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;
169 static struct kmem_cache *bfq_pool;
171 /* Below this threshold (in ns), we consider thinktime immediate. */
172 #define BFQ_MIN_TT (2 * NSEC_PER_MSEC)
174 /* hw_tag detection: parallel requests threshold and min samples needed. */
175 #define BFQ_HW_QUEUE_THRESHOLD 4
176 #define BFQ_HW_QUEUE_SAMPLES 32
178 #define BFQQ_SEEK_THR (sector_t)(8 * 100)
179 #define BFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
180 #define BFQQ_CLOSE_THR (sector_t)(8 * 1024)
181 #define BFQQ_SEEKY(bfqq) (hweight32(bfqq->seek_history) > 32/8)
183 /* Min number of samples required to perform peak-rate update */
184 #define BFQ_RATE_MIN_SAMPLES 32
185 /* Min observation time interval required to perform a peak-rate update (ns) */
186 #define BFQ_RATE_MIN_INTERVAL (300*NSEC_PER_MSEC)
187 /* Target observation time interval for a peak-rate update (ns) */
188 #define BFQ_RATE_REF_INTERVAL NSEC_PER_SEC
190 /* Shift used for peak rate fixed precision calculations. */
191 #define BFQ_RATE_SHIFT 16
194 * By default, BFQ computes the duration of the weight raising for
195 * interactive applications automatically, using the following formula:
196 * duration = (R / r) * T, where r is the peak rate of the device, and
197 * R and T are two reference parameters.
198 * In particular, R is the peak rate of the reference device (see below),
199 * and T is a reference time: given the systems that are likely to be
200 * installed on the reference device according to its speed class, T is
201 * about the maximum time needed, under BFQ and while reading two files in
202 * parallel, to load typical large applications on these systems.
203 * In practice, the slower/faster the device at hand is, the more/less it
204 * takes to load applications with respect to the reference device.
205 * Accordingly, the longer/shorter BFQ grants weight raising to interactive
208 * BFQ uses four different reference pairs (R, T), depending on:
209 * . whether the device is rotational or non-rotational;
210 * . whether the device is slow, such as old or portable HDDs, as well as
211 * SD cards, or fast, such as newer HDDs and SSDs.
213 * The device's speed class is dynamically (re)detected in
214 * bfq_update_peak_rate() every time the estimated peak rate is updated.
216 * In the following definitions, R_slow[0]/R_fast[0] and
217 * T_slow[0]/T_fast[0] are the reference values for a slow/fast
218 * rotational device, whereas R_slow[1]/R_fast[1] and
219 * T_slow[1]/T_fast[1] are the reference values for a slow/fast
220 * non-rotational device. Finally, device_speed_thresh are the
221 * thresholds used to switch between speed classes. The reference
222 * rates are not the actual peak rates of the devices used as a
223 * reference, but slightly lower values. The reason for using these
224 * slightly lower values is that the peak-rate estimator tends to
225 * yield slightly lower values than the actual peak rate (it can yield
226 * the actual peak rate only if there is only one process doing I/O,
227 * and the process does sequential I/O).
229 * Both the reference peak rates and the thresholds are measured in
230 * sectors/usec, left-shifted by BFQ_RATE_SHIFT.
232 static int R_slow[2] = {1000, 10700};
233 static int R_fast[2] = {14000, 33000};
235 * To improve readability, a conversion function is used to initialize the
236 * following arrays, which entails that they can be initialized only in a
239 static int T_slow[2];
240 static int T_fast[2];
241 static int device_speed_thresh[2];
243 #define RQ_BIC(rq) icq_to_bic((rq)->elv.priv[0])
244 #define RQ_BFQQ(rq) ((rq)->elv.priv[1])
246 struct bfq_queue *bic_to_bfqq(struct bfq_io_cq *bic, bool is_sync)
248 return bic->bfqq[is_sync];
251 void bic_set_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq, bool is_sync)
253 bic->bfqq[is_sync] = bfqq;
256 struct bfq_data *bic_to_bfqd(struct bfq_io_cq *bic)
258 return bic->icq.q->elevator->elevator_data;
262 * icq_to_bic - convert iocontext queue structure to bfq_io_cq.
263 * @icq: the iocontext queue.
265 static struct bfq_io_cq *icq_to_bic(struct io_cq *icq)
267 /* bic->icq is the first member, %NULL will convert to %NULL */
268 return container_of(icq, struct bfq_io_cq, icq);
272 * bfq_bic_lookup - search into @ioc a bic associated to @bfqd.
273 * @bfqd: the lookup key.
274 * @ioc: the io_context of the process doing I/O.
275 * @q: the request queue.
277 static struct bfq_io_cq *bfq_bic_lookup(struct bfq_data *bfqd,
278 struct io_context *ioc,
279 struct request_queue *q)
283 struct bfq_io_cq *icq;
285 spin_lock_irqsave(q->queue_lock, flags);
286 icq = icq_to_bic(ioc_lookup_icq(ioc, q));
287 spin_unlock_irqrestore(q->queue_lock, flags);
296 * Scheduler run of queue, if there are requests pending and no one in the
297 * driver that will restart queueing.
299 void bfq_schedule_dispatch(struct bfq_data *bfqd)
301 if (bfqd->queued != 0) {
302 bfq_log(bfqd, "schedule dispatch");
303 blk_mq_run_hw_queues(bfqd->queue, true);
307 #define bfq_class_idle(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
308 #define bfq_class_rt(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_RT)
310 #define bfq_sample_valid(samples) ((samples) > 80)
313 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
314 * We choose the request that is closesr to the head right now. Distance
315 * behind the head is penalized and only allowed to a certain extent.
317 static struct request *bfq_choose_req(struct bfq_data *bfqd,
322 sector_t s1, s2, d1 = 0, d2 = 0;
323 unsigned long back_max;
324 #define BFQ_RQ1_WRAP 0x01 /* request 1 wraps */
325 #define BFQ_RQ2_WRAP 0x02 /* request 2 wraps */
326 unsigned int wrap = 0; /* bit mask: requests behind the disk head? */
328 if (!rq1 || rq1 == rq2)
333 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
335 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
337 if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
339 else if ((rq2->cmd_flags & REQ_META) && !(rq1->cmd_flags & REQ_META))
342 s1 = blk_rq_pos(rq1);
343 s2 = blk_rq_pos(rq2);
346 * By definition, 1KiB is 2 sectors.
348 back_max = bfqd->bfq_back_max * 2;
351 * Strict one way elevator _except_ in the case where we allow
352 * short backward seeks which are biased as twice the cost of a
353 * similar forward seek.
357 else if (s1 + back_max >= last)
358 d1 = (last - s1) * bfqd->bfq_back_penalty;
360 wrap |= BFQ_RQ1_WRAP;
364 else if (s2 + back_max >= last)
365 d2 = (last - s2) * bfqd->bfq_back_penalty;
367 wrap |= BFQ_RQ2_WRAP;
369 /* Found required data */
372 * By doing switch() on the bit mask "wrap" we avoid having to
373 * check two variables for all permutations: --> faster!
376 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
391 case BFQ_RQ1_WRAP|BFQ_RQ2_WRAP: /* both rqs wrapped */
394 * Since both rqs are wrapped,
395 * start with the one that's further behind head
396 * (--> only *one* back seek required),
397 * since back seek takes more time than forward.
406 static struct bfq_queue *
407 bfq_rq_pos_tree_lookup(struct bfq_data *bfqd, struct rb_root *root,
408 sector_t sector, struct rb_node **ret_parent,
409 struct rb_node ***rb_link)
411 struct rb_node **p, *parent;
412 struct bfq_queue *bfqq = NULL;
420 bfqq = rb_entry(parent, struct bfq_queue, pos_node);
423 * Sort strictly based on sector. Smallest to the left,
424 * largest to the right.
426 if (sector > blk_rq_pos(bfqq->next_rq))
428 else if (sector < blk_rq_pos(bfqq->next_rq))
436 *ret_parent = parent;
440 bfq_log(bfqd, "rq_pos_tree_lookup %llu: returning %d",
441 (unsigned long long)sector,
442 bfqq ? bfqq->pid : 0);
447 void bfq_pos_tree_add_move(struct bfq_data *bfqd, struct bfq_queue *bfqq)
449 struct rb_node **p, *parent;
450 struct bfq_queue *__bfqq;
452 if (bfqq->pos_root) {
453 rb_erase(&bfqq->pos_node, bfqq->pos_root);
454 bfqq->pos_root = NULL;
457 if (bfq_class_idle(bfqq))
462 bfqq->pos_root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree;
463 __bfqq = bfq_rq_pos_tree_lookup(bfqd, bfqq->pos_root,
464 blk_rq_pos(bfqq->next_rq), &parent, &p);
466 rb_link_node(&bfqq->pos_node, parent, p);
467 rb_insert_color(&bfqq->pos_node, bfqq->pos_root);
469 bfqq->pos_root = NULL;
473 * Tell whether there are active queues or groups with differentiated weights.
475 static bool bfq_differentiated_weights(struct bfq_data *bfqd)
478 * For weights to differ, at least one of the trees must contain
479 * at least two nodes.
481 return (!RB_EMPTY_ROOT(&bfqd->queue_weights_tree) &&
482 (bfqd->queue_weights_tree.rb_node->rb_left ||
483 bfqd->queue_weights_tree.rb_node->rb_right)
484 #ifdef CONFIG_BFQ_GROUP_IOSCHED
486 (!RB_EMPTY_ROOT(&bfqd->group_weights_tree) &&
487 (bfqd->group_weights_tree.rb_node->rb_left ||
488 bfqd->group_weights_tree.rb_node->rb_right)
494 * The following function returns true if every queue must receive the
495 * same share of the throughput (this condition is used when deciding
496 * whether idling may be disabled, see the comments in the function
497 * bfq_bfqq_may_idle()).
499 * Such a scenario occurs when:
500 * 1) all active queues have the same weight,
501 * 2) all active groups at the same level in the groups tree have the same
503 * 3) all active groups at the same level in the groups tree have the same
504 * number of children.
506 * Unfortunately, keeping the necessary state for evaluating exactly the
507 * above symmetry conditions would be quite complex and time-consuming.
508 * Therefore this function evaluates, instead, the following stronger
509 * sub-conditions, for which it is much easier to maintain the needed
511 * 1) all active queues have the same weight,
512 * 2) all active groups have the same weight,
513 * 3) all active groups have at most one active child each.
514 * In particular, the last two conditions are always true if hierarchical
515 * support and the cgroups interface are not enabled, thus no state needs
516 * to be maintained in this case.
518 static bool bfq_symmetric_scenario(struct bfq_data *bfqd)
520 return !bfq_differentiated_weights(bfqd);
524 * If the weight-counter tree passed as input contains no counter for
525 * the weight of the input entity, then add that counter; otherwise just
526 * increment the existing counter.
528 * Note that weight-counter trees contain few nodes in mostly symmetric
529 * scenarios. For example, if all queues have the same weight, then the
530 * weight-counter tree for the queues may contain at most one node.
531 * This holds even if low_latency is on, because weight-raised queues
532 * are not inserted in the tree.
533 * In most scenarios, the rate at which nodes are created/destroyed
536 void bfq_weights_tree_add(struct bfq_data *bfqd, struct bfq_entity *entity,
537 struct rb_root *root)
539 struct rb_node **new = &(root->rb_node), *parent = NULL;
542 * Do not insert if the entity is already associated with a
543 * counter, which happens if:
544 * 1) the entity is associated with a queue,
545 * 2) a request arrival has caused the queue to become both
546 * non-weight-raised, and hence change its weight, and
547 * backlogged; in this respect, each of the two events
548 * causes an invocation of this function,
549 * 3) this is the invocation of this function caused by the
550 * second event. This second invocation is actually useless,
551 * and we handle this fact by exiting immediately. More
552 * efficient or clearer solutions might possibly be adopted.
554 if (entity->weight_counter)
558 struct bfq_weight_counter *__counter = container_of(*new,
559 struct bfq_weight_counter,
563 if (entity->weight == __counter->weight) {
564 entity->weight_counter = __counter;
567 if (entity->weight < __counter->weight)
568 new = &((*new)->rb_left);
570 new = &((*new)->rb_right);
573 entity->weight_counter = kzalloc(sizeof(struct bfq_weight_counter),
577 * In the unlucky event of an allocation failure, we just
578 * exit. This will cause the weight of entity to not be
579 * considered in bfq_differentiated_weights, which, in its
580 * turn, causes the scenario to be deemed wrongly symmetric in
581 * case entity's weight would have been the only weight making
582 * the scenario asymmetric. On the bright side, no unbalance
583 * will however occur when entity becomes inactive again (the
584 * invocation of this function is triggered by an activation
585 * of entity). In fact, bfq_weights_tree_remove does nothing
586 * if !entity->weight_counter.
588 if (unlikely(!entity->weight_counter))
591 entity->weight_counter->weight = entity->weight;
592 rb_link_node(&entity->weight_counter->weights_node, parent, new);
593 rb_insert_color(&entity->weight_counter->weights_node, root);
596 entity->weight_counter->num_active++;
600 * Decrement the weight counter associated with the entity, and, if the
601 * counter reaches 0, remove the counter from the tree.
602 * See the comments to the function bfq_weights_tree_add() for considerations
605 void bfq_weights_tree_remove(struct bfq_data *bfqd, struct bfq_entity *entity,
606 struct rb_root *root)
608 if (!entity->weight_counter)
611 entity->weight_counter->num_active--;
612 if (entity->weight_counter->num_active > 0)
613 goto reset_entity_pointer;
615 rb_erase(&entity->weight_counter->weights_node, root);
616 kfree(entity->weight_counter);
618 reset_entity_pointer:
619 entity->weight_counter = NULL;
623 * Return expired entry, or NULL to just start from scratch in rbtree.
625 static struct request *bfq_check_fifo(struct bfq_queue *bfqq,
626 struct request *last)
630 if (bfq_bfqq_fifo_expire(bfqq))
633 bfq_mark_bfqq_fifo_expire(bfqq);
635 rq = rq_entry_fifo(bfqq->fifo.next);
637 if (rq == last || ktime_get_ns() < rq->fifo_time)
640 bfq_log_bfqq(bfqq->bfqd, bfqq, "check_fifo: returned %p", rq);
644 static struct request *bfq_find_next_rq(struct bfq_data *bfqd,
645 struct bfq_queue *bfqq,
646 struct request *last)
648 struct rb_node *rbnext = rb_next(&last->rb_node);
649 struct rb_node *rbprev = rb_prev(&last->rb_node);
650 struct request *next, *prev = NULL;
652 /* Follow expired path, else get first next available. */
653 next = bfq_check_fifo(bfqq, last);
658 prev = rb_entry_rq(rbprev);
661 next = rb_entry_rq(rbnext);
663 rbnext = rb_first(&bfqq->sort_list);
664 if (rbnext && rbnext != &last->rb_node)
665 next = rb_entry_rq(rbnext);
668 return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last));
671 /* see the definition of bfq_async_charge_factor for details */
672 static unsigned long bfq_serv_to_charge(struct request *rq,
673 struct bfq_queue *bfqq)
675 if (bfq_bfqq_sync(bfqq) || bfqq->wr_coeff > 1)
676 return blk_rq_sectors(rq);
679 * If there are no weight-raised queues, then amplify service
680 * by just the async charge factor; otherwise amplify service
681 * by twice the async charge factor, to further reduce latency
682 * for weight-raised queues.
684 if (bfqq->bfqd->wr_busy_queues == 0)
685 return blk_rq_sectors(rq) * bfq_async_charge_factor;
687 return blk_rq_sectors(rq) * 2 * bfq_async_charge_factor;
691 * bfq_updated_next_req - update the queue after a new next_rq selection.
692 * @bfqd: the device data the queue belongs to.
693 * @bfqq: the queue to update.
695 * If the first request of a queue changes we make sure that the queue
696 * has enough budget to serve at least its first request (if the
697 * request has grown). We do this because if the queue has not enough
698 * budget for its first request, it has to go through two dispatch
699 * rounds to actually get it dispatched.
701 static void bfq_updated_next_req(struct bfq_data *bfqd,
702 struct bfq_queue *bfqq)
704 struct bfq_entity *entity = &bfqq->entity;
705 struct request *next_rq = bfqq->next_rq;
706 unsigned long new_budget;
711 if (bfqq == bfqd->in_service_queue)
713 * In order not to break guarantees, budgets cannot be
714 * changed after an entity has been selected.
718 new_budget = max_t(unsigned long, bfqq->max_budget,
719 bfq_serv_to_charge(next_rq, bfqq));
720 if (entity->budget != new_budget) {
721 entity->budget = new_budget;
722 bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu",
724 bfq_requeue_bfqq(bfqd, bfqq, false);
728 static unsigned int bfq_wr_duration(struct bfq_data *bfqd)
732 if (bfqd->bfq_wr_max_time > 0)
733 return bfqd->bfq_wr_max_time;
736 do_div(dur, bfqd->peak_rate);
739 * Limit duration between 3 and 13 seconds. Tests show that
740 * higher values than 13 seconds often yield the opposite of
741 * the desired result, i.e., worsen responsiveness by letting
742 * non-interactive and non-soft-real-time applications
743 * preserve weight raising for a too long time interval.
745 * On the other end, lower values than 3 seconds make it
746 * difficult for most interactive tasks to complete their jobs
747 * before weight-raising finishes.
749 if (dur > msecs_to_jiffies(13000))
750 dur = msecs_to_jiffies(13000);
751 else if (dur < msecs_to_jiffies(3000))
752 dur = msecs_to_jiffies(3000);
757 /* switch back from soft real-time to interactive weight raising */
758 static void switch_back_to_interactive_wr(struct bfq_queue *bfqq,
759 struct bfq_data *bfqd)
761 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
762 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
763 bfqq->last_wr_start_finish = bfqq->wr_start_at_switch_to_srt;
767 bfq_bfqq_resume_state(struct bfq_queue *bfqq, struct bfq_data *bfqd,
768 struct bfq_io_cq *bic, bool bfq_already_existing)
770 unsigned int old_wr_coeff = bfqq->wr_coeff;
771 bool busy = bfq_already_existing && bfq_bfqq_busy(bfqq);
773 if (bic->saved_has_short_ttime)
774 bfq_mark_bfqq_has_short_ttime(bfqq);
776 bfq_clear_bfqq_has_short_ttime(bfqq);
778 if (bic->saved_IO_bound)
779 bfq_mark_bfqq_IO_bound(bfqq);
781 bfq_clear_bfqq_IO_bound(bfqq);
783 bfqq->ttime = bic->saved_ttime;
784 bfqq->wr_coeff = bic->saved_wr_coeff;
785 bfqq->wr_start_at_switch_to_srt = bic->saved_wr_start_at_switch_to_srt;
786 bfqq->last_wr_start_finish = bic->saved_last_wr_start_finish;
787 bfqq->wr_cur_max_time = bic->saved_wr_cur_max_time;
789 if (bfqq->wr_coeff > 1 && (bfq_bfqq_in_large_burst(bfqq) ||
790 time_is_before_jiffies(bfqq->last_wr_start_finish +
791 bfqq->wr_cur_max_time))) {
792 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
793 !bfq_bfqq_in_large_burst(bfqq) &&
794 time_is_after_eq_jiffies(bfqq->wr_start_at_switch_to_srt +
795 bfq_wr_duration(bfqd))) {
796 switch_back_to_interactive_wr(bfqq, bfqd);
799 bfq_log_bfqq(bfqq->bfqd, bfqq,
800 "resume state: switching off wr");
804 /* make sure weight will be updated, however we got here */
805 bfqq->entity.prio_changed = 1;
810 if (old_wr_coeff == 1 && bfqq->wr_coeff > 1)
811 bfqd->wr_busy_queues++;
812 else if (old_wr_coeff > 1 && bfqq->wr_coeff == 1)
813 bfqd->wr_busy_queues--;
816 static int bfqq_process_refs(struct bfq_queue *bfqq)
818 return bfqq->ref - bfqq->allocated - bfqq->entity.on_st;
821 /* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */
822 static void bfq_reset_burst_list(struct bfq_data *bfqd, struct bfq_queue *bfqq)
824 struct bfq_queue *item;
825 struct hlist_node *n;
827 hlist_for_each_entry_safe(item, n, &bfqd->burst_list, burst_list_node)
828 hlist_del_init(&item->burst_list_node);
829 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
830 bfqd->burst_size = 1;
831 bfqd->burst_parent_entity = bfqq->entity.parent;
834 /* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */
835 static void bfq_add_to_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
837 /* Increment burst size to take into account also bfqq */
840 if (bfqd->burst_size == bfqd->bfq_large_burst_thresh) {
841 struct bfq_queue *pos, *bfqq_item;
842 struct hlist_node *n;
845 * Enough queues have been activated shortly after each
846 * other to consider this burst as large.
848 bfqd->large_burst = true;
851 * We can now mark all queues in the burst list as
852 * belonging to a large burst.
854 hlist_for_each_entry(bfqq_item, &bfqd->burst_list,
856 bfq_mark_bfqq_in_large_burst(bfqq_item);
857 bfq_mark_bfqq_in_large_burst(bfqq);
860 * From now on, and until the current burst finishes, any
861 * new queue being activated shortly after the last queue
862 * was inserted in the burst can be immediately marked as
863 * belonging to a large burst. So the burst list is not
864 * needed any more. Remove it.
866 hlist_for_each_entry_safe(pos, n, &bfqd->burst_list,
868 hlist_del_init(&pos->burst_list_node);
870 * Burst not yet large: add bfqq to the burst list. Do
871 * not increment the ref counter for bfqq, because bfqq
872 * is removed from the burst list before freeing bfqq
875 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
879 * If many queues belonging to the same group happen to be created
880 * shortly after each other, then the processes associated with these
881 * queues have typically a common goal. In particular, bursts of queue
882 * creations are usually caused by services or applications that spawn
883 * many parallel threads/processes. Examples are systemd during boot,
884 * or git grep. To help these processes get their job done as soon as
885 * possible, it is usually better to not grant either weight-raising
886 * or device idling to their queues.
888 * In this comment we describe, firstly, the reasons why this fact
889 * holds, and, secondly, the next function, which implements the main
890 * steps needed to properly mark these queues so that they can then be
891 * treated in a different way.
893 * The above services or applications benefit mostly from a high
894 * throughput: the quicker the requests of the activated queues are
895 * cumulatively served, the sooner the target job of these queues gets
896 * completed. As a consequence, weight-raising any of these queues,
897 * which also implies idling the device for it, is almost always
898 * counterproductive. In most cases it just lowers throughput.
900 * On the other hand, a burst of queue creations may be caused also by
901 * the start of an application that does not consist of a lot of
902 * parallel I/O-bound threads. In fact, with a complex application,
903 * several short processes may need to be executed to start-up the
904 * application. In this respect, to start an application as quickly as
905 * possible, the best thing to do is in any case to privilege the I/O
906 * related to the application with respect to all other
907 * I/O. Therefore, the best strategy to start as quickly as possible
908 * an application that causes a burst of queue creations is to
909 * weight-raise all the queues created during the burst. This is the
910 * exact opposite of the best strategy for the other type of bursts.
912 * In the end, to take the best action for each of the two cases, the
913 * two types of bursts need to be distinguished. Fortunately, this
914 * seems relatively easy, by looking at the sizes of the bursts. In
915 * particular, we found a threshold such that only bursts with a
916 * larger size than that threshold are apparently caused by
917 * services or commands such as systemd or git grep. For brevity,
918 * hereafter we call just 'large' these bursts. BFQ *does not*
919 * weight-raise queues whose creation occurs in a large burst. In
920 * addition, for each of these queues BFQ performs or does not perform
921 * idling depending on which choice boosts the throughput more. The
922 * exact choice depends on the device and request pattern at
925 * Unfortunately, false positives may occur while an interactive task
926 * is starting (e.g., an application is being started). The
927 * consequence is that the queues associated with the task do not
928 * enjoy weight raising as expected. Fortunately these false positives
929 * are very rare. They typically occur if some service happens to
930 * start doing I/O exactly when the interactive task starts.
932 * Turning back to the next function, it implements all the steps
933 * needed to detect the occurrence of a large burst and to properly
934 * mark all the queues belonging to it (so that they can then be
935 * treated in a different way). This goal is achieved by maintaining a
936 * "burst list" that holds, temporarily, the queues that belong to the
937 * burst in progress. The list is then used to mark these queues as
938 * belonging to a large burst if the burst does become large. The main
939 * steps are the following.
941 * . when the very first queue is created, the queue is inserted into the
942 * list (as it could be the first queue in a possible burst)
944 * . if the current burst has not yet become large, and a queue Q that does
945 * not yet belong to the burst is activated shortly after the last time
946 * at which a new queue entered the burst list, then the function appends
947 * Q to the burst list
949 * . if, as a consequence of the previous step, the burst size reaches
950 * the large-burst threshold, then
952 * . all the queues in the burst list are marked as belonging to a
955 * . the burst list is deleted; in fact, the burst list already served
956 * its purpose (keeping temporarily track of the queues in a burst,
957 * so as to be able to mark them as belonging to a large burst in the
958 * previous sub-step), and now is not needed any more
960 * . the device enters a large-burst mode
962 * . if a queue Q that does not belong to the burst is created while
963 * the device is in large-burst mode and shortly after the last time
964 * at which a queue either entered the burst list or was marked as
965 * belonging to the current large burst, then Q is immediately marked
966 * as belonging to a large burst.
968 * . if a queue Q that does not belong to the burst is created a while
969 * later, i.e., not shortly after, than the last time at which a queue
970 * either entered the burst list or was marked as belonging to the
971 * current large burst, then the current burst is deemed as finished and:
973 * . the large-burst mode is reset if set
975 * . the burst list is emptied
977 * . Q is inserted in the burst list, as Q may be the first queue
978 * in a possible new burst (then the burst list contains just Q
981 static void bfq_handle_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
984 * If bfqq is already in the burst list or is part of a large
985 * burst, or finally has just been split, then there is
986 * nothing else to do.
988 if (!hlist_unhashed(&bfqq->burst_list_node) ||
989 bfq_bfqq_in_large_burst(bfqq) ||
990 time_is_after_eq_jiffies(bfqq->split_time +
991 msecs_to_jiffies(10)))
995 * If bfqq's creation happens late enough, or bfqq belongs to
996 * a different group than the burst group, then the current
997 * burst is finished, and related data structures must be
1000 * In this respect, consider the special case where bfqq is
1001 * the very first queue created after BFQ is selected for this
1002 * device. In this case, last_ins_in_burst and
1003 * burst_parent_entity are not yet significant when we get
1004 * here. But it is easy to verify that, whether or not the
1005 * following condition is true, bfqq will end up being
1006 * inserted into the burst list. In particular the list will
1007 * happen to contain only bfqq. And this is exactly what has
1008 * to happen, as bfqq may be the first queue of the first
1011 if (time_is_before_jiffies(bfqd->last_ins_in_burst +
1012 bfqd->bfq_burst_interval) ||
1013 bfqq->entity.parent != bfqd->burst_parent_entity) {
1014 bfqd->large_burst = false;
1015 bfq_reset_burst_list(bfqd, bfqq);
1020 * If we get here, then bfqq is being activated shortly after the
1021 * last queue. So, if the current burst is also large, we can mark
1022 * bfqq as belonging to this large burst immediately.
1024 if (bfqd->large_burst) {
1025 bfq_mark_bfqq_in_large_burst(bfqq);
1030 * If we get here, then a large-burst state has not yet been
1031 * reached, but bfqq is being activated shortly after the last
1032 * queue. Then we add bfqq to the burst.
1034 bfq_add_to_burst(bfqd, bfqq);
1037 * At this point, bfqq either has been added to the current
1038 * burst or has caused the current burst to terminate and a
1039 * possible new burst to start. In particular, in the second
1040 * case, bfqq has become the first queue in the possible new
1041 * burst. In both cases last_ins_in_burst needs to be moved
1044 bfqd->last_ins_in_burst = jiffies;
1047 static int bfq_bfqq_budget_left(struct bfq_queue *bfqq)
1049 struct bfq_entity *entity = &bfqq->entity;
1051 return entity->budget - entity->service;
1055 * If enough samples have been computed, return the current max budget
1056 * stored in bfqd, which is dynamically updated according to the
1057 * estimated disk peak rate; otherwise return the default max budget
1059 static int bfq_max_budget(struct bfq_data *bfqd)
1061 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1062 return bfq_default_max_budget;
1064 return bfqd->bfq_max_budget;
1068 * Return min budget, which is a fraction of the current or default
1069 * max budget (trying with 1/32)
1071 static int bfq_min_budget(struct bfq_data *bfqd)
1073 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1074 return bfq_default_max_budget / 32;
1076 return bfqd->bfq_max_budget / 32;
1080 * The next function, invoked after the input queue bfqq switches from
1081 * idle to busy, updates the budget of bfqq. The function also tells
1082 * whether the in-service queue should be expired, by returning
1083 * true. The purpose of expiring the in-service queue is to give bfqq
1084 * the chance to possibly preempt the in-service queue, and the reason
1085 * for preempting the in-service queue is to achieve one of the two
1088 * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has
1089 * expired because it has remained idle. In particular, bfqq may have
1090 * expired for one of the following two reasons:
1092 * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling
1093 * and did not make it to issue a new request before its last
1094 * request was served;
1096 * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue
1097 * a new request before the expiration of the idling-time.
1099 * Even if bfqq has expired for one of the above reasons, the process
1100 * associated with the queue may be however issuing requests greedily,
1101 * and thus be sensitive to the bandwidth it receives (bfqq may have
1102 * remained idle for other reasons: CPU high load, bfqq not enjoying
1103 * idling, I/O throttling somewhere in the path from the process to
1104 * the I/O scheduler, ...). But if, after every expiration for one of
1105 * the above two reasons, bfqq has to wait for the service of at least
1106 * one full budget of another queue before being served again, then
1107 * bfqq is likely to get a much lower bandwidth or resource time than
1108 * its reserved ones. To address this issue, two countermeasures need
1111 * First, the budget and the timestamps of bfqq need to be updated in
1112 * a special way on bfqq reactivation: they need to be updated as if
1113 * bfqq did not remain idle and did not expire. In fact, if they are
1114 * computed as if bfqq expired and remained idle until reactivation,
1115 * then the process associated with bfqq is treated as if, instead of
1116 * being greedy, it stopped issuing requests when bfqq remained idle,
1117 * and restarts issuing requests only on this reactivation. In other
1118 * words, the scheduler does not help the process recover the "service
1119 * hole" between bfqq expiration and reactivation. As a consequence,
1120 * the process receives a lower bandwidth than its reserved one. In
1121 * contrast, to recover this hole, the budget must be updated as if
1122 * bfqq was not expired at all before this reactivation, i.e., it must
1123 * be set to the value of the remaining budget when bfqq was
1124 * expired. Along the same line, timestamps need to be assigned the
1125 * value they had the last time bfqq was selected for service, i.e.,
1126 * before last expiration. Thus timestamps need to be back-shifted
1127 * with respect to their normal computation (see [1] for more details
1128 * on this tricky aspect).
1130 * Secondly, to allow the process to recover the hole, the in-service
1131 * queue must be expired too, to give bfqq the chance to preempt it
1132 * immediately. In fact, if bfqq has to wait for a full budget of the
1133 * in-service queue to be completed, then it may become impossible to
1134 * let the process recover the hole, even if the back-shifted
1135 * timestamps of bfqq are lower than those of the in-service queue. If
1136 * this happens for most or all of the holes, then the process may not
1137 * receive its reserved bandwidth. In this respect, it is worth noting
1138 * that, being the service of outstanding requests unpreemptible, a
1139 * little fraction of the holes may however be unrecoverable, thereby
1140 * causing a little loss of bandwidth.
1142 * The last important point is detecting whether bfqq does need this
1143 * bandwidth recovery. In this respect, the next function deems the
1144 * process associated with bfqq greedy, and thus allows it to recover
1145 * the hole, if: 1) the process is waiting for the arrival of a new
1146 * request (which implies that bfqq expired for one of the above two
1147 * reasons), and 2) such a request has arrived soon. The first
1148 * condition is controlled through the flag non_blocking_wait_rq,
1149 * while the second through the flag arrived_in_time. If both
1150 * conditions hold, then the function computes the budget in the
1151 * above-described special way, and signals that the in-service queue
1152 * should be expired. Timestamp back-shifting is done later in
1153 * __bfq_activate_entity.
1155 * 2. Reduce latency. Even if timestamps are not backshifted to let
1156 * the process associated with bfqq recover a service hole, bfqq may
1157 * however happen to have, after being (re)activated, a lower finish
1158 * timestamp than the in-service queue. That is, the next budget of
1159 * bfqq may have to be completed before the one of the in-service
1160 * queue. If this is the case, then preempting the in-service queue
1161 * allows this goal to be achieved, apart from the unpreemptible,
1162 * outstanding requests mentioned above.
1164 * Unfortunately, regardless of which of the above two goals one wants
1165 * to achieve, service trees need first to be updated to know whether
1166 * the in-service queue must be preempted. To have service trees
1167 * correctly updated, the in-service queue must be expired and
1168 * rescheduled, and bfqq must be scheduled too. This is one of the
1169 * most costly operations (in future versions, the scheduling
1170 * mechanism may be re-designed in such a way to make it possible to
1171 * know whether preemption is needed without needing to update service
1172 * trees). In addition, queue preemptions almost always cause random
1173 * I/O, and thus loss of throughput. Because of these facts, the next
1174 * function adopts the following simple scheme to avoid both costly
1175 * operations and too frequent preemptions: it requests the expiration
1176 * of the in-service queue (unconditionally) only for queues that need
1177 * to recover a hole, or that either are weight-raised or deserve to
1180 static bool bfq_bfqq_update_budg_for_activation(struct bfq_data *bfqd,
1181 struct bfq_queue *bfqq,
1182 bool arrived_in_time,
1183 bool wr_or_deserves_wr)
1185 struct bfq_entity *entity = &bfqq->entity;
1187 if (bfq_bfqq_non_blocking_wait_rq(bfqq) && arrived_in_time) {
1189 * We do not clear the flag non_blocking_wait_rq here, as
1190 * the latter is used in bfq_activate_bfqq to signal
1191 * that timestamps need to be back-shifted (and is
1192 * cleared right after).
1196 * In next assignment we rely on that either
1197 * entity->service or entity->budget are not updated
1198 * on expiration if bfqq is empty (see
1199 * __bfq_bfqq_recalc_budget). Thus both quantities
1200 * remain unchanged after such an expiration, and the
1201 * following statement therefore assigns to
1202 * entity->budget the remaining budget on such an
1203 * expiration. For clarity, entity->service is not
1204 * updated on expiration in any case, and, in normal
1205 * operation, is reset only when bfqq is selected for
1206 * service (see bfq_get_next_queue).
1208 entity->budget = min_t(unsigned long,
1209 bfq_bfqq_budget_left(bfqq),
1215 entity->budget = max_t(unsigned long, bfqq->max_budget,
1216 bfq_serv_to_charge(bfqq->next_rq, bfqq));
1217 bfq_clear_bfqq_non_blocking_wait_rq(bfqq);
1218 return wr_or_deserves_wr;
1222 * Return the farthest future time instant according to jiffies
1225 static unsigned long bfq_greatest_from_now(void)
1227 return jiffies + MAX_JIFFY_OFFSET;
1231 * Return the farthest past time instant according to jiffies
1234 static unsigned long bfq_smallest_from_now(void)
1236 return jiffies - MAX_JIFFY_OFFSET;
1239 static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data *bfqd,
1240 struct bfq_queue *bfqq,
1241 unsigned int old_wr_coeff,
1242 bool wr_or_deserves_wr,
1247 if (old_wr_coeff == 1 && wr_or_deserves_wr) {
1248 /* start a weight-raising period */
1250 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1251 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1254 * No interactive weight raising in progress
1255 * here: assign minus infinity to
1256 * wr_start_at_switch_to_srt, to make sure
1257 * that, at the end of the soft-real-time
1258 * weight raising periods that is starting
1259 * now, no interactive weight-raising period
1260 * may be wrongly considered as still in
1261 * progress (and thus actually started by
1264 bfqq->wr_start_at_switch_to_srt =
1265 bfq_smallest_from_now();
1266 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1267 BFQ_SOFTRT_WEIGHT_FACTOR;
1268 bfqq->wr_cur_max_time =
1269 bfqd->bfq_wr_rt_max_time;
1273 * If needed, further reduce budget to make sure it is
1274 * close to bfqq's backlog, so as to reduce the
1275 * scheduling-error component due to a too large
1276 * budget. Do not care about throughput consequences,
1277 * but only about latency. Finally, do not assign a
1278 * too small budget either, to avoid increasing
1279 * latency by causing too frequent expirations.
1281 bfqq->entity.budget = min_t(unsigned long,
1282 bfqq->entity.budget,
1283 2 * bfq_min_budget(bfqd));
1284 } else if (old_wr_coeff > 1) {
1285 if (interactive) { /* update wr coeff and duration */
1286 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1287 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1288 } else if (in_burst)
1292 * The application is now or still meeting the
1293 * requirements for being deemed soft rt. We
1294 * can then correctly and safely (re)charge
1295 * the weight-raising duration for the
1296 * application with the weight-raising
1297 * duration for soft rt applications.
1299 * In particular, doing this recharge now, i.e.,
1300 * before the weight-raising period for the
1301 * application finishes, reduces the probability
1302 * of the following negative scenario:
1303 * 1) the weight of a soft rt application is
1304 * raised at startup (as for any newly
1305 * created application),
1306 * 2) since the application is not interactive,
1307 * at a certain time weight-raising is
1308 * stopped for the application,
1309 * 3) at that time the application happens to
1310 * still have pending requests, and hence
1311 * is destined to not have a chance to be
1312 * deemed soft rt before these requests are
1313 * completed (see the comments to the
1314 * function bfq_bfqq_softrt_next_start()
1315 * for details on soft rt detection),
1316 * 4) these pending requests experience a high
1317 * latency because the application is not
1318 * weight-raised while they are pending.
1320 if (bfqq->wr_cur_max_time !=
1321 bfqd->bfq_wr_rt_max_time) {
1322 bfqq->wr_start_at_switch_to_srt =
1323 bfqq->last_wr_start_finish;
1325 bfqq->wr_cur_max_time =
1326 bfqd->bfq_wr_rt_max_time;
1327 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1328 BFQ_SOFTRT_WEIGHT_FACTOR;
1330 bfqq->last_wr_start_finish = jiffies;
1335 static bool bfq_bfqq_idle_for_long_time(struct bfq_data *bfqd,
1336 struct bfq_queue *bfqq)
1338 return bfqq->dispatched == 0 &&
1339 time_is_before_jiffies(
1340 bfqq->budget_timeout +
1341 bfqd->bfq_wr_min_idle_time);
1344 static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data *bfqd,
1345 struct bfq_queue *bfqq,
1350 bool soft_rt, in_burst, wr_or_deserves_wr,
1351 bfqq_wants_to_preempt,
1352 idle_for_long_time = bfq_bfqq_idle_for_long_time(bfqd, bfqq),
1354 * See the comments on
1355 * bfq_bfqq_update_budg_for_activation for
1356 * details on the usage of the next variable.
1358 arrived_in_time = ktime_get_ns() <=
1359 bfqq->ttime.last_end_request +
1360 bfqd->bfq_slice_idle * 3;
1362 bfqg_stats_update_io_add(bfqq_group(RQ_BFQQ(rq)), bfqq, rq->cmd_flags);
1365 * bfqq deserves to be weight-raised if:
1367 * - it does not belong to a large burst,
1368 * - it has been idle for enough time or is soft real-time,
1369 * - is linked to a bfq_io_cq (it is not shared in any sense).
1371 in_burst = bfq_bfqq_in_large_burst(bfqq);
1372 soft_rt = bfqd->bfq_wr_max_softrt_rate > 0 &&
1374 time_is_before_jiffies(bfqq->soft_rt_next_start);
1375 *interactive = !in_burst && idle_for_long_time;
1376 wr_or_deserves_wr = bfqd->low_latency &&
1377 (bfqq->wr_coeff > 1 ||
1378 (bfq_bfqq_sync(bfqq) &&
1379 bfqq->bic && (*interactive || soft_rt)));
1382 * Using the last flag, update budget and check whether bfqq
1383 * may want to preempt the in-service queue.
1385 bfqq_wants_to_preempt =
1386 bfq_bfqq_update_budg_for_activation(bfqd, bfqq,
1391 * If bfqq happened to be activated in a burst, but has been
1392 * idle for much more than an interactive queue, then we
1393 * assume that, in the overall I/O initiated in the burst, the
1394 * I/O associated with bfqq is finished. So bfqq does not need
1395 * to be treated as a queue belonging to a burst
1396 * anymore. Accordingly, we reset bfqq's in_large_burst flag
1397 * if set, and remove bfqq from the burst list if it's
1398 * there. We do not decrement burst_size, because the fact
1399 * that bfqq does not need to belong to the burst list any
1400 * more does not invalidate the fact that bfqq was created in
1403 if (likely(!bfq_bfqq_just_created(bfqq)) &&
1404 idle_for_long_time &&
1405 time_is_before_jiffies(
1406 bfqq->budget_timeout +
1407 msecs_to_jiffies(10000))) {
1408 hlist_del_init(&bfqq->burst_list_node);
1409 bfq_clear_bfqq_in_large_burst(bfqq);
1412 bfq_clear_bfqq_just_created(bfqq);
1415 if (!bfq_bfqq_IO_bound(bfqq)) {
1416 if (arrived_in_time) {
1417 bfqq->requests_within_timer++;
1418 if (bfqq->requests_within_timer >=
1419 bfqd->bfq_requests_within_timer)
1420 bfq_mark_bfqq_IO_bound(bfqq);
1422 bfqq->requests_within_timer = 0;
1425 if (bfqd->low_latency) {
1426 if (unlikely(time_is_after_jiffies(bfqq->split_time)))
1429 jiffies - bfqd->bfq_wr_min_idle_time - 1;
1431 if (time_is_before_jiffies(bfqq->split_time +
1432 bfqd->bfq_wr_min_idle_time)) {
1433 bfq_update_bfqq_wr_on_rq_arrival(bfqd, bfqq,
1440 if (old_wr_coeff != bfqq->wr_coeff)
1441 bfqq->entity.prio_changed = 1;
1445 bfqq->last_idle_bklogged = jiffies;
1446 bfqq->service_from_backlogged = 0;
1447 bfq_clear_bfqq_softrt_update(bfqq);
1449 bfq_add_bfqq_busy(bfqd, bfqq);
1452 * Expire in-service queue only if preemption may be needed
1453 * for guarantees. In this respect, the function
1454 * next_queue_may_preempt just checks a simple, necessary
1455 * condition, and not a sufficient condition based on
1456 * timestamps. In fact, for the latter condition to be
1457 * evaluated, timestamps would need first to be updated, and
1458 * this operation is quite costly (see the comments on the
1459 * function bfq_bfqq_update_budg_for_activation).
1461 if (bfqd->in_service_queue && bfqq_wants_to_preempt &&
1462 bfqd->in_service_queue->wr_coeff < bfqq->wr_coeff &&
1463 next_queue_may_preempt(bfqd))
1464 bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
1465 false, BFQQE_PREEMPTED);
1468 static void bfq_add_request(struct request *rq)
1470 struct bfq_queue *bfqq = RQ_BFQQ(rq);
1471 struct bfq_data *bfqd = bfqq->bfqd;
1472 struct request *next_rq, *prev;
1473 unsigned int old_wr_coeff = bfqq->wr_coeff;
1474 bool interactive = false;
1476 bfq_log_bfqq(bfqd, bfqq, "add_request %d", rq_is_sync(rq));
1477 bfqq->queued[rq_is_sync(rq)]++;
1480 elv_rb_add(&bfqq->sort_list, rq);
1483 * Check if this request is a better next-serve candidate.
1485 prev = bfqq->next_rq;
1486 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position);
1487 bfqq->next_rq = next_rq;
1490 * Adjust priority tree position, if next_rq changes.
1492 if (prev != bfqq->next_rq)
1493 bfq_pos_tree_add_move(bfqd, bfqq);
1495 if (!bfq_bfqq_busy(bfqq)) /* switching to busy ... */
1496 bfq_bfqq_handle_idle_busy_switch(bfqd, bfqq, old_wr_coeff,
1499 if (bfqd->low_latency && old_wr_coeff == 1 && !rq_is_sync(rq) &&
1500 time_is_before_jiffies(
1501 bfqq->last_wr_start_finish +
1502 bfqd->bfq_wr_min_inter_arr_async)) {
1503 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1504 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1506 bfqd->wr_busy_queues++;
1507 bfqq->entity.prio_changed = 1;
1509 if (prev != bfqq->next_rq)
1510 bfq_updated_next_req(bfqd, bfqq);
1514 * Assign jiffies to last_wr_start_finish in the following
1517 * . if bfqq is not going to be weight-raised, because, for
1518 * non weight-raised queues, last_wr_start_finish stores the
1519 * arrival time of the last request; as of now, this piece
1520 * of information is used only for deciding whether to
1521 * weight-raise async queues
1523 * . if bfqq is not weight-raised, because, if bfqq is now
1524 * switching to weight-raised, then last_wr_start_finish
1525 * stores the time when weight-raising starts
1527 * . if bfqq is interactive, because, regardless of whether
1528 * bfqq is currently weight-raised, the weight-raising
1529 * period must start or restart (this case is considered
1530 * separately because it is not detected by the above
1531 * conditions, if bfqq is already weight-raised)
1533 * last_wr_start_finish has to be updated also if bfqq is soft
1534 * real-time, because the weight-raising period is constantly
1535 * restarted on idle-to-busy transitions for these queues, but
1536 * this is already done in bfq_bfqq_handle_idle_busy_switch if
1539 if (bfqd->low_latency &&
1540 (old_wr_coeff == 1 || bfqq->wr_coeff == 1 || interactive))
1541 bfqq->last_wr_start_finish = jiffies;
1544 static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd,
1546 struct request_queue *q)
1548 struct bfq_queue *bfqq = bfqd->bio_bfqq;
1552 return elv_rb_find(&bfqq->sort_list, bio_end_sector(bio));
1557 static sector_t get_sdist(sector_t last_pos, struct request *rq)
1560 return abs(blk_rq_pos(rq) - last_pos);
1565 #if 0 /* Still not clear if we can do without next two functions */
1566 static void bfq_activate_request(struct request_queue *q, struct request *rq)
1568 struct bfq_data *bfqd = q->elevator->elevator_data;
1570 bfqd->rq_in_driver++;
1573 static void bfq_deactivate_request(struct request_queue *q, struct request *rq)
1575 struct bfq_data *bfqd = q->elevator->elevator_data;
1577 bfqd->rq_in_driver--;
1581 static void bfq_remove_request(struct request_queue *q,
1584 struct bfq_queue *bfqq = RQ_BFQQ(rq);
1585 struct bfq_data *bfqd = bfqq->bfqd;
1586 const int sync = rq_is_sync(rq);
1588 if (bfqq->next_rq == rq) {
1589 bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq);
1590 bfq_updated_next_req(bfqd, bfqq);
1593 if (rq->queuelist.prev != &rq->queuelist)
1594 list_del_init(&rq->queuelist);
1595 bfqq->queued[sync]--;
1597 elv_rb_del(&bfqq->sort_list, rq);
1599 elv_rqhash_del(q, rq);
1600 if (q->last_merge == rq)
1601 q->last_merge = NULL;
1603 if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
1604 bfqq->next_rq = NULL;
1606 if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->in_service_queue) {
1607 bfq_del_bfqq_busy(bfqd, bfqq, false);
1609 * bfqq emptied. In normal operation, when
1610 * bfqq is empty, bfqq->entity.service and
1611 * bfqq->entity.budget must contain,
1612 * respectively, the service received and the
1613 * budget used last time bfqq emptied. These
1614 * facts do not hold in this case, as at least
1615 * this last removal occurred while bfqq is
1616 * not in service. To avoid inconsistencies,
1617 * reset both bfqq->entity.service and
1618 * bfqq->entity.budget, if bfqq has still a
1619 * process that may issue I/O requests to it.
1621 bfqq->entity.budget = bfqq->entity.service = 0;
1625 * Remove queue from request-position tree as it is empty.
1627 if (bfqq->pos_root) {
1628 rb_erase(&bfqq->pos_node, bfqq->pos_root);
1629 bfqq->pos_root = NULL;
1633 if (rq->cmd_flags & REQ_META)
1634 bfqq->meta_pending--;
1636 bfqg_stats_update_io_remove(bfqq_group(bfqq), rq->cmd_flags);
1639 static bool bfq_bio_merge(struct blk_mq_hw_ctx *hctx, struct bio *bio)
1641 struct request_queue *q = hctx->queue;
1642 struct bfq_data *bfqd = q->elevator->elevator_data;
1643 struct request *free = NULL;
1645 * bfq_bic_lookup grabs the queue_lock: invoke it now and
1646 * store its return value for later use, to avoid nesting
1647 * queue_lock inside the bfqd->lock. We assume that the bic
1648 * returned by bfq_bic_lookup does not go away before
1649 * bfqd->lock is taken.
1651 struct bfq_io_cq *bic = bfq_bic_lookup(bfqd, current->io_context, q);
1654 spin_lock_irq(&bfqd->lock);
1657 bfqd->bio_bfqq = bic_to_bfqq(bic, op_is_sync(bio->bi_opf));
1659 bfqd->bio_bfqq = NULL;
1660 bfqd->bio_bic = bic;
1662 ret = blk_mq_sched_try_merge(q, bio, &free);
1665 blk_mq_free_request(free);
1666 spin_unlock_irq(&bfqd->lock);
1671 static int bfq_request_merge(struct request_queue *q, struct request **req,
1674 struct bfq_data *bfqd = q->elevator->elevator_data;
1675 struct request *__rq;
1677 __rq = bfq_find_rq_fmerge(bfqd, bio, q);
1678 if (__rq && elv_bio_merge_ok(__rq, bio)) {
1680 return ELEVATOR_FRONT_MERGE;
1683 return ELEVATOR_NO_MERGE;
1686 static void bfq_request_merged(struct request_queue *q, struct request *req,
1687 enum elv_merge type)
1689 if (type == ELEVATOR_FRONT_MERGE &&
1690 rb_prev(&req->rb_node) &&
1692 blk_rq_pos(container_of(rb_prev(&req->rb_node),
1693 struct request, rb_node))) {
1694 struct bfq_queue *bfqq = RQ_BFQQ(req);
1695 struct bfq_data *bfqd = bfqq->bfqd;
1696 struct request *prev, *next_rq;
1698 /* Reposition request in its sort_list */
1699 elv_rb_del(&bfqq->sort_list, req);
1700 elv_rb_add(&bfqq->sort_list, req);
1702 /* Choose next request to be served for bfqq */
1703 prev = bfqq->next_rq;
1704 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, req,
1705 bfqd->last_position);
1706 bfqq->next_rq = next_rq;
1708 * If next_rq changes, update both the queue's budget to
1709 * fit the new request and the queue's position in its
1712 if (prev != bfqq->next_rq) {
1713 bfq_updated_next_req(bfqd, bfqq);
1714 bfq_pos_tree_add_move(bfqd, bfqq);
1719 static void bfq_requests_merged(struct request_queue *q, struct request *rq,
1720 struct request *next)
1722 struct bfq_queue *bfqq = RQ_BFQQ(rq), *next_bfqq = RQ_BFQQ(next);
1724 if (!RB_EMPTY_NODE(&rq->rb_node))
1726 spin_lock_irq(&bfqq->bfqd->lock);
1729 * If next and rq belong to the same bfq_queue and next is older
1730 * than rq, then reposition rq in the fifo (by substituting next
1731 * with rq). Otherwise, if next and rq belong to different
1732 * bfq_queues, never reposition rq: in fact, we would have to
1733 * reposition it with respect to next's position in its own fifo,
1734 * which would most certainly be too expensive with respect to
1737 if (bfqq == next_bfqq &&
1738 !list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1739 next->fifo_time < rq->fifo_time) {
1740 list_del_init(&rq->queuelist);
1741 list_replace_init(&next->queuelist, &rq->queuelist);
1742 rq->fifo_time = next->fifo_time;
1745 if (bfqq->next_rq == next)
1748 bfq_remove_request(q, next);
1750 spin_unlock_irq(&bfqq->bfqd->lock);
1752 bfqg_stats_update_io_merged(bfqq_group(bfqq), next->cmd_flags);
1755 /* Must be called with bfqq != NULL */
1756 static void bfq_bfqq_end_wr(struct bfq_queue *bfqq)
1758 if (bfq_bfqq_busy(bfqq))
1759 bfqq->bfqd->wr_busy_queues--;
1761 bfqq->wr_cur_max_time = 0;
1762 bfqq->last_wr_start_finish = jiffies;
1764 * Trigger a weight change on the next invocation of
1765 * __bfq_entity_update_weight_prio.
1767 bfqq->entity.prio_changed = 1;
1770 void bfq_end_wr_async_queues(struct bfq_data *bfqd,
1771 struct bfq_group *bfqg)
1775 for (i = 0; i < 2; i++)
1776 for (j = 0; j < IOPRIO_BE_NR; j++)
1777 if (bfqg->async_bfqq[i][j])
1778 bfq_bfqq_end_wr(bfqg->async_bfqq[i][j]);
1779 if (bfqg->async_idle_bfqq)
1780 bfq_bfqq_end_wr(bfqg->async_idle_bfqq);
1783 static void bfq_end_wr(struct bfq_data *bfqd)
1785 struct bfq_queue *bfqq;
1787 spin_lock_irq(&bfqd->lock);
1789 list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
1790 bfq_bfqq_end_wr(bfqq);
1791 list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list)
1792 bfq_bfqq_end_wr(bfqq);
1793 bfq_end_wr_async(bfqd);
1795 spin_unlock_irq(&bfqd->lock);
1798 static sector_t bfq_io_struct_pos(void *io_struct, bool request)
1801 return blk_rq_pos(io_struct);
1803 return ((struct bio *)io_struct)->bi_iter.bi_sector;
1806 static int bfq_rq_close_to_sector(void *io_struct, bool request,
1809 return abs(bfq_io_struct_pos(io_struct, request) - sector) <=
1813 static struct bfq_queue *bfqq_find_close(struct bfq_data *bfqd,
1814 struct bfq_queue *bfqq,
1817 struct rb_root *root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree;
1818 struct rb_node *parent, *node;
1819 struct bfq_queue *__bfqq;
1821 if (RB_EMPTY_ROOT(root))
1825 * First, if we find a request starting at the end of the last
1826 * request, choose it.
1828 __bfqq = bfq_rq_pos_tree_lookup(bfqd, root, sector, &parent, NULL);
1833 * If the exact sector wasn't found, the parent of the NULL leaf
1834 * will contain the closest sector (rq_pos_tree sorted by
1835 * next_request position).
1837 __bfqq = rb_entry(parent, struct bfq_queue, pos_node);
1838 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
1841 if (blk_rq_pos(__bfqq->next_rq) < sector)
1842 node = rb_next(&__bfqq->pos_node);
1844 node = rb_prev(&__bfqq->pos_node);
1848 __bfqq = rb_entry(node, struct bfq_queue, pos_node);
1849 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
1855 static struct bfq_queue *bfq_find_close_cooperator(struct bfq_data *bfqd,
1856 struct bfq_queue *cur_bfqq,
1859 struct bfq_queue *bfqq;
1862 * We shall notice if some of the queues are cooperating,
1863 * e.g., working closely on the same area of the device. In
1864 * that case, we can group them together and: 1) don't waste
1865 * time idling, and 2) serve the union of their requests in
1866 * the best possible order for throughput.
1868 bfqq = bfqq_find_close(bfqd, cur_bfqq, sector);
1869 if (!bfqq || bfqq == cur_bfqq)
1875 static struct bfq_queue *
1876 bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
1878 int process_refs, new_process_refs;
1879 struct bfq_queue *__bfqq;
1882 * If there are no process references on the new_bfqq, then it is
1883 * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain
1884 * may have dropped their last reference (not just their last process
1887 if (!bfqq_process_refs(new_bfqq))
1890 /* Avoid a circular list and skip interim queue merges. */
1891 while ((__bfqq = new_bfqq->new_bfqq)) {
1897 process_refs = bfqq_process_refs(bfqq);
1898 new_process_refs = bfqq_process_refs(new_bfqq);
1900 * If the process for the bfqq has gone away, there is no
1901 * sense in merging the queues.
1903 if (process_refs == 0 || new_process_refs == 0)
1906 bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d",
1910 * Merging is just a redirection: the requests of the process
1911 * owning one of the two queues are redirected to the other queue.
1912 * The latter queue, in its turn, is set as shared if this is the
1913 * first time that the requests of some process are redirected to
1916 * We redirect bfqq to new_bfqq and not the opposite, because
1917 * we are in the context of the process owning bfqq, thus we
1918 * have the io_cq of this process. So we can immediately
1919 * configure this io_cq to redirect the requests of the
1920 * process to new_bfqq. In contrast, the io_cq of new_bfqq is
1921 * not available any more (new_bfqq->bic == NULL).
1923 * Anyway, even in case new_bfqq coincides with the in-service
1924 * queue, redirecting requests the in-service queue is the
1925 * best option, as we feed the in-service queue with new
1926 * requests close to the last request served and, by doing so,
1927 * are likely to increase the throughput.
1929 bfqq->new_bfqq = new_bfqq;
1930 new_bfqq->ref += process_refs;
1934 static bool bfq_may_be_close_cooperator(struct bfq_queue *bfqq,
1935 struct bfq_queue *new_bfqq)
1937 if (bfq_class_idle(bfqq) || bfq_class_idle(new_bfqq) ||
1938 (bfqq->ioprio_class != new_bfqq->ioprio_class))
1942 * If either of the queues has already been detected as seeky,
1943 * then merging it with the other queue is unlikely to lead to
1946 if (BFQQ_SEEKY(bfqq) || BFQQ_SEEKY(new_bfqq))
1950 * Interleaved I/O is known to be done by (some) applications
1951 * only for reads, so it does not make sense to merge async
1954 if (!bfq_bfqq_sync(bfqq) || !bfq_bfqq_sync(new_bfqq))
1961 * If this function returns true, then bfqq cannot be merged. The idea
1962 * is that true cooperation happens very early after processes start
1963 * to do I/O. Usually, late cooperations are just accidental false
1964 * positives. In case bfqq is weight-raised, such false positives
1965 * would evidently degrade latency guarantees for bfqq.
1967 static bool wr_from_too_long(struct bfq_queue *bfqq)
1969 return bfqq->wr_coeff > 1 &&
1970 time_is_before_jiffies(bfqq->last_wr_start_finish +
1971 msecs_to_jiffies(100));
1975 * Attempt to schedule a merge of bfqq with the currently in-service
1976 * queue or with a close queue among the scheduled queues. Return
1977 * NULL if no merge was scheduled, a pointer to the shared bfq_queue
1978 * structure otherwise.
1980 * The OOM queue is not allowed to participate to cooperation: in fact, since
1981 * the requests temporarily redirected to the OOM queue could be redirected
1982 * again to dedicated queues at any time, the state needed to correctly
1983 * handle merging with the OOM queue would be quite complex and expensive
1984 * to maintain. Besides, in such a critical condition as an out of memory,
1985 * the benefits of queue merging may be little relevant, or even negligible.
1987 * Weight-raised queues can be merged only if their weight-raising
1988 * period has just started. In fact cooperating processes are usually
1989 * started together. Thus, with this filter we avoid false positives
1990 * that would jeopardize low-latency guarantees.
1992 * WARNING: queue merging may impair fairness among non-weight raised
1993 * queues, for at least two reasons: 1) the original weight of a
1994 * merged queue may change during the merged state, 2) even being the
1995 * weight the same, a merged queue may be bloated with many more
1996 * requests than the ones produced by its originally-associated
1999 static struct bfq_queue *
2000 bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2001 void *io_struct, bool request)
2003 struct bfq_queue *in_service_bfqq, *new_bfqq;
2006 return bfqq->new_bfqq;
2009 wr_from_too_long(bfqq) ||
2010 unlikely(bfqq == &bfqd->oom_bfqq))
2013 /* If there is only one backlogged queue, don't search. */
2014 if (bfqd->busy_queues == 1)
2017 in_service_bfqq = bfqd->in_service_queue;
2019 if (!in_service_bfqq || in_service_bfqq == bfqq
2020 || wr_from_too_long(in_service_bfqq) ||
2021 unlikely(in_service_bfqq == &bfqd->oom_bfqq))
2022 goto check_scheduled;
2024 if (bfq_rq_close_to_sector(io_struct, request, bfqd->last_position) &&
2025 bfqq->entity.parent == in_service_bfqq->entity.parent &&
2026 bfq_may_be_close_cooperator(bfqq, in_service_bfqq)) {
2027 new_bfqq = bfq_setup_merge(bfqq, in_service_bfqq);
2032 * Check whether there is a cooperator among currently scheduled
2033 * queues. The only thing we need is that the bio/request is not
2034 * NULL, as we need it to establish whether a cooperator exists.
2037 new_bfqq = bfq_find_close_cooperator(bfqd, bfqq,
2038 bfq_io_struct_pos(io_struct, request));
2040 if (new_bfqq && !wr_from_too_long(new_bfqq) &&
2041 likely(new_bfqq != &bfqd->oom_bfqq) &&
2042 bfq_may_be_close_cooperator(bfqq, new_bfqq))
2043 return bfq_setup_merge(bfqq, new_bfqq);
2048 static void bfq_bfqq_save_state(struct bfq_queue *bfqq)
2050 struct bfq_io_cq *bic = bfqq->bic;
2053 * If !bfqq->bic, the queue is already shared or its requests
2054 * have already been redirected to a shared queue; both idle window
2055 * and weight raising state have already been saved. Do nothing.
2060 bic->saved_ttime = bfqq->ttime;
2061 bic->saved_has_short_ttime = bfq_bfqq_has_short_ttime(bfqq);
2062 bic->saved_IO_bound = bfq_bfqq_IO_bound(bfqq);
2063 bic->saved_in_large_burst = bfq_bfqq_in_large_burst(bfqq);
2064 bic->was_in_burst_list = !hlist_unhashed(&bfqq->burst_list_node);
2065 if (unlikely(bfq_bfqq_just_created(bfqq) &&
2066 !bfq_bfqq_in_large_burst(bfqq))) {
2068 * bfqq being merged right after being created: bfqq
2069 * would have deserved interactive weight raising, but
2070 * did not make it to be set in a weight-raised state,
2071 * because of this early merge. Store directly the
2072 * weight-raising state that would have been assigned
2073 * to bfqq, so that to avoid that bfqq unjustly fails
2074 * to enjoy weight raising if split soon.
2076 bic->saved_wr_coeff = bfqq->bfqd->bfq_wr_coeff;
2077 bic->saved_wr_cur_max_time = bfq_wr_duration(bfqq->bfqd);
2078 bic->saved_last_wr_start_finish = jiffies;
2080 bic->saved_wr_coeff = bfqq->wr_coeff;
2081 bic->saved_wr_start_at_switch_to_srt =
2082 bfqq->wr_start_at_switch_to_srt;
2083 bic->saved_last_wr_start_finish = bfqq->last_wr_start_finish;
2084 bic->saved_wr_cur_max_time = bfqq->wr_cur_max_time;
2089 bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic,
2090 struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2092 bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu",
2093 (unsigned long)new_bfqq->pid);
2094 /* Save weight raising and idle window of the merged queues */
2095 bfq_bfqq_save_state(bfqq);
2096 bfq_bfqq_save_state(new_bfqq);
2097 if (bfq_bfqq_IO_bound(bfqq))
2098 bfq_mark_bfqq_IO_bound(new_bfqq);
2099 bfq_clear_bfqq_IO_bound(bfqq);
2102 * If bfqq is weight-raised, then let new_bfqq inherit
2103 * weight-raising. To reduce false positives, neglect the case
2104 * where bfqq has just been created, but has not yet made it
2105 * to be weight-raised (which may happen because EQM may merge
2106 * bfqq even before bfq_add_request is executed for the first
2107 * time for bfqq). Handling this case would however be very
2108 * easy, thanks to the flag just_created.
2110 if (new_bfqq->wr_coeff == 1 && bfqq->wr_coeff > 1) {
2111 new_bfqq->wr_coeff = bfqq->wr_coeff;
2112 new_bfqq->wr_cur_max_time = bfqq->wr_cur_max_time;
2113 new_bfqq->last_wr_start_finish = bfqq->last_wr_start_finish;
2114 new_bfqq->wr_start_at_switch_to_srt =
2115 bfqq->wr_start_at_switch_to_srt;
2116 if (bfq_bfqq_busy(new_bfqq))
2117 bfqd->wr_busy_queues++;
2118 new_bfqq->entity.prio_changed = 1;
2121 if (bfqq->wr_coeff > 1) { /* bfqq has given its wr to new_bfqq */
2123 bfqq->entity.prio_changed = 1;
2124 if (bfq_bfqq_busy(bfqq))
2125 bfqd->wr_busy_queues--;
2128 bfq_log_bfqq(bfqd, new_bfqq, "merge_bfqqs: wr_busy %d",
2129 bfqd->wr_busy_queues);
2132 * Merge queues (that is, let bic redirect its requests to new_bfqq)
2134 bic_set_bfqq(bic, new_bfqq, 1);
2135 bfq_mark_bfqq_coop(new_bfqq);
2137 * new_bfqq now belongs to at least two bics (it is a shared queue):
2138 * set new_bfqq->bic to NULL. bfqq either:
2139 * - does not belong to any bic any more, and hence bfqq->bic must
2140 * be set to NULL, or
2141 * - is a queue whose owning bics have already been redirected to a
2142 * different queue, hence the queue is destined to not belong to
2143 * any bic soon and bfqq->bic is already NULL (therefore the next
2144 * assignment causes no harm).
2146 new_bfqq->bic = NULL;
2148 /* release process reference to bfqq */
2149 bfq_put_queue(bfqq);
2152 static bool bfq_allow_bio_merge(struct request_queue *q, struct request *rq,
2155 struct bfq_data *bfqd = q->elevator->elevator_data;
2156 bool is_sync = op_is_sync(bio->bi_opf);
2157 struct bfq_queue *bfqq = bfqd->bio_bfqq, *new_bfqq;
2160 * Disallow merge of a sync bio into an async request.
2162 if (is_sync && !rq_is_sync(rq))
2166 * Lookup the bfqq that this bio will be queued with. Allow
2167 * merge only if rq is queued there.
2173 * We take advantage of this function to perform an early merge
2174 * of the queues of possible cooperating processes.
2176 new_bfqq = bfq_setup_cooperator(bfqd, bfqq, bio, false);
2179 * bic still points to bfqq, then it has not yet been
2180 * redirected to some other bfq_queue, and a queue
2181 * merge beween bfqq and new_bfqq can be safely
2182 * fulfillled, i.e., bic can be redirected to new_bfqq
2183 * and bfqq can be put.
2185 bfq_merge_bfqqs(bfqd, bfqd->bio_bic, bfqq,
2188 * If we get here, bio will be queued into new_queue,
2189 * so use new_bfqq to decide whether bio and rq can be
2195 * Change also bqfd->bio_bfqq, as
2196 * bfqd->bio_bic now points to new_bfqq, and
2197 * this function may be invoked again (and then may
2198 * use again bqfd->bio_bfqq).
2200 bfqd->bio_bfqq = bfqq;
2203 return bfqq == RQ_BFQQ(rq);
2207 * Set the maximum time for the in-service queue to consume its
2208 * budget. This prevents seeky processes from lowering the throughput.
2209 * In practice, a time-slice service scheme is used with seeky
2212 static void bfq_set_budget_timeout(struct bfq_data *bfqd,
2213 struct bfq_queue *bfqq)
2215 unsigned int timeout_coeff;
2217 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time)
2220 timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight;
2222 bfqd->last_budget_start = ktime_get();
2224 bfqq->budget_timeout = jiffies +
2225 bfqd->bfq_timeout * timeout_coeff;
2228 static void __bfq_set_in_service_queue(struct bfq_data *bfqd,
2229 struct bfq_queue *bfqq)
2232 bfqg_stats_update_avg_queue_size(bfqq_group(bfqq));
2233 bfq_clear_bfqq_fifo_expire(bfqq);
2235 bfqd->budgets_assigned = (bfqd->budgets_assigned * 7 + 256) / 8;
2237 if (time_is_before_jiffies(bfqq->last_wr_start_finish) &&
2238 bfqq->wr_coeff > 1 &&
2239 bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
2240 time_is_before_jiffies(bfqq->budget_timeout)) {
2242 * For soft real-time queues, move the start
2243 * of the weight-raising period forward by the
2244 * time the queue has not received any
2245 * service. Otherwise, a relatively long
2246 * service delay is likely to cause the
2247 * weight-raising period of the queue to end,
2248 * because of the short duration of the
2249 * weight-raising period of a soft real-time
2250 * queue. It is worth noting that this move
2251 * is not so dangerous for the other queues,
2252 * because soft real-time queues are not
2255 * To not add a further variable, we use the
2256 * overloaded field budget_timeout to
2257 * determine for how long the queue has not
2258 * received service, i.e., how much time has
2259 * elapsed since the queue expired. However,
2260 * this is a little imprecise, because
2261 * budget_timeout is set to jiffies if bfqq
2262 * not only expires, but also remains with no
2265 if (time_after(bfqq->budget_timeout,
2266 bfqq->last_wr_start_finish))
2267 bfqq->last_wr_start_finish +=
2268 jiffies - bfqq->budget_timeout;
2270 bfqq->last_wr_start_finish = jiffies;
2273 bfq_set_budget_timeout(bfqd, bfqq);
2274 bfq_log_bfqq(bfqd, bfqq,
2275 "set_in_service_queue, cur-budget = %d",
2276 bfqq->entity.budget);
2279 bfqd->in_service_queue = bfqq;
2283 * Get and set a new queue for service.
2285 static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd)
2287 struct bfq_queue *bfqq = bfq_get_next_queue(bfqd);
2289 __bfq_set_in_service_queue(bfqd, bfqq);
2293 static void bfq_arm_slice_timer(struct bfq_data *bfqd)
2295 struct bfq_queue *bfqq = bfqd->in_service_queue;
2298 bfq_mark_bfqq_wait_request(bfqq);
2301 * We don't want to idle for seeks, but we do want to allow
2302 * fair distribution of slice time for a process doing back-to-back
2303 * seeks. So allow a little bit of time for him to submit a new rq.
2305 sl = bfqd->bfq_slice_idle;
2307 * Unless the queue is being weight-raised or the scenario is
2308 * asymmetric, grant only minimum idle time if the queue
2309 * is seeky. A long idling is preserved for a weight-raised
2310 * queue, or, more in general, in an asymmetric scenario,
2311 * because a long idling is needed for guaranteeing to a queue
2312 * its reserved share of the throughput (in particular, it is
2313 * needed if the queue has a higher weight than some other
2316 if (BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 &&
2317 bfq_symmetric_scenario(bfqd))
2318 sl = min_t(u64, sl, BFQ_MIN_TT);
2320 bfqd->last_idling_start = ktime_get();
2321 hrtimer_start(&bfqd->idle_slice_timer, ns_to_ktime(sl),
2323 bfqg_stats_set_start_idle_time(bfqq_group(bfqq));
2327 * In autotuning mode, max_budget is dynamically recomputed as the
2328 * amount of sectors transferred in timeout at the estimated peak
2329 * rate. This enables BFQ to utilize a full timeslice with a full
2330 * budget, even if the in-service queue is served at peak rate. And
2331 * this maximises throughput with sequential workloads.
2333 static unsigned long bfq_calc_max_budget(struct bfq_data *bfqd)
2335 return (u64)bfqd->peak_rate * USEC_PER_MSEC *
2336 jiffies_to_msecs(bfqd->bfq_timeout)>>BFQ_RATE_SHIFT;
2340 * Update parameters related to throughput and responsiveness, as a
2341 * function of the estimated peak rate. See comments on
2342 * bfq_calc_max_budget(), and on T_slow and T_fast arrays.
2344 static void update_thr_responsiveness_params(struct bfq_data *bfqd)
2346 int dev_type = blk_queue_nonrot(bfqd->queue);
2348 if (bfqd->bfq_user_max_budget == 0)
2349 bfqd->bfq_max_budget =
2350 bfq_calc_max_budget(bfqd);
2352 if (bfqd->device_speed == BFQ_BFQD_FAST &&
2353 bfqd->peak_rate < device_speed_thresh[dev_type]) {
2354 bfqd->device_speed = BFQ_BFQD_SLOW;
2355 bfqd->RT_prod = R_slow[dev_type] *
2357 } else if (bfqd->device_speed == BFQ_BFQD_SLOW &&
2358 bfqd->peak_rate > device_speed_thresh[dev_type]) {
2359 bfqd->device_speed = BFQ_BFQD_FAST;
2360 bfqd->RT_prod = R_fast[dev_type] *
2365 "dev_type %s dev_speed_class = %s (%llu sects/sec), thresh %llu setcs/sec",
2366 dev_type == 0 ? "ROT" : "NONROT",
2367 bfqd->device_speed == BFQ_BFQD_FAST ? "FAST" : "SLOW",
2368 bfqd->device_speed == BFQ_BFQD_FAST ?
2369 (USEC_PER_SEC*(u64)R_fast[dev_type])>>BFQ_RATE_SHIFT :
2370 (USEC_PER_SEC*(u64)R_slow[dev_type])>>BFQ_RATE_SHIFT,
2371 (USEC_PER_SEC*(u64)device_speed_thresh[dev_type])>>
2375 static void bfq_reset_rate_computation(struct bfq_data *bfqd,
2378 if (rq != NULL) { /* new rq dispatch now, reset accordingly */
2379 bfqd->last_dispatch = bfqd->first_dispatch = ktime_get_ns();
2380 bfqd->peak_rate_samples = 1;
2381 bfqd->sequential_samples = 0;
2382 bfqd->tot_sectors_dispatched = bfqd->last_rq_max_size =
2384 } else /* no new rq dispatched, just reset the number of samples */
2385 bfqd->peak_rate_samples = 0; /* full re-init on next disp. */
2388 "reset_rate_computation at end, sample %u/%u tot_sects %llu",
2389 bfqd->peak_rate_samples, bfqd->sequential_samples,
2390 bfqd->tot_sectors_dispatched);
2393 static void bfq_update_rate_reset(struct bfq_data *bfqd, struct request *rq)
2395 u32 rate, weight, divisor;
2398 * For the convergence property to hold (see comments on
2399 * bfq_update_peak_rate()) and for the assessment to be
2400 * reliable, a minimum number of samples must be present, and
2401 * a minimum amount of time must have elapsed. If not so, do
2402 * not compute new rate. Just reset parameters, to get ready
2403 * for a new evaluation attempt.
2405 if (bfqd->peak_rate_samples < BFQ_RATE_MIN_SAMPLES ||
2406 bfqd->delta_from_first < BFQ_RATE_MIN_INTERVAL)
2407 goto reset_computation;
2410 * If a new request completion has occurred after last
2411 * dispatch, then, to approximate the rate at which requests
2412 * have been served by the device, it is more precise to
2413 * extend the observation interval to the last completion.
2415 bfqd->delta_from_first =
2416 max_t(u64, bfqd->delta_from_first,
2417 bfqd->last_completion - bfqd->first_dispatch);
2420 * Rate computed in sects/usec, and not sects/nsec, for
2423 rate = div64_ul(bfqd->tot_sectors_dispatched<<BFQ_RATE_SHIFT,
2424 div_u64(bfqd->delta_from_first, NSEC_PER_USEC));
2427 * Peak rate not updated if:
2428 * - the percentage of sequential dispatches is below 3/4 of the
2429 * total, and rate is below the current estimated peak rate
2430 * - rate is unreasonably high (> 20M sectors/sec)
2432 if ((bfqd->sequential_samples < (3 * bfqd->peak_rate_samples)>>2 &&
2433 rate <= bfqd->peak_rate) ||
2434 rate > 20<<BFQ_RATE_SHIFT)
2435 goto reset_computation;
2438 * We have to update the peak rate, at last! To this purpose,
2439 * we use a low-pass filter. We compute the smoothing constant
2440 * of the filter as a function of the 'weight' of the new
2443 * As can be seen in next formulas, we define this weight as a
2444 * quantity proportional to how sequential the workload is,
2445 * and to how long the observation time interval is.
2447 * The weight runs from 0 to 8. The maximum value of the
2448 * weight, 8, yields the minimum value for the smoothing
2449 * constant. At this minimum value for the smoothing constant,
2450 * the measured rate contributes for half of the next value of
2451 * the estimated peak rate.
2453 * So, the first step is to compute the weight as a function
2454 * of how sequential the workload is. Note that the weight
2455 * cannot reach 9, because bfqd->sequential_samples cannot
2456 * become equal to bfqd->peak_rate_samples, which, in its
2457 * turn, holds true because bfqd->sequential_samples is not
2458 * incremented for the first sample.
2460 weight = (9 * bfqd->sequential_samples) / bfqd->peak_rate_samples;
2463 * Second step: further refine the weight as a function of the
2464 * duration of the observation interval.
2466 weight = min_t(u32, 8,
2467 div_u64(weight * bfqd->delta_from_first,
2468 BFQ_RATE_REF_INTERVAL));
2471 * Divisor ranging from 10, for minimum weight, to 2, for
2474 divisor = 10 - weight;
2477 * Finally, update peak rate:
2479 * peak_rate = peak_rate * (divisor-1) / divisor + rate / divisor
2481 bfqd->peak_rate *= divisor-1;
2482 bfqd->peak_rate /= divisor;
2483 rate /= divisor; /* smoothing constant alpha = 1/divisor */
2485 bfqd->peak_rate += rate;
2486 update_thr_responsiveness_params(bfqd);
2489 bfq_reset_rate_computation(bfqd, rq);
2493 * Update the read/write peak rate (the main quantity used for
2494 * auto-tuning, see update_thr_responsiveness_params()).
2496 * It is not trivial to estimate the peak rate (correctly): because of
2497 * the presence of sw and hw queues between the scheduler and the
2498 * device components that finally serve I/O requests, it is hard to
2499 * say exactly when a given dispatched request is served inside the
2500 * device, and for how long. As a consequence, it is hard to know
2501 * precisely at what rate a given set of requests is actually served
2504 * On the opposite end, the dispatch time of any request is trivially
2505 * available, and, from this piece of information, the "dispatch rate"
2506 * of requests can be immediately computed. So, the idea in the next
2507 * function is to use what is known, namely request dispatch times
2508 * (plus, when useful, request completion times), to estimate what is
2509 * unknown, namely in-device request service rate.
2511 * The main issue is that, because of the above facts, the rate at
2512 * which a certain set of requests is dispatched over a certain time
2513 * interval can vary greatly with respect to the rate at which the
2514 * same requests are then served. But, since the size of any
2515 * intermediate queue is limited, and the service scheme is lossless
2516 * (no request is silently dropped), the following obvious convergence
2517 * property holds: the number of requests dispatched MUST become
2518 * closer and closer to the number of requests completed as the
2519 * observation interval grows. This is the key property used in
2520 * the next function to estimate the peak service rate as a function
2521 * of the observed dispatch rate. The function assumes to be invoked
2522 * on every request dispatch.
2524 static void bfq_update_peak_rate(struct bfq_data *bfqd, struct request *rq)
2526 u64 now_ns = ktime_get_ns();
2528 if (bfqd->peak_rate_samples == 0) { /* first dispatch */
2529 bfq_log(bfqd, "update_peak_rate: goto reset, samples %d",
2530 bfqd->peak_rate_samples);
2531 bfq_reset_rate_computation(bfqd, rq);
2532 goto update_last_values; /* will add one sample */
2536 * Device idle for very long: the observation interval lasting
2537 * up to this dispatch cannot be a valid observation interval
2538 * for computing a new peak rate (similarly to the late-
2539 * completion event in bfq_completed_request()). Go to
2540 * update_rate_and_reset to have the following three steps
2542 * - close the observation interval at the last (previous)
2543 * request dispatch or completion
2544 * - compute rate, if possible, for that observation interval
2545 * - start a new observation interval with this dispatch
2547 if (now_ns - bfqd->last_dispatch > 100*NSEC_PER_MSEC &&
2548 bfqd->rq_in_driver == 0)
2549 goto update_rate_and_reset;
2551 /* Update sampling information */
2552 bfqd->peak_rate_samples++;
2554 if ((bfqd->rq_in_driver > 0 ||
2555 now_ns - bfqd->last_completion < BFQ_MIN_TT)
2556 && get_sdist(bfqd->last_position, rq) < BFQQ_SEEK_THR)
2557 bfqd->sequential_samples++;
2559 bfqd->tot_sectors_dispatched += blk_rq_sectors(rq);
2561 /* Reset max observed rq size every 32 dispatches */
2562 if (likely(bfqd->peak_rate_samples % 32))
2563 bfqd->last_rq_max_size =
2564 max_t(u32, blk_rq_sectors(rq), bfqd->last_rq_max_size);
2566 bfqd->last_rq_max_size = blk_rq_sectors(rq);
2568 bfqd->delta_from_first = now_ns - bfqd->first_dispatch;
2570 /* Target observation interval not yet reached, go on sampling */
2571 if (bfqd->delta_from_first < BFQ_RATE_REF_INTERVAL)
2572 goto update_last_values;
2574 update_rate_and_reset:
2575 bfq_update_rate_reset(bfqd, rq);
2577 bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2578 bfqd->last_dispatch = now_ns;
2582 * Remove request from internal lists.
2584 static void bfq_dispatch_remove(struct request_queue *q, struct request *rq)
2586 struct bfq_queue *bfqq = RQ_BFQQ(rq);
2589 * For consistency, the next instruction should have been
2590 * executed after removing the request from the queue and
2591 * dispatching it. We execute instead this instruction before
2592 * bfq_remove_request() (and hence introduce a temporary
2593 * inconsistency), for efficiency. In fact, should this
2594 * dispatch occur for a non in-service bfqq, this anticipated
2595 * increment prevents two counters related to bfqq->dispatched
2596 * from risking to be, first, uselessly decremented, and then
2597 * incremented again when the (new) value of bfqq->dispatched
2598 * happens to be taken into account.
2601 bfq_update_peak_rate(q->elevator->elevator_data, rq);
2603 bfq_remove_request(q, rq);
2606 static void __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq)
2609 * If this bfqq is shared between multiple processes, check
2610 * to make sure that those processes are still issuing I/Os
2611 * within the mean seek distance. If not, it may be time to
2612 * break the queues apart again.
2614 if (bfq_bfqq_coop(bfqq) && BFQQ_SEEKY(bfqq))
2615 bfq_mark_bfqq_split_coop(bfqq);
2617 if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
2618 if (bfqq->dispatched == 0)
2620 * Overloading budget_timeout field to store
2621 * the time at which the queue remains with no
2622 * backlog and no outstanding request; used by
2623 * the weight-raising mechanism.
2625 bfqq->budget_timeout = jiffies;
2627 bfq_del_bfqq_busy(bfqd, bfqq, true);
2629 bfq_requeue_bfqq(bfqd, bfqq, true);
2631 * Resort priority tree of potential close cooperators.
2633 bfq_pos_tree_add_move(bfqd, bfqq);
2637 * All in-service entities must have been properly deactivated
2638 * or requeued before executing the next function, which
2639 * resets all in-service entites as no more in service.
2641 __bfq_bfqd_reset_in_service(bfqd);
2645 * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior.
2646 * @bfqd: device data.
2647 * @bfqq: queue to update.
2648 * @reason: reason for expiration.
2650 * Handle the feedback on @bfqq budget at queue expiration.
2651 * See the body for detailed comments.
2653 static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd,
2654 struct bfq_queue *bfqq,
2655 enum bfqq_expiration reason)
2657 struct request *next_rq;
2658 int budget, min_budget;
2660 min_budget = bfq_min_budget(bfqd);
2662 if (bfqq->wr_coeff == 1)
2663 budget = bfqq->max_budget;
2665 * Use a constant, low budget for weight-raised queues,
2666 * to help achieve a low latency. Keep it slightly higher
2667 * than the minimum possible budget, to cause a little
2668 * bit fewer expirations.
2670 budget = 2 * min_budget;
2672 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %d, budg left %d",
2673 bfqq->entity.budget, bfq_bfqq_budget_left(bfqq));
2674 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %d, min budg %d",
2675 budget, bfq_min_budget(bfqd));
2676 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d",
2677 bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->in_service_queue));
2679 if (bfq_bfqq_sync(bfqq) && bfqq->wr_coeff == 1) {
2682 * Caveat: in all the following cases we trade latency
2685 case BFQQE_TOO_IDLE:
2687 * This is the only case where we may reduce
2688 * the budget: if there is no request of the
2689 * process still waiting for completion, then
2690 * we assume (tentatively) that the timer has
2691 * expired because the batch of requests of
2692 * the process could have been served with a
2693 * smaller budget. Hence, betting that
2694 * process will behave in the same way when it
2695 * becomes backlogged again, we reduce its
2696 * next budget. As long as we guess right,
2697 * this budget cut reduces the latency
2698 * experienced by the process.
2700 * However, if there are still outstanding
2701 * requests, then the process may have not yet
2702 * issued its next request just because it is
2703 * still waiting for the completion of some of
2704 * the still outstanding ones. So in this
2705 * subcase we do not reduce its budget, on the
2706 * contrary we increase it to possibly boost
2707 * the throughput, as discussed in the
2708 * comments to the BUDGET_TIMEOUT case.
2710 if (bfqq->dispatched > 0) /* still outstanding reqs */
2711 budget = min(budget * 2, bfqd->bfq_max_budget);
2713 if (budget > 5 * min_budget)
2714 budget -= 4 * min_budget;
2716 budget = min_budget;
2719 case BFQQE_BUDGET_TIMEOUT:
2721 * We double the budget here because it gives
2722 * the chance to boost the throughput if this
2723 * is not a seeky process (and has bumped into
2724 * this timeout because of, e.g., ZBR).
2726 budget = min(budget * 2, bfqd->bfq_max_budget);
2728 case BFQQE_BUDGET_EXHAUSTED:
2730 * The process still has backlog, and did not
2731 * let either the budget timeout or the disk
2732 * idling timeout expire. Hence it is not
2733 * seeky, has a short thinktime and may be
2734 * happy with a higher budget too. So
2735 * definitely increase the budget of this good
2736 * candidate to boost the disk throughput.
2738 budget = min(budget * 4, bfqd->bfq_max_budget);
2740 case BFQQE_NO_MORE_REQUESTS:
2742 * For queues that expire for this reason, it
2743 * is particularly important to keep the
2744 * budget close to the actual service they
2745 * need. Doing so reduces the timestamp
2746 * misalignment problem described in the
2747 * comments in the body of
2748 * __bfq_activate_entity. In fact, suppose
2749 * that a queue systematically expires for
2750 * BFQQE_NO_MORE_REQUESTS and presents a
2751 * new request in time to enjoy timestamp
2752 * back-shifting. The larger the budget of the
2753 * queue is with respect to the service the
2754 * queue actually requests in each service
2755 * slot, the more times the queue can be
2756 * reactivated with the same virtual finish
2757 * time. It follows that, even if this finish
2758 * time is pushed to the system virtual time
2759 * to reduce the consequent timestamp
2760 * misalignment, the queue unjustly enjoys for
2761 * many re-activations a lower finish time
2762 * than all newly activated queues.
2764 * The service needed by bfqq is measured
2765 * quite precisely by bfqq->entity.service.
2766 * Since bfqq does not enjoy device idling,
2767 * bfqq->entity.service is equal to the number
2768 * of sectors that the process associated with
2769 * bfqq requested to read/write before waiting
2770 * for request completions, or blocking for
2773 budget = max_t(int, bfqq->entity.service, min_budget);
2778 } else if (!bfq_bfqq_sync(bfqq)) {
2780 * Async queues get always the maximum possible
2781 * budget, as for them we do not care about latency
2782 * (in addition, their ability to dispatch is limited
2783 * by the charging factor).
2785 budget = bfqd->bfq_max_budget;
2788 bfqq->max_budget = budget;
2790 if (bfqd->budgets_assigned >= bfq_stats_min_budgets &&
2791 !bfqd->bfq_user_max_budget)
2792 bfqq->max_budget = min(bfqq->max_budget, bfqd->bfq_max_budget);
2795 * If there is still backlog, then assign a new budget, making
2796 * sure that it is large enough for the next request. Since
2797 * the finish time of bfqq must be kept in sync with the
2798 * budget, be sure to call __bfq_bfqq_expire() *after* this
2801 * If there is no backlog, then no need to update the budget;
2802 * it will be updated on the arrival of a new request.
2804 next_rq = bfqq->next_rq;
2806 bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget,
2807 bfq_serv_to_charge(next_rq, bfqq));
2809 bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %d",
2810 next_rq ? blk_rq_sectors(next_rq) : 0,
2811 bfqq->entity.budget);
2815 * Return true if the process associated with bfqq is "slow". The slow
2816 * flag is used, in addition to the budget timeout, to reduce the
2817 * amount of service provided to seeky processes, and thus reduce
2818 * their chances to lower the throughput. More details in the comments
2819 * on the function bfq_bfqq_expire().
2821 * An important observation is in order: as discussed in the comments
2822 * on the function bfq_update_peak_rate(), with devices with internal
2823 * queues, it is hard if ever possible to know when and for how long
2824 * an I/O request is processed by the device (apart from the trivial
2825 * I/O pattern where a new request is dispatched only after the
2826 * previous one has been completed). This makes it hard to evaluate
2827 * the real rate at which the I/O requests of each bfq_queue are
2828 * served. In fact, for an I/O scheduler like BFQ, serving a
2829 * bfq_queue means just dispatching its requests during its service
2830 * slot (i.e., until the budget of the queue is exhausted, or the
2831 * queue remains idle, or, finally, a timeout fires). But, during the
2832 * service slot of a bfq_queue, around 100 ms at most, the device may
2833 * be even still processing requests of bfq_queues served in previous
2834 * service slots. On the opposite end, the requests of the in-service
2835 * bfq_queue may be completed after the service slot of the queue
2838 * Anyway, unless more sophisticated solutions are used
2839 * (where possible), the sum of the sizes of the requests dispatched
2840 * during the service slot of a bfq_queue is probably the only
2841 * approximation available for the service received by the bfq_queue
2842 * during its service slot. And this sum is the quantity used in this
2843 * function to evaluate the I/O speed of a process.
2845 static bool bfq_bfqq_is_slow(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2846 bool compensate, enum bfqq_expiration reason,
2847 unsigned long *delta_ms)
2849 ktime_t delta_ktime;
2851 bool slow = BFQQ_SEEKY(bfqq); /* if delta too short, use seekyness */
2853 if (!bfq_bfqq_sync(bfqq))
2857 delta_ktime = bfqd->last_idling_start;
2859 delta_ktime = ktime_get();
2860 delta_ktime = ktime_sub(delta_ktime, bfqd->last_budget_start);
2861 delta_usecs = ktime_to_us(delta_ktime);
2863 /* don't use too short time intervals */
2864 if (delta_usecs < 1000) {
2865 if (blk_queue_nonrot(bfqd->queue))
2867 * give same worst-case guarantees as idling
2870 *delta_ms = BFQ_MIN_TT / NSEC_PER_MSEC;
2871 else /* charge at least one seek */
2872 *delta_ms = bfq_slice_idle / NSEC_PER_MSEC;
2877 *delta_ms = delta_usecs / USEC_PER_MSEC;
2880 * Use only long (> 20ms) intervals to filter out excessive
2881 * spikes in service rate estimation.
2883 if (delta_usecs > 20000) {
2885 * Caveat for rotational devices: processes doing I/O
2886 * in the slower disk zones tend to be slow(er) even
2887 * if not seeky. In this respect, the estimated peak
2888 * rate is likely to be an average over the disk
2889 * surface. Accordingly, to not be too harsh with
2890 * unlucky processes, a process is deemed slow only if
2891 * its rate has been lower than half of the estimated
2894 slow = bfqq->entity.service < bfqd->bfq_max_budget / 2;
2897 bfq_log_bfqq(bfqd, bfqq, "bfq_bfqq_is_slow: slow %d", slow);
2903 * To be deemed as soft real-time, an application must meet two
2904 * requirements. First, the application must not require an average
2905 * bandwidth higher than the approximate bandwidth required to playback or
2906 * record a compressed high-definition video.
2907 * The next function is invoked on the completion of the last request of a
2908 * batch, to compute the next-start time instant, soft_rt_next_start, such
2909 * that, if the next request of the application does not arrive before
2910 * soft_rt_next_start, then the above requirement on the bandwidth is met.
2912 * The second requirement is that the request pattern of the application is
2913 * isochronous, i.e., that, after issuing a request or a batch of requests,
2914 * the application stops issuing new requests until all its pending requests
2915 * have been completed. After that, the application may issue a new batch,
2917 * For this reason the next function is invoked to compute
2918 * soft_rt_next_start only for applications that meet this requirement,
2919 * whereas soft_rt_next_start is set to infinity for applications that do
2922 * Unfortunately, even a greedy application may happen to behave in an
2923 * isochronous way if the CPU load is high. In fact, the application may
2924 * stop issuing requests while the CPUs are busy serving other processes,
2925 * then restart, then stop again for a while, and so on. In addition, if
2926 * the disk achieves a low enough throughput with the request pattern
2927 * issued by the application (e.g., because the request pattern is random
2928 * and/or the device is slow), then the application may meet the above
2929 * bandwidth requirement too. To prevent such a greedy application to be
2930 * deemed as soft real-time, a further rule is used in the computation of
2931 * soft_rt_next_start: soft_rt_next_start must be higher than the current
2932 * time plus the maximum time for which the arrival of a request is waited
2933 * for when a sync queue becomes idle, namely bfqd->bfq_slice_idle.
2934 * This filters out greedy applications, as the latter issue instead their
2935 * next request as soon as possible after the last one has been completed
2936 * (in contrast, when a batch of requests is completed, a soft real-time
2937 * application spends some time processing data).
2939 * Unfortunately, the last filter may easily generate false positives if
2940 * only bfqd->bfq_slice_idle is used as a reference time interval and one
2941 * or both the following cases occur:
2942 * 1) HZ is so low that the duration of a jiffy is comparable to or higher
2943 * than bfqd->bfq_slice_idle. This happens, e.g., on slow devices with
2945 * 2) jiffies, instead of increasing at a constant rate, may stop increasing
2946 * for a while, then suddenly 'jump' by several units to recover the lost
2947 * increments. This seems to happen, e.g., inside virtual machines.
2948 * To address this issue, we do not use as a reference time interval just
2949 * bfqd->bfq_slice_idle, but bfqd->bfq_slice_idle plus a few jiffies. In
2950 * particular we add the minimum number of jiffies for which the filter
2951 * seems to be quite precise also in embedded systems and KVM/QEMU virtual
2954 static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data *bfqd,
2955 struct bfq_queue *bfqq)
2957 return max(bfqq->last_idle_bklogged +
2958 HZ * bfqq->service_from_backlogged /
2959 bfqd->bfq_wr_max_softrt_rate,
2960 jiffies + nsecs_to_jiffies(bfqq->bfqd->bfq_slice_idle) + 4);
2964 * bfq_bfqq_expire - expire a queue.
2965 * @bfqd: device owning the queue.
2966 * @bfqq: the queue to expire.
2967 * @compensate: if true, compensate for the time spent idling.
2968 * @reason: the reason causing the expiration.
2970 * If the process associated with bfqq does slow I/O (e.g., because it
2971 * issues random requests), we charge bfqq with the time it has been
2972 * in service instead of the service it has received (see
2973 * bfq_bfqq_charge_time for details on how this goal is achieved). As
2974 * a consequence, bfqq will typically get higher timestamps upon
2975 * reactivation, and hence it will be rescheduled as if it had
2976 * received more service than what it has actually received. In the
2977 * end, bfqq receives less service in proportion to how slowly its
2978 * associated process consumes its budgets (and hence how seriously it
2979 * tends to lower the throughput). In addition, this time-charging
2980 * strategy guarantees time fairness among slow processes. In
2981 * contrast, if the process associated with bfqq is not slow, we
2982 * charge bfqq exactly with the service it has received.
2984 * Charging time to the first type of queues and the exact service to
2985 * the other has the effect of using the WF2Q+ policy to schedule the
2986 * former on a timeslice basis, without violating service domain
2987 * guarantees among the latter.
2989 void bfq_bfqq_expire(struct bfq_data *bfqd,
2990 struct bfq_queue *bfqq,
2992 enum bfqq_expiration reason)
2995 unsigned long delta = 0;
2996 struct bfq_entity *entity = &bfqq->entity;
3000 * Check whether the process is slow (see bfq_bfqq_is_slow).
3002 slow = bfq_bfqq_is_slow(bfqd, bfqq, compensate, reason, &delta);
3005 * Increase service_from_backlogged before next statement,
3006 * because the possible next invocation of
3007 * bfq_bfqq_charge_time would likely inflate
3008 * entity->service. In contrast, service_from_backlogged must
3009 * contain real service, to enable the soft real-time
3010 * heuristic to correctly compute the bandwidth consumed by
3013 bfqq->service_from_backlogged += entity->service;
3016 * As above explained, charge slow (typically seeky) and
3017 * timed-out queues with the time and not the service
3018 * received, to favor sequential workloads.
3020 * Processes doing I/O in the slower disk zones will tend to
3021 * be slow(er) even if not seeky. Therefore, since the
3022 * estimated peak rate is actually an average over the disk
3023 * surface, these processes may timeout just for bad luck. To
3024 * avoid punishing them, do not charge time to processes that
3025 * succeeded in consuming at least 2/3 of their budget. This
3026 * allows BFQ to preserve enough elasticity to still perform
3027 * bandwidth, and not time, distribution with little unlucky
3028 * or quasi-sequential processes.
3030 if (bfqq->wr_coeff == 1 &&
3032 (reason == BFQQE_BUDGET_TIMEOUT &&
3033 bfq_bfqq_budget_left(bfqq) >= entity->budget / 3)))
3034 bfq_bfqq_charge_time(bfqd, bfqq, delta);
3036 if (reason == BFQQE_TOO_IDLE &&
3037 entity->service <= 2 * entity->budget / 10)
3038 bfq_clear_bfqq_IO_bound(bfqq);
3040 if (bfqd->low_latency && bfqq->wr_coeff == 1)
3041 bfqq->last_wr_start_finish = jiffies;
3043 if (bfqd->low_latency && bfqd->bfq_wr_max_softrt_rate > 0 &&
3044 RB_EMPTY_ROOT(&bfqq->sort_list)) {
3046 * If we get here, and there are no outstanding
3047 * requests, then the request pattern is isochronous
3048 * (see the comments on the function
3049 * bfq_bfqq_softrt_next_start()). Thus we can compute
3050 * soft_rt_next_start. If, instead, the queue still
3051 * has outstanding requests, then we have to wait for
3052 * the completion of all the outstanding requests to
3053 * discover whether the request pattern is actually
3056 if (bfqq->dispatched == 0)
3057 bfqq->soft_rt_next_start =
3058 bfq_bfqq_softrt_next_start(bfqd, bfqq);
3061 * The application is still waiting for the
3062 * completion of one or more requests:
3063 * prevent it from possibly being incorrectly
3064 * deemed as soft real-time by setting its
3065 * soft_rt_next_start to infinity. In fact,
3066 * without this assignment, the application
3067 * would be incorrectly deemed as soft
3069 * 1) it issued a new request before the
3070 * completion of all its in-flight
3072 * 2) at that time, its soft_rt_next_start
3073 * happened to be in the past.
3075 bfqq->soft_rt_next_start =
3076 bfq_greatest_from_now();
3078 * Schedule an update of soft_rt_next_start to when
3079 * the task may be discovered to be isochronous.
3081 bfq_mark_bfqq_softrt_update(bfqq);
3085 bfq_log_bfqq(bfqd, bfqq,
3086 "expire (%d, slow %d, num_disp %d, short_ttime %d)", reason,
3087 slow, bfqq->dispatched, bfq_bfqq_has_short_ttime(bfqq));
3090 * Increase, decrease or leave budget unchanged according to
3093 __bfq_bfqq_recalc_budget(bfqd, bfqq, reason);
3095 __bfq_bfqq_expire(bfqd, bfqq);
3097 /* mark bfqq as waiting a request only if a bic still points to it */
3098 if (ref > 1 && !bfq_bfqq_busy(bfqq) &&
3099 reason != BFQQE_BUDGET_TIMEOUT &&
3100 reason != BFQQE_BUDGET_EXHAUSTED)
3101 bfq_mark_bfqq_non_blocking_wait_rq(bfqq);
3105 * Budget timeout is not implemented through a dedicated timer, but
3106 * just checked on request arrivals and completions, as well as on
3107 * idle timer expirations.
3109 static bool bfq_bfqq_budget_timeout(struct bfq_queue *bfqq)
3111 return time_is_before_eq_jiffies(bfqq->budget_timeout);
3115 * If we expire a queue that is actively waiting (i.e., with the
3116 * device idled) for the arrival of a new request, then we may incur
3117 * the timestamp misalignment problem described in the body of the
3118 * function __bfq_activate_entity. Hence we return true only if this
3119 * condition does not hold, or if the queue is slow enough to deserve
3120 * only to be kicked off for preserving a high throughput.
3122 static bool bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq)
3124 bfq_log_bfqq(bfqq->bfqd, bfqq,
3125 "may_budget_timeout: wait_request %d left %d timeout %d",
3126 bfq_bfqq_wait_request(bfqq),
3127 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3,
3128 bfq_bfqq_budget_timeout(bfqq));
3130 return (!bfq_bfqq_wait_request(bfqq) ||
3131 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3)
3133 bfq_bfqq_budget_timeout(bfqq);
3137 * For a queue that becomes empty, device idling is allowed only if
3138 * this function returns true for the queue. As a consequence, since
3139 * device idling plays a critical role in both throughput boosting and
3140 * service guarantees, the return value of this function plays a
3141 * critical role in both these aspects as well.
3143 * In a nutshell, this function returns true only if idling is
3144 * beneficial for throughput or, even if detrimental for throughput,
3145 * idling is however necessary to preserve service guarantees (low
3146 * latency, desired throughput distribution, ...). In particular, on
3147 * NCQ-capable devices, this function tries to return false, so as to
3148 * help keep the drives' internal queues full, whenever this helps the
3149 * device boost the throughput without causing any service-guarantee
3152 * In more detail, the return value of this function is obtained by,
3153 * first, computing a number of boolean variables that take into
3154 * account throughput and service-guarantee issues, and, then,
3155 * combining these variables in a logical expression. Most of the
3156 * issues taken into account are not trivial. We discuss these issues
3157 * individually while introducing the variables.
3159 static bool bfq_bfqq_may_idle(struct bfq_queue *bfqq)
3161 struct bfq_data *bfqd = bfqq->bfqd;
3162 bool rot_without_queueing =
3163 !blk_queue_nonrot(bfqd->queue) && !bfqd->hw_tag,
3164 bfqq_sequential_and_IO_bound,
3165 idling_boosts_thr, idling_boosts_thr_without_issues,
3166 idling_needed_for_service_guarantees,
3167 asymmetric_scenario;
3169 if (bfqd->strict_guarantees)
3173 * Idling is performed only if slice_idle > 0. In addition, we
3176 * (b) bfqq is in the idle io prio class: in this case we do
3177 * not idle because we want to minimize the bandwidth that
3178 * queues in this class can steal to higher-priority queues
3180 if (bfqd->bfq_slice_idle == 0 || !bfq_bfqq_sync(bfqq) ||
3181 bfq_class_idle(bfqq))
3184 bfqq_sequential_and_IO_bound = !BFQQ_SEEKY(bfqq) &&
3185 bfq_bfqq_IO_bound(bfqq) && bfq_bfqq_has_short_ttime(bfqq);
3188 * The next variable takes into account the cases where idling
3189 * boosts the throughput.
3191 * The value of the variable is computed considering, first, that
3192 * idling is virtually always beneficial for the throughput if:
3193 * (a) the device is not NCQ-capable and rotational, or
3194 * (b) regardless of the presence of NCQ, the device is rotational and
3195 * the request pattern for bfqq is I/O-bound and sequential, or
3196 * (c) regardless of whether it is rotational, the device is
3197 * not NCQ-capable and the request pattern for bfqq is
3198 * I/O-bound and sequential.
3200 * Secondly, and in contrast to the above item (b), idling an
3201 * NCQ-capable flash-based device would not boost the
3202 * throughput even with sequential I/O; rather it would lower
3203 * the throughput in proportion to how fast the device
3204 * is. Accordingly, the next variable is true if any of the
3205 * above conditions (a), (b) or (c) is true, and, in
3206 * particular, happens to be false if bfqd is an NCQ-capable
3207 * flash-based device.
3209 idling_boosts_thr = rot_without_queueing ||
3210 ((!blk_queue_nonrot(bfqd->queue) || !bfqd->hw_tag) &&
3211 bfqq_sequential_and_IO_bound);
3214 * The value of the next variable,
3215 * idling_boosts_thr_without_issues, is equal to that of
3216 * idling_boosts_thr, unless a special case holds. In this
3217 * special case, described below, idling may cause problems to
3218 * weight-raised queues.
3220 * When the request pool is saturated (e.g., in the presence
3221 * of write hogs), if the processes associated with
3222 * non-weight-raised queues ask for requests at a lower rate,
3223 * then processes associated with weight-raised queues have a
3224 * higher probability to get a request from the pool
3225 * immediately (or at least soon) when they need one. Thus
3226 * they have a higher probability to actually get a fraction
3227 * of the device throughput proportional to their high
3228 * weight. This is especially true with NCQ-capable drives,
3229 * which enqueue several requests in advance, and further
3230 * reorder internally-queued requests.
3232 * For this reason, we force to false the value of
3233 * idling_boosts_thr_without_issues if there are weight-raised
3234 * busy queues. In this case, and if bfqq is not weight-raised,
3235 * this guarantees that the device is not idled for bfqq (if,
3236 * instead, bfqq is weight-raised, then idling will be
3237 * guaranteed by another variable, see below). Combined with
3238 * the timestamping rules of BFQ (see [1] for details), this
3239 * behavior causes bfqq, and hence any sync non-weight-raised
3240 * queue, to get a lower number of requests served, and thus
3241 * to ask for a lower number of requests from the request
3242 * pool, before the busy weight-raised queues get served
3243 * again. This often mitigates starvation problems in the
3244 * presence of heavy write workloads and NCQ, thereby
3245 * guaranteeing a higher application and system responsiveness
3246 * in these hostile scenarios.
3248 idling_boosts_thr_without_issues = idling_boosts_thr &&
3249 bfqd->wr_busy_queues == 0;
3252 * There is then a case where idling must be performed not
3253 * for throughput concerns, but to preserve service
3256 * To introduce this case, we can note that allowing the drive
3257 * to enqueue more than one request at a time, and hence
3258 * delegating de facto final scheduling decisions to the
3259 * drive's internal scheduler, entails loss of control on the
3260 * actual request service order. In particular, the critical
3261 * situation is when requests from different processes happen
3262 * to be present, at the same time, in the internal queue(s)
3263 * of the drive. In such a situation, the drive, by deciding
3264 * the service order of the internally-queued requests, does
3265 * determine also the actual throughput distribution among
3266 * these processes. But the drive typically has no notion or
3267 * concern about per-process throughput distribution, and
3268 * makes its decisions only on a per-request basis. Therefore,
3269 * the service distribution enforced by the drive's internal
3270 * scheduler is likely to coincide with the desired
3271 * device-throughput distribution only in a completely
3272 * symmetric scenario where:
3273 * (i) each of these processes must get the same throughput as
3275 * (ii) all these processes have the same I/O pattern
3276 (either sequential or random).
3277 * In fact, in such a scenario, the drive will tend to treat
3278 * the requests of each of these processes in about the same
3279 * way as the requests of the others, and thus to provide
3280 * each of these processes with about the same throughput
3281 * (which is exactly the desired throughput distribution). In
3282 * contrast, in any asymmetric scenario, device idling is
3283 * certainly needed to guarantee that bfqq receives its
3284 * assigned fraction of the device throughput (see [1] for
3287 * We address this issue by controlling, actually, only the
3288 * symmetry sub-condition (i), i.e., provided that
3289 * sub-condition (i) holds, idling is not performed,
3290 * regardless of whether sub-condition (ii) holds. In other
3291 * words, only if sub-condition (i) holds, then idling is
3292 * allowed, and the device tends to be prevented from queueing
3293 * many requests, possibly of several processes. The reason
3294 * for not controlling also sub-condition (ii) is that we
3295 * exploit preemption to preserve guarantees in case of
3296 * symmetric scenarios, even if (ii) does not hold, as
3297 * explained in the next two paragraphs.
3299 * Even if a queue, say Q, is expired when it remains idle, Q
3300 * can still preempt the new in-service queue if the next
3301 * request of Q arrives soon (see the comments on
3302 * bfq_bfqq_update_budg_for_activation). If all queues and
3303 * groups have the same weight, this form of preemption,
3304 * combined with the hole-recovery heuristic described in the
3305 * comments on function bfq_bfqq_update_budg_for_activation,
3306 * are enough to preserve a correct bandwidth distribution in
3307 * the mid term, even without idling. In fact, even if not
3308 * idling allows the internal queues of the device to contain
3309 * many requests, and thus to reorder requests, we can rather
3310 * safely assume that the internal scheduler still preserves a
3311 * minimum of mid-term fairness. The motivation for using
3312 * preemption instead of idling is that, by not idling,
3313 * service guarantees are preserved without minimally
3314 * sacrificing throughput. In other words, both a high
3315 * throughput and its desired distribution are obtained.
3317 * More precisely, this preemption-based, idleless approach
3318 * provides fairness in terms of IOPS, and not sectors per
3319 * second. This can be seen with a simple example. Suppose
3320 * that there are two queues with the same weight, but that
3321 * the first queue receives requests of 8 sectors, while the
3322 * second queue receives requests of 1024 sectors. In
3323 * addition, suppose that each of the two queues contains at
3324 * most one request at a time, which implies that each queue
3325 * always remains idle after it is served. Finally, after
3326 * remaining idle, each queue receives very quickly a new
3327 * request. It follows that the two queues are served
3328 * alternatively, preempting each other if needed. This
3329 * implies that, although both queues have the same weight,
3330 * the queue with large requests receives a service that is
3331 * 1024/8 times as high as the service received by the other
3334 * On the other hand, device idling is performed, and thus
3335 * pure sector-domain guarantees are provided, for the
3336 * following queues, which are likely to need stronger
3337 * throughput guarantees: weight-raised queues, and queues
3338 * with a higher weight than other queues. When such queues
3339 * are active, sub-condition (i) is false, which triggers
3342 * According to the above considerations, the next variable is
3343 * true (only) if sub-condition (i) holds. To compute the
3344 * value of this variable, we not only use the return value of
3345 * the function bfq_symmetric_scenario(), but also check
3346 * whether bfqq is being weight-raised, because
3347 * bfq_symmetric_scenario() does not take into account also
3348 * weight-raised queues (see comments on
3349 * bfq_weights_tree_add()).
3351 * As a side note, it is worth considering that the above
3352 * device-idling countermeasures may however fail in the
3353 * following unlucky scenario: if idling is (correctly)
3354 * disabled in a time period during which all symmetry
3355 * sub-conditions hold, and hence the device is allowed to
3356 * enqueue many requests, but at some later point in time some
3357 * sub-condition stops to hold, then it may become impossible
3358 * to let requests be served in the desired order until all
3359 * the requests already queued in the device have been served.
3361 asymmetric_scenario = bfqq->wr_coeff > 1 ||
3362 !bfq_symmetric_scenario(bfqd);
3365 * Finally, there is a case where maximizing throughput is the
3366 * best choice even if it may cause unfairness toward
3367 * bfqq. Such a case is when bfqq became active in a burst of
3368 * queue activations. Queues that became active during a large
3369 * burst benefit only from throughput, as discussed in the
3370 * comments on bfq_handle_burst. Thus, if bfqq became active
3371 * in a burst and not idling the device maximizes throughput,
3372 * then the device must no be idled, because not idling the
3373 * device provides bfqq and all other queues in the burst with
3374 * maximum benefit. Combining this and the above case, we can
3375 * now establish when idling is actually needed to preserve
3376 * service guarantees.
3378 idling_needed_for_service_guarantees =
3379 asymmetric_scenario && !bfq_bfqq_in_large_burst(bfqq);
3382 * We have now all the components we need to compute the
3383 * return value of the function, which is true only if idling
3384 * either boosts the throughput (without issues), or is
3385 * necessary to preserve service guarantees.
3387 return idling_boosts_thr_without_issues ||
3388 idling_needed_for_service_guarantees;
3392 * If the in-service queue is empty but the function bfq_bfqq_may_idle
3393 * returns true, then:
3394 * 1) the queue must remain in service and cannot be expired, and
3395 * 2) the device must be idled to wait for the possible arrival of a new
3396 * request for the queue.
3397 * See the comments on the function bfq_bfqq_may_idle for the reasons
3398 * why performing device idling is the best choice to boost the throughput
3399 * and preserve service guarantees when bfq_bfqq_may_idle itself
3402 static bool bfq_bfqq_must_idle(struct bfq_queue *bfqq)
3404 return RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_bfqq_may_idle(bfqq);
3408 * Select a queue for service. If we have a current queue in service,
3409 * check whether to continue servicing it, or retrieve and set a new one.
3411 static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd)
3413 struct bfq_queue *bfqq;
3414 struct request *next_rq;
3415 enum bfqq_expiration reason = BFQQE_BUDGET_TIMEOUT;
3417 bfqq = bfqd->in_service_queue;
3421 bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue");
3423 if (bfq_may_expire_for_budg_timeout(bfqq) &&
3424 !bfq_bfqq_wait_request(bfqq) &&
3425 !bfq_bfqq_must_idle(bfqq))
3430 * This loop is rarely executed more than once. Even when it
3431 * happens, it is much more convenient to re-execute this loop
3432 * than to return NULL and trigger a new dispatch to get a
3435 next_rq = bfqq->next_rq;
3437 * If bfqq has requests queued and it has enough budget left to
3438 * serve them, keep the queue, otherwise expire it.
3441 if (bfq_serv_to_charge(next_rq, bfqq) >
3442 bfq_bfqq_budget_left(bfqq)) {
3444 * Expire the queue for budget exhaustion,
3445 * which makes sure that the next budget is
3446 * enough to serve the next request, even if
3447 * it comes from the fifo expired path.
3449 reason = BFQQE_BUDGET_EXHAUSTED;
3453 * The idle timer may be pending because we may
3454 * not disable disk idling even when a new request
3457 if (bfq_bfqq_wait_request(bfqq)) {
3459 * If we get here: 1) at least a new request
3460 * has arrived but we have not disabled the
3461 * timer because the request was too small,
3462 * 2) then the block layer has unplugged
3463 * the device, causing the dispatch to be
3466 * Since the device is unplugged, now the
3467 * requests are probably large enough to
3468 * provide a reasonable throughput.
3469 * So we disable idling.
3471 bfq_clear_bfqq_wait_request(bfqq);
3472 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
3473 bfqg_stats_update_idle_time(bfqq_group(bfqq));
3480 * No requests pending. However, if the in-service queue is idling
3481 * for a new request, or has requests waiting for a completion and
3482 * may idle after their completion, then keep it anyway.
3484 if (bfq_bfqq_wait_request(bfqq) ||
3485 (bfqq->dispatched != 0 && bfq_bfqq_may_idle(bfqq))) {
3490 reason = BFQQE_NO_MORE_REQUESTS;
3492 bfq_bfqq_expire(bfqd, bfqq, false, reason);
3494 bfqq = bfq_set_in_service_queue(bfqd);
3496 bfq_log_bfqq(bfqd, bfqq, "select_queue: checking new queue");
3501 bfq_log_bfqq(bfqd, bfqq, "select_queue: returned this queue");
3503 bfq_log(bfqd, "select_queue: no queue returned");
3508 static void bfq_update_wr_data(struct bfq_data *bfqd, struct bfq_queue *bfqq)
3510 struct bfq_entity *entity = &bfqq->entity;
3512 if (bfqq->wr_coeff > 1) { /* queue is being weight-raised */
3513 bfq_log_bfqq(bfqd, bfqq,
3514 "raising period dur %u/%u msec, old coeff %u, w %d(%d)",
3515 jiffies_to_msecs(jiffies - bfqq->last_wr_start_finish),
3516 jiffies_to_msecs(bfqq->wr_cur_max_time),
3518 bfqq->entity.weight, bfqq->entity.orig_weight);
3520 if (entity->prio_changed)
3521 bfq_log_bfqq(bfqd, bfqq, "WARN: pending prio change");
3524 * If the queue was activated in a burst, or too much
3525 * time has elapsed from the beginning of this
3526 * weight-raising period, then end weight raising.
3528 if (bfq_bfqq_in_large_burst(bfqq))
3529 bfq_bfqq_end_wr(bfqq);
3530 else if (time_is_before_jiffies(bfqq->last_wr_start_finish +
3531 bfqq->wr_cur_max_time)) {
3532 if (bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time ||
3533 time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
3534 bfq_wr_duration(bfqd)))
3535 bfq_bfqq_end_wr(bfqq);
3537 switch_back_to_interactive_wr(bfqq, bfqd);
3538 bfqq->entity.prio_changed = 1;
3543 * To improve latency (for this or other queues), immediately
3544 * update weight both if it must be raised and if it must be
3545 * lowered. Since, entity may be on some active tree here, and
3546 * might have a pending change of its ioprio class, invoke
3547 * next function with the last parameter unset (see the
3548 * comments on the function).
3550 if ((entity->weight > entity->orig_weight) != (bfqq->wr_coeff > 1))
3551 __bfq_entity_update_weight_prio(bfq_entity_service_tree(entity),
3556 * Dispatch next request from bfqq.
3558 static struct request *bfq_dispatch_rq_from_bfqq(struct bfq_data *bfqd,
3559 struct bfq_queue *bfqq)
3561 struct request *rq = bfqq->next_rq;
3562 unsigned long service_to_charge;
3564 service_to_charge = bfq_serv_to_charge(rq, bfqq);
3566 bfq_bfqq_served(bfqq, service_to_charge);
3568 bfq_dispatch_remove(bfqd->queue, rq);
3571 * If weight raising has to terminate for bfqq, then next
3572 * function causes an immediate update of bfqq's weight,
3573 * without waiting for next activation. As a consequence, on
3574 * expiration, bfqq will be timestamped as if has never been
3575 * weight-raised during this service slot, even if it has
3576 * received part or even most of the service as a
3577 * weight-raised queue. This inflates bfqq's timestamps, which
3578 * is beneficial, as bfqq is then more willing to leave the
3579 * device immediately to possible other weight-raised queues.
3581 bfq_update_wr_data(bfqd, bfqq);
3584 * Expire bfqq, pretending that its budget expired, if bfqq
3585 * belongs to CLASS_IDLE and other queues are waiting for
3588 if (bfqd->busy_queues > 1 && bfq_class_idle(bfqq))
3594 bfq_bfqq_expire(bfqd, bfqq, false, BFQQE_BUDGET_EXHAUSTED);
3598 static bool bfq_has_work(struct blk_mq_hw_ctx *hctx)
3600 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
3603 * Avoiding lock: a race on bfqd->busy_queues should cause at
3604 * most a call to dispatch for nothing
3606 return !list_empty_careful(&bfqd->dispatch) ||
3607 bfqd->busy_queues > 0;
3610 static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
3612 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
3613 struct request *rq = NULL;
3614 struct bfq_queue *bfqq = NULL;
3616 if (!list_empty(&bfqd->dispatch)) {
3617 rq = list_first_entry(&bfqd->dispatch, struct request,
3619 list_del_init(&rq->queuelist);
3625 * Increment counters here, because this
3626 * dispatch does not follow the standard
3627 * dispatch flow (where counters are
3632 goto inc_in_driver_start_rq;
3636 * We exploit the put_rq_private hook to decrement
3637 * rq_in_driver, but put_rq_private will not be
3638 * invoked on this request. So, to avoid unbalance,
3639 * just start this request, without incrementing
3640 * rq_in_driver. As a negative consequence,
3641 * rq_in_driver is deceptively lower than it should be
3642 * while this request is in service. This may cause
3643 * bfq_schedule_dispatch to be invoked uselessly.
3645 * As for implementing an exact solution, the
3646 * put_request hook, if defined, is probably invoked
3647 * also on this request. So, by exploiting this hook,
3648 * we could 1) increment rq_in_driver here, and 2)
3649 * decrement it in put_request. Such a solution would
3650 * let the value of the counter be always accurate,
3651 * but it would entail using an extra interface
3652 * function. This cost seems higher than the benefit,
3653 * being the frequency of non-elevator-private
3654 * requests very low.
3659 bfq_log(bfqd, "dispatch requests: %d busy queues", bfqd->busy_queues);
3661 if (bfqd->busy_queues == 0)
3665 * Force device to serve one request at a time if
3666 * strict_guarantees is true. Forcing this service scheme is
3667 * currently the ONLY way to guarantee that the request
3668 * service order enforced by the scheduler is respected by a
3669 * queueing device. Otherwise the device is free even to make
3670 * some unlucky request wait for as long as the device
3673 * Of course, serving one request at at time may cause loss of
3676 if (bfqd->strict_guarantees && bfqd->rq_in_driver > 0)
3679 bfqq = bfq_select_queue(bfqd);
3683 rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq);
3686 inc_in_driver_start_rq:
3687 bfqd->rq_in_driver++;
3689 rq->rq_flags |= RQF_STARTED;
3695 static struct request *bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
3697 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
3700 spin_lock_irq(&bfqd->lock);
3702 rq = __bfq_dispatch_request(hctx);
3703 spin_unlock_irq(&bfqd->lock);
3709 * Task holds one reference to the queue, dropped when task exits. Each rq
3710 * in-flight on this queue also holds a reference, dropped when rq is freed.
3712 * Scheduler lock must be held here. Recall not to use bfqq after calling
3713 * this function on it.
3715 void bfq_put_queue(struct bfq_queue *bfqq)
3717 #ifdef CONFIG_BFQ_GROUP_IOSCHED
3718 struct bfq_group *bfqg = bfqq_group(bfqq);
3722 bfq_log_bfqq(bfqq->bfqd, bfqq, "put_queue: %p %d",
3729 if (!hlist_unhashed(&bfqq->burst_list_node)) {
3730 hlist_del_init(&bfqq->burst_list_node);
3732 * Decrement also burst size after the removal, if the
3733 * process associated with bfqq is exiting, and thus
3734 * does not contribute to the burst any longer. This
3735 * decrement helps filter out false positives of large
3736 * bursts, when some short-lived process (often due to
3737 * the execution of commands by some service) happens
3738 * to start and exit while a complex application is
3739 * starting, and thus spawning several processes that
3740 * do I/O (and that *must not* be treated as a large
3741 * burst, see comments on bfq_handle_burst).
3743 * In particular, the decrement is performed only if:
3744 * 1) bfqq is not a merged queue, because, if it is,
3745 * then this free of bfqq is not triggered by the exit
3746 * of the process bfqq is associated with, but exactly
3747 * by the fact that bfqq has just been merged.
3748 * 2) burst_size is greater than 0, to handle
3749 * unbalanced decrements. Unbalanced decrements may
3750 * happen in te following case: bfqq is inserted into
3751 * the current burst list--without incrementing
3752 * bust_size--because of a split, but the current
3753 * burst list is not the burst list bfqq belonged to
3754 * (see comments on the case of a split in
3757 if (bfqq->bic && bfqq->bfqd->burst_size > 0)
3758 bfqq->bfqd->burst_size--;
3761 kmem_cache_free(bfq_pool, bfqq);
3762 #ifdef CONFIG_BFQ_GROUP_IOSCHED
3763 bfqg_and_blkg_put(bfqg);
3767 static void bfq_put_cooperator(struct bfq_queue *bfqq)
3769 struct bfq_queue *__bfqq, *next;
3772 * If this queue was scheduled to merge with another queue, be
3773 * sure to drop the reference taken on that queue (and others in
3774 * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs.
3776 __bfqq = bfqq->new_bfqq;
3780 next = __bfqq->new_bfqq;
3781 bfq_put_queue(__bfqq);
3786 static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq)
3788 if (bfqq == bfqd->in_service_queue) {
3789 __bfq_bfqq_expire(bfqd, bfqq);
3790 bfq_schedule_dispatch(bfqd);
3793 bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, bfqq->ref);
3795 bfq_put_cooperator(bfqq);
3797 bfq_put_queue(bfqq); /* release process reference */
3800 static void bfq_exit_icq_bfqq(struct bfq_io_cq *bic, bool is_sync)
3802 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
3803 struct bfq_data *bfqd;
3806 bfqd = bfqq->bfqd; /* NULL if scheduler already exited */
3809 unsigned long flags;
3811 spin_lock_irqsave(&bfqd->lock, flags);
3812 bfq_exit_bfqq(bfqd, bfqq);
3813 bic_set_bfqq(bic, NULL, is_sync);
3814 spin_unlock_irqrestore(&bfqd->lock, flags);
3818 static void bfq_exit_icq(struct io_cq *icq)
3820 struct bfq_io_cq *bic = icq_to_bic(icq);
3822 bfq_exit_icq_bfqq(bic, true);
3823 bfq_exit_icq_bfqq(bic, false);
3827 * Update the entity prio values; note that the new values will not
3828 * be used until the next (re)activation.
3831 bfq_set_next_ioprio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic)
3833 struct task_struct *tsk = current;
3835 struct bfq_data *bfqd = bfqq->bfqd;
3840 ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
3841 switch (ioprio_class) {
3843 dev_err(bfqq->bfqd->queue->backing_dev_info->dev,
3844 "bfq: bad prio class %d\n", ioprio_class);
3846 case IOPRIO_CLASS_NONE:
3848 * No prio set, inherit CPU scheduling settings.
3850 bfqq->new_ioprio = task_nice_ioprio(tsk);
3851 bfqq->new_ioprio_class = task_nice_ioclass(tsk);
3853 case IOPRIO_CLASS_RT:
3854 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
3855 bfqq->new_ioprio_class = IOPRIO_CLASS_RT;
3857 case IOPRIO_CLASS_BE:
3858 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
3859 bfqq->new_ioprio_class = IOPRIO_CLASS_BE;
3861 case IOPRIO_CLASS_IDLE:
3862 bfqq->new_ioprio_class = IOPRIO_CLASS_IDLE;
3863 bfqq->new_ioprio = 7;
3867 if (bfqq->new_ioprio >= IOPRIO_BE_NR) {
3868 pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n",
3870 bfqq->new_ioprio = IOPRIO_BE_NR;
3873 bfqq->entity.new_weight = bfq_ioprio_to_weight(bfqq->new_ioprio);
3874 bfqq->entity.prio_changed = 1;
3877 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
3878 struct bio *bio, bool is_sync,
3879 struct bfq_io_cq *bic);
3881 static void bfq_check_ioprio_change(struct bfq_io_cq *bic, struct bio *bio)
3883 struct bfq_data *bfqd = bic_to_bfqd(bic);
3884 struct bfq_queue *bfqq;
3885 int ioprio = bic->icq.ioc->ioprio;
3888 * This condition may trigger on a newly created bic, be sure to
3889 * drop the lock before returning.
3891 if (unlikely(!bfqd) || likely(bic->ioprio == ioprio))
3894 bic->ioprio = ioprio;
3896 bfqq = bic_to_bfqq(bic, false);
3898 /* release process reference on this queue */
3899 bfq_put_queue(bfqq);
3900 bfqq = bfq_get_queue(bfqd, bio, BLK_RW_ASYNC, bic);
3901 bic_set_bfqq(bic, bfqq, false);
3904 bfqq = bic_to_bfqq(bic, true);
3906 bfq_set_next_ioprio_data(bfqq, bic);
3909 static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
3910 struct bfq_io_cq *bic, pid_t pid, int is_sync)
3912 RB_CLEAR_NODE(&bfqq->entity.rb_node);
3913 INIT_LIST_HEAD(&bfqq->fifo);
3914 INIT_HLIST_NODE(&bfqq->burst_list_node);
3920 bfq_set_next_ioprio_data(bfqq, bic);
3924 * No need to mark as has_short_ttime if in
3925 * idle_class, because no device idling is performed
3926 * for queues in idle class
3928 if (!bfq_class_idle(bfqq))
3929 /* tentatively mark as has_short_ttime */
3930 bfq_mark_bfqq_has_short_ttime(bfqq);
3931 bfq_mark_bfqq_sync(bfqq);
3932 bfq_mark_bfqq_just_created(bfqq);
3934 bfq_clear_bfqq_sync(bfqq);
3936 /* set end request to minus infinity from now */
3937 bfqq->ttime.last_end_request = ktime_get_ns() + 1;
3939 bfq_mark_bfqq_IO_bound(bfqq);
3943 /* Tentative initial value to trade off between thr and lat */
3944 bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3;
3945 bfqq->budget_timeout = bfq_smallest_from_now();
3948 bfqq->last_wr_start_finish = jiffies;
3949 bfqq->wr_start_at_switch_to_srt = bfq_smallest_from_now();
3950 bfqq->split_time = bfq_smallest_from_now();
3953 * Set to the value for which bfqq will not be deemed as
3954 * soft rt when it becomes backlogged.
3956 bfqq->soft_rt_next_start = bfq_greatest_from_now();
3958 /* first request is almost certainly seeky */
3959 bfqq->seek_history = 1;
3962 static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd,
3963 struct bfq_group *bfqg,
3964 int ioprio_class, int ioprio)
3966 switch (ioprio_class) {
3967 case IOPRIO_CLASS_RT:
3968 return &bfqg->async_bfqq[0][ioprio];
3969 case IOPRIO_CLASS_NONE:
3970 ioprio = IOPRIO_NORM;
3972 case IOPRIO_CLASS_BE:
3973 return &bfqg->async_bfqq[1][ioprio];
3974 case IOPRIO_CLASS_IDLE:
3975 return &bfqg->async_idle_bfqq;
3981 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
3982 struct bio *bio, bool is_sync,
3983 struct bfq_io_cq *bic)
3985 const int ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
3986 const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
3987 struct bfq_queue **async_bfqq = NULL;
3988 struct bfq_queue *bfqq;
3989 struct bfq_group *bfqg;
3993 bfqg = bfq_find_set_group(bfqd, bio_blkcg(bio));
3995 bfqq = &bfqd->oom_bfqq;
4000 async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class,
4007 bfqq = kmem_cache_alloc_node(bfq_pool,
4008 GFP_NOWAIT | __GFP_ZERO | __GFP_NOWARN,
4012 bfq_init_bfqq(bfqd, bfqq, bic, current->pid,
4014 bfq_init_entity(&bfqq->entity, bfqg);
4015 bfq_log_bfqq(bfqd, bfqq, "allocated");
4017 bfqq = &bfqd->oom_bfqq;
4018 bfq_log_bfqq(bfqd, bfqq, "using oom bfqq");
4023 * Pin the queue now that it's allocated, scheduler exit will
4028 * Extra group reference, w.r.t. sync
4029 * queue. This extra reference is removed
4030 * only if bfqq->bfqg disappears, to
4031 * guarantee that this queue is not freed
4032 * until its group goes away.
4034 bfq_log_bfqq(bfqd, bfqq, "get_queue, bfqq not in async: %p, %d",
4040 bfqq->ref++; /* get a process reference to this queue */
4041 bfq_log_bfqq(bfqd, bfqq, "get_queue, at end: %p, %d", bfqq, bfqq->ref);
4046 static void bfq_update_io_thinktime(struct bfq_data *bfqd,
4047 struct bfq_queue *bfqq)
4049 struct bfq_ttime *ttime = &bfqq->ttime;
4050 u64 elapsed = ktime_get_ns() - bfqq->ttime.last_end_request;
4052 elapsed = min_t(u64, elapsed, 2ULL * bfqd->bfq_slice_idle);
4054 ttime->ttime_samples = (7*bfqq->ttime.ttime_samples + 256) / 8;
4055 ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed, 8);
4056 ttime->ttime_mean = div64_ul(ttime->ttime_total + 128,
4057 ttime->ttime_samples);
4061 bfq_update_io_seektime(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4064 bfqq->seek_history <<= 1;
4065 bfqq->seek_history |=
4066 get_sdist(bfqq->last_request_pos, rq) > BFQQ_SEEK_THR &&
4067 (!blk_queue_nonrot(bfqd->queue) ||
4068 blk_rq_sectors(rq) < BFQQ_SECT_THR_NONROT);
4071 static void bfq_update_has_short_ttime(struct bfq_data *bfqd,
4072 struct bfq_queue *bfqq,
4073 struct bfq_io_cq *bic)
4075 bool has_short_ttime = true;
4078 * No need to update has_short_ttime if bfqq is async or in
4079 * idle io prio class, or if bfq_slice_idle is zero, because
4080 * no device idling is performed for bfqq in this case.
4082 if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq) ||
4083 bfqd->bfq_slice_idle == 0)
4086 /* Idle window just restored, statistics are meaningless. */
4087 if (time_is_after_eq_jiffies(bfqq->split_time +
4088 bfqd->bfq_wr_min_idle_time))
4091 /* Think time is infinite if no process is linked to
4092 * bfqq. Otherwise check average think time to
4093 * decide whether to mark as has_short_ttime
4095 if (atomic_read(&bic->icq.ioc->active_ref) == 0 ||
4096 (bfq_sample_valid(bfqq->ttime.ttime_samples) &&
4097 bfqq->ttime.ttime_mean > bfqd->bfq_slice_idle))
4098 has_short_ttime = false;
4100 bfq_log_bfqq(bfqd, bfqq, "update_has_short_ttime: has_short_ttime %d",
4103 if (has_short_ttime)
4104 bfq_mark_bfqq_has_short_ttime(bfqq);
4106 bfq_clear_bfqq_has_short_ttime(bfqq);
4110 * Called when a new fs request (rq) is added to bfqq. Check if there's
4111 * something we should do about it.
4113 static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4116 struct bfq_io_cq *bic = RQ_BIC(rq);
4118 if (rq->cmd_flags & REQ_META)
4119 bfqq->meta_pending++;
4121 bfq_update_io_thinktime(bfqd, bfqq);
4122 bfq_update_has_short_ttime(bfqd, bfqq, bic);
4123 bfq_update_io_seektime(bfqd, bfqq, rq);
4125 bfq_log_bfqq(bfqd, bfqq,
4126 "rq_enqueued: has_short_ttime=%d (seeky %d)",
4127 bfq_bfqq_has_short_ttime(bfqq), BFQQ_SEEKY(bfqq));
4129 bfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
4131 if (bfqq == bfqd->in_service_queue && bfq_bfqq_wait_request(bfqq)) {
4132 bool small_req = bfqq->queued[rq_is_sync(rq)] == 1 &&
4133 blk_rq_sectors(rq) < 32;
4134 bool budget_timeout = bfq_bfqq_budget_timeout(bfqq);
4137 * There is just this request queued: if the request
4138 * is small and the queue is not to be expired, then
4141 * In this way, if the device is being idled to wait
4142 * for a new request from the in-service queue, we
4143 * avoid unplugging the device and committing the
4144 * device to serve just a small request. On the
4145 * contrary, we wait for the block layer to decide
4146 * when to unplug the device: hopefully, new requests
4147 * will be merged to this one quickly, then the device
4148 * will be unplugged and larger requests will be
4151 if (small_req && !budget_timeout)
4155 * A large enough request arrived, or the queue is to
4156 * be expired: in both cases disk idling is to be
4157 * stopped, so clear wait_request flag and reset
4160 bfq_clear_bfqq_wait_request(bfqq);
4161 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
4162 bfqg_stats_update_idle_time(bfqq_group(bfqq));
4165 * The queue is not empty, because a new request just
4166 * arrived. Hence we can safely expire the queue, in
4167 * case of budget timeout, without risking that the
4168 * timestamps of the queue are not updated correctly.
4169 * See [1] for more details.
4172 bfq_bfqq_expire(bfqd, bfqq, false,
4173 BFQQE_BUDGET_TIMEOUT);
4177 static void __bfq_insert_request(struct bfq_data *bfqd, struct request *rq)
4179 struct bfq_queue *bfqq = RQ_BFQQ(rq),
4180 *new_bfqq = bfq_setup_cooperator(bfqd, bfqq, rq, true);
4183 if (bic_to_bfqq(RQ_BIC(rq), 1) != bfqq)
4184 new_bfqq = bic_to_bfqq(RQ_BIC(rq), 1);
4186 * Release the request's reference to the old bfqq
4187 * and make sure one is taken to the shared queue.
4189 new_bfqq->allocated++;
4193 * If the bic associated with the process
4194 * issuing this request still points to bfqq
4195 * (and thus has not been already redirected
4196 * to new_bfqq or even some other bfq_queue),
4197 * then complete the merge and redirect it to
4200 if (bic_to_bfqq(RQ_BIC(rq), 1) == bfqq)
4201 bfq_merge_bfqqs(bfqd, RQ_BIC(rq),
4204 bfq_clear_bfqq_just_created(bfqq);
4206 * rq is about to be enqueued into new_bfqq,
4207 * release rq reference on bfqq
4209 bfq_put_queue(bfqq);
4210 rq->elv.priv[1] = new_bfqq;
4214 bfq_add_request(rq);
4216 rq->fifo_time = ktime_get_ns() + bfqd->bfq_fifo_expire[rq_is_sync(rq)];
4217 list_add_tail(&rq->queuelist, &bfqq->fifo);
4219 bfq_rq_enqueued(bfqd, bfqq, rq);
4222 static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
4225 struct request_queue *q = hctx->queue;
4226 struct bfq_data *bfqd = q->elevator->elevator_data;
4228 spin_lock_irq(&bfqd->lock);
4229 if (blk_mq_sched_try_insert_merge(q, rq)) {
4230 spin_unlock_irq(&bfqd->lock);
4234 spin_unlock_irq(&bfqd->lock);
4236 blk_mq_sched_request_inserted(rq);
4238 spin_lock_irq(&bfqd->lock);
4239 if (at_head || blk_rq_is_passthrough(rq)) {
4241 list_add(&rq->queuelist, &bfqd->dispatch);
4243 list_add_tail(&rq->queuelist, &bfqd->dispatch);
4245 __bfq_insert_request(bfqd, rq);
4247 if (rq_mergeable(rq)) {
4248 elv_rqhash_add(q, rq);
4254 spin_unlock_irq(&bfqd->lock);
4257 static void bfq_insert_requests(struct blk_mq_hw_ctx *hctx,
4258 struct list_head *list, bool at_head)
4260 while (!list_empty(list)) {
4263 rq = list_first_entry(list, struct request, queuelist);
4264 list_del_init(&rq->queuelist);
4265 bfq_insert_request(hctx, rq, at_head);
4269 static void bfq_update_hw_tag(struct bfq_data *bfqd)
4271 bfqd->max_rq_in_driver = max_t(int, bfqd->max_rq_in_driver,
4272 bfqd->rq_in_driver);
4274 if (bfqd->hw_tag == 1)
4278 * This sample is valid if the number of outstanding requests
4279 * is large enough to allow a queueing behavior. Note that the
4280 * sum is not exact, as it's not taking into account deactivated
4283 if (bfqd->rq_in_driver + bfqd->queued < BFQ_HW_QUEUE_THRESHOLD)
4286 if (bfqd->hw_tag_samples++ < BFQ_HW_QUEUE_SAMPLES)
4289 bfqd->hw_tag = bfqd->max_rq_in_driver > BFQ_HW_QUEUE_THRESHOLD;
4290 bfqd->max_rq_in_driver = 0;
4291 bfqd->hw_tag_samples = 0;
4294 static void bfq_completed_request(struct bfq_queue *bfqq, struct bfq_data *bfqd)
4299 bfq_update_hw_tag(bfqd);
4301 bfqd->rq_in_driver--;
4304 if (!bfqq->dispatched && !bfq_bfqq_busy(bfqq)) {
4306 * Set budget_timeout (which we overload to store the
4307 * time at which the queue remains with no backlog and
4308 * no outstanding request; used by the weight-raising
4311 bfqq->budget_timeout = jiffies;
4313 bfq_weights_tree_remove(bfqd, &bfqq->entity,
4314 &bfqd->queue_weights_tree);
4317 now_ns = ktime_get_ns();
4319 bfqq->ttime.last_end_request = now_ns;
4322 * Using us instead of ns, to get a reasonable precision in
4323 * computing rate in next check.
4325 delta_us = div_u64(now_ns - bfqd->last_completion, NSEC_PER_USEC);
4328 * If the request took rather long to complete, and, according
4329 * to the maximum request size recorded, this completion latency
4330 * implies that the request was certainly served at a very low
4331 * rate (less than 1M sectors/sec), then the whole observation
4332 * interval that lasts up to this time instant cannot be a
4333 * valid time interval for computing a new peak rate. Invoke
4334 * bfq_update_rate_reset to have the following three steps
4336 * - close the observation interval at the last (previous)
4337 * request dispatch or completion
4338 * - compute rate, if possible, for that observation interval
4339 * - reset to zero samples, which will trigger a proper
4340 * re-initialization of the observation interval on next
4343 if (delta_us > BFQ_MIN_TT/NSEC_PER_USEC &&
4344 (bfqd->last_rq_max_size<<BFQ_RATE_SHIFT)/delta_us <
4345 1UL<<(BFQ_RATE_SHIFT - 10))
4346 bfq_update_rate_reset(bfqd, NULL);
4347 bfqd->last_completion = now_ns;
4350 * If we are waiting to discover whether the request pattern
4351 * of the task associated with the queue is actually
4352 * isochronous, and both requisites for this condition to hold
4353 * are now satisfied, then compute soft_rt_next_start (see the
4354 * comments on the function bfq_bfqq_softrt_next_start()). We
4355 * schedule this delayed check when bfqq expires, if it still
4356 * has in-flight requests.
4358 if (bfq_bfqq_softrt_update(bfqq) && bfqq->dispatched == 0 &&
4359 RB_EMPTY_ROOT(&bfqq->sort_list))
4360 bfqq->soft_rt_next_start =
4361 bfq_bfqq_softrt_next_start(bfqd, bfqq);
4364 * If this is the in-service queue, check if it needs to be expired,
4365 * or if we want to idle in case it has no pending requests.
4367 if (bfqd->in_service_queue == bfqq) {
4368 if (bfqq->dispatched == 0 && bfq_bfqq_must_idle(bfqq)) {
4369 bfq_arm_slice_timer(bfqd);
4371 } else if (bfq_may_expire_for_budg_timeout(bfqq))
4372 bfq_bfqq_expire(bfqd, bfqq, false,
4373 BFQQE_BUDGET_TIMEOUT);
4374 else if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
4375 (bfqq->dispatched == 0 ||
4376 !bfq_bfqq_may_idle(bfqq)))
4377 bfq_bfqq_expire(bfqd, bfqq, false,
4378 BFQQE_NO_MORE_REQUESTS);
4381 if (!bfqd->rq_in_driver)
4382 bfq_schedule_dispatch(bfqd);
4385 static void bfq_put_rq_priv_body(struct bfq_queue *bfqq)
4389 bfq_put_queue(bfqq);
4392 static void bfq_finish_request(struct request *rq)
4394 struct bfq_queue *bfqq;
4395 struct bfq_data *bfqd;
4403 if (rq->rq_flags & RQF_STARTED)
4404 bfqg_stats_update_completion(bfqq_group(bfqq),
4405 rq_start_time_ns(rq),
4406 rq_io_start_time_ns(rq),
4409 if (likely(rq->rq_flags & RQF_STARTED)) {
4410 unsigned long flags;
4412 spin_lock_irqsave(&bfqd->lock, flags);
4414 bfq_completed_request(bfqq, bfqd);
4415 bfq_put_rq_priv_body(bfqq);
4417 spin_unlock_irqrestore(&bfqd->lock, flags);
4420 * Request rq may be still/already in the scheduler,
4421 * in which case we need to remove it. And we cannot
4422 * defer such a check and removal, to avoid
4423 * inconsistencies in the time interval from the end
4424 * of this function to the start of the deferred work.
4425 * This situation seems to occur only in process
4426 * context, as a consequence of a merge. In the
4427 * current version of the code, this implies that the
4431 if (!RB_EMPTY_NODE(&rq->rb_node))
4432 bfq_remove_request(rq->q, rq);
4433 bfq_put_rq_priv_body(bfqq);
4436 rq->elv.priv[0] = NULL;
4437 rq->elv.priv[1] = NULL;
4441 * Returns NULL if a new bfqq should be allocated, or the old bfqq if this
4442 * was the last process referring to that bfqq.
4444 static struct bfq_queue *
4445 bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq)
4447 bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue");
4449 if (bfqq_process_refs(bfqq) == 1) {
4450 bfqq->pid = current->pid;
4451 bfq_clear_bfqq_coop(bfqq);
4452 bfq_clear_bfqq_split_coop(bfqq);
4456 bic_set_bfqq(bic, NULL, 1);
4458 bfq_put_cooperator(bfqq);
4460 bfq_put_queue(bfqq);
4464 static struct bfq_queue *bfq_get_bfqq_handle_split(struct bfq_data *bfqd,
4465 struct bfq_io_cq *bic,
4467 bool split, bool is_sync,
4470 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
4472 if (likely(bfqq && bfqq != &bfqd->oom_bfqq))
4479 bfq_put_queue(bfqq);
4480 bfqq = bfq_get_queue(bfqd, bio, is_sync, bic);
4482 bic_set_bfqq(bic, bfqq, is_sync);
4483 if (split && is_sync) {
4484 if ((bic->was_in_burst_list && bfqd->large_burst) ||
4485 bic->saved_in_large_burst)
4486 bfq_mark_bfqq_in_large_burst(bfqq);
4488 bfq_clear_bfqq_in_large_burst(bfqq);
4489 if (bic->was_in_burst_list)
4491 * If bfqq was in the current
4492 * burst list before being
4493 * merged, then we have to add
4494 * it back. And we do not need
4495 * to increase burst_size, as
4496 * we did not decrement
4497 * burst_size when we removed
4498 * bfqq from the burst list as
4499 * a consequence of a merge
4501 * bfq_put_queue). In this
4502 * respect, it would be rather
4503 * costly to know whether the
4504 * current burst list is still
4505 * the same burst list from
4506 * which bfqq was removed on
4507 * the merge. To avoid this
4508 * cost, if bfqq was in a
4509 * burst list, then we add
4510 * bfqq to the current burst
4511 * list without any further
4512 * check. This can cause
4513 * inappropriate insertions,
4514 * but rarely enough to not
4515 * harm the detection of large
4516 * bursts significantly.
4518 hlist_add_head(&bfqq->burst_list_node,
4521 bfqq->split_time = jiffies;
4528 * Allocate bfq data structures associated with this request.
4530 static void bfq_prepare_request(struct request *rq, struct bio *bio)
4532 struct request_queue *q = rq->q;
4533 struct bfq_data *bfqd = q->elevator->elevator_data;
4534 struct bfq_io_cq *bic;
4535 const int is_sync = rq_is_sync(rq);
4536 struct bfq_queue *bfqq;
4537 bool new_queue = false;
4538 bool bfqq_already_existing = false, split = false;
4542 bic = icq_to_bic(rq->elv.icq);
4544 spin_lock_irq(&bfqd->lock);
4546 bfq_check_ioprio_change(bic, bio);
4548 bfq_bic_update_cgroup(bic, bio);
4550 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio, false, is_sync,
4553 if (likely(!new_queue)) {
4554 /* If the queue was seeky for too long, break it apart. */
4555 if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq)) {
4556 bfq_log_bfqq(bfqd, bfqq, "breaking apart bfqq");
4558 /* Update bic before losing reference to bfqq */
4559 if (bfq_bfqq_in_large_burst(bfqq))
4560 bic->saved_in_large_burst = true;
4562 bfqq = bfq_split_bfqq(bic, bfqq);
4566 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio,
4570 bfqq_already_existing = true;
4576 bfq_log_bfqq(bfqd, bfqq, "get_request %p: bfqq %p, %d",
4577 rq, bfqq, bfqq->ref);
4579 rq->elv.priv[0] = bic;
4580 rq->elv.priv[1] = bfqq;
4583 * If a bfq_queue has only one process reference, it is owned
4584 * by only this bic: we can then set bfqq->bic = bic. in
4585 * addition, if the queue has also just been split, we have to
4588 if (likely(bfqq != &bfqd->oom_bfqq) && bfqq_process_refs(bfqq) == 1) {
4592 * The queue has just been split from a shared
4593 * queue: restore the idle window and the
4594 * possible weight raising period.
4596 bfq_bfqq_resume_state(bfqq, bfqd, bic,
4597 bfqq_already_existing);
4601 if (unlikely(bfq_bfqq_just_created(bfqq)))
4602 bfq_handle_burst(bfqd, bfqq);
4604 spin_unlock_irq(&bfqd->lock);
4607 static void bfq_idle_slice_timer_body(struct bfq_queue *bfqq)
4609 struct bfq_data *bfqd = bfqq->bfqd;
4610 enum bfqq_expiration reason;
4611 unsigned long flags;
4613 spin_lock_irqsave(&bfqd->lock, flags);
4614 bfq_clear_bfqq_wait_request(bfqq);
4616 if (bfqq != bfqd->in_service_queue) {
4617 spin_unlock_irqrestore(&bfqd->lock, flags);
4621 if (bfq_bfqq_budget_timeout(bfqq))
4623 * Also here the queue can be safely expired
4624 * for budget timeout without wasting
4627 reason = BFQQE_BUDGET_TIMEOUT;
4628 else if (bfqq->queued[0] == 0 && bfqq->queued[1] == 0)
4630 * The queue may not be empty upon timer expiration,
4631 * because we may not disable the timer when the
4632 * first request of the in-service queue arrives
4633 * during disk idling.
4635 reason = BFQQE_TOO_IDLE;
4637 goto schedule_dispatch;
4639 bfq_bfqq_expire(bfqd, bfqq, true, reason);
4642 spin_unlock_irqrestore(&bfqd->lock, flags);
4643 bfq_schedule_dispatch(bfqd);
4647 * Handler of the expiration of the timer running if the in-service queue
4648 * is idling inside its time slice.
4650 static enum hrtimer_restart bfq_idle_slice_timer(struct hrtimer *timer)
4652 struct bfq_data *bfqd = container_of(timer, struct bfq_data,
4654 struct bfq_queue *bfqq = bfqd->in_service_queue;
4657 * Theoretical race here: the in-service queue can be NULL or
4658 * different from the queue that was idling if a new request
4659 * arrives for the current queue and there is a full dispatch
4660 * cycle that changes the in-service queue. This can hardly
4661 * happen, but in the worst case we just expire a queue too
4665 bfq_idle_slice_timer_body(bfqq);
4667 return HRTIMER_NORESTART;
4670 static void __bfq_put_async_bfqq(struct bfq_data *bfqd,
4671 struct bfq_queue **bfqq_ptr)
4673 struct bfq_queue *bfqq = *bfqq_ptr;
4675 bfq_log(bfqd, "put_async_bfqq: %p", bfqq);
4677 bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
4679 bfq_log_bfqq(bfqd, bfqq, "put_async_bfqq: putting %p, %d",
4681 bfq_put_queue(bfqq);
4687 * Release all the bfqg references to its async queues. If we are
4688 * deallocating the group these queues may still contain requests, so
4689 * we reparent them to the root cgroup (i.e., the only one that will
4690 * exist for sure until all the requests on a device are gone).
4692 void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg)
4696 for (i = 0; i < 2; i++)
4697 for (j = 0; j < IOPRIO_BE_NR; j++)
4698 __bfq_put_async_bfqq(bfqd, &bfqg->async_bfqq[i][j]);
4700 __bfq_put_async_bfqq(bfqd, &bfqg->async_idle_bfqq);
4703 static void bfq_exit_queue(struct elevator_queue *e)
4705 struct bfq_data *bfqd = e->elevator_data;
4706 struct bfq_queue *bfqq, *n;
4708 hrtimer_cancel(&bfqd->idle_slice_timer);
4710 spin_lock_irq(&bfqd->lock);
4711 list_for_each_entry_safe(bfqq, n, &bfqd->idle_list, bfqq_list)
4712 bfq_deactivate_bfqq(bfqd, bfqq, false, false);
4713 spin_unlock_irq(&bfqd->lock);
4715 hrtimer_cancel(&bfqd->idle_slice_timer);
4717 #ifdef CONFIG_BFQ_GROUP_IOSCHED
4718 blkcg_deactivate_policy(bfqd->queue, &blkcg_policy_bfq);
4720 spin_lock_irq(&bfqd->lock);
4721 bfq_put_async_queues(bfqd, bfqd->root_group);
4722 kfree(bfqd->root_group);
4723 spin_unlock_irq(&bfqd->lock);
4729 static void bfq_init_root_group(struct bfq_group *root_group,
4730 struct bfq_data *bfqd)
4734 #ifdef CONFIG_BFQ_GROUP_IOSCHED
4735 root_group->entity.parent = NULL;
4736 root_group->my_entity = NULL;
4737 root_group->bfqd = bfqd;
4739 root_group->rq_pos_tree = RB_ROOT;
4740 for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
4741 root_group->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
4742 root_group->sched_data.bfq_class_idle_last_service = jiffies;
4745 static int bfq_init_queue(struct request_queue *q, struct elevator_type *e)
4747 struct bfq_data *bfqd;
4748 struct elevator_queue *eq;
4750 eq = elevator_alloc(q, e);
4754 bfqd = kzalloc_node(sizeof(*bfqd), GFP_KERNEL, q->node);
4756 kobject_put(&eq->kobj);
4759 eq->elevator_data = bfqd;
4761 spin_lock_irq(q->queue_lock);
4763 spin_unlock_irq(q->queue_lock);
4766 * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues.
4767 * Grab a permanent reference to it, so that the normal code flow
4768 * will not attempt to free it.
4770 bfq_init_bfqq(bfqd, &bfqd->oom_bfqq, NULL, 1, 0);
4771 bfqd->oom_bfqq.ref++;
4772 bfqd->oom_bfqq.new_ioprio = BFQ_DEFAULT_QUEUE_IOPRIO;
4773 bfqd->oom_bfqq.new_ioprio_class = IOPRIO_CLASS_BE;
4774 bfqd->oom_bfqq.entity.new_weight =
4775 bfq_ioprio_to_weight(bfqd->oom_bfqq.new_ioprio);
4777 /* oom_bfqq does not participate to bursts */
4778 bfq_clear_bfqq_just_created(&bfqd->oom_bfqq);
4781 * Trigger weight initialization, according to ioprio, at the
4782 * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio
4783 * class won't be changed any more.
4785 bfqd->oom_bfqq.entity.prio_changed = 1;
4789 INIT_LIST_HEAD(&bfqd->dispatch);
4791 hrtimer_init(&bfqd->idle_slice_timer, CLOCK_MONOTONIC,
4793 bfqd->idle_slice_timer.function = bfq_idle_slice_timer;
4795 bfqd->queue_weights_tree = RB_ROOT;
4796 bfqd->group_weights_tree = RB_ROOT;
4798 INIT_LIST_HEAD(&bfqd->active_list);
4799 INIT_LIST_HEAD(&bfqd->idle_list);
4800 INIT_HLIST_HEAD(&bfqd->burst_list);
4804 bfqd->bfq_max_budget = bfq_default_max_budget;
4806 bfqd->bfq_fifo_expire[0] = bfq_fifo_expire[0];
4807 bfqd->bfq_fifo_expire[1] = bfq_fifo_expire[1];
4808 bfqd->bfq_back_max = bfq_back_max;
4809 bfqd->bfq_back_penalty = bfq_back_penalty;
4810 bfqd->bfq_slice_idle = bfq_slice_idle;
4811 bfqd->bfq_timeout = bfq_timeout;
4813 bfqd->bfq_requests_within_timer = 120;
4815 bfqd->bfq_large_burst_thresh = 8;
4816 bfqd->bfq_burst_interval = msecs_to_jiffies(180);
4818 bfqd->low_latency = true;
4821 * Trade-off between responsiveness and fairness.
4823 bfqd->bfq_wr_coeff = 30;
4824 bfqd->bfq_wr_rt_max_time = msecs_to_jiffies(300);
4825 bfqd->bfq_wr_max_time = 0;
4826 bfqd->bfq_wr_min_idle_time = msecs_to_jiffies(2000);
4827 bfqd->bfq_wr_min_inter_arr_async = msecs_to_jiffies(500);
4828 bfqd->bfq_wr_max_softrt_rate = 7000; /*
4829 * Approximate rate required
4830 * to playback or record a
4831 * high-definition compressed
4834 bfqd->wr_busy_queues = 0;
4837 * Begin by assuming, optimistically, that the device is a
4838 * high-speed one, and that its peak rate is equal to 2/3 of
4839 * the highest reference rate.
4841 bfqd->RT_prod = R_fast[blk_queue_nonrot(bfqd->queue)] *
4842 T_fast[blk_queue_nonrot(bfqd->queue)];
4843 bfqd->peak_rate = R_fast[blk_queue_nonrot(bfqd->queue)] * 2 / 3;
4844 bfqd->device_speed = BFQ_BFQD_FAST;
4846 spin_lock_init(&bfqd->lock);
4849 * The invocation of the next bfq_create_group_hierarchy
4850 * function is the head of a chain of function calls
4851 * (bfq_create_group_hierarchy->blkcg_activate_policy->
4852 * blk_mq_freeze_queue) that may lead to the invocation of the
4853 * has_work hook function. For this reason,
4854 * bfq_create_group_hierarchy is invoked only after all
4855 * scheduler data has been initialized, apart from the fields
4856 * that can be initialized only after invoking
4857 * bfq_create_group_hierarchy. This, in particular, enables
4858 * has_work to correctly return false. Of course, to avoid
4859 * other inconsistencies, the blk-mq stack must then refrain
4860 * from invoking further scheduler hooks before this init
4861 * function is finished.
4863 bfqd->root_group = bfq_create_group_hierarchy(bfqd, q->node);
4864 if (!bfqd->root_group)
4866 bfq_init_root_group(bfqd->root_group, bfqd);
4867 bfq_init_entity(&bfqd->oom_bfqq.entity, bfqd->root_group);
4869 wbt_disable_default(q);
4874 kobject_put(&eq->kobj);
4878 static void bfq_slab_kill(void)
4880 kmem_cache_destroy(bfq_pool);
4883 static int __init bfq_slab_setup(void)
4885 bfq_pool = KMEM_CACHE(bfq_queue, 0);
4891 static ssize_t bfq_var_show(unsigned int var, char *page)
4893 return sprintf(page, "%u\n", var);
4896 static int bfq_var_store(unsigned long *var, const char *page)
4898 unsigned long new_val;
4899 int ret = kstrtoul(page, 10, &new_val);
4907 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
4908 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
4910 struct bfq_data *bfqd = e->elevator_data; \
4911 u64 __data = __VAR; \
4913 __data = jiffies_to_msecs(__data); \
4914 else if (__CONV == 2) \
4915 __data = div_u64(__data, NSEC_PER_MSEC); \
4916 return bfq_var_show(__data, (page)); \
4918 SHOW_FUNCTION(bfq_fifo_expire_sync_show, bfqd->bfq_fifo_expire[1], 2);
4919 SHOW_FUNCTION(bfq_fifo_expire_async_show, bfqd->bfq_fifo_expire[0], 2);
4920 SHOW_FUNCTION(bfq_back_seek_max_show, bfqd->bfq_back_max, 0);
4921 SHOW_FUNCTION(bfq_back_seek_penalty_show, bfqd->bfq_back_penalty, 0);
4922 SHOW_FUNCTION(bfq_slice_idle_show, bfqd->bfq_slice_idle, 2);
4923 SHOW_FUNCTION(bfq_max_budget_show, bfqd->bfq_user_max_budget, 0);
4924 SHOW_FUNCTION(bfq_timeout_sync_show, bfqd->bfq_timeout, 1);
4925 SHOW_FUNCTION(bfq_strict_guarantees_show, bfqd->strict_guarantees, 0);
4926 SHOW_FUNCTION(bfq_low_latency_show, bfqd->low_latency, 0);
4927 #undef SHOW_FUNCTION
4929 #define USEC_SHOW_FUNCTION(__FUNC, __VAR) \
4930 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
4932 struct bfq_data *bfqd = e->elevator_data; \
4933 u64 __data = __VAR; \
4934 __data = div_u64(__data, NSEC_PER_USEC); \
4935 return bfq_var_show(__data, (page)); \
4937 USEC_SHOW_FUNCTION(bfq_slice_idle_us_show, bfqd->bfq_slice_idle);
4938 #undef USEC_SHOW_FUNCTION
4940 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
4942 __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4944 struct bfq_data *bfqd = e->elevator_data; \
4945 unsigned long __data, __min = (MIN), __max = (MAX); \
4948 ret = bfq_var_store(&__data, (page)); \
4951 if (__data < __min) \
4953 else if (__data > __max) \
4956 *(__PTR) = msecs_to_jiffies(__data); \
4957 else if (__CONV == 2) \
4958 *(__PTR) = (u64)__data * NSEC_PER_MSEC; \
4960 *(__PTR) = __data; \
4963 STORE_FUNCTION(bfq_fifo_expire_sync_store, &bfqd->bfq_fifo_expire[1], 1,
4965 STORE_FUNCTION(bfq_fifo_expire_async_store, &bfqd->bfq_fifo_expire[0], 1,
4967 STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0);
4968 STORE_FUNCTION(bfq_back_seek_penalty_store, &bfqd->bfq_back_penalty, 1,
4970 STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 2);
4971 #undef STORE_FUNCTION
4973 #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
4974 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\
4976 struct bfq_data *bfqd = e->elevator_data; \
4977 unsigned long __data, __min = (MIN), __max = (MAX); \
4980 ret = bfq_var_store(&__data, (page)); \
4983 if (__data < __min) \
4985 else if (__data > __max) \
4987 *(__PTR) = (u64)__data * NSEC_PER_USEC; \
4990 USEC_STORE_FUNCTION(bfq_slice_idle_us_store, &bfqd->bfq_slice_idle, 0,
4992 #undef USEC_STORE_FUNCTION
4994 static ssize_t bfq_max_budget_store(struct elevator_queue *e,
4995 const char *page, size_t count)
4997 struct bfq_data *bfqd = e->elevator_data;
4998 unsigned long __data;
5001 ret = bfq_var_store(&__data, (page));
5006 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
5008 if (__data > INT_MAX)
5010 bfqd->bfq_max_budget = __data;
5013 bfqd->bfq_user_max_budget = __data;
5019 * Leaving this name to preserve name compatibility with cfq
5020 * parameters, but this timeout is used for both sync and async.
5022 static ssize_t bfq_timeout_sync_store(struct elevator_queue *e,
5023 const char *page, size_t count)
5025 struct bfq_data *bfqd = e->elevator_data;
5026 unsigned long __data;
5029 ret = bfq_var_store(&__data, (page));
5035 else if (__data > INT_MAX)
5038 bfqd->bfq_timeout = msecs_to_jiffies(__data);
5039 if (bfqd->bfq_user_max_budget == 0)
5040 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
5045 static ssize_t bfq_strict_guarantees_store(struct elevator_queue *e,
5046 const char *page, size_t count)
5048 struct bfq_data *bfqd = e->elevator_data;
5049 unsigned long __data;
5052 ret = bfq_var_store(&__data, (page));
5058 if (!bfqd->strict_guarantees && __data == 1
5059 && bfqd->bfq_slice_idle < 8 * NSEC_PER_MSEC)
5060 bfqd->bfq_slice_idle = 8 * NSEC_PER_MSEC;
5062 bfqd->strict_guarantees = __data;
5067 static ssize_t bfq_low_latency_store(struct elevator_queue *e,
5068 const char *page, size_t count)
5070 struct bfq_data *bfqd = e->elevator_data;
5071 unsigned long __data;
5074 ret = bfq_var_store(&__data, (page));
5080 if (__data == 0 && bfqd->low_latency != 0)
5082 bfqd->low_latency = __data;
5087 #define BFQ_ATTR(name) \
5088 __ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store)
5090 static struct elv_fs_entry bfq_attrs[] = {
5091 BFQ_ATTR(fifo_expire_sync),
5092 BFQ_ATTR(fifo_expire_async),
5093 BFQ_ATTR(back_seek_max),
5094 BFQ_ATTR(back_seek_penalty),
5095 BFQ_ATTR(slice_idle),
5096 BFQ_ATTR(slice_idle_us),
5097 BFQ_ATTR(max_budget),
5098 BFQ_ATTR(timeout_sync),
5099 BFQ_ATTR(strict_guarantees),
5100 BFQ_ATTR(low_latency),
5104 static struct elevator_type iosched_bfq_mq = {
5106 .prepare_request = bfq_prepare_request,
5107 .finish_request = bfq_finish_request,
5108 .exit_icq = bfq_exit_icq,
5109 .insert_requests = bfq_insert_requests,
5110 .dispatch_request = bfq_dispatch_request,
5111 .next_request = elv_rb_latter_request,
5112 .former_request = elv_rb_former_request,
5113 .allow_merge = bfq_allow_bio_merge,
5114 .bio_merge = bfq_bio_merge,
5115 .request_merge = bfq_request_merge,
5116 .requests_merged = bfq_requests_merged,
5117 .request_merged = bfq_request_merged,
5118 .has_work = bfq_has_work,
5119 .init_sched = bfq_init_queue,
5120 .exit_sched = bfq_exit_queue,
5124 .icq_size = sizeof(struct bfq_io_cq),
5125 .icq_align = __alignof__(struct bfq_io_cq),
5126 .elevator_attrs = bfq_attrs,
5127 .elevator_name = "bfq",
5128 .elevator_owner = THIS_MODULE,
5130 MODULE_ALIAS("bfq-iosched");
5132 static int __init bfq_init(void)
5136 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5137 ret = blkcg_policy_register(&blkcg_policy_bfq);
5143 if (bfq_slab_setup())
5147 * Times to load large popular applications for the typical
5148 * systems installed on the reference devices (see the
5149 * comments before the definitions of the next two
5150 * arrays). Actually, we use slightly slower values, as the
5151 * estimated peak rate tends to be smaller than the actual
5152 * peak rate. The reason for this last fact is that estimates
5153 * are computed over much shorter time intervals than the long
5154 * intervals typically used for benchmarking. Why? First, to
5155 * adapt more quickly to variations. Second, because an I/O
5156 * scheduler cannot rely on a peak-rate-evaluation workload to
5157 * be run for a long time.
5159 T_slow[0] = msecs_to_jiffies(3500); /* actually 4 sec */
5160 T_slow[1] = msecs_to_jiffies(6000); /* actually 6.5 sec */
5161 T_fast[0] = msecs_to_jiffies(7000); /* actually 8 sec */
5162 T_fast[1] = msecs_to_jiffies(2500); /* actually 3 sec */
5165 * Thresholds that determine the switch between speed classes
5166 * (see the comments before the definition of the array
5167 * device_speed_thresh). These thresholds are biased towards
5168 * transitions to the fast class. This is safer than the
5169 * opposite bias. In fact, a wrong transition to the slow
5170 * class results in short weight-raising periods, because the
5171 * speed of the device then tends to be higher that the
5172 * reference peak rate. On the opposite end, a wrong
5173 * transition to the fast class tends to increase
5174 * weight-raising periods, because of the opposite reason.
5176 device_speed_thresh[0] = (4 * R_slow[0]) / 3;
5177 device_speed_thresh[1] = (4 * R_slow[1]) / 3;
5179 ret = elv_register(&iosched_bfq_mq);
5188 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5189 blkcg_policy_unregister(&blkcg_policy_bfq);
5194 static void __exit bfq_exit(void)
5196 elv_unregister(&iosched_bfq_mq);
5197 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5198 blkcg_policy_unregister(&blkcg_policy_bfq);
5203 module_init(bfq_init);
5204 module_exit(bfq_exit);
5206 MODULE_AUTHOR("Paolo Valente");
5207 MODULE_LICENSE("GPL");
5208 MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler");