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 * Invoke __bfq_weights_tree_remove on bfqq and decrement the number
969 * of active groups for each queue's inactive parent entity.
971 void bfq_weights_tree_remove(struct bfq_queue *bfqq)
973 __bfq_weights_tree_remove(bfqq);
977 * Return expired entry, or NULL to just start from scratch in rbtree.
979 static struct request *bfq_check_fifo(struct bfq_queue *bfqq,
980 struct request *last)
984 if (bfq_bfqq_fifo_expire(bfqq))
987 bfq_mark_bfqq_fifo_expire(bfqq);
989 rq = rq_entry_fifo(bfqq->fifo.next);
991 if (rq == last || ktime_get_ns() < rq->fifo_time)
994 bfq_log_bfqq(bfqq->bfqd, bfqq, "check_fifo: returned %p", rq);
998 static struct request *bfq_find_next_rq(struct bfq_data *bfqd,
999 struct bfq_queue *bfqq,
1000 struct request *last)
1002 struct rb_node *rbnext = rb_next(&last->rb_node);
1003 struct rb_node *rbprev = rb_prev(&last->rb_node);
1004 struct request *next, *prev = NULL;
1006 /* Follow expired path, else get first next available. */
1007 next = bfq_check_fifo(bfqq, last);
1012 prev = rb_entry_rq(rbprev);
1015 next = rb_entry_rq(rbnext);
1017 rbnext = rb_first(&bfqq->sort_list);
1018 if (rbnext && rbnext != &last->rb_node)
1019 next = rb_entry_rq(rbnext);
1022 return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last));
1025 /* see the definition of bfq_async_charge_factor for details */
1026 static unsigned long bfq_serv_to_charge(struct request *rq,
1027 struct bfq_queue *bfqq)
1029 if (bfq_bfqq_sync(bfqq) || bfqq->wr_coeff > 1 ||
1030 bfq_asymmetric_scenario(bfqq->bfqd, bfqq))
1031 return blk_rq_sectors(rq);
1033 return blk_rq_sectors(rq) * bfq_async_charge_factor;
1037 * bfq_updated_next_req - update the queue after a new next_rq selection.
1038 * @bfqd: the device data the queue belongs to.
1039 * @bfqq: the queue to update.
1041 * If the first request of a queue changes we make sure that the queue
1042 * has enough budget to serve at least its first request (if the
1043 * request has grown). We do this because if the queue has not enough
1044 * budget for its first request, it has to go through two dispatch
1045 * rounds to actually get it dispatched.
1047 static void bfq_updated_next_req(struct bfq_data *bfqd,
1048 struct bfq_queue *bfqq)
1050 struct bfq_entity *entity = &bfqq->entity;
1051 struct request *next_rq = bfqq->next_rq;
1052 unsigned long new_budget;
1057 if (bfqq == bfqd->in_service_queue)
1059 * In order not to break guarantees, budgets cannot be
1060 * changed after an entity has been selected.
1064 new_budget = max_t(unsigned long,
1065 max_t(unsigned long, bfqq->max_budget,
1066 bfq_serv_to_charge(next_rq, bfqq)),
1068 if (entity->budget != new_budget) {
1069 entity->budget = new_budget;
1070 bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu",
1072 bfq_requeue_bfqq(bfqd, bfqq, false);
1076 static unsigned int bfq_wr_duration(struct bfq_data *bfqd)
1080 if (bfqd->bfq_wr_max_time > 0)
1081 return bfqd->bfq_wr_max_time;
1083 dur = bfqd->rate_dur_prod;
1084 do_div(dur, bfqd->peak_rate);
1087 * Limit duration between 3 and 25 seconds. The upper limit
1088 * has been conservatively set after the following worst case:
1089 * on a QEMU/KVM virtual machine
1090 * - running in a slow PC
1091 * - with a virtual disk stacked on a slow low-end 5400rpm HDD
1092 * - serving a heavy I/O workload, such as the sequential reading
1094 * mplayer took 23 seconds to start, if constantly weight-raised.
1096 * As for higher values than that accommodating the above bad
1097 * scenario, tests show that higher values would often yield
1098 * the opposite of the desired result, i.e., would worsen
1099 * responsiveness by allowing non-interactive applications to
1100 * preserve weight raising for too long.
1102 * On the other end, lower values than 3 seconds make it
1103 * difficult for most interactive tasks to complete their jobs
1104 * before weight-raising finishes.
1106 return clamp_val(dur, msecs_to_jiffies(3000), msecs_to_jiffies(25000));
1109 /* switch back from soft real-time to interactive weight raising */
1110 static void switch_back_to_interactive_wr(struct bfq_queue *bfqq,
1111 struct bfq_data *bfqd)
1113 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1114 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1115 bfqq->last_wr_start_finish = bfqq->wr_start_at_switch_to_srt;
1119 bfq_bfqq_resume_state(struct bfq_queue *bfqq, struct bfq_data *bfqd,
1120 struct bfq_io_cq *bic, bool bfq_already_existing)
1122 unsigned int old_wr_coeff = 1;
1123 bool busy = bfq_already_existing && bfq_bfqq_busy(bfqq);
1125 if (bic->saved_has_short_ttime)
1126 bfq_mark_bfqq_has_short_ttime(bfqq);
1128 bfq_clear_bfqq_has_short_ttime(bfqq);
1130 if (bic->saved_IO_bound)
1131 bfq_mark_bfqq_IO_bound(bfqq);
1133 bfq_clear_bfqq_IO_bound(bfqq);
1135 bfqq->last_serv_time_ns = bic->saved_last_serv_time_ns;
1136 bfqq->inject_limit = bic->saved_inject_limit;
1137 bfqq->decrease_time_jif = bic->saved_decrease_time_jif;
1139 bfqq->entity.new_weight = bic->saved_weight;
1140 bfqq->ttime = bic->saved_ttime;
1141 bfqq->io_start_time = bic->saved_io_start_time;
1142 bfqq->tot_idle_time = bic->saved_tot_idle_time;
1144 * Restore weight coefficient only if low_latency is on
1146 if (bfqd->low_latency) {
1147 old_wr_coeff = bfqq->wr_coeff;
1148 bfqq->wr_coeff = bic->saved_wr_coeff;
1150 bfqq->service_from_wr = bic->saved_service_from_wr;
1151 bfqq->wr_start_at_switch_to_srt = bic->saved_wr_start_at_switch_to_srt;
1152 bfqq->last_wr_start_finish = bic->saved_last_wr_start_finish;
1153 bfqq->wr_cur_max_time = bic->saved_wr_cur_max_time;
1155 if (bfqq->wr_coeff > 1 && (bfq_bfqq_in_large_burst(bfqq) ||
1156 time_is_before_jiffies(bfqq->last_wr_start_finish +
1157 bfqq->wr_cur_max_time))) {
1158 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
1159 !bfq_bfqq_in_large_burst(bfqq) &&
1160 time_is_after_eq_jiffies(bfqq->wr_start_at_switch_to_srt +
1161 bfq_wr_duration(bfqd))) {
1162 switch_back_to_interactive_wr(bfqq, bfqd);
1165 bfq_log_bfqq(bfqq->bfqd, bfqq,
1166 "resume state: switching off wr");
1170 /* make sure weight will be updated, however we got here */
1171 bfqq->entity.prio_changed = 1;
1176 if (old_wr_coeff == 1 && bfqq->wr_coeff > 1)
1177 bfqd->wr_busy_queues++;
1178 else if (old_wr_coeff > 1 && bfqq->wr_coeff == 1)
1179 bfqd->wr_busy_queues--;
1182 static int bfqq_process_refs(struct bfq_queue *bfqq)
1184 return bfqq->ref - bfqq->entity.allocated -
1185 bfqq->entity.on_st_or_in_serv -
1186 (bfqq->weight_counter != NULL) - bfqq->stable_ref;
1189 /* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */
1190 static void bfq_reset_burst_list(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1192 struct bfq_queue *item;
1193 struct hlist_node *n;
1195 hlist_for_each_entry_safe(item, n, &bfqd->burst_list, burst_list_node)
1196 hlist_del_init(&item->burst_list_node);
1199 * Start the creation of a new burst list only if there is no
1200 * active queue. See comments on the conditional invocation of
1201 * bfq_handle_burst().
1203 if (bfq_tot_busy_queues(bfqd) == 0) {
1204 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1205 bfqd->burst_size = 1;
1207 bfqd->burst_size = 0;
1209 bfqd->burst_parent_entity = bfqq->entity.parent;
1212 /* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */
1213 static void bfq_add_to_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1215 /* Increment burst size to take into account also bfqq */
1218 if (bfqd->burst_size == bfqd->bfq_large_burst_thresh) {
1219 struct bfq_queue *pos, *bfqq_item;
1220 struct hlist_node *n;
1223 * Enough queues have been activated shortly after each
1224 * other to consider this burst as large.
1226 bfqd->large_burst = true;
1229 * We can now mark all queues in the burst list as
1230 * belonging to a large burst.
1232 hlist_for_each_entry(bfqq_item, &bfqd->burst_list,
1234 bfq_mark_bfqq_in_large_burst(bfqq_item);
1235 bfq_mark_bfqq_in_large_burst(bfqq);
1238 * From now on, and until the current burst finishes, any
1239 * new queue being activated shortly after the last queue
1240 * was inserted in the burst can be immediately marked as
1241 * belonging to a large burst. So the burst list is not
1242 * needed any more. Remove it.
1244 hlist_for_each_entry_safe(pos, n, &bfqd->burst_list,
1246 hlist_del_init(&pos->burst_list_node);
1248 * Burst not yet large: add bfqq to the burst list. Do
1249 * not increment the ref counter for bfqq, because bfqq
1250 * is removed from the burst list before freeing bfqq
1253 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1257 * If many queues belonging to the same group happen to be created
1258 * shortly after each other, then the processes associated with these
1259 * queues have typically a common goal. In particular, bursts of queue
1260 * creations are usually caused by services or applications that spawn
1261 * many parallel threads/processes. Examples are systemd during boot,
1262 * or git grep. To help these processes get their job done as soon as
1263 * possible, it is usually better to not grant either weight-raising
1264 * or device idling to their queues, unless these queues must be
1265 * protected from the I/O flowing through other active queues.
1267 * In this comment we describe, firstly, the reasons why this fact
1268 * holds, and, secondly, the next function, which implements the main
1269 * steps needed to properly mark these queues so that they can then be
1270 * treated in a different way.
1272 * The above services or applications benefit mostly from a high
1273 * throughput: the quicker the requests of the activated queues are
1274 * cumulatively served, the sooner the target job of these queues gets
1275 * completed. As a consequence, weight-raising any of these queues,
1276 * which also implies idling the device for it, is almost always
1277 * counterproductive, unless there are other active queues to isolate
1278 * these new queues from. If there no other active queues, then
1279 * weight-raising these new queues just lowers throughput in most
1282 * On the other hand, a burst of queue creations may be caused also by
1283 * the start of an application that does not consist of a lot of
1284 * parallel I/O-bound threads. In fact, with a complex application,
1285 * several short processes may need to be executed to start-up the
1286 * application. In this respect, to start an application as quickly as
1287 * possible, the best thing to do is in any case to privilege the I/O
1288 * related to the application with respect to all other
1289 * I/O. Therefore, the best strategy to start as quickly as possible
1290 * an application that causes a burst of queue creations is to
1291 * weight-raise all the queues created during the burst. This is the
1292 * exact opposite of the best strategy for the other type of bursts.
1294 * In the end, to take the best action for each of the two cases, the
1295 * two types of bursts need to be distinguished. Fortunately, this
1296 * seems relatively easy, by looking at the sizes of the bursts. In
1297 * particular, we found a threshold such that only bursts with a
1298 * larger size than that threshold are apparently caused by
1299 * services or commands such as systemd or git grep. For brevity,
1300 * hereafter we call just 'large' these bursts. BFQ *does not*
1301 * weight-raise queues whose creation occurs in a large burst. In
1302 * addition, for each of these queues BFQ performs or does not perform
1303 * idling depending on which choice boosts the throughput more. The
1304 * exact choice depends on the device and request pattern at
1307 * Unfortunately, false positives may occur while an interactive task
1308 * is starting (e.g., an application is being started). The
1309 * consequence is that the queues associated with the task do not
1310 * enjoy weight raising as expected. Fortunately these false positives
1311 * are very rare. They typically occur if some service happens to
1312 * start doing I/O exactly when the interactive task starts.
1314 * Turning back to the next function, it is invoked only if there are
1315 * no active queues (apart from active queues that would belong to the
1316 * same, possible burst bfqq would belong to), and it implements all
1317 * the steps needed to detect the occurrence of a large burst and to
1318 * properly mark all the queues belonging to it (so that they can then
1319 * be treated in a different way). This goal is achieved by
1320 * maintaining a "burst list" that holds, temporarily, the queues that
1321 * belong to the burst in progress. The list is then used to mark
1322 * these queues as belonging to a large burst if the burst does become
1323 * large. The main steps are the following.
1325 * . when the very first queue is created, the queue is inserted into the
1326 * list (as it could be the first queue in a possible burst)
1328 * . if the current burst has not yet become large, and a queue Q that does
1329 * not yet belong to the burst is activated shortly after the last time
1330 * at which a new queue entered the burst list, then the function appends
1331 * Q to the burst list
1333 * . if, as a consequence of the previous step, the burst size reaches
1334 * the large-burst threshold, then
1336 * . all the queues in the burst list are marked as belonging to a
1339 * . the burst list is deleted; in fact, the burst list already served
1340 * its purpose (keeping temporarily track of the queues in a burst,
1341 * so as to be able to mark them as belonging to a large burst in the
1342 * previous sub-step), and now is not needed any more
1344 * . the device enters a large-burst mode
1346 * . if a queue Q that does not belong to the burst is created while
1347 * the device is in large-burst mode and shortly after the last time
1348 * at which a queue either entered the burst list or was marked as
1349 * belonging to the current large burst, then Q is immediately marked
1350 * as belonging to a large burst.
1352 * . if a queue Q that does not belong to the burst is created a while
1353 * later, i.e., not shortly after, than the last time at which a queue
1354 * either entered the burst list or was marked as belonging to the
1355 * current large burst, then the current burst is deemed as finished and:
1357 * . the large-burst mode is reset if set
1359 * . the burst list is emptied
1361 * . Q is inserted in the burst list, as Q may be the first queue
1362 * in a possible new burst (then the burst list contains just Q
1365 static void bfq_handle_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1368 * If bfqq is already in the burst list or is part of a large
1369 * burst, or finally has just been split, then there is
1370 * nothing else to do.
1372 if (!hlist_unhashed(&bfqq->burst_list_node) ||
1373 bfq_bfqq_in_large_burst(bfqq) ||
1374 time_is_after_eq_jiffies(bfqq->split_time +
1375 msecs_to_jiffies(10)))
1379 * If bfqq's creation happens late enough, or bfqq belongs to
1380 * a different group than the burst group, then the current
1381 * burst is finished, and related data structures must be
1384 * In this respect, consider the special case where bfqq is
1385 * the very first queue created after BFQ is selected for this
1386 * device. In this case, last_ins_in_burst and
1387 * burst_parent_entity are not yet significant when we get
1388 * here. But it is easy to verify that, whether or not the
1389 * following condition is true, bfqq will end up being
1390 * inserted into the burst list. In particular the list will
1391 * happen to contain only bfqq. And this is exactly what has
1392 * to happen, as bfqq may be the first queue of the first
1395 if (time_is_before_jiffies(bfqd->last_ins_in_burst +
1396 bfqd->bfq_burst_interval) ||
1397 bfqq->entity.parent != bfqd->burst_parent_entity) {
1398 bfqd->large_burst = false;
1399 bfq_reset_burst_list(bfqd, bfqq);
1404 * If we get here, then bfqq is being activated shortly after the
1405 * last queue. So, if the current burst is also large, we can mark
1406 * bfqq as belonging to this large burst immediately.
1408 if (bfqd->large_burst) {
1409 bfq_mark_bfqq_in_large_burst(bfqq);
1414 * If we get here, then a large-burst state has not yet been
1415 * reached, but bfqq is being activated shortly after the last
1416 * queue. Then we add bfqq to the burst.
1418 bfq_add_to_burst(bfqd, bfqq);
1421 * At this point, bfqq either has been added to the current
1422 * burst or has caused the current burst to terminate and a
1423 * possible new burst to start. In particular, in the second
1424 * case, bfqq has become the first queue in the possible new
1425 * burst. In both cases last_ins_in_burst needs to be moved
1428 bfqd->last_ins_in_burst = jiffies;
1431 static int bfq_bfqq_budget_left(struct bfq_queue *bfqq)
1433 struct bfq_entity *entity = &bfqq->entity;
1435 return entity->budget - entity->service;
1439 * If enough samples have been computed, return the current max budget
1440 * stored in bfqd, which is dynamically updated according to the
1441 * estimated disk peak rate; otherwise return the default max budget
1443 static int bfq_max_budget(struct bfq_data *bfqd)
1445 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1446 return bfq_default_max_budget;
1448 return bfqd->bfq_max_budget;
1452 * Return min budget, which is a fraction of the current or default
1453 * max budget (trying with 1/32)
1455 static int bfq_min_budget(struct bfq_data *bfqd)
1457 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1458 return bfq_default_max_budget / 32;
1460 return bfqd->bfq_max_budget / 32;
1464 * The next function, invoked after the input queue bfqq switches from
1465 * idle to busy, updates the budget of bfqq. The function also tells
1466 * whether the in-service queue should be expired, by returning
1467 * true. The purpose of expiring the in-service queue is to give bfqq
1468 * the chance to possibly preempt the in-service queue, and the reason
1469 * for preempting the in-service queue is to achieve one of the two
1472 * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has
1473 * expired because it has remained idle. In particular, bfqq may have
1474 * expired for one of the following two reasons:
1476 * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling
1477 * and did not make it to issue a new request before its last
1478 * request was served;
1480 * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue
1481 * a new request before the expiration of the idling-time.
1483 * Even if bfqq has expired for one of the above reasons, the process
1484 * associated with the queue may be however issuing requests greedily,
1485 * and thus be sensitive to the bandwidth it receives (bfqq may have
1486 * remained idle for other reasons: CPU high load, bfqq not enjoying
1487 * idling, I/O throttling somewhere in the path from the process to
1488 * the I/O scheduler, ...). But if, after every expiration for one of
1489 * the above two reasons, bfqq has to wait for the service of at least
1490 * one full budget of another queue before being served again, then
1491 * bfqq is likely to get a much lower bandwidth or resource time than
1492 * its reserved ones. To address this issue, two countermeasures need
1495 * First, the budget and the timestamps of bfqq need to be updated in
1496 * a special way on bfqq reactivation: they need to be updated as if
1497 * bfqq did not remain idle and did not expire. In fact, if they are
1498 * computed as if bfqq expired and remained idle until reactivation,
1499 * then the process associated with bfqq is treated as if, instead of
1500 * being greedy, it stopped issuing requests when bfqq remained idle,
1501 * and restarts issuing requests only on this reactivation. In other
1502 * words, the scheduler does not help the process recover the "service
1503 * hole" between bfqq expiration and reactivation. As a consequence,
1504 * the process receives a lower bandwidth than its reserved one. In
1505 * contrast, to recover this hole, the budget must be updated as if
1506 * bfqq was not expired at all before this reactivation, i.e., it must
1507 * be set to the value of the remaining budget when bfqq was
1508 * expired. Along the same line, timestamps need to be assigned the
1509 * value they had the last time bfqq was selected for service, i.e.,
1510 * before last expiration. Thus timestamps need to be back-shifted
1511 * with respect to their normal computation (see [1] for more details
1512 * on this tricky aspect).
1514 * Secondly, to allow the process to recover the hole, the in-service
1515 * queue must be expired too, to give bfqq the chance to preempt it
1516 * immediately. In fact, if bfqq has to wait for a full budget of the
1517 * in-service queue to be completed, then it may become impossible to
1518 * let the process recover the hole, even if the back-shifted
1519 * timestamps of bfqq are lower than those of the in-service queue. If
1520 * this happens for most or all of the holes, then the process may not
1521 * receive its reserved bandwidth. In this respect, it is worth noting
1522 * that, being the service of outstanding requests unpreemptible, a
1523 * little fraction of the holes may however be unrecoverable, thereby
1524 * causing a little loss of bandwidth.
1526 * The last important point is detecting whether bfqq does need this
1527 * bandwidth recovery. In this respect, the next function deems the
1528 * process associated with bfqq greedy, and thus allows it to recover
1529 * the hole, if: 1) the process is waiting for the arrival of a new
1530 * request (which implies that bfqq expired for one of the above two
1531 * reasons), and 2) such a request has arrived soon. The first
1532 * condition is controlled through the flag non_blocking_wait_rq,
1533 * while the second through the flag arrived_in_time. If both
1534 * conditions hold, then the function computes the budget in the
1535 * above-described special way, and signals that the in-service queue
1536 * should be expired. Timestamp back-shifting is done later in
1537 * __bfq_activate_entity.
1539 * 2. Reduce latency. Even if timestamps are not backshifted to let
1540 * the process associated with bfqq recover a service hole, bfqq may
1541 * however happen to have, after being (re)activated, a lower finish
1542 * timestamp than the in-service queue. That is, the next budget of
1543 * bfqq may have to be completed before the one of the in-service
1544 * queue. If this is the case, then preempting the in-service queue
1545 * allows this goal to be achieved, apart from the unpreemptible,
1546 * outstanding requests mentioned above.
1548 * Unfortunately, regardless of which of the above two goals one wants
1549 * to achieve, service trees need first to be updated to know whether
1550 * the in-service queue must be preempted. To have service trees
1551 * correctly updated, the in-service queue must be expired and
1552 * rescheduled, and bfqq must be scheduled too. This is one of the
1553 * most costly operations (in future versions, the scheduling
1554 * mechanism may be re-designed in such a way to make it possible to
1555 * know whether preemption is needed without needing to update service
1556 * trees). In addition, queue preemptions almost always cause random
1557 * I/O, which may in turn cause loss of throughput. Finally, there may
1558 * even be no in-service queue when the next function is invoked (so,
1559 * no queue to compare timestamps with). Because of these facts, the
1560 * next function adopts the following simple scheme to avoid costly
1561 * operations, too frequent preemptions and too many dependencies on
1562 * the state of the scheduler: it requests the expiration of the
1563 * in-service queue (unconditionally) only for queues that need to
1564 * recover a hole. Then it delegates to other parts of the code the
1565 * responsibility of handling the above case 2.
1567 static bool bfq_bfqq_update_budg_for_activation(struct bfq_data *bfqd,
1568 struct bfq_queue *bfqq,
1569 bool arrived_in_time)
1571 struct bfq_entity *entity = &bfqq->entity;
1574 * In the next compound condition, we check also whether there
1575 * is some budget left, because otherwise there is no point in
1576 * trying to go on serving bfqq with this same budget: bfqq
1577 * would be expired immediately after being selected for
1578 * service. This would only cause useless overhead.
1580 if (bfq_bfqq_non_blocking_wait_rq(bfqq) && arrived_in_time &&
1581 bfq_bfqq_budget_left(bfqq) > 0) {
1583 * We do not clear the flag non_blocking_wait_rq here, as
1584 * the latter is used in bfq_activate_bfqq to signal
1585 * that timestamps need to be back-shifted (and is
1586 * cleared right after).
1590 * In next assignment we rely on that either
1591 * entity->service or entity->budget are not updated
1592 * on expiration if bfqq is empty (see
1593 * __bfq_bfqq_recalc_budget). Thus both quantities
1594 * remain unchanged after such an expiration, and the
1595 * following statement therefore assigns to
1596 * entity->budget the remaining budget on such an
1599 entity->budget = min_t(unsigned long,
1600 bfq_bfqq_budget_left(bfqq),
1604 * At this point, we have used entity->service to get
1605 * the budget left (needed for updating
1606 * entity->budget). Thus we finally can, and have to,
1607 * reset entity->service. The latter must be reset
1608 * because bfqq would otherwise be charged again for
1609 * the service it has received during its previous
1612 entity->service = 0;
1618 * We can finally complete expiration, by setting service to 0.
1620 entity->service = 0;
1621 entity->budget = max_t(unsigned long, bfqq->max_budget,
1622 bfq_serv_to_charge(bfqq->next_rq, bfqq));
1623 bfq_clear_bfqq_non_blocking_wait_rq(bfqq);
1628 * Return the farthest past time instant according to jiffies
1631 static unsigned long bfq_smallest_from_now(void)
1633 return jiffies - MAX_JIFFY_OFFSET;
1636 static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data *bfqd,
1637 struct bfq_queue *bfqq,
1638 unsigned int old_wr_coeff,
1639 bool wr_or_deserves_wr,
1644 if (old_wr_coeff == 1 && wr_or_deserves_wr) {
1645 /* start a weight-raising period */
1647 bfqq->service_from_wr = 0;
1648 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1649 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1652 * No interactive weight raising in progress
1653 * here: assign minus infinity to
1654 * wr_start_at_switch_to_srt, to make sure
1655 * that, at the end of the soft-real-time
1656 * weight raising periods that is starting
1657 * now, no interactive weight-raising period
1658 * may be wrongly considered as still in
1659 * progress (and thus actually started by
1662 bfqq->wr_start_at_switch_to_srt =
1663 bfq_smallest_from_now();
1664 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1665 BFQ_SOFTRT_WEIGHT_FACTOR;
1666 bfqq->wr_cur_max_time =
1667 bfqd->bfq_wr_rt_max_time;
1671 * If needed, further reduce budget to make sure it is
1672 * close to bfqq's backlog, so as to reduce the
1673 * scheduling-error component due to a too large
1674 * budget. Do not care about throughput consequences,
1675 * but only about latency. Finally, do not assign a
1676 * too small budget either, to avoid increasing
1677 * latency by causing too frequent expirations.
1679 bfqq->entity.budget = min_t(unsigned long,
1680 bfqq->entity.budget,
1681 2 * bfq_min_budget(bfqd));
1682 } else if (old_wr_coeff > 1) {
1683 if (interactive) { /* update wr coeff and duration */
1684 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1685 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1686 } else if (in_burst)
1690 * The application is now or still meeting the
1691 * requirements for being deemed soft rt. We
1692 * can then correctly and safely (re)charge
1693 * the weight-raising duration for the
1694 * application with the weight-raising
1695 * duration for soft rt applications.
1697 * In particular, doing this recharge now, i.e.,
1698 * before the weight-raising period for the
1699 * application finishes, reduces the probability
1700 * of the following negative scenario:
1701 * 1) the weight of a soft rt application is
1702 * raised at startup (as for any newly
1703 * created application),
1704 * 2) since the application is not interactive,
1705 * at a certain time weight-raising is
1706 * stopped for the application,
1707 * 3) at that time the application happens to
1708 * still have pending requests, and hence
1709 * is destined to not have a chance to be
1710 * deemed soft rt before these requests are
1711 * completed (see the comments to the
1712 * function bfq_bfqq_softrt_next_start()
1713 * for details on soft rt detection),
1714 * 4) these pending requests experience a high
1715 * latency because the application is not
1716 * weight-raised while they are pending.
1718 if (bfqq->wr_cur_max_time !=
1719 bfqd->bfq_wr_rt_max_time) {
1720 bfqq->wr_start_at_switch_to_srt =
1721 bfqq->last_wr_start_finish;
1723 bfqq->wr_cur_max_time =
1724 bfqd->bfq_wr_rt_max_time;
1725 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1726 BFQ_SOFTRT_WEIGHT_FACTOR;
1728 bfqq->last_wr_start_finish = jiffies;
1733 static bool bfq_bfqq_idle_for_long_time(struct bfq_data *bfqd,
1734 struct bfq_queue *bfqq)
1736 return bfqq->dispatched == 0 &&
1737 time_is_before_jiffies(
1738 bfqq->budget_timeout +
1739 bfqd->bfq_wr_min_idle_time);
1744 * Return true if bfqq is in a higher priority class, or has a higher
1745 * weight than the in-service queue.
1747 static bool bfq_bfqq_higher_class_or_weight(struct bfq_queue *bfqq,
1748 struct bfq_queue *in_serv_bfqq)
1750 int bfqq_weight, in_serv_weight;
1752 if (bfqq->ioprio_class < in_serv_bfqq->ioprio_class)
1755 if (in_serv_bfqq->entity.parent == bfqq->entity.parent) {
1756 bfqq_weight = bfqq->entity.weight;
1757 in_serv_weight = in_serv_bfqq->entity.weight;
1759 if (bfqq->entity.parent)
1760 bfqq_weight = bfqq->entity.parent->weight;
1762 bfqq_weight = bfqq->entity.weight;
1763 if (in_serv_bfqq->entity.parent)
1764 in_serv_weight = in_serv_bfqq->entity.parent->weight;
1766 in_serv_weight = in_serv_bfqq->entity.weight;
1769 return bfqq_weight > in_serv_weight;
1772 static bool bfq_better_to_idle(struct bfq_queue *bfqq);
1774 static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data *bfqd,
1775 struct bfq_queue *bfqq,
1780 bool soft_rt, in_burst, wr_or_deserves_wr,
1781 bfqq_wants_to_preempt,
1782 idle_for_long_time = bfq_bfqq_idle_for_long_time(bfqd, bfqq),
1784 * See the comments on
1785 * bfq_bfqq_update_budg_for_activation for
1786 * details on the usage of the next variable.
1788 arrived_in_time = ktime_get_ns() <=
1789 bfqq->ttime.last_end_request +
1790 bfqd->bfq_slice_idle * 3;
1794 * bfqq deserves to be weight-raised if:
1796 * - it does not belong to a large burst,
1797 * - it has been idle for enough time or is soft real-time,
1798 * - is linked to a bfq_io_cq (it is not shared in any sense),
1799 * - has a default weight (otherwise we assume the user wanted
1800 * to control its weight explicitly)
1802 in_burst = bfq_bfqq_in_large_burst(bfqq);
1803 soft_rt = bfqd->bfq_wr_max_softrt_rate > 0 &&
1804 !BFQQ_TOTALLY_SEEKY(bfqq) &&
1806 time_is_before_jiffies(bfqq->soft_rt_next_start) &&
1807 bfqq->dispatched == 0 &&
1808 bfqq->entity.new_weight == 40;
1809 *interactive = !in_burst && idle_for_long_time &&
1810 bfqq->entity.new_weight == 40;
1812 * Merged bfq_queues are kept out of weight-raising
1813 * (low-latency) mechanisms. The reason is that these queues
1814 * are usually created for non-interactive and
1815 * non-soft-real-time tasks. Yet this is not the case for
1816 * stably-merged queues. These queues are merged just because
1817 * they are created shortly after each other. So they may
1818 * easily serve the I/O of an interactive or soft-real time
1819 * application, if the application happens to spawn multiple
1820 * processes. So let also stably-merged queued enjoy weight
1823 wr_or_deserves_wr = bfqd->low_latency &&
1824 (bfqq->wr_coeff > 1 ||
1825 (bfq_bfqq_sync(bfqq) &&
1826 (bfqq->bic || RQ_BIC(rq)->stably_merged) &&
1827 (*interactive || soft_rt)));
1830 * Using the last flag, update budget and check whether bfqq
1831 * may want to preempt the in-service queue.
1833 bfqq_wants_to_preempt =
1834 bfq_bfqq_update_budg_for_activation(bfqd, bfqq,
1838 * If bfqq happened to be activated in a burst, but has been
1839 * idle for much more than an interactive queue, then we
1840 * assume that, in the overall I/O initiated in the burst, the
1841 * I/O associated with bfqq is finished. So bfqq does not need
1842 * to be treated as a queue belonging to a burst
1843 * anymore. Accordingly, we reset bfqq's in_large_burst flag
1844 * if set, and remove bfqq from the burst list if it's
1845 * there. We do not decrement burst_size, because the fact
1846 * that bfqq does not need to belong to the burst list any
1847 * more does not invalidate the fact that bfqq was created in
1850 if (likely(!bfq_bfqq_just_created(bfqq)) &&
1851 idle_for_long_time &&
1852 time_is_before_jiffies(
1853 bfqq->budget_timeout +
1854 msecs_to_jiffies(10000))) {
1855 hlist_del_init(&bfqq->burst_list_node);
1856 bfq_clear_bfqq_in_large_burst(bfqq);
1859 bfq_clear_bfqq_just_created(bfqq);
1861 if (bfqd->low_latency) {
1862 if (unlikely(time_is_after_jiffies(bfqq->split_time)))
1865 jiffies - bfqd->bfq_wr_min_idle_time - 1;
1867 if (time_is_before_jiffies(bfqq->split_time +
1868 bfqd->bfq_wr_min_idle_time)) {
1869 bfq_update_bfqq_wr_on_rq_arrival(bfqd, bfqq,
1876 if (old_wr_coeff != bfqq->wr_coeff)
1877 bfqq->entity.prio_changed = 1;
1881 bfqq->last_idle_bklogged = jiffies;
1882 bfqq->service_from_backlogged = 0;
1883 bfq_clear_bfqq_softrt_update(bfqq);
1885 bfq_add_bfqq_busy(bfqq);
1888 * Expire in-service queue if preemption may be needed for
1889 * guarantees or throughput. As for guarantees, we care
1890 * explicitly about two cases. The first is that bfqq has to
1891 * recover a service hole, as explained in the comments on
1892 * bfq_bfqq_update_budg_for_activation(), i.e., that
1893 * bfqq_wants_to_preempt is true. However, if bfqq does not
1894 * carry time-critical I/O, then bfqq's bandwidth is less
1895 * important than that of queues that carry time-critical I/O.
1896 * So, as a further constraint, we consider this case only if
1897 * bfqq is at least as weight-raised, i.e., at least as time
1898 * critical, as the in-service queue.
1900 * The second case is that bfqq is in a higher priority class,
1901 * or has a higher weight than the in-service queue. If this
1902 * condition does not hold, we don't care because, even if
1903 * bfqq does not start to be served immediately, the resulting
1904 * delay for bfqq's I/O is however lower or much lower than
1905 * the ideal completion time to be guaranteed to bfqq's I/O.
1907 * In both cases, preemption is needed only if, according to
1908 * the timestamps of both bfqq and of the in-service queue,
1909 * bfqq actually is the next queue to serve. So, to reduce
1910 * useless preemptions, the return value of
1911 * next_queue_may_preempt() is considered in the next compound
1912 * condition too. Yet next_queue_may_preempt() just checks a
1913 * simple, necessary condition for bfqq to be the next queue
1914 * to serve. In fact, to evaluate a sufficient condition, the
1915 * timestamps of the in-service queue would need to be
1916 * updated, and this operation is quite costly (see the
1917 * comments on bfq_bfqq_update_budg_for_activation()).
1919 * As for throughput, we ask bfq_better_to_idle() whether we
1920 * still need to plug I/O dispatching. If bfq_better_to_idle()
1921 * says no, then plugging is not needed any longer, either to
1922 * boost throughput or to perserve service guarantees. Then
1923 * the best option is to stop plugging I/O, as not doing so
1924 * would certainly lower throughput. We may end up in this
1925 * case if: (1) upon a dispatch attempt, we detected that it
1926 * was better to plug I/O dispatch, and to wait for a new
1927 * request to arrive for the currently in-service queue, but
1928 * (2) this switch of bfqq to busy changes the scenario.
1930 if (bfqd->in_service_queue &&
1931 ((bfqq_wants_to_preempt &&
1932 bfqq->wr_coeff >= bfqd->in_service_queue->wr_coeff) ||
1933 bfq_bfqq_higher_class_or_weight(bfqq, bfqd->in_service_queue) ||
1934 !bfq_better_to_idle(bfqd->in_service_queue)) &&
1935 next_queue_may_preempt(bfqd))
1936 bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
1937 false, BFQQE_PREEMPTED);
1940 static void bfq_reset_inject_limit(struct bfq_data *bfqd,
1941 struct bfq_queue *bfqq)
1943 /* invalidate baseline total service time */
1944 bfqq->last_serv_time_ns = 0;
1947 * Reset pointer in case we are waiting for
1948 * some request completion.
1950 bfqd->waited_rq = NULL;
1953 * If bfqq has a short think time, then start by setting the
1954 * inject limit to 0 prudentially, because the service time of
1955 * an injected I/O request may be higher than the think time
1956 * of bfqq, and therefore, if one request was injected when
1957 * bfqq remains empty, this injected request might delay the
1958 * service of the next I/O request for bfqq significantly. In
1959 * case bfqq can actually tolerate some injection, then the
1960 * adaptive update will however raise the limit soon. This
1961 * lucky circumstance holds exactly because bfqq has a short
1962 * think time, and thus, after remaining empty, is likely to
1963 * get new I/O enqueued---and then completed---before being
1964 * expired. This is the very pattern that gives the
1965 * limit-update algorithm the chance to measure the effect of
1966 * injection on request service times, and then to update the
1967 * limit accordingly.
1969 * However, in the following special case, the inject limit is
1970 * left to 1 even if the think time is short: bfqq's I/O is
1971 * synchronized with that of some other queue, i.e., bfqq may
1972 * receive new I/O only after the I/O of the other queue is
1973 * completed. Keeping the inject limit to 1 allows the
1974 * blocking I/O to be served while bfqq is in service. And
1975 * this is very convenient both for bfqq and for overall
1976 * throughput, as explained in detail in the comments in
1977 * bfq_update_has_short_ttime().
1979 * On the opposite end, if bfqq has a long think time, then
1980 * start directly by 1, because:
1981 * a) on the bright side, keeping at most one request in
1982 * service in the drive is unlikely to cause any harm to the
1983 * latency of bfqq's requests, as the service time of a single
1984 * request is likely to be lower than the think time of bfqq;
1985 * b) on the downside, after becoming empty, bfqq is likely to
1986 * expire before getting its next request. With this request
1987 * arrival pattern, it is very hard to sample total service
1988 * times and update the inject limit accordingly (see comments
1989 * on bfq_update_inject_limit()). So the limit is likely to be
1990 * never, or at least seldom, updated. As a consequence, by
1991 * setting the limit to 1, we avoid that no injection ever
1992 * occurs with bfqq. On the downside, this proactive step
1993 * further reduces chances to actually compute the baseline
1994 * total service time. Thus it reduces chances to execute the
1995 * limit-update algorithm and possibly raise the limit to more
1998 if (bfq_bfqq_has_short_ttime(bfqq))
1999 bfqq->inject_limit = 0;
2001 bfqq->inject_limit = 1;
2003 bfqq->decrease_time_jif = jiffies;
2006 static void bfq_update_io_intensity(struct bfq_queue *bfqq, u64 now_ns)
2008 u64 tot_io_time = now_ns - bfqq->io_start_time;
2010 if (RB_EMPTY_ROOT(&bfqq->sort_list) && bfqq->dispatched == 0)
2011 bfqq->tot_idle_time +=
2012 now_ns - bfqq->ttime.last_end_request;
2014 if (unlikely(bfq_bfqq_just_created(bfqq)))
2018 * Must be busy for at least about 80% of the time to be
2019 * considered I/O bound.
2021 if (bfqq->tot_idle_time * 5 > tot_io_time)
2022 bfq_clear_bfqq_IO_bound(bfqq);
2024 bfq_mark_bfqq_IO_bound(bfqq);
2027 * Keep an observation window of at most 200 ms in the past
2030 if (tot_io_time > 200 * NSEC_PER_MSEC) {
2031 bfqq->io_start_time = now_ns - (tot_io_time>>1);
2032 bfqq->tot_idle_time >>= 1;
2037 * Detect whether bfqq's I/O seems synchronized with that of some
2038 * other queue, i.e., whether bfqq, after remaining empty, happens to
2039 * receive new I/O only right after some I/O request of the other
2040 * queue has been completed. We call waker queue the other queue, and
2041 * we assume, for simplicity, that bfqq may have at most one waker
2044 * A remarkable throughput boost can be reached by unconditionally
2045 * injecting the I/O of the waker queue, every time a new
2046 * bfq_dispatch_request happens to be invoked while I/O is being
2047 * plugged for bfqq. In addition to boosting throughput, this
2048 * unblocks bfqq's I/O, thereby improving bandwidth and latency for
2049 * bfqq. Note that these same results may be achieved with the general
2050 * injection mechanism, but less effectively. For details on this
2051 * aspect, see the comments on the choice of the queue for injection
2052 * in bfq_select_queue().
2054 * Turning back to the detection of a waker queue, a queue Q is deemed as a
2055 * waker queue for bfqq if, for three consecutive times, bfqq happens to become
2056 * non empty right after a request of Q has been completed within given
2057 * timeout. In this respect, even if bfqq is empty, we do not check for a waker
2058 * if it still has some in-flight I/O. In fact, in this case bfqq is actually
2059 * still being served by the drive, and may receive new I/O on the completion
2060 * of some of the in-flight requests. In particular, on the first time, Q is
2061 * tentatively set as a candidate waker queue, while on the third consecutive
2062 * time that Q is detected, the field waker_bfqq is set to Q, to confirm that Q
2063 * is a waker queue for bfqq. These detection steps are performed only if bfqq
2064 * has a long think time, so as to make it more likely that bfqq's I/O is
2065 * actually being blocked by a synchronization. This last filter, plus the
2066 * above three-times requirement and time limit for detection, make false
2067 * positives less likely.
2071 * The sooner a waker queue is detected, the sooner throughput can be
2072 * boosted by injecting I/O from the waker queue. Fortunately,
2073 * detection is likely to be actually fast, for the following
2074 * reasons. While blocked by synchronization, bfqq has a long think
2075 * time. This implies that bfqq's inject limit is at least equal to 1
2076 * (see the comments in bfq_update_inject_limit()). So, thanks to
2077 * injection, the waker queue is likely to be served during the very
2078 * first I/O-plugging time interval for bfqq. This triggers the first
2079 * step of the detection mechanism. Thanks again to injection, the
2080 * candidate waker queue is then likely to be confirmed no later than
2081 * during the next I/O-plugging interval for bfqq.
2085 * On queue merging all waker information is lost.
2087 static void bfq_check_waker(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2090 char waker_name[MAX_BFQQ_NAME_LENGTH];
2092 if (!bfqd->last_completed_rq_bfqq ||
2093 bfqd->last_completed_rq_bfqq == bfqq ||
2094 bfq_bfqq_has_short_ttime(bfqq) ||
2095 now_ns - bfqd->last_completion >= 4 * NSEC_PER_MSEC)
2099 * We reset waker detection logic also if too much time has passed
2100 * since the first detection. If wakeups are rare, pointless idling
2101 * doesn't hurt throughput that much. The condition below makes sure
2102 * we do not uselessly idle blocking waker in more than 1/64 cases.
2104 if (bfqd->last_completed_rq_bfqq !=
2105 bfqq->tentative_waker_bfqq ||
2106 now_ns > bfqq->waker_detection_started +
2107 128 * (u64)bfqd->bfq_slice_idle) {
2109 * First synchronization detected with a
2110 * candidate waker queue, or with a different
2111 * candidate waker queue from the current one.
2113 bfqq->tentative_waker_bfqq =
2114 bfqd->last_completed_rq_bfqq;
2115 bfqq->num_waker_detections = 1;
2116 bfqq->waker_detection_started = now_ns;
2117 bfq_bfqq_name(bfqq->tentative_waker_bfqq, waker_name,
2118 MAX_BFQQ_NAME_LENGTH);
2119 bfq_log_bfqq(bfqd, bfqq, "set tentative waker %s", waker_name);
2120 } else /* Same tentative waker queue detected again */
2121 bfqq->num_waker_detections++;
2123 if (bfqq->num_waker_detections == 3) {
2124 bfqq->waker_bfqq = bfqd->last_completed_rq_bfqq;
2125 bfqq->tentative_waker_bfqq = NULL;
2126 bfq_bfqq_name(bfqq->waker_bfqq, waker_name,
2127 MAX_BFQQ_NAME_LENGTH);
2128 bfq_log_bfqq(bfqd, bfqq, "set waker %s", waker_name);
2131 * If the waker queue disappears, then
2132 * bfqq->waker_bfqq must be reset. To
2133 * this goal, we maintain in each
2134 * waker queue a list, woken_list, of
2135 * all the queues that reference the
2136 * waker queue through their
2137 * waker_bfqq pointer. When the waker
2138 * queue exits, the waker_bfqq pointer
2139 * of all the queues in the woken_list
2142 * In addition, if bfqq is already in
2143 * the woken_list of a waker queue,
2144 * then, before being inserted into
2145 * the woken_list of a new waker
2146 * queue, bfqq must be removed from
2147 * the woken_list of the old waker
2150 if (!hlist_unhashed(&bfqq->woken_list_node))
2151 hlist_del_init(&bfqq->woken_list_node);
2152 hlist_add_head(&bfqq->woken_list_node,
2153 &bfqd->last_completed_rq_bfqq->woken_list);
2157 static void bfq_add_request(struct request *rq)
2159 struct bfq_queue *bfqq = RQ_BFQQ(rq);
2160 struct bfq_data *bfqd = bfqq->bfqd;
2161 struct request *next_rq, *prev;
2162 unsigned int old_wr_coeff = bfqq->wr_coeff;
2163 bool interactive = false;
2164 u64 now_ns = ktime_get_ns();
2166 bfq_log_bfqq(bfqd, bfqq, "add_request %d", rq_is_sync(rq));
2167 bfqq->queued[rq_is_sync(rq)]++;
2169 * Updating of 'bfqd->queued' is protected by 'bfqd->lock', however, it
2170 * may be read without holding the lock in bfq_has_work().
2172 WRITE_ONCE(bfqd->queued, bfqd->queued + 1);
2174 if (bfq_bfqq_sync(bfqq) && RQ_BIC(rq)->requests <= 1) {
2175 bfq_check_waker(bfqd, bfqq, now_ns);
2178 * Periodically reset inject limit, to make sure that
2179 * the latter eventually drops in case workload
2180 * changes, see step (3) in the comments on
2181 * bfq_update_inject_limit().
2183 if (time_is_before_eq_jiffies(bfqq->decrease_time_jif +
2184 msecs_to_jiffies(1000)))
2185 bfq_reset_inject_limit(bfqd, bfqq);
2188 * The following conditions must hold to setup a new
2189 * sampling of total service time, and then a new
2190 * update of the inject limit:
2191 * - bfqq is in service, because the total service
2192 * time is evaluated only for the I/O requests of
2193 * the queues in service;
2194 * - this is the right occasion to compute or to
2195 * lower the baseline total service time, because
2196 * there are actually no requests in the drive,
2198 * the baseline total service time is available, and
2199 * this is the right occasion to compute the other
2200 * quantity needed to update the inject limit, i.e.,
2201 * the total service time caused by the amount of
2202 * injection allowed by the current value of the
2203 * limit. It is the right occasion because injection
2204 * has actually been performed during the service
2205 * hole, and there are still in-flight requests,
2206 * which are very likely to be exactly the injected
2207 * requests, or part of them;
2208 * - the minimum interval for sampling the total
2209 * service time and updating the inject limit has
2212 if (bfqq == bfqd->in_service_queue &&
2213 (bfqd->rq_in_driver == 0 ||
2214 (bfqq->last_serv_time_ns > 0 &&
2215 bfqd->rqs_injected && bfqd->rq_in_driver > 0)) &&
2216 time_is_before_eq_jiffies(bfqq->decrease_time_jif +
2217 msecs_to_jiffies(10))) {
2218 bfqd->last_empty_occupied_ns = ktime_get_ns();
2220 * Start the state machine for measuring the
2221 * total service time of rq: setting
2222 * wait_dispatch will cause bfqd->waited_rq to
2223 * be set when rq will be dispatched.
2225 bfqd->wait_dispatch = true;
2227 * If there is no I/O in service in the drive,
2228 * then possible injection occurred before the
2229 * arrival of rq will not affect the total
2230 * service time of rq. So the injection limit
2231 * must not be updated as a function of such
2232 * total service time, unless new injection
2233 * occurs before rq is completed. To have the
2234 * injection limit updated only in the latter
2235 * case, reset rqs_injected here (rqs_injected
2236 * will be set in case injection is performed
2237 * on bfqq before rq is completed).
2239 if (bfqd->rq_in_driver == 0)
2240 bfqd->rqs_injected = false;
2244 if (bfq_bfqq_sync(bfqq))
2245 bfq_update_io_intensity(bfqq, now_ns);
2247 elv_rb_add(&bfqq->sort_list, rq);
2250 * Check if this request is a better next-serve candidate.
2252 prev = bfqq->next_rq;
2253 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position);
2254 bfqq->next_rq = next_rq;
2257 * Adjust priority tree position, if next_rq changes.
2258 * See comments on bfq_pos_tree_add_move() for the unlikely().
2260 if (unlikely(!bfqd->nonrot_with_queueing && prev != bfqq->next_rq))
2261 bfq_pos_tree_add_move(bfqd, bfqq);
2263 if (!bfq_bfqq_busy(bfqq)) /* switching to busy ... */
2264 bfq_bfqq_handle_idle_busy_switch(bfqd, bfqq, old_wr_coeff,
2267 if (bfqd->low_latency && old_wr_coeff == 1 && !rq_is_sync(rq) &&
2268 time_is_before_jiffies(
2269 bfqq->last_wr_start_finish +
2270 bfqd->bfq_wr_min_inter_arr_async)) {
2271 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
2272 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
2274 bfqd->wr_busy_queues++;
2275 bfqq->entity.prio_changed = 1;
2277 if (prev != bfqq->next_rq)
2278 bfq_updated_next_req(bfqd, bfqq);
2282 * Assign jiffies to last_wr_start_finish in the following
2285 * . if bfqq is not going to be weight-raised, because, for
2286 * non weight-raised queues, last_wr_start_finish stores the
2287 * arrival time of the last request; as of now, this piece
2288 * of information is used only for deciding whether to
2289 * weight-raise async queues
2291 * . if bfqq is not weight-raised, because, if bfqq is now
2292 * switching to weight-raised, then last_wr_start_finish
2293 * stores the time when weight-raising starts
2295 * . if bfqq is interactive, because, regardless of whether
2296 * bfqq is currently weight-raised, the weight-raising
2297 * period must start or restart (this case is considered
2298 * separately because it is not detected by the above
2299 * conditions, if bfqq is already weight-raised)
2301 * last_wr_start_finish has to be updated also if bfqq is soft
2302 * real-time, because the weight-raising period is constantly
2303 * restarted on idle-to-busy transitions for these queues, but
2304 * this is already done in bfq_bfqq_handle_idle_busy_switch if
2307 if (bfqd->low_latency &&
2308 (old_wr_coeff == 1 || bfqq->wr_coeff == 1 || interactive))
2309 bfqq->last_wr_start_finish = jiffies;
2312 static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd,
2314 struct request_queue *q)
2316 struct bfq_queue *bfqq = bfqd->bio_bfqq;
2320 return elv_rb_find(&bfqq->sort_list, bio_end_sector(bio));
2325 static sector_t get_sdist(sector_t last_pos, struct request *rq)
2328 return abs(blk_rq_pos(rq) - last_pos);
2333 #if 0 /* Still not clear if we can do without next two functions */
2334 static void bfq_activate_request(struct request_queue *q, struct request *rq)
2336 struct bfq_data *bfqd = q->elevator->elevator_data;
2338 bfqd->rq_in_driver++;
2341 static void bfq_deactivate_request(struct request_queue *q, struct request *rq)
2343 struct bfq_data *bfqd = q->elevator->elevator_data;
2345 bfqd->rq_in_driver--;
2349 static void bfq_remove_request(struct request_queue *q,
2352 struct bfq_queue *bfqq = RQ_BFQQ(rq);
2353 struct bfq_data *bfqd = bfqq->bfqd;
2354 const int sync = rq_is_sync(rq);
2356 if (bfqq->next_rq == rq) {
2357 bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq);
2358 bfq_updated_next_req(bfqd, bfqq);
2361 if (rq->queuelist.prev != &rq->queuelist)
2362 list_del_init(&rq->queuelist);
2363 bfqq->queued[sync]--;
2365 * Updating of 'bfqd->queued' is protected by 'bfqd->lock', however, it
2366 * may be read without holding the lock in bfq_has_work().
2368 WRITE_ONCE(bfqd->queued, bfqd->queued - 1);
2369 elv_rb_del(&bfqq->sort_list, rq);
2371 elv_rqhash_del(q, rq);
2372 if (q->last_merge == rq)
2373 q->last_merge = NULL;
2375 if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
2376 bfqq->next_rq = NULL;
2378 if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->in_service_queue) {
2379 bfq_del_bfqq_busy(bfqq, false);
2381 * bfqq emptied. In normal operation, when
2382 * bfqq is empty, bfqq->entity.service and
2383 * bfqq->entity.budget must contain,
2384 * respectively, the service received and the
2385 * budget used last time bfqq emptied. These
2386 * facts do not hold in this case, as at least
2387 * this last removal occurred while bfqq is
2388 * not in service. To avoid inconsistencies,
2389 * reset both bfqq->entity.service and
2390 * bfqq->entity.budget, if bfqq has still a
2391 * process that may issue I/O requests to it.
2393 bfqq->entity.budget = bfqq->entity.service = 0;
2397 * Remove queue from request-position tree as it is empty.
2399 if (bfqq->pos_root) {
2400 rb_erase(&bfqq->pos_node, bfqq->pos_root);
2401 bfqq->pos_root = NULL;
2404 /* see comments on bfq_pos_tree_add_move() for the unlikely() */
2405 if (unlikely(!bfqd->nonrot_with_queueing))
2406 bfq_pos_tree_add_move(bfqd, bfqq);
2409 if (rq->cmd_flags & REQ_META)
2410 bfqq->meta_pending--;
2414 static bool bfq_bio_merge(struct request_queue *q, struct bio *bio,
2415 unsigned int nr_segs)
2417 struct bfq_data *bfqd = q->elevator->elevator_data;
2418 struct request *free = NULL;
2420 * bfq_bic_lookup grabs the queue_lock: invoke it now and
2421 * store its return value for later use, to avoid nesting
2422 * queue_lock inside the bfqd->lock. We assume that the bic
2423 * returned by bfq_bic_lookup does not go away before
2424 * bfqd->lock is taken.
2426 struct bfq_io_cq *bic = bfq_bic_lookup(q);
2429 spin_lock_irq(&bfqd->lock);
2433 * Make sure cgroup info is uptodate for current process before
2434 * considering the merge.
2436 bfq_bic_update_cgroup(bic, bio);
2438 bfqd->bio_bfqq = bic_to_bfqq(bic, op_is_sync(bio->bi_opf));
2440 bfqd->bio_bfqq = NULL;
2442 bfqd->bio_bic = bic;
2444 ret = blk_mq_sched_try_merge(q, bio, nr_segs, &free);
2446 spin_unlock_irq(&bfqd->lock);
2448 blk_mq_free_request(free);
2453 static int bfq_request_merge(struct request_queue *q, struct request **req,
2456 struct bfq_data *bfqd = q->elevator->elevator_data;
2457 struct request *__rq;
2459 __rq = bfq_find_rq_fmerge(bfqd, bio, q);
2460 if (__rq && elv_bio_merge_ok(__rq, bio)) {
2463 if (blk_discard_mergable(__rq))
2464 return ELEVATOR_DISCARD_MERGE;
2465 return ELEVATOR_FRONT_MERGE;
2468 return ELEVATOR_NO_MERGE;
2471 static void bfq_request_merged(struct request_queue *q, struct request *req,
2472 enum elv_merge type)
2474 if (type == ELEVATOR_FRONT_MERGE &&
2475 rb_prev(&req->rb_node) &&
2477 blk_rq_pos(container_of(rb_prev(&req->rb_node),
2478 struct request, rb_node))) {
2479 struct bfq_queue *bfqq = RQ_BFQQ(req);
2480 struct bfq_data *bfqd;
2481 struct request *prev, *next_rq;
2488 /* Reposition request in its sort_list */
2489 elv_rb_del(&bfqq->sort_list, req);
2490 elv_rb_add(&bfqq->sort_list, req);
2492 /* Choose next request to be served for bfqq */
2493 prev = bfqq->next_rq;
2494 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, req,
2495 bfqd->last_position);
2496 bfqq->next_rq = next_rq;
2498 * If next_rq changes, update both the queue's budget to
2499 * fit the new request and the queue's position in its
2502 if (prev != bfqq->next_rq) {
2503 bfq_updated_next_req(bfqd, bfqq);
2505 * See comments on bfq_pos_tree_add_move() for
2508 if (unlikely(!bfqd->nonrot_with_queueing))
2509 bfq_pos_tree_add_move(bfqd, bfqq);
2515 * This function is called to notify the scheduler that the requests
2516 * rq and 'next' have been merged, with 'next' going away. BFQ
2517 * exploits this hook to address the following issue: if 'next' has a
2518 * fifo_time lower that rq, then the fifo_time of rq must be set to
2519 * the value of 'next', to not forget the greater age of 'next'.
2521 * NOTE: in this function we assume that rq is in a bfq_queue, basing
2522 * on that rq is picked from the hash table q->elevator->hash, which,
2523 * in its turn, is filled only with I/O requests present in
2524 * bfq_queues, while BFQ is in use for the request queue q. In fact,
2525 * the function that fills this hash table (elv_rqhash_add) is called
2526 * only by bfq_insert_request.
2528 static void bfq_requests_merged(struct request_queue *q, struct request *rq,
2529 struct request *next)
2531 struct bfq_queue *bfqq = RQ_BFQQ(rq),
2532 *next_bfqq = RQ_BFQQ(next);
2538 * If next and rq belong to the same bfq_queue and next is older
2539 * than rq, then reposition rq in the fifo (by substituting next
2540 * with rq). Otherwise, if next and rq belong to different
2541 * bfq_queues, never reposition rq: in fact, we would have to
2542 * reposition it with respect to next's position in its own fifo,
2543 * which would most certainly be too expensive with respect to
2546 if (bfqq == next_bfqq &&
2547 !list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2548 next->fifo_time < rq->fifo_time) {
2549 list_del_init(&rq->queuelist);
2550 list_replace_init(&next->queuelist, &rq->queuelist);
2551 rq->fifo_time = next->fifo_time;
2554 if (bfqq->next_rq == next)
2557 bfqg_stats_update_io_merged(bfqq_group(bfqq), next->cmd_flags);
2559 /* Merged request may be in the IO scheduler. Remove it. */
2560 if (!RB_EMPTY_NODE(&next->rb_node)) {
2561 bfq_remove_request(next->q, next);
2563 bfqg_stats_update_io_remove(bfqq_group(next_bfqq),
2568 /* Must be called with bfqq != NULL */
2569 static void bfq_bfqq_end_wr(struct bfq_queue *bfqq)
2572 * If bfqq has been enjoying interactive weight-raising, then
2573 * reset soft_rt_next_start. We do it for the following
2574 * reason. bfqq may have been conveying the I/O needed to load
2575 * a soft real-time application. Such an application actually
2576 * exhibits a soft real-time I/O pattern after it finishes
2577 * loading, and finally starts doing its job. But, if bfqq has
2578 * been receiving a lot of bandwidth so far (likely to happen
2579 * on a fast device), then soft_rt_next_start now contains a
2580 * high value that. So, without this reset, bfqq would be
2581 * prevented from being possibly considered as soft_rt for a
2585 if (bfqq->wr_cur_max_time !=
2586 bfqq->bfqd->bfq_wr_rt_max_time)
2587 bfqq->soft_rt_next_start = jiffies;
2589 if (bfq_bfqq_busy(bfqq))
2590 bfqq->bfqd->wr_busy_queues--;
2592 bfqq->wr_cur_max_time = 0;
2593 bfqq->last_wr_start_finish = jiffies;
2595 * Trigger a weight change on the next invocation of
2596 * __bfq_entity_update_weight_prio.
2598 bfqq->entity.prio_changed = 1;
2601 void bfq_end_wr_async_queues(struct bfq_data *bfqd,
2602 struct bfq_group *bfqg)
2606 for (i = 0; i < 2; i++)
2607 for (j = 0; j < IOPRIO_NR_LEVELS; j++)
2608 if (bfqg->async_bfqq[i][j])
2609 bfq_bfqq_end_wr(bfqg->async_bfqq[i][j]);
2610 if (bfqg->async_idle_bfqq)
2611 bfq_bfqq_end_wr(bfqg->async_idle_bfqq);
2614 static void bfq_end_wr(struct bfq_data *bfqd)
2616 struct bfq_queue *bfqq;
2618 spin_lock_irq(&bfqd->lock);
2620 list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
2621 bfq_bfqq_end_wr(bfqq);
2622 list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list)
2623 bfq_bfqq_end_wr(bfqq);
2624 bfq_end_wr_async(bfqd);
2626 spin_unlock_irq(&bfqd->lock);
2629 static sector_t bfq_io_struct_pos(void *io_struct, bool request)
2632 return blk_rq_pos(io_struct);
2634 return ((struct bio *)io_struct)->bi_iter.bi_sector;
2637 static int bfq_rq_close_to_sector(void *io_struct, bool request,
2640 return abs(bfq_io_struct_pos(io_struct, request) - sector) <=
2644 static struct bfq_queue *bfqq_find_close(struct bfq_data *bfqd,
2645 struct bfq_queue *bfqq,
2648 struct rb_root *root = &bfqq_group(bfqq)->rq_pos_tree;
2649 struct rb_node *parent, *node;
2650 struct bfq_queue *__bfqq;
2652 if (RB_EMPTY_ROOT(root))
2656 * First, if we find a request starting at the end of the last
2657 * request, choose it.
2659 __bfqq = bfq_rq_pos_tree_lookup(bfqd, root, sector, &parent, NULL);
2664 * If the exact sector wasn't found, the parent of the NULL leaf
2665 * will contain the closest sector (rq_pos_tree sorted by
2666 * next_request position).
2668 __bfqq = rb_entry(parent, struct bfq_queue, pos_node);
2669 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2672 if (blk_rq_pos(__bfqq->next_rq) < sector)
2673 node = rb_next(&__bfqq->pos_node);
2675 node = rb_prev(&__bfqq->pos_node);
2679 __bfqq = rb_entry(node, struct bfq_queue, pos_node);
2680 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2686 static struct bfq_queue *bfq_find_close_cooperator(struct bfq_data *bfqd,
2687 struct bfq_queue *cur_bfqq,
2690 struct bfq_queue *bfqq;
2693 * We shall notice if some of the queues are cooperating,
2694 * e.g., working closely on the same area of the device. In
2695 * that case, we can group them together and: 1) don't waste
2696 * time idling, and 2) serve the union of their requests in
2697 * the best possible order for throughput.
2699 bfqq = bfqq_find_close(bfqd, cur_bfqq, sector);
2700 if (!bfqq || bfqq == cur_bfqq)
2706 static struct bfq_queue *
2707 bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2709 int process_refs, new_process_refs;
2710 struct bfq_queue *__bfqq;
2713 * If there are no process references on the new_bfqq, then it is
2714 * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain
2715 * may have dropped their last reference (not just their last process
2718 if (!bfqq_process_refs(new_bfqq))
2721 /* Avoid a circular list and skip interim queue merges. */
2722 while ((__bfqq = new_bfqq->new_bfqq)) {
2728 process_refs = bfqq_process_refs(bfqq);
2729 new_process_refs = bfqq_process_refs(new_bfqq);
2731 * If the process for the bfqq has gone away, there is no
2732 * sense in merging the queues.
2734 if (process_refs == 0 || new_process_refs == 0)
2738 * Make sure merged queues belong to the same parent. Parents could
2739 * have changed since the time we decided the two queues are suitable
2742 if (new_bfqq->entity.parent != bfqq->entity.parent)
2745 bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d",
2749 * Merging is just a redirection: the requests of the process
2750 * owning one of the two queues are redirected to the other queue.
2751 * The latter queue, in its turn, is set as shared if this is the
2752 * first time that the requests of some process are redirected to
2755 * We redirect bfqq to new_bfqq and not the opposite, because
2756 * we are in the context of the process owning bfqq, thus we
2757 * have the io_cq of this process. So we can immediately
2758 * configure this io_cq to redirect the requests of the
2759 * process to new_bfqq. In contrast, the io_cq of new_bfqq is
2760 * not available any more (new_bfqq->bic == NULL).
2762 * Anyway, even in case new_bfqq coincides with the in-service
2763 * queue, redirecting requests the in-service queue is the
2764 * best option, as we feed the in-service queue with new
2765 * requests close to the last request served and, by doing so,
2766 * are likely to increase the throughput.
2768 bfqq->new_bfqq = new_bfqq;
2770 * The above assignment schedules the following redirections:
2771 * each time some I/O for bfqq arrives, the process that
2772 * generated that I/O is disassociated from bfqq and
2773 * associated with new_bfqq. Here we increases new_bfqq->ref
2774 * in advance, adding the number of processes that are
2775 * expected to be associated with new_bfqq as they happen to
2778 new_bfqq->ref += process_refs;
2782 static bool bfq_may_be_close_cooperator(struct bfq_queue *bfqq,
2783 struct bfq_queue *new_bfqq)
2785 if (bfq_too_late_for_merging(new_bfqq))
2788 if (bfq_class_idle(bfqq) || bfq_class_idle(new_bfqq) ||
2789 (bfqq->ioprio_class != new_bfqq->ioprio_class))
2793 * If either of the queues has already been detected as seeky,
2794 * then merging it with the other queue is unlikely to lead to
2797 if (BFQQ_SEEKY(bfqq) || BFQQ_SEEKY(new_bfqq))
2801 * Interleaved I/O is known to be done by (some) applications
2802 * only for reads, so it does not make sense to merge async
2805 if (!bfq_bfqq_sync(bfqq) || !bfq_bfqq_sync(new_bfqq))
2811 static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd,
2812 struct bfq_queue *bfqq);
2815 * Attempt to schedule a merge of bfqq with the currently in-service
2816 * queue or with a close queue among the scheduled queues. Return
2817 * NULL if no merge was scheduled, a pointer to the shared bfq_queue
2818 * structure otherwise.
2820 * The OOM queue is not allowed to participate to cooperation: in fact, since
2821 * the requests temporarily redirected to the OOM queue could be redirected
2822 * again to dedicated queues at any time, the state needed to correctly
2823 * handle merging with the OOM queue would be quite complex and expensive
2824 * to maintain. Besides, in such a critical condition as an out of memory,
2825 * the benefits of queue merging may be little relevant, or even negligible.
2827 * WARNING: queue merging may impair fairness among non-weight raised
2828 * queues, for at least two reasons: 1) the original weight of a
2829 * merged queue may change during the merged state, 2) even being the
2830 * weight the same, a merged queue may be bloated with many more
2831 * requests than the ones produced by its originally-associated
2834 static struct bfq_queue *
2835 bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2836 void *io_struct, bool request, struct bfq_io_cq *bic)
2838 struct bfq_queue *in_service_bfqq, *new_bfqq;
2840 /* if a merge has already been setup, then proceed with that first */
2842 return bfqq->new_bfqq;
2845 * Check delayed stable merge for rotational or non-queueing
2846 * devs. For this branch to be executed, bfqq must not be
2847 * currently merged with some other queue (i.e., bfqq->bic
2848 * must be non null). If we considered also merged queues,
2849 * then we should also check whether bfqq has already been
2850 * merged with bic->stable_merge_bfqq. But this would be
2851 * costly and complicated.
2853 if (unlikely(!bfqd->nonrot_with_queueing)) {
2855 * Make sure also that bfqq is sync, because
2856 * bic->stable_merge_bfqq may point to some queue (for
2857 * stable merging) also if bic is associated with a
2858 * sync queue, but this bfqq is async
2860 if (bfq_bfqq_sync(bfqq) && bic->stable_merge_bfqq &&
2861 !bfq_bfqq_just_created(bfqq) &&
2862 time_is_before_jiffies(bfqq->split_time +
2863 msecs_to_jiffies(bfq_late_stable_merging)) &&
2864 time_is_before_jiffies(bfqq->creation_time +
2865 msecs_to_jiffies(bfq_late_stable_merging))) {
2866 struct bfq_queue *stable_merge_bfqq =
2867 bic->stable_merge_bfqq;
2868 int proc_ref = min(bfqq_process_refs(bfqq),
2869 bfqq_process_refs(stable_merge_bfqq));
2871 /* deschedule stable merge, because done or aborted here */
2872 bfq_put_stable_ref(stable_merge_bfqq);
2874 bic->stable_merge_bfqq = NULL;
2876 if (!idling_boosts_thr_without_issues(bfqd, bfqq) &&
2878 /* next function will take at least one ref */
2879 struct bfq_queue *new_bfqq =
2880 bfq_setup_merge(bfqq, stable_merge_bfqq);
2883 bic->stably_merged = true;
2885 new_bfqq->bic->stably_merged =
2895 * Do not perform queue merging if the device is non
2896 * rotational and performs internal queueing. In fact, such a
2897 * device reaches a high speed through internal parallelism
2898 * and pipelining. This means that, to reach a high
2899 * throughput, it must have many requests enqueued at the same
2900 * time. But, in this configuration, the internal scheduling
2901 * algorithm of the device does exactly the job of queue
2902 * merging: it reorders requests so as to obtain as much as
2903 * possible a sequential I/O pattern. As a consequence, with
2904 * the workload generated by processes doing interleaved I/O,
2905 * the throughput reached by the device is likely to be the
2906 * same, with and without queue merging.
2908 * Disabling merging also provides a remarkable benefit in
2909 * terms of throughput. Merging tends to make many workloads
2910 * artificially more uneven, because of shared queues
2911 * remaining non empty for incomparably more time than
2912 * non-merged queues. This may accentuate workload
2913 * asymmetries. For example, if one of the queues in a set of
2914 * merged queues has a higher weight than a normal queue, then
2915 * the shared queue may inherit such a high weight and, by
2916 * staying almost always active, may force BFQ to perform I/O
2917 * plugging most of the time. This evidently makes it harder
2918 * for BFQ to let the device reach a high throughput.
2920 * Finally, the likely() macro below is not used because one
2921 * of the two branches is more likely than the other, but to
2922 * have the code path after the following if() executed as
2923 * fast as possible for the case of a non rotational device
2924 * with queueing. We want it because this is the fastest kind
2925 * of device. On the opposite end, the likely() may lengthen
2926 * the execution time of BFQ for the case of slower devices
2927 * (rotational or at least without queueing). But in this case
2928 * the execution time of BFQ matters very little, if not at
2931 if (likely(bfqd->nonrot_with_queueing))
2935 * Prevent bfqq from being merged if it has been created too
2936 * long ago. The idea is that true cooperating processes, and
2937 * thus their associated bfq_queues, are supposed to be
2938 * created shortly after each other. This is the case, e.g.,
2939 * for KVM/QEMU and dump I/O threads. Basing on this
2940 * assumption, the following filtering greatly reduces the
2941 * probability that two non-cooperating processes, which just
2942 * happen to do close I/O for some short time interval, have
2943 * their queues merged by mistake.
2945 if (bfq_too_late_for_merging(bfqq))
2948 if (!io_struct || unlikely(bfqq == &bfqd->oom_bfqq))
2951 /* If there is only one backlogged queue, don't search. */
2952 if (bfq_tot_busy_queues(bfqd) == 1)
2955 in_service_bfqq = bfqd->in_service_queue;
2957 if (in_service_bfqq && in_service_bfqq != bfqq &&
2958 likely(in_service_bfqq != &bfqd->oom_bfqq) &&
2959 bfq_rq_close_to_sector(io_struct, request,
2960 bfqd->in_serv_last_pos) &&
2961 bfqq->entity.parent == in_service_bfqq->entity.parent &&
2962 bfq_may_be_close_cooperator(bfqq, in_service_bfqq)) {
2963 new_bfqq = bfq_setup_merge(bfqq, in_service_bfqq);
2968 * Check whether there is a cooperator among currently scheduled
2969 * queues. The only thing we need is that the bio/request is not
2970 * NULL, as we need it to establish whether a cooperator exists.
2972 new_bfqq = bfq_find_close_cooperator(bfqd, bfqq,
2973 bfq_io_struct_pos(io_struct, request));
2975 if (new_bfqq && likely(new_bfqq != &bfqd->oom_bfqq) &&
2976 bfq_may_be_close_cooperator(bfqq, new_bfqq))
2977 return bfq_setup_merge(bfqq, new_bfqq);
2982 static void bfq_bfqq_save_state(struct bfq_queue *bfqq)
2984 struct bfq_io_cq *bic = bfqq->bic;
2987 * If !bfqq->bic, the queue is already shared or its requests
2988 * have already been redirected to a shared queue; both idle window
2989 * and weight raising state have already been saved. Do nothing.
2994 bic->saved_last_serv_time_ns = bfqq->last_serv_time_ns;
2995 bic->saved_inject_limit = bfqq->inject_limit;
2996 bic->saved_decrease_time_jif = bfqq->decrease_time_jif;
2998 bic->saved_weight = bfqq->entity.orig_weight;
2999 bic->saved_ttime = bfqq->ttime;
3000 bic->saved_has_short_ttime = bfq_bfqq_has_short_ttime(bfqq);
3001 bic->saved_IO_bound = bfq_bfqq_IO_bound(bfqq);
3002 bic->saved_io_start_time = bfqq->io_start_time;
3003 bic->saved_tot_idle_time = bfqq->tot_idle_time;
3004 bic->saved_in_large_burst = bfq_bfqq_in_large_burst(bfqq);
3005 bic->was_in_burst_list = !hlist_unhashed(&bfqq->burst_list_node);
3006 if (unlikely(bfq_bfqq_just_created(bfqq) &&
3007 !bfq_bfqq_in_large_burst(bfqq) &&
3008 bfqq->bfqd->low_latency)) {
3010 * bfqq being merged right after being created: bfqq
3011 * would have deserved interactive weight raising, but
3012 * did not make it to be set in a weight-raised state,
3013 * because of this early merge. Store directly the
3014 * weight-raising state that would have been assigned
3015 * to bfqq, so that to avoid that bfqq unjustly fails
3016 * to enjoy weight raising if split soon.
3018 bic->saved_wr_coeff = bfqq->bfqd->bfq_wr_coeff;
3019 bic->saved_wr_start_at_switch_to_srt = bfq_smallest_from_now();
3020 bic->saved_wr_cur_max_time = bfq_wr_duration(bfqq->bfqd);
3021 bic->saved_last_wr_start_finish = jiffies;
3023 bic->saved_wr_coeff = bfqq->wr_coeff;
3024 bic->saved_wr_start_at_switch_to_srt =
3025 bfqq->wr_start_at_switch_to_srt;
3026 bic->saved_service_from_wr = bfqq->service_from_wr;
3027 bic->saved_last_wr_start_finish = bfqq->last_wr_start_finish;
3028 bic->saved_wr_cur_max_time = bfqq->wr_cur_max_time;
3034 bfq_reassign_last_bfqq(struct bfq_queue *cur_bfqq, struct bfq_queue *new_bfqq)
3036 if (cur_bfqq->entity.parent &&
3037 cur_bfqq->entity.parent->last_bfqq_created == cur_bfqq)
3038 cur_bfqq->entity.parent->last_bfqq_created = new_bfqq;
3039 else if (cur_bfqq->bfqd && cur_bfqq->bfqd->last_bfqq_created == cur_bfqq)
3040 cur_bfqq->bfqd->last_bfqq_created = new_bfqq;
3043 void bfq_release_process_ref(struct bfq_data *bfqd, struct bfq_queue *bfqq)
3046 * To prevent bfqq's service guarantees from being violated,
3047 * bfqq may be left busy, i.e., queued for service, even if
3048 * empty (see comments in __bfq_bfqq_expire() for
3049 * details). But, if no process will send requests to bfqq any
3050 * longer, then there is no point in keeping bfqq queued for
3051 * service. In addition, keeping bfqq queued for service, but
3052 * with no process ref any longer, may have caused bfqq to be
3053 * freed when dequeued from service. But this is assumed to
3056 if (bfq_bfqq_busy(bfqq) && RB_EMPTY_ROOT(&bfqq->sort_list) &&
3057 bfqq != bfqd->in_service_queue)
3058 bfq_del_bfqq_busy(bfqq, false);
3060 bfq_reassign_last_bfqq(bfqq, NULL);
3062 bfq_put_queue(bfqq);
3066 bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic,
3067 struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
3069 bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu",
3070 (unsigned long)new_bfqq->pid);
3071 /* Save weight raising and idle window of the merged queues */
3072 bfq_bfqq_save_state(bfqq);
3073 bfq_bfqq_save_state(new_bfqq);
3074 if (bfq_bfqq_IO_bound(bfqq))
3075 bfq_mark_bfqq_IO_bound(new_bfqq);
3076 bfq_clear_bfqq_IO_bound(bfqq);
3079 * The processes associated with bfqq are cooperators of the
3080 * processes associated with new_bfqq. So, if bfqq has a
3081 * waker, then assume that all these processes will be happy
3082 * to let bfqq's waker freely inject I/O when they have no
3085 if (bfqq->waker_bfqq && !new_bfqq->waker_bfqq &&
3086 bfqq->waker_bfqq != new_bfqq) {
3087 new_bfqq->waker_bfqq = bfqq->waker_bfqq;
3088 new_bfqq->tentative_waker_bfqq = NULL;
3091 * If the waker queue disappears, then
3092 * new_bfqq->waker_bfqq must be reset. So insert
3093 * new_bfqq into the woken_list of the waker. See
3094 * bfq_check_waker for details.
3096 hlist_add_head(&new_bfqq->woken_list_node,
3097 &new_bfqq->waker_bfqq->woken_list);
3102 * If bfqq is weight-raised, then let new_bfqq inherit
3103 * weight-raising. To reduce false positives, neglect the case
3104 * where bfqq has just been created, but has not yet made it
3105 * to be weight-raised (which may happen because EQM may merge
3106 * bfqq even before bfq_add_request is executed for the first
3107 * time for bfqq). Handling this case would however be very
3108 * easy, thanks to the flag just_created.
3110 if (new_bfqq->wr_coeff == 1 && bfqq->wr_coeff > 1) {
3111 new_bfqq->wr_coeff = bfqq->wr_coeff;
3112 new_bfqq->wr_cur_max_time = bfqq->wr_cur_max_time;
3113 new_bfqq->last_wr_start_finish = bfqq->last_wr_start_finish;
3114 new_bfqq->wr_start_at_switch_to_srt =
3115 bfqq->wr_start_at_switch_to_srt;
3116 if (bfq_bfqq_busy(new_bfqq))
3117 bfqd->wr_busy_queues++;
3118 new_bfqq->entity.prio_changed = 1;
3121 if (bfqq->wr_coeff > 1) { /* bfqq has given its wr to new_bfqq */
3123 bfqq->entity.prio_changed = 1;
3124 if (bfq_bfqq_busy(bfqq))
3125 bfqd->wr_busy_queues--;
3128 bfq_log_bfqq(bfqd, new_bfqq, "merge_bfqqs: wr_busy %d",
3129 bfqd->wr_busy_queues);
3132 * Merge queues (that is, let bic redirect its requests to new_bfqq)
3134 bic_set_bfqq(bic, new_bfqq, 1);
3135 bfq_mark_bfqq_coop(new_bfqq);
3137 * new_bfqq now belongs to at least two bics (it is a shared queue):
3138 * set new_bfqq->bic to NULL. bfqq either:
3139 * - does not belong to any bic any more, and hence bfqq->bic must
3140 * be set to NULL, or
3141 * - is a queue whose owning bics have already been redirected to a
3142 * different queue, hence the queue is destined to not belong to
3143 * any bic soon and bfqq->bic is already NULL (therefore the next
3144 * assignment causes no harm).
3146 new_bfqq->bic = NULL;
3148 * If the queue is shared, the pid is the pid of one of the associated
3149 * processes. Which pid depends on the exact sequence of merge events
3150 * the queue underwent. So printing such a pid is useless and confusing
3151 * because it reports a random pid between those of the associated
3153 * We mark such a queue with a pid -1, and then print SHARED instead of
3154 * a pid in logging messages.
3159 bfq_reassign_last_bfqq(bfqq, new_bfqq);
3161 bfq_release_process_ref(bfqd, bfqq);
3164 static bool bfq_allow_bio_merge(struct request_queue *q, struct request *rq,
3167 struct bfq_data *bfqd = q->elevator->elevator_data;
3168 bool is_sync = op_is_sync(bio->bi_opf);
3169 struct bfq_queue *bfqq = bfqd->bio_bfqq, *new_bfqq;
3172 * Disallow merge of a sync bio into an async request.
3174 if (is_sync && !rq_is_sync(rq))
3178 * Lookup the bfqq that this bio will be queued with. Allow
3179 * merge only if rq is queued there.
3185 * We take advantage of this function to perform an early merge
3186 * of the queues of possible cooperating processes.
3188 new_bfqq = bfq_setup_cooperator(bfqd, bfqq, bio, false, bfqd->bio_bic);
3191 * bic still points to bfqq, then it has not yet been
3192 * redirected to some other bfq_queue, and a queue
3193 * merge between bfqq and new_bfqq can be safely
3194 * fulfilled, i.e., bic can be redirected to new_bfqq
3195 * and bfqq can be put.
3197 bfq_merge_bfqqs(bfqd, bfqd->bio_bic, bfqq,
3200 * If we get here, bio will be queued into new_queue,
3201 * so use new_bfqq to decide whether bio and rq can be
3207 * Change also bqfd->bio_bfqq, as
3208 * bfqd->bio_bic now points to new_bfqq, and
3209 * this function may be invoked again (and then may
3210 * use again bqfd->bio_bfqq).
3212 bfqd->bio_bfqq = bfqq;
3215 return bfqq == RQ_BFQQ(rq);
3219 * Set the maximum time for the in-service queue to consume its
3220 * budget. This prevents seeky processes from lowering the throughput.
3221 * In practice, a time-slice service scheme is used with seeky
3224 static void bfq_set_budget_timeout(struct bfq_data *bfqd,
3225 struct bfq_queue *bfqq)
3227 unsigned int timeout_coeff;
3229 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time)
3232 timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight;
3234 bfqd->last_budget_start = ktime_get();
3236 bfqq->budget_timeout = jiffies +
3237 bfqd->bfq_timeout * timeout_coeff;
3240 static void __bfq_set_in_service_queue(struct bfq_data *bfqd,
3241 struct bfq_queue *bfqq)
3244 bfq_clear_bfqq_fifo_expire(bfqq);
3246 bfqd->budgets_assigned = (bfqd->budgets_assigned * 7 + 256) / 8;
3248 if (time_is_before_jiffies(bfqq->last_wr_start_finish) &&
3249 bfqq->wr_coeff > 1 &&
3250 bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
3251 time_is_before_jiffies(bfqq->budget_timeout)) {
3253 * For soft real-time queues, move the start
3254 * of the weight-raising period forward by the
3255 * time the queue has not received any
3256 * service. Otherwise, a relatively long
3257 * service delay is likely to cause the
3258 * weight-raising period of the queue to end,
3259 * because of the short duration of the
3260 * weight-raising period of a soft real-time
3261 * queue. It is worth noting that this move
3262 * is not so dangerous for the other queues,
3263 * because soft real-time queues are not
3266 * To not add a further variable, we use the
3267 * overloaded field budget_timeout to
3268 * determine for how long the queue has not
3269 * received service, i.e., how much time has
3270 * elapsed since the queue expired. However,
3271 * this is a little imprecise, because
3272 * budget_timeout is set to jiffies if bfqq
3273 * not only expires, but also remains with no
3276 if (time_after(bfqq->budget_timeout,
3277 bfqq->last_wr_start_finish))
3278 bfqq->last_wr_start_finish +=
3279 jiffies - bfqq->budget_timeout;
3281 bfqq->last_wr_start_finish = jiffies;
3284 bfq_set_budget_timeout(bfqd, bfqq);
3285 bfq_log_bfqq(bfqd, bfqq,
3286 "set_in_service_queue, cur-budget = %d",
3287 bfqq->entity.budget);
3290 bfqd->in_service_queue = bfqq;
3291 bfqd->in_serv_last_pos = 0;
3295 * Get and set a new queue for service.
3297 static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd)
3299 struct bfq_queue *bfqq = bfq_get_next_queue(bfqd);
3301 __bfq_set_in_service_queue(bfqd, bfqq);
3305 static void bfq_arm_slice_timer(struct bfq_data *bfqd)
3307 struct bfq_queue *bfqq = bfqd->in_service_queue;
3310 bfq_mark_bfqq_wait_request(bfqq);
3313 * We don't want to idle for seeks, but we do want to allow
3314 * fair distribution of slice time for a process doing back-to-back
3315 * seeks. So allow a little bit of time for him to submit a new rq.
3317 sl = bfqd->bfq_slice_idle;
3319 * Unless the queue is being weight-raised or the scenario is
3320 * asymmetric, grant only minimum idle time if the queue
3321 * is seeky. A long idling is preserved for a weight-raised
3322 * queue, or, more in general, in an asymmetric scenario,
3323 * because a long idling is needed for guaranteeing to a queue
3324 * its reserved share of the throughput (in particular, it is
3325 * needed if the queue has a higher weight than some other
3328 if (BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 &&
3329 !bfq_asymmetric_scenario(bfqd, bfqq))
3330 sl = min_t(u64, sl, BFQ_MIN_TT);
3331 else if (bfqq->wr_coeff > 1)
3332 sl = max_t(u32, sl, 20ULL * NSEC_PER_MSEC);
3334 bfqd->last_idling_start = ktime_get();
3335 bfqd->last_idling_start_jiffies = jiffies;
3337 hrtimer_start(&bfqd->idle_slice_timer, ns_to_ktime(sl),
3339 bfqg_stats_set_start_idle_time(bfqq_group(bfqq));
3343 * In autotuning mode, max_budget is dynamically recomputed as the
3344 * amount of sectors transferred in timeout at the estimated peak
3345 * rate. This enables BFQ to utilize a full timeslice with a full
3346 * budget, even if the in-service queue is served at peak rate. And
3347 * this maximises throughput with sequential workloads.
3349 static unsigned long bfq_calc_max_budget(struct bfq_data *bfqd)
3351 return (u64)bfqd->peak_rate * USEC_PER_MSEC *
3352 jiffies_to_msecs(bfqd->bfq_timeout)>>BFQ_RATE_SHIFT;
3356 * Update parameters related to throughput and responsiveness, as a
3357 * function of the estimated peak rate. See comments on
3358 * bfq_calc_max_budget(), and on the ref_wr_duration array.
3360 static void update_thr_responsiveness_params(struct bfq_data *bfqd)
3362 if (bfqd->bfq_user_max_budget == 0) {
3363 bfqd->bfq_max_budget =
3364 bfq_calc_max_budget(bfqd);
3365 bfq_log(bfqd, "new max_budget = %d", bfqd->bfq_max_budget);
3369 static void bfq_reset_rate_computation(struct bfq_data *bfqd,
3372 if (rq != NULL) { /* new rq dispatch now, reset accordingly */
3373 bfqd->last_dispatch = bfqd->first_dispatch = ktime_get_ns();
3374 bfqd->peak_rate_samples = 1;
3375 bfqd->sequential_samples = 0;
3376 bfqd->tot_sectors_dispatched = bfqd->last_rq_max_size =
3378 } else /* no new rq dispatched, just reset the number of samples */
3379 bfqd->peak_rate_samples = 0; /* full re-init on next disp. */
3382 "reset_rate_computation at end, sample %u/%u tot_sects %llu",
3383 bfqd->peak_rate_samples, bfqd->sequential_samples,
3384 bfqd->tot_sectors_dispatched);
3387 static void bfq_update_rate_reset(struct bfq_data *bfqd, struct request *rq)
3389 u32 rate, weight, divisor;
3392 * For the convergence property to hold (see comments on
3393 * bfq_update_peak_rate()) and for the assessment to be
3394 * reliable, a minimum number of samples must be present, and
3395 * a minimum amount of time must have elapsed. If not so, do
3396 * not compute new rate. Just reset parameters, to get ready
3397 * for a new evaluation attempt.
3399 if (bfqd->peak_rate_samples < BFQ_RATE_MIN_SAMPLES ||
3400 bfqd->delta_from_first < BFQ_RATE_MIN_INTERVAL)
3401 goto reset_computation;
3404 * If a new request completion has occurred after last
3405 * dispatch, then, to approximate the rate at which requests
3406 * have been served by the device, it is more precise to
3407 * extend the observation interval to the last completion.
3409 bfqd->delta_from_first =
3410 max_t(u64, bfqd->delta_from_first,
3411 bfqd->last_completion - bfqd->first_dispatch);
3414 * Rate computed in sects/usec, and not sects/nsec, for
3417 rate = div64_ul(bfqd->tot_sectors_dispatched<<BFQ_RATE_SHIFT,
3418 div_u64(bfqd->delta_from_first, NSEC_PER_USEC));
3421 * Peak rate not updated if:
3422 * - the percentage of sequential dispatches is below 3/4 of the
3423 * total, and rate is below the current estimated peak rate
3424 * - rate is unreasonably high (> 20M sectors/sec)
3426 if ((bfqd->sequential_samples < (3 * bfqd->peak_rate_samples)>>2 &&
3427 rate <= bfqd->peak_rate) ||
3428 rate > 20<<BFQ_RATE_SHIFT)
3429 goto reset_computation;
3432 * We have to update the peak rate, at last! To this purpose,
3433 * we use a low-pass filter. We compute the smoothing constant
3434 * of the filter as a function of the 'weight' of the new
3437 * As can be seen in next formulas, we define this weight as a
3438 * quantity proportional to how sequential the workload is,
3439 * and to how long the observation time interval is.
3441 * The weight runs from 0 to 8. The maximum value of the
3442 * weight, 8, yields the minimum value for the smoothing
3443 * constant. At this minimum value for the smoothing constant,
3444 * the measured rate contributes for half of the next value of
3445 * the estimated peak rate.
3447 * So, the first step is to compute the weight as a function
3448 * of how sequential the workload is. Note that the weight
3449 * cannot reach 9, because bfqd->sequential_samples cannot
3450 * become equal to bfqd->peak_rate_samples, which, in its
3451 * turn, holds true because bfqd->sequential_samples is not
3452 * incremented for the first sample.
3454 weight = (9 * bfqd->sequential_samples) / bfqd->peak_rate_samples;
3457 * Second step: further refine the weight as a function of the
3458 * duration of the observation interval.
3460 weight = min_t(u32, 8,
3461 div_u64(weight * bfqd->delta_from_first,
3462 BFQ_RATE_REF_INTERVAL));
3465 * Divisor ranging from 10, for minimum weight, to 2, for
3468 divisor = 10 - weight;
3471 * Finally, update peak rate:
3473 * peak_rate = peak_rate * (divisor-1) / divisor + rate / divisor
3475 bfqd->peak_rate *= divisor-1;
3476 bfqd->peak_rate /= divisor;
3477 rate /= divisor; /* smoothing constant alpha = 1/divisor */
3479 bfqd->peak_rate += rate;
3482 * For a very slow device, bfqd->peak_rate can reach 0 (see
3483 * the minimum representable values reported in the comments
3484 * on BFQ_RATE_SHIFT). Push to 1 if this happens, to avoid
3485 * divisions by zero where bfqd->peak_rate is used as a
3488 bfqd->peak_rate = max_t(u32, 1, bfqd->peak_rate);
3490 update_thr_responsiveness_params(bfqd);
3493 bfq_reset_rate_computation(bfqd, rq);
3497 * Update the read/write peak rate (the main quantity used for
3498 * auto-tuning, see update_thr_responsiveness_params()).
3500 * It is not trivial to estimate the peak rate (correctly): because of
3501 * the presence of sw and hw queues between the scheduler and the
3502 * device components that finally serve I/O requests, it is hard to
3503 * say exactly when a given dispatched request is served inside the
3504 * device, and for how long. As a consequence, it is hard to know
3505 * precisely at what rate a given set of requests is actually served
3508 * On the opposite end, the dispatch time of any request is trivially
3509 * available, and, from this piece of information, the "dispatch rate"
3510 * of requests can be immediately computed. So, the idea in the next
3511 * function is to use what is known, namely request dispatch times
3512 * (plus, when useful, request completion times), to estimate what is
3513 * unknown, namely in-device request service rate.
3515 * The main issue is that, because of the above facts, the rate at
3516 * which a certain set of requests is dispatched over a certain time
3517 * interval can vary greatly with respect to the rate at which the
3518 * same requests are then served. But, since the size of any
3519 * intermediate queue is limited, and the service scheme is lossless
3520 * (no request is silently dropped), the following obvious convergence
3521 * property holds: the number of requests dispatched MUST become
3522 * closer and closer to the number of requests completed as the
3523 * observation interval grows. This is the key property used in
3524 * the next function to estimate the peak service rate as a function
3525 * of the observed dispatch rate. The function assumes to be invoked
3526 * on every request dispatch.
3528 static void bfq_update_peak_rate(struct bfq_data *bfqd, struct request *rq)
3530 u64 now_ns = ktime_get_ns();
3532 if (bfqd->peak_rate_samples == 0) { /* first dispatch */
3533 bfq_log(bfqd, "update_peak_rate: goto reset, samples %d",
3534 bfqd->peak_rate_samples);
3535 bfq_reset_rate_computation(bfqd, rq);
3536 goto update_last_values; /* will add one sample */
3540 * Device idle for very long: the observation interval lasting
3541 * up to this dispatch cannot be a valid observation interval
3542 * for computing a new peak rate (similarly to the late-
3543 * completion event in bfq_completed_request()). Go to
3544 * update_rate_and_reset to have the following three steps
3546 * - close the observation interval at the last (previous)
3547 * request dispatch or completion
3548 * - compute rate, if possible, for that observation interval
3549 * - start a new observation interval with this dispatch
3551 if (now_ns - bfqd->last_dispatch > 100*NSEC_PER_MSEC &&
3552 bfqd->rq_in_driver == 0)
3553 goto update_rate_and_reset;
3555 /* Update sampling information */
3556 bfqd->peak_rate_samples++;
3558 if ((bfqd->rq_in_driver > 0 ||
3559 now_ns - bfqd->last_completion < BFQ_MIN_TT)
3560 && !BFQ_RQ_SEEKY(bfqd, bfqd->last_position, rq))
3561 bfqd->sequential_samples++;
3563 bfqd->tot_sectors_dispatched += blk_rq_sectors(rq);
3565 /* Reset max observed rq size every 32 dispatches */
3566 if (likely(bfqd->peak_rate_samples % 32))
3567 bfqd->last_rq_max_size =
3568 max_t(u32, blk_rq_sectors(rq), bfqd->last_rq_max_size);
3570 bfqd->last_rq_max_size = blk_rq_sectors(rq);
3572 bfqd->delta_from_first = now_ns - bfqd->first_dispatch;
3574 /* Target observation interval not yet reached, go on sampling */
3575 if (bfqd->delta_from_first < BFQ_RATE_REF_INTERVAL)
3576 goto update_last_values;
3578 update_rate_and_reset:
3579 bfq_update_rate_reset(bfqd, rq);
3581 bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
3582 if (RQ_BFQQ(rq) == bfqd->in_service_queue)
3583 bfqd->in_serv_last_pos = bfqd->last_position;
3584 bfqd->last_dispatch = now_ns;
3588 * Remove request from internal lists.
3590 static void bfq_dispatch_remove(struct request_queue *q, struct request *rq)
3592 struct bfq_queue *bfqq = RQ_BFQQ(rq);
3595 * For consistency, the next instruction should have been
3596 * executed after removing the request from the queue and
3597 * dispatching it. We execute instead this instruction before
3598 * bfq_remove_request() (and hence introduce a temporary
3599 * inconsistency), for efficiency. In fact, should this
3600 * dispatch occur for a non in-service bfqq, this anticipated
3601 * increment prevents two counters related to bfqq->dispatched
3602 * from risking to be, first, uselessly decremented, and then
3603 * incremented again when the (new) value of bfqq->dispatched
3604 * happens to be taken into account.
3607 bfq_update_peak_rate(q->elevator->elevator_data, rq);
3609 bfq_remove_request(q, rq);
3613 * There is a case where idling does not have to be performed for
3614 * throughput concerns, but to preserve the throughput share of
3615 * the process associated with bfqq.
3617 * To introduce this case, we can note that allowing the drive
3618 * to enqueue more than one request at a time, and hence
3619 * delegating de facto final scheduling decisions to the
3620 * drive's internal scheduler, entails loss of control on the
3621 * actual request service order. In particular, the critical
3622 * situation is when requests from different processes happen
3623 * to be present, at the same time, in the internal queue(s)
3624 * of the drive. In such a situation, the drive, by deciding
3625 * the service order of the internally-queued requests, does
3626 * determine also the actual throughput distribution among
3627 * these processes. But the drive typically has no notion or
3628 * concern about per-process throughput distribution, and
3629 * makes its decisions only on a per-request basis. Therefore,
3630 * the service distribution enforced by the drive's internal
3631 * scheduler is likely to coincide with the desired throughput
3632 * distribution only in a completely symmetric, or favorably
3633 * skewed scenario where:
3634 * (i-a) each of these processes must get the same throughput as
3636 * (i-b) in case (i-a) does not hold, it holds that the process
3637 * associated with bfqq must receive a lower or equal
3638 * throughput than any of the other processes;
3639 * (ii) the I/O of each process has the same properties, in
3640 * terms of locality (sequential or random), direction
3641 * (reads or writes), request sizes, greediness
3642 * (from I/O-bound to sporadic), and so on;
3644 * In fact, in such a scenario, the drive tends to treat the requests
3645 * of each process in about the same way as the requests of the
3646 * others, and thus to provide each of these processes with about the
3647 * same throughput. This is exactly the desired throughput
3648 * distribution if (i-a) holds, or, if (i-b) holds instead, this is an
3649 * even more convenient distribution for (the process associated with)
3652 * In contrast, in any asymmetric or unfavorable scenario, device
3653 * idling (I/O-dispatch plugging) is certainly needed to guarantee
3654 * that bfqq receives its assigned fraction of the device throughput
3655 * (see [1] for details).
3657 * The problem is that idling may significantly reduce throughput with
3658 * certain combinations of types of I/O and devices. An important
3659 * example is sync random I/O on flash storage with command
3660 * queueing. So, unless bfqq falls in cases where idling also boosts
3661 * throughput, it is important to check conditions (i-a), i(-b) and
3662 * (ii) accurately, so as to avoid idling when not strictly needed for
3663 * service guarantees.
3665 * Unfortunately, it is extremely difficult to thoroughly check
3666 * condition (ii). And, in case there are active groups, it becomes
3667 * very difficult to check conditions (i-a) and (i-b) too. In fact,
3668 * if there are active groups, then, for conditions (i-a) or (i-b) to
3669 * become false 'indirectly', it is enough that an active group
3670 * contains more active processes or sub-groups than some other active
3671 * group. More precisely, for conditions (i-a) or (i-b) to become
3672 * false because of such a group, it is not even necessary that the
3673 * group is (still) active: it is sufficient that, even if the group
3674 * has become inactive, some of its descendant processes still have
3675 * some request already dispatched but still waiting for
3676 * completion. In fact, requests have still to be guaranteed their
3677 * share of the throughput even after being dispatched. In this
3678 * respect, it is easy to show that, if a group frequently becomes
3679 * inactive while still having in-flight requests, and if, when this
3680 * happens, the group is not considered in the calculation of whether
3681 * the scenario is asymmetric, then the group may fail to be
3682 * guaranteed its fair share of the throughput (basically because
3683 * idling may not be performed for the descendant processes of the
3684 * group, but it had to be). We address this issue with the following
3685 * bi-modal behavior, implemented in the function
3686 * bfq_asymmetric_scenario().
3688 * If there are groups with requests waiting for completion
3689 * (as commented above, some of these groups may even be
3690 * already inactive), then the scenario is tagged as
3691 * asymmetric, conservatively, without checking any of the
3692 * conditions (i-a), (i-b) or (ii). So the device is idled for bfqq.
3693 * This behavior matches also the fact that groups are created
3694 * exactly if controlling I/O is a primary concern (to
3695 * preserve bandwidth and latency guarantees).
3697 * On the opposite end, if there are no groups with requests waiting
3698 * for completion, then only conditions (i-a) and (i-b) are actually
3699 * controlled, i.e., provided that conditions (i-a) or (i-b) holds,
3700 * idling is not performed, regardless of whether condition (ii)
3701 * holds. In other words, only if conditions (i-a) and (i-b) do not
3702 * hold, then idling is allowed, and the device tends to be prevented
3703 * from queueing many requests, possibly of several processes. Since
3704 * there are no groups with requests waiting for completion, then, to
3705 * control conditions (i-a) and (i-b) it is enough to check just
3706 * whether all the queues with requests waiting for completion also
3707 * have the same weight.
3709 * Not checking condition (ii) evidently exposes bfqq to the
3710 * risk of getting less throughput than its fair share.
3711 * However, for queues with the same weight, a further
3712 * mechanism, preemption, mitigates or even eliminates this
3713 * problem. And it does so without consequences on overall
3714 * throughput. This mechanism and its benefits are explained
3715 * in the next three paragraphs.
3717 * Even if a queue, say Q, is expired when it remains idle, Q
3718 * can still preempt the new in-service queue if the next
3719 * request of Q arrives soon (see the comments on
3720 * bfq_bfqq_update_budg_for_activation). If all queues and
3721 * groups have the same weight, this form of preemption,
3722 * combined with the hole-recovery heuristic described in the
3723 * comments on function bfq_bfqq_update_budg_for_activation,
3724 * are enough to preserve a correct bandwidth distribution in
3725 * the mid term, even without idling. In fact, even if not
3726 * idling allows the internal queues of the device to contain
3727 * many requests, and thus to reorder requests, we can rather
3728 * safely assume that the internal scheduler still preserves a
3729 * minimum of mid-term fairness.
3731 * More precisely, this preemption-based, idleless approach
3732 * provides fairness in terms of IOPS, and not sectors per
3733 * second. This can be seen with a simple example. Suppose
3734 * that there are two queues with the same weight, but that
3735 * the first queue receives requests of 8 sectors, while the
3736 * second queue receives requests of 1024 sectors. In
3737 * addition, suppose that each of the two queues contains at
3738 * most one request at a time, which implies that each queue
3739 * always remains idle after it is served. Finally, after
3740 * remaining idle, each queue receives very quickly a new
3741 * request. It follows that the two queues are served
3742 * alternatively, preempting each other if needed. This
3743 * implies that, although both queues have the same weight,
3744 * the queue with large requests receives a service that is
3745 * 1024/8 times as high as the service received by the other
3748 * The motivation for using preemption instead of idling (for
3749 * queues with the same weight) is that, by not idling,
3750 * service guarantees are preserved (completely or at least in
3751 * part) without minimally sacrificing throughput. And, if
3752 * there is no active group, then the primary expectation for
3753 * this device is probably a high throughput.
3755 * We are now left only with explaining the two sub-conditions in the
3756 * additional compound condition that is checked below for deciding
3757 * whether the scenario is asymmetric. To explain the first
3758 * sub-condition, we need to add that the function
3759 * bfq_asymmetric_scenario checks the weights of only
3760 * non-weight-raised queues, for efficiency reasons (see comments on
3761 * bfq_weights_tree_add()). Then the fact that bfqq is weight-raised
3762 * is checked explicitly here. More precisely, the compound condition
3763 * below takes into account also the fact that, even if bfqq is being
3764 * weight-raised, the scenario is still symmetric if all queues with
3765 * requests waiting for completion happen to be
3766 * weight-raised. Actually, we should be even more precise here, and
3767 * differentiate between interactive weight raising and soft real-time
3770 * The second sub-condition checked in the compound condition is
3771 * whether there is a fair amount of already in-flight I/O not
3772 * belonging to bfqq. If so, I/O dispatching is to be plugged, for the
3773 * following reason. The drive may decide to serve in-flight
3774 * non-bfqq's I/O requests before bfqq's ones, thereby delaying the
3775 * arrival of new I/O requests for bfqq (recall that bfqq is sync). If
3776 * I/O-dispatching is not plugged, then, while bfqq remains empty, a
3777 * basically uncontrolled amount of I/O from other queues may be
3778 * dispatched too, possibly causing the service of bfqq's I/O to be
3779 * delayed even longer in the drive. This problem gets more and more
3780 * serious as the speed and the queue depth of the drive grow,
3781 * because, as these two quantities grow, the probability to find no
3782 * queue busy but many requests in flight grows too. By contrast,
3783 * plugging I/O dispatching minimizes the delay induced by already
3784 * in-flight I/O, and enables bfqq to recover the bandwidth it may
3785 * lose because of this delay.
3787 * As a side note, it is worth considering that the above
3788 * device-idling countermeasures may however fail in the following
3789 * unlucky scenario: if I/O-dispatch plugging is (correctly) disabled
3790 * in a time period during which all symmetry sub-conditions hold, and
3791 * therefore the device is allowed to enqueue many requests, but at
3792 * some later point in time some sub-condition stops to hold, then it
3793 * may become impossible to make requests be served in the desired
3794 * order until all the requests already queued in the device have been
3795 * served. The last sub-condition commented above somewhat mitigates
3796 * this problem for weight-raised queues.
3798 * However, as an additional mitigation for this problem, we preserve
3799 * plugging for a special symmetric case that may suddenly turn into
3800 * asymmetric: the case where only bfqq is busy. In this case, not
3801 * expiring bfqq does not cause any harm to any other queues in terms
3802 * of service guarantees. In contrast, it avoids the following unlucky
3803 * sequence of events: (1) bfqq is expired, (2) a new queue with a
3804 * lower weight than bfqq becomes busy (or more queues), (3) the new
3805 * queue is served until a new request arrives for bfqq, (4) when bfqq
3806 * is finally served, there are so many requests of the new queue in
3807 * the drive that the pending requests for bfqq take a lot of time to
3808 * be served. In particular, event (2) may case even already
3809 * dispatched requests of bfqq to be delayed, inside the drive. So, to
3810 * avoid this series of events, the scenario is preventively declared
3811 * as asymmetric also if bfqq is the only busy queues
3813 static bool idling_needed_for_service_guarantees(struct bfq_data *bfqd,
3814 struct bfq_queue *bfqq)
3816 int tot_busy_queues = bfq_tot_busy_queues(bfqd);
3818 /* No point in idling for bfqq if it won't get requests any longer */
3819 if (unlikely(!bfqq_process_refs(bfqq)))
3822 return (bfqq->wr_coeff > 1 &&
3823 (bfqd->wr_busy_queues <
3825 bfqd->rq_in_driver >=
3826 bfqq->dispatched + 4)) ||
3827 bfq_asymmetric_scenario(bfqd, bfqq) ||
3828 tot_busy_queues == 1;
3831 static bool __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq,
3832 enum bfqq_expiration reason)
3835 * If this bfqq is shared between multiple processes, check
3836 * to make sure that those processes are still issuing I/Os
3837 * within the mean seek distance. If not, it may be time to
3838 * break the queues apart again.
3840 if (bfq_bfqq_coop(bfqq) && BFQQ_SEEKY(bfqq))
3841 bfq_mark_bfqq_split_coop(bfqq);
3844 * Consider queues with a higher finish virtual time than
3845 * bfqq. If idling_needed_for_service_guarantees(bfqq) returns
3846 * true, then bfqq's bandwidth would be violated if an
3847 * uncontrolled amount of I/O from these queues were
3848 * dispatched while bfqq is waiting for its new I/O to
3849 * arrive. This is exactly what may happen if this is a forced
3850 * expiration caused by a preemption attempt, and if bfqq is
3851 * not re-scheduled. To prevent this from happening, re-queue
3852 * bfqq if it needs I/O-dispatch plugging, even if it is
3853 * empty. By doing so, bfqq is granted to be served before the
3854 * above queues (provided that bfqq is of course eligible).
3856 if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
3857 !(reason == BFQQE_PREEMPTED &&
3858 idling_needed_for_service_guarantees(bfqd, bfqq))) {
3859 if (bfqq->dispatched == 0)
3861 * Overloading budget_timeout field to store
3862 * the time at which the queue remains with no
3863 * backlog and no outstanding request; used by
3864 * the weight-raising mechanism.
3866 bfqq->budget_timeout = jiffies;
3868 bfq_del_bfqq_busy(bfqq, true);
3870 bfq_requeue_bfqq(bfqd, bfqq, true);
3872 * Resort priority tree of potential close cooperators.
3873 * See comments on bfq_pos_tree_add_move() for the unlikely().
3875 if (unlikely(!bfqd->nonrot_with_queueing &&
3876 !RB_EMPTY_ROOT(&bfqq->sort_list)))
3877 bfq_pos_tree_add_move(bfqd, bfqq);
3881 * All in-service entities must have been properly deactivated
3882 * or requeued before executing the next function, which
3883 * resets all in-service entities as no more in service. This
3884 * may cause bfqq to be freed. If this happens, the next
3885 * function returns true.
3887 return __bfq_bfqd_reset_in_service(bfqd);
3891 * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior.
3892 * @bfqd: device data.
3893 * @bfqq: queue to update.
3894 * @reason: reason for expiration.
3896 * Handle the feedback on @bfqq budget at queue expiration.
3897 * See the body for detailed comments.
3899 static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd,
3900 struct bfq_queue *bfqq,
3901 enum bfqq_expiration reason)
3903 struct request *next_rq;
3904 int budget, min_budget;
3906 min_budget = bfq_min_budget(bfqd);
3908 if (bfqq->wr_coeff == 1)
3909 budget = bfqq->max_budget;
3911 * Use a constant, low budget for weight-raised queues,
3912 * to help achieve a low latency. Keep it slightly higher
3913 * than the minimum possible budget, to cause a little
3914 * bit fewer expirations.
3916 budget = 2 * min_budget;
3918 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %d, budg left %d",
3919 bfqq->entity.budget, bfq_bfqq_budget_left(bfqq));
3920 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %d, min budg %d",
3921 budget, bfq_min_budget(bfqd));
3922 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d",
3923 bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->in_service_queue));
3925 if (bfq_bfqq_sync(bfqq) && bfqq->wr_coeff == 1) {
3928 * Caveat: in all the following cases we trade latency
3931 case BFQQE_TOO_IDLE:
3933 * This is the only case where we may reduce
3934 * the budget: if there is no request of the
3935 * process still waiting for completion, then
3936 * we assume (tentatively) that the timer has
3937 * expired because the batch of requests of
3938 * the process could have been served with a
3939 * smaller budget. Hence, betting that
3940 * process will behave in the same way when it
3941 * becomes backlogged again, we reduce its
3942 * next budget. As long as we guess right,
3943 * this budget cut reduces the latency
3944 * experienced by the process.
3946 * However, if there are still outstanding
3947 * requests, then the process may have not yet
3948 * issued its next request just because it is
3949 * still waiting for the completion of some of
3950 * the still outstanding ones. So in this
3951 * subcase we do not reduce its budget, on the
3952 * contrary we increase it to possibly boost
3953 * the throughput, as discussed in the
3954 * comments to the BUDGET_TIMEOUT case.
3956 if (bfqq->dispatched > 0) /* still outstanding reqs */
3957 budget = min(budget * 2, bfqd->bfq_max_budget);
3959 if (budget > 5 * min_budget)
3960 budget -= 4 * min_budget;
3962 budget = min_budget;
3965 case BFQQE_BUDGET_TIMEOUT:
3967 * We double the budget here because it gives
3968 * the chance to boost the throughput if this
3969 * is not a seeky process (and has bumped into
3970 * this timeout because of, e.g., ZBR).
3972 budget = min(budget * 2, bfqd->bfq_max_budget);
3974 case BFQQE_BUDGET_EXHAUSTED:
3976 * The process still has backlog, and did not
3977 * let either the budget timeout or the disk
3978 * idling timeout expire. Hence it is not
3979 * seeky, has a short thinktime and may be
3980 * happy with a higher budget too. So
3981 * definitely increase the budget of this good
3982 * candidate to boost the disk throughput.
3984 budget = min(budget * 4, bfqd->bfq_max_budget);
3986 case BFQQE_NO_MORE_REQUESTS:
3988 * For queues that expire for this reason, it
3989 * is particularly important to keep the
3990 * budget close to the actual service they
3991 * need. Doing so reduces the timestamp
3992 * misalignment problem described in the
3993 * comments in the body of
3994 * __bfq_activate_entity. In fact, suppose
3995 * that a queue systematically expires for
3996 * BFQQE_NO_MORE_REQUESTS and presents a
3997 * new request in time to enjoy timestamp
3998 * back-shifting. The larger the budget of the
3999 * queue is with respect to the service the
4000 * queue actually requests in each service
4001 * slot, the more times the queue can be
4002 * reactivated with the same virtual finish
4003 * time. It follows that, even if this finish
4004 * time is pushed to the system virtual time
4005 * to reduce the consequent timestamp
4006 * misalignment, the queue unjustly enjoys for
4007 * many re-activations a lower finish time
4008 * than all newly activated queues.
4010 * The service needed by bfqq is measured
4011 * quite precisely by bfqq->entity.service.
4012 * Since bfqq does not enjoy device idling,
4013 * bfqq->entity.service is equal to the number
4014 * of sectors that the process associated with
4015 * bfqq requested to read/write before waiting
4016 * for request completions, or blocking for
4019 budget = max_t(int, bfqq->entity.service, min_budget);
4024 } else if (!bfq_bfqq_sync(bfqq)) {
4026 * Async queues get always the maximum possible
4027 * budget, as for them we do not care about latency
4028 * (in addition, their ability to dispatch is limited
4029 * by the charging factor).
4031 budget = bfqd->bfq_max_budget;
4034 bfqq->max_budget = budget;
4036 if (bfqd->budgets_assigned >= bfq_stats_min_budgets &&
4037 !bfqd->bfq_user_max_budget)
4038 bfqq->max_budget = min(bfqq->max_budget, bfqd->bfq_max_budget);
4041 * If there is still backlog, then assign a new budget, making
4042 * sure that it is large enough for the next request. Since
4043 * the finish time of bfqq must be kept in sync with the
4044 * budget, be sure to call __bfq_bfqq_expire() *after* this
4047 * If there is no backlog, then no need to update the budget;
4048 * it will be updated on the arrival of a new request.
4050 next_rq = bfqq->next_rq;
4052 bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget,
4053 bfq_serv_to_charge(next_rq, bfqq));
4055 bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %d",
4056 next_rq ? blk_rq_sectors(next_rq) : 0,
4057 bfqq->entity.budget);
4061 * Return true if the process associated with bfqq is "slow". The slow
4062 * flag is used, in addition to the budget timeout, to reduce the
4063 * amount of service provided to seeky processes, and thus reduce
4064 * their chances to lower the throughput. More details in the comments
4065 * on the function bfq_bfqq_expire().
4067 * An important observation is in order: as discussed in the comments
4068 * on the function bfq_update_peak_rate(), with devices with internal
4069 * queues, it is hard if ever possible to know when and for how long
4070 * an I/O request is processed by the device (apart from the trivial
4071 * I/O pattern where a new request is dispatched only after the
4072 * previous one has been completed). This makes it hard to evaluate
4073 * the real rate at which the I/O requests of each bfq_queue are
4074 * served. In fact, for an I/O scheduler like BFQ, serving a
4075 * bfq_queue means just dispatching its requests during its service
4076 * slot (i.e., until the budget of the queue is exhausted, or the
4077 * queue remains idle, or, finally, a timeout fires). But, during the
4078 * service slot of a bfq_queue, around 100 ms at most, the device may
4079 * be even still processing requests of bfq_queues served in previous
4080 * service slots. On the opposite end, the requests of the in-service
4081 * bfq_queue may be completed after the service slot of the queue
4084 * Anyway, unless more sophisticated solutions are used
4085 * (where possible), the sum of the sizes of the requests dispatched
4086 * during the service slot of a bfq_queue is probably the only
4087 * approximation available for the service received by the bfq_queue
4088 * during its service slot. And this sum is the quantity used in this
4089 * function to evaluate the I/O speed of a process.
4091 static bool bfq_bfqq_is_slow(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4092 bool compensate, enum bfqq_expiration reason,
4093 unsigned long *delta_ms)
4095 ktime_t delta_ktime;
4097 bool slow = BFQQ_SEEKY(bfqq); /* if delta too short, use seekyness */
4099 if (!bfq_bfqq_sync(bfqq))
4103 delta_ktime = bfqd->last_idling_start;
4105 delta_ktime = ktime_get();
4106 delta_ktime = ktime_sub(delta_ktime, bfqd->last_budget_start);
4107 delta_usecs = ktime_to_us(delta_ktime);
4109 /* don't use too short time intervals */
4110 if (delta_usecs < 1000) {
4111 if (blk_queue_nonrot(bfqd->queue))
4113 * give same worst-case guarantees as idling
4116 *delta_ms = BFQ_MIN_TT / NSEC_PER_MSEC;
4117 else /* charge at least one seek */
4118 *delta_ms = bfq_slice_idle / NSEC_PER_MSEC;
4123 *delta_ms = delta_usecs / USEC_PER_MSEC;
4126 * Use only long (> 20ms) intervals to filter out excessive
4127 * spikes in service rate estimation.
4129 if (delta_usecs > 20000) {
4131 * Caveat for rotational devices: processes doing I/O
4132 * in the slower disk zones tend to be slow(er) even
4133 * if not seeky. In this respect, the estimated peak
4134 * rate is likely to be an average over the disk
4135 * surface. Accordingly, to not be too harsh with
4136 * unlucky processes, a process is deemed slow only if
4137 * its rate has been lower than half of the estimated
4140 slow = bfqq->entity.service < bfqd->bfq_max_budget / 2;
4143 bfq_log_bfqq(bfqd, bfqq, "bfq_bfqq_is_slow: slow %d", slow);
4149 * To be deemed as soft real-time, an application must meet two
4150 * requirements. First, the application must not require an average
4151 * bandwidth higher than the approximate bandwidth required to playback or
4152 * record a compressed high-definition video.
4153 * The next function is invoked on the completion of the last request of a
4154 * batch, to compute the next-start time instant, soft_rt_next_start, such
4155 * that, if the next request of the application does not arrive before
4156 * soft_rt_next_start, then the above requirement on the bandwidth is met.
4158 * The second requirement is that the request pattern of the application is
4159 * isochronous, i.e., that, after issuing a request or a batch of requests,
4160 * the application stops issuing new requests until all its pending requests
4161 * have been completed. After that, the application may issue a new batch,
4163 * For this reason the next function is invoked to compute
4164 * soft_rt_next_start only for applications that meet this requirement,
4165 * whereas soft_rt_next_start is set to infinity for applications that do
4168 * Unfortunately, even a greedy (i.e., I/O-bound) application may
4169 * happen to meet, occasionally or systematically, both the above
4170 * bandwidth and isochrony requirements. This may happen at least in
4171 * the following circumstances. First, if the CPU load is high. The
4172 * application may stop issuing requests while the CPUs are busy
4173 * serving other processes, then restart, then stop again for a while,
4174 * and so on. The other circumstances are related to the storage
4175 * device: the storage device is highly loaded or reaches a low-enough
4176 * throughput with the I/O of the application (e.g., because the I/O
4177 * is random and/or the device is slow). In all these cases, the
4178 * I/O of the application may be simply slowed down enough to meet
4179 * the bandwidth and isochrony requirements. To reduce the probability
4180 * that greedy applications are deemed as soft real-time in these
4181 * corner cases, a further rule is used in the computation of
4182 * soft_rt_next_start: the return value of this function is forced to
4183 * be higher than the maximum between the following two quantities.
4185 * (a) Current time plus: (1) the maximum time for which the arrival
4186 * of a request is waited for when a sync queue becomes idle,
4187 * namely bfqd->bfq_slice_idle, and (2) a few extra jiffies. We
4188 * postpone for a moment the reason for adding a few extra
4189 * jiffies; we get back to it after next item (b). Lower-bounding
4190 * the return value of this function with the current time plus
4191 * bfqd->bfq_slice_idle tends to filter out greedy applications,
4192 * because the latter issue their next request as soon as possible
4193 * after the last one has been completed. In contrast, a soft
4194 * real-time application spends some time processing data, after a
4195 * batch of its requests has been completed.
4197 * (b) Current value of bfqq->soft_rt_next_start. As pointed out
4198 * above, greedy applications may happen to meet both the
4199 * bandwidth and isochrony requirements under heavy CPU or
4200 * storage-device load. In more detail, in these scenarios, these
4201 * applications happen, only for limited time periods, to do I/O
4202 * slowly enough to meet all the requirements described so far,
4203 * including the filtering in above item (a). These slow-speed
4204 * time intervals are usually interspersed between other time
4205 * intervals during which these applications do I/O at a very high
4206 * speed. Fortunately, exactly because of the high speed of the
4207 * I/O in the high-speed intervals, the values returned by this
4208 * function happen to be so high, near the end of any such
4209 * high-speed interval, to be likely to fall *after* the end of
4210 * the low-speed time interval that follows. These high values are
4211 * stored in bfqq->soft_rt_next_start after each invocation of
4212 * this function. As a consequence, if the last value of
4213 * bfqq->soft_rt_next_start is constantly used to lower-bound the
4214 * next value that this function may return, then, from the very
4215 * beginning of a low-speed interval, bfqq->soft_rt_next_start is
4216 * likely to be constantly kept so high that any I/O request
4217 * issued during the low-speed interval is considered as arriving
4218 * to soon for the application to be deemed as soft
4219 * real-time. Then, in the high-speed interval that follows, the
4220 * application will not be deemed as soft real-time, just because
4221 * it will do I/O at a high speed. And so on.
4223 * Getting back to the filtering in item (a), in the following two
4224 * cases this filtering might be easily passed by a greedy
4225 * application, if the reference quantity was just
4226 * bfqd->bfq_slice_idle:
4227 * 1) HZ is so low that the duration of a jiffy is comparable to or
4228 * higher than bfqd->bfq_slice_idle. This happens, e.g., on slow
4229 * devices with HZ=100. The time granularity may be so coarse
4230 * that the approximation, in jiffies, of bfqd->bfq_slice_idle
4231 * is rather lower than the exact value.
4232 * 2) jiffies, instead of increasing at a constant rate, may stop increasing
4233 * for a while, then suddenly 'jump' by several units to recover the lost
4234 * increments. This seems to happen, e.g., inside virtual machines.
4235 * To address this issue, in the filtering in (a) we do not use as a
4236 * reference time interval just bfqd->bfq_slice_idle, but
4237 * bfqd->bfq_slice_idle plus a few jiffies. In particular, we add the
4238 * minimum number of jiffies for which the filter seems to be quite
4239 * precise also in embedded systems and KVM/QEMU virtual machines.
4241 static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data *bfqd,
4242 struct bfq_queue *bfqq)
4244 return max3(bfqq->soft_rt_next_start,
4245 bfqq->last_idle_bklogged +
4246 HZ * bfqq->service_from_backlogged /
4247 bfqd->bfq_wr_max_softrt_rate,
4248 jiffies + nsecs_to_jiffies(bfqq->bfqd->bfq_slice_idle) + 4);
4252 * bfq_bfqq_expire - expire a queue.
4253 * @bfqd: device owning the queue.
4254 * @bfqq: the queue to expire.
4255 * @compensate: if true, compensate for the time spent idling.
4256 * @reason: the reason causing the expiration.
4258 * If the process associated with bfqq does slow I/O (e.g., because it
4259 * issues random requests), we charge bfqq with the time it has been
4260 * in service instead of the service it has received (see
4261 * bfq_bfqq_charge_time for details on how this goal is achieved). As
4262 * a consequence, bfqq will typically get higher timestamps upon
4263 * reactivation, and hence it will be rescheduled as if it had
4264 * received more service than what it has actually received. In the
4265 * end, bfqq receives less service in proportion to how slowly its
4266 * associated process consumes its budgets (and hence how seriously it
4267 * tends to lower the throughput). In addition, this time-charging
4268 * strategy guarantees time fairness among slow processes. In
4269 * contrast, if the process associated with bfqq is not slow, we
4270 * charge bfqq exactly with the service it has received.
4272 * Charging time to the first type of queues and the exact service to
4273 * the other has the effect of using the WF2Q+ policy to schedule the
4274 * former on a timeslice basis, without violating service domain
4275 * guarantees among the latter.
4277 void bfq_bfqq_expire(struct bfq_data *bfqd,
4278 struct bfq_queue *bfqq,
4280 enum bfqq_expiration reason)
4283 unsigned long delta = 0;
4284 struct bfq_entity *entity = &bfqq->entity;
4287 * Check whether the process is slow (see bfq_bfqq_is_slow).
4289 slow = bfq_bfqq_is_slow(bfqd, bfqq, compensate, reason, &delta);
4292 * As above explained, charge slow (typically seeky) and
4293 * timed-out queues with the time and not the service
4294 * received, to favor sequential workloads.
4296 * Processes doing I/O in the slower disk zones will tend to
4297 * be slow(er) even if not seeky. Therefore, since the
4298 * estimated peak rate is actually an average over the disk
4299 * surface, these processes may timeout just for bad luck. To
4300 * avoid punishing them, do not charge time to processes that
4301 * succeeded in consuming at least 2/3 of their budget. This
4302 * allows BFQ to preserve enough elasticity to still perform
4303 * bandwidth, and not time, distribution with little unlucky
4304 * or quasi-sequential processes.
4306 if (bfqq->wr_coeff == 1 &&
4308 (reason == BFQQE_BUDGET_TIMEOUT &&
4309 bfq_bfqq_budget_left(bfqq) >= entity->budget / 3)))
4310 bfq_bfqq_charge_time(bfqd, bfqq, delta);
4312 if (bfqd->low_latency && bfqq->wr_coeff == 1)
4313 bfqq->last_wr_start_finish = jiffies;
4315 if (bfqd->low_latency && bfqd->bfq_wr_max_softrt_rate > 0 &&
4316 RB_EMPTY_ROOT(&bfqq->sort_list)) {
4318 * If we get here, and there are no outstanding
4319 * requests, then the request pattern is isochronous
4320 * (see the comments on the function
4321 * bfq_bfqq_softrt_next_start()). Therefore we can
4322 * compute soft_rt_next_start.
4324 * If, instead, the queue still has outstanding
4325 * requests, then we have to wait for the completion
4326 * of all the outstanding requests to discover whether
4327 * the request pattern is actually isochronous.
4329 if (bfqq->dispatched == 0)
4330 bfqq->soft_rt_next_start =
4331 bfq_bfqq_softrt_next_start(bfqd, bfqq);
4332 else if (bfqq->dispatched > 0) {
4334 * Schedule an update of soft_rt_next_start to when
4335 * the task may be discovered to be isochronous.
4337 bfq_mark_bfqq_softrt_update(bfqq);
4341 bfq_log_bfqq(bfqd, bfqq,
4342 "expire (%d, slow %d, num_disp %d, short_ttime %d)", reason,
4343 slow, bfqq->dispatched, bfq_bfqq_has_short_ttime(bfqq));
4346 * bfqq expired, so no total service time needs to be computed
4347 * any longer: reset state machine for measuring total service
4350 bfqd->rqs_injected = bfqd->wait_dispatch = false;
4351 bfqd->waited_rq = NULL;
4354 * Increase, decrease or leave budget unchanged according to
4357 __bfq_bfqq_recalc_budget(bfqd, bfqq, reason);
4358 if (__bfq_bfqq_expire(bfqd, bfqq, reason))
4359 /* bfqq is gone, no more actions on it */
4362 /* mark bfqq as waiting a request only if a bic still points to it */
4363 if (!bfq_bfqq_busy(bfqq) &&
4364 reason != BFQQE_BUDGET_TIMEOUT &&
4365 reason != BFQQE_BUDGET_EXHAUSTED) {
4366 bfq_mark_bfqq_non_blocking_wait_rq(bfqq);
4368 * Not setting service to 0, because, if the next rq
4369 * arrives in time, the queue will go on receiving
4370 * service with this same budget (as if it never expired)
4373 entity->service = 0;
4376 * Reset the received-service counter for every parent entity.
4377 * Differently from what happens with bfqq->entity.service,
4378 * the resetting of this counter never needs to be postponed
4379 * for parent entities. In fact, in case bfqq may have a
4380 * chance to go on being served using the last, partially
4381 * consumed budget, bfqq->entity.service needs to be kept,
4382 * because if bfqq then actually goes on being served using
4383 * the same budget, the last value of bfqq->entity.service is
4384 * needed to properly decrement bfqq->entity.budget by the
4385 * portion already consumed. In contrast, it is not necessary
4386 * to keep entity->service for parent entities too, because
4387 * the bubble up of the new value of bfqq->entity.budget will
4388 * make sure that the budgets of parent entities are correct,
4389 * even in case bfqq and thus parent entities go on receiving
4390 * service with the same budget.
4392 entity = entity->parent;
4393 for_each_entity(entity)
4394 entity->service = 0;
4398 * Budget timeout is not implemented through a dedicated timer, but
4399 * just checked on request arrivals and completions, as well as on
4400 * idle timer expirations.
4402 static bool bfq_bfqq_budget_timeout(struct bfq_queue *bfqq)
4404 return time_is_before_eq_jiffies(bfqq->budget_timeout);
4408 * If we expire a queue that is actively waiting (i.e., with the
4409 * device idled) for the arrival of a new request, then we may incur
4410 * the timestamp misalignment problem described in the body of the
4411 * function __bfq_activate_entity. Hence we return true only if this
4412 * condition does not hold, or if the queue is slow enough to deserve
4413 * only to be kicked off for preserving a high throughput.
4415 static bool bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq)
4417 bfq_log_bfqq(bfqq->bfqd, bfqq,
4418 "may_budget_timeout: wait_request %d left %d timeout %d",
4419 bfq_bfqq_wait_request(bfqq),
4420 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3,
4421 bfq_bfqq_budget_timeout(bfqq));
4423 return (!bfq_bfqq_wait_request(bfqq) ||
4424 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3)
4426 bfq_bfqq_budget_timeout(bfqq);
4429 static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd,
4430 struct bfq_queue *bfqq)
4432 bool rot_without_queueing =
4433 !blk_queue_nonrot(bfqd->queue) && !bfqd->hw_tag,
4434 bfqq_sequential_and_IO_bound,
4437 /* No point in idling for bfqq if it won't get requests any longer */
4438 if (unlikely(!bfqq_process_refs(bfqq)))
4441 bfqq_sequential_and_IO_bound = !BFQQ_SEEKY(bfqq) &&
4442 bfq_bfqq_IO_bound(bfqq) && bfq_bfqq_has_short_ttime(bfqq);
4445 * The next variable takes into account the cases where idling
4446 * boosts the throughput.
4448 * The value of the variable is computed considering, first, that
4449 * idling is virtually always beneficial for the throughput if:
4450 * (a) the device is not NCQ-capable and rotational, or
4451 * (b) regardless of the presence of NCQ, the device is rotational and
4452 * the request pattern for bfqq is I/O-bound and sequential, or
4453 * (c) regardless of whether it is rotational, the device is
4454 * not NCQ-capable and the request pattern for bfqq is
4455 * I/O-bound and sequential.
4457 * Secondly, and in contrast to the above item (b), idling an
4458 * NCQ-capable flash-based device would not boost the
4459 * throughput even with sequential I/O; rather it would lower
4460 * the throughput in proportion to how fast the device
4461 * is. Accordingly, the next variable is true if any of the
4462 * above conditions (a), (b) or (c) is true, and, in
4463 * particular, happens to be false if bfqd is an NCQ-capable
4464 * flash-based device.
4466 idling_boosts_thr = rot_without_queueing ||
4467 ((!blk_queue_nonrot(bfqd->queue) || !bfqd->hw_tag) &&
4468 bfqq_sequential_and_IO_bound);
4471 * The return value of this function is equal to that of
4472 * idling_boosts_thr, unless a special case holds. In this
4473 * special case, described below, idling may cause problems to
4474 * weight-raised queues.
4476 * When the request pool is saturated (e.g., in the presence
4477 * of write hogs), if the processes associated with
4478 * non-weight-raised queues ask for requests at a lower rate,
4479 * then processes associated with weight-raised queues have a
4480 * higher probability to get a request from the pool
4481 * immediately (or at least soon) when they need one. Thus
4482 * they have a higher probability to actually get a fraction
4483 * of the device throughput proportional to their high
4484 * weight. This is especially true with NCQ-capable drives,
4485 * which enqueue several requests in advance, and further
4486 * reorder internally-queued requests.
4488 * For this reason, we force to false the return value if
4489 * there are weight-raised busy queues. In this case, and if
4490 * bfqq is not weight-raised, this guarantees that the device
4491 * is not idled for bfqq (if, instead, bfqq is weight-raised,
4492 * then idling will be guaranteed by another variable, see
4493 * below). Combined with the timestamping rules of BFQ (see
4494 * [1] for details), this behavior causes bfqq, and hence any
4495 * sync non-weight-raised queue, to get a lower number of
4496 * requests served, and thus to ask for a lower number of
4497 * requests from the request pool, before the busy
4498 * weight-raised queues get served again. This often mitigates
4499 * starvation problems in the presence of heavy write
4500 * workloads and NCQ, thereby guaranteeing a higher
4501 * application and system responsiveness in these hostile
4504 return idling_boosts_thr &&
4505 bfqd->wr_busy_queues == 0;
4509 * For a queue that becomes empty, device idling is allowed only if
4510 * this function returns true for that queue. As a consequence, since
4511 * device idling plays a critical role for both throughput boosting
4512 * and service guarantees, the return value of this function plays a
4513 * critical role as well.
4515 * In a nutshell, this function returns true only if idling is
4516 * beneficial for throughput or, even if detrimental for throughput,
4517 * idling is however necessary to preserve service guarantees (low
4518 * latency, desired throughput distribution, ...). In particular, on
4519 * NCQ-capable devices, this function tries to return false, so as to
4520 * help keep the drives' internal queues full, whenever this helps the
4521 * device boost the throughput without causing any service-guarantee
4524 * Most of the issues taken into account to get the return value of
4525 * this function are not trivial. We discuss these issues in the two
4526 * functions providing the main pieces of information needed by this
4529 static bool bfq_better_to_idle(struct bfq_queue *bfqq)
4531 struct bfq_data *bfqd = bfqq->bfqd;
4532 bool idling_boosts_thr_with_no_issue, idling_needed_for_service_guar;
4534 /* No point in idling for bfqq if it won't get requests any longer */
4535 if (unlikely(!bfqq_process_refs(bfqq)))
4538 if (unlikely(bfqd->strict_guarantees))
4542 * Idling is performed only if slice_idle > 0. In addition, we
4545 * (b) bfqq is in the idle io prio class: in this case we do
4546 * not idle because we want to minimize the bandwidth that
4547 * queues in this class can steal to higher-priority queues
4549 if (bfqd->bfq_slice_idle == 0 || !bfq_bfqq_sync(bfqq) ||
4550 bfq_class_idle(bfqq))
4553 idling_boosts_thr_with_no_issue =
4554 idling_boosts_thr_without_issues(bfqd, bfqq);
4556 idling_needed_for_service_guar =
4557 idling_needed_for_service_guarantees(bfqd, bfqq);
4560 * We have now the two components we need to compute the
4561 * return value of the function, which is true only if idling
4562 * either boosts the throughput (without issues), or is
4563 * necessary to preserve service guarantees.
4565 return idling_boosts_thr_with_no_issue ||
4566 idling_needed_for_service_guar;
4570 * If the in-service queue is empty but the function bfq_better_to_idle
4571 * returns true, then:
4572 * 1) the queue must remain in service and cannot be expired, and
4573 * 2) the device must be idled to wait for the possible arrival of a new
4574 * request for the queue.
4575 * See the comments on the function bfq_better_to_idle for the reasons
4576 * why performing device idling is the best choice to boost the throughput
4577 * and preserve service guarantees when bfq_better_to_idle itself
4580 static bool bfq_bfqq_must_idle(struct bfq_queue *bfqq)
4582 return RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_better_to_idle(bfqq);
4586 * This function chooses the queue from which to pick the next extra
4587 * I/O request to inject, if it finds a compatible queue. See the
4588 * comments on bfq_update_inject_limit() for details on the injection
4589 * mechanism, and for the definitions of the quantities mentioned
4592 static struct bfq_queue *
4593 bfq_choose_bfqq_for_injection(struct bfq_data *bfqd)
4595 struct bfq_queue *bfqq, *in_serv_bfqq = bfqd->in_service_queue;
4596 unsigned int limit = in_serv_bfqq->inject_limit;
4599 * - bfqq is not weight-raised and therefore does not carry
4600 * time-critical I/O,
4602 * - regardless of whether bfqq is weight-raised, bfqq has
4603 * however a long think time, during which it can absorb the
4604 * effect of an appropriate number of extra I/O requests
4605 * from other queues (see bfq_update_inject_limit for
4606 * details on the computation of this number);
4607 * then injection can be performed without restrictions.
4609 bool in_serv_always_inject = in_serv_bfqq->wr_coeff == 1 ||
4610 !bfq_bfqq_has_short_ttime(in_serv_bfqq);
4614 * - the baseline total service time could not be sampled yet,
4615 * so the inject limit happens to be still 0, and
4616 * - a lot of time has elapsed since the plugging of I/O
4617 * dispatching started, so drive speed is being wasted
4619 * then temporarily raise inject limit to one request.
4621 if (limit == 0 && in_serv_bfqq->last_serv_time_ns == 0 &&
4622 bfq_bfqq_wait_request(in_serv_bfqq) &&
4623 time_is_before_eq_jiffies(bfqd->last_idling_start_jiffies +
4624 bfqd->bfq_slice_idle)
4628 if (bfqd->rq_in_driver >= limit)
4632 * Linear search of the source queue for injection; but, with
4633 * a high probability, very few steps are needed to find a
4634 * candidate queue, i.e., a queue with enough budget left for
4635 * its next request. In fact:
4636 * - BFQ dynamically updates the budget of every queue so as
4637 * to accommodate the expected backlog of the queue;
4638 * - if a queue gets all its requests dispatched as injected
4639 * service, then the queue is removed from the active list
4640 * (and re-added only if it gets new requests, but then it
4641 * is assigned again enough budget for its new backlog).
4643 list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
4644 if (!RB_EMPTY_ROOT(&bfqq->sort_list) &&
4645 (in_serv_always_inject || bfqq->wr_coeff > 1) &&
4646 bfq_serv_to_charge(bfqq->next_rq, bfqq) <=
4647 bfq_bfqq_budget_left(bfqq)) {
4649 * Allow for only one large in-flight request
4650 * on non-rotational devices, for the
4651 * following reason. On non-rotationl drives,
4652 * large requests take much longer than
4653 * smaller requests to be served. In addition,
4654 * the drive prefers to serve large requests
4655 * w.r.t. to small ones, if it can choose. So,
4656 * having more than one large requests queued
4657 * in the drive may easily make the next first
4658 * request of the in-service queue wait for so
4659 * long to break bfqq's service guarantees. On
4660 * the bright side, large requests let the
4661 * drive reach a very high throughput, even if
4662 * there is only one in-flight large request
4665 if (blk_queue_nonrot(bfqd->queue) &&
4666 blk_rq_sectors(bfqq->next_rq) >=
4667 BFQQ_SECT_THR_NONROT)
4668 limit = min_t(unsigned int, 1, limit);
4670 limit = in_serv_bfqq->inject_limit;
4672 if (bfqd->rq_in_driver < limit) {
4673 bfqd->rqs_injected = true;
4682 * Select a queue for service. If we have a current queue in service,
4683 * check whether to continue servicing it, or retrieve and set a new one.
4685 static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd)
4687 struct bfq_queue *bfqq;
4688 struct request *next_rq;
4689 enum bfqq_expiration reason = BFQQE_BUDGET_TIMEOUT;
4691 bfqq = bfqd->in_service_queue;
4695 bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue");
4698 * Do not expire bfqq for budget timeout if bfqq may be about
4699 * to enjoy device idling. The reason why, in this case, we
4700 * prevent bfqq from expiring is the same as in the comments
4701 * on the case where bfq_bfqq_must_idle() returns true, in
4702 * bfq_completed_request().
4704 if (bfq_may_expire_for_budg_timeout(bfqq) &&
4705 !bfq_bfqq_must_idle(bfqq))
4710 * This loop is rarely executed more than once. Even when it
4711 * happens, it is much more convenient to re-execute this loop
4712 * than to return NULL and trigger a new dispatch to get a
4715 next_rq = bfqq->next_rq;
4717 * If bfqq has requests queued and it has enough budget left to
4718 * serve them, keep the queue, otherwise expire it.
4721 if (bfq_serv_to_charge(next_rq, bfqq) >
4722 bfq_bfqq_budget_left(bfqq)) {
4724 * Expire the queue for budget exhaustion,
4725 * which makes sure that the next budget is
4726 * enough to serve the next request, even if
4727 * it comes from the fifo expired path.
4729 reason = BFQQE_BUDGET_EXHAUSTED;
4733 * The idle timer may be pending because we may
4734 * not disable disk idling even when a new request
4737 if (bfq_bfqq_wait_request(bfqq)) {
4739 * If we get here: 1) at least a new request
4740 * has arrived but we have not disabled the
4741 * timer because the request was too small,
4742 * 2) then the block layer has unplugged
4743 * the device, causing the dispatch to be
4746 * Since the device is unplugged, now the
4747 * requests are probably large enough to
4748 * provide a reasonable throughput.
4749 * So we disable idling.
4751 bfq_clear_bfqq_wait_request(bfqq);
4752 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
4759 * No requests pending. However, if the in-service queue is idling
4760 * for a new request, or has requests waiting for a completion and
4761 * may idle after their completion, then keep it anyway.
4763 * Yet, inject service from other queues if it boosts
4764 * throughput and is possible.
4766 if (bfq_bfqq_wait_request(bfqq) ||
4767 (bfqq->dispatched != 0 && bfq_better_to_idle(bfqq))) {
4768 struct bfq_queue *async_bfqq =
4769 bfqq->bic && bfqq->bic->bfqq[0] &&
4770 bfq_bfqq_busy(bfqq->bic->bfqq[0]) &&
4771 bfqq->bic->bfqq[0]->next_rq ?
4772 bfqq->bic->bfqq[0] : NULL;
4773 struct bfq_queue *blocked_bfqq =
4774 !hlist_empty(&bfqq->woken_list) ?
4775 container_of(bfqq->woken_list.first,
4781 * The next four mutually-exclusive ifs decide
4782 * whether to try injection, and choose the queue to
4783 * pick an I/O request from.
4785 * The first if checks whether the process associated
4786 * with bfqq has also async I/O pending. If so, it
4787 * injects such I/O unconditionally. Injecting async
4788 * I/O from the same process can cause no harm to the
4789 * process. On the contrary, it can only increase
4790 * bandwidth and reduce latency for the process.
4792 * The second if checks whether there happens to be a
4793 * non-empty waker queue for bfqq, i.e., a queue whose
4794 * I/O needs to be completed for bfqq to receive new
4795 * I/O. This happens, e.g., if bfqq is associated with
4796 * a process that does some sync. A sync generates
4797 * extra blocking I/O, which must be completed before
4798 * the process associated with bfqq can go on with its
4799 * I/O. If the I/O of the waker queue is not served,
4800 * then bfqq remains empty, and no I/O is dispatched,
4801 * until the idle timeout fires for bfqq. This is
4802 * likely to result in lower bandwidth and higher
4803 * latencies for bfqq, and in a severe loss of total
4804 * throughput. The best action to take is therefore to
4805 * serve the waker queue as soon as possible. So do it
4806 * (without relying on the third alternative below for
4807 * eventually serving waker_bfqq's I/O; see the last
4808 * paragraph for further details). This systematic
4809 * injection of I/O from the waker queue does not
4810 * cause any delay to bfqq's I/O. On the contrary,
4811 * next bfqq's I/O is brought forward dramatically,
4812 * for it is not blocked for milliseconds.
4814 * The third if checks whether there is a queue woken
4815 * by bfqq, and currently with pending I/O. Such a
4816 * woken queue does not steal bandwidth from bfqq,
4817 * because it remains soon without I/O if bfqq is not
4818 * served. So there is virtually no risk of loss of
4819 * bandwidth for bfqq if this woken queue has I/O
4820 * dispatched while bfqq is waiting for new I/O.
4822 * The fourth if checks whether bfqq is a queue for
4823 * which it is better to avoid injection. It is so if
4824 * bfqq delivers more throughput when served without
4825 * any further I/O from other queues in the middle, or
4826 * if the service times of bfqq's I/O requests both
4827 * count more than overall throughput, and may be
4828 * easily increased by injection (this happens if bfqq
4829 * has a short think time). If none of these
4830 * conditions holds, then a candidate queue for
4831 * injection is looked for through
4832 * bfq_choose_bfqq_for_injection(). Note that the
4833 * latter may return NULL (for example if the inject
4834 * limit for bfqq is currently 0).
4836 * NOTE: motivation for the second alternative
4838 * Thanks to the way the inject limit is updated in
4839 * bfq_update_has_short_ttime(), it is rather likely
4840 * that, if I/O is being plugged for bfqq and the
4841 * waker queue has pending I/O requests that are
4842 * blocking bfqq's I/O, then the fourth alternative
4843 * above lets the waker queue get served before the
4844 * I/O-plugging timeout fires. So one may deem the
4845 * second alternative superfluous. It is not, because
4846 * the fourth alternative may be way less effective in
4847 * case of a synchronization. For two main
4848 * reasons. First, throughput may be low because the
4849 * inject limit may be too low to guarantee the same
4850 * amount of injected I/O, from the waker queue or
4851 * other queues, that the second alternative
4852 * guarantees (the second alternative unconditionally
4853 * injects a pending I/O request of the waker queue
4854 * for each bfq_dispatch_request()). Second, with the
4855 * fourth alternative, the duration of the plugging,
4856 * i.e., the time before bfqq finally receives new I/O,
4857 * may not be minimized, because the waker queue may
4858 * happen to be served only after other queues.
4861 icq_to_bic(async_bfqq->next_rq->elv.icq) == bfqq->bic &&
4862 bfq_serv_to_charge(async_bfqq->next_rq, async_bfqq) <=
4863 bfq_bfqq_budget_left(async_bfqq))
4864 bfqq = bfqq->bic->bfqq[0];
4865 else if (bfqq->waker_bfqq &&
4866 bfq_bfqq_busy(bfqq->waker_bfqq) &&
4867 bfqq->waker_bfqq->next_rq &&
4868 bfq_serv_to_charge(bfqq->waker_bfqq->next_rq,
4869 bfqq->waker_bfqq) <=
4870 bfq_bfqq_budget_left(bfqq->waker_bfqq)
4872 bfqq = bfqq->waker_bfqq;
4873 else if (blocked_bfqq &&
4874 bfq_bfqq_busy(blocked_bfqq) &&
4875 blocked_bfqq->next_rq &&
4876 bfq_serv_to_charge(blocked_bfqq->next_rq,
4878 bfq_bfqq_budget_left(blocked_bfqq)
4880 bfqq = blocked_bfqq;
4881 else if (!idling_boosts_thr_without_issues(bfqd, bfqq) &&
4882 (bfqq->wr_coeff == 1 || bfqd->wr_busy_queues > 1 ||
4883 !bfq_bfqq_has_short_ttime(bfqq)))
4884 bfqq = bfq_choose_bfqq_for_injection(bfqd);
4891 reason = BFQQE_NO_MORE_REQUESTS;
4893 bfq_bfqq_expire(bfqd, bfqq, false, reason);
4895 bfqq = bfq_set_in_service_queue(bfqd);
4897 bfq_log_bfqq(bfqd, bfqq, "select_queue: checking new queue");
4902 bfq_log_bfqq(bfqd, bfqq, "select_queue: returned this queue");
4904 bfq_log(bfqd, "select_queue: no queue returned");
4909 static void bfq_update_wr_data(struct bfq_data *bfqd, struct bfq_queue *bfqq)
4911 struct bfq_entity *entity = &bfqq->entity;
4913 if (bfqq->wr_coeff > 1) { /* queue is being weight-raised */
4914 bfq_log_bfqq(bfqd, bfqq,
4915 "raising period dur %u/%u msec, old coeff %u, w %d(%d)",
4916 jiffies_to_msecs(jiffies - bfqq->last_wr_start_finish),
4917 jiffies_to_msecs(bfqq->wr_cur_max_time),
4919 bfqq->entity.weight, bfqq->entity.orig_weight);
4921 if (entity->prio_changed)
4922 bfq_log_bfqq(bfqd, bfqq, "WARN: pending prio change");
4925 * If the queue was activated in a burst, or too much
4926 * time has elapsed from the beginning of this
4927 * weight-raising period, then end weight raising.
4929 if (bfq_bfqq_in_large_burst(bfqq))
4930 bfq_bfqq_end_wr(bfqq);
4931 else if (time_is_before_jiffies(bfqq->last_wr_start_finish +
4932 bfqq->wr_cur_max_time)) {
4933 if (bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time ||
4934 time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
4935 bfq_wr_duration(bfqd))) {
4937 * Either in interactive weight
4938 * raising, or in soft_rt weight
4940 * interactive-weight-raising period
4941 * elapsed (so no switch back to
4942 * interactive weight raising).
4944 bfq_bfqq_end_wr(bfqq);
4946 * soft_rt finishing while still in
4947 * interactive period, switch back to
4948 * interactive weight raising
4950 switch_back_to_interactive_wr(bfqq, bfqd);
4951 bfqq->entity.prio_changed = 1;
4954 if (bfqq->wr_coeff > 1 &&
4955 bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time &&
4956 bfqq->service_from_wr > max_service_from_wr) {
4957 /* see comments on max_service_from_wr */
4958 bfq_bfqq_end_wr(bfqq);
4962 * To improve latency (for this or other queues), immediately
4963 * update weight both if it must be raised and if it must be
4964 * lowered. Since, entity may be on some active tree here, and
4965 * might have a pending change of its ioprio class, invoke
4966 * next function with the last parameter unset (see the
4967 * comments on the function).
4969 if ((entity->weight > entity->orig_weight) != (bfqq->wr_coeff > 1))
4970 __bfq_entity_update_weight_prio(bfq_entity_service_tree(entity),
4975 * Dispatch next request from bfqq.
4977 static struct request *bfq_dispatch_rq_from_bfqq(struct bfq_data *bfqd,
4978 struct bfq_queue *bfqq)
4980 struct request *rq = bfqq->next_rq;
4981 unsigned long service_to_charge;
4983 service_to_charge = bfq_serv_to_charge(rq, bfqq);
4985 bfq_bfqq_served(bfqq, service_to_charge);
4987 if (bfqq == bfqd->in_service_queue && bfqd->wait_dispatch) {
4988 bfqd->wait_dispatch = false;
4989 bfqd->waited_rq = rq;
4992 bfq_dispatch_remove(bfqd->queue, rq);
4994 if (bfqq != bfqd->in_service_queue)
4998 * If weight raising has to terminate for bfqq, then next
4999 * function causes an immediate update of bfqq's weight,
5000 * without waiting for next activation. As a consequence, on
5001 * expiration, bfqq will be timestamped as if has never been
5002 * weight-raised during this service slot, even if it has
5003 * received part or even most of the service as a
5004 * weight-raised queue. This inflates bfqq's timestamps, which
5005 * is beneficial, as bfqq is then more willing to leave the
5006 * device immediately to possible other weight-raised queues.
5008 bfq_update_wr_data(bfqd, bfqq);
5011 * Expire bfqq, pretending that its budget expired, if bfqq
5012 * belongs to CLASS_IDLE and other queues are waiting for
5015 if (!(bfq_tot_busy_queues(bfqd) > 1 && bfq_class_idle(bfqq)))
5018 bfq_bfqq_expire(bfqd, bfqq, false, BFQQE_BUDGET_EXHAUSTED);
5024 static bool bfq_has_work(struct blk_mq_hw_ctx *hctx)
5026 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5029 * Avoiding lock: a race on bfqd->queued should cause at
5030 * most a call to dispatch for nothing
5032 return !list_empty_careful(&bfqd->dispatch) ||
5033 READ_ONCE(bfqd->queued);
5036 static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
5038 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5039 struct request *rq = NULL;
5040 struct bfq_queue *bfqq = NULL;
5042 if (!list_empty(&bfqd->dispatch)) {
5043 rq = list_first_entry(&bfqd->dispatch, struct request,
5045 list_del_init(&rq->queuelist);
5051 * Increment counters here, because this
5052 * dispatch does not follow the standard
5053 * dispatch flow (where counters are
5058 goto inc_in_driver_start_rq;
5062 * We exploit the bfq_finish_requeue_request hook to
5063 * decrement rq_in_driver, but
5064 * bfq_finish_requeue_request will not be invoked on
5065 * this request. So, to avoid unbalance, just start
5066 * this request, without incrementing rq_in_driver. As
5067 * a negative consequence, rq_in_driver is deceptively
5068 * lower than it should be while this request is in
5069 * service. This may cause bfq_schedule_dispatch to be
5070 * invoked uselessly.
5072 * As for implementing an exact solution, the
5073 * bfq_finish_requeue_request hook, if defined, is
5074 * probably invoked also on this request. So, by
5075 * exploiting this hook, we could 1) increment
5076 * rq_in_driver here, and 2) decrement it in
5077 * bfq_finish_requeue_request. Such a solution would
5078 * let the value of the counter be always accurate,
5079 * but it would entail using an extra interface
5080 * function. This cost seems higher than the benefit,
5081 * being the frequency of non-elevator-private
5082 * requests very low.
5087 bfq_log(bfqd, "dispatch requests: %d busy queues",
5088 bfq_tot_busy_queues(bfqd));
5090 if (bfq_tot_busy_queues(bfqd) == 0)
5094 * Force device to serve one request at a time if
5095 * strict_guarantees is true. Forcing this service scheme is
5096 * currently the ONLY way to guarantee that the request
5097 * service order enforced by the scheduler is respected by a
5098 * queueing device. Otherwise the device is free even to make
5099 * some unlucky request wait for as long as the device
5102 * Of course, serving one request at a time may cause loss of
5105 if (bfqd->strict_guarantees && bfqd->rq_in_driver > 0)
5108 bfqq = bfq_select_queue(bfqd);
5112 rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq);
5115 inc_in_driver_start_rq:
5116 bfqd->rq_in_driver++;
5118 rq->rq_flags |= RQF_STARTED;
5124 #ifdef CONFIG_BFQ_CGROUP_DEBUG
5125 static void bfq_update_dispatch_stats(struct request_queue *q,
5127 struct bfq_queue *in_serv_queue,
5128 bool idle_timer_disabled)
5130 struct bfq_queue *bfqq = rq ? RQ_BFQQ(rq) : NULL;
5132 if (!idle_timer_disabled && !bfqq)
5136 * rq and bfqq are guaranteed to exist until this function
5137 * ends, for the following reasons. First, rq can be
5138 * dispatched to the device, and then can be completed and
5139 * freed, only after this function ends. Second, rq cannot be
5140 * merged (and thus freed because of a merge) any longer,
5141 * because it has already started. Thus rq cannot be freed
5142 * before this function ends, and, since rq has a reference to
5143 * bfqq, the same guarantee holds for bfqq too.
5145 * In addition, the following queue lock guarantees that
5146 * bfqq_group(bfqq) exists as well.
5148 spin_lock_irq(&q->queue_lock);
5149 if (idle_timer_disabled)
5151 * Since the idle timer has been disabled,
5152 * in_serv_queue contained some request when
5153 * __bfq_dispatch_request was invoked above, which
5154 * implies that rq was picked exactly from
5155 * in_serv_queue. Thus in_serv_queue == bfqq, and is
5156 * therefore guaranteed to exist because of the above
5159 bfqg_stats_update_idle_time(bfqq_group(in_serv_queue));
5161 struct bfq_group *bfqg = bfqq_group(bfqq);
5163 bfqg_stats_update_avg_queue_size(bfqg);
5164 bfqg_stats_set_start_empty_time(bfqg);
5165 bfqg_stats_update_io_remove(bfqg, rq->cmd_flags);
5167 spin_unlock_irq(&q->queue_lock);
5170 static inline void bfq_update_dispatch_stats(struct request_queue *q,
5172 struct bfq_queue *in_serv_queue,
5173 bool idle_timer_disabled) {}
5174 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
5176 static struct request *bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
5178 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5180 struct bfq_queue *in_serv_queue;
5181 bool waiting_rq, idle_timer_disabled = false;
5183 spin_lock_irq(&bfqd->lock);
5185 in_serv_queue = bfqd->in_service_queue;
5186 waiting_rq = in_serv_queue && bfq_bfqq_wait_request(in_serv_queue);
5188 rq = __bfq_dispatch_request(hctx);
5189 if (in_serv_queue == bfqd->in_service_queue) {
5190 idle_timer_disabled =
5191 waiting_rq && !bfq_bfqq_wait_request(in_serv_queue);
5194 spin_unlock_irq(&bfqd->lock);
5195 bfq_update_dispatch_stats(hctx->queue, rq,
5196 idle_timer_disabled ? in_serv_queue : NULL,
5197 idle_timer_disabled);
5203 * Task holds one reference to the queue, dropped when task exits. Each rq
5204 * in-flight on this queue also holds a reference, dropped when rq is freed.
5206 * Scheduler lock must be held here. Recall not to use bfqq after calling
5207 * this function on it.
5209 void bfq_put_queue(struct bfq_queue *bfqq)
5211 struct bfq_queue *item;
5212 struct hlist_node *n;
5213 struct bfq_group *bfqg = bfqq_group(bfqq);
5215 bfq_log_bfqq(bfqq->bfqd, bfqq, "put_queue: %p %d", bfqq, bfqq->ref);
5221 if (!hlist_unhashed(&bfqq->burst_list_node)) {
5222 hlist_del_init(&bfqq->burst_list_node);
5224 * Decrement also burst size after the removal, if the
5225 * process associated with bfqq is exiting, and thus
5226 * does not contribute to the burst any longer. This
5227 * decrement helps filter out false positives of large
5228 * bursts, when some short-lived process (often due to
5229 * the execution of commands by some service) happens
5230 * to start and exit while a complex application is
5231 * starting, and thus spawning several processes that
5232 * do I/O (and that *must not* be treated as a large
5233 * burst, see comments on bfq_handle_burst).
5235 * In particular, the decrement is performed only if:
5236 * 1) bfqq is not a merged queue, because, if it is,
5237 * then this free of bfqq is not triggered by the exit
5238 * of the process bfqq is associated with, but exactly
5239 * by the fact that bfqq has just been merged.
5240 * 2) burst_size is greater than 0, to handle
5241 * unbalanced decrements. Unbalanced decrements may
5242 * happen in te following case: bfqq is inserted into
5243 * the current burst list--without incrementing
5244 * bust_size--because of a split, but the current
5245 * burst list is not the burst list bfqq belonged to
5246 * (see comments on the case of a split in
5249 if (bfqq->bic && bfqq->bfqd->burst_size > 0)
5250 bfqq->bfqd->burst_size--;
5254 * bfqq does not exist any longer, so it cannot be woken by
5255 * any other queue, and cannot wake any other queue. Then bfqq
5256 * must be removed from the woken list of its possible waker
5257 * queue, and all queues in the woken list of bfqq must stop
5258 * having a waker queue. Strictly speaking, these updates
5259 * should be performed when bfqq remains with no I/O source
5260 * attached to it, which happens before bfqq gets freed. In
5261 * particular, this happens when the last process associated
5262 * with bfqq exits or gets associated with a different
5263 * queue. However, both events lead to bfqq being freed soon,
5264 * and dangling references would come out only after bfqq gets
5265 * freed. So these updates are done here, as a simple and safe
5266 * way to handle all cases.
5268 /* remove bfqq from woken list */
5269 if (!hlist_unhashed(&bfqq->woken_list_node))
5270 hlist_del_init(&bfqq->woken_list_node);
5272 /* reset waker for all queues in woken list */
5273 hlist_for_each_entry_safe(item, n, &bfqq->woken_list,
5275 item->waker_bfqq = NULL;
5276 hlist_del_init(&item->woken_list_node);
5279 if (bfqq->bfqd->last_completed_rq_bfqq == bfqq)
5280 bfqq->bfqd->last_completed_rq_bfqq = NULL;
5282 kmem_cache_free(bfq_pool, bfqq);
5283 bfqg_and_blkg_put(bfqg);
5286 static void bfq_put_stable_ref(struct bfq_queue *bfqq)
5289 bfq_put_queue(bfqq);
5292 void bfq_put_cooperator(struct bfq_queue *bfqq)
5294 struct bfq_queue *__bfqq, *next;
5297 * If this queue was scheduled to merge with another queue, be
5298 * sure to drop the reference taken on that queue (and others in
5299 * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs.
5301 __bfqq = bfqq->new_bfqq;
5305 next = __bfqq->new_bfqq;
5306 bfq_put_queue(__bfqq);
5311 static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq)
5313 if (bfqq == bfqd->in_service_queue) {
5314 __bfq_bfqq_expire(bfqd, bfqq, BFQQE_BUDGET_TIMEOUT);
5315 bfq_schedule_dispatch(bfqd);
5318 bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, bfqq->ref);
5320 bfq_put_cooperator(bfqq);
5322 bfq_release_process_ref(bfqd, bfqq);
5325 static void bfq_exit_icq_bfqq(struct bfq_io_cq *bic, bool is_sync)
5327 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
5328 struct bfq_data *bfqd;
5331 bfqd = bfqq->bfqd; /* NULL if scheduler already exited */
5334 unsigned long flags;
5336 spin_lock_irqsave(&bfqd->lock, flags);
5338 bfq_exit_bfqq(bfqd, bfqq);
5339 bic_set_bfqq(bic, NULL, is_sync);
5340 spin_unlock_irqrestore(&bfqd->lock, flags);
5344 static void bfq_exit_icq(struct io_cq *icq)
5346 struct bfq_io_cq *bic = icq_to_bic(icq);
5348 if (bic->stable_merge_bfqq) {
5349 struct bfq_data *bfqd = bic->stable_merge_bfqq->bfqd;
5352 * bfqd is NULL if scheduler already exited, and in
5353 * that case this is the last time bfqq is accessed.
5356 unsigned long flags;
5358 spin_lock_irqsave(&bfqd->lock, flags);
5359 bfq_put_stable_ref(bic->stable_merge_bfqq);
5360 spin_unlock_irqrestore(&bfqd->lock, flags);
5362 bfq_put_stable_ref(bic->stable_merge_bfqq);
5366 bfq_exit_icq_bfqq(bic, true);
5367 bfq_exit_icq_bfqq(bic, false);
5371 * Update the entity prio values; note that the new values will not
5372 * be used until the next (re)activation.
5375 bfq_set_next_ioprio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic)
5377 struct task_struct *tsk = current;
5379 struct bfq_data *bfqd = bfqq->bfqd;
5384 ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
5385 switch (ioprio_class) {
5387 pr_err("bdi %s: bfq: bad prio class %d\n",
5388 bdi_dev_name(bfqq->bfqd->queue->disk->bdi),
5391 case IOPRIO_CLASS_NONE:
5393 * No prio set, inherit CPU scheduling settings.
5395 bfqq->new_ioprio = task_nice_ioprio(tsk);
5396 bfqq->new_ioprio_class = task_nice_ioclass(tsk);
5398 case IOPRIO_CLASS_RT:
5399 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
5400 bfqq->new_ioprio_class = IOPRIO_CLASS_RT;
5402 case IOPRIO_CLASS_BE:
5403 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
5404 bfqq->new_ioprio_class = IOPRIO_CLASS_BE;
5406 case IOPRIO_CLASS_IDLE:
5407 bfqq->new_ioprio_class = IOPRIO_CLASS_IDLE;
5408 bfqq->new_ioprio = 7;
5412 if (bfqq->new_ioprio >= IOPRIO_NR_LEVELS) {
5413 pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n",
5415 bfqq->new_ioprio = IOPRIO_NR_LEVELS - 1;
5418 bfqq->entity.new_weight = bfq_ioprio_to_weight(bfqq->new_ioprio);
5419 bfq_log_bfqq(bfqd, bfqq, "new_ioprio %d new_weight %d",
5420 bfqq->new_ioprio, bfqq->entity.new_weight);
5421 bfqq->entity.prio_changed = 1;
5424 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
5425 struct bio *bio, bool is_sync,
5426 struct bfq_io_cq *bic,
5429 static void bfq_check_ioprio_change(struct bfq_io_cq *bic, struct bio *bio)
5431 struct bfq_data *bfqd = bic_to_bfqd(bic);
5432 struct bfq_queue *bfqq;
5433 int ioprio = bic->icq.ioc->ioprio;
5436 * This condition may trigger on a newly created bic, be sure to
5437 * drop the lock before returning.
5439 if (unlikely(!bfqd) || likely(bic->ioprio == ioprio))
5442 bic->ioprio = ioprio;
5444 bfqq = bic_to_bfqq(bic, false);
5446 bfq_release_process_ref(bfqd, bfqq);
5447 bfqq = bfq_get_queue(bfqd, bio, false, bic, true);
5448 bic_set_bfqq(bic, bfqq, false);
5451 bfqq = bic_to_bfqq(bic, true);
5453 bfq_set_next_ioprio_data(bfqq, bic);
5456 static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5457 struct bfq_io_cq *bic, pid_t pid, int is_sync)
5459 u64 now_ns = ktime_get_ns();
5461 RB_CLEAR_NODE(&bfqq->entity.rb_node);
5462 INIT_LIST_HEAD(&bfqq->fifo);
5463 INIT_HLIST_NODE(&bfqq->burst_list_node);
5464 INIT_HLIST_NODE(&bfqq->woken_list_node);
5465 INIT_HLIST_HEAD(&bfqq->woken_list);
5471 bfq_set_next_ioprio_data(bfqq, bic);
5475 * No need to mark as has_short_ttime if in
5476 * idle_class, because no device idling is performed
5477 * for queues in idle class
5479 if (!bfq_class_idle(bfqq))
5480 /* tentatively mark as has_short_ttime */
5481 bfq_mark_bfqq_has_short_ttime(bfqq);
5482 bfq_mark_bfqq_sync(bfqq);
5483 bfq_mark_bfqq_just_created(bfqq);
5485 bfq_clear_bfqq_sync(bfqq);
5487 /* set end request to minus infinity from now */
5488 bfqq->ttime.last_end_request = now_ns + 1;
5490 bfqq->creation_time = jiffies;
5492 bfqq->io_start_time = now_ns;
5494 bfq_mark_bfqq_IO_bound(bfqq);
5498 /* Tentative initial value to trade off between thr and lat */
5499 bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3;
5500 bfqq->budget_timeout = bfq_smallest_from_now();
5503 bfqq->last_wr_start_finish = jiffies;
5504 bfqq->wr_start_at_switch_to_srt = bfq_smallest_from_now();
5505 bfqq->split_time = bfq_smallest_from_now();
5508 * To not forget the possibly high bandwidth consumed by a
5509 * process/queue in the recent past,
5510 * bfq_bfqq_softrt_next_start() returns a value at least equal
5511 * to the current value of bfqq->soft_rt_next_start (see
5512 * comments on bfq_bfqq_softrt_next_start). Set
5513 * soft_rt_next_start to now, to mean that bfqq has consumed
5514 * no bandwidth so far.
5516 bfqq->soft_rt_next_start = jiffies;
5518 /* first request is almost certainly seeky */
5519 bfqq->seek_history = 1;
5522 static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd,
5523 struct bfq_group *bfqg,
5524 int ioprio_class, int ioprio)
5526 switch (ioprio_class) {
5527 case IOPRIO_CLASS_RT:
5528 return &bfqg->async_bfqq[0][ioprio];
5529 case IOPRIO_CLASS_NONE:
5530 ioprio = IOPRIO_BE_NORM;
5532 case IOPRIO_CLASS_BE:
5533 return &bfqg->async_bfqq[1][ioprio];
5534 case IOPRIO_CLASS_IDLE:
5535 return &bfqg->async_idle_bfqq;
5541 static struct bfq_queue *
5542 bfq_do_early_stable_merge(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5543 struct bfq_io_cq *bic,
5544 struct bfq_queue *last_bfqq_created)
5546 struct bfq_queue *new_bfqq =
5547 bfq_setup_merge(bfqq, last_bfqq_created);
5553 new_bfqq->bic->stably_merged = true;
5554 bic->stably_merged = true;
5557 * Reusing merge functions. This implies that
5558 * bfqq->bic must be set too, for
5559 * bfq_merge_bfqqs to correctly save bfqq's
5560 * state before killing it.
5563 bfq_merge_bfqqs(bfqd, bic, bfqq, new_bfqq);
5569 * Many throughput-sensitive workloads are made of several parallel
5570 * I/O flows, with all flows generated by the same application, or
5571 * more generically by the same task (e.g., system boot). The most
5572 * counterproductive action with these workloads is plugging I/O
5573 * dispatch when one of the bfq_queues associated with these flows
5574 * remains temporarily empty.
5576 * To avoid this plugging, BFQ has been using a burst-handling
5577 * mechanism for years now. This mechanism has proven effective for
5578 * throughput, and not detrimental for service guarantees. The
5579 * following function pushes this mechanism a little bit further,
5580 * basing on the following two facts.
5582 * First, all the I/O flows of a the same application or task
5583 * contribute to the execution/completion of that common application
5584 * or task. So the performance figures that matter are total
5585 * throughput of the flows and task-wide I/O latency. In particular,
5586 * these flows do not need to be protected from each other, in terms
5587 * of individual bandwidth or latency.
5589 * Second, the above fact holds regardless of the number of flows.
5591 * Putting these two facts together, this commits merges stably the
5592 * bfq_queues associated with these I/O flows, i.e., with the
5593 * processes that generate these IO/ flows, regardless of how many the
5594 * involved processes are.
5596 * To decide whether a set of bfq_queues is actually associated with
5597 * the I/O flows of a common application or task, and to merge these
5598 * queues stably, this function operates as follows: given a bfq_queue,
5599 * say Q2, currently being created, and the last bfq_queue, say Q1,
5600 * created before Q2, Q2 is merged stably with Q1 if
5601 * - very little time has elapsed since when Q1 was created
5602 * - Q2 has the same ioprio as Q1
5603 * - Q2 belongs to the same group as Q1
5605 * Merging bfq_queues also reduces scheduling overhead. A fio test
5606 * with ten random readers on /dev/nullb shows a throughput boost of
5607 * 40%, with a quadcore. Since BFQ's execution time amounts to ~50% of
5608 * the total per-request processing time, the above throughput boost
5609 * implies that BFQ's overhead is reduced by more than 50%.
5611 * This new mechanism most certainly obsoletes the current
5612 * burst-handling heuristics. We keep those heuristics for the moment.
5614 static struct bfq_queue *bfq_do_or_sched_stable_merge(struct bfq_data *bfqd,
5615 struct bfq_queue *bfqq,
5616 struct bfq_io_cq *bic)
5618 struct bfq_queue **source_bfqq = bfqq->entity.parent ?
5619 &bfqq->entity.parent->last_bfqq_created :
5620 &bfqd->last_bfqq_created;
5622 struct bfq_queue *last_bfqq_created = *source_bfqq;
5625 * If last_bfqq_created has not been set yet, then init it. If
5626 * it has been set already, but too long ago, then move it
5627 * forward to bfqq. Finally, move also if bfqq belongs to a
5628 * different group than last_bfqq_created, or if bfqq has a
5629 * different ioprio or ioprio_class. If none of these
5630 * conditions holds true, then try an early stable merge or
5631 * schedule a delayed stable merge.
5633 * A delayed merge is scheduled (instead of performing an
5634 * early merge), in case bfqq might soon prove to be more
5635 * throughput-beneficial if not merged. Currently this is
5636 * possible only if bfqd is rotational with no queueing. For
5637 * such a drive, not merging bfqq is better for throughput if
5638 * bfqq happens to contain sequential I/O. So, we wait a
5639 * little bit for enough I/O to flow through bfqq. After that,
5640 * if such an I/O is sequential, then the merge is
5641 * canceled. Otherwise the merge is finally performed.
5643 if (!last_bfqq_created ||
5644 time_before(last_bfqq_created->creation_time +
5645 msecs_to_jiffies(bfq_activation_stable_merging),
5646 bfqq->creation_time) ||
5647 bfqq->entity.parent != last_bfqq_created->entity.parent ||
5648 bfqq->ioprio != last_bfqq_created->ioprio ||
5649 bfqq->ioprio_class != last_bfqq_created->ioprio_class)
5650 *source_bfqq = bfqq;
5651 else if (time_after_eq(last_bfqq_created->creation_time +
5652 bfqd->bfq_burst_interval,
5653 bfqq->creation_time)) {
5654 if (likely(bfqd->nonrot_with_queueing))
5656 * With this type of drive, leaving
5657 * bfqq alone may provide no
5658 * throughput benefits compared with
5659 * merging bfqq. So merge bfqq now.
5661 bfqq = bfq_do_early_stable_merge(bfqd, bfqq,
5664 else { /* schedule tentative stable merge */
5666 * get reference on last_bfqq_created,
5667 * to prevent it from being freed,
5668 * until we decide whether to merge
5670 last_bfqq_created->ref++;
5672 * need to keep track of stable refs, to
5673 * compute process refs correctly
5675 last_bfqq_created->stable_ref++;
5677 * Record the bfqq to merge to.
5679 bic->stable_merge_bfqq = last_bfqq_created;
5687 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
5688 struct bio *bio, bool is_sync,
5689 struct bfq_io_cq *bic,
5692 const int ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
5693 const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
5694 struct bfq_queue **async_bfqq = NULL;
5695 struct bfq_queue *bfqq;
5696 struct bfq_group *bfqg;
5698 bfqg = bfq_bio_bfqg(bfqd, bio);
5700 async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class,
5707 bfqq = kmem_cache_alloc_node(bfq_pool,
5708 GFP_NOWAIT | __GFP_ZERO | __GFP_NOWARN,
5712 bfq_init_bfqq(bfqd, bfqq, bic, current->pid,
5714 bfq_init_entity(&bfqq->entity, bfqg);
5715 bfq_log_bfqq(bfqd, bfqq, "allocated");
5717 bfqq = &bfqd->oom_bfqq;
5718 bfq_log_bfqq(bfqd, bfqq, "using oom bfqq");
5723 * Pin the queue now that it's allocated, scheduler exit will
5728 * Extra group reference, w.r.t. sync
5729 * queue. This extra reference is removed
5730 * only if bfqq->bfqg disappears, to
5731 * guarantee that this queue is not freed
5732 * until its group goes away.
5734 bfq_log_bfqq(bfqd, bfqq, "get_queue, bfqq not in async: %p, %d",
5740 bfqq->ref++; /* get a process reference to this queue */
5742 if (bfqq != &bfqd->oom_bfqq && is_sync && !respawn)
5743 bfqq = bfq_do_or_sched_stable_merge(bfqd, bfqq, bic);
5747 static void bfq_update_io_thinktime(struct bfq_data *bfqd,
5748 struct bfq_queue *bfqq)
5750 struct bfq_ttime *ttime = &bfqq->ttime;
5754 * We are really interested in how long it takes for the queue to
5755 * become busy when there is no outstanding IO for this queue. So
5756 * ignore cases when the bfq queue has already IO queued.
5758 if (bfqq->dispatched || bfq_bfqq_busy(bfqq))
5760 elapsed = ktime_get_ns() - bfqq->ttime.last_end_request;
5761 elapsed = min_t(u64, elapsed, 2ULL * bfqd->bfq_slice_idle);
5763 ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
5764 ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed, 8);
5765 ttime->ttime_mean = div64_ul(ttime->ttime_total + 128,
5766 ttime->ttime_samples);
5770 bfq_update_io_seektime(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5773 bfqq->seek_history <<= 1;
5774 bfqq->seek_history |= BFQ_RQ_SEEKY(bfqd, bfqq->last_request_pos, rq);
5776 if (bfqq->wr_coeff > 1 &&
5777 bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
5778 BFQQ_TOTALLY_SEEKY(bfqq)) {
5779 if (time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
5780 bfq_wr_duration(bfqd))) {
5782 * In soft_rt weight raising with the
5783 * interactive-weight-raising period
5784 * elapsed (so no switch back to
5785 * interactive weight raising).
5787 bfq_bfqq_end_wr(bfqq);
5789 * stopping soft_rt weight raising
5790 * while still in interactive period,
5791 * switch back to interactive weight
5794 switch_back_to_interactive_wr(bfqq, bfqd);
5795 bfqq->entity.prio_changed = 1;
5800 static void bfq_update_has_short_ttime(struct bfq_data *bfqd,
5801 struct bfq_queue *bfqq,
5802 struct bfq_io_cq *bic)
5804 bool has_short_ttime = true, state_changed;
5807 * No need to update has_short_ttime if bfqq is async or in
5808 * idle io prio class, or if bfq_slice_idle is zero, because
5809 * no device idling is performed for bfqq in this case.
5811 if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq) ||
5812 bfqd->bfq_slice_idle == 0)
5815 /* Idle window just restored, statistics are meaningless. */
5816 if (time_is_after_eq_jiffies(bfqq->split_time +
5817 bfqd->bfq_wr_min_idle_time))
5820 /* Think time is infinite if no process is linked to
5821 * bfqq. Otherwise check average think time to decide whether
5822 * to mark as has_short_ttime. To this goal, compare average
5823 * think time with half the I/O-plugging timeout.
5825 if (atomic_read(&bic->icq.ioc->active_ref) == 0 ||
5826 (bfq_sample_valid(bfqq->ttime.ttime_samples) &&
5827 bfqq->ttime.ttime_mean > bfqd->bfq_slice_idle>>1))
5828 has_short_ttime = false;
5830 state_changed = has_short_ttime != bfq_bfqq_has_short_ttime(bfqq);
5832 if (has_short_ttime)
5833 bfq_mark_bfqq_has_short_ttime(bfqq);
5835 bfq_clear_bfqq_has_short_ttime(bfqq);
5838 * Until the base value for the total service time gets
5839 * finally computed for bfqq, the inject limit does depend on
5840 * the think-time state (short|long). In particular, the limit
5841 * is 0 or 1 if the think time is deemed, respectively, as
5842 * short or long (details in the comments in
5843 * bfq_update_inject_limit()). Accordingly, the next
5844 * instructions reset the inject limit if the think-time state
5845 * has changed and the above base value is still to be
5848 * However, the reset is performed only if more than 100 ms
5849 * have elapsed since the last update of the inject limit, or
5850 * (inclusive) if the change is from short to long think
5851 * time. The reason for this waiting is as follows.
5853 * bfqq may have a long think time because of a
5854 * synchronization with some other queue, i.e., because the
5855 * I/O of some other queue may need to be completed for bfqq
5856 * to receive new I/O. Details in the comments on the choice
5857 * of the queue for injection in bfq_select_queue().
5859 * As stressed in those comments, if such a synchronization is
5860 * actually in place, then, without injection on bfqq, the
5861 * blocking I/O cannot happen to served while bfqq is in
5862 * service. As a consequence, if bfqq is granted
5863 * I/O-dispatch-plugging, then bfqq remains empty, and no I/O
5864 * is dispatched, until the idle timeout fires. This is likely
5865 * to result in lower bandwidth and higher latencies for bfqq,
5866 * and in a severe loss of total throughput.
5868 * On the opposite end, a non-zero inject limit may allow the
5869 * I/O that blocks bfqq to be executed soon, and therefore
5870 * bfqq to receive new I/O soon.
5872 * But, if the blocking gets actually eliminated, then the
5873 * next think-time sample for bfqq may be very low. This in
5874 * turn may cause bfqq's think time to be deemed
5875 * short. Without the 100 ms barrier, this new state change
5876 * would cause the body of the next if to be executed
5877 * immediately. But this would set to 0 the inject
5878 * limit. Without injection, the blocking I/O would cause the
5879 * think time of bfqq to become long again, and therefore the
5880 * inject limit to be raised again, and so on. The only effect
5881 * of such a steady oscillation between the two think-time
5882 * states would be to prevent effective injection on bfqq.
5884 * In contrast, if the inject limit is not reset during such a
5885 * long time interval as 100 ms, then the number of short
5886 * think time samples can grow significantly before the reset
5887 * is performed. As a consequence, the think time state can
5888 * become stable before the reset. Therefore there will be no
5889 * state change when the 100 ms elapse, and no reset of the
5890 * inject limit. The inject limit remains steadily equal to 1
5891 * both during and after the 100 ms. So injection can be
5892 * performed at all times, and throughput gets boosted.
5894 * An inject limit equal to 1 is however in conflict, in
5895 * general, with the fact that the think time of bfqq is
5896 * short, because injection may be likely to delay bfqq's I/O
5897 * (as explained in the comments in
5898 * bfq_update_inject_limit()). But this does not happen in
5899 * this special case, because bfqq's low think time is due to
5900 * an effective handling of a synchronization, through
5901 * injection. In this special case, bfqq's I/O does not get
5902 * delayed by injection; on the contrary, bfqq's I/O is
5903 * brought forward, because it is not blocked for
5906 * In addition, serving the blocking I/O much sooner, and much
5907 * more frequently than once per I/O-plugging timeout, makes
5908 * it much quicker to detect a waker queue (the concept of
5909 * waker queue is defined in the comments in
5910 * bfq_add_request()). This makes it possible to start sooner
5911 * to boost throughput more effectively, by injecting the I/O
5912 * of the waker queue unconditionally on every
5913 * bfq_dispatch_request().
5915 * One last, important benefit of not resetting the inject
5916 * limit before 100 ms is that, during this time interval, the
5917 * base value for the total service time is likely to get
5918 * finally computed for bfqq, freeing the inject limit from
5919 * its relation with the think time.
5921 if (state_changed && bfqq->last_serv_time_ns == 0 &&
5922 (time_is_before_eq_jiffies(bfqq->decrease_time_jif +
5923 msecs_to_jiffies(100)) ||
5925 bfq_reset_inject_limit(bfqd, bfqq);
5929 * Called when a new fs request (rq) is added to bfqq. Check if there's
5930 * something we should do about it.
5932 static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5935 if (rq->cmd_flags & REQ_META)
5936 bfqq->meta_pending++;
5938 bfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
5940 if (bfqq == bfqd->in_service_queue && bfq_bfqq_wait_request(bfqq)) {
5941 bool small_req = bfqq->queued[rq_is_sync(rq)] == 1 &&
5942 blk_rq_sectors(rq) < 32;
5943 bool budget_timeout = bfq_bfqq_budget_timeout(bfqq);
5946 * There is just this request queued: if
5947 * - the request is small, and
5948 * - we are idling to boost throughput, and
5949 * - the queue is not to be expired,
5952 * In this way, if the device is being idled to wait
5953 * for a new request from the in-service queue, we
5954 * avoid unplugging the device and committing the
5955 * device to serve just a small request. In contrast
5956 * we wait for the block layer to decide when to
5957 * unplug the device: hopefully, new requests will be
5958 * merged to this one quickly, then the device will be
5959 * unplugged and larger requests will be dispatched.
5961 if (small_req && idling_boosts_thr_without_issues(bfqd, bfqq) &&
5966 * A large enough request arrived, or idling is being
5967 * performed to preserve service guarantees, or
5968 * finally the queue is to be expired: in all these
5969 * cases disk idling is to be stopped, so clear
5970 * wait_request flag and reset timer.
5972 bfq_clear_bfqq_wait_request(bfqq);
5973 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
5976 * The queue is not empty, because a new request just
5977 * arrived. Hence we can safely expire the queue, in
5978 * case of budget timeout, without risking that the
5979 * timestamps of the queue are not updated correctly.
5980 * See [1] for more details.
5983 bfq_bfqq_expire(bfqd, bfqq, false,
5984 BFQQE_BUDGET_TIMEOUT);
5988 static void bfqq_request_allocated(struct bfq_queue *bfqq)
5990 struct bfq_entity *entity = &bfqq->entity;
5992 for_each_entity(entity)
5993 entity->allocated++;
5996 static void bfqq_request_freed(struct bfq_queue *bfqq)
5998 struct bfq_entity *entity = &bfqq->entity;
6000 for_each_entity(entity)
6001 entity->allocated--;
6004 /* returns true if it causes the idle timer to be disabled */
6005 static bool __bfq_insert_request(struct bfq_data *bfqd, struct request *rq)
6007 struct bfq_queue *bfqq = RQ_BFQQ(rq),
6008 *new_bfqq = bfq_setup_cooperator(bfqd, bfqq, rq, true,
6010 bool waiting, idle_timer_disabled = false;
6014 * Release the request's reference to the old bfqq
6015 * and make sure one is taken to the shared queue.
6017 bfqq_request_allocated(new_bfqq);
6018 bfqq_request_freed(bfqq);
6021 * If the bic associated with the process
6022 * issuing this request still points to bfqq
6023 * (and thus has not been already redirected
6024 * to new_bfqq or even some other bfq_queue),
6025 * then complete the merge and redirect it to
6028 if (bic_to_bfqq(RQ_BIC(rq), 1) == bfqq)
6029 bfq_merge_bfqqs(bfqd, RQ_BIC(rq),
6032 bfq_clear_bfqq_just_created(bfqq);
6034 * rq is about to be enqueued into new_bfqq,
6035 * release rq reference on bfqq
6037 bfq_put_queue(bfqq);
6038 rq->elv.priv[1] = new_bfqq;
6042 bfq_update_io_thinktime(bfqd, bfqq);
6043 bfq_update_has_short_ttime(bfqd, bfqq, RQ_BIC(rq));
6044 bfq_update_io_seektime(bfqd, bfqq, rq);
6046 waiting = bfqq && bfq_bfqq_wait_request(bfqq);
6047 bfq_add_request(rq);
6048 idle_timer_disabled = waiting && !bfq_bfqq_wait_request(bfqq);
6050 rq->fifo_time = ktime_get_ns() + bfqd->bfq_fifo_expire[rq_is_sync(rq)];
6051 list_add_tail(&rq->queuelist, &bfqq->fifo);
6053 bfq_rq_enqueued(bfqd, bfqq, rq);
6055 return idle_timer_disabled;
6058 #ifdef CONFIG_BFQ_CGROUP_DEBUG
6059 static void bfq_update_insert_stats(struct request_queue *q,
6060 struct bfq_queue *bfqq,
6061 bool idle_timer_disabled,
6062 blk_opf_t cmd_flags)
6068 * bfqq still exists, because it can disappear only after
6069 * either it is merged with another queue, or the process it
6070 * is associated with exits. But both actions must be taken by
6071 * the same process currently executing this flow of
6074 * In addition, the following queue lock guarantees that
6075 * bfqq_group(bfqq) exists as well.
6077 spin_lock_irq(&q->queue_lock);
6078 bfqg_stats_update_io_add(bfqq_group(bfqq), bfqq, cmd_flags);
6079 if (idle_timer_disabled)
6080 bfqg_stats_update_idle_time(bfqq_group(bfqq));
6081 spin_unlock_irq(&q->queue_lock);
6084 static inline void bfq_update_insert_stats(struct request_queue *q,
6085 struct bfq_queue *bfqq,
6086 bool idle_timer_disabled,
6087 blk_opf_t cmd_flags) {}
6088 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
6090 static struct bfq_queue *bfq_init_rq(struct request *rq);
6092 static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
6095 struct request_queue *q = hctx->queue;
6096 struct bfq_data *bfqd = q->elevator->elevator_data;
6097 struct bfq_queue *bfqq;
6098 bool idle_timer_disabled = false;
6099 blk_opf_t cmd_flags;
6102 #ifdef CONFIG_BFQ_GROUP_IOSCHED
6103 if (!cgroup_subsys_on_dfl(io_cgrp_subsys) && rq->bio)
6104 bfqg_stats_update_legacy_io(q, rq);
6106 spin_lock_irq(&bfqd->lock);
6107 bfqq = bfq_init_rq(rq);
6108 if (blk_mq_sched_try_insert_merge(q, rq, &free)) {
6109 spin_unlock_irq(&bfqd->lock);
6110 blk_mq_free_requests(&free);
6114 trace_block_rq_insert(rq);
6116 if (!bfqq || at_head) {
6118 list_add(&rq->queuelist, &bfqd->dispatch);
6120 list_add_tail(&rq->queuelist, &bfqd->dispatch);
6122 idle_timer_disabled = __bfq_insert_request(bfqd, rq);
6124 * Update bfqq, because, if a queue merge has occurred
6125 * in __bfq_insert_request, then rq has been
6126 * redirected into a new queue.
6130 if (rq_mergeable(rq)) {
6131 elv_rqhash_add(q, rq);
6138 * Cache cmd_flags before releasing scheduler lock, because rq
6139 * may disappear afterwards (for example, because of a request
6142 cmd_flags = rq->cmd_flags;
6143 spin_unlock_irq(&bfqd->lock);
6145 bfq_update_insert_stats(q, bfqq, idle_timer_disabled,
6149 static void bfq_insert_requests(struct blk_mq_hw_ctx *hctx,
6150 struct list_head *list, bool at_head)
6152 while (!list_empty(list)) {
6155 rq = list_first_entry(list, struct request, queuelist);
6156 list_del_init(&rq->queuelist);
6157 bfq_insert_request(hctx, rq, at_head);
6161 static void bfq_update_hw_tag(struct bfq_data *bfqd)
6163 struct bfq_queue *bfqq = bfqd->in_service_queue;
6165 bfqd->max_rq_in_driver = max_t(int, bfqd->max_rq_in_driver,
6166 bfqd->rq_in_driver);
6168 if (bfqd->hw_tag == 1)
6172 * This sample is valid if the number of outstanding requests
6173 * is large enough to allow a queueing behavior. Note that the
6174 * sum is not exact, as it's not taking into account deactivated
6177 if (bfqd->rq_in_driver + bfqd->queued <= BFQ_HW_QUEUE_THRESHOLD)
6181 * If active queue hasn't enough requests and can idle, bfq might not
6182 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
6185 if (bfqq && bfq_bfqq_has_short_ttime(bfqq) &&
6186 bfqq->dispatched + bfqq->queued[0] + bfqq->queued[1] <
6187 BFQ_HW_QUEUE_THRESHOLD &&
6188 bfqd->rq_in_driver < BFQ_HW_QUEUE_THRESHOLD)
6191 if (bfqd->hw_tag_samples++ < BFQ_HW_QUEUE_SAMPLES)
6194 bfqd->hw_tag = bfqd->max_rq_in_driver > BFQ_HW_QUEUE_THRESHOLD;
6195 bfqd->max_rq_in_driver = 0;
6196 bfqd->hw_tag_samples = 0;
6198 bfqd->nonrot_with_queueing =
6199 blk_queue_nonrot(bfqd->queue) && bfqd->hw_tag;
6202 static void bfq_completed_request(struct bfq_queue *bfqq, struct bfq_data *bfqd)
6207 bfq_update_hw_tag(bfqd);
6209 bfqd->rq_in_driver--;
6212 if (!bfqq->dispatched && !bfq_bfqq_busy(bfqq)) {
6214 * Set budget_timeout (which we overload to store the
6215 * time at which the queue remains with no backlog and
6216 * no outstanding request; used by the weight-raising
6219 bfqq->budget_timeout = jiffies;
6221 bfq_del_bfqq_in_groups_with_pending_reqs(bfqq);
6222 bfq_weights_tree_remove(bfqq);
6225 now_ns = ktime_get_ns();
6227 bfqq->ttime.last_end_request = now_ns;
6230 * Using us instead of ns, to get a reasonable precision in
6231 * computing rate in next check.
6233 delta_us = div_u64(now_ns - bfqd->last_completion, NSEC_PER_USEC);
6236 * If the request took rather long to complete, and, according
6237 * to the maximum request size recorded, this completion latency
6238 * implies that the request was certainly served at a very low
6239 * rate (less than 1M sectors/sec), then the whole observation
6240 * interval that lasts up to this time instant cannot be a
6241 * valid time interval for computing a new peak rate. Invoke
6242 * bfq_update_rate_reset to have the following three steps
6244 * - close the observation interval at the last (previous)
6245 * request dispatch or completion
6246 * - compute rate, if possible, for that observation interval
6247 * - reset to zero samples, which will trigger a proper
6248 * re-initialization of the observation interval on next
6251 if (delta_us > BFQ_MIN_TT/NSEC_PER_USEC &&
6252 (bfqd->last_rq_max_size<<BFQ_RATE_SHIFT)/delta_us <
6253 1UL<<(BFQ_RATE_SHIFT - 10))
6254 bfq_update_rate_reset(bfqd, NULL);
6255 bfqd->last_completion = now_ns;
6257 * Shared queues are likely to receive I/O at a high
6258 * rate. This may deceptively let them be considered as wakers
6259 * of other queues. But a false waker will unjustly steal
6260 * bandwidth to its supposedly woken queue. So considering
6261 * also shared queues in the waking mechanism may cause more
6262 * control troubles than throughput benefits. Then reset
6263 * last_completed_rq_bfqq if bfqq is a shared queue.
6265 if (!bfq_bfqq_coop(bfqq))
6266 bfqd->last_completed_rq_bfqq = bfqq;
6268 bfqd->last_completed_rq_bfqq = NULL;
6271 * If we are waiting to discover whether the request pattern
6272 * of the task associated with the queue is actually
6273 * isochronous, and both requisites for this condition to hold
6274 * are now satisfied, then compute soft_rt_next_start (see the
6275 * comments on the function bfq_bfqq_softrt_next_start()). We
6276 * do not compute soft_rt_next_start if bfqq is in interactive
6277 * weight raising (see the comments in bfq_bfqq_expire() for
6278 * an explanation). We schedule this delayed update when bfqq
6279 * expires, if it still has in-flight requests.
6281 if (bfq_bfqq_softrt_update(bfqq) && bfqq->dispatched == 0 &&
6282 RB_EMPTY_ROOT(&bfqq->sort_list) &&
6283 bfqq->wr_coeff != bfqd->bfq_wr_coeff)
6284 bfqq->soft_rt_next_start =
6285 bfq_bfqq_softrt_next_start(bfqd, bfqq);
6288 * If this is the in-service queue, check if it needs to be expired,
6289 * or if we want to idle in case it has no pending requests.
6291 if (bfqd->in_service_queue == bfqq) {
6292 if (bfq_bfqq_must_idle(bfqq)) {
6293 if (bfqq->dispatched == 0)
6294 bfq_arm_slice_timer(bfqd);
6296 * If we get here, we do not expire bfqq, even
6297 * if bfqq was in budget timeout or had no
6298 * more requests (as controlled in the next
6299 * conditional instructions). The reason for
6300 * not expiring bfqq is as follows.
6302 * Here bfqq->dispatched > 0 holds, but
6303 * bfq_bfqq_must_idle() returned true. This
6304 * implies that, even if no request arrives
6305 * for bfqq before bfqq->dispatched reaches 0,
6306 * bfqq will, however, not be expired on the
6307 * completion event that causes bfqq->dispatch
6308 * to reach zero. In contrast, on this event,
6309 * bfqq will start enjoying device idling
6310 * (I/O-dispatch plugging).
6312 * But, if we expired bfqq here, bfqq would
6313 * not have the chance to enjoy device idling
6314 * when bfqq->dispatched finally reaches
6315 * zero. This would expose bfqq to violation
6316 * of its reserved service guarantees.
6319 } else if (bfq_may_expire_for_budg_timeout(bfqq))
6320 bfq_bfqq_expire(bfqd, bfqq, false,
6321 BFQQE_BUDGET_TIMEOUT);
6322 else if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
6323 (bfqq->dispatched == 0 ||
6324 !bfq_better_to_idle(bfqq)))
6325 bfq_bfqq_expire(bfqd, bfqq, false,
6326 BFQQE_NO_MORE_REQUESTS);
6329 if (!bfqd->rq_in_driver)
6330 bfq_schedule_dispatch(bfqd);
6334 * The processes associated with bfqq may happen to generate their
6335 * cumulative I/O at a lower rate than the rate at which the device
6336 * could serve the same I/O. This is rather probable, e.g., if only
6337 * one process is associated with bfqq and the device is an SSD. It
6338 * results in bfqq becoming often empty while in service. In this
6339 * respect, if BFQ is allowed to switch to another queue when bfqq
6340 * remains empty, then the device goes on being fed with I/O requests,
6341 * and the throughput is not affected. In contrast, if BFQ is not
6342 * allowed to switch to another queue---because bfqq is sync and
6343 * I/O-dispatch needs to be plugged while bfqq is temporarily
6344 * empty---then, during the service of bfqq, there will be frequent
6345 * "service holes", i.e., time intervals during which bfqq gets empty
6346 * and the device can only consume the I/O already queued in its
6347 * hardware queues. During service holes, the device may even get to
6348 * remaining idle. In the end, during the service of bfqq, the device
6349 * is driven at a lower speed than the one it can reach with the kind
6350 * of I/O flowing through bfqq.
6352 * To counter this loss of throughput, BFQ implements a "request
6353 * injection mechanism", which tries to fill the above service holes
6354 * with I/O requests taken from other queues. The hard part in this
6355 * mechanism is finding the right amount of I/O to inject, so as to
6356 * both boost throughput and not break bfqq's bandwidth and latency
6357 * guarantees. In this respect, the mechanism maintains a per-queue
6358 * inject limit, computed as below. While bfqq is empty, the injection
6359 * mechanism dispatches extra I/O requests only until the total number
6360 * of I/O requests in flight---i.e., already dispatched but not yet
6361 * completed---remains lower than this limit.
6363 * A first definition comes in handy to introduce the algorithm by
6364 * which the inject limit is computed. We define as first request for
6365 * bfqq, an I/O request for bfqq that arrives while bfqq is in
6366 * service, and causes bfqq to switch from empty to non-empty. The
6367 * algorithm updates the limit as a function of the effect of
6368 * injection on the service times of only the first requests of
6369 * bfqq. The reason for this restriction is that these are the
6370 * requests whose service time is affected most, because they are the
6371 * first to arrive after injection possibly occurred.
6373 * To evaluate the effect of injection, the algorithm measures the
6374 * "total service time" of first requests. We define as total service
6375 * time of an I/O request, the time that elapses since when the
6376 * request is enqueued into bfqq, to when it is completed. This
6377 * quantity allows the whole effect of injection to be measured. It is
6378 * easy to see why. Suppose that some requests of other queues are
6379 * actually injected while bfqq is empty, and that a new request R
6380 * then arrives for bfqq. If the device does start to serve all or
6381 * part of the injected requests during the service hole, then,
6382 * because of this extra service, it may delay the next invocation of
6383 * the dispatch hook of BFQ. Then, even after R gets eventually
6384 * dispatched, the device may delay the actual service of R if it is
6385 * still busy serving the extra requests, or if it decides to serve,
6386 * before R, some extra request still present in its queues. As a
6387 * conclusion, the cumulative extra delay caused by injection can be
6388 * easily evaluated by just comparing the total service time of first
6389 * requests with and without injection.
6391 * The limit-update algorithm works as follows. On the arrival of a
6392 * first request of bfqq, the algorithm measures the total time of the
6393 * request only if one of the three cases below holds, and, for each
6394 * case, it updates the limit as described below:
6396 * (1) If there is no in-flight request. This gives a baseline for the
6397 * total service time of the requests of bfqq. If the baseline has
6398 * not been computed yet, then, after computing it, the limit is
6399 * set to 1, to start boosting throughput, and to prepare the
6400 * ground for the next case. If the baseline has already been
6401 * computed, then it is updated, in case it results to be lower
6402 * than the previous value.
6404 * (2) If the limit is higher than 0 and there are in-flight
6405 * requests. By comparing the total service time in this case with
6406 * the above baseline, it is possible to know at which extent the
6407 * current value of the limit is inflating the total service
6408 * time. If the inflation is below a certain threshold, then bfqq
6409 * is assumed to be suffering from no perceivable loss of its
6410 * service guarantees, and the limit is even tentatively
6411 * increased. If the inflation is above the threshold, then the
6412 * limit is decreased. Due to the lack of any hysteresis, this
6413 * logic makes the limit oscillate even in steady workload
6414 * conditions. Yet we opted for it, because it is fast in reaching
6415 * the best value for the limit, as a function of the current I/O
6416 * workload. To reduce oscillations, this step is disabled for a
6417 * short time interval after the limit happens to be decreased.
6419 * (3) Periodically, after resetting the limit, to make sure that the
6420 * limit eventually drops in case the workload changes. This is
6421 * needed because, after the limit has gone safely up for a
6422 * certain workload, it is impossible to guess whether the
6423 * baseline total service time may have changed, without measuring
6424 * it again without injection. A more effective version of this
6425 * step might be to just sample the baseline, by interrupting
6426 * injection only once, and then to reset/lower the limit only if
6427 * the total service time with the current limit does happen to be
6430 * More details on each step are provided in the comments on the
6431 * pieces of code that implement these steps: the branch handling the
6432 * transition from empty to non empty in bfq_add_request(), the branch
6433 * handling injection in bfq_select_queue(), and the function
6434 * bfq_choose_bfqq_for_injection(). These comments also explain some
6435 * exceptions, made by the injection mechanism in some special cases.
6437 static void bfq_update_inject_limit(struct bfq_data *bfqd,
6438 struct bfq_queue *bfqq)
6440 u64 tot_time_ns = ktime_get_ns() - bfqd->last_empty_occupied_ns;
6441 unsigned int old_limit = bfqq->inject_limit;
6443 if (bfqq->last_serv_time_ns > 0 && bfqd->rqs_injected) {
6444 u64 threshold = (bfqq->last_serv_time_ns * 3)>>1;
6446 if (tot_time_ns >= threshold && old_limit > 0) {
6447 bfqq->inject_limit--;
6448 bfqq->decrease_time_jif = jiffies;
6449 } else if (tot_time_ns < threshold &&
6450 old_limit <= bfqd->max_rq_in_driver)
6451 bfqq->inject_limit++;
6455 * Either we still have to compute the base value for the
6456 * total service time, and there seem to be the right
6457 * conditions to do it, or we can lower the last base value
6460 * NOTE: (bfqd->rq_in_driver == 1) means that there is no I/O
6461 * request in flight, because this function is in the code
6462 * path that handles the completion of a request of bfqq, and,
6463 * in particular, this function is executed before
6464 * bfqd->rq_in_driver is decremented in such a code path.
6466 if ((bfqq->last_serv_time_ns == 0 && bfqd->rq_in_driver == 1) ||
6467 tot_time_ns < bfqq->last_serv_time_ns) {
6468 if (bfqq->last_serv_time_ns == 0) {
6470 * Now we certainly have a base value: make sure we
6471 * start trying injection.
6473 bfqq->inject_limit = max_t(unsigned int, 1, old_limit);
6475 bfqq->last_serv_time_ns = tot_time_ns;
6476 } else if (!bfqd->rqs_injected && bfqd->rq_in_driver == 1)
6478 * No I/O injected and no request still in service in
6479 * the drive: these are the exact conditions for
6480 * computing the base value of the total service time
6481 * for bfqq. So let's update this value, because it is
6482 * rather variable. For example, it varies if the size
6483 * or the spatial locality of the I/O requests in bfqq
6486 bfqq->last_serv_time_ns = tot_time_ns;
6489 /* update complete, not waiting for any request completion any longer */
6490 bfqd->waited_rq = NULL;
6491 bfqd->rqs_injected = false;
6495 * Handle either a requeue or a finish for rq. The things to do are
6496 * the same in both cases: all references to rq are to be dropped. In
6497 * particular, rq is considered completed from the point of view of
6500 static void bfq_finish_requeue_request(struct request *rq)
6502 struct bfq_queue *bfqq = RQ_BFQQ(rq);
6503 struct bfq_data *bfqd;
6504 unsigned long flags;
6507 * rq either is not associated with any icq, or is an already
6508 * requeued request that has not (yet) been re-inserted into
6511 if (!rq->elv.icq || !bfqq)
6516 if (rq->rq_flags & RQF_STARTED)
6517 bfqg_stats_update_completion(bfqq_group(bfqq),
6519 rq->io_start_time_ns,
6522 spin_lock_irqsave(&bfqd->lock, flags);
6523 if (likely(rq->rq_flags & RQF_STARTED)) {
6524 if (rq == bfqd->waited_rq)
6525 bfq_update_inject_limit(bfqd, bfqq);
6527 bfq_completed_request(bfqq, bfqd);
6529 bfqq_request_freed(bfqq);
6530 bfq_put_queue(bfqq);
6531 RQ_BIC(rq)->requests--;
6532 spin_unlock_irqrestore(&bfqd->lock, flags);
6535 * Reset private fields. In case of a requeue, this allows
6536 * this function to correctly do nothing if it is spuriously
6537 * invoked again on this same request (see the check at the
6538 * beginning of the function). Probably, a better general
6539 * design would be to prevent blk-mq from invoking the requeue
6540 * or finish hooks of an elevator, for a request that is not
6541 * referred by that elevator.
6543 * Resetting the following fields would break the
6544 * request-insertion logic if rq is re-inserted into a bfq
6545 * internal queue, without a re-preparation. Here we assume
6546 * that re-insertions of requeued requests, without
6547 * re-preparation, can happen only for pass_through or at_head
6548 * requests (which are not re-inserted into bfq internal
6551 rq->elv.priv[0] = NULL;
6552 rq->elv.priv[1] = NULL;
6555 static void bfq_finish_request(struct request *rq)
6557 bfq_finish_requeue_request(rq);
6560 put_io_context(rq->elv.icq->ioc);
6566 * Removes the association between the current task and bfqq, assuming
6567 * that bic points to the bfq iocontext of the task.
6568 * Returns NULL if a new bfqq should be allocated, or the old bfqq if this
6569 * was the last process referring to that bfqq.
6571 static struct bfq_queue *
6572 bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq)
6574 bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue");
6576 if (bfqq_process_refs(bfqq) == 1) {
6577 bfqq->pid = current->pid;
6578 bfq_clear_bfqq_coop(bfqq);
6579 bfq_clear_bfqq_split_coop(bfqq);
6583 bic_set_bfqq(bic, NULL, 1);
6585 bfq_put_cooperator(bfqq);
6587 bfq_release_process_ref(bfqq->bfqd, bfqq);
6591 static struct bfq_queue *bfq_get_bfqq_handle_split(struct bfq_data *bfqd,
6592 struct bfq_io_cq *bic,
6594 bool split, bool is_sync,
6597 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
6599 if (likely(bfqq && bfqq != &bfqd->oom_bfqq))
6606 bfq_put_queue(bfqq);
6607 bfqq = bfq_get_queue(bfqd, bio, is_sync, bic, split);
6609 bic_set_bfqq(bic, bfqq, is_sync);
6610 if (split && is_sync) {
6611 if ((bic->was_in_burst_list && bfqd->large_burst) ||
6612 bic->saved_in_large_burst)
6613 bfq_mark_bfqq_in_large_burst(bfqq);
6615 bfq_clear_bfqq_in_large_burst(bfqq);
6616 if (bic->was_in_burst_list)
6618 * If bfqq was in the current
6619 * burst list before being
6620 * merged, then we have to add
6621 * it back. And we do not need
6622 * to increase burst_size, as
6623 * we did not decrement
6624 * burst_size when we removed
6625 * bfqq from the burst list as
6626 * a consequence of a merge
6628 * bfq_put_queue). In this
6629 * respect, it would be rather
6630 * costly to know whether the
6631 * current burst list is still
6632 * the same burst list from
6633 * which bfqq was removed on
6634 * the merge. To avoid this
6635 * cost, if bfqq was in a
6636 * burst list, then we add
6637 * bfqq to the current burst
6638 * list without any further
6639 * check. This can cause
6640 * inappropriate insertions,
6641 * but rarely enough to not
6642 * harm the detection of large
6643 * bursts significantly.
6645 hlist_add_head(&bfqq->burst_list_node,
6648 bfqq->split_time = jiffies;
6655 * Only reset private fields. The actual request preparation will be
6656 * performed by bfq_init_rq, when rq is either inserted or merged. See
6657 * comments on bfq_init_rq for the reason behind this delayed
6660 static void bfq_prepare_request(struct request *rq)
6662 rq->elv.icq = ioc_find_get_icq(rq->q);
6665 * Regardless of whether we have an icq attached, we have to
6666 * clear the scheduler pointers, as they might point to
6667 * previously allocated bic/bfqq structs.
6669 rq->elv.priv[0] = rq->elv.priv[1] = NULL;
6673 * If needed, init rq, allocate bfq data structures associated with
6674 * rq, and increment reference counters in the destination bfq_queue
6675 * for rq. Return the destination bfq_queue for rq, or NULL is rq is
6676 * not associated with any bfq_queue.
6678 * This function is invoked by the functions that perform rq insertion
6679 * or merging. One may have expected the above preparation operations
6680 * to be performed in bfq_prepare_request, and not delayed to when rq
6681 * is inserted or merged. The rationale behind this delayed
6682 * preparation is that, after the prepare_request hook is invoked for
6683 * rq, rq may still be transformed into a request with no icq, i.e., a
6684 * request not associated with any queue. No bfq hook is invoked to
6685 * signal this transformation. As a consequence, should these
6686 * preparation operations be performed when the prepare_request hook
6687 * is invoked, and should rq be transformed one moment later, bfq
6688 * would end up in an inconsistent state, because it would have
6689 * incremented some queue counters for an rq destined to
6690 * transformation, without any chance to correctly lower these
6691 * counters back. In contrast, no transformation can still happen for
6692 * rq after rq has been inserted or merged. So, it is safe to execute
6693 * these preparation operations when rq is finally inserted or merged.
6695 static struct bfq_queue *bfq_init_rq(struct request *rq)
6697 struct request_queue *q = rq->q;
6698 struct bio *bio = rq->bio;
6699 struct bfq_data *bfqd = q->elevator->elevator_data;
6700 struct bfq_io_cq *bic;
6701 const int is_sync = rq_is_sync(rq);
6702 struct bfq_queue *bfqq;
6703 bool new_queue = false;
6704 bool bfqq_already_existing = false, split = false;
6706 if (unlikely(!rq->elv.icq))
6710 * Assuming that elv.priv[1] is set only if everything is set
6711 * for this rq. This holds true, because this function is
6712 * invoked only for insertion or merging, and, after such
6713 * events, a request cannot be manipulated any longer before
6714 * being removed from bfq.
6716 if (rq->elv.priv[1])
6717 return rq->elv.priv[1];
6719 bic = icq_to_bic(rq->elv.icq);
6721 bfq_check_ioprio_change(bic, bio);
6723 bfq_bic_update_cgroup(bic, bio);
6725 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio, false, is_sync,
6728 if (likely(!new_queue)) {
6729 /* If the queue was seeky for too long, break it apart. */
6730 if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq) &&
6731 !bic->stably_merged) {
6732 struct bfq_queue *old_bfqq = bfqq;
6734 /* Update bic before losing reference to bfqq */
6735 if (bfq_bfqq_in_large_burst(bfqq))
6736 bic->saved_in_large_burst = true;
6738 bfqq = bfq_split_bfqq(bic, bfqq);
6742 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio,
6745 bfqq->waker_bfqq = old_bfqq->waker_bfqq;
6746 bfqq->tentative_waker_bfqq = NULL;
6749 * If the waker queue disappears, then
6750 * new_bfqq->waker_bfqq must be
6751 * reset. So insert new_bfqq into the
6752 * woken_list of the waker. See
6753 * bfq_check_waker for details.
6755 if (bfqq->waker_bfqq)
6756 hlist_add_head(&bfqq->woken_list_node,
6757 &bfqq->waker_bfqq->woken_list);
6759 bfqq_already_existing = true;
6763 bfqq_request_allocated(bfqq);
6766 bfq_log_bfqq(bfqd, bfqq, "get_request %p: bfqq %p, %d",
6767 rq, bfqq, bfqq->ref);
6769 rq->elv.priv[0] = bic;
6770 rq->elv.priv[1] = bfqq;
6773 * If a bfq_queue has only one process reference, it is owned
6774 * by only this bic: we can then set bfqq->bic = bic. in
6775 * addition, if the queue has also just been split, we have to
6778 if (likely(bfqq != &bfqd->oom_bfqq) && bfqq_process_refs(bfqq) == 1) {
6782 * The queue has just been split from a shared
6783 * queue: restore the idle window and the
6784 * possible weight raising period.
6786 bfq_bfqq_resume_state(bfqq, bfqd, bic,
6787 bfqq_already_existing);
6792 * Consider bfqq as possibly belonging to a burst of newly
6793 * created queues only if:
6794 * 1) A burst is actually happening (bfqd->burst_size > 0)
6796 * 2) There is no other active queue. In fact, if, in
6797 * contrast, there are active queues not belonging to the
6798 * possible burst bfqq may belong to, then there is no gain
6799 * in considering bfqq as belonging to a burst, and
6800 * therefore in not weight-raising bfqq. See comments on
6801 * bfq_handle_burst().
6803 * This filtering also helps eliminating false positives,
6804 * occurring when bfqq does not belong to an actual large
6805 * burst, but some background task (e.g., a service) happens
6806 * to trigger the creation of new queues very close to when
6807 * bfqq and its possible companion queues are created. See
6808 * comments on bfq_handle_burst() for further details also on
6811 if (unlikely(bfq_bfqq_just_created(bfqq) &&
6812 (bfqd->burst_size > 0 ||
6813 bfq_tot_busy_queues(bfqd) == 0)))
6814 bfq_handle_burst(bfqd, bfqq);
6820 bfq_idle_slice_timer_body(struct bfq_data *bfqd, struct bfq_queue *bfqq)
6822 enum bfqq_expiration reason;
6823 unsigned long flags;
6825 spin_lock_irqsave(&bfqd->lock, flags);
6828 * Considering that bfqq may be in race, we should firstly check
6829 * whether bfqq is in service before doing something on it. If
6830 * the bfqq in race is not in service, it has already been expired
6831 * through __bfq_bfqq_expire func and its wait_request flags has
6832 * been cleared in __bfq_bfqd_reset_in_service func.
6834 if (bfqq != bfqd->in_service_queue) {
6835 spin_unlock_irqrestore(&bfqd->lock, flags);
6839 bfq_clear_bfqq_wait_request(bfqq);
6841 if (bfq_bfqq_budget_timeout(bfqq))
6843 * Also here the queue can be safely expired
6844 * for budget timeout without wasting
6847 reason = BFQQE_BUDGET_TIMEOUT;
6848 else if (bfqq->queued[0] == 0 && bfqq->queued[1] == 0)
6850 * The queue may not be empty upon timer expiration,
6851 * because we may not disable the timer when the
6852 * first request of the in-service queue arrives
6853 * during disk idling.
6855 reason = BFQQE_TOO_IDLE;
6857 goto schedule_dispatch;
6859 bfq_bfqq_expire(bfqd, bfqq, true, reason);
6862 bfq_schedule_dispatch(bfqd);
6863 spin_unlock_irqrestore(&bfqd->lock, flags);
6867 * Handler of the expiration of the timer running if the in-service queue
6868 * is idling inside its time slice.
6870 static enum hrtimer_restart bfq_idle_slice_timer(struct hrtimer *timer)
6872 struct bfq_data *bfqd = container_of(timer, struct bfq_data,
6874 struct bfq_queue *bfqq = bfqd->in_service_queue;
6877 * Theoretical race here: the in-service queue can be NULL or
6878 * different from the queue that was idling if a new request
6879 * arrives for the current queue and there is a full dispatch
6880 * cycle that changes the in-service queue. This can hardly
6881 * happen, but in the worst case we just expire a queue too
6885 bfq_idle_slice_timer_body(bfqd, bfqq);
6887 return HRTIMER_NORESTART;
6890 static void __bfq_put_async_bfqq(struct bfq_data *bfqd,
6891 struct bfq_queue **bfqq_ptr)
6893 struct bfq_queue *bfqq = *bfqq_ptr;
6895 bfq_log(bfqd, "put_async_bfqq: %p", bfqq);
6897 bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
6899 bfq_log_bfqq(bfqd, bfqq, "put_async_bfqq: putting %p, %d",
6901 bfq_put_queue(bfqq);
6907 * Release all the bfqg references to its async queues. If we are
6908 * deallocating the group these queues may still contain requests, so
6909 * we reparent them to the root cgroup (i.e., the only one that will
6910 * exist for sure until all the requests on a device are gone).
6912 void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg)
6916 for (i = 0; i < 2; i++)
6917 for (j = 0; j < IOPRIO_NR_LEVELS; j++)
6918 __bfq_put_async_bfqq(bfqd, &bfqg->async_bfqq[i][j]);
6920 __bfq_put_async_bfqq(bfqd, &bfqg->async_idle_bfqq);
6924 * See the comments on bfq_limit_depth for the purpose of
6925 * the depths set in the function. Return minimum shallow depth we'll use.
6927 static void bfq_update_depths(struct bfq_data *bfqd, struct sbitmap_queue *bt)
6929 unsigned int depth = 1U << bt->sb.shift;
6931 bfqd->full_depth_shift = bt->sb.shift;
6933 * In-word depths if no bfq_queue is being weight-raised:
6934 * leaving 25% of tags only for sync reads.
6936 * In next formulas, right-shift the value
6937 * (1U<<bt->sb.shift), instead of computing directly
6938 * (1U<<(bt->sb.shift - something)), to be robust against
6939 * any possible value of bt->sb.shift, without having to
6940 * limit 'something'.
6942 /* no more than 50% of tags for async I/O */
6943 bfqd->word_depths[0][0] = max(depth >> 1, 1U);
6945 * no more than 75% of tags for sync writes (25% extra tags
6946 * w.r.t. async I/O, to prevent async I/O from starving sync
6949 bfqd->word_depths[0][1] = max((depth * 3) >> 2, 1U);
6952 * In-word depths in case some bfq_queue is being weight-
6953 * raised: leaving ~63% of tags for sync reads. This is the
6954 * highest percentage for which, in our tests, application
6955 * start-up times didn't suffer from any regression due to tag
6958 /* no more than ~18% of tags for async I/O */
6959 bfqd->word_depths[1][0] = max((depth * 3) >> 4, 1U);
6960 /* no more than ~37% of tags for sync writes (~20% extra tags) */
6961 bfqd->word_depths[1][1] = max((depth * 6) >> 4, 1U);
6964 static void bfq_depth_updated(struct blk_mq_hw_ctx *hctx)
6966 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
6967 struct blk_mq_tags *tags = hctx->sched_tags;
6969 bfq_update_depths(bfqd, &tags->bitmap_tags);
6970 sbitmap_queue_min_shallow_depth(&tags->bitmap_tags, 1);
6973 static int bfq_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int index)
6975 bfq_depth_updated(hctx);
6979 static void bfq_exit_queue(struct elevator_queue *e)
6981 struct bfq_data *bfqd = e->elevator_data;
6982 struct bfq_queue *bfqq, *n;
6984 hrtimer_cancel(&bfqd->idle_slice_timer);
6986 spin_lock_irq(&bfqd->lock);
6987 list_for_each_entry_safe(bfqq, n, &bfqd->idle_list, bfqq_list)
6988 bfq_deactivate_bfqq(bfqd, bfqq, false, false);
6989 spin_unlock_irq(&bfqd->lock);
6991 hrtimer_cancel(&bfqd->idle_slice_timer);
6993 /* release oom-queue reference to root group */
6994 bfqg_and_blkg_put(bfqd->root_group);
6996 #ifdef CONFIG_BFQ_GROUP_IOSCHED
6997 blkcg_deactivate_policy(bfqd->queue, &blkcg_policy_bfq);
6999 spin_lock_irq(&bfqd->lock);
7000 bfq_put_async_queues(bfqd, bfqd->root_group);
7001 kfree(bfqd->root_group);
7002 spin_unlock_irq(&bfqd->lock);
7005 blk_stat_disable_accounting(bfqd->queue);
7006 clear_bit(ELEVATOR_FLAG_DISABLE_WBT, &e->flags);
7007 wbt_enable_default(bfqd->queue);
7012 static void bfq_init_root_group(struct bfq_group *root_group,
7013 struct bfq_data *bfqd)
7017 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7018 root_group->entity.parent = NULL;
7019 root_group->my_entity = NULL;
7020 root_group->bfqd = bfqd;
7022 root_group->rq_pos_tree = RB_ROOT;
7023 for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
7024 root_group->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
7025 root_group->sched_data.bfq_class_idle_last_service = jiffies;
7028 static int bfq_init_queue(struct request_queue *q, struct elevator_type *e)
7030 struct bfq_data *bfqd;
7031 struct elevator_queue *eq;
7033 eq = elevator_alloc(q, e);
7037 bfqd = kzalloc_node(sizeof(*bfqd), GFP_KERNEL, q->node);
7039 kobject_put(&eq->kobj);
7042 eq->elevator_data = bfqd;
7044 spin_lock_irq(&q->queue_lock);
7046 spin_unlock_irq(&q->queue_lock);
7049 * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues.
7050 * Grab a permanent reference to it, so that the normal code flow
7051 * will not attempt to free it.
7053 bfq_init_bfqq(bfqd, &bfqd->oom_bfqq, NULL, 1, 0);
7054 bfqd->oom_bfqq.ref++;
7055 bfqd->oom_bfqq.new_ioprio = BFQ_DEFAULT_QUEUE_IOPRIO;
7056 bfqd->oom_bfqq.new_ioprio_class = IOPRIO_CLASS_BE;
7057 bfqd->oom_bfqq.entity.new_weight =
7058 bfq_ioprio_to_weight(bfqd->oom_bfqq.new_ioprio);
7060 /* oom_bfqq does not participate to bursts */
7061 bfq_clear_bfqq_just_created(&bfqd->oom_bfqq);
7064 * Trigger weight initialization, according to ioprio, at the
7065 * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio
7066 * class won't be changed any more.
7068 bfqd->oom_bfqq.entity.prio_changed = 1;
7072 INIT_LIST_HEAD(&bfqd->dispatch);
7074 hrtimer_init(&bfqd->idle_slice_timer, CLOCK_MONOTONIC,
7076 bfqd->idle_slice_timer.function = bfq_idle_slice_timer;
7078 bfqd->queue_weights_tree = RB_ROOT_CACHED;
7079 bfqd->num_groups_with_pending_reqs = 0;
7081 INIT_LIST_HEAD(&bfqd->active_list);
7082 INIT_LIST_HEAD(&bfqd->idle_list);
7083 INIT_HLIST_HEAD(&bfqd->burst_list);
7086 bfqd->nonrot_with_queueing = blk_queue_nonrot(bfqd->queue);
7088 bfqd->bfq_max_budget = bfq_default_max_budget;
7090 bfqd->bfq_fifo_expire[0] = bfq_fifo_expire[0];
7091 bfqd->bfq_fifo_expire[1] = bfq_fifo_expire[1];
7092 bfqd->bfq_back_max = bfq_back_max;
7093 bfqd->bfq_back_penalty = bfq_back_penalty;
7094 bfqd->bfq_slice_idle = bfq_slice_idle;
7095 bfqd->bfq_timeout = bfq_timeout;
7097 bfqd->bfq_large_burst_thresh = 8;
7098 bfqd->bfq_burst_interval = msecs_to_jiffies(180);
7100 bfqd->low_latency = true;
7103 * Trade-off between responsiveness and fairness.
7105 bfqd->bfq_wr_coeff = 30;
7106 bfqd->bfq_wr_rt_max_time = msecs_to_jiffies(300);
7107 bfqd->bfq_wr_max_time = 0;
7108 bfqd->bfq_wr_min_idle_time = msecs_to_jiffies(2000);
7109 bfqd->bfq_wr_min_inter_arr_async = msecs_to_jiffies(500);
7110 bfqd->bfq_wr_max_softrt_rate = 7000; /*
7111 * Approximate rate required
7112 * to playback or record a
7113 * high-definition compressed
7116 bfqd->wr_busy_queues = 0;
7119 * Begin by assuming, optimistically, that the device peak
7120 * rate is equal to 2/3 of the highest reference rate.
7122 bfqd->rate_dur_prod = ref_rate[blk_queue_nonrot(bfqd->queue)] *
7123 ref_wr_duration[blk_queue_nonrot(bfqd->queue)];
7124 bfqd->peak_rate = ref_rate[blk_queue_nonrot(bfqd->queue)] * 2 / 3;
7126 spin_lock_init(&bfqd->lock);
7129 * The invocation of the next bfq_create_group_hierarchy
7130 * function is the head of a chain of function calls
7131 * (bfq_create_group_hierarchy->blkcg_activate_policy->
7132 * blk_mq_freeze_queue) that may lead to the invocation of the
7133 * has_work hook function. For this reason,
7134 * bfq_create_group_hierarchy is invoked only after all
7135 * scheduler data has been initialized, apart from the fields
7136 * that can be initialized only after invoking
7137 * bfq_create_group_hierarchy. This, in particular, enables
7138 * has_work to correctly return false. Of course, to avoid
7139 * other inconsistencies, the blk-mq stack must then refrain
7140 * from invoking further scheduler hooks before this init
7141 * function is finished.
7143 bfqd->root_group = bfq_create_group_hierarchy(bfqd, q->node);
7144 if (!bfqd->root_group)
7146 bfq_init_root_group(bfqd->root_group, bfqd);
7147 bfq_init_entity(&bfqd->oom_bfqq.entity, bfqd->root_group);
7149 /* We dispatch from request queue wide instead of hw queue */
7150 blk_queue_flag_set(QUEUE_FLAG_SQ_SCHED, q);
7152 set_bit(ELEVATOR_FLAG_DISABLE_WBT, &eq->flags);
7153 wbt_disable_default(q);
7154 blk_stat_enable_accounting(q);
7160 kobject_put(&eq->kobj);
7164 static void bfq_slab_kill(void)
7166 kmem_cache_destroy(bfq_pool);
7169 static int __init bfq_slab_setup(void)
7171 bfq_pool = KMEM_CACHE(bfq_queue, 0);
7177 static ssize_t bfq_var_show(unsigned int var, char *page)
7179 return sprintf(page, "%u\n", var);
7182 static int bfq_var_store(unsigned long *var, const char *page)
7184 unsigned long new_val;
7185 int ret = kstrtoul(page, 10, &new_val);
7193 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
7194 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
7196 struct bfq_data *bfqd = e->elevator_data; \
7197 u64 __data = __VAR; \
7199 __data = jiffies_to_msecs(__data); \
7200 else if (__CONV == 2) \
7201 __data = div_u64(__data, NSEC_PER_MSEC); \
7202 return bfq_var_show(__data, (page)); \
7204 SHOW_FUNCTION(bfq_fifo_expire_sync_show, bfqd->bfq_fifo_expire[1], 2);
7205 SHOW_FUNCTION(bfq_fifo_expire_async_show, bfqd->bfq_fifo_expire[0], 2);
7206 SHOW_FUNCTION(bfq_back_seek_max_show, bfqd->bfq_back_max, 0);
7207 SHOW_FUNCTION(bfq_back_seek_penalty_show, bfqd->bfq_back_penalty, 0);
7208 SHOW_FUNCTION(bfq_slice_idle_show, bfqd->bfq_slice_idle, 2);
7209 SHOW_FUNCTION(bfq_max_budget_show, bfqd->bfq_user_max_budget, 0);
7210 SHOW_FUNCTION(bfq_timeout_sync_show, bfqd->bfq_timeout, 1);
7211 SHOW_FUNCTION(bfq_strict_guarantees_show, bfqd->strict_guarantees, 0);
7212 SHOW_FUNCTION(bfq_low_latency_show, bfqd->low_latency, 0);
7213 #undef SHOW_FUNCTION
7215 #define USEC_SHOW_FUNCTION(__FUNC, __VAR) \
7216 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
7218 struct bfq_data *bfqd = e->elevator_data; \
7219 u64 __data = __VAR; \
7220 __data = div_u64(__data, NSEC_PER_USEC); \
7221 return bfq_var_show(__data, (page)); \
7223 USEC_SHOW_FUNCTION(bfq_slice_idle_us_show, bfqd->bfq_slice_idle);
7224 #undef USEC_SHOW_FUNCTION
7226 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
7228 __FUNC(struct elevator_queue *e, const char *page, size_t count) \
7230 struct bfq_data *bfqd = e->elevator_data; \
7231 unsigned long __data, __min = (MIN), __max = (MAX); \
7234 ret = bfq_var_store(&__data, (page)); \
7237 if (__data < __min) \
7239 else if (__data > __max) \
7242 *(__PTR) = msecs_to_jiffies(__data); \
7243 else if (__CONV == 2) \
7244 *(__PTR) = (u64)__data * NSEC_PER_MSEC; \
7246 *(__PTR) = __data; \
7249 STORE_FUNCTION(bfq_fifo_expire_sync_store, &bfqd->bfq_fifo_expire[1], 1,
7251 STORE_FUNCTION(bfq_fifo_expire_async_store, &bfqd->bfq_fifo_expire[0], 1,
7253 STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0);
7254 STORE_FUNCTION(bfq_back_seek_penalty_store, &bfqd->bfq_back_penalty, 1,
7256 STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 2);
7257 #undef STORE_FUNCTION
7259 #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
7260 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\
7262 struct bfq_data *bfqd = e->elevator_data; \
7263 unsigned long __data, __min = (MIN), __max = (MAX); \
7266 ret = bfq_var_store(&__data, (page)); \
7269 if (__data < __min) \
7271 else if (__data > __max) \
7273 *(__PTR) = (u64)__data * NSEC_PER_USEC; \
7276 USEC_STORE_FUNCTION(bfq_slice_idle_us_store, &bfqd->bfq_slice_idle, 0,
7278 #undef USEC_STORE_FUNCTION
7280 static ssize_t bfq_max_budget_store(struct elevator_queue *e,
7281 const char *page, size_t count)
7283 struct bfq_data *bfqd = e->elevator_data;
7284 unsigned long __data;
7287 ret = bfq_var_store(&__data, (page));
7292 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
7294 if (__data > INT_MAX)
7296 bfqd->bfq_max_budget = __data;
7299 bfqd->bfq_user_max_budget = __data;
7305 * Leaving this name to preserve name compatibility with cfq
7306 * parameters, but this timeout is used for both sync and async.
7308 static ssize_t bfq_timeout_sync_store(struct elevator_queue *e,
7309 const char *page, size_t count)
7311 struct bfq_data *bfqd = e->elevator_data;
7312 unsigned long __data;
7315 ret = bfq_var_store(&__data, (page));
7321 else if (__data > INT_MAX)
7324 bfqd->bfq_timeout = msecs_to_jiffies(__data);
7325 if (bfqd->bfq_user_max_budget == 0)
7326 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
7331 static ssize_t bfq_strict_guarantees_store(struct elevator_queue *e,
7332 const char *page, size_t count)
7334 struct bfq_data *bfqd = e->elevator_data;
7335 unsigned long __data;
7338 ret = bfq_var_store(&__data, (page));
7344 if (!bfqd->strict_guarantees && __data == 1
7345 && bfqd->bfq_slice_idle < 8 * NSEC_PER_MSEC)
7346 bfqd->bfq_slice_idle = 8 * NSEC_PER_MSEC;
7348 bfqd->strict_guarantees = __data;
7353 static ssize_t bfq_low_latency_store(struct elevator_queue *e,
7354 const char *page, size_t count)
7356 struct bfq_data *bfqd = e->elevator_data;
7357 unsigned long __data;
7360 ret = bfq_var_store(&__data, (page));
7366 if (__data == 0 && bfqd->low_latency != 0)
7368 bfqd->low_latency = __data;
7373 #define BFQ_ATTR(name) \
7374 __ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store)
7376 static struct elv_fs_entry bfq_attrs[] = {
7377 BFQ_ATTR(fifo_expire_sync),
7378 BFQ_ATTR(fifo_expire_async),
7379 BFQ_ATTR(back_seek_max),
7380 BFQ_ATTR(back_seek_penalty),
7381 BFQ_ATTR(slice_idle),
7382 BFQ_ATTR(slice_idle_us),
7383 BFQ_ATTR(max_budget),
7384 BFQ_ATTR(timeout_sync),
7385 BFQ_ATTR(strict_guarantees),
7386 BFQ_ATTR(low_latency),
7390 static struct elevator_type iosched_bfq_mq = {
7392 .limit_depth = bfq_limit_depth,
7393 .prepare_request = bfq_prepare_request,
7394 .requeue_request = bfq_finish_requeue_request,
7395 .finish_request = bfq_finish_request,
7396 .exit_icq = bfq_exit_icq,
7397 .insert_requests = bfq_insert_requests,
7398 .dispatch_request = bfq_dispatch_request,
7399 .next_request = elv_rb_latter_request,
7400 .former_request = elv_rb_former_request,
7401 .allow_merge = bfq_allow_bio_merge,
7402 .bio_merge = bfq_bio_merge,
7403 .request_merge = bfq_request_merge,
7404 .requests_merged = bfq_requests_merged,
7405 .request_merged = bfq_request_merged,
7406 .has_work = bfq_has_work,
7407 .depth_updated = bfq_depth_updated,
7408 .init_hctx = bfq_init_hctx,
7409 .init_sched = bfq_init_queue,
7410 .exit_sched = bfq_exit_queue,
7413 .icq_size = sizeof(struct bfq_io_cq),
7414 .icq_align = __alignof__(struct bfq_io_cq),
7415 .elevator_attrs = bfq_attrs,
7416 .elevator_name = "bfq",
7417 .elevator_owner = THIS_MODULE,
7419 MODULE_ALIAS("bfq-iosched");
7421 static int __init bfq_init(void)
7425 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7426 ret = blkcg_policy_register(&blkcg_policy_bfq);
7432 if (bfq_slab_setup())
7436 * Times to load large popular applications for the typical
7437 * systems installed on the reference devices (see the
7438 * comments before the definition of the next
7439 * array). Actually, we use slightly lower values, as the
7440 * estimated peak rate tends to be smaller than the actual
7441 * peak rate. The reason for this last fact is that estimates
7442 * are computed over much shorter time intervals than the long
7443 * intervals typically used for benchmarking. Why? First, to
7444 * adapt more quickly to variations. Second, because an I/O
7445 * scheduler cannot rely on a peak-rate-evaluation workload to
7446 * be run for a long time.
7448 ref_wr_duration[0] = msecs_to_jiffies(7000); /* actually 8 sec */
7449 ref_wr_duration[1] = msecs_to_jiffies(2500); /* actually 3 sec */
7451 ret = elv_register(&iosched_bfq_mq);
7460 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7461 blkcg_policy_unregister(&blkcg_policy_bfq);
7466 static void __exit bfq_exit(void)
7468 elv_unregister(&iosched_bfq_mq);
7469 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7470 blkcg_policy_unregister(&blkcg_policy_bfq);
7475 module_init(bfq_init);
7476 module_exit(bfq_exit);
7478 MODULE_AUTHOR("Paolo Valente");
7479 MODULE_LICENSE("GPL");
7480 MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler");