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"
112 #define BFQ_BFQQ_FNS(name) \
113 void bfq_mark_bfqq_##name(struct bfq_queue *bfqq) \
115 __set_bit(BFQQF_##name, &(bfqq)->flags); \
117 void bfq_clear_bfqq_##name(struct bfq_queue *bfqq) \
119 __clear_bit(BFQQF_##name, &(bfqq)->flags); \
121 int bfq_bfqq_##name(const struct bfq_queue *bfqq) \
123 return test_bit(BFQQF_##name, &(bfqq)->flags); \
126 BFQ_BFQQ_FNS(just_created);
128 BFQ_BFQQ_FNS(wait_request);
129 BFQ_BFQQ_FNS(non_blocking_wait_rq);
130 BFQ_BFQQ_FNS(fifo_expire);
131 BFQ_BFQQ_FNS(has_short_ttime);
133 BFQ_BFQQ_FNS(IO_bound);
134 BFQ_BFQQ_FNS(in_large_burst);
136 BFQ_BFQQ_FNS(split_coop);
137 BFQ_BFQQ_FNS(softrt_update);
138 #undef BFQ_BFQQ_FNS \
140 /* Expiration time of sync (0) and async (1) requests, in ns. */
141 static const u64 bfq_fifo_expire[2] = { NSEC_PER_SEC / 4, NSEC_PER_SEC / 8 };
143 /* Maximum backwards seek (magic number lifted from CFQ), in KiB. */
144 static const int bfq_back_max = 16 * 1024;
146 /* Penalty of a backwards seek, in number of sectors. */
147 static const int bfq_back_penalty = 2;
149 /* Idling period duration, in ns. */
150 static u64 bfq_slice_idle = NSEC_PER_SEC / 125;
152 /* Minimum number of assigned budgets for which stats are safe to compute. */
153 static const int bfq_stats_min_budgets = 194;
155 /* Default maximum budget values, in sectors and number of requests. */
156 static const int bfq_default_max_budget = 16 * 1024;
159 * Async to sync throughput distribution is controlled as follows:
160 * when an async request is served, the entity is charged the number
161 * of sectors of the request, multiplied by the factor below
163 static const int bfq_async_charge_factor = 10;
165 /* Default timeout values, in jiffies, approximating CFQ defaults. */
166 const int bfq_timeout = HZ / 8;
168 static struct kmem_cache *bfq_pool;
170 /* Below this threshold (in ns), we consider thinktime immediate. */
171 #define BFQ_MIN_TT (2 * NSEC_PER_MSEC)
173 /* hw_tag detection: parallel requests threshold and min samples needed. */
174 #define BFQ_HW_QUEUE_THRESHOLD 4
175 #define BFQ_HW_QUEUE_SAMPLES 32
177 #define BFQQ_SEEK_THR (sector_t)(8 * 100)
178 #define BFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
179 #define BFQQ_CLOSE_THR (sector_t)(8 * 1024)
180 #define BFQQ_SEEKY(bfqq) (hweight32(bfqq->seek_history) > 32/8)
182 /* Min number of samples required to perform peak-rate update */
183 #define BFQ_RATE_MIN_SAMPLES 32
184 /* Min observation time interval required to perform a peak-rate update (ns) */
185 #define BFQ_RATE_MIN_INTERVAL (300*NSEC_PER_MSEC)
186 /* Target observation time interval for a peak-rate update (ns) */
187 #define BFQ_RATE_REF_INTERVAL NSEC_PER_SEC
189 /* Shift used for peak rate fixed precision calculations. */
190 #define BFQ_RATE_SHIFT 16
193 * By default, BFQ computes the duration of the weight raising for
194 * interactive applications automatically, using the following formula:
195 * duration = (R / r) * T, where r is the peak rate of the device, and
196 * R and T are two reference parameters.
197 * In particular, R is the peak rate of the reference device (see below),
198 * and T is a reference time: given the systems that are likely to be
199 * installed on the reference device according to its speed class, T is
200 * about the maximum time needed, under BFQ and while reading two files in
201 * parallel, to load typical large applications on these systems.
202 * In practice, the slower/faster the device at hand is, the more/less it
203 * takes to load applications with respect to the reference device.
204 * Accordingly, the longer/shorter BFQ grants weight raising to interactive
207 * BFQ uses four different reference pairs (R, T), depending on:
208 * . whether the device is rotational or non-rotational;
209 * . whether the device is slow, such as old or portable HDDs, as well as
210 * SD cards, or fast, such as newer HDDs and SSDs.
212 * The device's speed class is dynamically (re)detected in
213 * bfq_update_peak_rate() every time the estimated peak rate is updated.
215 * In the following definitions, R_slow[0]/R_fast[0] and
216 * T_slow[0]/T_fast[0] are the reference values for a slow/fast
217 * rotational device, whereas R_slow[1]/R_fast[1] and
218 * T_slow[1]/T_fast[1] are the reference values for a slow/fast
219 * non-rotational device. Finally, device_speed_thresh are the
220 * thresholds used to switch between speed classes. The reference
221 * rates are not the actual peak rates of the devices used as a
222 * reference, but slightly lower values. The reason for using these
223 * slightly lower values is that the peak-rate estimator tends to
224 * yield slightly lower values than the actual peak rate (it can yield
225 * the actual peak rate only if there is only one process doing I/O,
226 * and the process does sequential I/O).
228 * Both the reference peak rates and the thresholds are measured in
229 * sectors/usec, left-shifted by BFQ_RATE_SHIFT.
231 static int R_slow[2] = {1000, 10700};
232 static int R_fast[2] = {14000, 33000};
234 * To improve readability, a conversion function is used to initialize the
235 * following arrays, which entails that they can be initialized only in a
238 static int T_slow[2];
239 static int T_fast[2];
240 static int device_speed_thresh[2];
242 #define RQ_BIC(rq) icq_to_bic((rq)->elv.priv[0])
243 #define RQ_BFQQ(rq) ((rq)->elv.priv[1])
245 struct bfq_queue *bic_to_bfqq(struct bfq_io_cq *bic, bool is_sync)
247 return bic->bfqq[is_sync];
250 void bic_set_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq, bool is_sync)
252 bic->bfqq[is_sync] = bfqq;
255 struct bfq_data *bic_to_bfqd(struct bfq_io_cq *bic)
257 return bic->icq.q->elevator->elevator_data;
261 * icq_to_bic - convert iocontext queue structure to bfq_io_cq.
262 * @icq: the iocontext queue.
264 static struct bfq_io_cq *icq_to_bic(struct io_cq *icq)
266 /* bic->icq is the first member, %NULL will convert to %NULL */
267 return container_of(icq, struct bfq_io_cq, icq);
271 * bfq_bic_lookup - search into @ioc a bic associated to @bfqd.
272 * @bfqd: the lookup key.
273 * @ioc: the io_context of the process doing I/O.
274 * @q: the request queue.
276 static struct bfq_io_cq *bfq_bic_lookup(struct bfq_data *bfqd,
277 struct io_context *ioc,
278 struct request_queue *q)
282 struct bfq_io_cq *icq;
284 spin_lock_irqsave(q->queue_lock, flags);
285 icq = icq_to_bic(ioc_lookup_icq(ioc, q));
286 spin_unlock_irqrestore(q->queue_lock, flags);
295 * Scheduler run of queue, if there are requests pending and no one in the
296 * driver that will restart queueing.
298 void bfq_schedule_dispatch(struct bfq_data *bfqd)
300 if (bfqd->queued != 0) {
301 bfq_log(bfqd, "schedule dispatch");
302 blk_mq_run_hw_queues(bfqd->queue, true);
306 #define bfq_class_idle(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
307 #define bfq_class_rt(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_RT)
309 #define bfq_sample_valid(samples) ((samples) > 80)
312 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
313 * We choose the request that is closesr to the head right now. Distance
314 * behind the head is penalized and only allowed to a certain extent.
316 static struct request *bfq_choose_req(struct bfq_data *bfqd,
321 sector_t s1, s2, d1 = 0, d2 = 0;
322 unsigned long back_max;
323 #define BFQ_RQ1_WRAP 0x01 /* request 1 wraps */
324 #define BFQ_RQ2_WRAP 0x02 /* request 2 wraps */
325 unsigned int wrap = 0; /* bit mask: requests behind the disk head? */
327 if (!rq1 || rq1 == rq2)
332 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
334 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
336 if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
338 else if ((rq2->cmd_flags & REQ_META) && !(rq1->cmd_flags & REQ_META))
341 s1 = blk_rq_pos(rq1);
342 s2 = blk_rq_pos(rq2);
345 * By definition, 1KiB is 2 sectors.
347 back_max = bfqd->bfq_back_max * 2;
350 * Strict one way elevator _except_ in the case where we allow
351 * short backward seeks which are biased as twice the cost of a
352 * similar forward seek.
356 else if (s1 + back_max >= last)
357 d1 = (last - s1) * bfqd->bfq_back_penalty;
359 wrap |= BFQ_RQ1_WRAP;
363 else if (s2 + back_max >= last)
364 d2 = (last - s2) * bfqd->bfq_back_penalty;
366 wrap |= BFQ_RQ2_WRAP;
368 /* Found required data */
371 * By doing switch() on the bit mask "wrap" we avoid having to
372 * check two variables for all permutations: --> faster!
375 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
390 case BFQ_RQ1_WRAP|BFQ_RQ2_WRAP: /* both rqs wrapped */
393 * Since both rqs are wrapped,
394 * start with the one that's further behind head
395 * (--> only *one* back seek required),
396 * since back seek takes more time than forward.
405 static struct bfq_queue *
406 bfq_rq_pos_tree_lookup(struct bfq_data *bfqd, struct rb_root *root,
407 sector_t sector, struct rb_node **ret_parent,
408 struct rb_node ***rb_link)
410 struct rb_node **p, *parent;
411 struct bfq_queue *bfqq = NULL;
419 bfqq = rb_entry(parent, struct bfq_queue, pos_node);
422 * Sort strictly based on sector. Smallest to the left,
423 * largest to the right.
425 if (sector > blk_rq_pos(bfqq->next_rq))
427 else if (sector < blk_rq_pos(bfqq->next_rq))
435 *ret_parent = parent;
439 bfq_log(bfqd, "rq_pos_tree_lookup %llu: returning %d",
440 (unsigned long long)sector,
441 bfqq ? bfqq->pid : 0);
446 void bfq_pos_tree_add_move(struct bfq_data *bfqd, struct bfq_queue *bfqq)
448 struct rb_node **p, *parent;
449 struct bfq_queue *__bfqq;
451 if (bfqq->pos_root) {
452 rb_erase(&bfqq->pos_node, bfqq->pos_root);
453 bfqq->pos_root = NULL;
456 if (bfq_class_idle(bfqq))
461 bfqq->pos_root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree;
462 __bfqq = bfq_rq_pos_tree_lookup(bfqd, bfqq->pos_root,
463 blk_rq_pos(bfqq->next_rq), &parent, &p);
465 rb_link_node(&bfqq->pos_node, parent, p);
466 rb_insert_color(&bfqq->pos_node, bfqq->pos_root);
468 bfqq->pos_root = NULL;
472 * Tell whether there are active queues or groups with differentiated weights.
474 static bool bfq_differentiated_weights(struct bfq_data *bfqd)
477 * For weights to differ, at least one of the trees must contain
478 * at least two nodes.
480 return (!RB_EMPTY_ROOT(&bfqd->queue_weights_tree) &&
481 (bfqd->queue_weights_tree.rb_node->rb_left ||
482 bfqd->queue_weights_tree.rb_node->rb_right)
483 #ifdef CONFIG_BFQ_GROUP_IOSCHED
485 (!RB_EMPTY_ROOT(&bfqd->group_weights_tree) &&
486 (bfqd->group_weights_tree.rb_node->rb_left ||
487 bfqd->group_weights_tree.rb_node->rb_right)
493 * The following function returns true if every queue must receive the
494 * same share of the throughput (this condition is used when deciding
495 * whether idling may be disabled, see the comments in the function
496 * bfq_bfqq_may_idle()).
498 * Such a scenario occurs when:
499 * 1) all active queues have the same weight,
500 * 2) all active groups at the same level in the groups tree have the same
502 * 3) all active groups at the same level in the groups tree have the same
503 * number of children.
505 * Unfortunately, keeping the necessary state for evaluating exactly the
506 * above symmetry conditions would be quite complex and time-consuming.
507 * Therefore this function evaluates, instead, the following stronger
508 * sub-conditions, for which it is much easier to maintain the needed
510 * 1) all active queues have the same weight,
511 * 2) all active groups have the same weight,
512 * 3) all active groups have at most one active child each.
513 * In particular, the last two conditions are always true if hierarchical
514 * support and the cgroups interface are not enabled, thus no state needs
515 * to be maintained in this case.
517 static bool bfq_symmetric_scenario(struct bfq_data *bfqd)
519 return !bfq_differentiated_weights(bfqd);
523 * If the weight-counter tree passed as input contains no counter for
524 * the weight of the input entity, then add that counter; otherwise just
525 * increment the existing counter.
527 * Note that weight-counter trees contain few nodes in mostly symmetric
528 * scenarios. For example, if all queues have the same weight, then the
529 * weight-counter tree for the queues may contain at most one node.
530 * This holds even if low_latency is on, because weight-raised queues
531 * are not inserted in the tree.
532 * In most scenarios, the rate at which nodes are created/destroyed
535 void bfq_weights_tree_add(struct bfq_data *bfqd, struct bfq_entity *entity,
536 struct rb_root *root)
538 struct rb_node **new = &(root->rb_node), *parent = NULL;
541 * Do not insert if the entity is already associated with a
542 * counter, which happens if:
543 * 1) the entity is associated with a queue,
544 * 2) a request arrival has caused the queue to become both
545 * non-weight-raised, and hence change its weight, and
546 * backlogged; in this respect, each of the two events
547 * causes an invocation of this function,
548 * 3) this is the invocation of this function caused by the
549 * second event. This second invocation is actually useless,
550 * and we handle this fact by exiting immediately. More
551 * efficient or clearer solutions might possibly be adopted.
553 if (entity->weight_counter)
557 struct bfq_weight_counter *__counter = container_of(*new,
558 struct bfq_weight_counter,
562 if (entity->weight == __counter->weight) {
563 entity->weight_counter = __counter;
566 if (entity->weight < __counter->weight)
567 new = &((*new)->rb_left);
569 new = &((*new)->rb_right);
572 entity->weight_counter = kzalloc(sizeof(struct bfq_weight_counter),
576 * In the unlucky event of an allocation failure, we just
577 * exit. This will cause the weight of entity to not be
578 * considered in bfq_differentiated_weights, which, in its
579 * turn, causes the scenario to be deemed wrongly symmetric in
580 * case entity's weight would have been the only weight making
581 * the scenario asymmetric. On the bright side, no unbalance
582 * will however occur when entity becomes inactive again (the
583 * invocation of this function is triggered by an activation
584 * of entity). In fact, bfq_weights_tree_remove does nothing
585 * if !entity->weight_counter.
587 if (unlikely(!entity->weight_counter))
590 entity->weight_counter->weight = entity->weight;
591 rb_link_node(&entity->weight_counter->weights_node, parent, new);
592 rb_insert_color(&entity->weight_counter->weights_node, root);
595 entity->weight_counter->num_active++;
599 * Decrement the weight counter associated with the entity, and, if the
600 * counter reaches 0, remove the counter from the tree.
601 * See the comments to the function bfq_weights_tree_add() for considerations
604 void bfq_weights_tree_remove(struct bfq_data *bfqd, struct bfq_entity *entity,
605 struct rb_root *root)
607 if (!entity->weight_counter)
610 entity->weight_counter->num_active--;
611 if (entity->weight_counter->num_active > 0)
612 goto reset_entity_pointer;
614 rb_erase(&entity->weight_counter->weights_node, root);
615 kfree(entity->weight_counter);
617 reset_entity_pointer:
618 entity->weight_counter = NULL;
622 * Return expired entry, or NULL to just start from scratch in rbtree.
624 static struct request *bfq_check_fifo(struct bfq_queue *bfqq,
625 struct request *last)
629 if (bfq_bfqq_fifo_expire(bfqq))
632 bfq_mark_bfqq_fifo_expire(bfqq);
634 rq = rq_entry_fifo(bfqq->fifo.next);
636 if (rq == last || ktime_get_ns() < rq->fifo_time)
639 bfq_log_bfqq(bfqq->bfqd, bfqq, "check_fifo: returned %p", rq);
643 static struct request *bfq_find_next_rq(struct bfq_data *bfqd,
644 struct bfq_queue *bfqq,
645 struct request *last)
647 struct rb_node *rbnext = rb_next(&last->rb_node);
648 struct rb_node *rbprev = rb_prev(&last->rb_node);
649 struct request *next, *prev = NULL;
651 /* Follow expired path, else get first next available. */
652 next = bfq_check_fifo(bfqq, last);
657 prev = rb_entry_rq(rbprev);
660 next = rb_entry_rq(rbnext);
662 rbnext = rb_first(&bfqq->sort_list);
663 if (rbnext && rbnext != &last->rb_node)
664 next = rb_entry_rq(rbnext);
667 return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last));
670 /* see the definition of bfq_async_charge_factor for details */
671 static unsigned long bfq_serv_to_charge(struct request *rq,
672 struct bfq_queue *bfqq)
674 if (bfq_bfqq_sync(bfqq) || bfqq->wr_coeff > 1)
675 return blk_rq_sectors(rq);
678 * If there are no weight-raised queues, then amplify service
679 * by just the async charge factor; otherwise amplify service
680 * by twice the async charge factor, to further reduce latency
681 * for weight-raised queues.
683 if (bfqq->bfqd->wr_busy_queues == 0)
684 return blk_rq_sectors(rq) * bfq_async_charge_factor;
686 return blk_rq_sectors(rq) * 2 * bfq_async_charge_factor;
690 * bfq_updated_next_req - update the queue after a new next_rq selection.
691 * @bfqd: the device data the queue belongs to.
692 * @bfqq: the queue to update.
694 * If the first request of a queue changes we make sure that the queue
695 * has enough budget to serve at least its first request (if the
696 * request has grown). We do this because if the queue has not enough
697 * budget for its first request, it has to go through two dispatch
698 * rounds to actually get it dispatched.
700 static void bfq_updated_next_req(struct bfq_data *bfqd,
701 struct bfq_queue *bfqq)
703 struct bfq_entity *entity = &bfqq->entity;
704 struct request *next_rq = bfqq->next_rq;
705 unsigned long new_budget;
710 if (bfqq == bfqd->in_service_queue)
712 * In order not to break guarantees, budgets cannot be
713 * changed after an entity has been selected.
717 new_budget = max_t(unsigned long, bfqq->max_budget,
718 bfq_serv_to_charge(next_rq, bfqq));
719 if (entity->budget != new_budget) {
720 entity->budget = new_budget;
721 bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu",
723 bfq_requeue_bfqq(bfqd, bfqq, false);
728 bfq_bfqq_resume_state(struct bfq_queue *bfqq, struct bfq_data *bfqd,
729 struct bfq_io_cq *bic, bool bfq_already_existing)
731 unsigned int old_wr_coeff = bfqq->wr_coeff;
732 bool busy = bfq_already_existing && bfq_bfqq_busy(bfqq);
734 if (bic->saved_has_short_ttime)
735 bfq_mark_bfqq_has_short_ttime(bfqq);
737 bfq_clear_bfqq_has_short_ttime(bfqq);
739 if (bic->saved_IO_bound)
740 bfq_mark_bfqq_IO_bound(bfqq);
742 bfq_clear_bfqq_IO_bound(bfqq);
744 bfqq->ttime = bic->saved_ttime;
745 bfqq->wr_coeff = bic->saved_wr_coeff;
746 bfqq->wr_start_at_switch_to_srt = bic->saved_wr_start_at_switch_to_srt;
747 bfqq->last_wr_start_finish = bic->saved_last_wr_start_finish;
748 bfqq->wr_cur_max_time = bic->saved_wr_cur_max_time;
750 if (bfqq->wr_coeff > 1 && (bfq_bfqq_in_large_burst(bfqq) ||
751 time_is_before_jiffies(bfqq->last_wr_start_finish +
752 bfqq->wr_cur_max_time))) {
753 bfq_log_bfqq(bfqq->bfqd, bfqq,
754 "resume state: switching off wr");
759 /* make sure weight will be updated, however we got here */
760 bfqq->entity.prio_changed = 1;
765 if (old_wr_coeff == 1 && bfqq->wr_coeff > 1)
766 bfqd->wr_busy_queues++;
767 else if (old_wr_coeff > 1 && bfqq->wr_coeff == 1)
768 bfqd->wr_busy_queues--;
771 static int bfqq_process_refs(struct bfq_queue *bfqq)
773 return bfqq->ref - bfqq->allocated - bfqq->entity.on_st;
776 /* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */
777 static void bfq_reset_burst_list(struct bfq_data *bfqd, struct bfq_queue *bfqq)
779 struct bfq_queue *item;
780 struct hlist_node *n;
782 hlist_for_each_entry_safe(item, n, &bfqd->burst_list, burst_list_node)
783 hlist_del_init(&item->burst_list_node);
784 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
785 bfqd->burst_size = 1;
786 bfqd->burst_parent_entity = bfqq->entity.parent;
789 /* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */
790 static void bfq_add_to_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
792 /* Increment burst size to take into account also bfqq */
795 if (bfqd->burst_size == bfqd->bfq_large_burst_thresh) {
796 struct bfq_queue *pos, *bfqq_item;
797 struct hlist_node *n;
800 * Enough queues have been activated shortly after each
801 * other to consider this burst as large.
803 bfqd->large_burst = true;
806 * We can now mark all queues in the burst list as
807 * belonging to a large burst.
809 hlist_for_each_entry(bfqq_item, &bfqd->burst_list,
811 bfq_mark_bfqq_in_large_burst(bfqq_item);
812 bfq_mark_bfqq_in_large_burst(bfqq);
815 * From now on, and until the current burst finishes, any
816 * new queue being activated shortly after the last queue
817 * was inserted in the burst can be immediately marked as
818 * belonging to a large burst. So the burst list is not
819 * needed any more. Remove it.
821 hlist_for_each_entry_safe(pos, n, &bfqd->burst_list,
823 hlist_del_init(&pos->burst_list_node);
825 * Burst not yet large: add bfqq to the burst list. Do
826 * not increment the ref counter for bfqq, because bfqq
827 * is removed from the burst list before freeing bfqq
830 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
834 * If many queues belonging to the same group happen to be created
835 * shortly after each other, then the processes associated with these
836 * queues have typically a common goal. In particular, bursts of queue
837 * creations are usually caused by services or applications that spawn
838 * many parallel threads/processes. Examples are systemd during boot,
839 * or git grep. To help these processes get their job done as soon as
840 * possible, it is usually better to not grant either weight-raising
841 * or device idling to their queues.
843 * In this comment we describe, firstly, the reasons why this fact
844 * holds, and, secondly, the next function, which implements the main
845 * steps needed to properly mark these queues so that they can then be
846 * treated in a different way.
848 * The above services or applications benefit mostly from a high
849 * throughput: the quicker the requests of the activated queues are
850 * cumulatively served, the sooner the target job of these queues gets
851 * completed. As a consequence, weight-raising any of these queues,
852 * which also implies idling the device for it, is almost always
853 * counterproductive. In most cases it just lowers throughput.
855 * On the other hand, a burst of queue creations may be caused also by
856 * the start of an application that does not consist of a lot of
857 * parallel I/O-bound threads. In fact, with a complex application,
858 * several short processes may need to be executed to start-up the
859 * application. In this respect, to start an application as quickly as
860 * possible, the best thing to do is in any case to privilege the I/O
861 * related to the application with respect to all other
862 * I/O. Therefore, the best strategy to start as quickly as possible
863 * an application that causes a burst of queue creations is to
864 * weight-raise all the queues created during the burst. This is the
865 * exact opposite of the best strategy for the other type of bursts.
867 * In the end, to take the best action for each of the two cases, the
868 * two types of bursts need to be distinguished. Fortunately, this
869 * seems relatively easy, by looking at the sizes of the bursts. In
870 * particular, we found a threshold such that only bursts with a
871 * larger size than that threshold are apparently caused by
872 * services or commands such as systemd or git grep. For brevity,
873 * hereafter we call just 'large' these bursts. BFQ *does not*
874 * weight-raise queues whose creation occurs in a large burst. In
875 * addition, for each of these queues BFQ performs or does not perform
876 * idling depending on which choice boosts the throughput more. The
877 * exact choice depends on the device and request pattern at
880 * Unfortunately, false positives may occur while an interactive task
881 * is starting (e.g., an application is being started). The
882 * consequence is that the queues associated with the task do not
883 * enjoy weight raising as expected. Fortunately these false positives
884 * are very rare. They typically occur if some service happens to
885 * start doing I/O exactly when the interactive task starts.
887 * Turning back to the next function, it implements all the steps
888 * needed to detect the occurrence of a large burst and to properly
889 * mark all the queues belonging to it (so that they can then be
890 * treated in a different way). This goal is achieved by maintaining a
891 * "burst list" that holds, temporarily, the queues that belong to the
892 * burst in progress. The list is then used to mark these queues as
893 * belonging to a large burst if the burst does become large. The main
894 * steps are the following.
896 * . when the very first queue is created, the queue is inserted into the
897 * list (as it could be the first queue in a possible burst)
899 * . if the current burst has not yet become large, and a queue Q that does
900 * not yet belong to the burst is activated shortly after the last time
901 * at which a new queue entered the burst list, then the function appends
902 * Q to the burst list
904 * . if, as a consequence of the previous step, the burst size reaches
905 * the large-burst threshold, then
907 * . all the queues in the burst list are marked as belonging to a
910 * . the burst list is deleted; in fact, the burst list already served
911 * its purpose (keeping temporarily track of the queues in a burst,
912 * so as to be able to mark them as belonging to a large burst in the
913 * previous sub-step), and now is not needed any more
915 * . the device enters a large-burst mode
917 * . if a queue Q that does not belong to the burst is created while
918 * the device is in large-burst mode and shortly after the last time
919 * at which a queue either entered the burst list or was marked as
920 * belonging to the current large burst, then Q is immediately marked
921 * as belonging to a large burst.
923 * . if a queue Q that does not belong to the burst is created a while
924 * later, i.e., not shortly after, than the last time at which a queue
925 * either entered the burst list or was marked as belonging to the
926 * current large burst, then the current burst is deemed as finished and:
928 * . the large-burst mode is reset if set
930 * . the burst list is emptied
932 * . Q is inserted in the burst list, as Q may be the first queue
933 * in a possible new burst (then the burst list contains just Q
936 static void bfq_handle_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
939 * If bfqq is already in the burst list or is part of a large
940 * burst, or finally has just been split, then there is
941 * nothing else to do.
943 if (!hlist_unhashed(&bfqq->burst_list_node) ||
944 bfq_bfqq_in_large_burst(bfqq) ||
945 time_is_after_eq_jiffies(bfqq->split_time +
946 msecs_to_jiffies(10)))
950 * If bfqq's creation happens late enough, or bfqq belongs to
951 * a different group than the burst group, then the current
952 * burst is finished, and related data structures must be
955 * In this respect, consider the special case where bfqq is
956 * the very first queue created after BFQ is selected for this
957 * device. In this case, last_ins_in_burst and
958 * burst_parent_entity are not yet significant when we get
959 * here. But it is easy to verify that, whether or not the
960 * following condition is true, bfqq will end up being
961 * inserted into the burst list. In particular the list will
962 * happen to contain only bfqq. And this is exactly what has
963 * to happen, as bfqq may be the first queue of the first
966 if (time_is_before_jiffies(bfqd->last_ins_in_burst +
967 bfqd->bfq_burst_interval) ||
968 bfqq->entity.parent != bfqd->burst_parent_entity) {
969 bfqd->large_burst = false;
970 bfq_reset_burst_list(bfqd, bfqq);
975 * If we get here, then bfqq is being activated shortly after the
976 * last queue. So, if the current burst is also large, we can mark
977 * bfqq as belonging to this large burst immediately.
979 if (bfqd->large_burst) {
980 bfq_mark_bfqq_in_large_burst(bfqq);
985 * If we get here, then a large-burst state has not yet been
986 * reached, but bfqq is being activated shortly after the last
987 * queue. Then we add bfqq to the burst.
989 bfq_add_to_burst(bfqd, bfqq);
992 * At this point, bfqq either has been added to the current
993 * burst or has caused the current burst to terminate and a
994 * possible new burst to start. In particular, in the second
995 * case, bfqq has become the first queue in the possible new
996 * burst. In both cases last_ins_in_burst needs to be moved
999 bfqd->last_ins_in_burst = jiffies;
1002 static int bfq_bfqq_budget_left(struct bfq_queue *bfqq)
1004 struct bfq_entity *entity = &bfqq->entity;
1006 return entity->budget - entity->service;
1010 * If enough samples have been computed, return the current max budget
1011 * stored in bfqd, which is dynamically updated according to the
1012 * estimated disk peak rate; otherwise return the default max budget
1014 static int bfq_max_budget(struct bfq_data *bfqd)
1016 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1017 return bfq_default_max_budget;
1019 return bfqd->bfq_max_budget;
1023 * Return min budget, which is a fraction of the current or default
1024 * max budget (trying with 1/32)
1026 static int bfq_min_budget(struct bfq_data *bfqd)
1028 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1029 return bfq_default_max_budget / 32;
1031 return bfqd->bfq_max_budget / 32;
1035 * The next function, invoked after the input queue bfqq switches from
1036 * idle to busy, updates the budget of bfqq. The function also tells
1037 * whether the in-service queue should be expired, by returning
1038 * true. The purpose of expiring the in-service queue is to give bfqq
1039 * the chance to possibly preempt the in-service queue, and the reason
1040 * for preempting the in-service queue is to achieve one of the two
1043 * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has
1044 * expired because it has remained idle. In particular, bfqq may have
1045 * expired for one of the following two reasons:
1047 * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling
1048 * and did not make it to issue a new request before its last
1049 * request was served;
1051 * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue
1052 * a new request before the expiration of the idling-time.
1054 * Even if bfqq has expired for one of the above reasons, the process
1055 * associated with the queue may be however issuing requests greedily,
1056 * and thus be sensitive to the bandwidth it receives (bfqq may have
1057 * remained idle for other reasons: CPU high load, bfqq not enjoying
1058 * idling, I/O throttling somewhere in the path from the process to
1059 * the I/O scheduler, ...). But if, after every expiration for one of
1060 * the above two reasons, bfqq has to wait for the service of at least
1061 * one full budget of another queue before being served again, then
1062 * bfqq is likely to get a much lower bandwidth or resource time than
1063 * its reserved ones. To address this issue, two countermeasures need
1066 * First, the budget and the timestamps of bfqq need to be updated in
1067 * a special way on bfqq reactivation: they need to be updated as if
1068 * bfqq did not remain idle and did not expire. In fact, if they are
1069 * computed as if bfqq expired and remained idle until reactivation,
1070 * then the process associated with bfqq is treated as if, instead of
1071 * being greedy, it stopped issuing requests when bfqq remained idle,
1072 * and restarts issuing requests only on this reactivation. In other
1073 * words, the scheduler does not help the process recover the "service
1074 * hole" between bfqq expiration and reactivation. As a consequence,
1075 * the process receives a lower bandwidth than its reserved one. In
1076 * contrast, to recover this hole, the budget must be updated as if
1077 * bfqq was not expired at all before this reactivation, i.e., it must
1078 * be set to the value of the remaining budget when bfqq was
1079 * expired. Along the same line, timestamps need to be assigned the
1080 * value they had the last time bfqq was selected for service, i.e.,
1081 * before last expiration. Thus timestamps need to be back-shifted
1082 * with respect to their normal computation (see [1] for more details
1083 * on this tricky aspect).
1085 * Secondly, to allow the process to recover the hole, the in-service
1086 * queue must be expired too, to give bfqq the chance to preempt it
1087 * immediately. In fact, if bfqq has to wait for a full budget of the
1088 * in-service queue to be completed, then it may become impossible to
1089 * let the process recover the hole, even if the back-shifted
1090 * timestamps of bfqq are lower than those of the in-service queue. If
1091 * this happens for most or all of the holes, then the process may not
1092 * receive its reserved bandwidth. In this respect, it is worth noting
1093 * that, being the service of outstanding requests unpreemptible, a
1094 * little fraction of the holes may however be unrecoverable, thereby
1095 * causing a little loss of bandwidth.
1097 * The last important point is detecting whether bfqq does need this
1098 * bandwidth recovery. In this respect, the next function deems the
1099 * process associated with bfqq greedy, and thus allows it to recover
1100 * the hole, if: 1) the process is waiting for the arrival of a new
1101 * request (which implies that bfqq expired for one of the above two
1102 * reasons), and 2) such a request has arrived soon. The first
1103 * condition is controlled through the flag non_blocking_wait_rq,
1104 * while the second through the flag arrived_in_time. If both
1105 * conditions hold, then the function computes the budget in the
1106 * above-described special way, and signals that the in-service queue
1107 * should be expired. Timestamp back-shifting is done later in
1108 * __bfq_activate_entity.
1110 * 2. Reduce latency. Even if timestamps are not backshifted to let
1111 * the process associated with bfqq recover a service hole, bfqq may
1112 * however happen to have, after being (re)activated, a lower finish
1113 * timestamp than the in-service queue. That is, the next budget of
1114 * bfqq may have to be completed before the one of the in-service
1115 * queue. If this is the case, then preempting the in-service queue
1116 * allows this goal to be achieved, apart from the unpreemptible,
1117 * outstanding requests mentioned above.
1119 * Unfortunately, regardless of which of the above two goals one wants
1120 * to achieve, service trees need first to be updated to know whether
1121 * the in-service queue must be preempted. To have service trees
1122 * correctly updated, the in-service queue must be expired and
1123 * rescheduled, and bfqq must be scheduled too. This is one of the
1124 * most costly operations (in future versions, the scheduling
1125 * mechanism may be re-designed in such a way to make it possible to
1126 * know whether preemption is needed without needing to update service
1127 * trees). In addition, queue preemptions almost always cause random
1128 * I/O, and thus loss of throughput. Because of these facts, the next
1129 * function adopts the following simple scheme to avoid both costly
1130 * operations and too frequent preemptions: it requests the expiration
1131 * of the in-service queue (unconditionally) only for queues that need
1132 * to recover a hole, or that either are weight-raised or deserve to
1135 static bool bfq_bfqq_update_budg_for_activation(struct bfq_data *bfqd,
1136 struct bfq_queue *bfqq,
1137 bool arrived_in_time,
1138 bool wr_or_deserves_wr)
1140 struct bfq_entity *entity = &bfqq->entity;
1142 if (bfq_bfqq_non_blocking_wait_rq(bfqq) && arrived_in_time) {
1144 * We do not clear the flag non_blocking_wait_rq here, as
1145 * the latter is used in bfq_activate_bfqq to signal
1146 * that timestamps need to be back-shifted (and is
1147 * cleared right after).
1151 * In next assignment we rely on that either
1152 * entity->service or entity->budget are not updated
1153 * on expiration if bfqq is empty (see
1154 * __bfq_bfqq_recalc_budget). Thus both quantities
1155 * remain unchanged after such an expiration, and the
1156 * following statement therefore assigns to
1157 * entity->budget the remaining budget on such an
1158 * expiration. For clarity, entity->service is not
1159 * updated on expiration in any case, and, in normal
1160 * operation, is reset only when bfqq is selected for
1161 * service (see bfq_get_next_queue).
1163 entity->budget = min_t(unsigned long,
1164 bfq_bfqq_budget_left(bfqq),
1170 entity->budget = max_t(unsigned long, bfqq->max_budget,
1171 bfq_serv_to_charge(bfqq->next_rq, bfqq));
1172 bfq_clear_bfqq_non_blocking_wait_rq(bfqq);
1173 return wr_or_deserves_wr;
1176 static unsigned int bfq_wr_duration(struct bfq_data *bfqd)
1180 if (bfqd->bfq_wr_max_time > 0)
1181 return bfqd->bfq_wr_max_time;
1183 dur = bfqd->RT_prod;
1184 do_div(dur, bfqd->peak_rate);
1187 * Limit duration between 3 and 13 seconds. Tests show that
1188 * higher values than 13 seconds often yield the opposite of
1189 * the desired result, i.e., worsen responsiveness by letting
1190 * non-interactive and non-soft-real-time applications
1191 * preserve weight raising for a too long time interval.
1193 * On the other end, lower values than 3 seconds make it
1194 * difficult for most interactive tasks to complete their jobs
1195 * before weight-raising finishes.
1197 if (dur > msecs_to_jiffies(13000))
1198 dur = msecs_to_jiffies(13000);
1199 else if (dur < msecs_to_jiffies(3000))
1200 dur = msecs_to_jiffies(3000);
1205 static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data *bfqd,
1206 struct bfq_queue *bfqq,
1207 unsigned int old_wr_coeff,
1208 bool wr_or_deserves_wr,
1213 if (old_wr_coeff == 1 && wr_or_deserves_wr) {
1214 /* start a weight-raising period */
1216 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1217 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1219 bfqq->wr_start_at_switch_to_srt = jiffies;
1220 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1221 BFQ_SOFTRT_WEIGHT_FACTOR;
1222 bfqq->wr_cur_max_time =
1223 bfqd->bfq_wr_rt_max_time;
1227 * If needed, further reduce budget to make sure it is
1228 * close to bfqq's backlog, so as to reduce the
1229 * scheduling-error component due to a too large
1230 * budget. Do not care about throughput consequences,
1231 * but only about latency. Finally, do not assign a
1232 * too small budget either, to avoid increasing
1233 * latency by causing too frequent expirations.
1235 bfqq->entity.budget = min_t(unsigned long,
1236 bfqq->entity.budget,
1237 2 * bfq_min_budget(bfqd));
1238 } else if (old_wr_coeff > 1) {
1239 if (interactive) { /* update wr coeff and duration */
1240 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1241 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1242 } else if (in_burst)
1246 * The application is now or still meeting the
1247 * requirements for being deemed soft rt. We
1248 * can then correctly and safely (re)charge
1249 * the weight-raising duration for the
1250 * application with the weight-raising
1251 * duration for soft rt applications.
1253 * In particular, doing this recharge now, i.e.,
1254 * before the weight-raising period for the
1255 * application finishes, reduces the probability
1256 * of the following negative scenario:
1257 * 1) the weight of a soft rt application is
1258 * raised at startup (as for any newly
1259 * created application),
1260 * 2) since the application is not interactive,
1261 * at a certain time weight-raising is
1262 * stopped for the application,
1263 * 3) at that time the application happens to
1264 * still have pending requests, and hence
1265 * is destined to not have a chance to be
1266 * deemed soft rt before these requests are
1267 * completed (see the comments to the
1268 * function bfq_bfqq_softrt_next_start()
1269 * for details on soft rt detection),
1270 * 4) these pending requests experience a high
1271 * latency because the application is not
1272 * weight-raised while they are pending.
1274 if (bfqq->wr_cur_max_time !=
1275 bfqd->bfq_wr_rt_max_time) {
1276 bfqq->wr_start_at_switch_to_srt =
1277 bfqq->last_wr_start_finish;
1279 bfqq->wr_cur_max_time =
1280 bfqd->bfq_wr_rt_max_time;
1281 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1282 BFQ_SOFTRT_WEIGHT_FACTOR;
1284 bfqq->last_wr_start_finish = jiffies;
1289 static bool bfq_bfqq_idle_for_long_time(struct bfq_data *bfqd,
1290 struct bfq_queue *bfqq)
1292 return bfqq->dispatched == 0 &&
1293 time_is_before_jiffies(
1294 bfqq->budget_timeout +
1295 bfqd->bfq_wr_min_idle_time);
1298 static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data *bfqd,
1299 struct bfq_queue *bfqq,
1304 bool soft_rt, in_burst, wr_or_deserves_wr,
1305 bfqq_wants_to_preempt,
1306 idle_for_long_time = bfq_bfqq_idle_for_long_time(bfqd, bfqq),
1308 * See the comments on
1309 * bfq_bfqq_update_budg_for_activation for
1310 * details on the usage of the next variable.
1312 arrived_in_time = ktime_get_ns() <=
1313 bfqq->ttime.last_end_request +
1314 bfqd->bfq_slice_idle * 3;
1316 bfqg_stats_update_io_add(bfqq_group(RQ_BFQQ(rq)), bfqq, rq->cmd_flags);
1319 * bfqq deserves to be weight-raised if:
1321 * - it does not belong to a large burst,
1322 * - it has been idle for enough time or is soft real-time,
1323 * - is linked to a bfq_io_cq (it is not shared in any sense).
1325 in_burst = bfq_bfqq_in_large_burst(bfqq);
1326 soft_rt = bfqd->bfq_wr_max_softrt_rate > 0 &&
1328 time_is_before_jiffies(bfqq->soft_rt_next_start);
1329 *interactive = !in_burst && idle_for_long_time;
1330 wr_or_deserves_wr = bfqd->low_latency &&
1331 (bfqq->wr_coeff > 1 ||
1332 (bfq_bfqq_sync(bfqq) &&
1333 bfqq->bic && (*interactive || soft_rt)));
1336 * Using the last flag, update budget and check whether bfqq
1337 * may want to preempt the in-service queue.
1339 bfqq_wants_to_preempt =
1340 bfq_bfqq_update_budg_for_activation(bfqd, bfqq,
1345 * If bfqq happened to be activated in a burst, but has been
1346 * idle for much more than an interactive queue, then we
1347 * assume that, in the overall I/O initiated in the burst, the
1348 * I/O associated with bfqq is finished. So bfqq does not need
1349 * to be treated as a queue belonging to a burst
1350 * anymore. Accordingly, we reset bfqq's in_large_burst flag
1351 * if set, and remove bfqq from the burst list if it's
1352 * there. We do not decrement burst_size, because the fact
1353 * that bfqq does not need to belong to the burst list any
1354 * more does not invalidate the fact that bfqq was created in
1357 if (likely(!bfq_bfqq_just_created(bfqq)) &&
1358 idle_for_long_time &&
1359 time_is_before_jiffies(
1360 bfqq->budget_timeout +
1361 msecs_to_jiffies(10000))) {
1362 hlist_del_init(&bfqq->burst_list_node);
1363 bfq_clear_bfqq_in_large_burst(bfqq);
1366 bfq_clear_bfqq_just_created(bfqq);
1369 if (!bfq_bfqq_IO_bound(bfqq)) {
1370 if (arrived_in_time) {
1371 bfqq->requests_within_timer++;
1372 if (bfqq->requests_within_timer >=
1373 bfqd->bfq_requests_within_timer)
1374 bfq_mark_bfqq_IO_bound(bfqq);
1376 bfqq->requests_within_timer = 0;
1379 if (bfqd->low_latency) {
1380 if (unlikely(time_is_after_jiffies(bfqq->split_time)))
1383 jiffies - bfqd->bfq_wr_min_idle_time - 1;
1385 if (time_is_before_jiffies(bfqq->split_time +
1386 bfqd->bfq_wr_min_idle_time)) {
1387 bfq_update_bfqq_wr_on_rq_arrival(bfqd, bfqq,
1394 if (old_wr_coeff != bfqq->wr_coeff)
1395 bfqq->entity.prio_changed = 1;
1399 bfqq->last_idle_bklogged = jiffies;
1400 bfqq->service_from_backlogged = 0;
1401 bfq_clear_bfqq_softrt_update(bfqq);
1403 bfq_add_bfqq_busy(bfqd, bfqq);
1406 * Expire in-service queue only if preemption may be needed
1407 * for guarantees. In this respect, the function
1408 * next_queue_may_preempt just checks a simple, necessary
1409 * condition, and not a sufficient condition based on
1410 * timestamps. In fact, for the latter condition to be
1411 * evaluated, timestamps would need first to be updated, and
1412 * this operation is quite costly (see the comments on the
1413 * function bfq_bfqq_update_budg_for_activation).
1415 if (bfqd->in_service_queue && bfqq_wants_to_preempt &&
1416 bfqd->in_service_queue->wr_coeff < bfqq->wr_coeff &&
1417 next_queue_may_preempt(bfqd))
1418 bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
1419 false, BFQQE_PREEMPTED);
1422 static void bfq_add_request(struct request *rq)
1424 struct bfq_queue *bfqq = RQ_BFQQ(rq);
1425 struct bfq_data *bfqd = bfqq->bfqd;
1426 struct request *next_rq, *prev;
1427 unsigned int old_wr_coeff = bfqq->wr_coeff;
1428 bool interactive = false;
1430 bfq_log_bfqq(bfqd, bfqq, "add_request %d", rq_is_sync(rq));
1431 bfqq->queued[rq_is_sync(rq)]++;
1434 elv_rb_add(&bfqq->sort_list, rq);
1437 * Check if this request is a better next-serve candidate.
1439 prev = bfqq->next_rq;
1440 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position);
1441 bfqq->next_rq = next_rq;
1444 * Adjust priority tree position, if next_rq changes.
1446 if (prev != bfqq->next_rq)
1447 bfq_pos_tree_add_move(bfqd, bfqq);
1449 if (!bfq_bfqq_busy(bfqq)) /* switching to busy ... */
1450 bfq_bfqq_handle_idle_busy_switch(bfqd, bfqq, old_wr_coeff,
1453 if (bfqd->low_latency && old_wr_coeff == 1 && !rq_is_sync(rq) &&
1454 time_is_before_jiffies(
1455 bfqq->last_wr_start_finish +
1456 bfqd->bfq_wr_min_inter_arr_async)) {
1457 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1458 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1460 bfqd->wr_busy_queues++;
1461 bfqq->entity.prio_changed = 1;
1463 if (prev != bfqq->next_rq)
1464 bfq_updated_next_req(bfqd, bfqq);
1468 * Assign jiffies to last_wr_start_finish in the following
1471 * . if bfqq is not going to be weight-raised, because, for
1472 * non weight-raised queues, last_wr_start_finish stores the
1473 * arrival time of the last request; as of now, this piece
1474 * of information is used only for deciding whether to
1475 * weight-raise async queues
1477 * . if bfqq is not weight-raised, because, if bfqq is now
1478 * switching to weight-raised, then last_wr_start_finish
1479 * stores the time when weight-raising starts
1481 * . if bfqq is interactive, because, regardless of whether
1482 * bfqq is currently weight-raised, the weight-raising
1483 * period must start or restart (this case is considered
1484 * separately because it is not detected by the above
1485 * conditions, if bfqq is already weight-raised)
1487 * last_wr_start_finish has to be updated also if bfqq is soft
1488 * real-time, because the weight-raising period is constantly
1489 * restarted on idle-to-busy transitions for these queues, but
1490 * this is already done in bfq_bfqq_handle_idle_busy_switch if
1493 if (bfqd->low_latency &&
1494 (old_wr_coeff == 1 || bfqq->wr_coeff == 1 || interactive))
1495 bfqq->last_wr_start_finish = jiffies;
1498 static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd,
1500 struct request_queue *q)
1502 struct bfq_queue *bfqq = bfqd->bio_bfqq;
1506 return elv_rb_find(&bfqq->sort_list, bio_end_sector(bio));
1511 static sector_t get_sdist(sector_t last_pos, struct request *rq)
1514 return abs(blk_rq_pos(rq) - last_pos);
1519 #if 0 /* Still not clear if we can do without next two functions */
1520 static void bfq_activate_request(struct request_queue *q, struct request *rq)
1522 struct bfq_data *bfqd = q->elevator->elevator_data;
1524 bfqd->rq_in_driver++;
1527 static void bfq_deactivate_request(struct request_queue *q, struct request *rq)
1529 struct bfq_data *bfqd = q->elevator->elevator_data;
1531 bfqd->rq_in_driver--;
1535 static void bfq_remove_request(struct request_queue *q,
1538 struct bfq_queue *bfqq = RQ_BFQQ(rq);
1539 struct bfq_data *bfqd = bfqq->bfqd;
1540 const int sync = rq_is_sync(rq);
1542 if (bfqq->next_rq == rq) {
1543 bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq);
1544 bfq_updated_next_req(bfqd, bfqq);
1547 if (rq->queuelist.prev != &rq->queuelist)
1548 list_del_init(&rq->queuelist);
1549 bfqq->queued[sync]--;
1551 elv_rb_del(&bfqq->sort_list, rq);
1553 elv_rqhash_del(q, rq);
1554 if (q->last_merge == rq)
1555 q->last_merge = NULL;
1557 if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
1558 bfqq->next_rq = NULL;
1560 if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->in_service_queue) {
1561 bfq_del_bfqq_busy(bfqd, bfqq, false);
1563 * bfqq emptied. In normal operation, when
1564 * bfqq is empty, bfqq->entity.service and
1565 * bfqq->entity.budget must contain,
1566 * respectively, the service received and the
1567 * budget used last time bfqq emptied. These
1568 * facts do not hold in this case, as at least
1569 * this last removal occurred while bfqq is
1570 * not in service. To avoid inconsistencies,
1571 * reset both bfqq->entity.service and
1572 * bfqq->entity.budget, if bfqq has still a
1573 * process that may issue I/O requests to it.
1575 bfqq->entity.budget = bfqq->entity.service = 0;
1579 * Remove queue from request-position tree as it is empty.
1581 if (bfqq->pos_root) {
1582 rb_erase(&bfqq->pos_node, bfqq->pos_root);
1583 bfqq->pos_root = NULL;
1587 if (rq->cmd_flags & REQ_META)
1588 bfqq->meta_pending--;
1590 bfqg_stats_update_io_remove(bfqq_group(bfqq), rq->cmd_flags);
1593 static bool bfq_bio_merge(struct blk_mq_hw_ctx *hctx, struct bio *bio)
1595 struct request_queue *q = hctx->queue;
1596 struct bfq_data *bfqd = q->elevator->elevator_data;
1597 struct request *free = NULL;
1599 * bfq_bic_lookup grabs the queue_lock: invoke it now and
1600 * store its return value for later use, to avoid nesting
1601 * queue_lock inside the bfqd->lock. We assume that the bic
1602 * returned by bfq_bic_lookup does not go away before
1603 * bfqd->lock is taken.
1605 struct bfq_io_cq *bic = bfq_bic_lookup(bfqd, current->io_context, q);
1608 spin_lock_irq(&bfqd->lock);
1611 bfqd->bio_bfqq = bic_to_bfqq(bic, op_is_sync(bio->bi_opf));
1613 bfqd->bio_bfqq = NULL;
1614 bfqd->bio_bic = bic;
1616 ret = blk_mq_sched_try_merge(q, bio, &free);
1619 blk_mq_free_request(free);
1620 spin_unlock_irq(&bfqd->lock);
1625 static int bfq_request_merge(struct request_queue *q, struct request **req,
1628 struct bfq_data *bfqd = q->elevator->elevator_data;
1629 struct request *__rq;
1631 __rq = bfq_find_rq_fmerge(bfqd, bio, q);
1632 if (__rq && elv_bio_merge_ok(__rq, bio)) {
1634 return ELEVATOR_FRONT_MERGE;
1637 return ELEVATOR_NO_MERGE;
1640 static void bfq_request_merged(struct request_queue *q, struct request *req,
1641 enum elv_merge type)
1643 if (type == ELEVATOR_FRONT_MERGE &&
1644 rb_prev(&req->rb_node) &&
1646 blk_rq_pos(container_of(rb_prev(&req->rb_node),
1647 struct request, rb_node))) {
1648 struct bfq_queue *bfqq = RQ_BFQQ(req);
1649 struct bfq_data *bfqd = bfqq->bfqd;
1650 struct request *prev, *next_rq;
1652 /* Reposition request in its sort_list */
1653 elv_rb_del(&bfqq->sort_list, req);
1654 elv_rb_add(&bfqq->sort_list, req);
1656 /* Choose next request to be served for bfqq */
1657 prev = bfqq->next_rq;
1658 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, req,
1659 bfqd->last_position);
1660 bfqq->next_rq = next_rq;
1662 * If next_rq changes, update both the queue's budget to
1663 * fit the new request and the queue's position in its
1666 if (prev != bfqq->next_rq) {
1667 bfq_updated_next_req(bfqd, bfqq);
1668 bfq_pos_tree_add_move(bfqd, bfqq);
1673 static void bfq_requests_merged(struct request_queue *q, struct request *rq,
1674 struct request *next)
1676 struct bfq_queue *bfqq = RQ_BFQQ(rq), *next_bfqq = RQ_BFQQ(next);
1678 if (!RB_EMPTY_NODE(&rq->rb_node))
1680 spin_lock_irq(&bfqq->bfqd->lock);
1683 * If next and rq belong to the same bfq_queue and next is older
1684 * than rq, then reposition rq in the fifo (by substituting next
1685 * with rq). Otherwise, if next and rq belong to different
1686 * bfq_queues, never reposition rq: in fact, we would have to
1687 * reposition it with respect to next's position in its own fifo,
1688 * which would most certainly be too expensive with respect to
1691 if (bfqq == next_bfqq &&
1692 !list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1693 next->fifo_time < rq->fifo_time) {
1694 list_del_init(&rq->queuelist);
1695 list_replace_init(&next->queuelist, &rq->queuelist);
1696 rq->fifo_time = next->fifo_time;
1699 if (bfqq->next_rq == next)
1702 bfq_remove_request(q, next);
1704 spin_unlock_irq(&bfqq->bfqd->lock);
1706 bfqg_stats_update_io_merged(bfqq_group(bfqq), next->cmd_flags);
1709 /* Must be called with bfqq != NULL */
1710 static void bfq_bfqq_end_wr(struct bfq_queue *bfqq)
1712 if (bfq_bfqq_busy(bfqq))
1713 bfqq->bfqd->wr_busy_queues--;
1715 bfqq->wr_cur_max_time = 0;
1716 bfqq->last_wr_start_finish = jiffies;
1718 * Trigger a weight change on the next invocation of
1719 * __bfq_entity_update_weight_prio.
1721 bfqq->entity.prio_changed = 1;
1724 void bfq_end_wr_async_queues(struct bfq_data *bfqd,
1725 struct bfq_group *bfqg)
1729 for (i = 0; i < 2; i++)
1730 for (j = 0; j < IOPRIO_BE_NR; j++)
1731 if (bfqg->async_bfqq[i][j])
1732 bfq_bfqq_end_wr(bfqg->async_bfqq[i][j]);
1733 if (bfqg->async_idle_bfqq)
1734 bfq_bfqq_end_wr(bfqg->async_idle_bfqq);
1737 static void bfq_end_wr(struct bfq_data *bfqd)
1739 struct bfq_queue *bfqq;
1741 spin_lock_irq(&bfqd->lock);
1743 list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
1744 bfq_bfqq_end_wr(bfqq);
1745 list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list)
1746 bfq_bfqq_end_wr(bfqq);
1747 bfq_end_wr_async(bfqd);
1749 spin_unlock_irq(&bfqd->lock);
1752 static sector_t bfq_io_struct_pos(void *io_struct, bool request)
1755 return blk_rq_pos(io_struct);
1757 return ((struct bio *)io_struct)->bi_iter.bi_sector;
1760 static int bfq_rq_close_to_sector(void *io_struct, bool request,
1763 return abs(bfq_io_struct_pos(io_struct, request) - sector) <=
1767 static struct bfq_queue *bfqq_find_close(struct bfq_data *bfqd,
1768 struct bfq_queue *bfqq,
1771 struct rb_root *root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree;
1772 struct rb_node *parent, *node;
1773 struct bfq_queue *__bfqq;
1775 if (RB_EMPTY_ROOT(root))
1779 * First, if we find a request starting at the end of the last
1780 * request, choose it.
1782 __bfqq = bfq_rq_pos_tree_lookup(bfqd, root, sector, &parent, NULL);
1787 * If the exact sector wasn't found, the parent of the NULL leaf
1788 * will contain the closest sector (rq_pos_tree sorted by
1789 * next_request position).
1791 __bfqq = rb_entry(parent, struct bfq_queue, pos_node);
1792 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
1795 if (blk_rq_pos(__bfqq->next_rq) < sector)
1796 node = rb_next(&__bfqq->pos_node);
1798 node = rb_prev(&__bfqq->pos_node);
1802 __bfqq = rb_entry(node, struct bfq_queue, pos_node);
1803 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
1809 static struct bfq_queue *bfq_find_close_cooperator(struct bfq_data *bfqd,
1810 struct bfq_queue *cur_bfqq,
1813 struct bfq_queue *bfqq;
1816 * We shall notice if some of the queues are cooperating,
1817 * e.g., working closely on the same area of the device. In
1818 * that case, we can group them together and: 1) don't waste
1819 * time idling, and 2) serve the union of their requests in
1820 * the best possible order for throughput.
1822 bfqq = bfqq_find_close(bfqd, cur_bfqq, sector);
1823 if (!bfqq || bfqq == cur_bfqq)
1829 static struct bfq_queue *
1830 bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
1832 int process_refs, new_process_refs;
1833 struct bfq_queue *__bfqq;
1836 * If there are no process references on the new_bfqq, then it is
1837 * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain
1838 * may have dropped their last reference (not just their last process
1841 if (!bfqq_process_refs(new_bfqq))
1844 /* Avoid a circular list and skip interim queue merges. */
1845 while ((__bfqq = new_bfqq->new_bfqq)) {
1851 process_refs = bfqq_process_refs(bfqq);
1852 new_process_refs = bfqq_process_refs(new_bfqq);
1854 * If the process for the bfqq has gone away, there is no
1855 * sense in merging the queues.
1857 if (process_refs == 0 || new_process_refs == 0)
1860 bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d",
1864 * Merging is just a redirection: the requests of the process
1865 * owning one of the two queues are redirected to the other queue.
1866 * The latter queue, in its turn, is set as shared if this is the
1867 * first time that the requests of some process are redirected to
1870 * We redirect bfqq to new_bfqq and not the opposite, because
1871 * we are in the context of the process owning bfqq, thus we
1872 * have the io_cq of this process. So we can immediately
1873 * configure this io_cq to redirect the requests of the
1874 * process to new_bfqq. In contrast, the io_cq of new_bfqq is
1875 * not available any more (new_bfqq->bic == NULL).
1877 * Anyway, even in case new_bfqq coincides with the in-service
1878 * queue, redirecting requests the in-service queue is the
1879 * best option, as we feed the in-service queue with new
1880 * requests close to the last request served and, by doing so,
1881 * are likely to increase the throughput.
1883 bfqq->new_bfqq = new_bfqq;
1884 new_bfqq->ref += process_refs;
1888 static bool bfq_may_be_close_cooperator(struct bfq_queue *bfqq,
1889 struct bfq_queue *new_bfqq)
1891 if (bfq_class_idle(bfqq) || bfq_class_idle(new_bfqq) ||
1892 (bfqq->ioprio_class != new_bfqq->ioprio_class))
1896 * If either of the queues has already been detected as seeky,
1897 * then merging it with the other queue is unlikely to lead to
1900 if (BFQQ_SEEKY(bfqq) || BFQQ_SEEKY(new_bfqq))
1904 * Interleaved I/O is known to be done by (some) applications
1905 * only for reads, so it does not make sense to merge async
1908 if (!bfq_bfqq_sync(bfqq) || !bfq_bfqq_sync(new_bfqq))
1915 * If this function returns true, then bfqq cannot be merged. The idea
1916 * is that true cooperation happens very early after processes start
1917 * to do I/O. Usually, late cooperations are just accidental false
1918 * positives. In case bfqq is weight-raised, such false positives
1919 * would evidently degrade latency guarantees for bfqq.
1921 static bool wr_from_too_long(struct bfq_queue *bfqq)
1923 return bfqq->wr_coeff > 1 &&
1924 time_is_before_jiffies(bfqq->last_wr_start_finish +
1925 msecs_to_jiffies(100));
1929 * Attempt to schedule a merge of bfqq with the currently in-service
1930 * queue or with a close queue among the scheduled queues. Return
1931 * NULL if no merge was scheduled, a pointer to the shared bfq_queue
1932 * structure otherwise.
1934 * The OOM queue is not allowed to participate to cooperation: in fact, since
1935 * the requests temporarily redirected to the OOM queue could be redirected
1936 * again to dedicated queues at any time, the state needed to correctly
1937 * handle merging with the OOM queue would be quite complex and expensive
1938 * to maintain. Besides, in such a critical condition as an out of memory,
1939 * the benefits of queue merging may be little relevant, or even negligible.
1941 * Weight-raised queues can be merged only if their weight-raising
1942 * period has just started. In fact cooperating processes are usually
1943 * started together. Thus, with this filter we avoid false positives
1944 * that would jeopardize low-latency guarantees.
1946 * WARNING: queue merging may impair fairness among non-weight raised
1947 * queues, for at least two reasons: 1) the original weight of a
1948 * merged queue may change during the merged state, 2) even being the
1949 * weight the same, a merged queue may be bloated with many more
1950 * requests than the ones produced by its originally-associated
1953 static struct bfq_queue *
1954 bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq,
1955 void *io_struct, bool request)
1957 struct bfq_queue *in_service_bfqq, *new_bfqq;
1960 return bfqq->new_bfqq;
1963 wr_from_too_long(bfqq) ||
1964 unlikely(bfqq == &bfqd->oom_bfqq))
1967 /* If there is only one backlogged queue, don't search. */
1968 if (bfqd->busy_queues == 1)
1971 in_service_bfqq = bfqd->in_service_queue;
1973 if (!in_service_bfqq || in_service_bfqq == bfqq
1974 || wr_from_too_long(in_service_bfqq) ||
1975 unlikely(in_service_bfqq == &bfqd->oom_bfqq))
1976 goto check_scheduled;
1978 if (bfq_rq_close_to_sector(io_struct, request, bfqd->last_position) &&
1979 bfqq->entity.parent == in_service_bfqq->entity.parent &&
1980 bfq_may_be_close_cooperator(bfqq, in_service_bfqq)) {
1981 new_bfqq = bfq_setup_merge(bfqq, in_service_bfqq);
1986 * Check whether there is a cooperator among currently scheduled
1987 * queues. The only thing we need is that the bio/request is not
1988 * NULL, as we need it to establish whether a cooperator exists.
1991 new_bfqq = bfq_find_close_cooperator(bfqd, bfqq,
1992 bfq_io_struct_pos(io_struct, request));
1994 if (new_bfqq && !wr_from_too_long(new_bfqq) &&
1995 likely(new_bfqq != &bfqd->oom_bfqq) &&
1996 bfq_may_be_close_cooperator(bfqq, new_bfqq))
1997 return bfq_setup_merge(bfqq, new_bfqq);
2002 static void bfq_bfqq_save_state(struct bfq_queue *bfqq)
2004 struct bfq_io_cq *bic = bfqq->bic;
2007 * If !bfqq->bic, the queue is already shared or its requests
2008 * have already been redirected to a shared queue; both idle window
2009 * and weight raising state have already been saved. Do nothing.
2014 bic->saved_ttime = bfqq->ttime;
2015 bic->saved_has_short_ttime = bfq_bfqq_has_short_ttime(bfqq);
2016 bic->saved_IO_bound = bfq_bfqq_IO_bound(bfqq);
2017 bic->saved_in_large_burst = bfq_bfqq_in_large_burst(bfqq);
2018 bic->was_in_burst_list = !hlist_unhashed(&bfqq->burst_list_node);
2019 bic->saved_wr_coeff = bfqq->wr_coeff;
2020 bic->saved_wr_start_at_switch_to_srt = bfqq->wr_start_at_switch_to_srt;
2021 bic->saved_last_wr_start_finish = bfqq->last_wr_start_finish;
2022 bic->saved_wr_cur_max_time = bfqq->wr_cur_max_time;
2026 bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic,
2027 struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2029 bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu",
2030 (unsigned long)new_bfqq->pid);
2031 /* Save weight raising and idle window of the merged queues */
2032 bfq_bfqq_save_state(bfqq);
2033 bfq_bfqq_save_state(new_bfqq);
2034 if (bfq_bfqq_IO_bound(bfqq))
2035 bfq_mark_bfqq_IO_bound(new_bfqq);
2036 bfq_clear_bfqq_IO_bound(bfqq);
2039 * If bfqq is weight-raised, then let new_bfqq inherit
2040 * weight-raising. To reduce false positives, neglect the case
2041 * where bfqq has just been created, but has not yet made it
2042 * to be weight-raised (which may happen because EQM may merge
2043 * bfqq even before bfq_add_request is executed for the first
2044 * time for bfqq). Handling this case would however be very
2045 * easy, thanks to the flag just_created.
2047 if (new_bfqq->wr_coeff == 1 && bfqq->wr_coeff > 1) {
2048 new_bfqq->wr_coeff = bfqq->wr_coeff;
2049 new_bfqq->wr_cur_max_time = bfqq->wr_cur_max_time;
2050 new_bfqq->last_wr_start_finish = bfqq->last_wr_start_finish;
2051 new_bfqq->wr_start_at_switch_to_srt =
2052 bfqq->wr_start_at_switch_to_srt;
2053 if (bfq_bfqq_busy(new_bfqq))
2054 bfqd->wr_busy_queues++;
2055 new_bfqq->entity.prio_changed = 1;
2058 if (bfqq->wr_coeff > 1) { /* bfqq has given its wr to new_bfqq */
2060 bfqq->entity.prio_changed = 1;
2061 if (bfq_bfqq_busy(bfqq))
2062 bfqd->wr_busy_queues--;
2065 bfq_log_bfqq(bfqd, new_bfqq, "merge_bfqqs: wr_busy %d",
2066 bfqd->wr_busy_queues);
2069 * Merge queues (that is, let bic redirect its requests to new_bfqq)
2071 bic_set_bfqq(bic, new_bfqq, 1);
2072 bfq_mark_bfqq_coop(new_bfqq);
2074 * new_bfqq now belongs to at least two bics (it is a shared queue):
2075 * set new_bfqq->bic to NULL. bfqq either:
2076 * - does not belong to any bic any more, and hence bfqq->bic must
2077 * be set to NULL, or
2078 * - is a queue whose owning bics have already been redirected to a
2079 * different queue, hence the queue is destined to not belong to
2080 * any bic soon and bfqq->bic is already NULL (therefore the next
2081 * assignment causes no harm).
2083 new_bfqq->bic = NULL;
2085 /* release process reference to bfqq */
2086 bfq_put_queue(bfqq);
2089 static bool bfq_allow_bio_merge(struct request_queue *q, struct request *rq,
2092 struct bfq_data *bfqd = q->elevator->elevator_data;
2093 bool is_sync = op_is_sync(bio->bi_opf);
2094 struct bfq_queue *bfqq = bfqd->bio_bfqq, *new_bfqq;
2097 * Disallow merge of a sync bio into an async request.
2099 if (is_sync && !rq_is_sync(rq))
2103 * Lookup the bfqq that this bio will be queued with. Allow
2104 * merge only if rq is queued there.
2110 * We take advantage of this function to perform an early merge
2111 * of the queues of possible cooperating processes.
2113 new_bfqq = bfq_setup_cooperator(bfqd, bfqq, bio, false);
2116 * bic still points to bfqq, then it has not yet been
2117 * redirected to some other bfq_queue, and a queue
2118 * merge beween bfqq and new_bfqq can be safely
2119 * fulfillled, i.e., bic can be redirected to new_bfqq
2120 * and bfqq can be put.
2122 bfq_merge_bfqqs(bfqd, bfqd->bio_bic, bfqq,
2125 * If we get here, bio will be queued into new_queue,
2126 * so use new_bfqq to decide whether bio and rq can be
2132 * Change also bqfd->bio_bfqq, as
2133 * bfqd->bio_bic now points to new_bfqq, and
2134 * this function may be invoked again (and then may
2135 * use again bqfd->bio_bfqq).
2137 bfqd->bio_bfqq = bfqq;
2140 return bfqq == RQ_BFQQ(rq);
2144 * Set the maximum time for the in-service queue to consume its
2145 * budget. This prevents seeky processes from lowering the throughput.
2146 * In practice, a time-slice service scheme is used with seeky
2149 static void bfq_set_budget_timeout(struct bfq_data *bfqd,
2150 struct bfq_queue *bfqq)
2152 unsigned int timeout_coeff;
2154 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time)
2157 timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight;
2159 bfqd->last_budget_start = ktime_get();
2161 bfqq->budget_timeout = jiffies +
2162 bfqd->bfq_timeout * timeout_coeff;
2165 static void __bfq_set_in_service_queue(struct bfq_data *bfqd,
2166 struct bfq_queue *bfqq)
2169 bfqg_stats_update_avg_queue_size(bfqq_group(bfqq));
2170 bfq_clear_bfqq_fifo_expire(bfqq);
2172 bfqd->budgets_assigned = (bfqd->budgets_assigned * 7 + 256) / 8;
2174 if (time_is_before_jiffies(bfqq->last_wr_start_finish) &&
2175 bfqq->wr_coeff > 1 &&
2176 bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
2177 time_is_before_jiffies(bfqq->budget_timeout)) {
2179 * For soft real-time queues, move the start
2180 * of the weight-raising period forward by the
2181 * time the queue has not received any
2182 * service. Otherwise, a relatively long
2183 * service delay is likely to cause the
2184 * weight-raising period of the queue to end,
2185 * because of the short duration of the
2186 * weight-raising period of a soft real-time
2187 * queue. It is worth noting that this move
2188 * is not so dangerous for the other queues,
2189 * because soft real-time queues are not
2192 * To not add a further variable, we use the
2193 * overloaded field budget_timeout to
2194 * determine for how long the queue has not
2195 * received service, i.e., how much time has
2196 * elapsed since the queue expired. However,
2197 * this is a little imprecise, because
2198 * budget_timeout is set to jiffies if bfqq
2199 * not only expires, but also remains with no
2202 if (time_after(bfqq->budget_timeout,
2203 bfqq->last_wr_start_finish))
2204 bfqq->last_wr_start_finish +=
2205 jiffies - bfqq->budget_timeout;
2207 bfqq->last_wr_start_finish = jiffies;
2210 bfq_set_budget_timeout(bfqd, bfqq);
2211 bfq_log_bfqq(bfqd, bfqq,
2212 "set_in_service_queue, cur-budget = %d",
2213 bfqq->entity.budget);
2216 bfqd->in_service_queue = bfqq;
2220 * Get and set a new queue for service.
2222 static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd)
2224 struct bfq_queue *bfqq = bfq_get_next_queue(bfqd);
2226 __bfq_set_in_service_queue(bfqd, bfqq);
2230 static void bfq_arm_slice_timer(struct bfq_data *bfqd)
2232 struct bfq_queue *bfqq = bfqd->in_service_queue;
2235 bfq_mark_bfqq_wait_request(bfqq);
2238 * We don't want to idle for seeks, but we do want to allow
2239 * fair distribution of slice time for a process doing back-to-back
2240 * seeks. So allow a little bit of time for him to submit a new rq.
2242 sl = bfqd->bfq_slice_idle;
2244 * Unless the queue is being weight-raised or the scenario is
2245 * asymmetric, grant only minimum idle time if the queue
2246 * is seeky. A long idling is preserved for a weight-raised
2247 * queue, or, more in general, in an asymmetric scenario,
2248 * because a long idling is needed for guaranteeing to a queue
2249 * its reserved share of the throughput (in particular, it is
2250 * needed if the queue has a higher weight than some other
2253 if (BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 &&
2254 bfq_symmetric_scenario(bfqd))
2255 sl = min_t(u64, sl, BFQ_MIN_TT);
2257 bfqd->last_idling_start = ktime_get();
2258 hrtimer_start(&bfqd->idle_slice_timer, ns_to_ktime(sl),
2260 bfqg_stats_set_start_idle_time(bfqq_group(bfqq));
2264 * In autotuning mode, max_budget is dynamically recomputed as the
2265 * amount of sectors transferred in timeout at the estimated peak
2266 * rate. This enables BFQ to utilize a full timeslice with a full
2267 * budget, even if the in-service queue is served at peak rate. And
2268 * this maximises throughput with sequential workloads.
2270 static unsigned long bfq_calc_max_budget(struct bfq_data *bfqd)
2272 return (u64)bfqd->peak_rate * USEC_PER_MSEC *
2273 jiffies_to_msecs(bfqd->bfq_timeout)>>BFQ_RATE_SHIFT;
2277 * Update parameters related to throughput and responsiveness, as a
2278 * function of the estimated peak rate. See comments on
2279 * bfq_calc_max_budget(), and on T_slow and T_fast arrays.
2281 static void update_thr_responsiveness_params(struct bfq_data *bfqd)
2283 int dev_type = blk_queue_nonrot(bfqd->queue);
2285 if (bfqd->bfq_user_max_budget == 0)
2286 bfqd->bfq_max_budget =
2287 bfq_calc_max_budget(bfqd);
2289 if (bfqd->device_speed == BFQ_BFQD_FAST &&
2290 bfqd->peak_rate < device_speed_thresh[dev_type]) {
2291 bfqd->device_speed = BFQ_BFQD_SLOW;
2292 bfqd->RT_prod = R_slow[dev_type] *
2294 } else if (bfqd->device_speed == BFQ_BFQD_SLOW &&
2295 bfqd->peak_rate > device_speed_thresh[dev_type]) {
2296 bfqd->device_speed = BFQ_BFQD_FAST;
2297 bfqd->RT_prod = R_fast[dev_type] *
2302 "dev_type %s dev_speed_class = %s (%llu sects/sec), thresh %llu setcs/sec",
2303 dev_type == 0 ? "ROT" : "NONROT",
2304 bfqd->device_speed == BFQ_BFQD_FAST ? "FAST" : "SLOW",
2305 bfqd->device_speed == BFQ_BFQD_FAST ?
2306 (USEC_PER_SEC*(u64)R_fast[dev_type])>>BFQ_RATE_SHIFT :
2307 (USEC_PER_SEC*(u64)R_slow[dev_type])>>BFQ_RATE_SHIFT,
2308 (USEC_PER_SEC*(u64)device_speed_thresh[dev_type])>>
2312 static void bfq_reset_rate_computation(struct bfq_data *bfqd,
2315 if (rq != NULL) { /* new rq dispatch now, reset accordingly */
2316 bfqd->last_dispatch = bfqd->first_dispatch = ktime_get_ns();
2317 bfqd->peak_rate_samples = 1;
2318 bfqd->sequential_samples = 0;
2319 bfqd->tot_sectors_dispatched = bfqd->last_rq_max_size =
2321 } else /* no new rq dispatched, just reset the number of samples */
2322 bfqd->peak_rate_samples = 0; /* full re-init on next disp. */
2325 "reset_rate_computation at end, sample %u/%u tot_sects %llu",
2326 bfqd->peak_rate_samples, bfqd->sequential_samples,
2327 bfqd->tot_sectors_dispatched);
2330 static void bfq_update_rate_reset(struct bfq_data *bfqd, struct request *rq)
2332 u32 rate, weight, divisor;
2335 * For the convergence property to hold (see comments on
2336 * bfq_update_peak_rate()) and for the assessment to be
2337 * reliable, a minimum number of samples must be present, and
2338 * a minimum amount of time must have elapsed. If not so, do
2339 * not compute new rate. Just reset parameters, to get ready
2340 * for a new evaluation attempt.
2342 if (bfqd->peak_rate_samples < BFQ_RATE_MIN_SAMPLES ||
2343 bfqd->delta_from_first < BFQ_RATE_MIN_INTERVAL)
2344 goto reset_computation;
2347 * If a new request completion has occurred after last
2348 * dispatch, then, to approximate the rate at which requests
2349 * have been served by the device, it is more precise to
2350 * extend the observation interval to the last completion.
2352 bfqd->delta_from_first =
2353 max_t(u64, bfqd->delta_from_first,
2354 bfqd->last_completion - bfqd->first_dispatch);
2357 * Rate computed in sects/usec, and not sects/nsec, for
2360 rate = div64_ul(bfqd->tot_sectors_dispatched<<BFQ_RATE_SHIFT,
2361 div_u64(bfqd->delta_from_first, NSEC_PER_USEC));
2364 * Peak rate not updated if:
2365 * - the percentage of sequential dispatches is below 3/4 of the
2366 * total, and rate is below the current estimated peak rate
2367 * - rate is unreasonably high (> 20M sectors/sec)
2369 if ((bfqd->sequential_samples < (3 * bfqd->peak_rate_samples)>>2 &&
2370 rate <= bfqd->peak_rate) ||
2371 rate > 20<<BFQ_RATE_SHIFT)
2372 goto reset_computation;
2375 * We have to update the peak rate, at last! To this purpose,
2376 * we use a low-pass filter. We compute the smoothing constant
2377 * of the filter as a function of the 'weight' of the new
2380 * As can be seen in next formulas, we define this weight as a
2381 * quantity proportional to how sequential the workload is,
2382 * and to how long the observation time interval is.
2384 * The weight runs from 0 to 8. The maximum value of the
2385 * weight, 8, yields the minimum value for the smoothing
2386 * constant. At this minimum value for the smoothing constant,
2387 * the measured rate contributes for half of the next value of
2388 * the estimated peak rate.
2390 * So, the first step is to compute the weight as a function
2391 * of how sequential the workload is. Note that the weight
2392 * cannot reach 9, because bfqd->sequential_samples cannot
2393 * become equal to bfqd->peak_rate_samples, which, in its
2394 * turn, holds true because bfqd->sequential_samples is not
2395 * incremented for the first sample.
2397 weight = (9 * bfqd->sequential_samples) / bfqd->peak_rate_samples;
2400 * Second step: further refine the weight as a function of the
2401 * duration of the observation interval.
2403 weight = min_t(u32, 8,
2404 div_u64(weight * bfqd->delta_from_first,
2405 BFQ_RATE_REF_INTERVAL));
2408 * Divisor ranging from 10, for minimum weight, to 2, for
2411 divisor = 10 - weight;
2414 * Finally, update peak rate:
2416 * peak_rate = peak_rate * (divisor-1) / divisor + rate / divisor
2418 bfqd->peak_rate *= divisor-1;
2419 bfqd->peak_rate /= divisor;
2420 rate /= divisor; /* smoothing constant alpha = 1/divisor */
2422 bfqd->peak_rate += rate;
2423 update_thr_responsiveness_params(bfqd);
2426 bfq_reset_rate_computation(bfqd, rq);
2430 * Update the read/write peak rate (the main quantity used for
2431 * auto-tuning, see update_thr_responsiveness_params()).
2433 * It is not trivial to estimate the peak rate (correctly): because of
2434 * the presence of sw and hw queues between the scheduler and the
2435 * device components that finally serve I/O requests, it is hard to
2436 * say exactly when a given dispatched request is served inside the
2437 * device, and for how long. As a consequence, it is hard to know
2438 * precisely at what rate a given set of requests is actually served
2441 * On the opposite end, the dispatch time of any request is trivially
2442 * available, and, from this piece of information, the "dispatch rate"
2443 * of requests can be immediately computed. So, the idea in the next
2444 * function is to use what is known, namely request dispatch times
2445 * (plus, when useful, request completion times), to estimate what is
2446 * unknown, namely in-device request service rate.
2448 * The main issue is that, because of the above facts, the rate at
2449 * which a certain set of requests is dispatched over a certain time
2450 * interval can vary greatly with respect to the rate at which the
2451 * same requests are then served. But, since the size of any
2452 * intermediate queue is limited, and the service scheme is lossless
2453 * (no request is silently dropped), the following obvious convergence
2454 * property holds: the number of requests dispatched MUST become
2455 * closer and closer to the number of requests completed as the
2456 * observation interval grows. This is the key property used in
2457 * the next function to estimate the peak service rate as a function
2458 * of the observed dispatch rate. The function assumes to be invoked
2459 * on every request dispatch.
2461 static void bfq_update_peak_rate(struct bfq_data *bfqd, struct request *rq)
2463 u64 now_ns = ktime_get_ns();
2465 if (bfqd->peak_rate_samples == 0) { /* first dispatch */
2466 bfq_log(bfqd, "update_peak_rate: goto reset, samples %d",
2467 bfqd->peak_rate_samples);
2468 bfq_reset_rate_computation(bfqd, rq);
2469 goto update_last_values; /* will add one sample */
2473 * Device idle for very long: the observation interval lasting
2474 * up to this dispatch cannot be a valid observation interval
2475 * for computing a new peak rate (similarly to the late-
2476 * completion event in bfq_completed_request()). Go to
2477 * update_rate_and_reset to have the following three steps
2479 * - close the observation interval at the last (previous)
2480 * request dispatch or completion
2481 * - compute rate, if possible, for that observation interval
2482 * - start a new observation interval with this dispatch
2484 if (now_ns - bfqd->last_dispatch > 100*NSEC_PER_MSEC &&
2485 bfqd->rq_in_driver == 0)
2486 goto update_rate_and_reset;
2488 /* Update sampling information */
2489 bfqd->peak_rate_samples++;
2491 if ((bfqd->rq_in_driver > 0 ||
2492 now_ns - bfqd->last_completion < BFQ_MIN_TT)
2493 && get_sdist(bfqd->last_position, rq) < BFQQ_SEEK_THR)
2494 bfqd->sequential_samples++;
2496 bfqd->tot_sectors_dispatched += blk_rq_sectors(rq);
2498 /* Reset max observed rq size every 32 dispatches */
2499 if (likely(bfqd->peak_rate_samples % 32))
2500 bfqd->last_rq_max_size =
2501 max_t(u32, blk_rq_sectors(rq), bfqd->last_rq_max_size);
2503 bfqd->last_rq_max_size = blk_rq_sectors(rq);
2505 bfqd->delta_from_first = now_ns - bfqd->first_dispatch;
2507 /* Target observation interval not yet reached, go on sampling */
2508 if (bfqd->delta_from_first < BFQ_RATE_REF_INTERVAL)
2509 goto update_last_values;
2511 update_rate_and_reset:
2512 bfq_update_rate_reset(bfqd, rq);
2514 bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2515 bfqd->last_dispatch = now_ns;
2519 * Remove request from internal lists.
2521 static void bfq_dispatch_remove(struct request_queue *q, struct request *rq)
2523 struct bfq_queue *bfqq = RQ_BFQQ(rq);
2526 * For consistency, the next instruction should have been
2527 * executed after removing the request from the queue and
2528 * dispatching it. We execute instead this instruction before
2529 * bfq_remove_request() (and hence introduce a temporary
2530 * inconsistency), for efficiency. In fact, should this
2531 * dispatch occur for a non in-service bfqq, this anticipated
2532 * increment prevents two counters related to bfqq->dispatched
2533 * from risking to be, first, uselessly decremented, and then
2534 * incremented again when the (new) value of bfqq->dispatched
2535 * happens to be taken into account.
2538 bfq_update_peak_rate(q->elevator->elevator_data, rq);
2540 bfq_remove_request(q, rq);
2543 static void __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq)
2546 * If this bfqq is shared between multiple processes, check
2547 * to make sure that those processes are still issuing I/Os
2548 * within the mean seek distance. If not, it may be time to
2549 * break the queues apart again.
2551 if (bfq_bfqq_coop(bfqq) && BFQQ_SEEKY(bfqq))
2552 bfq_mark_bfqq_split_coop(bfqq);
2554 if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
2555 if (bfqq->dispatched == 0)
2557 * Overloading budget_timeout field to store
2558 * the time at which the queue remains with no
2559 * backlog and no outstanding request; used by
2560 * the weight-raising mechanism.
2562 bfqq->budget_timeout = jiffies;
2564 bfq_del_bfqq_busy(bfqd, bfqq, true);
2566 bfq_requeue_bfqq(bfqd, bfqq, true);
2568 * Resort priority tree of potential close cooperators.
2570 bfq_pos_tree_add_move(bfqd, bfqq);
2574 * All in-service entities must have been properly deactivated
2575 * or requeued before executing the next function, which
2576 * resets all in-service entites as no more in service.
2578 __bfq_bfqd_reset_in_service(bfqd);
2582 * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior.
2583 * @bfqd: device data.
2584 * @bfqq: queue to update.
2585 * @reason: reason for expiration.
2587 * Handle the feedback on @bfqq budget at queue expiration.
2588 * See the body for detailed comments.
2590 static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd,
2591 struct bfq_queue *bfqq,
2592 enum bfqq_expiration reason)
2594 struct request *next_rq;
2595 int budget, min_budget;
2597 min_budget = bfq_min_budget(bfqd);
2599 if (bfqq->wr_coeff == 1)
2600 budget = bfqq->max_budget;
2602 * Use a constant, low budget for weight-raised queues,
2603 * to help achieve a low latency. Keep it slightly higher
2604 * than the minimum possible budget, to cause a little
2605 * bit fewer expirations.
2607 budget = 2 * min_budget;
2609 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %d, budg left %d",
2610 bfqq->entity.budget, bfq_bfqq_budget_left(bfqq));
2611 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %d, min budg %d",
2612 budget, bfq_min_budget(bfqd));
2613 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d",
2614 bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->in_service_queue));
2616 if (bfq_bfqq_sync(bfqq) && bfqq->wr_coeff == 1) {
2619 * Caveat: in all the following cases we trade latency
2622 case BFQQE_TOO_IDLE:
2624 * This is the only case where we may reduce
2625 * the budget: if there is no request of the
2626 * process still waiting for completion, then
2627 * we assume (tentatively) that the timer has
2628 * expired because the batch of requests of
2629 * the process could have been served with a
2630 * smaller budget. Hence, betting that
2631 * process will behave in the same way when it
2632 * becomes backlogged again, we reduce its
2633 * next budget. As long as we guess right,
2634 * this budget cut reduces the latency
2635 * experienced by the process.
2637 * However, if there are still outstanding
2638 * requests, then the process may have not yet
2639 * issued its next request just because it is
2640 * still waiting for the completion of some of
2641 * the still outstanding ones. So in this
2642 * subcase we do not reduce its budget, on the
2643 * contrary we increase it to possibly boost
2644 * the throughput, as discussed in the
2645 * comments to the BUDGET_TIMEOUT case.
2647 if (bfqq->dispatched > 0) /* still outstanding reqs */
2648 budget = min(budget * 2, bfqd->bfq_max_budget);
2650 if (budget > 5 * min_budget)
2651 budget -= 4 * min_budget;
2653 budget = min_budget;
2656 case BFQQE_BUDGET_TIMEOUT:
2658 * We double the budget here because it gives
2659 * the chance to boost the throughput if this
2660 * is not a seeky process (and has bumped into
2661 * this timeout because of, e.g., ZBR).
2663 budget = min(budget * 2, bfqd->bfq_max_budget);
2665 case BFQQE_BUDGET_EXHAUSTED:
2667 * The process still has backlog, and did not
2668 * let either the budget timeout or the disk
2669 * idling timeout expire. Hence it is not
2670 * seeky, has a short thinktime and may be
2671 * happy with a higher budget too. So
2672 * definitely increase the budget of this good
2673 * candidate to boost the disk throughput.
2675 budget = min(budget * 4, bfqd->bfq_max_budget);
2677 case BFQQE_NO_MORE_REQUESTS:
2679 * For queues that expire for this reason, it
2680 * is particularly important to keep the
2681 * budget close to the actual service they
2682 * need. Doing so reduces the timestamp
2683 * misalignment problem described in the
2684 * comments in the body of
2685 * __bfq_activate_entity. In fact, suppose
2686 * that a queue systematically expires for
2687 * BFQQE_NO_MORE_REQUESTS and presents a
2688 * new request in time to enjoy timestamp
2689 * back-shifting. The larger the budget of the
2690 * queue is with respect to the service the
2691 * queue actually requests in each service
2692 * slot, the more times the queue can be
2693 * reactivated with the same virtual finish
2694 * time. It follows that, even if this finish
2695 * time is pushed to the system virtual time
2696 * to reduce the consequent timestamp
2697 * misalignment, the queue unjustly enjoys for
2698 * many re-activations a lower finish time
2699 * than all newly activated queues.
2701 * The service needed by bfqq is measured
2702 * quite precisely by bfqq->entity.service.
2703 * Since bfqq does not enjoy device idling,
2704 * bfqq->entity.service is equal to the number
2705 * of sectors that the process associated with
2706 * bfqq requested to read/write before waiting
2707 * for request completions, or blocking for
2710 budget = max_t(int, bfqq->entity.service, min_budget);
2715 } else if (!bfq_bfqq_sync(bfqq)) {
2717 * Async queues get always the maximum possible
2718 * budget, as for them we do not care about latency
2719 * (in addition, their ability to dispatch is limited
2720 * by the charging factor).
2722 budget = bfqd->bfq_max_budget;
2725 bfqq->max_budget = budget;
2727 if (bfqd->budgets_assigned >= bfq_stats_min_budgets &&
2728 !bfqd->bfq_user_max_budget)
2729 bfqq->max_budget = min(bfqq->max_budget, bfqd->bfq_max_budget);
2732 * If there is still backlog, then assign a new budget, making
2733 * sure that it is large enough for the next request. Since
2734 * the finish time of bfqq must be kept in sync with the
2735 * budget, be sure to call __bfq_bfqq_expire() *after* this
2738 * If there is no backlog, then no need to update the budget;
2739 * it will be updated on the arrival of a new request.
2741 next_rq = bfqq->next_rq;
2743 bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget,
2744 bfq_serv_to_charge(next_rq, bfqq));
2746 bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %d",
2747 next_rq ? blk_rq_sectors(next_rq) : 0,
2748 bfqq->entity.budget);
2752 * Return true if the process associated with bfqq is "slow". The slow
2753 * flag is used, in addition to the budget timeout, to reduce the
2754 * amount of service provided to seeky processes, and thus reduce
2755 * their chances to lower the throughput. More details in the comments
2756 * on the function bfq_bfqq_expire().
2758 * An important observation is in order: as discussed in the comments
2759 * on the function bfq_update_peak_rate(), with devices with internal
2760 * queues, it is hard if ever possible to know when and for how long
2761 * an I/O request is processed by the device (apart from the trivial
2762 * I/O pattern where a new request is dispatched only after the
2763 * previous one has been completed). This makes it hard to evaluate
2764 * the real rate at which the I/O requests of each bfq_queue are
2765 * served. In fact, for an I/O scheduler like BFQ, serving a
2766 * bfq_queue means just dispatching its requests during its service
2767 * slot (i.e., until the budget of the queue is exhausted, or the
2768 * queue remains idle, or, finally, a timeout fires). But, during the
2769 * service slot of a bfq_queue, around 100 ms at most, the device may
2770 * be even still processing requests of bfq_queues served in previous
2771 * service slots. On the opposite end, the requests of the in-service
2772 * bfq_queue may be completed after the service slot of the queue
2775 * Anyway, unless more sophisticated solutions are used
2776 * (where possible), the sum of the sizes of the requests dispatched
2777 * during the service slot of a bfq_queue is probably the only
2778 * approximation available for the service received by the bfq_queue
2779 * during its service slot. And this sum is the quantity used in this
2780 * function to evaluate the I/O speed of a process.
2782 static bool bfq_bfqq_is_slow(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2783 bool compensate, enum bfqq_expiration reason,
2784 unsigned long *delta_ms)
2786 ktime_t delta_ktime;
2788 bool slow = BFQQ_SEEKY(bfqq); /* if delta too short, use seekyness */
2790 if (!bfq_bfqq_sync(bfqq))
2794 delta_ktime = bfqd->last_idling_start;
2796 delta_ktime = ktime_get();
2797 delta_ktime = ktime_sub(delta_ktime, bfqd->last_budget_start);
2798 delta_usecs = ktime_to_us(delta_ktime);
2800 /* don't use too short time intervals */
2801 if (delta_usecs < 1000) {
2802 if (blk_queue_nonrot(bfqd->queue))
2804 * give same worst-case guarantees as idling
2807 *delta_ms = BFQ_MIN_TT / NSEC_PER_MSEC;
2808 else /* charge at least one seek */
2809 *delta_ms = bfq_slice_idle / NSEC_PER_MSEC;
2814 *delta_ms = delta_usecs / USEC_PER_MSEC;
2817 * Use only long (> 20ms) intervals to filter out excessive
2818 * spikes in service rate estimation.
2820 if (delta_usecs > 20000) {
2822 * Caveat for rotational devices: processes doing I/O
2823 * in the slower disk zones tend to be slow(er) even
2824 * if not seeky. In this respect, the estimated peak
2825 * rate is likely to be an average over the disk
2826 * surface. Accordingly, to not be too harsh with
2827 * unlucky processes, a process is deemed slow only if
2828 * its rate has been lower than half of the estimated
2831 slow = bfqq->entity.service < bfqd->bfq_max_budget / 2;
2834 bfq_log_bfqq(bfqd, bfqq, "bfq_bfqq_is_slow: slow %d", slow);
2840 * To be deemed as soft real-time, an application must meet two
2841 * requirements. First, the application must not require an average
2842 * bandwidth higher than the approximate bandwidth required to playback or
2843 * record a compressed high-definition video.
2844 * The next function is invoked on the completion of the last request of a
2845 * batch, to compute the next-start time instant, soft_rt_next_start, such
2846 * that, if the next request of the application does not arrive before
2847 * soft_rt_next_start, then the above requirement on the bandwidth is met.
2849 * The second requirement is that the request pattern of the application is
2850 * isochronous, i.e., that, after issuing a request or a batch of requests,
2851 * the application stops issuing new requests until all its pending requests
2852 * have been completed. After that, the application may issue a new batch,
2854 * For this reason the next function is invoked to compute
2855 * soft_rt_next_start only for applications that meet this requirement,
2856 * whereas soft_rt_next_start is set to infinity for applications that do
2859 * Unfortunately, even a greedy application may happen to behave in an
2860 * isochronous way if the CPU load is high. In fact, the application may
2861 * stop issuing requests while the CPUs are busy serving other processes,
2862 * then restart, then stop again for a while, and so on. In addition, if
2863 * the disk achieves a low enough throughput with the request pattern
2864 * issued by the application (e.g., because the request pattern is random
2865 * and/or the device is slow), then the application may meet the above
2866 * bandwidth requirement too. To prevent such a greedy application to be
2867 * deemed as soft real-time, a further rule is used in the computation of
2868 * soft_rt_next_start: soft_rt_next_start must be higher than the current
2869 * time plus the maximum time for which the arrival of a request is waited
2870 * for when a sync queue becomes idle, namely bfqd->bfq_slice_idle.
2871 * This filters out greedy applications, as the latter issue instead their
2872 * next request as soon as possible after the last one has been completed
2873 * (in contrast, when a batch of requests is completed, a soft real-time
2874 * application spends some time processing data).
2876 * Unfortunately, the last filter may easily generate false positives if
2877 * only bfqd->bfq_slice_idle is used as a reference time interval and one
2878 * or both the following cases occur:
2879 * 1) HZ is so low that the duration of a jiffy is comparable to or higher
2880 * than bfqd->bfq_slice_idle. This happens, e.g., on slow devices with
2882 * 2) jiffies, instead of increasing at a constant rate, may stop increasing
2883 * for a while, then suddenly 'jump' by several units to recover the lost
2884 * increments. This seems to happen, e.g., inside virtual machines.
2885 * To address this issue, we do not use as a reference time interval just
2886 * bfqd->bfq_slice_idle, but bfqd->bfq_slice_idle plus a few jiffies. In
2887 * particular we add the minimum number of jiffies for which the filter
2888 * seems to be quite precise also in embedded systems and KVM/QEMU virtual
2891 static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data *bfqd,
2892 struct bfq_queue *bfqq)
2894 return max(bfqq->last_idle_bklogged +
2895 HZ * bfqq->service_from_backlogged /
2896 bfqd->bfq_wr_max_softrt_rate,
2897 jiffies + nsecs_to_jiffies(bfqq->bfqd->bfq_slice_idle) + 4);
2901 * Return the farthest future time instant according to jiffies
2904 static unsigned long bfq_greatest_from_now(void)
2906 return jiffies + MAX_JIFFY_OFFSET;
2910 * Return the farthest past time instant according to jiffies
2913 static unsigned long bfq_smallest_from_now(void)
2915 return jiffies - MAX_JIFFY_OFFSET;
2919 * bfq_bfqq_expire - expire a queue.
2920 * @bfqd: device owning the queue.
2921 * @bfqq: the queue to expire.
2922 * @compensate: if true, compensate for the time spent idling.
2923 * @reason: the reason causing the expiration.
2925 * If the process associated with bfqq does slow I/O (e.g., because it
2926 * issues random requests), we charge bfqq with the time it has been
2927 * in service instead of the service it has received (see
2928 * bfq_bfqq_charge_time for details on how this goal is achieved). As
2929 * a consequence, bfqq will typically get higher timestamps upon
2930 * reactivation, and hence it will be rescheduled as if it had
2931 * received more service than what it has actually received. In the
2932 * end, bfqq receives less service in proportion to how slowly its
2933 * associated process consumes its budgets (and hence how seriously it
2934 * tends to lower the throughput). In addition, this time-charging
2935 * strategy guarantees time fairness among slow processes. In
2936 * contrast, if the process associated with bfqq is not slow, we
2937 * charge bfqq exactly with the service it has received.
2939 * Charging time to the first type of queues and the exact service to
2940 * the other has the effect of using the WF2Q+ policy to schedule the
2941 * former on a timeslice basis, without violating service domain
2942 * guarantees among the latter.
2944 void bfq_bfqq_expire(struct bfq_data *bfqd,
2945 struct bfq_queue *bfqq,
2947 enum bfqq_expiration reason)
2950 unsigned long delta = 0;
2951 struct bfq_entity *entity = &bfqq->entity;
2955 * Check whether the process is slow (see bfq_bfqq_is_slow).
2957 slow = bfq_bfqq_is_slow(bfqd, bfqq, compensate, reason, &delta);
2960 * Increase service_from_backlogged before next statement,
2961 * because the possible next invocation of
2962 * bfq_bfqq_charge_time would likely inflate
2963 * entity->service. In contrast, service_from_backlogged must
2964 * contain real service, to enable the soft real-time
2965 * heuristic to correctly compute the bandwidth consumed by
2968 bfqq->service_from_backlogged += entity->service;
2971 * As above explained, charge slow (typically seeky) and
2972 * timed-out queues with the time and not the service
2973 * received, to favor sequential workloads.
2975 * Processes doing I/O in the slower disk zones will tend to
2976 * be slow(er) even if not seeky. Therefore, since the
2977 * estimated peak rate is actually an average over the disk
2978 * surface, these processes may timeout just for bad luck. To
2979 * avoid punishing them, do not charge time to processes that
2980 * succeeded in consuming at least 2/3 of their budget. This
2981 * allows BFQ to preserve enough elasticity to still perform
2982 * bandwidth, and not time, distribution with little unlucky
2983 * or quasi-sequential processes.
2985 if (bfqq->wr_coeff == 1 &&
2987 (reason == BFQQE_BUDGET_TIMEOUT &&
2988 bfq_bfqq_budget_left(bfqq) >= entity->budget / 3)))
2989 bfq_bfqq_charge_time(bfqd, bfqq, delta);
2991 if (reason == BFQQE_TOO_IDLE &&
2992 entity->service <= 2 * entity->budget / 10)
2993 bfq_clear_bfqq_IO_bound(bfqq);
2995 if (bfqd->low_latency && bfqq->wr_coeff == 1)
2996 bfqq->last_wr_start_finish = jiffies;
2998 if (bfqd->low_latency && bfqd->bfq_wr_max_softrt_rate > 0 &&
2999 RB_EMPTY_ROOT(&bfqq->sort_list)) {
3001 * If we get here, and there are no outstanding
3002 * requests, then the request pattern is isochronous
3003 * (see the comments on the function
3004 * bfq_bfqq_softrt_next_start()). Thus we can compute
3005 * soft_rt_next_start. If, instead, the queue still
3006 * has outstanding requests, then we have to wait for
3007 * the completion of all the outstanding requests to
3008 * discover whether the request pattern is actually
3011 if (bfqq->dispatched == 0)
3012 bfqq->soft_rt_next_start =
3013 bfq_bfqq_softrt_next_start(bfqd, bfqq);
3016 * The application is still waiting for the
3017 * completion of one or more requests:
3018 * prevent it from possibly being incorrectly
3019 * deemed as soft real-time by setting its
3020 * soft_rt_next_start to infinity. In fact,
3021 * without this assignment, the application
3022 * would be incorrectly deemed as soft
3024 * 1) it issued a new request before the
3025 * completion of all its in-flight
3027 * 2) at that time, its soft_rt_next_start
3028 * happened to be in the past.
3030 bfqq->soft_rt_next_start =
3031 bfq_greatest_from_now();
3033 * Schedule an update of soft_rt_next_start to when
3034 * the task may be discovered to be isochronous.
3036 bfq_mark_bfqq_softrt_update(bfqq);
3040 bfq_log_bfqq(bfqd, bfqq,
3041 "expire (%d, slow %d, num_disp %d, short_ttime %d)", reason,
3042 slow, bfqq->dispatched, bfq_bfqq_has_short_ttime(bfqq));
3045 * Increase, decrease or leave budget unchanged according to
3048 __bfq_bfqq_recalc_budget(bfqd, bfqq, reason);
3050 __bfq_bfqq_expire(bfqd, bfqq);
3052 /* mark bfqq as waiting a request only if a bic still points to it */
3053 if (ref > 1 && !bfq_bfqq_busy(bfqq) &&
3054 reason != BFQQE_BUDGET_TIMEOUT &&
3055 reason != BFQQE_BUDGET_EXHAUSTED)
3056 bfq_mark_bfqq_non_blocking_wait_rq(bfqq);
3060 * Budget timeout is not implemented through a dedicated timer, but
3061 * just checked on request arrivals and completions, as well as on
3062 * idle timer expirations.
3064 static bool bfq_bfqq_budget_timeout(struct bfq_queue *bfqq)
3066 return time_is_before_eq_jiffies(bfqq->budget_timeout);
3070 * If we expire a queue that is actively waiting (i.e., with the
3071 * device idled) for the arrival of a new request, then we may incur
3072 * the timestamp misalignment problem described in the body of the
3073 * function __bfq_activate_entity. Hence we return true only if this
3074 * condition does not hold, or if the queue is slow enough to deserve
3075 * only to be kicked off for preserving a high throughput.
3077 static bool bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq)
3079 bfq_log_bfqq(bfqq->bfqd, bfqq,
3080 "may_budget_timeout: wait_request %d left %d timeout %d",
3081 bfq_bfqq_wait_request(bfqq),
3082 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3,
3083 bfq_bfqq_budget_timeout(bfqq));
3085 return (!bfq_bfqq_wait_request(bfqq) ||
3086 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3)
3088 bfq_bfqq_budget_timeout(bfqq);
3092 * For a queue that becomes empty, device idling is allowed only if
3093 * this function returns true for the queue. As a consequence, since
3094 * device idling plays a critical role in both throughput boosting and
3095 * service guarantees, the return value of this function plays a
3096 * critical role in both these aspects as well.
3098 * In a nutshell, this function returns true only if idling is
3099 * beneficial for throughput or, even if detrimental for throughput,
3100 * idling is however necessary to preserve service guarantees (low
3101 * latency, desired throughput distribution, ...). In particular, on
3102 * NCQ-capable devices, this function tries to return false, so as to
3103 * help keep the drives' internal queues full, whenever this helps the
3104 * device boost the throughput without causing any service-guarantee
3107 * In more detail, the return value of this function is obtained by,
3108 * first, computing a number of boolean variables that take into
3109 * account throughput and service-guarantee issues, and, then,
3110 * combining these variables in a logical expression. Most of the
3111 * issues taken into account are not trivial. We discuss these issues
3112 * individually while introducing the variables.
3114 static bool bfq_bfqq_may_idle(struct bfq_queue *bfqq)
3116 struct bfq_data *bfqd = bfqq->bfqd;
3117 bool rot_without_queueing =
3118 !blk_queue_nonrot(bfqd->queue) && !bfqd->hw_tag,
3119 bfqq_sequential_and_IO_bound,
3120 idling_boosts_thr, idling_boosts_thr_without_issues,
3121 idling_needed_for_service_guarantees,
3122 asymmetric_scenario;
3124 if (bfqd->strict_guarantees)
3128 * Idling is performed only if slice_idle > 0. In addition, we
3131 * (b) bfqq is in the idle io prio class: in this case we do
3132 * not idle because we want to minimize the bandwidth that
3133 * queues in this class can steal to higher-priority queues
3135 if (bfqd->bfq_slice_idle == 0 || !bfq_bfqq_sync(bfqq) ||
3136 bfq_class_idle(bfqq))
3139 bfqq_sequential_and_IO_bound = !BFQQ_SEEKY(bfqq) &&
3140 bfq_bfqq_IO_bound(bfqq) && bfq_bfqq_has_short_ttime(bfqq);
3143 * The next variable takes into account the cases where idling
3144 * boosts the throughput.
3146 * The value of the variable is computed considering, first, that
3147 * idling is virtually always beneficial for the throughput if:
3148 * (a) the device is not NCQ-capable and rotational, or
3149 * (b) regardless of the presence of NCQ, the device is rotational and
3150 * the request pattern for bfqq is I/O-bound and sequential, or
3151 * (c) regardless of whether it is rotational, the device is
3152 * not NCQ-capable and the request pattern for bfqq is
3153 * I/O-bound and sequential.
3155 * Secondly, and in contrast to the above item (b), idling an
3156 * NCQ-capable flash-based device would not boost the
3157 * throughput even with sequential I/O; rather it would lower
3158 * the throughput in proportion to how fast the device
3159 * is. Accordingly, the next variable is true if any of the
3160 * above conditions (a), (b) or (c) is true, and, in
3161 * particular, happens to be false if bfqd is an NCQ-capable
3162 * flash-based device.
3164 idling_boosts_thr = rot_without_queueing ||
3165 ((!blk_queue_nonrot(bfqd->queue) || !bfqd->hw_tag) &&
3166 bfqq_sequential_and_IO_bound);
3169 * The value of the next variable,
3170 * idling_boosts_thr_without_issues, is equal to that of
3171 * idling_boosts_thr, unless a special case holds. In this
3172 * special case, described below, idling may cause problems to
3173 * weight-raised queues.
3175 * When the request pool is saturated (e.g., in the presence
3176 * of write hogs), if the processes associated with
3177 * non-weight-raised queues ask for requests at a lower rate,
3178 * then processes associated with weight-raised queues have a
3179 * higher probability to get a request from the pool
3180 * immediately (or at least soon) when they need one. Thus
3181 * they have a higher probability to actually get a fraction
3182 * of the device throughput proportional to their high
3183 * weight. This is especially true with NCQ-capable drives,
3184 * which enqueue several requests in advance, and further
3185 * reorder internally-queued requests.
3187 * For this reason, we force to false the value of
3188 * idling_boosts_thr_without_issues if there are weight-raised
3189 * busy queues. In this case, and if bfqq is not weight-raised,
3190 * this guarantees that the device is not idled for bfqq (if,
3191 * instead, bfqq is weight-raised, then idling will be
3192 * guaranteed by another variable, see below). Combined with
3193 * the timestamping rules of BFQ (see [1] for details), this
3194 * behavior causes bfqq, and hence any sync non-weight-raised
3195 * queue, to get a lower number of requests served, and thus
3196 * to ask for a lower number of requests from the request
3197 * pool, before the busy weight-raised queues get served
3198 * again. This often mitigates starvation problems in the
3199 * presence of heavy write workloads and NCQ, thereby
3200 * guaranteeing a higher application and system responsiveness
3201 * in these hostile scenarios.
3203 idling_boosts_thr_without_issues = idling_boosts_thr &&
3204 bfqd->wr_busy_queues == 0;
3207 * There is then a case where idling must be performed not
3208 * for throughput concerns, but to preserve service
3211 * To introduce this case, we can note that allowing the drive
3212 * to enqueue more than one request at a time, and hence
3213 * delegating de facto final scheduling decisions to the
3214 * drive's internal scheduler, entails loss of control on the
3215 * actual request service order. In particular, the critical
3216 * situation is when requests from different processes happen
3217 * to be present, at the same time, in the internal queue(s)
3218 * of the drive. In such a situation, the drive, by deciding
3219 * the service order of the internally-queued requests, does
3220 * determine also the actual throughput distribution among
3221 * these processes. But the drive typically has no notion or
3222 * concern about per-process throughput distribution, and
3223 * makes its decisions only on a per-request basis. Therefore,
3224 * the service distribution enforced by the drive's internal
3225 * scheduler is likely to coincide with the desired
3226 * device-throughput distribution only in a completely
3227 * symmetric scenario where:
3228 * (i) each of these processes must get the same throughput as
3230 * (ii) all these processes have the same I/O pattern
3231 (either sequential or random).
3232 * In fact, in such a scenario, the drive will tend to treat
3233 * the requests of each of these processes in about the same
3234 * way as the requests of the others, and thus to provide
3235 * each of these processes with about the same throughput
3236 * (which is exactly the desired throughput distribution). In
3237 * contrast, in any asymmetric scenario, device idling is
3238 * certainly needed to guarantee that bfqq receives its
3239 * assigned fraction of the device throughput (see [1] for
3242 * We address this issue by controlling, actually, only the
3243 * symmetry sub-condition (i), i.e., provided that
3244 * sub-condition (i) holds, idling is not performed,
3245 * regardless of whether sub-condition (ii) holds. In other
3246 * words, only if sub-condition (i) holds, then idling is
3247 * allowed, and the device tends to be prevented from queueing
3248 * many requests, possibly of several processes. The reason
3249 * for not controlling also sub-condition (ii) is that we
3250 * exploit preemption to preserve guarantees in case of
3251 * symmetric scenarios, even if (ii) does not hold, as
3252 * explained in the next two paragraphs.
3254 * Even if a queue, say Q, is expired when it remains idle, Q
3255 * can still preempt the new in-service queue if the next
3256 * request of Q arrives soon (see the comments on
3257 * bfq_bfqq_update_budg_for_activation). If all queues and
3258 * groups have the same weight, this form of preemption,
3259 * combined with the hole-recovery heuristic described in the
3260 * comments on function bfq_bfqq_update_budg_for_activation,
3261 * are enough to preserve a correct bandwidth distribution in
3262 * the mid term, even without idling. In fact, even if not
3263 * idling allows the internal queues of the device to contain
3264 * many requests, and thus to reorder requests, we can rather
3265 * safely assume that the internal scheduler still preserves a
3266 * minimum of mid-term fairness. The motivation for using
3267 * preemption instead of idling is that, by not idling,
3268 * service guarantees are preserved without minimally
3269 * sacrificing throughput. In other words, both a high
3270 * throughput and its desired distribution are obtained.
3272 * More precisely, this preemption-based, idleless approach
3273 * provides fairness in terms of IOPS, and not sectors per
3274 * second. This can be seen with a simple example. Suppose
3275 * that there are two queues with the same weight, but that
3276 * the first queue receives requests of 8 sectors, while the
3277 * second queue receives requests of 1024 sectors. In
3278 * addition, suppose that each of the two queues contains at
3279 * most one request at a time, which implies that each queue
3280 * always remains idle after it is served. Finally, after
3281 * remaining idle, each queue receives very quickly a new
3282 * request. It follows that the two queues are served
3283 * alternatively, preempting each other if needed. This
3284 * implies that, although both queues have the same weight,
3285 * the queue with large requests receives a service that is
3286 * 1024/8 times as high as the service received by the other
3289 * On the other hand, device idling is performed, and thus
3290 * pure sector-domain guarantees are provided, for the
3291 * following queues, which are likely to need stronger
3292 * throughput guarantees: weight-raised queues, and queues
3293 * with a higher weight than other queues. When such queues
3294 * are active, sub-condition (i) is false, which triggers
3297 * According to the above considerations, the next variable is
3298 * true (only) if sub-condition (i) holds. To compute the
3299 * value of this variable, we not only use the return value of
3300 * the function bfq_symmetric_scenario(), but also check
3301 * whether bfqq is being weight-raised, because
3302 * bfq_symmetric_scenario() does not take into account also
3303 * weight-raised queues (see comments on
3304 * bfq_weights_tree_add()).
3306 * As a side note, it is worth considering that the above
3307 * device-idling countermeasures may however fail in the
3308 * following unlucky scenario: if idling is (correctly)
3309 * disabled in a time period during which all symmetry
3310 * sub-conditions hold, and hence the device is allowed to
3311 * enqueue many requests, but at some later point in time some
3312 * sub-condition stops to hold, then it may become impossible
3313 * to let requests be served in the desired order until all
3314 * the requests already queued in the device have been served.
3316 asymmetric_scenario = bfqq->wr_coeff > 1 ||
3317 !bfq_symmetric_scenario(bfqd);
3320 * Finally, there is a case where maximizing throughput is the
3321 * best choice even if it may cause unfairness toward
3322 * bfqq. Such a case is when bfqq became active in a burst of
3323 * queue activations. Queues that became active during a large
3324 * burst benefit only from throughput, as discussed in the
3325 * comments on bfq_handle_burst. Thus, if bfqq became active
3326 * in a burst and not idling the device maximizes throughput,
3327 * then the device must no be idled, because not idling the
3328 * device provides bfqq and all other queues in the burst with
3329 * maximum benefit. Combining this and the above case, we can
3330 * now establish when idling is actually needed to preserve
3331 * service guarantees.
3333 idling_needed_for_service_guarantees =
3334 asymmetric_scenario && !bfq_bfqq_in_large_burst(bfqq);
3337 * We have now all the components we need to compute the
3338 * return value of the function, which is true only if idling
3339 * either boosts the throughput (without issues), or is
3340 * necessary to preserve service guarantees.
3342 return idling_boosts_thr_without_issues ||
3343 idling_needed_for_service_guarantees;
3347 * If the in-service queue is empty but the function bfq_bfqq_may_idle
3348 * returns true, then:
3349 * 1) the queue must remain in service and cannot be expired, and
3350 * 2) the device must be idled to wait for the possible arrival of a new
3351 * request for the queue.
3352 * See the comments on the function bfq_bfqq_may_idle for the reasons
3353 * why performing device idling is the best choice to boost the throughput
3354 * and preserve service guarantees when bfq_bfqq_may_idle itself
3357 static bool bfq_bfqq_must_idle(struct bfq_queue *bfqq)
3359 return RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_bfqq_may_idle(bfqq);
3363 * Select a queue for service. If we have a current queue in service,
3364 * check whether to continue servicing it, or retrieve and set a new one.
3366 static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd)
3368 struct bfq_queue *bfqq;
3369 struct request *next_rq;
3370 enum bfqq_expiration reason = BFQQE_BUDGET_TIMEOUT;
3372 bfqq = bfqd->in_service_queue;
3376 bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue");
3378 if (bfq_may_expire_for_budg_timeout(bfqq) &&
3379 !bfq_bfqq_wait_request(bfqq) &&
3380 !bfq_bfqq_must_idle(bfqq))
3385 * This loop is rarely executed more than once. Even when it
3386 * happens, it is much more convenient to re-execute this loop
3387 * than to return NULL and trigger a new dispatch to get a
3390 next_rq = bfqq->next_rq;
3392 * If bfqq has requests queued and it has enough budget left to
3393 * serve them, keep the queue, otherwise expire it.
3396 if (bfq_serv_to_charge(next_rq, bfqq) >
3397 bfq_bfqq_budget_left(bfqq)) {
3399 * Expire the queue for budget exhaustion,
3400 * which makes sure that the next budget is
3401 * enough to serve the next request, even if
3402 * it comes from the fifo expired path.
3404 reason = BFQQE_BUDGET_EXHAUSTED;
3408 * The idle timer may be pending because we may
3409 * not disable disk idling even when a new request
3412 if (bfq_bfqq_wait_request(bfqq)) {
3414 * If we get here: 1) at least a new request
3415 * has arrived but we have not disabled the
3416 * timer because the request was too small,
3417 * 2) then the block layer has unplugged
3418 * the device, causing the dispatch to be
3421 * Since the device is unplugged, now the
3422 * requests are probably large enough to
3423 * provide a reasonable throughput.
3424 * So we disable idling.
3426 bfq_clear_bfqq_wait_request(bfqq);
3427 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
3428 bfqg_stats_update_idle_time(bfqq_group(bfqq));
3435 * No requests pending. However, if the in-service queue is idling
3436 * for a new request, or has requests waiting for a completion and
3437 * may idle after their completion, then keep it anyway.
3439 if (bfq_bfqq_wait_request(bfqq) ||
3440 (bfqq->dispatched != 0 && bfq_bfqq_may_idle(bfqq))) {
3445 reason = BFQQE_NO_MORE_REQUESTS;
3447 bfq_bfqq_expire(bfqd, bfqq, false, reason);
3449 bfqq = bfq_set_in_service_queue(bfqd);
3451 bfq_log_bfqq(bfqd, bfqq, "select_queue: checking new queue");
3456 bfq_log_bfqq(bfqd, bfqq, "select_queue: returned this queue");
3458 bfq_log(bfqd, "select_queue: no queue returned");
3463 static void bfq_update_wr_data(struct bfq_data *bfqd, struct bfq_queue *bfqq)
3465 struct bfq_entity *entity = &bfqq->entity;
3467 if (bfqq->wr_coeff > 1) { /* queue is being weight-raised */
3468 bfq_log_bfqq(bfqd, bfqq,
3469 "raising period dur %u/%u msec, old coeff %u, w %d(%d)",
3470 jiffies_to_msecs(jiffies - bfqq->last_wr_start_finish),
3471 jiffies_to_msecs(bfqq->wr_cur_max_time),
3473 bfqq->entity.weight, bfqq->entity.orig_weight);
3475 if (entity->prio_changed)
3476 bfq_log_bfqq(bfqd, bfqq, "WARN: pending prio change");
3479 * If the queue was activated in a burst, or too much
3480 * time has elapsed from the beginning of this
3481 * weight-raising period, then end weight raising.
3483 if (bfq_bfqq_in_large_burst(bfqq))
3484 bfq_bfqq_end_wr(bfqq);
3485 else if (time_is_before_jiffies(bfqq->last_wr_start_finish +
3486 bfqq->wr_cur_max_time)) {
3487 if (bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time ||
3488 time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
3489 bfq_wr_duration(bfqd)))
3490 bfq_bfqq_end_wr(bfqq);
3492 /* switch back to interactive wr */
3493 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
3494 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
3495 bfqq->last_wr_start_finish =
3496 bfqq->wr_start_at_switch_to_srt;
3497 bfqq->entity.prio_changed = 1;
3502 * To improve latency (for this or other queues), immediately
3503 * update weight both if it must be raised and if it must be
3504 * lowered. Since, entity may be on some active tree here, and
3505 * might have a pending change of its ioprio class, invoke
3506 * next function with the last parameter unset (see the
3507 * comments on the function).
3509 if ((entity->weight > entity->orig_weight) != (bfqq->wr_coeff > 1))
3510 __bfq_entity_update_weight_prio(bfq_entity_service_tree(entity),
3515 * Dispatch next request from bfqq.
3517 static struct request *bfq_dispatch_rq_from_bfqq(struct bfq_data *bfqd,
3518 struct bfq_queue *bfqq)
3520 struct request *rq = bfqq->next_rq;
3521 unsigned long service_to_charge;
3523 service_to_charge = bfq_serv_to_charge(rq, bfqq);
3525 bfq_bfqq_served(bfqq, service_to_charge);
3527 bfq_dispatch_remove(bfqd->queue, rq);
3530 * If weight raising has to terminate for bfqq, then next
3531 * function causes an immediate update of bfqq's weight,
3532 * without waiting for next activation. As a consequence, on
3533 * expiration, bfqq will be timestamped as if has never been
3534 * weight-raised during this service slot, even if it has
3535 * received part or even most of the service as a
3536 * weight-raised queue. This inflates bfqq's timestamps, which
3537 * is beneficial, as bfqq is then more willing to leave the
3538 * device immediately to possible other weight-raised queues.
3540 bfq_update_wr_data(bfqd, bfqq);
3543 * Expire bfqq, pretending that its budget expired, if bfqq
3544 * belongs to CLASS_IDLE and other queues are waiting for
3547 if (bfqd->busy_queues > 1 && bfq_class_idle(bfqq))
3553 bfq_bfqq_expire(bfqd, bfqq, false, BFQQE_BUDGET_EXHAUSTED);
3557 static bool bfq_has_work(struct blk_mq_hw_ctx *hctx)
3559 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
3562 * Avoiding lock: a race on bfqd->busy_queues should cause at
3563 * most a call to dispatch for nothing
3565 return !list_empty_careful(&bfqd->dispatch) ||
3566 bfqd->busy_queues > 0;
3569 static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
3571 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
3572 struct request *rq = NULL;
3573 struct bfq_queue *bfqq = NULL;
3575 if (!list_empty(&bfqd->dispatch)) {
3576 rq = list_first_entry(&bfqd->dispatch, struct request,
3578 list_del_init(&rq->queuelist);
3584 * Increment counters here, because this
3585 * dispatch does not follow the standard
3586 * dispatch flow (where counters are
3591 goto inc_in_driver_start_rq;
3595 * We exploit the put_rq_private hook to decrement
3596 * rq_in_driver, but put_rq_private will not be
3597 * invoked on this request. So, to avoid unbalance,
3598 * just start this request, without incrementing
3599 * rq_in_driver. As a negative consequence,
3600 * rq_in_driver is deceptively lower than it should be
3601 * while this request is in service. This may cause
3602 * bfq_schedule_dispatch to be invoked uselessly.
3604 * As for implementing an exact solution, the
3605 * put_request hook, if defined, is probably invoked
3606 * also on this request. So, by exploiting this hook,
3607 * we could 1) increment rq_in_driver here, and 2)
3608 * decrement it in put_request. Such a solution would
3609 * let the value of the counter be always accurate,
3610 * but it would entail using an extra interface
3611 * function. This cost seems higher than the benefit,
3612 * being the frequency of non-elevator-private
3613 * requests very low.
3618 bfq_log(bfqd, "dispatch requests: %d busy queues", bfqd->busy_queues);
3620 if (bfqd->busy_queues == 0)
3624 * Force device to serve one request at a time if
3625 * strict_guarantees is true. Forcing this service scheme is
3626 * currently the ONLY way to guarantee that the request
3627 * service order enforced by the scheduler is respected by a
3628 * queueing device. Otherwise the device is free even to make
3629 * some unlucky request wait for as long as the device
3632 * Of course, serving one request at at time may cause loss of
3635 if (bfqd->strict_guarantees && bfqd->rq_in_driver > 0)
3638 bfqq = bfq_select_queue(bfqd);
3642 rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq);
3645 inc_in_driver_start_rq:
3646 bfqd->rq_in_driver++;
3648 rq->rq_flags |= RQF_STARTED;
3654 static struct request *bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
3656 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
3659 spin_lock_irq(&bfqd->lock);
3661 rq = __bfq_dispatch_request(hctx);
3662 spin_unlock_irq(&bfqd->lock);
3668 * Task holds one reference to the queue, dropped when task exits. Each rq
3669 * in-flight on this queue also holds a reference, dropped when rq is freed.
3671 * Scheduler lock must be held here. Recall not to use bfqq after calling
3672 * this function on it.
3674 void bfq_put_queue(struct bfq_queue *bfqq)
3676 #ifdef CONFIG_BFQ_GROUP_IOSCHED
3677 struct bfq_group *bfqg = bfqq_group(bfqq);
3681 bfq_log_bfqq(bfqq->bfqd, bfqq, "put_queue: %p %d",
3688 if (bfq_bfqq_sync(bfqq))
3690 * The fact that this queue is being destroyed does not
3691 * invalidate the fact that this queue may have been
3692 * activated during the current burst. As a consequence,
3693 * although the queue does not exist anymore, and hence
3694 * needs to be removed from the burst list if there,
3695 * the burst size has not to be decremented.
3697 hlist_del_init(&bfqq->burst_list_node);
3699 kmem_cache_free(bfq_pool, bfqq);
3700 #ifdef CONFIG_BFQ_GROUP_IOSCHED
3701 bfqg_and_blkg_put(bfqg);
3705 static void bfq_put_cooperator(struct bfq_queue *bfqq)
3707 struct bfq_queue *__bfqq, *next;
3710 * If this queue was scheduled to merge with another queue, be
3711 * sure to drop the reference taken on that queue (and others in
3712 * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs.
3714 __bfqq = bfqq->new_bfqq;
3718 next = __bfqq->new_bfqq;
3719 bfq_put_queue(__bfqq);
3724 static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq)
3726 if (bfqq == bfqd->in_service_queue) {
3727 __bfq_bfqq_expire(bfqd, bfqq);
3728 bfq_schedule_dispatch(bfqd);
3731 bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, bfqq->ref);
3733 bfq_put_cooperator(bfqq);
3735 bfq_put_queue(bfqq); /* release process reference */
3738 static void bfq_exit_icq_bfqq(struct bfq_io_cq *bic, bool is_sync)
3740 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
3741 struct bfq_data *bfqd;
3744 bfqd = bfqq->bfqd; /* NULL if scheduler already exited */
3747 unsigned long flags;
3749 spin_lock_irqsave(&bfqd->lock, flags);
3750 bfq_exit_bfqq(bfqd, bfqq);
3751 bic_set_bfqq(bic, NULL, is_sync);
3752 spin_unlock_irqrestore(&bfqd->lock, flags);
3756 static void bfq_exit_icq(struct io_cq *icq)
3758 struct bfq_io_cq *bic = icq_to_bic(icq);
3760 bfq_exit_icq_bfqq(bic, true);
3761 bfq_exit_icq_bfqq(bic, false);
3765 * Update the entity prio values; note that the new values will not
3766 * be used until the next (re)activation.
3769 bfq_set_next_ioprio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic)
3771 struct task_struct *tsk = current;
3773 struct bfq_data *bfqd = bfqq->bfqd;
3778 ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
3779 switch (ioprio_class) {
3781 dev_err(bfqq->bfqd->queue->backing_dev_info->dev,
3782 "bfq: bad prio class %d\n", ioprio_class);
3784 case IOPRIO_CLASS_NONE:
3786 * No prio set, inherit CPU scheduling settings.
3788 bfqq->new_ioprio = task_nice_ioprio(tsk);
3789 bfqq->new_ioprio_class = task_nice_ioclass(tsk);
3791 case IOPRIO_CLASS_RT:
3792 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
3793 bfqq->new_ioprio_class = IOPRIO_CLASS_RT;
3795 case IOPRIO_CLASS_BE:
3796 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
3797 bfqq->new_ioprio_class = IOPRIO_CLASS_BE;
3799 case IOPRIO_CLASS_IDLE:
3800 bfqq->new_ioprio_class = IOPRIO_CLASS_IDLE;
3801 bfqq->new_ioprio = 7;
3805 if (bfqq->new_ioprio >= IOPRIO_BE_NR) {
3806 pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n",
3808 bfqq->new_ioprio = IOPRIO_BE_NR;
3811 bfqq->entity.new_weight = bfq_ioprio_to_weight(bfqq->new_ioprio);
3812 bfqq->entity.prio_changed = 1;
3815 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
3816 struct bio *bio, bool is_sync,
3817 struct bfq_io_cq *bic);
3819 static void bfq_check_ioprio_change(struct bfq_io_cq *bic, struct bio *bio)
3821 struct bfq_data *bfqd = bic_to_bfqd(bic);
3822 struct bfq_queue *bfqq;
3823 int ioprio = bic->icq.ioc->ioprio;
3826 * This condition may trigger on a newly created bic, be sure to
3827 * drop the lock before returning.
3829 if (unlikely(!bfqd) || likely(bic->ioprio == ioprio))
3832 bic->ioprio = ioprio;
3834 bfqq = bic_to_bfqq(bic, false);
3836 /* release process reference on this queue */
3837 bfq_put_queue(bfqq);
3838 bfqq = bfq_get_queue(bfqd, bio, BLK_RW_ASYNC, bic);
3839 bic_set_bfqq(bic, bfqq, false);
3842 bfqq = bic_to_bfqq(bic, true);
3844 bfq_set_next_ioprio_data(bfqq, bic);
3847 static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
3848 struct bfq_io_cq *bic, pid_t pid, int is_sync)
3850 RB_CLEAR_NODE(&bfqq->entity.rb_node);
3851 INIT_LIST_HEAD(&bfqq->fifo);
3852 INIT_HLIST_NODE(&bfqq->burst_list_node);
3858 bfq_set_next_ioprio_data(bfqq, bic);
3862 * No need to mark as has_short_ttime if in
3863 * idle_class, because no device idling is performed
3864 * for queues in idle class
3866 if (!bfq_class_idle(bfqq))
3867 /* tentatively mark as has_short_ttime */
3868 bfq_mark_bfqq_has_short_ttime(bfqq);
3869 bfq_mark_bfqq_sync(bfqq);
3870 bfq_mark_bfqq_just_created(bfqq);
3872 bfq_clear_bfqq_sync(bfqq);
3874 /* set end request to minus infinity from now */
3875 bfqq->ttime.last_end_request = ktime_get_ns() + 1;
3877 bfq_mark_bfqq_IO_bound(bfqq);
3881 /* Tentative initial value to trade off between thr and lat */
3882 bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3;
3883 bfqq->budget_timeout = bfq_smallest_from_now();
3886 bfqq->last_wr_start_finish = jiffies;
3887 bfqq->wr_start_at_switch_to_srt = bfq_smallest_from_now();
3888 bfqq->split_time = bfq_smallest_from_now();
3891 * Set to the value for which bfqq will not be deemed as
3892 * soft rt when it becomes backlogged.
3894 bfqq->soft_rt_next_start = bfq_greatest_from_now();
3896 /* first request is almost certainly seeky */
3897 bfqq->seek_history = 1;
3900 static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd,
3901 struct bfq_group *bfqg,
3902 int ioprio_class, int ioprio)
3904 switch (ioprio_class) {
3905 case IOPRIO_CLASS_RT:
3906 return &bfqg->async_bfqq[0][ioprio];
3907 case IOPRIO_CLASS_NONE:
3908 ioprio = IOPRIO_NORM;
3910 case IOPRIO_CLASS_BE:
3911 return &bfqg->async_bfqq[1][ioprio];
3912 case IOPRIO_CLASS_IDLE:
3913 return &bfqg->async_idle_bfqq;
3919 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
3920 struct bio *bio, bool is_sync,
3921 struct bfq_io_cq *bic)
3923 const int ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
3924 const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
3925 struct bfq_queue **async_bfqq = NULL;
3926 struct bfq_queue *bfqq;
3927 struct bfq_group *bfqg;
3931 bfqg = bfq_find_set_group(bfqd, bio_blkcg(bio));
3933 bfqq = &bfqd->oom_bfqq;
3938 async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class,
3945 bfqq = kmem_cache_alloc_node(bfq_pool,
3946 GFP_NOWAIT | __GFP_ZERO | __GFP_NOWARN,
3950 bfq_init_bfqq(bfqd, bfqq, bic, current->pid,
3952 bfq_init_entity(&bfqq->entity, bfqg);
3953 bfq_log_bfqq(bfqd, bfqq, "allocated");
3955 bfqq = &bfqd->oom_bfqq;
3956 bfq_log_bfqq(bfqd, bfqq, "using oom bfqq");
3961 * Pin the queue now that it's allocated, scheduler exit will
3966 * Extra group reference, w.r.t. sync
3967 * queue. This extra reference is removed
3968 * only if bfqq->bfqg disappears, to
3969 * guarantee that this queue is not freed
3970 * until its group goes away.
3972 bfq_log_bfqq(bfqd, bfqq, "get_queue, bfqq not in async: %p, %d",
3978 bfqq->ref++; /* get a process reference to this queue */
3979 bfq_log_bfqq(bfqd, bfqq, "get_queue, at end: %p, %d", bfqq, bfqq->ref);
3984 static void bfq_update_io_thinktime(struct bfq_data *bfqd,
3985 struct bfq_queue *bfqq)
3987 struct bfq_ttime *ttime = &bfqq->ttime;
3988 u64 elapsed = ktime_get_ns() - bfqq->ttime.last_end_request;
3990 elapsed = min_t(u64, elapsed, 2ULL * bfqd->bfq_slice_idle);
3992 ttime->ttime_samples = (7*bfqq->ttime.ttime_samples + 256) / 8;
3993 ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed, 8);
3994 ttime->ttime_mean = div64_ul(ttime->ttime_total + 128,
3995 ttime->ttime_samples);
3999 bfq_update_io_seektime(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4002 bfqq->seek_history <<= 1;
4003 bfqq->seek_history |=
4004 get_sdist(bfqq->last_request_pos, rq) > BFQQ_SEEK_THR &&
4005 (!blk_queue_nonrot(bfqd->queue) ||
4006 blk_rq_sectors(rq) < BFQQ_SECT_THR_NONROT);
4009 static void bfq_update_has_short_ttime(struct bfq_data *bfqd,
4010 struct bfq_queue *bfqq,
4011 struct bfq_io_cq *bic)
4013 bool has_short_ttime = true;
4016 * No need to update has_short_ttime if bfqq is async or in
4017 * idle io prio class, or if bfq_slice_idle is zero, because
4018 * no device idling is performed for bfqq in this case.
4020 if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq) ||
4021 bfqd->bfq_slice_idle == 0)
4024 /* Idle window just restored, statistics are meaningless. */
4025 if (time_is_after_eq_jiffies(bfqq->split_time +
4026 bfqd->bfq_wr_min_idle_time))
4029 /* Think time is infinite if no process is linked to
4030 * bfqq. Otherwise check average think time to
4031 * decide whether to mark as has_short_ttime
4033 if (atomic_read(&bic->icq.ioc->active_ref) == 0 ||
4034 (bfq_sample_valid(bfqq->ttime.ttime_samples) &&
4035 bfqq->ttime.ttime_mean > bfqd->bfq_slice_idle))
4036 has_short_ttime = false;
4038 bfq_log_bfqq(bfqd, bfqq, "update_has_short_ttime: has_short_ttime %d",
4041 if (has_short_ttime)
4042 bfq_mark_bfqq_has_short_ttime(bfqq);
4044 bfq_clear_bfqq_has_short_ttime(bfqq);
4048 * Called when a new fs request (rq) is added to bfqq. Check if there's
4049 * something we should do about it.
4051 static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4054 struct bfq_io_cq *bic = RQ_BIC(rq);
4056 if (rq->cmd_flags & REQ_META)
4057 bfqq->meta_pending++;
4059 bfq_update_io_thinktime(bfqd, bfqq);
4060 bfq_update_has_short_ttime(bfqd, bfqq, bic);
4061 bfq_update_io_seektime(bfqd, bfqq, rq);
4063 bfq_log_bfqq(bfqd, bfqq,
4064 "rq_enqueued: has_short_ttime=%d (seeky %d)",
4065 bfq_bfqq_has_short_ttime(bfqq), BFQQ_SEEKY(bfqq));
4067 bfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
4069 if (bfqq == bfqd->in_service_queue && bfq_bfqq_wait_request(bfqq)) {
4070 bool small_req = bfqq->queued[rq_is_sync(rq)] == 1 &&
4071 blk_rq_sectors(rq) < 32;
4072 bool budget_timeout = bfq_bfqq_budget_timeout(bfqq);
4075 * There is just this request queued: if the request
4076 * is small and the queue is not to be expired, then
4079 * In this way, if the device is being idled to wait
4080 * for a new request from the in-service queue, we
4081 * avoid unplugging the device and committing the
4082 * device to serve just a small request. On the
4083 * contrary, we wait for the block layer to decide
4084 * when to unplug the device: hopefully, new requests
4085 * will be merged to this one quickly, then the device
4086 * will be unplugged and larger requests will be
4089 if (small_req && !budget_timeout)
4093 * A large enough request arrived, or the queue is to
4094 * be expired: in both cases disk idling is to be
4095 * stopped, so clear wait_request flag and reset
4098 bfq_clear_bfqq_wait_request(bfqq);
4099 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
4100 bfqg_stats_update_idle_time(bfqq_group(bfqq));
4103 * The queue is not empty, because a new request just
4104 * arrived. Hence we can safely expire the queue, in
4105 * case of budget timeout, without risking that the
4106 * timestamps of the queue are not updated correctly.
4107 * See [1] for more details.
4110 bfq_bfqq_expire(bfqd, bfqq, false,
4111 BFQQE_BUDGET_TIMEOUT);
4115 static void __bfq_insert_request(struct bfq_data *bfqd, struct request *rq)
4117 struct bfq_queue *bfqq = RQ_BFQQ(rq),
4118 *new_bfqq = bfq_setup_cooperator(bfqd, bfqq, rq, true);
4121 if (bic_to_bfqq(RQ_BIC(rq), 1) != bfqq)
4122 new_bfqq = bic_to_bfqq(RQ_BIC(rq), 1);
4124 * Release the request's reference to the old bfqq
4125 * and make sure one is taken to the shared queue.
4127 new_bfqq->allocated++;
4130 bfq_clear_bfqq_just_created(bfqq);
4132 * If the bic associated with the process
4133 * issuing this request still points to bfqq
4134 * (and thus has not been already redirected
4135 * to new_bfqq or even some other bfq_queue),
4136 * then complete the merge and redirect it to
4139 if (bic_to_bfqq(RQ_BIC(rq), 1) == bfqq)
4140 bfq_merge_bfqqs(bfqd, RQ_BIC(rq),
4143 * rq is about to be enqueued into new_bfqq,
4144 * release rq reference on bfqq
4146 bfq_put_queue(bfqq);
4147 rq->elv.priv[1] = new_bfqq;
4151 bfq_add_request(rq);
4153 rq->fifo_time = ktime_get_ns() + bfqd->bfq_fifo_expire[rq_is_sync(rq)];
4154 list_add_tail(&rq->queuelist, &bfqq->fifo);
4156 bfq_rq_enqueued(bfqd, bfqq, rq);
4159 static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
4162 struct request_queue *q = hctx->queue;
4163 struct bfq_data *bfqd = q->elevator->elevator_data;
4165 spin_lock_irq(&bfqd->lock);
4166 if (blk_mq_sched_try_insert_merge(q, rq)) {
4167 spin_unlock_irq(&bfqd->lock);
4171 spin_unlock_irq(&bfqd->lock);
4173 blk_mq_sched_request_inserted(rq);
4175 spin_lock_irq(&bfqd->lock);
4176 if (at_head || blk_rq_is_passthrough(rq)) {
4178 list_add(&rq->queuelist, &bfqd->dispatch);
4180 list_add_tail(&rq->queuelist, &bfqd->dispatch);
4182 __bfq_insert_request(bfqd, rq);
4184 if (rq_mergeable(rq)) {
4185 elv_rqhash_add(q, rq);
4191 spin_unlock_irq(&bfqd->lock);
4194 static void bfq_insert_requests(struct blk_mq_hw_ctx *hctx,
4195 struct list_head *list, bool at_head)
4197 while (!list_empty(list)) {
4200 rq = list_first_entry(list, struct request, queuelist);
4201 list_del_init(&rq->queuelist);
4202 bfq_insert_request(hctx, rq, at_head);
4206 static void bfq_update_hw_tag(struct bfq_data *bfqd)
4208 bfqd->max_rq_in_driver = max_t(int, bfqd->max_rq_in_driver,
4209 bfqd->rq_in_driver);
4211 if (bfqd->hw_tag == 1)
4215 * This sample is valid if the number of outstanding requests
4216 * is large enough to allow a queueing behavior. Note that the
4217 * sum is not exact, as it's not taking into account deactivated
4220 if (bfqd->rq_in_driver + bfqd->queued < BFQ_HW_QUEUE_THRESHOLD)
4223 if (bfqd->hw_tag_samples++ < BFQ_HW_QUEUE_SAMPLES)
4226 bfqd->hw_tag = bfqd->max_rq_in_driver > BFQ_HW_QUEUE_THRESHOLD;
4227 bfqd->max_rq_in_driver = 0;
4228 bfqd->hw_tag_samples = 0;
4231 static void bfq_completed_request(struct bfq_queue *bfqq, struct bfq_data *bfqd)
4236 bfq_update_hw_tag(bfqd);
4238 bfqd->rq_in_driver--;
4241 if (!bfqq->dispatched && !bfq_bfqq_busy(bfqq)) {
4243 * Set budget_timeout (which we overload to store the
4244 * time at which the queue remains with no backlog and
4245 * no outstanding request; used by the weight-raising
4248 bfqq->budget_timeout = jiffies;
4250 bfq_weights_tree_remove(bfqd, &bfqq->entity,
4251 &bfqd->queue_weights_tree);
4254 now_ns = ktime_get_ns();
4256 bfqq->ttime.last_end_request = now_ns;
4259 * Using us instead of ns, to get a reasonable precision in
4260 * computing rate in next check.
4262 delta_us = div_u64(now_ns - bfqd->last_completion, NSEC_PER_USEC);
4265 * If the request took rather long to complete, and, according
4266 * to the maximum request size recorded, this completion latency
4267 * implies that the request was certainly served at a very low
4268 * rate (less than 1M sectors/sec), then the whole observation
4269 * interval that lasts up to this time instant cannot be a
4270 * valid time interval for computing a new peak rate. Invoke
4271 * bfq_update_rate_reset to have the following three steps
4273 * - close the observation interval at the last (previous)
4274 * request dispatch or completion
4275 * - compute rate, if possible, for that observation interval
4276 * - reset to zero samples, which will trigger a proper
4277 * re-initialization of the observation interval on next
4280 if (delta_us > BFQ_MIN_TT/NSEC_PER_USEC &&
4281 (bfqd->last_rq_max_size<<BFQ_RATE_SHIFT)/delta_us <
4282 1UL<<(BFQ_RATE_SHIFT - 10))
4283 bfq_update_rate_reset(bfqd, NULL);
4284 bfqd->last_completion = now_ns;
4287 * If we are waiting to discover whether the request pattern
4288 * of the task associated with the queue is actually
4289 * isochronous, and both requisites for this condition to hold
4290 * are now satisfied, then compute soft_rt_next_start (see the
4291 * comments on the function bfq_bfqq_softrt_next_start()). We
4292 * schedule this delayed check when bfqq expires, if it still
4293 * has in-flight requests.
4295 if (bfq_bfqq_softrt_update(bfqq) && bfqq->dispatched == 0 &&
4296 RB_EMPTY_ROOT(&bfqq->sort_list))
4297 bfqq->soft_rt_next_start =
4298 bfq_bfqq_softrt_next_start(bfqd, bfqq);
4301 * If this is the in-service queue, check if it needs to be expired,
4302 * or if we want to idle in case it has no pending requests.
4304 if (bfqd->in_service_queue == bfqq) {
4305 if (bfqq->dispatched == 0 && bfq_bfqq_must_idle(bfqq)) {
4306 bfq_arm_slice_timer(bfqd);
4308 } else if (bfq_may_expire_for_budg_timeout(bfqq))
4309 bfq_bfqq_expire(bfqd, bfqq, false,
4310 BFQQE_BUDGET_TIMEOUT);
4311 else if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
4312 (bfqq->dispatched == 0 ||
4313 !bfq_bfqq_may_idle(bfqq)))
4314 bfq_bfqq_expire(bfqd, bfqq, false,
4315 BFQQE_NO_MORE_REQUESTS);
4318 if (!bfqd->rq_in_driver)
4319 bfq_schedule_dispatch(bfqd);
4322 static void bfq_put_rq_priv_body(struct bfq_queue *bfqq)
4326 bfq_put_queue(bfqq);
4329 static void bfq_finish_request(struct request *rq)
4331 struct bfq_queue *bfqq;
4332 struct bfq_data *bfqd;
4340 if (rq->rq_flags & RQF_STARTED)
4341 bfqg_stats_update_completion(bfqq_group(bfqq),
4342 rq_start_time_ns(rq),
4343 rq_io_start_time_ns(rq),
4346 if (likely(rq->rq_flags & RQF_STARTED)) {
4347 unsigned long flags;
4349 spin_lock_irqsave(&bfqd->lock, flags);
4351 bfq_completed_request(bfqq, bfqd);
4352 bfq_put_rq_priv_body(bfqq);
4354 spin_unlock_irqrestore(&bfqd->lock, flags);
4357 * Request rq may be still/already in the scheduler,
4358 * in which case we need to remove it. And we cannot
4359 * defer such a check and removal, to avoid
4360 * inconsistencies in the time interval from the end
4361 * of this function to the start of the deferred work.
4362 * This situation seems to occur only in process
4363 * context, as a consequence of a merge. In the
4364 * current version of the code, this implies that the
4368 if (!RB_EMPTY_NODE(&rq->rb_node))
4369 bfq_remove_request(rq->q, rq);
4370 bfq_put_rq_priv_body(bfqq);
4373 rq->elv.priv[0] = NULL;
4374 rq->elv.priv[1] = NULL;
4378 * Returns NULL if a new bfqq should be allocated, or the old bfqq if this
4379 * was the last process referring to that bfqq.
4381 static struct bfq_queue *
4382 bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq)
4384 bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue");
4386 if (bfqq_process_refs(bfqq) == 1) {
4387 bfqq->pid = current->pid;
4388 bfq_clear_bfqq_coop(bfqq);
4389 bfq_clear_bfqq_split_coop(bfqq);
4393 bic_set_bfqq(bic, NULL, 1);
4395 bfq_put_cooperator(bfqq);
4397 bfq_put_queue(bfqq);
4401 static struct bfq_queue *bfq_get_bfqq_handle_split(struct bfq_data *bfqd,
4402 struct bfq_io_cq *bic,
4404 bool split, bool is_sync,
4407 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
4409 if (likely(bfqq && bfqq != &bfqd->oom_bfqq))
4416 bfq_put_queue(bfqq);
4417 bfqq = bfq_get_queue(bfqd, bio, is_sync, bic);
4419 bic_set_bfqq(bic, bfqq, is_sync);
4420 if (split && is_sync) {
4421 if ((bic->was_in_burst_list && bfqd->large_burst) ||
4422 bic->saved_in_large_burst)
4423 bfq_mark_bfqq_in_large_burst(bfqq);
4425 bfq_clear_bfqq_in_large_burst(bfqq);
4426 if (bic->was_in_burst_list)
4427 hlist_add_head(&bfqq->burst_list_node,
4430 bfqq->split_time = jiffies;
4437 * Allocate bfq data structures associated with this request.
4439 static void bfq_prepare_request(struct request *rq, struct bio *bio)
4441 struct request_queue *q = rq->q;
4442 struct bfq_data *bfqd = q->elevator->elevator_data;
4443 struct bfq_io_cq *bic;
4444 const int is_sync = rq_is_sync(rq);
4445 struct bfq_queue *bfqq;
4446 bool new_queue = false;
4447 bool bfqq_already_existing = false, split = false;
4451 bic = icq_to_bic(rq->elv.icq);
4453 spin_lock_irq(&bfqd->lock);
4455 bfq_check_ioprio_change(bic, bio);
4457 bfq_bic_update_cgroup(bic, bio);
4459 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio, false, is_sync,
4462 if (likely(!new_queue)) {
4463 /* If the queue was seeky for too long, break it apart. */
4464 if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq)) {
4465 bfq_log_bfqq(bfqd, bfqq, "breaking apart bfqq");
4467 /* Update bic before losing reference to bfqq */
4468 if (bfq_bfqq_in_large_burst(bfqq))
4469 bic->saved_in_large_burst = true;
4471 bfqq = bfq_split_bfqq(bic, bfqq);
4475 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio,
4479 bfqq_already_existing = true;
4485 bfq_log_bfqq(bfqd, bfqq, "get_request %p: bfqq %p, %d",
4486 rq, bfqq, bfqq->ref);
4488 rq->elv.priv[0] = bic;
4489 rq->elv.priv[1] = bfqq;
4492 * If a bfq_queue has only one process reference, it is owned
4493 * by only this bic: we can then set bfqq->bic = bic. in
4494 * addition, if the queue has also just been split, we have to
4497 if (likely(bfqq != &bfqd->oom_bfqq) && bfqq_process_refs(bfqq) == 1) {
4501 * The queue has just been split from a shared
4502 * queue: restore the idle window and the
4503 * possible weight raising period.
4505 bfq_bfqq_resume_state(bfqq, bfqd, bic,
4506 bfqq_already_existing);
4510 if (unlikely(bfq_bfqq_just_created(bfqq)))
4511 bfq_handle_burst(bfqd, bfqq);
4513 spin_unlock_irq(&bfqd->lock);
4516 static void bfq_idle_slice_timer_body(struct bfq_queue *bfqq)
4518 struct bfq_data *bfqd = bfqq->bfqd;
4519 enum bfqq_expiration reason;
4520 unsigned long flags;
4522 spin_lock_irqsave(&bfqd->lock, flags);
4523 bfq_clear_bfqq_wait_request(bfqq);
4525 if (bfqq != bfqd->in_service_queue) {
4526 spin_unlock_irqrestore(&bfqd->lock, flags);
4530 if (bfq_bfqq_budget_timeout(bfqq))
4532 * Also here the queue can be safely expired
4533 * for budget timeout without wasting
4536 reason = BFQQE_BUDGET_TIMEOUT;
4537 else if (bfqq->queued[0] == 0 && bfqq->queued[1] == 0)
4539 * The queue may not be empty upon timer expiration,
4540 * because we may not disable the timer when the
4541 * first request of the in-service queue arrives
4542 * during disk idling.
4544 reason = BFQQE_TOO_IDLE;
4546 goto schedule_dispatch;
4548 bfq_bfqq_expire(bfqd, bfqq, true, reason);
4551 spin_unlock_irqrestore(&bfqd->lock, flags);
4552 bfq_schedule_dispatch(bfqd);
4556 * Handler of the expiration of the timer running if the in-service queue
4557 * is idling inside its time slice.
4559 static enum hrtimer_restart bfq_idle_slice_timer(struct hrtimer *timer)
4561 struct bfq_data *bfqd = container_of(timer, struct bfq_data,
4563 struct bfq_queue *bfqq = bfqd->in_service_queue;
4566 * Theoretical race here: the in-service queue can be NULL or
4567 * different from the queue that was idling if a new request
4568 * arrives for the current queue and there is a full dispatch
4569 * cycle that changes the in-service queue. This can hardly
4570 * happen, but in the worst case we just expire a queue too
4574 bfq_idle_slice_timer_body(bfqq);
4576 return HRTIMER_NORESTART;
4579 static void __bfq_put_async_bfqq(struct bfq_data *bfqd,
4580 struct bfq_queue **bfqq_ptr)
4582 struct bfq_queue *bfqq = *bfqq_ptr;
4584 bfq_log(bfqd, "put_async_bfqq: %p", bfqq);
4586 bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
4588 bfq_log_bfqq(bfqd, bfqq, "put_async_bfqq: putting %p, %d",
4590 bfq_put_queue(bfqq);
4596 * Release all the bfqg references to its async queues. If we are
4597 * deallocating the group these queues may still contain requests, so
4598 * we reparent them to the root cgroup (i.e., the only one that will
4599 * exist for sure until all the requests on a device are gone).
4601 void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg)
4605 for (i = 0; i < 2; i++)
4606 for (j = 0; j < IOPRIO_BE_NR; j++)
4607 __bfq_put_async_bfqq(bfqd, &bfqg->async_bfqq[i][j]);
4609 __bfq_put_async_bfqq(bfqd, &bfqg->async_idle_bfqq);
4612 static void bfq_exit_queue(struct elevator_queue *e)
4614 struct bfq_data *bfqd = e->elevator_data;
4615 struct bfq_queue *bfqq, *n;
4617 hrtimer_cancel(&bfqd->idle_slice_timer);
4619 spin_lock_irq(&bfqd->lock);
4620 list_for_each_entry_safe(bfqq, n, &bfqd->idle_list, bfqq_list)
4621 bfq_deactivate_bfqq(bfqd, bfqq, false, false);
4622 spin_unlock_irq(&bfqd->lock);
4624 hrtimer_cancel(&bfqd->idle_slice_timer);
4626 #ifdef CONFIG_BFQ_GROUP_IOSCHED
4627 blkcg_deactivate_policy(bfqd->queue, &blkcg_policy_bfq);
4629 spin_lock_irq(&bfqd->lock);
4630 bfq_put_async_queues(bfqd, bfqd->root_group);
4631 kfree(bfqd->root_group);
4632 spin_unlock_irq(&bfqd->lock);
4638 static void bfq_init_root_group(struct bfq_group *root_group,
4639 struct bfq_data *bfqd)
4643 #ifdef CONFIG_BFQ_GROUP_IOSCHED
4644 root_group->entity.parent = NULL;
4645 root_group->my_entity = NULL;
4646 root_group->bfqd = bfqd;
4648 root_group->rq_pos_tree = RB_ROOT;
4649 for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
4650 root_group->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
4651 root_group->sched_data.bfq_class_idle_last_service = jiffies;
4654 static int bfq_init_queue(struct request_queue *q, struct elevator_type *e)
4656 struct bfq_data *bfqd;
4657 struct elevator_queue *eq;
4659 eq = elevator_alloc(q, e);
4663 bfqd = kzalloc_node(sizeof(*bfqd), GFP_KERNEL, q->node);
4665 kobject_put(&eq->kobj);
4668 eq->elevator_data = bfqd;
4670 spin_lock_irq(q->queue_lock);
4672 spin_unlock_irq(q->queue_lock);
4675 * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues.
4676 * Grab a permanent reference to it, so that the normal code flow
4677 * will not attempt to free it.
4679 bfq_init_bfqq(bfqd, &bfqd->oom_bfqq, NULL, 1, 0);
4680 bfqd->oom_bfqq.ref++;
4681 bfqd->oom_bfqq.new_ioprio = BFQ_DEFAULT_QUEUE_IOPRIO;
4682 bfqd->oom_bfqq.new_ioprio_class = IOPRIO_CLASS_BE;
4683 bfqd->oom_bfqq.entity.new_weight =
4684 bfq_ioprio_to_weight(bfqd->oom_bfqq.new_ioprio);
4686 /* oom_bfqq does not participate to bursts */
4687 bfq_clear_bfqq_just_created(&bfqd->oom_bfqq);
4690 * Trigger weight initialization, according to ioprio, at the
4691 * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio
4692 * class won't be changed any more.
4694 bfqd->oom_bfqq.entity.prio_changed = 1;
4698 INIT_LIST_HEAD(&bfqd->dispatch);
4700 hrtimer_init(&bfqd->idle_slice_timer, CLOCK_MONOTONIC,
4702 bfqd->idle_slice_timer.function = bfq_idle_slice_timer;
4704 bfqd->queue_weights_tree = RB_ROOT;
4705 bfqd->group_weights_tree = RB_ROOT;
4707 INIT_LIST_HEAD(&bfqd->active_list);
4708 INIT_LIST_HEAD(&bfqd->idle_list);
4709 INIT_HLIST_HEAD(&bfqd->burst_list);
4713 bfqd->bfq_max_budget = bfq_default_max_budget;
4715 bfqd->bfq_fifo_expire[0] = bfq_fifo_expire[0];
4716 bfqd->bfq_fifo_expire[1] = bfq_fifo_expire[1];
4717 bfqd->bfq_back_max = bfq_back_max;
4718 bfqd->bfq_back_penalty = bfq_back_penalty;
4719 bfqd->bfq_slice_idle = bfq_slice_idle;
4720 bfqd->bfq_timeout = bfq_timeout;
4722 bfqd->bfq_requests_within_timer = 120;
4724 bfqd->bfq_large_burst_thresh = 8;
4725 bfqd->bfq_burst_interval = msecs_to_jiffies(180);
4727 bfqd->low_latency = true;
4730 * Trade-off between responsiveness and fairness.
4732 bfqd->bfq_wr_coeff = 30;
4733 bfqd->bfq_wr_rt_max_time = msecs_to_jiffies(300);
4734 bfqd->bfq_wr_max_time = 0;
4735 bfqd->bfq_wr_min_idle_time = msecs_to_jiffies(2000);
4736 bfqd->bfq_wr_min_inter_arr_async = msecs_to_jiffies(500);
4737 bfqd->bfq_wr_max_softrt_rate = 7000; /*
4738 * Approximate rate required
4739 * to playback or record a
4740 * high-definition compressed
4743 bfqd->wr_busy_queues = 0;
4746 * Begin by assuming, optimistically, that the device is a
4747 * high-speed one, and that its peak rate is equal to 2/3 of
4748 * the highest reference rate.
4750 bfqd->RT_prod = R_fast[blk_queue_nonrot(bfqd->queue)] *
4751 T_fast[blk_queue_nonrot(bfqd->queue)];
4752 bfqd->peak_rate = R_fast[blk_queue_nonrot(bfqd->queue)] * 2 / 3;
4753 bfqd->device_speed = BFQ_BFQD_FAST;
4755 spin_lock_init(&bfqd->lock);
4758 * The invocation of the next bfq_create_group_hierarchy
4759 * function is the head of a chain of function calls
4760 * (bfq_create_group_hierarchy->blkcg_activate_policy->
4761 * blk_mq_freeze_queue) that may lead to the invocation of the
4762 * has_work hook function. For this reason,
4763 * bfq_create_group_hierarchy is invoked only after all
4764 * scheduler data has been initialized, apart from the fields
4765 * that can be initialized only after invoking
4766 * bfq_create_group_hierarchy. This, in particular, enables
4767 * has_work to correctly return false. Of course, to avoid
4768 * other inconsistencies, the blk-mq stack must then refrain
4769 * from invoking further scheduler hooks before this init
4770 * function is finished.
4772 bfqd->root_group = bfq_create_group_hierarchy(bfqd, q->node);
4773 if (!bfqd->root_group)
4775 bfq_init_root_group(bfqd->root_group, bfqd);
4776 bfq_init_entity(&bfqd->oom_bfqq.entity, bfqd->root_group);
4783 kobject_put(&eq->kobj);
4787 static void bfq_slab_kill(void)
4789 kmem_cache_destroy(bfq_pool);
4792 static int __init bfq_slab_setup(void)
4794 bfq_pool = KMEM_CACHE(bfq_queue, 0);
4800 static ssize_t bfq_var_show(unsigned int var, char *page)
4802 return sprintf(page, "%u\n", var);
4805 static int bfq_var_store(unsigned long *var, const char *page)
4807 unsigned long new_val;
4808 int ret = kstrtoul(page, 10, &new_val);
4816 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
4817 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
4819 struct bfq_data *bfqd = e->elevator_data; \
4820 u64 __data = __VAR; \
4822 __data = jiffies_to_msecs(__data); \
4823 else if (__CONV == 2) \
4824 __data = div_u64(__data, NSEC_PER_MSEC); \
4825 return bfq_var_show(__data, (page)); \
4827 SHOW_FUNCTION(bfq_fifo_expire_sync_show, bfqd->bfq_fifo_expire[1], 2);
4828 SHOW_FUNCTION(bfq_fifo_expire_async_show, bfqd->bfq_fifo_expire[0], 2);
4829 SHOW_FUNCTION(bfq_back_seek_max_show, bfqd->bfq_back_max, 0);
4830 SHOW_FUNCTION(bfq_back_seek_penalty_show, bfqd->bfq_back_penalty, 0);
4831 SHOW_FUNCTION(bfq_slice_idle_show, bfqd->bfq_slice_idle, 2);
4832 SHOW_FUNCTION(bfq_max_budget_show, bfqd->bfq_user_max_budget, 0);
4833 SHOW_FUNCTION(bfq_timeout_sync_show, bfqd->bfq_timeout, 1);
4834 SHOW_FUNCTION(bfq_strict_guarantees_show, bfqd->strict_guarantees, 0);
4835 SHOW_FUNCTION(bfq_low_latency_show, bfqd->low_latency, 0);
4836 #undef SHOW_FUNCTION
4838 #define USEC_SHOW_FUNCTION(__FUNC, __VAR) \
4839 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
4841 struct bfq_data *bfqd = e->elevator_data; \
4842 u64 __data = __VAR; \
4843 __data = div_u64(__data, NSEC_PER_USEC); \
4844 return bfq_var_show(__data, (page)); \
4846 USEC_SHOW_FUNCTION(bfq_slice_idle_us_show, bfqd->bfq_slice_idle);
4847 #undef USEC_SHOW_FUNCTION
4849 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
4851 __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4853 struct bfq_data *bfqd = e->elevator_data; \
4854 unsigned long __data, __min = (MIN), __max = (MAX); \
4857 ret = bfq_var_store(&__data, (page)); \
4860 if (__data < __min) \
4862 else if (__data > __max) \
4865 *(__PTR) = msecs_to_jiffies(__data); \
4866 else if (__CONV == 2) \
4867 *(__PTR) = (u64)__data * NSEC_PER_MSEC; \
4869 *(__PTR) = __data; \
4872 STORE_FUNCTION(bfq_fifo_expire_sync_store, &bfqd->bfq_fifo_expire[1], 1,
4874 STORE_FUNCTION(bfq_fifo_expire_async_store, &bfqd->bfq_fifo_expire[0], 1,
4876 STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0);
4877 STORE_FUNCTION(bfq_back_seek_penalty_store, &bfqd->bfq_back_penalty, 1,
4879 STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 2);
4880 #undef STORE_FUNCTION
4882 #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
4883 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\
4885 struct bfq_data *bfqd = e->elevator_data; \
4886 unsigned long __data, __min = (MIN), __max = (MAX); \
4889 ret = bfq_var_store(&__data, (page)); \
4892 if (__data < __min) \
4894 else if (__data > __max) \
4896 *(__PTR) = (u64)__data * NSEC_PER_USEC; \
4899 USEC_STORE_FUNCTION(bfq_slice_idle_us_store, &bfqd->bfq_slice_idle, 0,
4901 #undef USEC_STORE_FUNCTION
4903 static ssize_t bfq_max_budget_store(struct elevator_queue *e,
4904 const char *page, size_t count)
4906 struct bfq_data *bfqd = e->elevator_data;
4907 unsigned long __data;
4910 ret = bfq_var_store(&__data, (page));
4915 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
4917 if (__data > INT_MAX)
4919 bfqd->bfq_max_budget = __data;
4922 bfqd->bfq_user_max_budget = __data;
4928 * Leaving this name to preserve name compatibility with cfq
4929 * parameters, but this timeout is used for both sync and async.
4931 static ssize_t bfq_timeout_sync_store(struct elevator_queue *e,
4932 const char *page, size_t count)
4934 struct bfq_data *bfqd = e->elevator_data;
4935 unsigned long __data;
4938 ret = bfq_var_store(&__data, (page));
4944 else if (__data > INT_MAX)
4947 bfqd->bfq_timeout = msecs_to_jiffies(__data);
4948 if (bfqd->bfq_user_max_budget == 0)
4949 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
4954 static ssize_t bfq_strict_guarantees_store(struct elevator_queue *e,
4955 const char *page, size_t count)
4957 struct bfq_data *bfqd = e->elevator_data;
4958 unsigned long __data;
4961 ret = bfq_var_store(&__data, (page));
4967 if (!bfqd->strict_guarantees && __data == 1
4968 && bfqd->bfq_slice_idle < 8 * NSEC_PER_MSEC)
4969 bfqd->bfq_slice_idle = 8 * NSEC_PER_MSEC;
4971 bfqd->strict_guarantees = __data;
4976 static ssize_t bfq_low_latency_store(struct elevator_queue *e,
4977 const char *page, size_t count)
4979 struct bfq_data *bfqd = e->elevator_data;
4980 unsigned long __data;
4983 ret = bfq_var_store(&__data, (page));
4989 if (__data == 0 && bfqd->low_latency != 0)
4991 bfqd->low_latency = __data;
4996 #define BFQ_ATTR(name) \
4997 __ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store)
4999 static struct elv_fs_entry bfq_attrs[] = {
5000 BFQ_ATTR(fifo_expire_sync),
5001 BFQ_ATTR(fifo_expire_async),
5002 BFQ_ATTR(back_seek_max),
5003 BFQ_ATTR(back_seek_penalty),
5004 BFQ_ATTR(slice_idle),
5005 BFQ_ATTR(slice_idle_us),
5006 BFQ_ATTR(max_budget),
5007 BFQ_ATTR(timeout_sync),
5008 BFQ_ATTR(strict_guarantees),
5009 BFQ_ATTR(low_latency),
5013 static struct elevator_type iosched_bfq_mq = {
5015 .prepare_request = bfq_prepare_request,
5016 .finish_request = bfq_finish_request,
5017 .exit_icq = bfq_exit_icq,
5018 .insert_requests = bfq_insert_requests,
5019 .dispatch_request = bfq_dispatch_request,
5020 .next_request = elv_rb_latter_request,
5021 .former_request = elv_rb_former_request,
5022 .allow_merge = bfq_allow_bio_merge,
5023 .bio_merge = bfq_bio_merge,
5024 .request_merge = bfq_request_merge,
5025 .requests_merged = bfq_requests_merged,
5026 .request_merged = bfq_request_merged,
5027 .has_work = bfq_has_work,
5028 .init_sched = bfq_init_queue,
5029 .exit_sched = bfq_exit_queue,
5033 .icq_size = sizeof(struct bfq_io_cq),
5034 .icq_align = __alignof__(struct bfq_io_cq),
5035 .elevator_attrs = bfq_attrs,
5036 .elevator_name = "bfq",
5037 .elevator_owner = THIS_MODULE,
5039 MODULE_ALIAS("bfq-iosched");
5041 static int __init bfq_init(void)
5045 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5046 ret = blkcg_policy_register(&blkcg_policy_bfq);
5052 if (bfq_slab_setup())
5056 * Times to load large popular applications for the typical
5057 * systems installed on the reference devices (see the
5058 * comments before the definitions of the next two
5059 * arrays). Actually, we use slightly slower values, as the
5060 * estimated peak rate tends to be smaller than the actual
5061 * peak rate. The reason for this last fact is that estimates
5062 * are computed over much shorter time intervals than the long
5063 * intervals typically used for benchmarking. Why? First, to
5064 * adapt more quickly to variations. Second, because an I/O
5065 * scheduler cannot rely on a peak-rate-evaluation workload to
5066 * be run for a long time.
5068 T_slow[0] = msecs_to_jiffies(3500); /* actually 4 sec */
5069 T_slow[1] = msecs_to_jiffies(6000); /* actually 6.5 sec */
5070 T_fast[0] = msecs_to_jiffies(7000); /* actually 8 sec */
5071 T_fast[1] = msecs_to_jiffies(2500); /* actually 3 sec */
5074 * Thresholds that determine the switch between speed classes
5075 * (see the comments before the definition of the array
5076 * device_speed_thresh). These thresholds are biased towards
5077 * transitions to the fast class. This is safer than the
5078 * opposite bias. In fact, a wrong transition to the slow
5079 * class results in short weight-raising periods, because the
5080 * speed of the device then tends to be higher that the
5081 * reference peak rate. On the opposite end, a wrong
5082 * transition to the fast class tends to increase
5083 * weight-raising periods, because of the opposite reason.
5085 device_speed_thresh[0] = (4 * R_slow[0]) / 3;
5086 device_speed_thresh[1] = (4 * R_slow[1]) / 3;
5088 ret = elv_register(&iosched_bfq_mq);
5097 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5098 blkcg_policy_unregister(&blkcg_policy_bfq);
5103 static void __exit bfq_exit(void)
5105 elv_unregister(&iosched_bfq_mq);
5106 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5107 blkcg_policy_unregister(&blkcg_policy_bfq);
5112 module_init(bfq_init);
5113 module_exit(bfq_exit);
5115 MODULE_AUTHOR("Paolo Valente");
5116 MODULE_LICENSE("GPL");
5117 MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler");