2 * Budget Fair Queueing (BFQ) I/O scheduler.
4 * Based on ideas and code from CFQ:
5 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
7 * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
8 * Paolo Valente <paolo.valente@unimore.it>
10 * Copyright (C) 2010 Paolo Valente <paolo.valente@unimore.it>
11 * Arianna Avanzini <avanzini@google.com>
13 * Copyright (C) 2017 Paolo Valente <paolo.valente@linaro.org>
15 * This program is free software; you can redistribute it and/or
16 * modify it under the terms of the GNU General Public License as
17 * published by the Free Software Foundation; either version 2 of the
18 * License, or (at your option) any later version.
20 * This program is distributed in the hope that it will be useful,
21 * but WITHOUT ANY WARRANTY; without even the implied warranty of
22 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
23 * General Public License for more details.
25 * BFQ is a proportional-share I/O scheduler, with some extra
26 * low-latency capabilities. BFQ also supports full hierarchical
27 * scheduling through cgroups. Next paragraphs provide an introduction
28 * on BFQ inner workings. Details on BFQ benefits, usage and
29 * limitations can be found in Documentation/block/bfq-iosched.txt.
31 * BFQ is a proportional-share storage-I/O scheduling algorithm based
32 * on the slice-by-slice service scheme of CFQ. But BFQ assigns
33 * budgets, measured in number of sectors, to processes instead of
34 * time slices. The device is not granted to the in-service process
35 * for a given time slice, but until it has exhausted its assigned
36 * budget. This change from the time to the service domain enables BFQ
37 * to distribute the device throughput among processes as desired,
38 * without any distortion due to throughput fluctuations, or to device
39 * internal queueing. BFQ uses an ad hoc internal scheduler, called
40 * B-WF2Q+, to schedule processes according to their budgets. More
41 * precisely, BFQ schedules queues associated with processes. Each
42 * process/queue is assigned a user-configurable weight, and B-WF2Q+
43 * guarantees that each queue receives a fraction of the throughput
44 * proportional to its weight. Thanks to the accurate policy of
45 * B-WF2Q+, BFQ can afford to assign high budgets to I/O-bound
46 * processes issuing sequential requests (to boost the throughput),
47 * and yet guarantee a low latency to interactive and soft real-time
50 * In particular, to provide these low-latency guarantees, BFQ
51 * explicitly privileges the I/O of two classes of time-sensitive
52 * applications: interactive and soft real-time. This feature enables
53 * BFQ to provide applications in these classes with a very low
54 * latency. Finally, BFQ also features additional heuristics for
55 * preserving both a low latency and a high throughput on NCQ-capable,
56 * rotational or flash-based devices, and to get the job done quickly
57 * for applications consisting in many I/O-bound processes.
59 * NOTE: if the main or only goal, with a given device, is to achieve
60 * the maximum-possible throughput at all times, then do switch off
61 * all low-latency heuristics for that device, by setting low_latency
64 * BFQ is described in [1], where also a reference to the initial, more
65 * theoretical paper on BFQ can be found. The interested reader can find
66 * in the latter paper full details on the main algorithm, as well as
67 * formulas of the guarantees and formal proofs of all the properties.
68 * With respect to the version of BFQ presented in these papers, this
69 * implementation adds a few more heuristics, such as the one that
70 * guarantees a low latency to soft real-time applications, and a
71 * hierarchical extension based on H-WF2Q+.
73 * B-WF2Q+ is based on WF2Q+, which is described in [2], together with
74 * H-WF2Q+, while the augmented tree used here to implement B-WF2Q+
75 * with O(log N) complexity derives from the one introduced with EEVDF
78 * [1] P. Valente, A. Avanzini, "Evolution of the BFQ Storage I/O
79 * Scheduler", Proceedings of the First Workshop on Mobile System
80 * Technologies (MST-2015), May 2015.
81 * http://algogroup.unimore.it/people/paolo/disk_sched/mst-2015.pdf
83 * [2] Jon C.R. Bennett and H. Zhang, "Hierarchical Packet Fair Queueing
84 * Algorithms", IEEE/ACM Transactions on Networking, 5(5):675-689,
87 * http://www.cs.cmu.edu/~hzhang/papers/TON-97-Oct.ps.gz
89 * [3] I. Stoica and H. Abdel-Wahab, "Earliest Eligible Virtual Deadline
90 * First: A Flexible and Accurate Mechanism for Proportional Share
91 * Resource Allocation", technical report.
93 * http://www.cs.berkeley.edu/~istoica/papers/eevdf-tr-95.pdf
95 #include <linux/module.h>
96 #include <linux/slab.h>
97 #include <linux/blkdev.h>
98 #include <linux/cgroup.h>
99 #include <linux/elevator.h>
100 #include <linux/ktime.h>
101 #include <linux/rbtree.h>
102 #include <linux/ioprio.h>
103 #include <linux/sbitmap.h>
104 #include <linux/delay.h>
108 #include "blk-mq-tag.h"
109 #include "blk-mq-sched.h"
110 #include "bfq-iosched.h"
113 #define BFQ_BFQQ_FNS(name) \
114 void bfq_mark_bfqq_##name(struct bfq_queue *bfqq) \
116 __set_bit(BFQQF_##name, &(bfqq)->flags); \
118 void bfq_clear_bfqq_##name(struct bfq_queue *bfqq) \
120 __clear_bit(BFQQF_##name, &(bfqq)->flags); \
122 int bfq_bfqq_##name(const struct bfq_queue *bfqq) \
124 return test_bit(BFQQF_##name, &(bfqq)->flags); \
127 BFQ_BFQQ_FNS(just_created);
129 BFQ_BFQQ_FNS(wait_request);
130 BFQ_BFQQ_FNS(non_blocking_wait_rq);
131 BFQ_BFQQ_FNS(fifo_expire);
132 BFQ_BFQQ_FNS(has_short_ttime);
134 BFQ_BFQQ_FNS(IO_bound);
135 BFQ_BFQQ_FNS(in_large_burst);
137 BFQ_BFQQ_FNS(split_coop);
138 BFQ_BFQQ_FNS(softrt_update);
139 #undef BFQ_BFQQ_FNS \
141 /* Expiration time of sync (0) and async (1) requests, in ns. */
142 static const u64 bfq_fifo_expire[2] = { NSEC_PER_SEC / 4, NSEC_PER_SEC / 8 };
144 /* Maximum backwards seek (magic number lifted from CFQ), in KiB. */
145 static const int bfq_back_max = 16 * 1024;
147 /* Penalty of a backwards seek, in number of sectors. */
148 static const int bfq_back_penalty = 2;
150 /* Idling period duration, in ns. */
151 static u64 bfq_slice_idle = NSEC_PER_SEC / 125;
153 /* Minimum number of assigned budgets for which stats are safe to compute. */
154 static const int bfq_stats_min_budgets = 194;
156 /* Default maximum budget values, in sectors and number of requests. */
157 static const int bfq_default_max_budget = 16 * 1024;
160 * Async to sync throughput distribution is controlled as follows:
161 * when an async request is served, the entity is charged the number
162 * of sectors of the request, multiplied by the factor below
164 static const int bfq_async_charge_factor = 10;
166 /* Default timeout values, in jiffies, approximating CFQ defaults. */
167 const int bfq_timeout = HZ / 8;
170 * Time limit for merging (see comments in bfq_setup_cooperator). Set
171 * to the slowest value that, in our tests, proved to be effective in
172 * removing false positives, while not causing true positives to miss
175 * As can be deduced from the low time limit below, queue merging, if
176 * successful, happens at the very beggining of the I/O of the involved
177 * cooperating processes, as a consequence of the arrival of the very
178 * first requests from each cooperator. After that, there is very
179 * little chance to find cooperators.
181 static const unsigned long bfq_merge_time_limit = HZ/10;
183 static struct kmem_cache *bfq_pool;
185 /* Below this threshold (in ns), we consider thinktime immediate. */
186 #define BFQ_MIN_TT (2 * NSEC_PER_MSEC)
188 /* hw_tag detection: parallel requests threshold and min samples needed. */
189 #define BFQ_HW_QUEUE_THRESHOLD 4
190 #define BFQ_HW_QUEUE_SAMPLES 32
192 #define BFQQ_SEEK_THR (sector_t)(8 * 100)
193 #define BFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
194 #define BFQQ_CLOSE_THR (sector_t)(8 * 1024)
195 #define BFQQ_SEEKY(bfqq) (hweight32(bfqq->seek_history) > 19)
197 /* Min number of samples required to perform peak-rate update */
198 #define BFQ_RATE_MIN_SAMPLES 32
199 /* Min observation time interval required to perform a peak-rate update (ns) */
200 #define BFQ_RATE_MIN_INTERVAL (300*NSEC_PER_MSEC)
201 /* Target observation time interval for a peak-rate update (ns) */
202 #define BFQ_RATE_REF_INTERVAL NSEC_PER_SEC
204 /* Shift used for peak rate fixed precision calculations. */
205 #define BFQ_RATE_SHIFT 16
208 * By default, BFQ computes the duration of the weight raising for
209 * interactive applications automatically, using the following formula:
210 * duration = (R / r) * T, where r is the peak rate of the device, and
211 * R and T are two reference parameters.
212 * In particular, R is the peak rate of the reference device (see below),
213 * and T is a reference time: given the systems that are likely to be
214 * installed on the reference device according to its speed class, T is
215 * about the maximum time needed, under BFQ and while reading two files in
216 * parallel, to load typical large applications on these systems.
217 * In practice, the slower/faster the device at hand is, the more/less it
218 * takes to load applications with respect to the reference device.
219 * Accordingly, the longer/shorter BFQ grants weight raising to interactive
222 * BFQ uses four different reference pairs (R, T), depending on:
223 * . whether the device is rotational or non-rotational;
224 * . whether the device is slow, such as old or portable HDDs, as well as
225 * SD cards, or fast, such as newer HDDs and SSDs.
227 * The device's speed class is dynamically (re)detected in
228 * bfq_update_peak_rate() every time the estimated peak rate is updated.
230 * In the following definitions, R_slow[0]/R_fast[0] and
231 * T_slow[0]/T_fast[0] are the reference values for a slow/fast
232 * rotational device, whereas R_slow[1]/R_fast[1] and
233 * T_slow[1]/T_fast[1] are the reference values for a slow/fast
234 * non-rotational device. Finally, device_speed_thresh are the
235 * thresholds used to switch between speed classes. The reference
236 * rates are not the actual peak rates of the devices used as a
237 * reference, but slightly lower values. The reason for using these
238 * slightly lower values is that the peak-rate estimator tends to
239 * yield slightly lower values than the actual peak rate (it can yield
240 * the actual peak rate only if there is only one process doing I/O,
241 * and the process does sequential I/O).
243 * Both the reference peak rates and the thresholds are measured in
244 * sectors/usec, left-shifted by BFQ_RATE_SHIFT.
246 static int R_slow[2] = {1000, 10700};
247 static int R_fast[2] = {14000, 33000};
249 * To improve readability, a conversion function is used to initialize the
250 * following arrays, which entails that they can be initialized only in a
253 static int T_slow[2];
254 static int T_fast[2];
255 static int device_speed_thresh[2];
257 #define RQ_BIC(rq) icq_to_bic((rq)->elv.priv[0])
258 #define RQ_BFQQ(rq) ((rq)->elv.priv[1])
260 struct bfq_queue *bic_to_bfqq(struct bfq_io_cq *bic, bool is_sync)
262 return bic->bfqq[is_sync];
265 void bic_set_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq, bool is_sync)
267 bic->bfqq[is_sync] = bfqq;
270 struct bfq_data *bic_to_bfqd(struct bfq_io_cq *bic)
272 return bic->icq.q->elevator->elevator_data;
276 * icq_to_bic - convert iocontext queue structure to bfq_io_cq.
277 * @icq: the iocontext queue.
279 static struct bfq_io_cq *icq_to_bic(struct io_cq *icq)
281 /* bic->icq is the first member, %NULL will convert to %NULL */
282 return container_of(icq, struct bfq_io_cq, icq);
286 * bfq_bic_lookup - search into @ioc a bic associated to @bfqd.
287 * @bfqd: the lookup key.
288 * @ioc: the io_context of the process doing I/O.
289 * @q: the request queue.
291 static struct bfq_io_cq *bfq_bic_lookup(struct bfq_data *bfqd,
292 struct io_context *ioc,
293 struct request_queue *q)
297 struct bfq_io_cq *icq;
299 spin_lock_irqsave(q->queue_lock, flags);
300 icq = icq_to_bic(ioc_lookup_icq(ioc, q));
301 spin_unlock_irqrestore(q->queue_lock, flags);
310 * Scheduler run of queue, if there are requests pending and no one in the
311 * driver that will restart queueing.
313 void bfq_schedule_dispatch(struct bfq_data *bfqd)
315 if (bfqd->queued != 0) {
316 bfq_log(bfqd, "schedule dispatch");
317 blk_mq_run_hw_queues(bfqd->queue, true);
321 #define bfq_class_idle(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
322 #define bfq_class_rt(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_RT)
324 #define bfq_sample_valid(samples) ((samples) > 80)
327 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
328 * We choose the request that is closesr to the head right now. Distance
329 * behind the head is penalized and only allowed to a certain extent.
331 static struct request *bfq_choose_req(struct bfq_data *bfqd,
336 sector_t s1, s2, d1 = 0, d2 = 0;
337 unsigned long back_max;
338 #define BFQ_RQ1_WRAP 0x01 /* request 1 wraps */
339 #define BFQ_RQ2_WRAP 0x02 /* request 2 wraps */
340 unsigned int wrap = 0; /* bit mask: requests behind the disk head? */
342 if (!rq1 || rq1 == rq2)
347 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
349 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
351 if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
353 else if ((rq2->cmd_flags & REQ_META) && !(rq1->cmd_flags & REQ_META))
356 s1 = blk_rq_pos(rq1);
357 s2 = blk_rq_pos(rq2);
360 * By definition, 1KiB is 2 sectors.
362 back_max = bfqd->bfq_back_max * 2;
365 * Strict one way elevator _except_ in the case where we allow
366 * short backward seeks which are biased as twice the cost of a
367 * similar forward seek.
371 else if (s1 + back_max >= last)
372 d1 = (last - s1) * bfqd->bfq_back_penalty;
374 wrap |= BFQ_RQ1_WRAP;
378 else if (s2 + back_max >= last)
379 d2 = (last - s2) * bfqd->bfq_back_penalty;
381 wrap |= BFQ_RQ2_WRAP;
383 /* Found required data */
386 * By doing switch() on the bit mask "wrap" we avoid having to
387 * check two variables for all permutations: --> faster!
390 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
405 case BFQ_RQ1_WRAP|BFQ_RQ2_WRAP: /* both rqs wrapped */
408 * Since both rqs are wrapped,
409 * start with the one that's further behind head
410 * (--> only *one* back seek required),
411 * since back seek takes more time than forward.
420 static struct bfq_queue *
421 bfq_rq_pos_tree_lookup(struct bfq_data *bfqd, struct rb_root *root,
422 sector_t sector, struct rb_node **ret_parent,
423 struct rb_node ***rb_link)
425 struct rb_node **p, *parent;
426 struct bfq_queue *bfqq = NULL;
434 bfqq = rb_entry(parent, struct bfq_queue, pos_node);
437 * Sort strictly based on sector. Smallest to the left,
438 * largest to the right.
440 if (sector > blk_rq_pos(bfqq->next_rq))
442 else if (sector < blk_rq_pos(bfqq->next_rq))
450 *ret_parent = parent;
454 bfq_log(bfqd, "rq_pos_tree_lookup %llu: returning %d",
455 (unsigned long long)sector,
456 bfqq ? bfqq->pid : 0);
461 static bool bfq_too_late_for_merging(struct bfq_queue *bfqq)
463 return bfqq->service_from_backlogged > 0 &&
464 time_is_before_jiffies(bfqq->first_IO_time +
465 bfq_merge_time_limit);
468 void bfq_pos_tree_add_move(struct bfq_data *bfqd, struct bfq_queue *bfqq)
470 struct rb_node **p, *parent;
471 struct bfq_queue *__bfqq;
473 if (bfqq->pos_root) {
474 rb_erase(&bfqq->pos_node, bfqq->pos_root);
475 bfqq->pos_root = NULL;
479 * bfqq cannot be merged any longer (see comments in
480 * bfq_setup_cooperator): no point in adding bfqq into the
483 if (bfq_too_late_for_merging(bfqq))
486 if (bfq_class_idle(bfqq))
491 bfqq->pos_root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree;
492 __bfqq = bfq_rq_pos_tree_lookup(bfqd, bfqq->pos_root,
493 blk_rq_pos(bfqq->next_rq), &parent, &p);
495 rb_link_node(&bfqq->pos_node, parent, p);
496 rb_insert_color(&bfqq->pos_node, bfqq->pos_root);
498 bfqq->pos_root = NULL;
502 * Tell whether there are active queues or groups with differentiated weights.
504 static bool bfq_differentiated_weights(struct bfq_data *bfqd)
507 * For weights to differ, at least one of the trees must contain
508 * at least two nodes.
510 return (!RB_EMPTY_ROOT(&bfqd->queue_weights_tree) &&
511 (bfqd->queue_weights_tree.rb_node->rb_left ||
512 bfqd->queue_weights_tree.rb_node->rb_right)
513 #ifdef CONFIG_BFQ_GROUP_IOSCHED
515 (!RB_EMPTY_ROOT(&bfqd->group_weights_tree) &&
516 (bfqd->group_weights_tree.rb_node->rb_left ||
517 bfqd->group_weights_tree.rb_node->rb_right)
523 * The following function returns true if every queue must receive the
524 * same share of the throughput (this condition is used when deciding
525 * whether idling may be disabled, see the comments in the function
526 * bfq_bfqq_may_idle()).
528 * Such a scenario occurs when:
529 * 1) all active queues have the same weight,
530 * 2) all active groups at the same level in the groups tree have the same
532 * 3) all active groups at the same level in the groups tree have the same
533 * number of children.
535 * Unfortunately, keeping the necessary state for evaluating exactly the
536 * above symmetry conditions would be quite complex and time-consuming.
537 * Therefore this function evaluates, instead, the following stronger
538 * sub-conditions, for which it is much easier to maintain the needed
540 * 1) all active queues have the same weight,
541 * 2) all active groups have the same weight,
542 * 3) all active groups have at most one active child each.
543 * In particular, the last two conditions are always true if hierarchical
544 * support and the cgroups interface are not enabled, thus no state needs
545 * to be maintained in this case.
547 static bool bfq_symmetric_scenario(struct bfq_data *bfqd)
549 return !bfq_differentiated_weights(bfqd);
553 * If the weight-counter tree passed as input contains no counter for
554 * the weight of the input entity, then add that counter; otherwise just
555 * increment the existing counter.
557 * Note that weight-counter trees contain few nodes in mostly symmetric
558 * scenarios. For example, if all queues have the same weight, then the
559 * weight-counter tree for the queues may contain at most one node.
560 * This holds even if low_latency is on, because weight-raised queues
561 * are not inserted in the tree.
562 * In most scenarios, the rate at which nodes are created/destroyed
565 void bfq_weights_tree_add(struct bfq_data *bfqd, struct bfq_entity *entity,
566 struct rb_root *root)
568 struct rb_node **new = &(root->rb_node), *parent = NULL;
571 * Do not insert if the entity is already associated with a
572 * counter, which happens if:
573 * 1) the entity is associated with a queue,
574 * 2) a request arrival has caused the queue to become both
575 * non-weight-raised, and hence change its weight, and
576 * backlogged; in this respect, each of the two events
577 * causes an invocation of this function,
578 * 3) this is the invocation of this function caused by the
579 * second event. This second invocation is actually useless,
580 * and we handle this fact by exiting immediately. More
581 * efficient or clearer solutions might possibly be adopted.
583 if (entity->weight_counter)
587 struct bfq_weight_counter *__counter = container_of(*new,
588 struct bfq_weight_counter,
592 if (entity->weight == __counter->weight) {
593 entity->weight_counter = __counter;
596 if (entity->weight < __counter->weight)
597 new = &((*new)->rb_left);
599 new = &((*new)->rb_right);
602 entity->weight_counter = kzalloc(sizeof(struct bfq_weight_counter),
606 * In the unlucky event of an allocation failure, we just
607 * exit. This will cause the weight of entity to not be
608 * considered in bfq_differentiated_weights, which, in its
609 * turn, causes the scenario to be deemed wrongly symmetric in
610 * case entity's weight would have been the only weight making
611 * the scenario asymmetric. On the bright side, no unbalance
612 * will however occur when entity becomes inactive again (the
613 * invocation of this function is triggered by an activation
614 * of entity). In fact, bfq_weights_tree_remove does nothing
615 * if !entity->weight_counter.
617 if (unlikely(!entity->weight_counter))
620 entity->weight_counter->weight = entity->weight;
621 rb_link_node(&entity->weight_counter->weights_node, parent, new);
622 rb_insert_color(&entity->weight_counter->weights_node, root);
625 entity->weight_counter->num_active++;
629 * Decrement the weight counter associated with the entity, and, if the
630 * counter reaches 0, remove the counter from the tree.
631 * See the comments to the function bfq_weights_tree_add() for considerations
634 void bfq_weights_tree_remove(struct bfq_data *bfqd, struct bfq_entity *entity,
635 struct rb_root *root)
637 if (!entity->weight_counter)
640 entity->weight_counter->num_active--;
641 if (entity->weight_counter->num_active > 0)
642 goto reset_entity_pointer;
644 rb_erase(&entity->weight_counter->weights_node, root);
645 kfree(entity->weight_counter);
647 reset_entity_pointer:
648 entity->weight_counter = NULL;
652 * Return expired entry, or NULL to just start from scratch in rbtree.
654 static struct request *bfq_check_fifo(struct bfq_queue *bfqq,
655 struct request *last)
659 if (bfq_bfqq_fifo_expire(bfqq))
662 bfq_mark_bfqq_fifo_expire(bfqq);
664 rq = rq_entry_fifo(bfqq->fifo.next);
666 if (rq == last || ktime_get_ns() < rq->fifo_time)
669 bfq_log_bfqq(bfqq->bfqd, bfqq, "check_fifo: returned %p", rq);
673 static struct request *bfq_find_next_rq(struct bfq_data *bfqd,
674 struct bfq_queue *bfqq,
675 struct request *last)
677 struct rb_node *rbnext = rb_next(&last->rb_node);
678 struct rb_node *rbprev = rb_prev(&last->rb_node);
679 struct request *next, *prev = NULL;
681 /* Follow expired path, else get first next available. */
682 next = bfq_check_fifo(bfqq, last);
687 prev = rb_entry_rq(rbprev);
690 next = rb_entry_rq(rbnext);
692 rbnext = rb_first(&bfqq->sort_list);
693 if (rbnext && rbnext != &last->rb_node)
694 next = rb_entry_rq(rbnext);
697 return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last));
700 /* see the definition of bfq_async_charge_factor for details */
701 static unsigned long bfq_serv_to_charge(struct request *rq,
702 struct bfq_queue *bfqq)
704 if (bfq_bfqq_sync(bfqq) || bfqq->wr_coeff > 1)
705 return blk_rq_sectors(rq);
708 * If there are no weight-raised queues, then amplify service
709 * by just the async charge factor; otherwise amplify service
710 * by twice the async charge factor, to further reduce latency
711 * for weight-raised queues.
713 if (bfqq->bfqd->wr_busy_queues == 0)
714 return blk_rq_sectors(rq) * bfq_async_charge_factor;
716 return blk_rq_sectors(rq) * 2 * bfq_async_charge_factor;
720 * bfq_updated_next_req - update the queue after a new next_rq selection.
721 * @bfqd: the device data the queue belongs to.
722 * @bfqq: the queue to update.
724 * If the first request of a queue changes we make sure that the queue
725 * has enough budget to serve at least its first request (if the
726 * request has grown). We do this because if the queue has not enough
727 * budget for its first request, it has to go through two dispatch
728 * rounds to actually get it dispatched.
730 static void bfq_updated_next_req(struct bfq_data *bfqd,
731 struct bfq_queue *bfqq)
733 struct bfq_entity *entity = &bfqq->entity;
734 struct request *next_rq = bfqq->next_rq;
735 unsigned long new_budget;
740 if (bfqq == bfqd->in_service_queue)
742 * In order not to break guarantees, budgets cannot be
743 * changed after an entity has been selected.
747 new_budget = max_t(unsigned long, bfqq->max_budget,
748 bfq_serv_to_charge(next_rq, bfqq));
749 if (entity->budget != new_budget) {
750 entity->budget = new_budget;
751 bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu",
753 bfq_requeue_bfqq(bfqd, bfqq, false);
757 static unsigned int bfq_wr_duration(struct bfq_data *bfqd)
761 if (bfqd->bfq_wr_max_time > 0)
762 return bfqd->bfq_wr_max_time;
765 do_div(dur, bfqd->peak_rate);
768 * Limit duration between 3 and 13 seconds. Tests show that
769 * higher values than 13 seconds often yield the opposite of
770 * the desired result, i.e., worsen responsiveness by letting
771 * non-interactive and non-soft-real-time applications
772 * preserve weight raising for a too long time interval.
774 * On the other end, lower values than 3 seconds make it
775 * difficult for most interactive tasks to complete their jobs
776 * before weight-raising finishes.
778 if (dur > msecs_to_jiffies(13000))
779 dur = msecs_to_jiffies(13000);
780 else if (dur < msecs_to_jiffies(3000))
781 dur = msecs_to_jiffies(3000);
786 /* switch back from soft real-time to interactive weight raising */
787 static void switch_back_to_interactive_wr(struct bfq_queue *bfqq,
788 struct bfq_data *bfqd)
790 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
791 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
792 bfqq->last_wr_start_finish = bfqq->wr_start_at_switch_to_srt;
796 bfq_bfqq_resume_state(struct bfq_queue *bfqq, struct bfq_data *bfqd,
797 struct bfq_io_cq *bic, bool bfq_already_existing)
799 unsigned int old_wr_coeff = bfqq->wr_coeff;
800 bool busy = bfq_already_existing && bfq_bfqq_busy(bfqq);
802 if (bic->saved_has_short_ttime)
803 bfq_mark_bfqq_has_short_ttime(bfqq);
805 bfq_clear_bfqq_has_short_ttime(bfqq);
807 if (bic->saved_IO_bound)
808 bfq_mark_bfqq_IO_bound(bfqq);
810 bfq_clear_bfqq_IO_bound(bfqq);
812 bfqq->ttime = bic->saved_ttime;
813 bfqq->wr_coeff = bic->saved_wr_coeff;
814 bfqq->wr_start_at_switch_to_srt = bic->saved_wr_start_at_switch_to_srt;
815 bfqq->last_wr_start_finish = bic->saved_last_wr_start_finish;
816 bfqq->wr_cur_max_time = bic->saved_wr_cur_max_time;
818 if (bfqq->wr_coeff > 1 && (bfq_bfqq_in_large_burst(bfqq) ||
819 time_is_before_jiffies(bfqq->last_wr_start_finish +
820 bfqq->wr_cur_max_time))) {
821 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
822 !bfq_bfqq_in_large_burst(bfqq) &&
823 time_is_after_eq_jiffies(bfqq->wr_start_at_switch_to_srt +
824 bfq_wr_duration(bfqd))) {
825 switch_back_to_interactive_wr(bfqq, bfqd);
828 bfq_log_bfqq(bfqq->bfqd, bfqq,
829 "resume state: switching off wr");
833 /* make sure weight will be updated, however we got here */
834 bfqq->entity.prio_changed = 1;
839 if (old_wr_coeff == 1 && bfqq->wr_coeff > 1)
840 bfqd->wr_busy_queues++;
841 else if (old_wr_coeff > 1 && bfqq->wr_coeff == 1)
842 bfqd->wr_busy_queues--;
845 static int bfqq_process_refs(struct bfq_queue *bfqq)
847 return bfqq->ref - bfqq->allocated - bfqq->entity.on_st;
850 /* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */
851 static void bfq_reset_burst_list(struct bfq_data *bfqd, struct bfq_queue *bfqq)
853 struct bfq_queue *item;
854 struct hlist_node *n;
856 hlist_for_each_entry_safe(item, n, &bfqd->burst_list, burst_list_node)
857 hlist_del_init(&item->burst_list_node);
858 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
859 bfqd->burst_size = 1;
860 bfqd->burst_parent_entity = bfqq->entity.parent;
863 /* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */
864 static void bfq_add_to_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
866 /* Increment burst size to take into account also bfqq */
869 if (bfqd->burst_size == bfqd->bfq_large_burst_thresh) {
870 struct bfq_queue *pos, *bfqq_item;
871 struct hlist_node *n;
874 * Enough queues have been activated shortly after each
875 * other to consider this burst as large.
877 bfqd->large_burst = true;
880 * We can now mark all queues in the burst list as
881 * belonging to a large burst.
883 hlist_for_each_entry(bfqq_item, &bfqd->burst_list,
885 bfq_mark_bfqq_in_large_burst(bfqq_item);
886 bfq_mark_bfqq_in_large_burst(bfqq);
889 * From now on, and until the current burst finishes, any
890 * new queue being activated shortly after the last queue
891 * was inserted in the burst can be immediately marked as
892 * belonging to a large burst. So the burst list is not
893 * needed any more. Remove it.
895 hlist_for_each_entry_safe(pos, n, &bfqd->burst_list,
897 hlist_del_init(&pos->burst_list_node);
899 * Burst not yet large: add bfqq to the burst list. Do
900 * not increment the ref counter for bfqq, because bfqq
901 * is removed from the burst list before freeing bfqq
904 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
908 * If many queues belonging to the same group happen to be created
909 * shortly after each other, then the processes associated with these
910 * queues have typically a common goal. In particular, bursts of queue
911 * creations are usually caused by services or applications that spawn
912 * many parallel threads/processes. Examples are systemd during boot,
913 * or git grep. To help these processes get their job done as soon as
914 * possible, it is usually better to not grant either weight-raising
915 * or device idling to their queues.
917 * In this comment we describe, firstly, the reasons why this fact
918 * holds, and, secondly, the next function, which implements the main
919 * steps needed to properly mark these queues so that they can then be
920 * treated in a different way.
922 * The above services or applications benefit mostly from a high
923 * throughput: the quicker the requests of the activated queues are
924 * cumulatively served, the sooner the target job of these queues gets
925 * completed. As a consequence, weight-raising any of these queues,
926 * which also implies idling the device for it, is almost always
927 * counterproductive. In most cases it just lowers throughput.
929 * On the other hand, a burst of queue creations may be caused also by
930 * the start of an application that does not consist of a lot of
931 * parallel I/O-bound threads. In fact, with a complex application,
932 * several short processes may need to be executed to start-up the
933 * application. In this respect, to start an application as quickly as
934 * possible, the best thing to do is in any case to privilege the I/O
935 * related to the application with respect to all other
936 * I/O. Therefore, the best strategy to start as quickly as possible
937 * an application that causes a burst of queue creations is to
938 * weight-raise all the queues created during the burst. This is the
939 * exact opposite of the best strategy for the other type of bursts.
941 * In the end, to take the best action for each of the two cases, the
942 * two types of bursts need to be distinguished. Fortunately, this
943 * seems relatively easy, by looking at the sizes of the bursts. In
944 * particular, we found a threshold such that only bursts with a
945 * larger size than that threshold are apparently caused by
946 * services or commands such as systemd or git grep. For brevity,
947 * hereafter we call just 'large' these bursts. BFQ *does not*
948 * weight-raise queues whose creation occurs in a large burst. In
949 * addition, for each of these queues BFQ performs or does not perform
950 * idling depending on which choice boosts the throughput more. The
951 * exact choice depends on the device and request pattern at
954 * Unfortunately, false positives may occur while an interactive task
955 * is starting (e.g., an application is being started). The
956 * consequence is that the queues associated with the task do not
957 * enjoy weight raising as expected. Fortunately these false positives
958 * are very rare. They typically occur if some service happens to
959 * start doing I/O exactly when the interactive task starts.
961 * Turning back to the next function, it implements all the steps
962 * needed to detect the occurrence of a large burst and to properly
963 * mark all the queues belonging to it (so that they can then be
964 * treated in a different way). This goal is achieved by maintaining a
965 * "burst list" that holds, temporarily, the queues that belong to the
966 * burst in progress. The list is then used to mark these queues as
967 * belonging to a large burst if the burst does become large. The main
968 * steps are the following.
970 * . when the very first queue is created, the queue is inserted into the
971 * list (as it could be the first queue in a possible burst)
973 * . if the current burst has not yet become large, and a queue Q that does
974 * not yet belong to the burst is activated shortly after the last time
975 * at which a new queue entered the burst list, then the function appends
976 * Q to the burst list
978 * . if, as a consequence of the previous step, the burst size reaches
979 * the large-burst threshold, then
981 * . all the queues in the burst list are marked as belonging to a
984 * . the burst list is deleted; in fact, the burst list already served
985 * its purpose (keeping temporarily track of the queues in a burst,
986 * so as to be able to mark them as belonging to a large burst in the
987 * previous sub-step), and now is not needed any more
989 * . the device enters a large-burst mode
991 * . if a queue Q that does not belong to the burst is created while
992 * the device is in large-burst mode and shortly after the last time
993 * at which a queue either entered the burst list or was marked as
994 * belonging to the current large burst, then Q is immediately marked
995 * as belonging to a large burst.
997 * . if a queue Q that does not belong to the burst is created a while
998 * later, i.e., not shortly after, than the last time at which a queue
999 * either entered the burst list or was marked as belonging to the
1000 * current large burst, then the current burst is deemed as finished and:
1002 * . the large-burst mode is reset if set
1004 * . the burst list is emptied
1006 * . Q is inserted in the burst list, as Q may be the first queue
1007 * in a possible new burst (then the burst list contains just Q
1010 static void bfq_handle_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1013 * If bfqq is already in the burst list or is part of a large
1014 * burst, or finally has just been split, then there is
1015 * nothing else to do.
1017 if (!hlist_unhashed(&bfqq->burst_list_node) ||
1018 bfq_bfqq_in_large_burst(bfqq) ||
1019 time_is_after_eq_jiffies(bfqq->split_time +
1020 msecs_to_jiffies(10)))
1024 * If bfqq's creation happens late enough, or bfqq belongs to
1025 * a different group than the burst group, then the current
1026 * burst is finished, and related data structures must be
1029 * In this respect, consider the special case where bfqq is
1030 * the very first queue created after BFQ is selected for this
1031 * device. In this case, last_ins_in_burst and
1032 * burst_parent_entity are not yet significant when we get
1033 * here. But it is easy to verify that, whether or not the
1034 * following condition is true, bfqq will end up being
1035 * inserted into the burst list. In particular the list will
1036 * happen to contain only bfqq. And this is exactly what has
1037 * to happen, as bfqq may be the first queue of the first
1040 if (time_is_before_jiffies(bfqd->last_ins_in_burst +
1041 bfqd->bfq_burst_interval) ||
1042 bfqq->entity.parent != bfqd->burst_parent_entity) {
1043 bfqd->large_burst = false;
1044 bfq_reset_burst_list(bfqd, bfqq);
1049 * If we get here, then bfqq is being activated shortly after the
1050 * last queue. So, if the current burst is also large, we can mark
1051 * bfqq as belonging to this large burst immediately.
1053 if (bfqd->large_burst) {
1054 bfq_mark_bfqq_in_large_burst(bfqq);
1059 * If we get here, then a large-burst state has not yet been
1060 * reached, but bfqq is being activated shortly after the last
1061 * queue. Then we add bfqq to the burst.
1063 bfq_add_to_burst(bfqd, bfqq);
1066 * At this point, bfqq either has been added to the current
1067 * burst or has caused the current burst to terminate and a
1068 * possible new burst to start. In particular, in the second
1069 * case, bfqq has become the first queue in the possible new
1070 * burst. In both cases last_ins_in_burst needs to be moved
1073 bfqd->last_ins_in_burst = jiffies;
1076 static int bfq_bfqq_budget_left(struct bfq_queue *bfqq)
1078 struct bfq_entity *entity = &bfqq->entity;
1080 return entity->budget - entity->service;
1084 * If enough samples have been computed, return the current max budget
1085 * stored in bfqd, which is dynamically updated according to the
1086 * estimated disk peak rate; otherwise return the default max budget
1088 static int bfq_max_budget(struct bfq_data *bfqd)
1090 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1091 return bfq_default_max_budget;
1093 return bfqd->bfq_max_budget;
1097 * Return min budget, which is a fraction of the current or default
1098 * max budget (trying with 1/32)
1100 static int bfq_min_budget(struct bfq_data *bfqd)
1102 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1103 return bfq_default_max_budget / 32;
1105 return bfqd->bfq_max_budget / 32;
1109 * The next function, invoked after the input queue bfqq switches from
1110 * idle to busy, updates the budget of bfqq. The function also tells
1111 * whether the in-service queue should be expired, by returning
1112 * true. The purpose of expiring the in-service queue is to give bfqq
1113 * the chance to possibly preempt the in-service queue, and the reason
1114 * for preempting the in-service queue is to achieve one of the two
1117 * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has
1118 * expired because it has remained idle. In particular, bfqq may have
1119 * expired for one of the following two reasons:
1121 * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling
1122 * and did not make it to issue a new request before its last
1123 * request was served;
1125 * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue
1126 * a new request before the expiration of the idling-time.
1128 * Even if bfqq has expired for one of the above reasons, the process
1129 * associated with the queue may be however issuing requests greedily,
1130 * and thus be sensitive to the bandwidth it receives (bfqq may have
1131 * remained idle for other reasons: CPU high load, bfqq not enjoying
1132 * idling, I/O throttling somewhere in the path from the process to
1133 * the I/O scheduler, ...). But if, after every expiration for one of
1134 * the above two reasons, bfqq has to wait for the service of at least
1135 * one full budget of another queue before being served again, then
1136 * bfqq is likely to get a much lower bandwidth or resource time than
1137 * its reserved ones. To address this issue, two countermeasures need
1140 * First, the budget and the timestamps of bfqq need to be updated in
1141 * a special way on bfqq reactivation: they need to be updated as if
1142 * bfqq did not remain idle and did not expire. In fact, if they are
1143 * computed as if bfqq expired and remained idle until reactivation,
1144 * then the process associated with bfqq is treated as if, instead of
1145 * being greedy, it stopped issuing requests when bfqq remained idle,
1146 * and restarts issuing requests only on this reactivation. In other
1147 * words, the scheduler does not help the process recover the "service
1148 * hole" between bfqq expiration and reactivation. As a consequence,
1149 * the process receives a lower bandwidth than its reserved one. In
1150 * contrast, to recover this hole, the budget must be updated as if
1151 * bfqq was not expired at all before this reactivation, i.e., it must
1152 * be set to the value of the remaining budget when bfqq was
1153 * expired. Along the same line, timestamps need to be assigned the
1154 * value they had the last time bfqq was selected for service, i.e.,
1155 * before last expiration. Thus timestamps need to be back-shifted
1156 * with respect to their normal computation (see [1] for more details
1157 * on this tricky aspect).
1159 * Secondly, to allow the process to recover the hole, the in-service
1160 * queue must be expired too, to give bfqq the chance to preempt it
1161 * immediately. In fact, if bfqq has to wait for a full budget of the
1162 * in-service queue to be completed, then it may become impossible to
1163 * let the process recover the hole, even if the back-shifted
1164 * timestamps of bfqq are lower than those of the in-service queue. If
1165 * this happens for most or all of the holes, then the process may not
1166 * receive its reserved bandwidth. In this respect, it is worth noting
1167 * that, being the service of outstanding requests unpreemptible, a
1168 * little fraction of the holes may however be unrecoverable, thereby
1169 * causing a little loss of bandwidth.
1171 * The last important point is detecting whether bfqq does need this
1172 * bandwidth recovery. In this respect, the next function deems the
1173 * process associated with bfqq greedy, and thus allows it to recover
1174 * the hole, if: 1) the process is waiting for the arrival of a new
1175 * request (which implies that bfqq expired for one of the above two
1176 * reasons), and 2) such a request has arrived soon. The first
1177 * condition is controlled through the flag non_blocking_wait_rq,
1178 * while the second through the flag arrived_in_time. If both
1179 * conditions hold, then the function computes the budget in the
1180 * above-described special way, and signals that the in-service queue
1181 * should be expired. Timestamp back-shifting is done later in
1182 * __bfq_activate_entity.
1184 * 2. Reduce latency. Even if timestamps are not backshifted to let
1185 * the process associated with bfqq recover a service hole, bfqq may
1186 * however happen to have, after being (re)activated, a lower finish
1187 * timestamp than the in-service queue. That is, the next budget of
1188 * bfqq may have to be completed before the one of the in-service
1189 * queue. If this is the case, then preempting the in-service queue
1190 * allows this goal to be achieved, apart from the unpreemptible,
1191 * outstanding requests mentioned above.
1193 * Unfortunately, regardless of which of the above two goals one wants
1194 * to achieve, service trees need first to be updated to know whether
1195 * the in-service queue must be preempted. To have service trees
1196 * correctly updated, the in-service queue must be expired and
1197 * rescheduled, and bfqq must be scheduled too. This is one of the
1198 * most costly operations (in future versions, the scheduling
1199 * mechanism may be re-designed in such a way to make it possible to
1200 * know whether preemption is needed without needing to update service
1201 * trees). In addition, queue preemptions almost always cause random
1202 * I/O, and thus loss of throughput. Because of these facts, the next
1203 * function adopts the following simple scheme to avoid both costly
1204 * operations and too frequent preemptions: it requests the expiration
1205 * of the in-service queue (unconditionally) only for queues that need
1206 * to recover a hole, or that either are weight-raised or deserve to
1209 static bool bfq_bfqq_update_budg_for_activation(struct bfq_data *bfqd,
1210 struct bfq_queue *bfqq,
1211 bool arrived_in_time,
1212 bool wr_or_deserves_wr)
1214 struct bfq_entity *entity = &bfqq->entity;
1216 if (bfq_bfqq_non_blocking_wait_rq(bfqq) && arrived_in_time) {
1218 * We do not clear the flag non_blocking_wait_rq here, as
1219 * the latter is used in bfq_activate_bfqq to signal
1220 * that timestamps need to be back-shifted (and is
1221 * cleared right after).
1225 * In next assignment we rely on that either
1226 * entity->service or entity->budget are not updated
1227 * on expiration if bfqq is empty (see
1228 * __bfq_bfqq_recalc_budget). Thus both quantities
1229 * remain unchanged after such an expiration, and the
1230 * following statement therefore assigns to
1231 * entity->budget the remaining budget on such an
1232 * expiration. For clarity, entity->service is not
1233 * updated on expiration in any case, and, in normal
1234 * operation, is reset only when bfqq is selected for
1235 * service (see bfq_get_next_queue).
1237 entity->budget = min_t(unsigned long,
1238 bfq_bfqq_budget_left(bfqq),
1244 entity->budget = max_t(unsigned long, bfqq->max_budget,
1245 bfq_serv_to_charge(bfqq->next_rq, bfqq));
1246 bfq_clear_bfqq_non_blocking_wait_rq(bfqq);
1247 return wr_or_deserves_wr;
1251 * Return the farthest future time instant according to jiffies
1254 static unsigned long bfq_greatest_from_now(void)
1256 return jiffies + MAX_JIFFY_OFFSET;
1260 * Return the farthest past time instant according to jiffies
1263 static unsigned long bfq_smallest_from_now(void)
1265 return jiffies - MAX_JIFFY_OFFSET;
1268 static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data *bfqd,
1269 struct bfq_queue *bfqq,
1270 unsigned int old_wr_coeff,
1271 bool wr_or_deserves_wr,
1276 if (old_wr_coeff == 1 && wr_or_deserves_wr) {
1277 /* start a weight-raising period */
1279 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1280 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1283 * No interactive weight raising in progress
1284 * here: assign minus infinity to
1285 * wr_start_at_switch_to_srt, to make sure
1286 * that, at the end of the soft-real-time
1287 * weight raising periods that is starting
1288 * now, no interactive weight-raising period
1289 * may be wrongly considered as still in
1290 * progress (and thus actually started by
1293 bfqq->wr_start_at_switch_to_srt =
1294 bfq_smallest_from_now();
1295 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1296 BFQ_SOFTRT_WEIGHT_FACTOR;
1297 bfqq->wr_cur_max_time =
1298 bfqd->bfq_wr_rt_max_time;
1302 * If needed, further reduce budget to make sure it is
1303 * close to bfqq's backlog, so as to reduce the
1304 * scheduling-error component due to a too large
1305 * budget. Do not care about throughput consequences,
1306 * but only about latency. Finally, do not assign a
1307 * too small budget either, to avoid increasing
1308 * latency by causing too frequent expirations.
1310 bfqq->entity.budget = min_t(unsigned long,
1311 bfqq->entity.budget,
1312 2 * bfq_min_budget(bfqd));
1313 } else if (old_wr_coeff > 1) {
1314 if (interactive) { /* update wr coeff and duration */
1315 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1316 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1317 } else if (in_burst)
1321 * The application is now or still meeting the
1322 * requirements for being deemed soft rt. We
1323 * can then correctly and safely (re)charge
1324 * the weight-raising duration for the
1325 * application with the weight-raising
1326 * duration for soft rt applications.
1328 * In particular, doing this recharge now, i.e.,
1329 * before the weight-raising period for the
1330 * application finishes, reduces the probability
1331 * of the following negative scenario:
1332 * 1) the weight of a soft rt application is
1333 * raised at startup (as for any newly
1334 * created application),
1335 * 2) since the application is not interactive,
1336 * at a certain time weight-raising is
1337 * stopped for the application,
1338 * 3) at that time the application happens to
1339 * still have pending requests, and hence
1340 * is destined to not have a chance to be
1341 * deemed soft rt before these requests are
1342 * completed (see the comments to the
1343 * function bfq_bfqq_softrt_next_start()
1344 * for details on soft rt detection),
1345 * 4) these pending requests experience a high
1346 * latency because the application is not
1347 * weight-raised while they are pending.
1349 if (bfqq->wr_cur_max_time !=
1350 bfqd->bfq_wr_rt_max_time) {
1351 bfqq->wr_start_at_switch_to_srt =
1352 bfqq->last_wr_start_finish;
1354 bfqq->wr_cur_max_time =
1355 bfqd->bfq_wr_rt_max_time;
1356 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1357 BFQ_SOFTRT_WEIGHT_FACTOR;
1359 bfqq->last_wr_start_finish = jiffies;
1364 static bool bfq_bfqq_idle_for_long_time(struct bfq_data *bfqd,
1365 struct bfq_queue *bfqq)
1367 return bfqq->dispatched == 0 &&
1368 time_is_before_jiffies(
1369 bfqq->budget_timeout +
1370 bfqd->bfq_wr_min_idle_time);
1373 static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data *bfqd,
1374 struct bfq_queue *bfqq,
1379 bool soft_rt, in_burst, wr_or_deserves_wr,
1380 bfqq_wants_to_preempt,
1381 idle_for_long_time = bfq_bfqq_idle_for_long_time(bfqd, bfqq),
1383 * See the comments on
1384 * bfq_bfqq_update_budg_for_activation for
1385 * details on the usage of the next variable.
1387 arrived_in_time = ktime_get_ns() <=
1388 bfqq->ttime.last_end_request +
1389 bfqd->bfq_slice_idle * 3;
1393 * bfqq deserves to be weight-raised if:
1395 * - it does not belong to a large burst,
1396 * - it has been idle for enough time or is soft real-time,
1397 * - is linked to a bfq_io_cq (it is not shared in any sense).
1399 in_burst = bfq_bfqq_in_large_burst(bfqq);
1400 soft_rt = bfqd->bfq_wr_max_softrt_rate > 0 &&
1402 time_is_before_jiffies(bfqq->soft_rt_next_start);
1403 *interactive = !in_burst && idle_for_long_time;
1404 wr_or_deserves_wr = bfqd->low_latency &&
1405 (bfqq->wr_coeff > 1 ||
1406 (bfq_bfqq_sync(bfqq) &&
1407 bfqq->bic && (*interactive || soft_rt)));
1410 * Using the last flag, update budget and check whether bfqq
1411 * may want to preempt the in-service queue.
1413 bfqq_wants_to_preempt =
1414 bfq_bfqq_update_budg_for_activation(bfqd, bfqq,
1419 * If bfqq happened to be activated in a burst, but has been
1420 * idle for much more than an interactive queue, then we
1421 * assume that, in the overall I/O initiated in the burst, the
1422 * I/O associated with bfqq is finished. So bfqq does not need
1423 * to be treated as a queue belonging to a burst
1424 * anymore. Accordingly, we reset bfqq's in_large_burst flag
1425 * if set, and remove bfqq from the burst list if it's
1426 * there. We do not decrement burst_size, because the fact
1427 * that bfqq does not need to belong to the burst list any
1428 * more does not invalidate the fact that bfqq was created in
1431 if (likely(!bfq_bfqq_just_created(bfqq)) &&
1432 idle_for_long_time &&
1433 time_is_before_jiffies(
1434 bfqq->budget_timeout +
1435 msecs_to_jiffies(10000))) {
1436 hlist_del_init(&bfqq->burst_list_node);
1437 bfq_clear_bfqq_in_large_burst(bfqq);
1440 bfq_clear_bfqq_just_created(bfqq);
1443 if (!bfq_bfqq_IO_bound(bfqq)) {
1444 if (arrived_in_time) {
1445 bfqq->requests_within_timer++;
1446 if (bfqq->requests_within_timer >=
1447 bfqd->bfq_requests_within_timer)
1448 bfq_mark_bfqq_IO_bound(bfqq);
1450 bfqq->requests_within_timer = 0;
1453 if (bfqd->low_latency) {
1454 if (unlikely(time_is_after_jiffies(bfqq->split_time)))
1457 jiffies - bfqd->bfq_wr_min_idle_time - 1;
1459 if (time_is_before_jiffies(bfqq->split_time +
1460 bfqd->bfq_wr_min_idle_time)) {
1461 bfq_update_bfqq_wr_on_rq_arrival(bfqd, bfqq,
1468 if (old_wr_coeff != bfqq->wr_coeff)
1469 bfqq->entity.prio_changed = 1;
1473 bfqq->last_idle_bklogged = jiffies;
1474 bfqq->service_from_backlogged = 0;
1475 bfq_clear_bfqq_softrt_update(bfqq);
1477 bfq_add_bfqq_busy(bfqd, bfqq);
1480 * Expire in-service queue only if preemption may be needed
1481 * for guarantees. In this respect, the function
1482 * next_queue_may_preempt just checks a simple, necessary
1483 * condition, and not a sufficient condition based on
1484 * timestamps. In fact, for the latter condition to be
1485 * evaluated, timestamps would need first to be updated, and
1486 * this operation is quite costly (see the comments on the
1487 * function bfq_bfqq_update_budg_for_activation).
1489 if (bfqd->in_service_queue && bfqq_wants_to_preempt &&
1490 bfqd->in_service_queue->wr_coeff < bfqq->wr_coeff &&
1491 next_queue_may_preempt(bfqd))
1492 bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
1493 false, BFQQE_PREEMPTED);
1496 static void bfq_add_request(struct request *rq)
1498 struct bfq_queue *bfqq = RQ_BFQQ(rq);
1499 struct bfq_data *bfqd = bfqq->bfqd;
1500 struct request *next_rq, *prev;
1501 unsigned int old_wr_coeff = bfqq->wr_coeff;
1502 bool interactive = false;
1504 bfq_log_bfqq(bfqd, bfqq, "add_request %d", rq_is_sync(rq));
1505 bfqq->queued[rq_is_sync(rq)]++;
1508 elv_rb_add(&bfqq->sort_list, rq);
1511 * Check if this request is a better next-serve candidate.
1513 prev = bfqq->next_rq;
1514 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position);
1515 bfqq->next_rq = next_rq;
1518 * Adjust priority tree position, if next_rq changes.
1520 if (prev != bfqq->next_rq)
1521 bfq_pos_tree_add_move(bfqd, bfqq);
1523 if (!bfq_bfqq_busy(bfqq)) /* switching to busy ... */
1524 bfq_bfqq_handle_idle_busy_switch(bfqd, bfqq, old_wr_coeff,
1527 if (bfqd->low_latency && old_wr_coeff == 1 && !rq_is_sync(rq) &&
1528 time_is_before_jiffies(
1529 bfqq->last_wr_start_finish +
1530 bfqd->bfq_wr_min_inter_arr_async)) {
1531 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1532 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1534 bfqd->wr_busy_queues++;
1535 bfqq->entity.prio_changed = 1;
1537 if (prev != bfqq->next_rq)
1538 bfq_updated_next_req(bfqd, bfqq);
1542 * Assign jiffies to last_wr_start_finish in the following
1545 * . if bfqq is not going to be weight-raised, because, for
1546 * non weight-raised queues, last_wr_start_finish stores the
1547 * arrival time of the last request; as of now, this piece
1548 * of information is used only for deciding whether to
1549 * weight-raise async queues
1551 * . if bfqq is not weight-raised, because, if bfqq is now
1552 * switching to weight-raised, then last_wr_start_finish
1553 * stores the time when weight-raising starts
1555 * . if bfqq is interactive, because, regardless of whether
1556 * bfqq is currently weight-raised, the weight-raising
1557 * period must start or restart (this case is considered
1558 * separately because it is not detected by the above
1559 * conditions, if bfqq is already weight-raised)
1561 * last_wr_start_finish has to be updated also if bfqq is soft
1562 * real-time, because the weight-raising period is constantly
1563 * restarted on idle-to-busy transitions for these queues, but
1564 * this is already done in bfq_bfqq_handle_idle_busy_switch if
1567 if (bfqd->low_latency &&
1568 (old_wr_coeff == 1 || bfqq->wr_coeff == 1 || interactive))
1569 bfqq->last_wr_start_finish = jiffies;
1572 static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd,
1574 struct request_queue *q)
1576 struct bfq_queue *bfqq = bfqd->bio_bfqq;
1580 return elv_rb_find(&bfqq->sort_list, bio_end_sector(bio));
1585 static sector_t get_sdist(sector_t last_pos, struct request *rq)
1588 return abs(blk_rq_pos(rq) - last_pos);
1593 #if 0 /* Still not clear if we can do without next two functions */
1594 static void bfq_activate_request(struct request_queue *q, struct request *rq)
1596 struct bfq_data *bfqd = q->elevator->elevator_data;
1598 bfqd->rq_in_driver++;
1601 static void bfq_deactivate_request(struct request_queue *q, struct request *rq)
1603 struct bfq_data *bfqd = q->elevator->elevator_data;
1605 bfqd->rq_in_driver--;
1609 static void bfq_remove_request(struct request_queue *q,
1612 struct bfq_queue *bfqq = RQ_BFQQ(rq);
1613 struct bfq_data *bfqd = bfqq->bfqd;
1614 const int sync = rq_is_sync(rq);
1616 if (bfqq->next_rq == rq) {
1617 bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq);
1618 bfq_updated_next_req(bfqd, bfqq);
1621 if (rq->queuelist.prev != &rq->queuelist)
1622 list_del_init(&rq->queuelist);
1623 bfqq->queued[sync]--;
1625 elv_rb_del(&bfqq->sort_list, rq);
1627 elv_rqhash_del(q, rq);
1628 if (q->last_merge == rq)
1629 q->last_merge = NULL;
1631 if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
1632 bfqq->next_rq = NULL;
1634 if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->in_service_queue) {
1635 bfq_del_bfqq_busy(bfqd, bfqq, false);
1637 * bfqq emptied. In normal operation, when
1638 * bfqq is empty, bfqq->entity.service and
1639 * bfqq->entity.budget must contain,
1640 * respectively, the service received and the
1641 * budget used last time bfqq emptied. These
1642 * facts do not hold in this case, as at least
1643 * this last removal occurred while bfqq is
1644 * not in service. To avoid inconsistencies,
1645 * reset both bfqq->entity.service and
1646 * bfqq->entity.budget, if bfqq has still a
1647 * process that may issue I/O requests to it.
1649 bfqq->entity.budget = bfqq->entity.service = 0;
1653 * Remove queue from request-position tree as it is empty.
1655 if (bfqq->pos_root) {
1656 rb_erase(&bfqq->pos_node, bfqq->pos_root);
1657 bfqq->pos_root = NULL;
1660 bfq_pos_tree_add_move(bfqd, bfqq);
1663 if (rq->cmd_flags & REQ_META)
1664 bfqq->meta_pending--;
1668 static bool bfq_bio_merge(struct blk_mq_hw_ctx *hctx, struct bio *bio)
1670 struct request_queue *q = hctx->queue;
1671 struct bfq_data *bfqd = q->elevator->elevator_data;
1672 struct request *free = NULL;
1674 * bfq_bic_lookup grabs the queue_lock: invoke it now and
1675 * store its return value for later use, to avoid nesting
1676 * queue_lock inside the bfqd->lock. We assume that the bic
1677 * returned by bfq_bic_lookup does not go away before
1678 * bfqd->lock is taken.
1680 struct bfq_io_cq *bic = bfq_bic_lookup(bfqd, current->io_context, q);
1683 spin_lock_irq(&bfqd->lock);
1686 bfqd->bio_bfqq = bic_to_bfqq(bic, op_is_sync(bio->bi_opf));
1688 bfqd->bio_bfqq = NULL;
1689 bfqd->bio_bic = bic;
1691 ret = blk_mq_sched_try_merge(q, bio, &free);
1694 blk_mq_free_request(free);
1695 spin_unlock_irq(&bfqd->lock);
1700 static int bfq_request_merge(struct request_queue *q, struct request **req,
1703 struct bfq_data *bfqd = q->elevator->elevator_data;
1704 struct request *__rq;
1706 __rq = bfq_find_rq_fmerge(bfqd, bio, q);
1707 if (__rq && elv_bio_merge_ok(__rq, bio)) {
1709 return ELEVATOR_FRONT_MERGE;
1712 return ELEVATOR_NO_MERGE;
1715 static void bfq_request_merged(struct request_queue *q, struct request *req,
1716 enum elv_merge type)
1718 if (type == ELEVATOR_FRONT_MERGE &&
1719 rb_prev(&req->rb_node) &&
1721 blk_rq_pos(container_of(rb_prev(&req->rb_node),
1722 struct request, rb_node))) {
1723 struct bfq_queue *bfqq = RQ_BFQQ(req);
1724 struct bfq_data *bfqd = bfqq->bfqd;
1725 struct request *prev, *next_rq;
1727 /* Reposition request in its sort_list */
1728 elv_rb_del(&bfqq->sort_list, req);
1729 elv_rb_add(&bfqq->sort_list, req);
1731 /* Choose next request to be served for bfqq */
1732 prev = bfqq->next_rq;
1733 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, req,
1734 bfqd->last_position);
1735 bfqq->next_rq = next_rq;
1737 * If next_rq changes, update both the queue's budget to
1738 * fit the new request and the queue's position in its
1741 if (prev != bfqq->next_rq) {
1742 bfq_updated_next_req(bfqd, bfqq);
1743 bfq_pos_tree_add_move(bfqd, bfqq);
1748 static void bfq_requests_merged(struct request_queue *q, struct request *rq,
1749 struct request *next)
1751 struct bfq_queue *bfqq = RQ_BFQQ(rq), *next_bfqq = RQ_BFQQ(next);
1753 if (!RB_EMPTY_NODE(&rq->rb_node))
1755 spin_lock_irq(&bfqq->bfqd->lock);
1758 * If next and rq belong to the same bfq_queue and next is older
1759 * than rq, then reposition rq in the fifo (by substituting next
1760 * with rq). Otherwise, if next and rq belong to different
1761 * bfq_queues, never reposition rq: in fact, we would have to
1762 * reposition it with respect to next's position in its own fifo,
1763 * which would most certainly be too expensive with respect to
1766 if (bfqq == next_bfqq &&
1767 !list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1768 next->fifo_time < rq->fifo_time) {
1769 list_del_init(&rq->queuelist);
1770 list_replace_init(&next->queuelist, &rq->queuelist);
1771 rq->fifo_time = next->fifo_time;
1774 if (bfqq->next_rq == next)
1777 bfq_remove_request(q, next);
1778 bfqg_stats_update_io_remove(bfqq_group(bfqq), next->cmd_flags);
1780 spin_unlock_irq(&bfqq->bfqd->lock);
1782 bfqg_stats_update_io_merged(bfqq_group(bfqq), next->cmd_flags);
1785 /* Must be called with bfqq != NULL */
1786 static void bfq_bfqq_end_wr(struct bfq_queue *bfqq)
1788 if (bfq_bfqq_busy(bfqq))
1789 bfqq->bfqd->wr_busy_queues--;
1791 bfqq->wr_cur_max_time = 0;
1792 bfqq->last_wr_start_finish = jiffies;
1794 * Trigger a weight change on the next invocation of
1795 * __bfq_entity_update_weight_prio.
1797 bfqq->entity.prio_changed = 1;
1800 void bfq_end_wr_async_queues(struct bfq_data *bfqd,
1801 struct bfq_group *bfqg)
1805 for (i = 0; i < 2; i++)
1806 for (j = 0; j < IOPRIO_BE_NR; j++)
1807 if (bfqg->async_bfqq[i][j])
1808 bfq_bfqq_end_wr(bfqg->async_bfqq[i][j]);
1809 if (bfqg->async_idle_bfqq)
1810 bfq_bfqq_end_wr(bfqg->async_idle_bfqq);
1813 static void bfq_end_wr(struct bfq_data *bfqd)
1815 struct bfq_queue *bfqq;
1817 spin_lock_irq(&bfqd->lock);
1819 list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
1820 bfq_bfqq_end_wr(bfqq);
1821 list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list)
1822 bfq_bfqq_end_wr(bfqq);
1823 bfq_end_wr_async(bfqd);
1825 spin_unlock_irq(&bfqd->lock);
1828 static sector_t bfq_io_struct_pos(void *io_struct, bool request)
1831 return blk_rq_pos(io_struct);
1833 return ((struct bio *)io_struct)->bi_iter.bi_sector;
1836 static int bfq_rq_close_to_sector(void *io_struct, bool request,
1839 return abs(bfq_io_struct_pos(io_struct, request) - sector) <=
1843 static struct bfq_queue *bfqq_find_close(struct bfq_data *bfqd,
1844 struct bfq_queue *bfqq,
1847 struct rb_root *root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree;
1848 struct rb_node *parent, *node;
1849 struct bfq_queue *__bfqq;
1851 if (RB_EMPTY_ROOT(root))
1855 * First, if we find a request starting at the end of the last
1856 * request, choose it.
1858 __bfqq = bfq_rq_pos_tree_lookup(bfqd, root, sector, &parent, NULL);
1863 * If the exact sector wasn't found, the parent of the NULL leaf
1864 * will contain the closest sector (rq_pos_tree sorted by
1865 * next_request position).
1867 __bfqq = rb_entry(parent, struct bfq_queue, pos_node);
1868 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
1871 if (blk_rq_pos(__bfqq->next_rq) < sector)
1872 node = rb_next(&__bfqq->pos_node);
1874 node = rb_prev(&__bfqq->pos_node);
1878 __bfqq = rb_entry(node, struct bfq_queue, pos_node);
1879 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
1885 static struct bfq_queue *bfq_find_close_cooperator(struct bfq_data *bfqd,
1886 struct bfq_queue *cur_bfqq,
1889 struct bfq_queue *bfqq;
1892 * We shall notice if some of the queues are cooperating,
1893 * e.g., working closely on the same area of the device. In
1894 * that case, we can group them together and: 1) don't waste
1895 * time idling, and 2) serve the union of their requests in
1896 * the best possible order for throughput.
1898 bfqq = bfqq_find_close(bfqd, cur_bfqq, sector);
1899 if (!bfqq || bfqq == cur_bfqq)
1905 static struct bfq_queue *
1906 bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
1908 int process_refs, new_process_refs;
1909 struct bfq_queue *__bfqq;
1912 * If there are no process references on the new_bfqq, then it is
1913 * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain
1914 * may have dropped their last reference (not just their last process
1917 if (!bfqq_process_refs(new_bfqq))
1920 /* Avoid a circular list and skip interim queue merges. */
1921 while ((__bfqq = new_bfqq->new_bfqq)) {
1927 process_refs = bfqq_process_refs(bfqq);
1928 new_process_refs = bfqq_process_refs(new_bfqq);
1930 * If the process for the bfqq has gone away, there is no
1931 * sense in merging the queues.
1933 if (process_refs == 0 || new_process_refs == 0)
1936 bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d",
1940 * Merging is just a redirection: the requests of the process
1941 * owning one of the two queues are redirected to the other queue.
1942 * The latter queue, in its turn, is set as shared if this is the
1943 * first time that the requests of some process are redirected to
1946 * We redirect bfqq to new_bfqq and not the opposite, because
1947 * we are in the context of the process owning bfqq, thus we
1948 * have the io_cq of this process. So we can immediately
1949 * configure this io_cq to redirect the requests of the
1950 * process to new_bfqq. In contrast, the io_cq of new_bfqq is
1951 * not available any more (new_bfqq->bic == NULL).
1953 * Anyway, even in case new_bfqq coincides with the in-service
1954 * queue, redirecting requests the in-service queue is the
1955 * best option, as we feed the in-service queue with new
1956 * requests close to the last request served and, by doing so,
1957 * are likely to increase the throughput.
1959 bfqq->new_bfqq = new_bfqq;
1960 new_bfqq->ref += process_refs;
1964 static bool bfq_may_be_close_cooperator(struct bfq_queue *bfqq,
1965 struct bfq_queue *new_bfqq)
1967 if (bfq_too_late_for_merging(new_bfqq))
1970 if (bfq_class_idle(bfqq) || bfq_class_idle(new_bfqq) ||
1971 (bfqq->ioprio_class != new_bfqq->ioprio_class))
1975 * If either of the queues has already been detected as seeky,
1976 * then merging it with the other queue is unlikely to lead to
1979 if (BFQQ_SEEKY(bfqq) || BFQQ_SEEKY(new_bfqq))
1983 * Interleaved I/O is known to be done by (some) applications
1984 * only for reads, so it does not make sense to merge async
1987 if (!bfq_bfqq_sync(bfqq) || !bfq_bfqq_sync(new_bfqq))
1994 * Attempt to schedule a merge of bfqq with the currently in-service
1995 * queue or with a close queue among the scheduled queues. Return
1996 * NULL if no merge was scheduled, a pointer to the shared bfq_queue
1997 * structure otherwise.
1999 * The OOM queue is not allowed to participate to cooperation: in fact, since
2000 * the requests temporarily redirected to the OOM queue could be redirected
2001 * again to dedicated queues at any time, the state needed to correctly
2002 * handle merging with the OOM queue would be quite complex and expensive
2003 * to maintain. Besides, in such a critical condition as an out of memory,
2004 * the benefits of queue merging may be little relevant, or even negligible.
2006 * WARNING: queue merging may impair fairness among non-weight raised
2007 * queues, for at least two reasons: 1) the original weight of a
2008 * merged queue may change during the merged state, 2) even being the
2009 * weight the same, a merged queue may be bloated with many more
2010 * requests than the ones produced by its originally-associated
2013 static struct bfq_queue *
2014 bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2015 void *io_struct, bool request)
2017 struct bfq_queue *in_service_bfqq, *new_bfqq;
2020 * Prevent bfqq from being merged if it has been created too
2021 * long ago. The idea is that true cooperating processes, and
2022 * thus their associated bfq_queues, are supposed to be
2023 * created shortly after each other. This is the case, e.g.,
2024 * for KVM/QEMU and dump I/O threads. Basing on this
2025 * assumption, the following filtering greatly reduces the
2026 * probability that two non-cooperating processes, which just
2027 * happen to do close I/O for some short time interval, have
2028 * their queues merged by mistake.
2030 if (bfq_too_late_for_merging(bfqq))
2034 return bfqq->new_bfqq;
2036 if (!io_struct || unlikely(bfqq == &bfqd->oom_bfqq))
2039 /* If there is only one backlogged queue, don't search. */
2040 if (bfqd->busy_queues == 1)
2043 in_service_bfqq = bfqd->in_service_queue;
2045 if (in_service_bfqq && in_service_bfqq != bfqq &&
2046 likely(in_service_bfqq != &bfqd->oom_bfqq) &&
2047 bfq_rq_close_to_sector(io_struct, request, bfqd->last_position) &&
2048 bfqq->entity.parent == in_service_bfqq->entity.parent &&
2049 bfq_may_be_close_cooperator(bfqq, in_service_bfqq)) {
2050 new_bfqq = bfq_setup_merge(bfqq, in_service_bfqq);
2055 * Check whether there is a cooperator among currently scheduled
2056 * queues. The only thing we need is that the bio/request is not
2057 * NULL, as we need it to establish whether a cooperator exists.
2059 new_bfqq = bfq_find_close_cooperator(bfqd, bfqq,
2060 bfq_io_struct_pos(io_struct, request));
2062 if (new_bfqq && likely(new_bfqq != &bfqd->oom_bfqq) &&
2063 bfq_may_be_close_cooperator(bfqq, new_bfqq))
2064 return bfq_setup_merge(bfqq, new_bfqq);
2069 static void bfq_bfqq_save_state(struct bfq_queue *bfqq)
2071 struct bfq_io_cq *bic = bfqq->bic;
2074 * If !bfqq->bic, the queue is already shared or its requests
2075 * have already been redirected to a shared queue; both idle window
2076 * and weight raising state have already been saved. Do nothing.
2081 bic->saved_ttime = bfqq->ttime;
2082 bic->saved_has_short_ttime = bfq_bfqq_has_short_ttime(bfqq);
2083 bic->saved_IO_bound = bfq_bfqq_IO_bound(bfqq);
2084 bic->saved_in_large_burst = bfq_bfqq_in_large_burst(bfqq);
2085 bic->was_in_burst_list = !hlist_unhashed(&bfqq->burst_list_node);
2086 if (unlikely(bfq_bfqq_just_created(bfqq) &&
2087 !bfq_bfqq_in_large_burst(bfqq) &&
2088 bfqq->bfqd->low_latency)) {
2090 * bfqq being merged right after being created: bfqq
2091 * would have deserved interactive weight raising, but
2092 * did not make it to be set in a weight-raised state,
2093 * because of this early merge. Store directly the
2094 * weight-raising state that would have been assigned
2095 * to bfqq, so that to avoid that bfqq unjustly fails
2096 * to enjoy weight raising if split soon.
2098 bic->saved_wr_coeff = bfqq->bfqd->bfq_wr_coeff;
2099 bic->saved_wr_cur_max_time = bfq_wr_duration(bfqq->bfqd);
2100 bic->saved_last_wr_start_finish = jiffies;
2102 bic->saved_wr_coeff = bfqq->wr_coeff;
2103 bic->saved_wr_start_at_switch_to_srt =
2104 bfqq->wr_start_at_switch_to_srt;
2105 bic->saved_last_wr_start_finish = bfqq->last_wr_start_finish;
2106 bic->saved_wr_cur_max_time = bfqq->wr_cur_max_time;
2111 bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic,
2112 struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2114 bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu",
2115 (unsigned long)new_bfqq->pid);
2116 /* Save weight raising and idle window of the merged queues */
2117 bfq_bfqq_save_state(bfqq);
2118 bfq_bfqq_save_state(new_bfqq);
2119 if (bfq_bfqq_IO_bound(bfqq))
2120 bfq_mark_bfqq_IO_bound(new_bfqq);
2121 bfq_clear_bfqq_IO_bound(bfqq);
2124 * If bfqq is weight-raised, then let new_bfqq inherit
2125 * weight-raising. To reduce false positives, neglect the case
2126 * where bfqq has just been created, but has not yet made it
2127 * to be weight-raised (which may happen because EQM may merge
2128 * bfqq even before bfq_add_request is executed for the first
2129 * time for bfqq). Handling this case would however be very
2130 * easy, thanks to the flag just_created.
2132 if (new_bfqq->wr_coeff == 1 && bfqq->wr_coeff > 1) {
2133 new_bfqq->wr_coeff = bfqq->wr_coeff;
2134 new_bfqq->wr_cur_max_time = bfqq->wr_cur_max_time;
2135 new_bfqq->last_wr_start_finish = bfqq->last_wr_start_finish;
2136 new_bfqq->wr_start_at_switch_to_srt =
2137 bfqq->wr_start_at_switch_to_srt;
2138 if (bfq_bfqq_busy(new_bfqq))
2139 bfqd->wr_busy_queues++;
2140 new_bfqq->entity.prio_changed = 1;
2143 if (bfqq->wr_coeff > 1) { /* bfqq has given its wr to new_bfqq */
2145 bfqq->entity.prio_changed = 1;
2146 if (bfq_bfqq_busy(bfqq))
2147 bfqd->wr_busy_queues--;
2150 bfq_log_bfqq(bfqd, new_bfqq, "merge_bfqqs: wr_busy %d",
2151 bfqd->wr_busy_queues);
2154 * Merge queues (that is, let bic redirect its requests to new_bfqq)
2156 bic_set_bfqq(bic, new_bfqq, 1);
2157 bfq_mark_bfqq_coop(new_bfqq);
2159 * new_bfqq now belongs to at least two bics (it is a shared queue):
2160 * set new_bfqq->bic to NULL. bfqq either:
2161 * - does not belong to any bic any more, and hence bfqq->bic must
2162 * be set to NULL, or
2163 * - is a queue whose owning bics have already been redirected to a
2164 * different queue, hence the queue is destined to not belong to
2165 * any bic soon and bfqq->bic is already NULL (therefore the next
2166 * assignment causes no harm).
2168 new_bfqq->bic = NULL;
2170 /* release process reference to bfqq */
2171 bfq_put_queue(bfqq);
2174 static bool bfq_allow_bio_merge(struct request_queue *q, struct request *rq,
2177 struct bfq_data *bfqd = q->elevator->elevator_data;
2178 bool is_sync = op_is_sync(bio->bi_opf);
2179 struct bfq_queue *bfqq = bfqd->bio_bfqq, *new_bfqq;
2182 * Disallow merge of a sync bio into an async request.
2184 if (is_sync && !rq_is_sync(rq))
2188 * Lookup the bfqq that this bio will be queued with. Allow
2189 * merge only if rq is queued there.
2195 * We take advantage of this function to perform an early merge
2196 * of the queues of possible cooperating processes.
2198 new_bfqq = bfq_setup_cooperator(bfqd, bfqq, bio, false);
2201 * bic still points to bfqq, then it has not yet been
2202 * redirected to some other bfq_queue, and a queue
2203 * merge beween bfqq and new_bfqq can be safely
2204 * fulfillled, i.e., bic can be redirected to new_bfqq
2205 * and bfqq can be put.
2207 bfq_merge_bfqqs(bfqd, bfqd->bio_bic, bfqq,
2210 * If we get here, bio will be queued into new_queue,
2211 * so use new_bfqq to decide whether bio and rq can be
2217 * Change also bqfd->bio_bfqq, as
2218 * bfqd->bio_bic now points to new_bfqq, and
2219 * this function may be invoked again (and then may
2220 * use again bqfd->bio_bfqq).
2222 bfqd->bio_bfqq = bfqq;
2225 return bfqq == RQ_BFQQ(rq);
2229 * Set the maximum time for the in-service queue to consume its
2230 * budget. This prevents seeky processes from lowering the throughput.
2231 * In practice, a time-slice service scheme is used with seeky
2234 static void bfq_set_budget_timeout(struct bfq_data *bfqd,
2235 struct bfq_queue *bfqq)
2237 unsigned int timeout_coeff;
2239 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time)
2242 timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight;
2244 bfqd->last_budget_start = ktime_get();
2246 bfqq->budget_timeout = jiffies +
2247 bfqd->bfq_timeout * timeout_coeff;
2250 static void __bfq_set_in_service_queue(struct bfq_data *bfqd,
2251 struct bfq_queue *bfqq)
2254 bfq_clear_bfqq_fifo_expire(bfqq);
2256 bfqd->budgets_assigned = (bfqd->budgets_assigned * 7 + 256) / 8;
2258 if (time_is_before_jiffies(bfqq->last_wr_start_finish) &&
2259 bfqq->wr_coeff > 1 &&
2260 bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
2261 time_is_before_jiffies(bfqq->budget_timeout)) {
2263 * For soft real-time queues, move the start
2264 * of the weight-raising period forward by the
2265 * time the queue has not received any
2266 * service. Otherwise, a relatively long
2267 * service delay is likely to cause the
2268 * weight-raising period of the queue to end,
2269 * because of the short duration of the
2270 * weight-raising period of a soft real-time
2271 * queue. It is worth noting that this move
2272 * is not so dangerous for the other queues,
2273 * because soft real-time queues are not
2276 * To not add a further variable, we use the
2277 * overloaded field budget_timeout to
2278 * determine for how long the queue has not
2279 * received service, i.e., how much time has
2280 * elapsed since the queue expired. However,
2281 * this is a little imprecise, because
2282 * budget_timeout is set to jiffies if bfqq
2283 * not only expires, but also remains with no
2286 if (time_after(bfqq->budget_timeout,
2287 bfqq->last_wr_start_finish))
2288 bfqq->last_wr_start_finish +=
2289 jiffies - bfqq->budget_timeout;
2291 bfqq->last_wr_start_finish = jiffies;
2294 bfq_set_budget_timeout(bfqd, bfqq);
2295 bfq_log_bfqq(bfqd, bfqq,
2296 "set_in_service_queue, cur-budget = %d",
2297 bfqq->entity.budget);
2300 bfqd->in_service_queue = bfqq;
2304 * Get and set a new queue for service.
2306 static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd)
2308 struct bfq_queue *bfqq = bfq_get_next_queue(bfqd);
2310 __bfq_set_in_service_queue(bfqd, bfqq);
2314 static void bfq_arm_slice_timer(struct bfq_data *bfqd)
2316 struct bfq_queue *bfqq = bfqd->in_service_queue;
2319 bfq_mark_bfqq_wait_request(bfqq);
2322 * We don't want to idle for seeks, but we do want to allow
2323 * fair distribution of slice time for a process doing back-to-back
2324 * seeks. So allow a little bit of time for him to submit a new rq.
2326 sl = bfqd->bfq_slice_idle;
2328 * Unless the queue is being weight-raised or the scenario is
2329 * asymmetric, grant only minimum idle time if the queue
2330 * is seeky. A long idling is preserved for a weight-raised
2331 * queue, or, more in general, in an asymmetric scenario,
2332 * because a long idling is needed for guaranteeing to a queue
2333 * its reserved share of the throughput (in particular, it is
2334 * needed if the queue has a higher weight than some other
2337 if (BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 &&
2338 bfq_symmetric_scenario(bfqd))
2339 sl = min_t(u64, sl, BFQ_MIN_TT);
2341 bfqd->last_idling_start = ktime_get();
2342 hrtimer_start(&bfqd->idle_slice_timer, ns_to_ktime(sl),
2344 bfqg_stats_set_start_idle_time(bfqq_group(bfqq));
2348 * In autotuning mode, max_budget is dynamically recomputed as the
2349 * amount of sectors transferred in timeout at the estimated peak
2350 * rate. This enables BFQ to utilize a full timeslice with a full
2351 * budget, even if the in-service queue is served at peak rate. And
2352 * this maximises throughput with sequential workloads.
2354 static unsigned long bfq_calc_max_budget(struct bfq_data *bfqd)
2356 return (u64)bfqd->peak_rate * USEC_PER_MSEC *
2357 jiffies_to_msecs(bfqd->bfq_timeout)>>BFQ_RATE_SHIFT;
2361 * Update parameters related to throughput and responsiveness, as a
2362 * function of the estimated peak rate. See comments on
2363 * bfq_calc_max_budget(), and on T_slow and T_fast arrays.
2365 static void update_thr_responsiveness_params(struct bfq_data *bfqd)
2367 int dev_type = blk_queue_nonrot(bfqd->queue);
2369 if (bfqd->bfq_user_max_budget == 0)
2370 bfqd->bfq_max_budget =
2371 bfq_calc_max_budget(bfqd);
2373 if (bfqd->device_speed == BFQ_BFQD_FAST &&
2374 bfqd->peak_rate < device_speed_thresh[dev_type]) {
2375 bfqd->device_speed = BFQ_BFQD_SLOW;
2376 bfqd->RT_prod = R_slow[dev_type] *
2378 } else if (bfqd->device_speed == BFQ_BFQD_SLOW &&
2379 bfqd->peak_rate > device_speed_thresh[dev_type]) {
2380 bfqd->device_speed = BFQ_BFQD_FAST;
2381 bfqd->RT_prod = R_fast[dev_type] *
2386 "dev_type %s dev_speed_class = %s (%llu sects/sec), thresh %llu setcs/sec",
2387 dev_type == 0 ? "ROT" : "NONROT",
2388 bfqd->device_speed == BFQ_BFQD_FAST ? "FAST" : "SLOW",
2389 bfqd->device_speed == BFQ_BFQD_FAST ?
2390 (USEC_PER_SEC*(u64)R_fast[dev_type])>>BFQ_RATE_SHIFT :
2391 (USEC_PER_SEC*(u64)R_slow[dev_type])>>BFQ_RATE_SHIFT,
2392 (USEC_PER_SEC*(u64)device_speed_thresh[dev_type])>>
2396 static void bfq_reset_rate_computation(struct bfq_data *bfqd,
2399 if (rq != NULL) { /* new rq dispatch now, reset accordingly */
2400 bfqd->last_dispatch = bfqd->first_dispatch = ktime_get_ns();
2401 bfqd->peak_rate_samples = 1;
2402 bfqd->sequential_samples = 0;
2403 bfqd->tot_sectors_dispatched = bfqd->last_rq_max_size =
2405 } else /* no new rq dispatched, just reset the number of samples */
2406 bfqd->peak_rate_samples = 0; /* full re-init on next disp. */
2409 "reset_rate_computation at end, sample %u/%u tot_sects %llu",
2410 bfqd->peak_rate_samples, bfqd->sequential_samples,
2411 bfqd->tot_sectors_dispatched);
2414 static void bfq_update_rate_reset(struct bfq_data *bfqd, struct request *rq)
2416 u32 rate, weight, divisor;
2419 * For the convergence property to hold (see comments on
2420 * bfq_update_peak_rate()) and for the assessment to be
2421 * reliable, a minimum number of samples must be present, and
2422 * a minimum amount of time must have elapsed. If not so, do
2423 * not compute new rate. Just reset parameters, to get ready
2424 * for a new evaluation attempt.
2426 if (bfqd->peak_rate_samples < BFQ_RATE_MIN_SAMPLES ||
2427 bfqd->delta_from_first < BFQ_RATE_MIN_INTERVAL)
2428 goto reset_computation;
2431 * If a new request completion has occurred after last
2432 * dispatch, then, to approximate the rate at which requests
2433 * have been served by the device, it is more precise to
2434 * extend the observation interval to the last completion.
2436 bfqd->delta_from_first =
2437 max_t(u64, bfqd->delta_from_first,
2438 bfqd->last_completion - bfqd->first_dispatch);
2441 * Rate computed in sects/usec, and not sects/nsec, for
2444 rate = div64_ul(bfqd->tot_sectors_dispatched<<BFQ_RATE_SHIFT,
2445 div_u64(bfqd->delta_from_first, NSEC_PER_USEC));
2448 * Peak rate not updated if:
2449 * - the percentage of sequential dispatches is below 3/4 of the
2450 * total, and rate is below the current estimated peak rate
2451 * - rate is unreasonably high (> 20M sectors/sec)
2453 if ((bfqd->sequential_samples < (3 * bfqd->peak_rate_samples)>>2 &&
2454 rate <= bfqd->peak_rate) ||
2455 rate > 20<<BFQ_RATE_SHIFT)
2456 goto reset_computation;
2459 * We have to update the peak rate, at last! To this purpose,
2460 * we use a low-pass filter. We compute the smoothing constant
2461 * of the filter as a function of the 'weight' of the new
2464 * As can be seen in next formulas, we define this weight as a
2465 * quantity proportional to how sequential the workload is,
2466 * and to how long the observation time interval is.
2468 * The weight runs from 0 to 8. The maximum value of the
2469 * weight, 8, yields the minimum value for the smoothing
2470 * constant. At this minimum value for the smoothing constant,
2471 * the measured rate contributes for half of the next value of
2472 * the estimated peak rate.
2474 * So, the first step is to compute the weight as a function
2475 * of how sequential the workload is. Note that the weight
2476 * cannot reach 9, because bfqd->sequential_samples cannot
2477 * become equal to bfqd->peak_rate_samples, which, in its
2478 * turn, holds true because bfqd->sequential_samples is not
2479 * incremented for the first sample.
2481 weight = (9 * bfqd->sequential_samples) / bfqd->peak_rate_samples;
2484 * Second step: further refine the weight as a function of the
2485 * duration of the observation interval.
2487 weight = min_t(u32, 8,
2488 div_u64(weight * bfqd->delta_from_first,
2489 BFQ_RATE_REF_INTERVAL));
2492 * Divisor ranging from 10, for minimum weight, to 2, for
2495 divisor = 10 - weight;
2498 * Finally, update peak rate:
2500 * peak_rate = peak_rate * (divisor-1) / divisor + rate / divisor
2502 bfqd->peak_rate *= divisor-1;
2503 bfqd->peak_rate /= divisor;
2504 rate /= divisor; /* smoothing constant alpha = 1/divisor */
2506 bfqd->peak_rate += rate;
2507 update_thr_responsiveness_params(bfqd);
2510 bfq_reset_rate_computation(bfqd, rq);
2514 * Update the read/write peak rate (the main quantity used for
2515 * auto-tuning, see update_thr_responsiveness_params()).
2517 * It is not trivial to estimate the peak rate (correctly): because of
2518 * the presence of sw and hw queues between the scheduler and the
2519 * device components that finally serve I/O requests, it is hard to
2520 * say exactly when a given dispatched request is served inside the
2521 * device, and for how long. As a consequence, it is hard to know
2522 * precisely at what rate a given set of requests is actually served
2525 * On the opposite end, the dispatch time of any request is trivially
2526 * available, and, from this piece of information, the "dispatch rate"
2527 * of requests can be immediately computed. So, the idea in the next
2528 * function is to use what is known, namely request dispatch times
2529 * (plus, when useful, request completion times), to estimate what is
2530 * unknown, namely in-device request service rate.
2532 * The main issue is that, because of the above facts, the rate at
2533 * which a certain set of requests is dispatched over a certain time
2534 * interval can vary greatly with respect to the rate at which the
2535 * same requests are then served. But, since the size of any
2536 * intermediate queue is limited, and the service scheme is lossless
2537 * (no request is silently dropped), the following obvious convergence
2538 * property holds: the number of requests dispatched MUST become
2539 * closer and closer to the number of requests completed as the
2540 * observation interval grows. This is the key property used in
2541 * the next function to estimate the peak service rate as a function
2542 * of the observed dispatch rate. The function assumes to be invoked
2543 * on every request dispatch.
2545 static void bfq_update_peak_rate(struct bfq_data *bfqd, struct request *rq)
2547 u64 now_ns = ktime_get_ns();
2549 if (bfqd->peak_rate_samples == 0) { /* first dispatch */
2550 bfq_log(bfqd, "update_peak_rate: goto reset, samples %d",
2551 bfqd->peak_rate_samples);
2552 bfq_reset_rate_computation(bfqd, rq);
2553 goto update_last_values; /* will add one sample */
2557 * Device idle for very long: the observation interval lasting
2558 * up to this dispatch cannot be a valid observation interval
2559 * for computing a new peak rate (similarly to the late-
2560 * completion event in bfq_completed_request()). Go to
2561 * update_rate_and_reset to have the following three steps
2563 * - close the observation interval at the last (previous)
2564 * request dispatch or completion
2565 * - compute rate, if possible, for that observation interval
2566 * - start a new observation interval with this dispatch
2568 if (now_ns - bfqd->last_dispatch > 100*NSEC_PER_MSEC &&
2569 bfqd->rq_in_driver == 0)
2570 goto update_rate_and_reset;
2572 /* Update sampling information */
2573 bfqd->peak_rate_samples++;
2575 if ((bfqd->rq_in_driver > 0 ||
2576 now_ns - bfqd->last_completion < BFQ_MIN_TT)
2577 && get_sdist(bfqd->last_position, rq) < BFQQ_SEEK_THR)
2578 bfqd->sequential_samples++;
2580 bfqd->tot_sectors_dispatched += blk_rq_sectors(rq);
2582 /* Reset max observed rq size every 32 dispatches */
2583 if (likely(bfqd->peak_rate_samples % 32))
2584 bfqd->last_rq_max_size =
2585 max_t(u32, blk_rq_sectors(rq), bfqd->last_rq_max_size);
2587 bfqd->last_rq_max_size = blk_rq_sectors(rq);
2589 bfqd->delta_from_first = now_ns - bfqd->first_dispatch;
2591 /* Target observation interval not yet reached, go on sampling */
2592 if (bfqd->delta_from_first < BFQ_RATE_REF_INTERVAL)
2593 goto update_last_values;
2595 update_rate_and_reset:
2596 bfq_update_rate_reset(bfqd, rq);
2598 bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2599 bfqd->last_dispatch = now_ns;
2603 * Remove request from internal lists.
2605 static void bfq_dispatch_remove(struct request_queue *q, struct request *rq)
2607 struct bfq_queue *bfqq = RQ_BFQQ(rq);
2610 * For consistency, the next instruction should have been
2611 * executed after removing the request from the queue and
2612 * dispatching it. We execute instead this instruction before
2613 * bfq_remove_request() (and hence introduce a temporary
2614 * inconsistency), for efficiency. In fact, should this
2615 * dispatch occur for a non in-service bfqq, this anticipated
2616 * increment prevents two counters related to bfqq->dispatched
2617 * from risking to be, first, uselessly decremented, and then
2618 * incremented again when the (new) value of bfqq->dispatched
2619 * happens to be taken into account.
2622 bfq_update_peak_rate(q->elevator->elevator_data, rq);
2624 bfq_remove_request(q, rq);
2627 static void __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq)
2630 * If this bfqq is shared between multiple processes, check
2631 * to make sure that those processes are still issuing I/Os
2632 * within the mean seek distance. If not, it may be time to
2633 * break the queues apart again.
2635 if (bfq_bfqq_coop(bfqq) && BFQQ_SEEKY(bfqq))
2636 bfq_mark_bfqq_split_coop(bfqq);
2638 if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
2639 if (bfqq->dispatched == 0)
2641 * Overloading budget_timeout field to store
2642 * the time at which the queue remains with no
2643 * backlog and no outstanding request; used by
2644 * the weight-raising mechanism.
2646 bfqq->budget_timeout = jiffies;
2648 bfq_del_bfqq_busy(bfqd, bfqq, true);
2650 bfq_requeue_bfqq(bfqd, bfqq, true);
2652 * Resort priority tree of potential close cooperators.
2654 bfq_pos_tree_add_move(bfqd, bfqq);
2658 * All in-service entities must have been properly deactivated
2659 * or requeued before executing the next function, which
2660 * resets all in-service entites as no more in service.
2662 __bfq_bfqd_reset_in_service(bfqd);
2666 * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior.
2667 * @bfqd: device data.
2668 * @bfqq: queue to update.
2669 * @reason: reason for expiration.
2671 * Handle the feedback on @bfqq budget at queue expiration.
2672 * See the body for detailed comments.
2674 static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd,
2675 struct bfq_queue *bfqq,
2676 enum bfqq_expiration reason)
2678 struct request *next_rq;
2679 int budget, min_budget;
2681 min_budget = bfq_min_budget(bfqd);
2683 if (bfqq->wr_coeff == 1)
2684 budget = bfqq->max_budget;
2686 * Use a constant, low budget for weight-raised queues,
2687 * to help achieve a low latency. Keep it slightly higher
2688 * than the minimum possible budget, to cause a little
2689 * bit fewer expirations.
2691 budget = 2 * min_budget;
2693 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %d, budg left %d",
2694 bfqq->entity.budget, bfq_bfqq_budget_left(bfqq));
2695 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %d, min budg %d",
2696 budget, bfq_min_budget(bfqd));
2697 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d",
2698 bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->in_service_queue));
2700 if (bfq_bfqq_sync(bfqq) && bfqq->wr_coeff == 1) {
2703 * Caveat: in all the following cases we trade latency
2706 case BFQQE_TOO_IDLE:
2708 * This is the only case where we may reduce
2709 * the budget: if there is no request of the
2710 * process still waiting for completion, then
2711 * we assume (tentatively) that the timer has
2712 * expired because the batch of requests of
2713 * the process could have been served with a
2714 * smaller budget. Hence, betting that
2715 * process will behave in the same way when it
2716 * becomes backlogged again, we reduce its
2717 * next budget. As long as we guess right,
2718 * this budget cut reduces the latency
2719 * experienced by the process.
2721 * However, if there are still outstanding
2722 * requests, then the process may have not yet
2723 * issued its next request just because it is
2724 * still waiting for the completion of some of
2725 * the still outstanding ones. So in this
2726 * subcase we do not reduce its budget, on the
2727 * contrary we increase it to possibly boost
2728 * the throughput, as discussed in the
2729 * comments to the BUDGET_TIMEOUT case.
2731 if (bfqq->dispatched > 0) /* still outstanding reqs */
2732 budget = min(budget * 2, bfqd->bfq_max_budget);
2734 if (budget > 5 * min_budget)
2735 budget -= 4 * min_budget;
2737 budget = min_budget;
2740 case BFQQE_BUDGET_TIMEOUT:
2742 * We double the budget here because it gives
2743 * the chance to boost the throughput if this
2744 * is not a seeky process (and has bumped into
2745 * this timeout because of, e.g., ZBR).
2747 budget = min(budget * 2, bfqd->bfq_max_budget);
2749 case BFQQE_BUDGET_EXHAUSTED:
2751 * The process still has backlog, and did not
2752 * let either the budget timeout or the disk
2753 * idling timeout expire. Hence it is not
2754 * seeky, has a short thinktime and may be
2755 * happy with a higher budget too. So
2756 * definitely increase the budget of this good
2757 * candidate to boost the disk throughput.
2759 budget = min(budget * 4, bfqd->bfq_max_budget);
2761 case BFQQE_NO_MORE_REQUESTS:
2763 * For queues that expire for this reason, it
2764 * is particularly important to keep the
2765 * budget close to the actual service they
2766 * need. Doing so reduces the timestamp
2767 * misalignment problem described in the
2768 * comments in the body of
2769 * __bfq_activate_entity. In fact, suppose
2770 * that a queue systematically expires for
2771 * BFQQE_NO_MORE_REQUESTS and presents a
2772 * new request in time to enjoy timestamp
2773 * back-shifting. The larger the budget of the
2774 * queue is with respect to the service the
2775 * queue actually requests in each service
2776 * slot, the more times the queue can be
2777 * reactivated with the same virtual finish
2778 * time. It follows that, even if this finish
2779 * time is pushed to the system virtual time
2780 * to reduce the consequent timestamp
2781 * misalignment, the queue unjustly enjoys for
2782 * many re-activations a lower finish time
2783 * than all newly activated queues.
2785 * The service needed by bfqq is measured
2786 * quite precisely by bfqq->entity.service.
2787 * Since bfqq does not enjoy device idling,
2788 * bfqq->entity.service is equal to the number
2789 * of sectors that the process associated with
2790 * bfqq requested to read/write before waiting
2791 * for request completions, or blocking for
2794 budget = max_t(int, bfqq->entity.service, min_budget);
2799 } else if (!bfq_bfqq_sync(bfqq)) {
2801 * Async queues get always the maximum possible
2802 * budget, as for them we do not care about latency
2803 * (in addition, their ability to dispatch is limited
2804 * by the charging factor).
2806 budget = bfqd->bfq_max_budget;
2809 bfqq->max_budget = budget;
2811 if (bfqd->budgets_assigned >= bfq_stats_min_budgets &&
2812 !bfqd->bfq_user_max_budget)
2813 bfqq->max_budget = min(bfqq->max_budget, bfqd->bfq_max_budget);
2816 * If there is still backlog, then assign a new budget, making
2817 * sure that it is large enough for the next request. Since
2818 * the finish time of bfqq must be kept in sync with the
2819 * budget, be sure to call __bfq_bfqq_expire() *after* this
2822 * If there is no backlog, then no need to update the budget;
2823 * it will be updated on the arrival of a new request.
2825 next_rq = bfqq->next_rq;
2827 bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget,
2828 bfq_serv_to_charge(next_rq, bfqq));
2830 bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %d",
2831 next_rq ? blk_rq_sectors(next_rq) : 0,
2832 bfqq->entity.budget);
2836 * Return true if the process associated with bfqq is "slow". The slow
2837 * flag is used, in addition to the budget timeout, to reduce the
2838 * amount of service provided to seeky processes, and thus reduce
2839 * their chances to lower the throughput. More details in the comments
2840 * on the function bfq_bfqq_expire().
2842 * An important observation is in order: as discussed in the comments
2843 * on the function bfq_update_peak_rate(), with devices with internal
2844 * queues, it is hard if ever possible to know when and for how long
2845 * an I/O request is processed by the device (apart from the trivial
2846 * I/O pattern where a new request is dispatched only after the
2847 * previous one has been completed). This makes it hard to evaluate
2848 * the real rate at which the I/O requests of each bfq_queue are
2849 * served. In fact, for an I/O scheduler like BFQ, serving a
2850 * bfq_queue means just dispatching its requests during its service
2851 * slot (i.e., until the budget of the queue is exhausted, or the
2852 * queue remains idle, or, finally, a timeout fires). But, during the
2853 * service slot of a bfq_queue, around 100 ms at most, the device may
2854 * be even still processing requests of bfq_queues served in previous
2855 * service slots. On the opposite end, the requests of the in-service
2856 * bfq_queue may be completed after the service slot of the queue
2859 * Anyway, unless more sophisticated solutions are used
2860 * (where possible), the sum of the sizes of the requests dispatched
2861 * during the service slot of a bfq_queue is probably the only
2862 * approximation available for the service received by the bfq_queue
2863 * during its service slot. And this sum is the quantity used in this
2864 * function to evaluate the I/O speed of a process.
2866 static bool bfq_bfqq_is_slow(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2867 bool compensate, enum bfqq_expiration reason,
2868 unsigned long *delta_ms)
2870 ktime_t delta_ktime;
2872 bool slow = BFQQ_SEEKY(bfqq); /* if delta too short, use seekyness */
2874 if (!bfq_bfqq_sync(bfqq))
2878 delta_ktime = bfqd->last_idling_start;
2880 delta_ktime = ktime_get();
2881 delta_ktime = ktime_sub(delta_ktime, bfqd->last_budget_start);
2882 delta_usecs = ktime_to_us(delta_ktime);
2884 /* don't use too short time intervals */
2885 if (delta_usecs < 1000) {
2886 if (blk_queue_nonrot(bfqd->queue))
2888 * give same worst-case guarantees as idling
2891 *delta_ms = BFQ_MIN_TT / NSEC_PER_MSEC;
2892 else /* charge at least one seek */
2893 *delta_ms = bfq_slice_idle / NSEC_PER_MSEC;
2898 *delta_ms = delta_usecs / USEC_PER_MSEC;
2901 * Use only long (> 20ms) intervals to filter out excessive
2902 * spikes in service rate estimation.
2904 if (delta_usecs > 20000) {
2906 * Caveat for rotational devices: processes doing I/O
2907 * in the slower disk zones tend to be slow(er) even
2908 * if not seeky. In this respect, the estimated peak
2909 * rate is likely to be an average over the disk
2910 * surface. Accordingly, to not be too harsh with
2911 * unlucky processes, a process is deemed slow only if
2912 * its rate has been lower than half of the estimated
2915 slow = bfqq->entity.service < bfqd->bfq_max_budget / 2;
2918 bfq_log_bfqq(bfqd, bfqq, "bfq_bfqq_is_slow: slow %d", slow);
2924 * To be deemed as soft real-time, an application must meet two
2925 * requirements. First, the application must not require an average
2926 * bandwidth higher than the approximate bandwidth required to playback or
2927 * record a compressed high-definition video.
2928 * The next function is invoked on the completion of the last request of a
2929 * batch, to compute the next-start time instant, soft_rt_next_start, such
2930 * that, if the next request of the application does not arrive before
2931 * soft_rt_next_start, then the above requirement on the bandwidth is met.
2933 * The second requirement is that the request pattern of the application is
2934 * isochronous, i.e., that, after issuing a request or a batch of requests,
2935 * the application stops issuing new requests until all its pending requests
2936 * have been completed. After that, the application may issue a new batch,
2938 * For this reason the next function is invoked to compute
2939 * soft_rt_next_start only for applications that meet this requirement,
2940 * whereas soft_rt_next_start is set to infinity for applications that do
2943 * Unfortunately, even a greedy (i.e., I/O-bound) application may
2944 * happen to meet, occasionally or systematically, both the above
2945 * bandwidth and isochrony requirements. This may happen at least in
2946 * the following circumstances. First, if the CPU load is high. The
2947 * application may stop issuing requests while the CPUs are busy
2948 * serving other processes, then restart, then stop again for a while,
2949 * and so on. The other circumstances are related to the storage
2950 * device: the storage device is highly loaded or reaches a low-enough
2951 * throughput with the I/O of the application (e.g., because the I/O
2952 * is random and/or the device is slow). In all these cases, the
2953 * I/O of the application may be simply slowed down enough to meet
2954 * the bandwidth and isochrony requirements. To reduce the probability
2955 * that greedy applications are deemed as soft real-time in these
2956 * corner cases, a further rule is used in the computation of
2957 * soft_rt_next_start: the return value of this function is forced to
2958 * be higher than the maximum between the following two quantities.
2960 * (a) Current time plus: (1) the maximum time for which the arrival
2961 * of a request is waited for when a sync queue becomes idle,
2962 * namely bfqd->bfq_slice_idle, and (2) a few extra jiffies. We
2963 * postpone for a moment the reason for adding a few extra
2964 * jiffies; we get back to it after next item (b). Lower-bounding
2965 * the return value of this function with the current time plus
2966 * bfqd->bfq_slice_idle tends to filter out greedy applications,
2967 * because the latter issue their next request as soon as possible
2968 * after the last one has been completed. In contrast, a soft
2969 * real-time application spends some time processing data, after a
2970 * batch of its requests has been completed.
2972 * (b) Current value of bfqq->soft_rt_next_start. As pointed out
2973 * above, greedy applications may happen to meet both the
2974 * bandwidth and isochrony requirements under heavy CPU or
2975 * storage-device load. In more detail, in these scenarios, these
2976 * applications happen, only for limited time periods, to do I/O
2977 * slowly enough to meet all the requirements described so far,
2978 * including the filtering in above item (a). These slow-speed
2979 * time intervals are usually interspersed between other time
2980 * intervals during which these applications do I/O at a very high
2981 * speed. Fortunately, exactly because of the high speed of the
2982 * I/O in the high-speed intervals, the values returned by this
2983 * function happen to be so high, near the end of any such
2984 * high-speed interval, to be likely to fall *after* the end of
2985 * the low-speed time interval that follows. These high values are
2986 * stored in bfqq->soft_rt_next_start after each invocation of
2987 * this function. As a consequence, if the last value of
2988 * bfqq->soft_rt_next_start is constantly used to lower-bound the
2989 * next value that this function may return, then, from the very
2990 * beginning of a low-speed interval, bfqq->soft_rt_next_start is
2991 * likely to be constantly kept so high that any I/O request
2992 * issued during the low-speed interval is considered as arriving
2993 * to soon for the application to be deemed as soft
2994 * real-time. Then, in the high-speed interval that follows, the
2995 * application will not be deemed as soft real-time, just because
2996 * it will do I/O at a high speed. And so on.
2998 * Getting back to the filtering in item (a), in the following two
2999 * cases this filtering might be easily passed by a greedy
3000 * application, if the reference quantity was just
3001 * bfqd->bfq_slice_idle:
3002 * 1) HZ is so low that the duration of a jiffy is comparable to or
3003 * higher than bfqd->bfq_slice_idle. This happens, e.g., on slow
3004 * devices with HZ=100. The time granularity may be so coarse
3005 * that the approximation, in jiffies, of bfqd->bfq_slice_idle
3006 * is rather lower than the exact value.
3007 * 2) jiffies, instead of increasing at a constant rate, may stop increasing
3008 * for a while, then suddenly 'jump' by several units to recover the lost
3009 * increments. This seems to happen, e.g., inside virtual machines.
3010 * To address this issue, in the filtering in (a) we do not use as a
3011 * reference time interval just bfqd->bfq_slice_idle, but
3012 * bfqd->bfq_slice_idle plus a few jiffies. In particular, we add the
3013 * minimum number of jiffies for which the filter seems to be quite
3014 * precise also in embedded systems and KVM/QEMU virtual machines.
3016 static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data *bfqd,
3017 struct bfq_queue *bfqq)
3019 return max3(bfqq->soft_rt_next_start,
3020 bfqq->last_idle_bklogged +
3021 HZ * bfqq->service_from_backlogged /
3022 bfqd->bfq_wr_max_softrt_rate,
3023 jiffies + nsecs_to_jiffies(bfqq->bfqd->bfq_slice_idle) + 4);
3027 * bfq_bfqq_expire - expire a queue.
3028 * @bfqd: device owning the queue.
3029 * @bfqq: the queue to expire.
3030 * @compensate: if true, compensate for the time spent idling.
3031 * @reason: the reason causing the expiration.
3033 * If the process associated with bfqq does slow I/O (e.g., because it
3034 * issues random requests), we charge bfqq with the time it has been
3035 * in service instead of the service it has received (see
3036 * bfq_bfqq_charge_time for details on how this goal is achieved). As
3037 * a consequence, bfqq will typically get higher timestamps upon
3038 * reactivation, and hence it will be rescheduled as if it had
3039 * received more service than what it has actually received. In the
3040 * end, bfqq receives less service in proportion to how slowly its
3041 * associated process consumes its budgets (and hence how seriously it
3042 * tends to lower the throughput). In addition, this time-charging
3043 * strategy guarantees time fairness among slow processes. In
3044 * contrast, if the process associated with bfqq is not slow, we
3045 * charge bfqq exactly with the service it has received.
3047 * Charging time to the first type of queues and the exact service to
3048 * the other has the effect of using the WF2Q+ policy to schedule the
3049 * former on a timeslice basis, without violating service domain
3050 * guarantees among the latter.
3052 void bfq_bfqq_expire(struct bfq_data *bfqd,
3053 struct bfq_queue *bfqq,
3055 enum bfqq_expiration reason)
3058 unsigned long delta = 0;
3059 struct bfq_entity *entity = &bfqq->entity;
3063 * Check whether the process is slow (see bfq_bfqq_is_slow).
3065 slow = bfq_bfqq_is_slow(bfqd, bfqq, compensate, reason, &delta);
3068 * As above explained, charge slow (typically seeky) and
3069 * timed-out queues with the time and not the service
3070 * received, to favor sequential workloads.
3072 * Processes doing I/O in the slower disk zones will tend to
3073 * be slow(er) even if not seeky. Therefore, since the
3074 * estimated peak rate is actually an average over the disk
3075 * surface, these processes may timeout just for bad luck. To
3076 * avoid punishing them, do not charge time to processes that
3077 * succeeded in consuming at least 2/3 of their budget. This
3078 * allows BFQ to preserve enough elasticity to still perform
3079 * bandwidth, and not time, distribution with little unlucky
3080 * or quasi-sequential processes.
3082 if (bfqq->wr_coeff == 1 &&
3084 (reason == BFQQE_BUDGET_TIMEOUT &&
3085 bfq_bfqq_budget_left(bfqq) >= entity->budget / 3)))
3086 bfq_bfqq_charge_time(bfqd, bfqq, delta);
3088 if (reason == BFQQE_TOO_IDLE &&
3089 entity->service <= 2 * entity->budget / 10)
3090 bfq_clear_bfqq_IO_bound(bfqq);
3092 if (bfqd->low_latency && bfqq->wr_coeff == 1)
3093 bfqq->last_wr_start_finish = jiffies;
3095 if (bfqd->low_latency && bfqd->bfq_wr_max_softrt_rate > 0 &&
3096 RB_EMPTY_ROOT(&bfqq->sort_list)) {
3098 * If we get here, and there are no outstanding
3099 * requests, then the request pattern is isochronous
3100 * (see the comments on the function
3101 * bfq_bfqq_softrt_next_start()). Thus we can compute
3102 * soft_rt_next_start. If, instead, the queue still
3103 * has outstanding requests, then we have to wait for
3104 * the completion of all the outstanding requests to
3105 * discover whether the request pattern is actually
3108 if (bfqq->dispatched == 0)
3109 bfqq->soft_rt_next_start =
3110 bfq_bfqq_softrt_next_start(bfqd, bfqq);
3113 * The application is still waiting for the
3114 * completion of one or more requests:
3115 * prevent it from possibly being incorrectly
3116 * deemed as soft real-time by setting its
3117 * soft_rt_next_start to infinity. In fact,
3118 * without this assignment, the application
3119 * would be incorrectly deemed as soft
3121 * 1) it issued a new request before the
3122 * completion of all its in-flight
3124 * 2) at that time, its soft_rt_next_start
3125 * happened to be in the past.
3127 bfqq->soft_rt_next_start =
3128 bfq_greatest_from_now();
3130 * Schedule an update of soft_rt_next_start to when
3131 * the task may be discovered to be isochronous.
3133 bfq_mark_bfqq_softrt_update(bfqq);
3137 bfq_log_bfqq(bfqd, bfqq,
3138 "expire (%d, slow %d, num_disp %d, short_ttime %d)", reason,
3139 slow, bfqq->dispatched, bfq_bfqq_has_short_ttime(bfqq));
3142 * Increase, decrease or leave budget unchanged according to
3145 __bfq_bfqq_recalc_budget(bfqd, bfqq, reason);
3147 __bfq_bfqq_expire(bfqd, bfqq);
3149 /* mark bfqq as waiting a request only if a bic still points to it */
3150 if (ref > 1 && !bfq_bfqq_busy(bfqq) &&
3151 reason != BFQQE_BUDGET_TIMEOUT &&
3152 reason != BFQQE_BUDGET_EXHAUSTED)
3153 bfq_mark_bfqq_non_blocking_wait_rq(bfqq);
3157 * Budget timeout is not implemented through a dedicated timer, but
3158 * just checked on request arrivals and completions, as well as on
3159 * idle timer expirations.
3161 static bool bfq_bfqq_budget_timeout(struct bfq_queue *bfqq)
3163 return time_is_before_eq_jiffies(bfqq->budget_timeout);
3167 * If we expire a queue that is actively waiting (i.e., with the
3168 * device idled) for the arrival of a new request, then we may incur
3169 * the timestamp misalignment problem described in the body of the
3170 * function __bfq_activate_entity. Hence we return true only if this
3171 * condition does not hold, or if the queue is slow enough to deserve
3172 * only to be kicked off for preserving a high throughput.
3174 static bool bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq)
3176 bfq_log_bfqq(bfqq->bfqd, bfqq,
3177 "may_budget_timeout: wait_request %d left %d timeout %d",
3178 bfq_bfqq_wait_request(bfqq),
3179 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3,
3180 bfq_bfqq_budget_timeout(bfqq));
3182 return (!bfq_bfqq_wait_request(bfqq) ||
3183 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3)
3185 bfq_bfqq_budget_timeout(bfqq);
3189 * For a queue that becomes empty, device idling is allowed only if
3190 * this function returns true for the queue. As a consequence, since
3191 * device idling plays a critical role in both throughput boosting and
3192 * service guarantees, the return value of this function plays a
3193 * critical role in both these aspects as well.
3195 * In a nutshell, this function returns true only if idling is
3196 * beneficial for throughput or, even if detrimental for throughput,
3197 * idling is however necessary to preserve service guarantees (low
3198 * latency, desired throughput distribution, ...). In particular, on
3199 * NCQ-capable devices, this function tries to return false, so as to
3200 * help keep the drives' internal queues full, whenever this helps the
3201 * device boost the throughput without causing any service-guarantee
3204 * In more detail, the return value of this function is obtained by,
3205 * first, computing a number of boolean variables that take into
3206 * account throughput and service-guarantee issues, and, then,
3207 * combining these variables in a logical expression. Most of the
3208 * issues taken into account are not trivial. We discuss these issues
3209 * individually while introducing the variables.
3211 static bool bfq_bfqq_may_idle(struct bfq_queue *bfqq)
3213 struct bfq_data *bfqd = bfqq->bfqd;
3214 bool rot_without_queueing =
3215 !blk_queue_nonrot(bfqd->queue) && !bfqd->hw_tag,
3216 bfqq_sequential_and_IO_bound,
3217 idling_boosts_thr, idling_boosts_thr_without_issues,
3218 idling_needed_for_service_guarantees,
3219 asymmetric_scenario;
3221 if (bfqd->strict_guarantees)
3225 * Idling is performed only if slice_idle > 0. In addition, we
3228 * (b) bfqq is in the idle io prio class: in this case we do
3229 * not idle because we want to minimize the bandwidth that
3230 * queues in this class can steal to higher-priority queues
3232 if (bfqd->bfq_slice_idle == 0 || !bfq_bfqq_sync(bfqq) ||
3233 bfq_class_idle(bfqq))
3236 bfqq_sequential_and_IO_bound = !BFQQ_SEEKY(bfqq) &&
3237 bfq_bfqq_IO_bound(bfqq) && bfq_bfqq_has_short_ttime(bfqq);
3240 * The next variable takes into account the cases where idling
3241 * boosts the throughput.
3243 * The value of the variable is computed considering, first, that
3244 * idling is virtually always beneficial for the throughput if:
3245 * (a) the device is not NCQ-capable and rotational, or
3246 * (b) regardless of the presence of NCQ, the device is rotational and
3247 * the request pattern for bfqq is I/O-bound and sequential, or
3248 * (c) regardless of whether it is rotational, the device is
3249 * not NCQ-capable and the request pattern for bfqq is
3250 * I/O-bound and sequential.
3252 * Secondly, and in contrast to the above item (b), idling an
3253 * NCQ-capable flash-based device would not boost the
3254 * throughput even with sequential I/O; rather it would lower
3255 * the throughput in proportion to how fast the device
3256 * is. Accordingly, the next variable is true if any of the
3257 * above conditions (a), (b) or (c) is true, and, in
3258 * particular, happens to be false if bfqd is an NCQ-capable
3259 * flash-based device.
3261 idling_boosts_thr = rot_without_queueing ||
3262 ((!blk_queue_nonrot(bfqd->queue) || !bfqd->hw_tag) &&
3263 bfqq_sequential_and_IO_bound);
3266 * The value of the next variable,
3267 * idling_boosts_thr_without_issues, is equal to that of
3268 * idling_boosts_thr, unless a special case holds. In this
3269 * special case, described below, idling may cause problems to
3270 * weight-raised queues.
3272 * When the request pool is saturated (e.g., in the presence
3273 * of write hogs), if the processes associated with
3274 * non-weight-raised queues ask for requests at a lower rate,
3275 * then processes associated with weight-raised queues have a
3276 * higher probability to get a request from the pool
3277 * immediately (or at least soon) when they need one. Thus
3278 * they have a higher probability to actually get a fraction
3279 * of the device throughput proportional to their high
3280 * weight. This is especially true with NCQ-capable drives,
3281 * which enqueue several requests in advance, and further
3282 * reorder internally-queued requests.
3284 * For this reason, we force to false the value of
3285 * idling_boosts_thr_without_issues if there are weight-raised
3286 * busy queues. In this case, and if bfqq is not weight-raised,
3287 * this guarantees that the device is not idled for bfqq (if,
3288 * instead, bfqq is weight-raised, then idling will be
3289 * guaranteed by another variable, see below). Combined with
3290 * the timestamping rules of BFQ (see [1] for details), this
3291 * behavior causes bfqq, and hence any sync non-weight-raised
3292 * queue, to get a lower number of requests served, and thus
3293 * to ask for a lower number of requests from the request
3294 * pool, before the busy weight-raised queues get served
3295 * again. This often mitigates starvation problems in the
3296 * presence of heavy write workloads and NCQ, thereby
3297 * guaranteeing a higher application and system responsiveness
3298 * in these hostile scenarios.
3300 idling_boosts_thr_without_issues = idling_boosts_thr &&
3301 bfqd->wr_busy_queues == 0;
3304 * There is then a case where idling must be performed not
3305 * for throughput concerns, but to preserve service
3308 * To introduce this case, we can note that allowing the drive
3309 * to enqueue more than one request at a time, and hence
3310 * delegating de facto final scheduling decisions to the
3311 * drive's internal scheduler, entails loss of control on the
3312 * actual request service order. In particular, the critical
3313 * situation is when requests from different processes happen
3314 * to be present, at the same time, in the internal queue(s)
3315 * of the drive. In such a situation, the drive, by deciding
3316 * the service order of the internally-queued requests, does
3317 * determine also the actual throughput distribution among
3318 * these processes. But the drive typically has no notion or
3319 * concern about per-process throughput distribution, and
3320 * makes its decisions only on a per-request basis. Therefore,
3321 * the service distribution enforced by the drive's internal
3322 * scheduler is likely to coincide with the desired
3323 * device-throughput distribution only in a completely
3324 * symmetric scenario where:
3325 * (i) each of these processes must get the same throughput as
3327 * (ii) all these processes have the same I/O pattern
3328 (either sequential or random).
3329 * In fact, in such a scenario, the drive will tend to treat
3330 * the requests of each of these processes in about the same
3331 * way as the requests of the others, and thus to provide
3332 * each of these processes with about the same throughput
3333 * (which is exactly the desired throughput distribution). In
3334 * contrast, in any asymmetric scenario, device idling is
3335 * certainly needed to guarantee that bfqq receives its
3336 * assigned fraction of the device throughput (see [1] for
3339 * We address this issue by controlling, actually, only the
3340 * symmetry sub-condition (i), i.e., provided that
3341 * sub-condition (i) holds, idling is not performed,
3342 * regardless of whether sub-condition (ii) holds. In other
3343 * words, only if sub-condition (i) holds, then idling is
3344 * allowed, and the device tends to be prevented from queueing
3345 * many requests, possibly of several processes. The reason
3346 * for not controlling also sub-condition (ii) is that we
3347 * exploit preemption to preserve guarantees in case of
3348 * symmetric scenarios, even if (ii) does not hold, as
3349 * explained in the next two paragraphs.
3351 * Even if a queue, say Q, is expired when it remains idle, Q
3352 * can still preempt the new in-service queue if the next
3353 * request of Q arrives soon (see the comments on
3354 * bfq_bfqq_update_budg_for_activation). If all queues and
3355 * groups have the same weight, this form of preemption,
3356 * combined with the hole-recovery heuristic described in the
3357 * comments on function bfq_bfqq_update_budg_for_activation,
3358 * are enough to preserve a correct bandwidth distribution in
3359 * the mid term, even without idling. In fact, even if not
3360 * idling allows the internal queues of the device to contain
3361 * many requests, and thus to reorder requests, we can rather
3362 * safely assume that the internal scheduler still preserves a
3363 * minimum of mid-term fairness. The motivation for using
3364 * preemption instead of idling is that, by not idling,
3365 * service guarantees are preserved without minimally
3366 * sacrificing throughput. In other words, both a high
3367 * throughput and its desired distribution are obtained.
3369 * More precisely, this preemption-based, idleless approach
3370 * provides fairness in terms of IOPS, and not sectors per
3371 * second. This can be seen with a simple example. Suppose
3372 * that there are two queues with the same weight, but that
3373 * the first queue receives requests of 8 sectors, while the
3374 * second queue receives requests of 1024 sectors. In
3375 * addition, suppose that each of the two queues contains at
3376 * most one request at a time, which implies that each queue
3377 * always remains idle after it is served. Finally, after
3378 * remaining idle, each queue receives very quickly a new
3379 * request. It follows that the two queues are served
3380 * alternatively, preempting each other if needed. This
3381 * implies that, although both queues have the same weight,
3382 * the queue with large requests receives a service that is
3383 * 1024/8 times as high as the service received by the other
3386 * On the other hand, device idling is performed, and thus
3387 * pure sector-domain guarantees are provided, for the
3388 * following queues, which are likely to need stronger
3389 * throughput guarantees: weight-raised queues, and queues
3390 * with a higher weight than other queues. When such queues
3391 * are active, sub-condition (i) is false, which triggers
3394 * According to the above considerations, the next variable is
3395 * true (only) if sub-condition (i) holds. To compute the
3396 * value of this variable, we not only use the return value of
3397 * the function bfq_symmetric_scenario(), but also check
3398 * whether bfqq is being weight-raised, because
3399 * bfq_symmetric_scenario() does not take into account also
3400 * weight-raised queues (see comments on
3401 * bfq_weights_tree_add()).
3403 * As a side note, it is worth considering that the above
3404 * device-idling countermeasures may however fail in the
3405 * following unlucky scenario: if idling is (correctly)
3406 * disabled in a time period during which all symmetry
3407 * sub-conditions hold, and hence the device is allowed to
3408 * enqueue many requests, but at some later point in time some
3409 * sub-condition stops to hold, then it may become impossible
3410 * to let requests be served in the desired order until all
3411 * the requests already queued in the device have been served.
3413 asymmetric_scenario = bfqq->wr_coeff > 1 ||
3414 !bfq_symmetric_scenario(bfqd);
3417 * Finally, there is a case where maximizing throughput is the
3418 * best choice even if it may cause unfairness toward
3419 * bfqq. Such a case is when bfqq became active in a burst of
3420 * queue activations. Queues that became active during a large
3421 * burst benefit only from throughput, as discussed in the
3422 * comments on bfq_handle_burst. Thus, if bfqq became active
3423 * in a burst and not idling the device maximizes throughput,
3424 * then the device must no be idled, because not idling the
3425 * device provides bfqq and all other queues in the burst with
3426 * maximum benefit. Combining this and the above case, we can
3427 * now establish when idling is actually needed to preserve
3428 * service guarantees.
3430 idling_needed_for_service_guarantees =
3431 asymmetric_scenario && !bfq_bfqq_in_large_burst(bfqq);
3434 * We have now all the components we need to compute the
3435 * return value of the function, which is true only if idling
3436 * either boosts the throughput (without issues), or is
3437 * necessary to preserve service guarantees.
3439 return idling_boosts_thr_without_issues ||
3440 idling_needed_for_service_guarantees;
3444 * If the in-service queue is empty but the function bfq_bfqq_may_idle
3445 * returns true, then:
3446 * 1) the queue must remain in service and cannot be expired, and
3447 * 2) the device must be idled to wait for the possible arrival of a new
3448 * request for the queue.
3449 * See the comments on the function bfq_bfqq_may_idle for the reasons
3450 * why performing device idling is the best choice to boost the throughput
3451 * and preserve service guarantees when bfq_bfqq_may_idle itself
3454 static bool bfq_bfqq_must_idle(struct bfq_queue *bfqq)
3456 return RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_bfqq_may_idle(bfqq);
3460 * Select a queue for service. If we have a current queue in service,
3461 * check whether to continue servicing it, or retrieve and set a new one.
3463 static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd)
3465 struct bfq_queue *bfqq;
3466 struct request *next_rq;
3467 enum bfqq_expiration reason = BFQQE_BUDGET_TIMEOUT;
3469 bfqq = bfqd->in_service_queue;
3473 bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue");
3475 if (bfq_may_expire_for_budg_timeout(bfqq) &&
3476 !bfq_bfqq_wait_request(bfqq) &&
3477 !bfq_bfqq_must_idle(bfqq))
3482 * This loop is rarely executed more than once. Even when it
3483 * happens, it is much more convenient to re-execute this loop
3484 * than to return NULL and trigger a new dispatch to get a
3487 next_rq = bfqq->next_rq;
3489 * If bfqq has requests queued and it has enough budget left to
3490 * serve them, keep the queue, otherwise expire it.
3493 if (bfq_serv_to_charge(next_rq, bfqq) >
3494 bfq_bfqq_budget_left(bfqq)) {
3496 * Expire the queue for budget exhaustion,
3497 * which makes sure that the next budget is
3498 * enough to serve the next request, even if
3499 * it comes from the fifo expired path.
3501 reason = BFQQE_BUDGET_EXHAUSTED;
3505 * The idle timer may be pending because we may
3506 * not disable disk idling even when a new request
3509 if (bfq_bfqq_wait_request(bfqq)) {
3511 * If we get here: 1) at least a new request
3512 * has arrived but we have not disabled the
3513 * timer because the request was too small,
3514 * 2) then the block layer has unplugged
3515 * the device, causing the dispatch to be
3518 * Since the device is unplugged, now the
3519 * requests are probably large enough to
3520 * provide a reasonable throughput.
3521 * So we disable idling.
3523 bfq_clear_bfqq_wait_request(bfqq);
3524 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
3531 * No requests pending. However, if the in-service queue is idling
3532 * for a new request, or has requests waiting for a completion and
3533 * may idle after their completion, then keep it anyway.
3535 if (bfq_bfqq_wait_request(bfqq) ||
3536 (bfqq->dispatched != 0 && bfq_bfqq_may_idle(bfqq))) {
3541 reason = BFQQE_NO_MORE_REQUESTS;
3543 bfq_bfqq_expire(bfqd, bfqq, false, reason);
3545 bfqq = bfq_set_in_service_queue(bfqd);
3547 bfq_log_bfqq(bfqd, bfqq, "select_queue: checking new queue");
3552 bfq_log_bfqq(bfqd, bfqq, "select_queue: returned this queue");
3554 bfq_log(bfqd, "select_queue: no queue returned");
3559 static void bfq_update_wr_data(struct bfq_data *bfqd, struct bfq_queue *bfqq)
3561 struct bfq_entity *entity = &bfqq->entity;
3563 if (bfqq->wr_coeff > 1) { /* queue is being weight-raised */
3564 bfq_log_bfqq(bfqd, bfqq,
3565 "raising period dur %u/%u msec, old coeff %u, w %d(%d)",
3566 jiffies_to_msecs(jiffies - bfqq->last_wr_start_finish),
3567 jiffies_to_msecs(bfqq->wr_cur_max_time),
3569 bfqq->entity.weight, bfqq->entity.orig_weight);
3571 if (entity->prio_changed)
3572 bfq_log_bfqq(bfqd, bfqq, "WARN: pending prio change");
3575 * If the queue was activated in a burst, or too much
3576 * time has elapsed from the beginning of this
3577 * weight-raising period, then end weight raising.
3579 if (bfq_bfqq_in_large_burst(bfqq))
3580 bfq_bfqq_end_wr(bfqq);
3581 else if (time_is_before_jiffies(bfqq->last_wr_start_finish +
3582 bfqq->wr_cur_max_time)) {
3583 if (bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time ||
3584 time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
3585 bfq_wr_duration(bfqd)))
3586 bfq_bfqq_end_wr(bfqq);
3588 switch_back_to_interactive_wr(bfqq, bfqd);
3589 bfqq->entity.prio_changed = 1;
3594 * To improve latency (for this or other queues), immediately
3595 * update weight both if it must be raised and if it must be
3596 * lowered. Since, entity may be on some active tree here, and
3597 * might have a pending change of its ioprio class, invoke
3598 * next function with the last parameter unset (see the
3599 * comments on the function).
3601 if ((entity->weight > entity->orig_weight) != (bfqq->wr_coeff > 1))
3602 __bfq_entity_update_weight_prio(bfq_entity_service_tree(entity),
3607 * Dispatch next request from bfqq.
3609 static struct request *bfq_dispatch_rq_from_bfqq(struct bfq_data *bfqd,
3610 struct bfq_queue *bfqq)
3612 struct request *rq = bfqq->next_rq;
3613 unsigned long service_to_charge;
3615 service_to_charge = bfq_serv_to_charge(rq, bfqq);
3617 bfq_bfqq_served(bfqq, service_to_charge);
3619 bfq_dispatch_remove(bfqd->queue, rq);
3622 * If weight raising has to terminate for bfqq, then next
3623 * function causes an immediate update of bfqq's weight,
3624 * without waiting for next activation. As a consequence, on
3625 * expiration, bfqq will be timestamped as if has never been
3626 * weight-raised during this service slot, even if it has
3627 * received part or even most of the service as a
3628 * weight-raised queue. This inflates bfqq's timestamps, which
3629 * is beneficial, as bfqq is then more willing to leave the
3630 * device immediately to possible other weight-raised queues.
3632 bfq_update_wr_data(bfqd, bfqq);
3635 * Expire bfqq, pretending that its budget expired, if bfqq
3636 * belongs to CLASS_IDLE and other queues are waiting for
3639 if (bfqd->busy_queues > 1 && bfq_class_idle(bfqq))
3645 bfq_bfqq_expire(bfqd, bfqq, false, BFQQE_BUDGET_EXHAUSTED);
3649 static bool bfq_has_work(struct blk_mq_hw_ctx *hctx)
3651 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
3654 * Avoiding lock: a race on bfqd->busy_queues should cause at
3655 * most a call to dispatch for nothing
3657 return !list_empty_careful(&bfqd->dispatch) ||
3658 bfqd->busy_queues > 0;
3661 static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
3663 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
3664 struct request *rq = NULL;
3665 struct bfq_queue *bfqq = NULL;
3667 if (!list_empty(&bfqd->dispatch)) {
3668 rq = list_first_entry(&bfqd->dispatch, struct request,
3670 list_del_init(&rq->queuelist);
3676 * Increment counters here, because this
3677 * dispatch does not follow the standard
3678 * dispatch flow (where counters are
3683 goto inc_in_driver_start_rq;
3687 * We exploit the bfq_finish_request hook to decrement
3688 * rq_in_driver, but bfq_finish_request will not be
3689 * invoked on this request. So, to avoid unbalance,
3690 * just start this request, without incrementing
3691 * rq_in_driver. As a negative consequence,
3692 * rq_in_driver is deceptively lower than it should be
3693 * while this request is in service. This may cause
3694 * bfq_schedule_dispatch to be invoked uselessly.
3696 * As for implementing an exact solution, the
3697 * bfq_finish_request hook, if defined, is probably
3698 * invoked also on this request. So, by exploiting
3699 * this hook, we could 1) increment rq_in_driver here,
3700 * and 2) decrement it in bfq_finish_request. Such a
3701 * solution would let the value of the counter be
3702 * always accurate, but it would entail using an extra
3703 * interface function. This cost seems higher than the
3704 * benefit, being the frequency of non-elevator-private
3705 * requests very low.
3710 bfq_log(bfqd, "dispatch requests: %d busy queues", bfqd->busy_queues);
3712 if (bfqd->busy_queues == 0)
3716 * Force device to serve one request at a time if
3717 * strict_guarantees is true. Forcing this service scheme is
3718 * currently the ONLY way to guarantee that the request
3719 * service order enforced by the scheduler is respected by a
3720 * queueing device. Otherwise the device is free even to make
3721 * some unlucky request wait for as long as the device
3724 * Of course, serving one request at at time may cause loss of
3727 if (bfqd->strict_guarantees && bfqd->rq_in_driver > 0)
3730 bfqq = bfq_select_queue(bfqd);
3734 rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq);
3737 inc_in_driver_start_rq:
3738 bfqd->rq_in_driver++;
3740 rq->rq_flags |= RQF_STARTED;
3746 #if defined(CONFIG_BFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
3747 static void bfq_update_dispatch_stats(struct request_queue *q,
3749 struct bfq_queue *in_serv_queue,
3750 bool idle_timer_disabled)
3752 struct bfq_queue *bfqq = rq ? RQ_BFQQ(rq) : NULL;
3754 if (!idle_timer_disabled && !bfqq)
3758 * rq and bfqq are guaranteed to exist until this function
3759 * ends, for the following reasons. First, rq can be
3760 * dispatched to the device, and then can be completed and
3761 * freed, only after this function ends. Second, rq cannot be
3762 * merged (and thus freed because of a merge) any longer,
3763 * because it has already started. Thus rq cannot be freed
3764 * before this function ends, and, since rq has a reference to
3765 * bfqq, the same guarantee holds for bfqq too.
3767 * In addition, the following queue lock guarantees that
3768 * bfqq_group(bfqq) exists as well.
3770 spin_lock_irq(q->queue_lock);
3771 if (idle_timer_disabled)
3773 * Since the idle timer has been disabled,
3774 * in_serv_queue contained some request when
3775 * __bfq_dispatch_request was invoked above, which
3776 * implies that rq was picked exactly from
3777 * in_serv_queue. Thus in_serv_queue == bfqq, and is
3778 * therefore guaranteed to exist because of the above
3781 bfqg_stats_update_idle_time(bfqq_group(in_serv_queue));
3783 struct bfq_group *bfqg = bfqq_group(bfqq);
3785 bfqg_stats_update_avg_queue_size(bfqg);
3786 bfqg_stats_set_start_empty_time(bfqg);
3787 bfqg_stats_update_io_remove(bfqg, rq->cmd_flags);
3789 spin_unlock_irq(q->queue_lock);
3792 static inline void bfq_update_dispatch_stats(struct request_queue *q,
3794 struct bfq_queue *in_serv_queue,
3795 bool idle_timer_disabled) {}
3798 static struct request *bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
3800 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
3802 struct bfq_queue *in_serv_queue;
3803 bool waiting_rq, idle_timer_disabled;
3805 spin_lock_irq(&bfqd->lock);
3807 in_serv_queue = bfqd->in_service_queue;
3808 waiting_rq = in_serv_queue && bfq_bfqq_wait_request(in_serv_queue);
3810 rq = __bfq_dispatch_request(hctx);
3812 idle_timer_disabled =
3813 waiting_rq && !bfq_bfqq_wait_request(in_serv_queue);
3815 spin_unlock_irq(&bfqd->lock);
3817 bfq_update_dispatch_stats(hctx->queue, rq, in_serv_queue,
3818 idle_timer_disabled);
3824 * Task holds one reference to the queue, dropped when task exits. Each rq
3825 * in-flight on this queue also holds a reference, dropped when rq is freed.
3827 * Scheduler lock must be held here. Recall not to use bfqq after calling
3828 * this function on it.
3830 void bfq_put_queue(struct bfq_queue *bfqq)
3832 #ifdef CONFIG_BFQ_GROUP_IOSCHED
3833 struct bfq_group *bfqg = bfqq_group(bfqq);
3837 bfq_log_bfqq(bfqq->bfqd, bfqq, "put_queue: %p %d",
3844 if (!hlist_unhashed(&bfqq->burst_list_node)) {
3845 hlist_del_init(&bfqq->burst_list_node);
3847 * Decrement also burst size after the removal, if the
3848 * process associated with bfqq is exiting, and thus
3849 * does not contribute to the burst any longer. This
3850 * decrement helps filter out false positives of large
3851 * bursts, when some short-lived process (often due to
3852 * the execution of commands by some service) happens
3853 * to start and exit while a complex application is
3854 * starting, and thus spawning several processes that
3855 * do I/O (and that *must not* be treated as a large
3856 * burst, see comments on bfq_handle_burst).
3858 * In particular, the decrement is performed only if:
3859 * 1) bfqq is not a merged queue, because, if it is,
3860 * then this free of bfqq is not triggered by the exit
3861 * of the process bfqq is associated with, but exactly
3862 * by the fact that bfqq has just been merged.
3863 * 2) burst_size is greater than 0, to handle
3864 * unbalanced decrements. Unbalanced decrements may
3865 * happen in te following case: bfqq is inserted into
3866 * the current burst list--without incrementing
3867 * bust_size--because of a split, but the current
3868 * burst list is not the burst list bfqq belonged to
3869 * (see comments on the case of a split in
3872 if (bfqq->bic && bfqq->bfqd->burst_size > 0)
3873 bfqq->bfqd->burst_size--;
3876 kmem_cache_free(bfq_pool, bfqq);
3877 #ifdef CONFIG_BFQ_GROUP_IOSCHED
3878 bfqg_and_blkg_put(bfqg);
3882 static void bfq_put_cooperator(struct bfq_queue *bfqq)
3884 struct bfq_queue *__bfqq, *next;
3887 * If this queue was scheduled to merge with another queue, be
3888 * sure to drop the reference taken on that queue (and others in
3889 * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs.
3891 __bfqq = bfqq->new_bfqq;
3895 next = __bfqq->new_bfqq;
3896 bfq_put_queue(__bfqq);
3901 static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq)
3903 if (bfqq == bfqd->in_service_queue) {
3904 __bfq_bfqq_expire(bfqd, bfqq);
3905 bfq_schedule_dispatch(bfqd);
3908 bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, bfqq->ref);
3910 bfq_put_cooperator(bfqq);
3912 bfq_put_queue(bfqq); /* release process reference */
3915 static void bfq_exit_icq_bfqq(struct bfq_io_cq *bic, bool is_sync)
3917 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
3918 struct bfq_data *bfqd;
3921 bfqd = bfqq->bfqd; /* NULL if scheduler already exited */
3924 unsigned long flags;
3926 spin_lock_irqsave(&bfqd->lock, flags);
3927 bfq_exit_bfqq(bfqd, bfqq);
3928 bic_set_bfqq(bic, NULL, is_sync);
3929 spin_unlock_irqrestore(&bfqd->lock, flags);
3933 static void bfq_exit_icq(struct io_cq *icq)
3935 struct bfq_io_cq *bic = icq_to_bic(icq);
3937 bfq_exit_icq_bfqq(bic, true);
3938 bfq_exit_icq_bfqq(bic, false);
3942 * Update the entity prio values; note that the new values will not
3943 * be used until the next (re)activation.
3946 bfq_set_next_ioprio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic)
3948 struct task_struct *tsk = current;
3950 struct bfq_data *bfqd = bfqq->bfqd;
3955 ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
3956 switch (ioprio_class) {
3958 dev_err(bfqq->bfqd->queue->backing_dev_info->dev,
3959 "bfq: bad prio class %d\n", ioprio_class);
3961 case IOPRIO_CLASS_NONE:
3963 * No prio set, inherit CPU scheduling settings.
3965 bfqq->new_ioprio = task_nice_ioprio(tsk);
3966 bfqq->new_ioprio_class = task_nice_ioclass(tsk);
3968 case IOPRIO_CLASS_RT:
3969 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
3970 bfqq->new_ioprio_class = IOPRIO_CLASS_RT;
3972 case IOPRIO_CLASS_BE:
3973 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
3974 bfqq->new_ioprio_class = IOPRIO_CLASS_BE;
3976 case IOPRIO_CLASS_IDLE:
3977 bfqq->new_ioprio_class = IOPRIO_CLASS_IDLE;
3978 bfqq->new_ioprio = 7;
3982 if (bfqq->new_ioprio >= IOPRIO_BE_NR) {
3983 pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n",
3985 bfqq->new_ioprio = IOPRIO_BE_NR;
3988 bfqq->entity.new_weight = bfq_ioprio_to_weight(bfqq->new_ioprio);
3989 bfqq->entity.prio_changed = 1;
3992 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
3993 struct bio *bio, bool is_sync,
3994 struct bfq_io_cq *bic);
3996 static void bfq_check_ioprio_change(struct bfq_io_cq *bic, struct bio *bio)
3998 struct bfq_data *bfqd = bic_to_bfqd(bic);
3999 struct bfq_queue *bfqq;
4000 int ioprio = bic->icq.ioc->ioprio;
4003 * This condition may trigger on a newly created bic, be sure to
4004 * drop the lock before returning.
4006 if (unlikely(!bfqd) || likely(bic->ioprio == ioprio))
4009 bic->ioprio = ioprio;
4011 bfqq = bic_to_bfqq(bic, false);
4013 /* release process reference on this queue */
4014 bfq_put_queue(bfqq);
4015 bfqq = bfq_get_queue(bfqd, bio, BLK_RW_ASYNC, bic);
4016 bic_set_bfqq(bic, bfqq, false);
4019 bfqq = bic_to_bfqq(bic, true);
4021 bfq_set_next_ioprio_data(bfqq, bic);
4024 static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4025 struct bfq_io_cq *bic, pid_t pid, int is_sync)
4027 RB_CLEAR_NODE(&bfqq->entity.rb_node);
4028 INIT_LIST_HEAD(&bfqq->fifo);
4029 INIT_HLIST_NODE(&bfqq->burst_list_node);
4035 bfq_set_next_ioprio_data(bfqq, bic);
4039 * No need to mark as has_short_ttime if in
4040 * idle_class, because no device idling is performed
4041 * for queues in idle class
4043 if (!bfq_class_idle(bfqq))
4044 /* tentatively mark as has_short_ttime */
4045 bfq_mark_bfqq_has_short_ttime(bfqq);
4046 bfq_mark_bfqq_sync(bfqq);
4047 bfq_mark_bfqq_just_created(bfqq);
4049 bfq_clear_bfqq_sync(bfqq);
4051 /* set end request to minus infinity from now */
4052 bfqq->ttime.last_end_request = ktime_get_ns() + 1;
4054 bfq_mark_bfqq_IO_bound(bfqq);
4058 /* Tentative initial value to trade off between thr and lat */
4059 bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3;
4060 bfqq->budget_timeout = bfq_smallest_from_now();
4063 bfqq->last_wr_start_finish = jiffies;
4064 bfqq->wr_start_at_switch_to_srt = bfq_smallest_from_now();
4065 bfqq->split_time = bfq_smallest_from_now();
4068 * To not forget the possibly high bandwidth consumed by a
4069 * process/queue in the recent past,
4070 * bfq_bfqq_softrt_next_start() returns a value at least equal
4071 * to the current value of bfqq->soft_rt_next_start (see
4072 * comments on bfq_bfqq_softrt_next_start). Set
4073 * soft_rt_next_start to now, to mean that bfqq has consumed
4074 * no bandwidth so far.
4076 bfqq->soft_rt_next_start = jiffies;
4078 /* first request is almost certainly seeky */
4079 bfqq->seek_history = 1;
4082 static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd,
4083 struct bfq_group *bfqg,
4084 int ioprio_class, int ioprio)
4086 switch (ioprio_class) {
4087 case IOPRIO_CLASS_RT:
4088 return &bfqg->async_bfqq[0][ioprio];
4089 case IOPRIO_CLASS_NONE:
4090 ioprio = IOPRIO_NORM;
4092 case IOPRIO_CLASS_BE:
4093 return &bfqg->async_bfqq[1][ioprio];
4094 case IOPRIO_CLASS_IDLE:
4095 return &bfqg->async_idle_bfqq;
4101 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
4102 struct bio *bio, bool is_sync,
4103 struct bfq_io_cq *bic)
4105 const int ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
4106 const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
4107 struct bfq_queue **async_bfqq = NULL;
4108 struct bfq_queue *bfqq;
4109 struct bfq_group *bfqg;
4113 bfqg = bfq_find_set_group(bfqd, bio_blkcg(bio));
4115 bfqq = &bfqd->oom_bfqq;
4120 async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class,
4127 bfqq = kmem_cache_alloc_node(bfq_pool,
4128 GFP_NOWAIT | __GFP_ZERO | __GFP_NOWARN,
4132 bfq_init_bfqq(bfqd, bfqq, bic, current->pid,
4134 bfq_init_entity(&bfqq->entity, bfqg);
4135 bfq_log_bfqq(bfqd, bfqq, "allocated");
4137 bfqq = &bfqd->oom_bfqq;
4138 bfq_log_bfqq(bfqd, bfqq, "using oom bfqq");
4143 * Pin the queue now that it's allocated, scheduler exit will
4148 * Extra group reference, w.r.t. sync
4149 * queue. This extra reference is removed
4150 * only if bfqq->bfqg disappears, to
4151 * guarantee that this queue is not freed
4152 * until its group goes away.
4154 bfq_log_bfqq(bfqd, bfqq, "get_queue, bfqq not in async: %p, %d",
4160 bfqq->ref++; /* get a process reference to this queue */
4161 bfq_log_bfqq(bfqd, bfqq, "get_queue, at end: %p, %d", bfqq, bfqq->ref);
4166 static void bfq_update_io_thinktime(struct bfq_data *bfqd,
4167 struct bfq_queue *bfqq)
4169 struct bfq_ttime *ttime = &bfqq->ttime;
4170 u64 elapsed = ktime_get_ns() - bfqq->ttime.last_end_request;
4172 elapsed = min_t(u64, elapsed, 2ULL * bfqd->bfq_slice_idle);
4174 ttime->ttime_samples = (7*bfqq->ttime.ttime_samples + 256) / 8;
4175 ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed, 8);
4176 ttime->ttime_mean = div64_ul(ttime->ttime_total + 128,
4177 ttime->ttime_samples);
4181 bfq_update_io_seektime(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4184 bfqq->seek_history <<= 1;
4185 bfqq->seek_history |=
4186 get_sdist(bfqq->last_request_pos, rq) > BFQQ_SEEK_THR &&
4187 (!blk_queue_nonrot(bfqd->queue) ||
4188 blk_rq_sectors(rq) < BFQQ_SECT_THR_NONROT);
4191 static void bfq_update_has_short_ttime(struct bfq_data *bfqd,
4192 struct bfq_queue *bfqq,
4193 struct bfq_io_cq *bic)
4195 bool has_short_ttime = true;
4198 * No need to update has_short_ttime if bfqq is async or in
4199 * idle io prio class, or if bfq_slice_idle is zero, because
4200 * no device idling is performed for bfqq in this case.
4202 if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq) ||
4203 bfqd->bfq_slice_idle == 0)
4206 /* Idle window just restored, statistics are meaningless. */
4207 if (time_is_after_eq_jiffies(bfqq->split_time +
4208 bfqd->bfq_wr_min_idle_time))
4211 /* Think time is infinite if no process is linked to
4212 * bfqq. Otherwise check average think time to
4213 * decide whether to mark as has_short_ttime
4215 if (atomic_read(&bic->icq.ioc->active_ref) == 0 ||
4216 (bfq_sample_valid(bfqq->ttime.ttime_samples) &&
4217 bfqq->ttime.ttime_mean > bfqd->bfq_slice_idle))
4218 has_short_ttime = false;
4220 bfq_log_bfqq(bfqd, bfqq, "update_has_short_ttime: has_short_ttime %d",
4223 if (has_short_ttime)
4224 bfq_mark_bfqq_has_short_ttime(bfqq);
4226 bfq_clear_bfqq_has_short_ttime(bfqq);
4230 * Called when a new fs request (rq) is added to bfqq. Check if there's
4231 * something we should do about it.
4233 static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4236 struct bfq_io_cq *bic = RQ_BIC(rq);
4238 if (rq->cmd_flags & REQ_META)
4239 bfqq->meta_pending++;
4241 bfq_update_io_thinktime(bfqd, bfqq);
4242 bfq_update_has_short_ttime(bfqd, bfqq, bic);
4243 bfq_update_io_seektime(bfqd, bfqq, rq);
4245 bfq_log_bfqq(bfqd, bfqq,
4246 "rq_enqueued: has_short_ttime=%d (seeky %d)",
4247 bfq_bfqq_has_short_ttime(bfqq), BFQQ_SEEKY(bfqq));
4249 bfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
4251 if (bfqq == bfqd->in_service_queue && bfq_bfqq_wait_request(bfqq)) {
4252 bool small_req = bfqq->queued[rq_is_sync(rq)] == 1 &&
4253 blk_rq_sectors(rq) < 32;
4254 bool budget_timeout = bfq_bfqq_budget_timeout(bfqq);
4257 * There is just this request queued: if the request
4258 * is small and the queue is not to be expired, then
4261 * In this way, if the device is being idled to wait
4262 * for a new request from the in-service queue, we
4263 * avoid unplugging the device and committing the
4264 * device to serve just a small request. On the
4265 * contrary, we wait for the block layer to decide
4266 * when to unplug the device: hopefully, new requests
4267 * will be merged to this one quickly, then the device
4268 * will be unplugged and larger requests will be
4271 if (small_req && !budget_timeout)
4275 * A large enough request arrived, or the queue is to
4276 * be expired: in both cases disk idling is to be
4277 * stopped, so clear wait_request flag and reset
4280 bfq_clear_bfqq_wait_request(bfqq);
4281 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
4284 * The queue is not empty, because a new request just
4285 * arrived. Hence we can safely expire the queue, in
4286 * case of budget timeout, without risking that the
4287 * timestamps of the queue are not updated correctly.
4288 * See [1] for more details.
4291 bfq_bfqq_expire(bfqd, bfqq, false,
4292 BFQQE_BUDGET_TIMEOUT);
4296 /* returns true if it causes the idle timer to be disabled */
4297 static bool __bfq_insert_request(struct bfq_data *bfqd, struct request *rq)
4299 struct bfq_queue *bfqq = RQ_BFQQ(rq),
4300 *new_bfqq = bfq_setup_cooperator(bfqd, bfqq, rq, true);
4301 bool waiting, idle_timer_disabled = false;
4304 if (bic_to_bfqq(RQ_BIC(rq), 1) != bfqq)
4305 new_bfqq = bic_to_bfqq(RQ_BIC(rq), 1);
4307 * Release the request's reference to the old bfqq
4308 * and make sure one is taken to the shared queue.
4310 new_bfqq->allocated++;
4314 * If the bic associated with the process
4315 * issuing this request still points to bfqq
4316 * (and thus has not been already redirected
4317 * to new_bfqq or even some other bfq_queue),
4318 * then complete the merge and redirect it to
4321 if (bic_to_bfqq(RQ_BIC(rq), 1) == bfqq)
4322 bfq_merge_bfqqs(bfqd, RQ_BIC(rq),
4325 bfq_clear_bfqq_just_created(bfqq);
4327 * rq is about to be enqueued into new_bfqq,
4328 * release rq reference on bfqq
4330 bfq_put_queue(bfqq);
4331 rq->elv.priv[1] = new_bfqq;
4335 waiting = bfqq && bfq_bfqq_wait_request(bfqq);
4336 bfq_add_request(rq);
4337 idle_timer_disabled = waiting && !bfq_bfqq_wait_request(bfqq);
4339 rq->fifo_time = ktime_get_ns() + bfqd->bfq_fifo_expire[rq_is_sync(rq)];
4340 list_add_tail(&rq->queuelist, &bfqq->fifo);
4342 bfq_rq_enqueued(bfqd, bfqq, rq);
4344 return idle_timer_disabled;
4347 #if defined(CONFIG_BFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
4348 static void bfq_update_insert_stats(struct request_queue *q,
4349 struct bfq_queue *bfqq,
4350 bool idle_timer_disabled,
4351 unsigned int cmd_flags)
4357 * bfqq still exists, because it can disappear only after
4358 * either it is merged with another queue, or the process it
4359 * is associated with exits. But both actions must be taken by
4360 * the same process currently executing this flow of
4363 * In addition, the following queue lock guarantees that
4364 * bfqq_group(bfqq) exists as well.
4366 spin_lock_irq(q->queue_lock);
4367 bfqg_stats_update_io_add(bfqq_group(bfqq), bfqq, cmd_flags);
4368 if (idle_timer_disabled)
4369 bfqg_stats_update_idle_time(bfqq_group(bfqq));
4370 spin_unlock_irq(q->queue_lock);
4373 static inline void bfq_update_insert_stats(struct request_queue *q,
4374 struct bfq_queue *bfqq,
4375 bool idle_timer_disabled,
4376 unsigned int cmd_flags) {}
4379 static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
4382 struct request_queue *q = hctx->queue;
4383 struct bfq_data *bfqd = q->elevator->elevator_data;
4384 struct bfq_queue *bfqq = RQ_BFQQ(rq);
4385 bool idle_timer_disabled = false;
4386 unsigned int cmd_flags;
4388 spin_lock_irq(&bfqd->lock);
4389 if (blk_mq_sched_try_insert_merge(q, rq)) {
4390 spin_unlock_irq(&bfqd->lock);
4394 spin_unlock_irq(&bfqd->lock);
4396 blk_mq_sched_request_inserted(rq);
4398 spin_lock_irq(&bfqd->lock);
4399 if (at_head || blk_rq_is_passthrough(rq)) {
4401 list_add(&rq->queuelist, &bfqd->dispatch);
4403 list_add_tail(&rq->queuelist, &bfqd->dispatch);
4405 idle_timer_disabled = __bfq_insert_request(bfqd, rq);
4407 * Update bfqq, because, if a queue merge has occurred
4408 * in __bfq_insert_request, then rq has been
4409 * redirected into a new queue.
4413 if (rq_mergeable(rq)) {
4414 elv_rqhash_add(q, rq);
4421 * Cache cmd_flags before releasing scheduler lock, because rq
4422 * may disappear afterwards (for example, because of a request
4425 cmd_flags = rq->cmd_flags;
4427 spin_unlock_irq(&bfqd->lock);
4429 bfq_update_insert_stats(q, bfqq, idle_timer_disabled,
4433 static void bfq_insert_requests(struct blk_mq_hw_ctx *hctx,
4434 struct list_head *list, bool at_head)
4436 while (!list_empty(list)) {
4439 rq = list_first_entry(list, struct request, queuelist);
4440 list_del_init(&rq->queuelist);
4441 bfq_insert_request(hctx, rq, at_head);
4445 static void bfq_update_hw_tag(struct bfq_data *bfqd)
4447 bfqd->max_rq_in_driver = max_t(int, bfqd->max_rq_in_driver,
4448 bfqd->rq_in_driver);
4450 if (bfqd->hw_tag == 1)
4454 * This sample is valid if the number of outstanding requests
4455 * is large enough to allow a queueing behavior. Note that the
4456 * sum is not exact, as it's not taking into account deactivated
4459 if (bfqd->rq_in_driver + bfqd->queued < BFQ_HW_QUEUE_THRESHOLD)
4462 if (bfqd->hw_tag_samples++ < BFQ_HW_QUEUE_SAMPLES)
4465 bfqd->hw_tag = bfqd->max_rq_in_driver > BFQ_HW_QUEUE_THRESHOLD;
4466 bfqd->max_rq_in_driver = 0;
4467 bfqd->hw_tag_samples = 0;
4470 static void bfq_completed_request(struct bfq_queue *bfqq, struct bfq_data *bfqd)
4475 bfq_update_hw_tag(bfqd);
4477 bfqd->rq_in_driver--;
4480 if (!bfqq->dispatched && !bfq_bfqq_busy(bfqq)) {
4482 * Set budget_timeout (which we overload to store the
4483 * time at which the queue remains with no backlog and
4484 * no outstanding request; used by the weight-raising
4487 bfqq->budget_timeout = jiffies;
4489 bfq_weights_tree_remove(bfqd, &bfqq->entity,
4490 &bfqd->queue_weights_tree);
4493 now_ns = ktime_get_ns();
4495 bfqq->ttime.last_end_request = now_ns;
4498 * Using us instead of ns, to get a reasonable precision in
4499 * computing rate in next check.
4501 delta_us = div_u64(now_ns - bfqd->last_completion, NSEC_PER_USEC);
4504 * If the request took rather long to complete, and, according
4505 * to the maximum request size recorded, this completion latency
4506 * implies that the request was certainly served at a very low
4507 * rate (less than 1M sectors/sec), then the whole observation
4508 * interval that lasts up to this time instant cannot be a
4509 * valid time interval for computing a new peak rate. Invoke
4510 * bfq_update_rate_reset to have the following three steps
4512 * - close the observation interval at the last (previous)
4513 * request dispatch or completion
4514 * - compute rate, if possible, for that observation interval
4515 * - reset to zero samples, which will trigger a proper
4516 * re-initialization of the observation interval on next
4519 if (delta_us > BFQ_MIN_TT/NSEC_PER_USEC &&
4520 (bfqd->last_rq_max_size<<BFQ_RATE_SHIFT)/delta_us <
4521 1UL<<(BFQ_RATE_SHIFT - 10))
4522 bfq_update_rate_reset(bfqd, NULL);
4523 bfqd->last_completion = now_ns;
4526 * If we are waiting to discover whether the request pattern
4527 * of the task associated with the queue is actually
4528 * isochronous, and both requisites for this condition to hold
4529 * are now satisfied, then compute soft_rt_next_start (see the
4530 * comments on the function bfq_bfqq_softrt_next_start()). We
4531 * schedule this delayed check when bfqq expires, if it still
4532 * has in-flight requests.
4534 if (bfq_bfqq_softrt_update(bfqq) && bfqq->dispatched == 0 &&
4535 RB_EMPTY_ROOT(&bfqq->sort_list))
4536 bfqq->soft_rt_next_start =
4537 bfq_bfqq_softrt_next_start(bfqd, bfqq);
4540 * If this is the in-service queue, check if it needs to be expired,
4541 * or if we want to idle in case it has no pending requests.
4543 if (bfqd->in_service_queue == bfqq) {
4544 if (bfqq->dispatched == 0 && bfq_bfqq_must_idle(bfqq)) {
4545 bfq_arm_slice_timer(bfqd);
4547 } else if (bfq_may_expire_for_budg_timeout(bfqq))
4548 bfq_bfqq_expire(bfqd, bfqq, false,
4549 BFQQE_BUDGET_TIMEOUT);
4550 else if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
4551 (bfqq->dispatched == 0 ||
4552 !bfq_bfqq_may_idle(bfqq)))
4553 bfq_bfqq_expire(bfqd, bfqq, false,
4554 BFQQE_NO_MORE_REQUESTS);
4557 if (!bfqd->rq_in_driver)
4558 bfq_schedule_dispatch(bfqd);
4561 static void bfq_finish_request_body(struct bfq_queue *bfqq)
4565 bfq_put_queue(bfqq);
4568 static void bfq_finish_request(struct request *rq)
4570 struct bfq_queue *bfqq;
4571 struct bfq_data *bfqd;
4579 if (rq->rq_flags & RQF_STARTED)
4580 bfqg_stats_update_completion(bfqq_group(bfqq),
4581 rq_start_time_ns(rq),
4582 rq_io_start_time_ns(rq),
4585 if (likely(rq->rq_flags & RQF_STARTED)) {
4586 unsigned long flags;
4588 spin_lock_irqsave(&bfqd->lock, flags);
4590 bfq_completed_request(bfqq, bfqd);
4591 bfq_finish_request_body(bfqq);
4593 spin_unlock_irqrestore(&bfqd->lock, flags);
4596 * Request rq may be still/already in the scheduler,
4597 * in which case we need to remove it. And we cannot
4598 * defer such a check and removal, to avoid
4599 * inconsistencies in the time interval from the end
4600 * of this function to the start of the deferred work.
4601 * This situation seems to occur only in process
4602 * context, as a consequence of a merge. In the
4603 * current version of the code, this implies that the
4607 if (!RB_EMPTY_NODE(&rq->rb_node)) {
4608 bfq_remove_request(rq->q, rq);
4609 bfqg_stats_update_io_remove(bfqq_group(bfqq),
4612 bfq_finish_request_body(bfqq);
4615 rq->elv.priv[0] = NULL;
4616 rq->elv.priv[1] = NULL;
4620 * Returns NULL if a new bfqq should be allocated, or the old bfqq if this
4621 * was the last process referring to that bfqq.
4623 static struct bfq_queue *
4624 bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq)
4626 bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue");
4628 if (bfqq_process_refs(bfqq) == 1) {
4629 bfqq->pid = current->pid;
4630 bfq_clear_bfqq_coop(bfqq);
4631 bfq_clear_bfqq_split_coop(bfqq);
4635 bic_set_bfqq(bic, NULL, 1);
4637 bfq_put_cooperator(bfqq);
4639 bfq_put_queue(bfqq);
4643 static struct bfq_queue *bfq_get_bfqq_handle_split(struct bfq_data *bfqd,
4644 struct bfq_io_cq *bic,
4646 bool split, bool is_sync,
4649 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
4651 if (likely(bfqq && bfqq != &bfqd->oom_bfqq))
4658 bfq_put_queue(bfqq);
4659 bfqq = bfq_get_queue(bfqd, bio, is_sync, bic);
4661 bic_set_bfqq(bic, bfqq, is_sync);
4662 if (split && is_sync) {
4663 if ((bic->was_in_burst_list && bfqd->large_burst) ||
4664 bic->saved_in_large_burst)
4665 bfq_mark_bfqq_in_large_burst(bfqq);
4667 bfq_clear_bfqq_in_large_burst(bfqq);
4668 if (bic->was_in_burst_list)
4670 * If bfqq was in the current
4671 * burst list before being
4672 * merged, then we have to add
4673 * it back. And we do not need
4674 * to increase burst_size, as
4675 * we did not decrement
4676 * burst_size when we removed
4677 * bfqq from the burst list as
4678 * a consequence of a merge
4680 * bfq_put_queue). In this
4681 * respect, it would be rather
4682 * costly to know whether the
4683 * current burst list is still
4684 * the same burst list from
4685 * which bfqq was removed on
4686 * the merge. To avoid this
4687 * cost, if bfqq was in a
4688 * burst list, then we add
4689 * bfqq to the current burst
4690 * list without any further
4691 * check. This can cause
4692 * inappropriate insertions,
4693 * but rarely enough to not
4694 * harm the detection of large
4695 * bursts significantly.
4697 hlist_add_head(&bfqq->burst_list_node,
4700 bfqq->split_time = jiffies;
4707 * Allocate bfq data structures associated with this request.
4709 static void bfq_prepare_request(struct request *rq, struct bio *bio)
4711 struct request_queue *q = rq->q;
4712 struct bfq_data *bfqd = q->elevator->elevator_data;
4713 struct bfq_io_cq *bic;
4714 const int is_sync = rq_is_sync(rq);
4715 struct bfq_queue *bfqq;
4716 bool new_queue = false;
4717 bool bfqq_already_existing = false, split = false;
4721 bic = icq_to_bic(rq->elv.icq);
4723 spin_lock_irq(&bfqd->lock);
4725 bfq_check_ioprio_change(bic, bio);
4727 bfq_bic_update_cgroup(bic, bio);
4729 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio, false, is_sync,
4732 if (likely(!new_queue)) {
4733 /* If the queue was seeky for too long, break it apart. */
4734 if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq)) {
4735 bfq_log_bfqq(bfqd, bfqq, "breaking apart bfqq");
4737 /* Update bic before losing reference to bfqq */
4738 if (bfq_bfqq_in_large_burst(bfqq))
4739 bic->saved_in_large_burst = true;
4741 bfqq = bfq_split_bfqq(bic, bfqq);
4745 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio,
4749 bfqq_already_existing = true;
4755 bfq_log_bfqq(bfqd, bfqq, "get_request %p: bfqq %p, %d",
4756 rq, bfqq, bfqq->ref);
4758 rq->elv.priv[0] = bic;
4759 rq->elv.priv[1] = bfqq;
4762 * If a bfq_queue has only one process reference, it is owned
4763 * by only this bic: we can then set bfqq->bic = bic. in
4764 * addition, if the queue has also just been split, we have to
4767 if (likely(bfqq != &bfqd->oom_bfqq) && bfqq_process_refs(bfqq) == 1) {
4771 * The queue has just been split from a shared
4772 * queue: restore the idle window and the
4773 * possible weight raising period.
4775 bfq_bfqq_resume_state(bfqq, bfqd, bic,
4776 bfqq_already_existing);
4780 if (unlikely(bfq_bfqq_just_created(bfqq)))
4781 bfq_handle_burst(bfqd, bfqq);
4783 spin_unlock_irq(&bfqd->lock);
4786 static void bfq_idle_slice_timer_body(struct bfq_queue *bfqq)
4788 struct bfq_data *bfqd = bfqq->bfqd;
4789 enum bfqq_expiration reason;
4790 unsigned long flags;
4792 spin_lock_irqsave(&bfqd->lock, flags);
4793 bfq_clear_bfqq_wait_request(bfqq);
4795 if (bfqq != bfqd->in_service_queue) {
4796 spin_unlock_irqrestore(&bfqd->lock, flags);
4800 if (bfq_bfqq_budget_timeout(bfqq))
4802 * Also here the queue can be safely expired
4803 * for budget timeout without wasting
4806 reason = BFQQE_BUDGET_TIMEOUT;
4807 else if (bfqq->queued[0] == 0 && bfqq->queued[1] == 0)
4809 * The queue may not be empty upon timer expiration,
4810 * because we may not disable the timer when the
4811 * first request of the in-service queue arrives
4812 * during disk idling.
4814 reason = BFQQE_TOO_IDLE;
4816 goto schedule_dispatch;
4818 bfq_bfqq_expire(bfqd, bfqq, true, reason);
4821 spin_unlock_irqrestore(&bfqd->lock, flags);
4822 bfq_schedule_dispatch(bfqd);
4826 * Handler of the expiration of the timer running if the in-service queue
4827 * is idling inside its time slice.
4829 static enum hrtimer_restart bfq_idle_slice_timer(struct hrtimer *timer)
4831 struct bfq_data *bfqd = container_of(timer, struct bfq_data,
4833 struct bfq_queue *bfqq = bfqd->in_service_queue;
4836 * Theoretical race here: the in-service queue can be NULL or
4837 * different from the queue that was idling if a new request
4838 * arrives for the current queue and there is a full dispatch
4839 * cycle that changes the in-service queue. This can hardly
4840 * happen, but in the worst case we just expire a queue too
4844 bfq_idle_slice_timer_body(bfqq);
4846 return HRTIMER_NORESTART;
4849 static void __bfq_put_async_bfqq(struct bfq_data *bfqd,
4850 struct bfq_queue **bfqq_ptr)
4852 struct bfq_queue *bfqq = *bfqq_ptr;
4854 bfq_log(bfqd, "put_async_bfqq: %p", bfqq);
4856 bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
4858 bfq_log_bfqq(bfqd, bfqq, "put_async_bfqq: putting %p, %d",
4860 bfq_put_queue(bfqq);
4866 * Release all the bfqg references to its async queues. If we are
4867 * deallocating the group these queues may still contain requests, so
4868 * we reparent them to the root cgroup (i.e., the only one that will
4869 * exist for sure until all the requests on a device are gone).
4871 void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg)
4875 for (i = 0; i < 2; i++)
4876 for (j = 0; j < IOPRIO_BE_NR; j++)
4877 __bfq_put_async_bfqq(bfqd, &bfqg->async_bfqq[i][j]);
4879 __bfq_put_async_bfqq(bfqd, &bfqg->async_idle_bfqq);
4882 static void bfq_exit_queue(struct elevator_queue *e)
4884 struct bfq_data *bfqd = e->elevator_data;
4885 struct bfq_queue *bfqq, *n;
4887 hrtimer_cancel(&bfqd->idle_slice_timer);
4889 spin_lock_irq(&bfqd->lock);
4890 list_for_each_entry_safe(bfqq, n, &bfqd->idle_list, bfqq_list)
4891 bfq_deactivate_bfqq(bfqd, bfqq, false, false);
4892 spin_unlock_irq(&bfqd->lock);
4894 hrtimer_cancel(&bfqd->idle_slice_timer);
4896 #ifdef CONFIG_BFQ_GROUP_IOSCHED
4897 /* release oom-queue reference to root group */
4898 bfqg_and_blkg_put(bfqd->root_group);
4900 blkcg_deactivate_policy(bfqd->queue, &blkcg_policy_bfq);
4902 spin_lock_irq(&bfqd->lock);
4903 bfq_put_async_queues(bfqd, bfqd->root_group);
4904 kfree(bfqd->root_group);
4905 spin_unlock_irq(&bfqd->lock);
4911 static void bfq_init_root_group(struct bfq_group *root_group,
4912 struct bfq_data *bfqd)
4916 #ifdef CONFIG_BFQ_GROUP_IOSCHED
4917 root_group->entity.parent = NULL;
4918 root_group->my_entity = NULL;
4919 root_group->bfqd = bfqd;
4921 root_group->rq_pos_tree = RB_ROOT;
4922 for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
4923 root_group->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
4924 root_group->sched_data.bfq_class_idle_last_service = jiffies;
4927 static int bfq_init_queue(struct request_queue *q, struct elevator_type *e)
4929 struct bfq_data *bfqd;
4930 struct elevator_queue *eq;
4932 eq = elevator_alloc(q, e);
4936 bfqd = kzalloc_node(sizeof(*bfqd), GFP_KERNEL, q->node);
4938 kobject_put(&eq->kobj);
4941 eq->elevator_data = bfqd;
4943 spin_lock_irq(q->queue_lock);
4945 spin_unlock_irq(q->queue_lock);
4948 * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues.
4949 * Grab a permanent reference to it, so that the normal code flow
4950 * will not attempt to free it.
4952 bfq_init_bfqq(bfqd, &bfqd->oom_bfqq, NULL, 1, 0);
4953 bfqd->oom_bfqq.ref++;
4954 bfqd->oom_bfqq.new_ioprio = BFQ_DEFAULT_QUEUE_IOPRIO;
4955 bfqd->oom_bfqq.new_ioprio_class = IOPRIO_CLASS_BE;
4956 bfqd->oom_bfqq.entity.new_weight =
4957 bfq_ioprio_to_weight(bfqd->oom_bfqq.new_ioprio);
4959 /* oom_bfqq does not participate to bursts */
4960 bfq_clear_bfqq_just_created(&bfqd->oom_bfqq);
4963 * Trigger weight initialization, according to ioprio, at the
4964 * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio
4965 * class won't be changed any more.
4967 bfqd->oom_bfqq.entity.prio_changed = 1;
4971 INIT_LIST_HEAD(&bfqd->dispatch);
4973 hrtimer_init(&bfqd->idle_slice_timer, CLOCK_MONOTONIC,
4975 bfqd->idle_slice_timer.function = bfq_idle_slice_timer;
4977 bfqd->queue_weights_tree = RB_ROOT;
4978 bfqd->group_weights_tree = RB_ROOT;
4980 INIT_LIST_HEAD(&bfqd->active_list);
4981 INIT_LIST_HEAD(&bfqd->idle_list);
4982 INIT_HLIST_HEAD(&bfqd->burst_list);
4986 bfqd->bfq_max_budget = bfq_default_max_budget;
4988 bfqd->bfq_fifo_expire[0] = bfq_fifo_expire[0];
4989 bfqd->bfq_fifo_expire[1] = bfq_fifo_expire[1];
4990 bfqd->bfq_back_max = bfq_back_max;
4991 bfqd->bfq_back_penalty = bfq_back_penalty;
4992 bfqd->bfq_slice_idle = bfq_slice_idle;
4993 bfqd->bfq_timeout = bfq_timeout;
4995 bfqd->bfq_requests_within_timer = 120;
4997 bfqd->bfq_large_burst_thresh = 8;
4998 bfqd->bfq_burst_interval = msecs_to_jiffies(180);
5000 bfqd->low_latency = true;
5003 * Trade-off between responsiveness and fairness.
5005 bfqd->bfq_wr_coeff = 30;
5006 bfqd->bfq_wr_rt_max_time = msecs_to_jiffies(300);
5007 bfqd->bfq_wr_max_time = 0;
5008 bfqd->bfq_wr_min_idle_time = msecs_to_jiffies(2000);
5009 bfqd->bfq_wr_min_inter_arr_async = msecs_to_jiffies(500);
5010 bfqd->bfq_wr_max_softrt_rate = 7000; /*
5011 * Approximate rate required
5012 * to playback or record a
5013 * high-definition compressed
5016 bfqd->wr_busy_queues = 0;
5019 * Begin by assuming, optimistically, that the device is a
5020 * high-speed one, and that its peak rate is equal to 2/3 of
5021 * the highest reference rate.
5023 bfqd->RT_prod = R_fast[blk_queue_nonrot(bfqd->queue)] *
5024 T_fast[blk_queue_nonrot(bfqd->queue)];
5025 bfqd->peak_rate = R_fast[blk_queue_nonrot(bfqd->queue)] * 2 / 3;
5026 bfqd->device_speed = BFQ_BFQD_FAST;
5028 spin_lock_init(&bfqd->lock);
5031 * The invocation of the next bfq_create_group_hierarchy
5032 * function is the head of a chain of function calls
5033 * (bfq_create_group_hierarchy->blkcg_activate_policy->
5034 * blk_mq_freeze_queue) that may lead to the invocation of the
5035 * has_work hook function. For this reason,
5036 * bfq_create_group_hierarchy is invoked only after all
5037 * scheduler data has been initialized, apart from the fields
5038 * that can be initialized only after invoking
5039 * bfq_create_group_hierarchy. This, in particular, enables
5040 * has_work to correctly return false. Of course, to avoid
5041 * other inconsistencies, the blk-mq stack must then refrain
5042 * from invoking further scheduler hooks before this init
5043 * function is finished.
5045 bfqd->root_group = bfq_create_group_hierarchy(bfqd, q->node);
5046 if (!bfqd->root_group)
5048 bfq_init_root_group(bfqd->root_group, bfqd);
5049 bfq_init_entity(&bfqd->oom_bfqq.entity, bfqd->root_group);
5051 wbt_disable_default(q);
5056 kobject_put(&eq->kobj);
5060 static void bfq_slab_kill(void)
5062 kmem_cache_destroy(bfq_pool);
5065 static int __init bfq_slab_setup(void)
5067 bfq_pool = KMEM_CACHE(bfq_queue, 0);
5073 static ssize_t bfq_var_show(unsigned int var, char *page)
5075 return sprintf(page, "%u\n", var);
5078 static int bfq_var_store(unsigned long *var, const char *page)
5080 unsigned long new_val;
5081 int ret = kstrtoul(page, 10, &new_val);
5089 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
5090 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
5092 struct bfq_data *bfqd = e->elevator_data; \
5093 u64 __data = __VAR; \
5095 __data = jiffies_to_msecs(__data); \
5096 else if (__CONV == 2) \
5097 __data = div_u64(__data, NSEC_PER_MSEC); \
5098 return bfq_var_show(__data, (page)); \
5100 SHOW_FUNCTION(bfq_fifo_expire_sync_show, bfqd->bfq_fifo_expire[1], 2);
5101 SHOW_FUNCTION(bfq_fifo_expire_async_show, bfqd->bfq_fifo_expire[0], 2);
5102 SHOW_FUNCTION(bfq_back_seek_max_show, bfqd->bfq_back_max, 0);
5103 SHOW_FUNCTION(bfq_back_seek_penalty_show, bfqd->bfq_back_penalty, 0);
5104 SHOW_FUNCTION(bfq_slice_idle_show, bfqd->bfq_slice_idle, 2);
5105 SHOW_FUNCTION(bfq_max_budget_show, bfqd->bfq_user_max_budget, 0);
5106 SHOW_FUNCTION(bfq_timeout_sync_show, bfqd->bfq_timeout, 1);
5107 SHOW_FUNCTION(bfq_strict_guarantees_show, bfqd->strict_guarantees, 0);
5108 SHOW_FUNCTION(bfq_low_latency_show, bfqd->low_latency, 0);
5109 #undef SHOW_FUNCTION
5111 #define USEC_SHOW_FUNCTION(__FUNC, __VAR) \
5112 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
5114 struct bfq_data *bfqd = e->elevator_data; \
5115 u64 __data = __VAR; \
5116 __data = div_u64(__data, NSEC_PER_USEC); \
5117 return bfq_var_show(__data, (page)); \
5119 USEC_SHOW_FUNCTION(bfq_slice_idle_us_show, bfqd->bfq_slice_idle);
5120 #undef USEC_SHOW_FUNCTION
5122 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
5124 __FUNC(struct elevator_queue *e, const char *page, size_t count) \
5126 struct bfq_data *bfqd = e->elevator_data; \
5127 unsigned long __data, __min = (MIN), __max = (MAX); \
5130 ret = bfq_var_store(&__data, (page)); \
5133 if (__data < __min) \
5135 else if (__data > __max) \
5138 *(__PTR) = msecs_to_jiffies(__data); \
5139 else if (__CONV == 2) \
5140 *(__PTR) = (u64)__data * NSEC_PER_MSEC; \
5142 *(__PTR) = __data; \
5145 STORE_FUNCTION(bfq_fifo_expire_sync_store, &bfqd->bfq_fifo_expire[1], 1,
5147 STORE_FUNCTION(bfq_fifo_expire_async_store, &bfqd->bfq_fifo_expire[0], 1,
5149 STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0);
5150 STORE_FUNCTION(bfq_back_seek_penalty_store, &bfqd->bfq_back_penalty, 1,
5152 STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 2);
5153 #undef STORE_FUNCTION
5155 #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
5156 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\
5158 struct bfq_data *bfqd = e->elevator_data; \
5159 unsigned long __data, __min = (MIN), __max = (MAX); \
5162 ret = bfq_var_store(&__data, (page)); \
5165 if (__data < __min) \
5167 else if (__data > __max) \
5169 *(__PTR) = (u64)__data * NSEC_PER_USEC; \
5172 USEC_STORE_FUNCTION(bfq_slice_idle_us_store, &bfqd->bfq_slice_idle, 0,
5174 #undef USEC_STORE_FUNCTION
5176 static ssize_t bfq_max_budget_store(struct elevator_queue *e,
5177 const char *page, size_t count)
5179 struct bfq_data *bfqd = e->elevator_data;
5180 unsigned long __data;
5183 ret = bfq_var_store(&__data, (page));
5188 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
5190 if (__data > INT_MAX)
5192 bfqd->bfq_max_budget = __data;
5195 bfqd->bfq_user_max_budget = __data;
5201 * Leaving this name to preserve name compatibility with cfq
5202 * parameters, but this timeout is used for both sync and async.
5204 static ssize_t bfq_timeout_sync_store(struct elevator_queue *e,
5205 const char *page, size_t count)
5207 struct bfq_data *bfqd = e->elevator_data;
5208 unsigned long __data;
5211 ret = bfq_var_store(&__data, (page));
5217 else if (__data > INT_MAX)
5220 bfqd->bfq_timeout = msecs_to_jiffies(__data);
5221 if (bfqd->bfq_user_max_budget == 0)
5222 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
5227 static ssize_t bfq_strict_guarantees_store(struct elevator_queue *e,
5228 const char *page, size_t count)
5230 struct bfq_data *bfqd = e->elevator_data;
5231 unsigned long __data;
5234 ret = bfq_var_store(&__data, (page));
5240 if (!bfqd->strict_guarantees && __data == 1
5241 && bfqd->bfq_slice_idle < 8 * NSEC_PER_MSEC)
5242 bfqd->bfq_slice_idle = 8 * NSEC_PER_MSEC;
5244 bfqd->strict_guarantees = __data;
5249 static ssize_t bfq_low_latency_store(struct elevator_queue *e,
5250 const char *page, size_t count)
5252 struct bfq_data *bfqd = e->elevator_data;
5253 unsigned long __data;
5256 ret = bfq_var_store(&__data, (page));
5262 if (__data == 0 && bfqd->low_latency != 0)
5264 bfqd->low_latency = __data;
5269 #define BFQ_ATTR(name) \
5270 __ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store)
5272 static struct elv_fs_entry bfq_attrs[] = {
5273 BFQ_ATTR(fifo_expire_sync),
5274 BFQ_ATTR(fifo_expire_async),
5275 BFQ_ATTR(back_seek_max),
5276 BFQ_ATTR(back_seek_penalty),
5277 BFQ_ATTR(slice_idle),
5278 BFQ_ATTR(slice_idle_us),
5279 BFQ_ATTR(max_budget),
5280 BFQ_ATTR(timeout_sync),
5281 BFQ_ATTR(strict_guarantees),
5282 BFQ_ATTR(low_latency),
5286 static struct elevator_type iosched_bfq_mq = {
5288 .prepare_request = bfq_prepare_request,
5289 .finish_request = bfq_finish_request,
5290 .exit_icq = bfq_exit_icq,
5291 .insert_requests = bfq_insert_requests,
5292 .dispatch_request = bfq_dispatch_request,
5293 .next_request = elv_rb_latter_request,
5294 .former_request = elv_rb_former_request,
5295 .allow_merge = bfq_allow_bio_merge,
5296 .bio_merge = bfq_bio_merge,
5297 .request_merge = bfq_request_merge,
5298 .requests_merged = bfq_requests_merged,
5299 .request_merged = bfq_request_merged,
5300 .has_work = bfq_has_work,
5301 .init_sched = bfq_init_queue,
5302 .exit_sched = bfq_exit_queue,
5306 .icq_size = sizeof(struct bfq_io_cq),
5307 .icq_align = __alignof__(struct bfq_io_cq),
5308 .elevator_attrs = bfq_attrs,
5309 .elevator_name = "bfq",
5310 .elevator_owner = THIS_MODULE,
5312 MODULE_ALIAS("bfq-iosched");
5314 static int __init bfq_init(void)
5318 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5319 ret = blkcg_policy_register(&blkcg_policy_bfq);
5325 if (bfq_slab_setup())
5329 * Times to load large popular applications for the typical
5330 * systems installed on the reference devices (see the
5331 * comments before the definitions of the next two
5332 * arrays). Actually, we use slightly slower values, as the
5333 * estimated peak rate tends to be smaller than the actual
5334 * peak rate. The reason for this last fact is that estimates
5335 * are computed over much shorter time intervals than the long
5336 * intervals typically used for benchmarking. Why? First, to
5337 * adapt more quickly to variations. Second, because an I/O
5338 * scheduler cannot rely on a peak-rate-evaluation workload to
5339 * be run for a long time.
5341 T_slow[0] = msecs_to_jiffies(3500); /* actually 4 sec */
5342 T_slow[1] = msecs_to_jiffies(6000); /* actually 6.5 sec */
5343 T_fast[0] = msecs_to_jiffies(7000); /* actually 8 sec */
5344 T_fast[1] = msecs_to_jiffies(2500); /* actually 3 sec */
5347 * Thresholds that determine the switch between speed classes
5348 * (see the comments before the definition of the array
5349 * device_speed_thresh). These thresholds are biased towards
5350 * transitions to the fast class. This is safer than the
5351 * opposite bias. In fact, a wrong transition to the slow
5352 * class results in short weight-raising periods, because the
5353 * speed of the device then tends to be higher that the
5354 * reference peak rate. On the opposite end, a wrong
5355 * transition to the fast class tends to increase
5356 * weight-raising periods, because of the opposite reason.
5358 device_speed_thresh[0] = (4 * R_slow[0]) / 3;
5359 device_speed_thresh[1] = (4 * R_slow[1]) / 3;
5361 ret = elv_register(&iosched_bfq_mq);
5370 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5371 blkcg_policy_unregister(&blkcg_policy_bfq);
5376 static void __exit bfq_exit(void)
5378 elv_unregister(&iosched_bfq_mq);
5379 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5380 blkcg_policy_unregister(&blkcg_policy_bfq);
5385 module_init(bfq_init);
5386 module_exit(bfq_exit);
5388 MODULE_AUTHOR("Paolo Valente");
5389 MODULE_LICENSE("GPL");
5390 MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler");