 struct bfq_queue *bfqq = *bfqq_ptr;
6877 bfq_log(bfqd, "put_async_bfqq: %p", bfqq);
6879 bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
6881 bfq_log_bfqq(bfqd, bfqq, "put_async_bfqq: putting %p, %d",
6883 bfq_put_queue(bfqq);
6889 * Release all the bfqg references to its async queues. If we are
6890 * deallocating the group these queues may still contain requests, so
6891 * we reparent them to the root cgroup (i.e., the only one that will
6892 * exist for sure until all the requests on a device are gone).
6894 void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg)
6898 for (i = 0; i < 2; i++)
6899 for (j = 0; j < IOPRIO_NR_LEVELS; j++)
6900 __bfq_put_async_bfqq(bfqd, &bfqg->async_bfqq[i][j]);
6902 __bfq_put_async_bfqq(bfqd, &bfqg->async_idle_bfqq);
6906 * See the comments on bfq_limit_depth for the purpose of
6907 * the depths set in the function. Return minimum shallow depth we'll use.
6909 static void bfq_update_depths(struct bfq_data *bfqd, struct sbitmap_queue *bt)
6911 unsigned int depth = 1U << bt->sb.shift;
6913 bfqd->full_depth_shift = bt->sb.shift;
6915 * In-word depths if no bfq_queue is being weight-raised:
6916 * leaving 25% of tags only for sync reads.
6918 * In next formulas, right-shift the value
6919 * (1U<<bt->sb.shift), instead of computing directly
6920 * (1U<<(bt->sb.shift - something)), to be robust against
6921 * any possible value of bt->sb.shift, without having to
6922 * limit 'something'.
6924 /* no more than 50% of tags for async I/O */
6925 bfqd->word_depths[0][0] = max(depth >> 1, 1U);
6927 * no more than 75% of tags for sync writes (25% extra tags
6928 * w.r.t. async I/O, to prevent async I/O from starving sync
6931 bfqd->word_depths[0][1] = max((depth * 3) >> 2, 1U);
6934 * In-word depths in case some bfq_queue is being weight-
6935 * raised: leaving ~63% of tags for sync reads. This is the
6936 * highest percentage for which, in our tests, application
6937 * start-up times didn't suffer from any regression due to tag
6940 /* no more than ~18% of tags for async I/O */
6941 bfqd->word_depths[1][0] = max((depth * 3) >> 4, 1U);
6942 /* no more than ~37% of tags for sync writes (~20% extra tags) */
6943 bfqd->word_depths[1][1] = max((depth * 6) >> 4, 1U);
6946 static void bfq_depth_updated(struct blk_mq_hw_ctx *hctx)
6948 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
6949 struct blk_mq_tags *tags = hctx->sched_tags;
6951 bfq_update_depths(bfqd, &tags->bitmap_tags);
6952 sbitmap_queue_min_shallow_depth(&tags->bitmap_tags, 1);
6955 static int bfq_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int index)
6957 bfq_depth_updated(hctx);
6961 static void bfq_exit_queue(struct elevator_queue *e)
6963 struct bfq_data *bfqd = e->elevator_data;
6964 struct bfq_queue *bfqq, *n;
6966 hrtimer_cancel(&bfqd->idle_slice_timer);
6968 spin_lock_irq(&bfqd->lock);
6969 list_for_each_entry_safe(bfqq, n, &bfqd->idle_list, bfqq_list)
6970 bfq_deactivate_bfqq(bfqd, bfqq, false, false);
6971 spin_unlock_irq(&bfqd->lock);
6973 hrtimer_cancel(&bfqd->idle_slice_timer);
6975 /* release oom-queue reference to root group */
6976 bfqg_and_blkg_put(bfqd->root_group);
6978 #ifdef CONFIG_BFQ_GROUP_IOSCHED
6979 blkcg_deactivate_policy(bfqd->queue, &blkcg_policy_bfq);
6981 spin_lock_irq(&bfqd->lock);
6982 bfq_put_async_queues(bfqd, bfqd->root_group);
6983 kfree(bfqd->root_group);
6984 spin_unlock_irq(&bfqd->lock);
6987 blk_stat_disable_accounting(bfqd->queue);
6988 clear_bit(ELEVATOR_FLAG_DISABLE_WBT, &e->flags);
6989 wbt_enable_default(bfqd->queue);
6994 static void bfq_init_root_group(struct bfq_group *root_group,
6995 struct bfq_data *bfqd)
6999 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7000 root_group->entity.parent = NULL;
7001 root_group->my_entity = NULL;
7002 root_group->bfqd = bfqd;
7004 root_group->rq_pos_tree = RB_ROOT;
7005 for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
7006 root_group->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
7007 root_group->sched_data.bfq_class_idle_last_service = jiffies;
7010 static int bfq_init_queue(struct request_queue *q, struct elevator_type *e)
7012 struct bfq_data *bfqd;
7013 struct elevator_queue *eq;
7015 eq = elevator_alloc(q, e);
7019 bfqd = kzalloc_node(sizeof(*bfqd), GFP_KERNEL, q->node);
7021 kobject_put(&eq->kobj);
7024 eq->elevator_data = bfqd;
7026 spin_lock_irq(&q->queue_lock);
7028 spin_unlock_irq(&q->queue_lock);
7031 * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues.
7032 * Grab a permanent reference to it, so that the normal code flow
7033 * will not attempt to free it.
7035 bfq_init_bfqq(bfqd, &bfqd->oom_bfqq, NULL, 1, 0);
7036 bfqd->oom_bfqq.ref++;
7037 bfqd->oom_bfqq.new_ioprio = BFQ_DEFAULT_QUEUE_IOPRIO;
7038 bfqd->oom_bfqq.new_ioprio_class = IOPRIO_CLASS_BE;
7039 bfqd->oom_bfqq.entity.new_weight =
7040 bfq_ioprio_to_weight(bfqd->oom_bfqq.new_ioprio);
7042 /* oom_bfqq does not participate to bursts */
7043 bfq_clear_bfqq_just_created(&bfqd->oom_bfqq);
7046 * Trigger weight initialization, according to ioprio, at the
7047 * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio
7048 * class won't be changed any more.
7050 bfqd->oom_bfqq.entity.prio_changed = 1;
7054 INIT_LIST_HEAD(&bfqd->dispatch);
7056 hrtimer_init(&bfqd->idle_slice_timer, CLOCK_MONOTONIC,
7058 bfqd->idle_slice_timer.function = bfq_idle_slice_timer;
7060 bfqd->queue_weights_tree = RB_ROOT_CACHED;
7061 bfqd->num_groups_with_pending_reqs = 0;
7063 INIT_LIST_HEAD(&bfqd->active_list);
7064 INIT_LIST_HEAD(&bfqd->idle_list);
7065 INIT_HLIST_HEAD(&bfqd->burst_list);
7068 bfqd->nonrot_with_queueing = blk_queue_nonrot(bfqd->queue);
7070 bfqd->bfq_max_budget = bfq_default_max_budget;
7072 bfqd->bfq_fifo_expire[0] = bfq_fifo_expire[0];
7073 bfqd->bfq_fifo_expire[1] = bfq_fifo_expire[1];
7074 bfqd->bfq_back_max = bfq_back_max;
7075 bfqd->bfq_back_penalty = bfq_back_penalty;
7076 bfqd->bfq_slice_idle = bfq_slice_idle;
7077 bfqd->bfq_timeout = bfq_timeout;
7079 bfqd->bfq_large_burst_thresh = 8;
7080 bfqd->bfq_burst_interval = msecs_to_jiffies(180);
7082 bfqd->low_latency = true;
7085 * Trade-off between responsiveness and fairness.
7087 bfqd->bfq_wr_coeff = 30;
7088 bfqd->bfq_wr_rt_max_time = msecs_to_jiffies(300);
7089 bfqd->bfq_wr_max_time = 0;
7090 bfqd->bfq_wr_min_idle_time = msecs_to_jiffies(2000);
7091 bfqd->bfq_wr_min_inter_arr_async = msecs_to_jiffies(500);
7092 bfqd->bfq_wr_max_softrt_rate = 7000; /*
7093 * Approximate rate required
7094 * to playback or record a
7095 * high-definition compressed
7098 bfqd->wr_busy_queues = 0;
7101 * Begin by assuming, optimistically, that the device peak
7102 * rate is equal to 2/3 of the highest reference rate.
7104 bfqd->rate_dur_prod = ref_rate[blk_queue_nonrot(bfqd->queue)] *
7105 ref_wr_duration[blk_queue_nonrot(bfqd->queue)];
7106 bfqd->peak_rate = ref_rate[blk_queue_nonrot(bfqd->queue)] * 2 / 3;
7108 spin_lock_init(&bfqd->lock);
7111 * The invocation of the next bfq_create_group_hierarchy
7112 * function is the head of a chain of function calls
7113 * (bfq_create_group_hierarchy->blkcg_activate_policy->
7114 * blk_mq_freeze_queue) that may lead to the invocation of the
7115 * has_work hook function. For this reason,
7116 * bfq_create_group_hierarchy is invoked only after all
7117 * scheduler data has been initialized, apart from the fields
7118 * that can be initialized only after invoking
7119 * bfq_create_group_hierarchy. This, in particular, enables
7120 * has_work to correctly return false. Of course, to avoid
7121 * other inconsistencies, the blk-mq stack must then refrain
7122 * from invoking further scheduler hooks before this init
7123 * function is finished.
7125 bfqd->root_group = bfq_create_group_hierarchy(bfqd, q->node);
7126 if (!bfqd->root_group)
7128 bfq_init_root_group(bfqd->root_group, bfqd);
7129 bfq_init_entity(&bfqd->oom_bfqq.entity, bfqd->root_group);
7131 /* We dispatch from request queue wide instead of hw queue */
7132 blk_queue_flag_set(QUEUE_FLAG_SQ_SCHED, q);
7134 set_bit(ELEVATOR_FLAG_DISABLE_WBT, &eq->flags);
7135 wbt_disable_default(q);
7136 blk_stat_enable_accounting(q);
7142 kobject_put(&eq->kobj);
7146 static void bfq_slab_kill(void)
7148 kmem_cache_destroy(bfq_pool);
7151 static int __init bfq_slab_setup(void)
7153 bfq_pool = KMEM_CACHE(bfq_queue, 0);
7159 static ssize_t bfq_var_show(unsigned int var, char *page)
7161 return sprintf(page, "%u\n", var);
7164 static int bfq_var_store(unsigned long *var, const char *page)
7166 unsigned long new_val;
7167 int ret = kstrtoul(page, 10, &new_val);
7175 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
7176 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
7178 struct bfq_data *bfqd = e->elevator_data; \
7179 u64 __data = __VAR; \
7181 __data = jiffies_to_msecs(__data); \
7182 else if (__CONV == 2) \
7183 __data = div_u64(__data, NSEC_PER_MSEC); \
7184 return bfq_var_show(__data, (page)); \
7186 SHOW_FUNCTION(bfq_fifo_expire_sync_show, bfqd->bfq_fifo_expire[1], 2);
7187 SHOW_FUNCTION(bfq_fifo_expire_async_show, bfqd->bfq_fifo_expire[0], 2);
7188 SHOW_FUNCTION(bfq_back_seek_max_show, bfqd->bfq_back_max, 0);
7189 SHOW_FUNCTION(bfq_back_seek_penalty_show, bfqd->bfq_back_penalty, 0);
7190 SHOW_FUNCTION(bfq_slice_idle_show, bfqd->bfq_slice_idle, 2);
7191 SHOW_FUNCTION(bfq_max_budget_show, bfqd->bfq_user_max_budget, 0);
7192 SHOW_FUNCTION(bfq_timeout_sync_show, bfqd->bfq_timeout, 1);
7193 SHOW_FUNCTION(bfq_strict_guarantees_show, bfqd->strict_guarantees, 0);
7194 SHOW_FUNCTION(bfq_low_latency_show, bfqd->low_latency, 0);
7195 #undef SHOW_FUNCTION
7197 #define USEC_SHOW_FUNCTION(__FUNC, __VAR) \
7198 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
7200 struct bfq_data *bfqd = e->elevator_data; \
7201 u64 __data = __VAR; \
7202 __data = div_u64(__data, NSEC_PER_USEC); \
7203 return bfq_var_show(__data, (page)); \
7205 USEC_SHOW_FUNCTION(bfq_slice_idle_us_show, bfqd->bfq_slice_idle);
7206 #undef USEC_SHOW_FUNCTION
7208 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
7210 __FUNC(struct elevator_queue *e, const char *page, size_t count) \
7212 struct bfq_data *bfqd = e->elevator_data; \
7213 unsigned long __data, __min = (MIN), __max = (MAX); \
7216 ret = bfq_var_store(&__data, (page)); \
7219 if (__data < __min) \
7221 else if (__data > __max) \
7224 *(__PTR) = msecs_to_jiffies(__data); \
7225 else if (__CONV == 2) \
7226 *(__PTR) = (u64)__data * NSEC_PER_MSEC; \
7228 *(__PTR) = __data; \
7231 STORE_FUNCTION(bfq_fifo_expire_sync_store, &bfqd->bfq_fifo_expire[1], 1,
7233 STORE_FUNCTION(bfq_fifo_expire_async_store, &bfqd->bfq_fifo_expire[0], 1,
7235 STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0);
7236 STORE_FUNCTION(bfq_back_seek_penalty_store, &bfqd->bfq_back_penalty, 1,
7238 STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 2);
7239 #undef STORE_FUNCTION
7241 #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
7242 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\
7244 struct bfq_data *bfqd = e->elevator_data; \
7245 unsigned long __data, __min = (MIN), __max = (MAX); \
7248 ret = bfq_var_store(&__data, (page)); \
7251 if (__data < __min) \
7253 else if (__data > __max) \
7255 *(__PTR) = (u64)__data * NSEC_PER_USEC; \
7258 USEC_STORE_FUNCTION(bfq_slice_idle_us_store, &bfqd->bfq_slice_idle, 0,
7260 #undef USEC_STORE_FUNCTION
7262 static ssize_t bfq_max_budget_store(struct elevator_queue *e,
7263 const char *page, size_t count)
7265 struct bfq_data *bfqd = e->elevator_data;
7266 unsigned long __data;
7269 ret = bfq_var_store(&__data, (page));
7274 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
7276 if (__data > INT_MAX)
7278 bfqd->bfq_max_budget = __data;
7281 bfqd->bfq_user_max_budget = __data;
7287 * Leaving this name to preserve name compatibility with cfq
7288 * parameters, but this timeout is used for both sync and async.
7290 static ssize_t bfq_timeout_sync_store(struct elevator_queue *e,
7291 const char *page, size_t count)
7293 struct bfq_data *bfqd = e->elevator_data;
7294 unsigned long __data;
7297 ret = bfq_var_store(&__data, (page));
7303 else if (__data > INT_MAX)
7306 bfqd->bfq_timeout = msecs_to_jiffies(__data);
7307 if (bfqd->bfq_user_max_budget == 0)
7308 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
7313 static ssize_t bfq_strict_guarantees_store(struct elevator_queue *e,
7314 const char *page, size_t count)
7316 struct bfq_data *bfqd = e->elevator_data;
7317 unsigned long __data;
7320 ret = bfq_var_store(&__data, (page));
7326 if (!bfqd->strict_guarantees && __data == 1
7327 && bfqd->bfq_slice_idle < 8 * NSEC_PER_MSEC)
7328 bfqd->bfq_slice_idle = 8 * NSEC_PER_MSEC;
7330 bfqd->strict_guarantees = __data;
7335 static ssize_t bfq_low_latency_store(struct elevator_queue *e,
7336 const char *page, size_t count)
7338 struct bfq_data *bfqd = e->elevator_data;
7339 unsigned long __data;
7342 ret = bfq_var_store(&__data, (page));
7348 if (__data == 0 && bfqd->low_latency != 0)
7350 bfqd->low_latency = __data;
7355 #define BFQ_ATTR(name) \
7356 __ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store)
7358 static struct elv_fs_entry bfq_attrs[] = {
7359 BFQ_ATTR(fifo_expire_sync),
7360 BFQ_ATTR(fifo_expire_async),
7361 BFQ_ATTR(back_seek_max),
7362 BFQ_ATTR(back_seek_penalty),
7363 BFQ_ATTR(slice_idle),
7364 BFQ_ATTR(slice_idle_us),
7365 BFQ_ATTR(max_budget),
7366 BFQ_ATTR(timeout_sync),
7367 BFQ_ATTR(strict_guarantees),
7368 BFQ_ATTR(low_latency),
7372 static struct elevator_type iosched_bfq_mq = {
7374 .limit_depth = bfq_limit_depth,
7375 .prepare_request = bfq_prepare_request,
7376 .requeue_request = bfq_finish_requeue_request,
7377 .finish_request = bfq_finish_request,
7378 .exit_icq = bfq_exit_icq,
7379 .insert_requests = bfq_insert_requests,
7380 .dispatch_request = bfq_dispatch_request,
7381 .next_request = elv_rb_latter_request,
7382 .former_request = elv_rb_former_request,
7383 .allow_merge = bfq_allow_bio_merge,
7384 .bio_merge = bfq_bio_merge,
7385 .request_merge = bfq_request_merge,
7386 .requests_merged = bfq_requests_merged,
7387 .request_merged = bfq_request_merged,
7388 .has_work = bfq_has_work,
7389 .depth_updated = bfq_depth_updated,
7390 .init_hctx = bfq_init_hctx,
7391 .init_sched = bfq_init_queue,
7392 .exit_sched = bfq_exit_queue,
7395 .icq_size = sizeof(struct bfq_io_cq),
7396 .icq_align = __alignof__(struct bfq_io_cq),
7397 .elevator_attrs = bfq_attrs,
7398 .elevator_name = "bfq",
7399 .elevator_owner = THIS_MODULE,
7401 MODULE_ALIAS("bfq-iosched");
7403 static int __init bfq_init(void)
7407 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7408 ret = blkcg_policy_register(&blkcg_policy_bfq);
7414 if (bfq_slab_setup())
7418 * Times to load large popular applications for the typical
7419 * systems installed on the reference devices (see the
7420 * comments before the definition of the next
7421 * array). Actually, we use slightly lower values, as the
7422 * estimated peak rate tends to be smaller than the actual
7423 * peak rate. The reason for this last fact is that estimates
7424 * are computed over much shorter time intervals than the long
7425 * intervals typically used for benchmarking. Why? First, to
7426 * adapt more quickly to variations. Second, because an I/O
7427 * scheduler cannot rely on a peak-rate-evaluation workload to
7428 * be run for a long time.
7430 ref_wr_duration[0] = msecs_to_jiffies(7000); /* actually 8 sec */
7431 ref_wr_duration[1] = msecs_to_jiffies(2500); /* actually 3 sec */
7433 ret = elv_register(&iosched_bfq_mq);
7442 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7443 blkcg_policy_unregister(&blkcg_policy_bfq);
7448 static void __exit bfq_exit(void)
7450 elv_unregister(&iosched_bfq_mq);
7451 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7452 blkcg_policy_unregister(&blkcg_policy_bfq);
7457 module_init(bfq_init);
7458 module_exit(bfq_exit);
7460 MODULE_AUTHOR("Paolo Valente");
7461 MODULE_LICENSE("GPL");
7462 MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler");