2 * Budget Fair Queueing (BFQ) I/O scheduler.
4 * Based on ideas and code from CFQ:
5 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
7 * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
8 * Paolo Valente <paolo.valente@unimore.it>
10 * Copyright (C) 2010 Paolo Valente <paolo.valente@unimore.it>
11 * Arianna Avanzini <avanzini@google.com>
13 * Copyright (C) 2017 Paolo Valente <paolo.valente@linaro.org>
15 * This program is free software; you can redistribute it and/or
16 * modify it under the terms of the GNU General Public License as
17 * published by the Free Software Foundation; either version 2 of the
18 * License, or (at your option) any later version.
20 * This program is distributed in the hope that it will be useful,
21 * but WITHOUT ANY WARRANTY; without even the implied warranty of
22 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
23 * General Public License for more details.
25 * BFQ is a proportional-share I/O scheduler, with some extra
26 * low-latency capabilities. BFQ also supports full hierarchical
27 * scheduling through cgroups. Next paragraphs provide an introduction
28 * on BFQ inner workings. Details on BFQ benefits, usage and
29 * limitations can be found in Documentation/block/bfq-iosched.txt.
31 * BFQ is a proportional-share storage-I/O scheduling algorithm based
32 * on the slice-by-slice service scheme of CFQ. But BFQ assigns
33 * budgets, measured in number of sectors, to processes instead of
34 * time slices. The device is not granted to the in-service process
35 * for a given time slice, but until it has exhausted its assigned
36 * budget. This change from the time to the service domain enables BFQ
37 * to distribute the device throughput among processes as desired,
38 * without any distortion due to throughput fluctuations, or to device
39 * internal queueing. BFQ uses an ad hoc internal scheduler, called
40 * B-WF2Q+, to schedule processes according to their budgets. More
41 * precisely, BFQ schedules queues associated with processes. Each
42 * process/queue is assigned a user-configurable weight, and B-WF2Q+
43 * guarantees that each queue receives a fraction of the throughput
44 * proportional to its weight. Thanks to the accurate policy of
45 * B-WF2Q+, BFQ can afford to assign high budgets to I/O-bound
46 * processes issuing sequential requests (to boost the throughput),
47 * and yet guarantee a low latency to interactive and soft real-time
50 * In particular, to provide these low-latency guarantees, BFQ
51 * explicitly privileges the I/O of two classes of time-sensitive
52 * applications: interactive and soft real-time. This feature enables
53 * BFQ to provide applications in these classes with a very low
54 * latency. Finally, BFQ also features additional heuristics for
55 * preserving both a low latency and a high throughput on NCQ-capable,
56 * rotational or flash-based devices, and to get the job done quickly
57 * for applications consisting in many I/O-bound processes.
59 * NOTE: if the main or only goal, with a given device, is to achieve
60 * the maximum-possible throughput at all times, then do switch off
61 * all low-latency heuristics for that device, by setting low_latency
64 * BFQ is described in [1], where also a reference to the initial, more
65 * theoretical paper on BFQ can be found. The interested reader can find
66 * in the latter paper full details on the main algorithm, as well as
67 * formulas of the guarantees and formal proofs of all the properties.
68 * With respect to the version of BFQ presented in these papers, this
69 * implementation adds a few more heuristics, such as the one that
70 * guarantees a low latency to soft real-time applications, and a
71 * hierarchical extension based on H-WF2Q+.
73 * B-WF2Q+ is based on WF2Q+, which is described in [2], together with
74 * H-WF2Q+, while the augmented tree used here to implement B-WF2Q+
75 * with O(log N) complexity derives from the one introduced with EEVDF
78 * [1] P. Valente, A. Avanzini, "Evolution of the BFQ Storage I/O
79 * Scheduler", Proceedings of the First Workshop on Mobile System
80 * Technologies (MST-2015), May 2015.
81 * http://algogroup.unimore.it/people/paolo/disk_sched/mst-2015.pdf
83 * [2] Jon C.R. Bennett and H. Zhang, "Hierarchical Packet Fair Queueing
84 * Algorithms", IEEE/ACM Transactions on Networking, 5(5):675-689,
87 * http://www.cs.cmu.edu/~hzhang/papers/TON-97-Oct.ps.gz
89 * [3] I. Stoica and H. Abdel-Wahab, "Earliest Eligible Virtual Deadline
90 * First: A Flexible and Accurate Mechanism for Proportional Share
91 * Resource Allocation", technical report.
93 * http://www.cs.berkeley.edu/~istoica/papers/eevdf-tr-95.pdf
95 #include <linux/module.h>
96 #include <linux/slab.h>
97 #include <linux/blkdev.h>
98 #include <linux/cgroup.h>
99 #include <linux/elevator.h>
100 #include <linux/ktime.h>
101 #include <linux/rbtree.h>
102 #include <linux/ioprio.h>
103 #include <linux/sbitmap.h>
104 #include <linux/delay.h>
108 #include "blk-mq-tag.h"
109 #include "blk-mq-sched.h"
110 #include "bfq-iosched.h"
113 #define BFQ_BFQQ_FNS(name) \
114 void bfq_mark_bfqq_##name(struct bfq_queue *bfqq) \
116 __set_bit(BFQQF_##name, &(bfqq)->flags); \
118 void bfq_clear_bfqq_##name(struct bfq_queue *bfqq) \
120 __clear_bit(BFQQF_##name, &(bfqq)->flags); \
122 int bfq_bfqq_##name(const struct bfq_queue *bfqq) \
124 return test_bit(BFQQF_##name, &(bfqq)->flags); \
127 BFQ_BFQQ_FNS(just_created
);
129 BFQ_BFQQ_FNS(wait_request
);
130 BFQ_BFQQ_FNS(non_blocking_wait_rq
);
131 BFQ_BFQQ_FNS(fifo_expire
);
132 BFQ_BFQQ_FNS(has_short_ttime
);
134 BFQ_BFQQ_FNS(IO_bound
);
135 BFQ_BFQQ_FNS(in_large_burst
);
137 BFQ_BFQQ_FNS(split_coop
);
138 BFQ_BFQQ_FNS(softrt_update
);
139 #undef BFQ_BFQQ_FNS \
141 /* Expiration time of sync (0) and async (1) requests, in ns. */
142 static const u64 bfq_fifo_expire
[2] = { NSEC_PER_SEC
/ 4, NSEC_PER_SEC
/ 8 };
144 /* Maximum backwards seek (magic number lifted from CFQ), in KiB. */
145 static const int bfq_back_max
= 16 * 1024;
147 /* Penalty of a backwards seek, in number of sectors. */
148 static const int bfq_back_penalty
= 2;
150 /* Idling period duration, in ns. */
151 static u64 bfq_slice_idle
= NSEC_PER_SEC
/ 125;
153 /* Minimum number of assigned budgets for which stats are safe to compute. */
154 static const int bfq_stats_min_budgets
= 194;
156 /* Default maximum budget values, in sectors and number of requests. */
157 static const int bfq_default_max_budget
= 16 * 1024;
160 * Async to sync throughput distribution is controlled as follows:
161 * when an async request is served, the entity is charged the number
162 * of sectors of the request, multiplied by the factor below
164 static const int bfq_async_charge_factor
= 10;
166 /* Default timeout values, in jiffies, approximating CFQ defaults. */
167 const int bfq_timeout
= HZ
/ 8;
169 static struct kmem_cache
*bfq_pool
;
171 /* Below this threshold (in ns), we consider thinktime immediate. */
172 #define BFQ_MIN_TT (2 * NSEC_PER_MSEC)
174 /* hw_tag detection: parallel requests threshold and min samples needed. */
175 #define BFQ_HW_QUEUE_THRESHOLD 4
176 #define BFQ_HW_QUEUE_SAMPLES 32
178 #define BFQQ_SEEK_THR (sector_t)(8 * 100)
179 #define BFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
180 #define BFQQ_CLOSE_THR (sector_t)(8 * 1024)
181 #define BFQQ_SEEKY(bfqq) (hweight32(bfqq->seek_history) > 32/8)
183 /* Min number of samples required to perform peak-rate update */
184 #define BFQ_RATE_MIN_SAMPLES 32
185 /* Min observation time interval required to perform a peak-rate update (ns) */
186 #define BFQ_RATE_MIN_INTERVAL (300*NSEC_PER_MSEC)
187 /* Target observation time interval for a peak-rate update (ns) */
188 #define BFQ_RATE_REF_INTERVAL NSEC_PER_SEC
190 /* Shift used for peak rate fixed precision calculations. */
191 #define BFQ_RATE_SHIFT 16
194 * By default, BFQ computes the duration of the weight raising for
195 * interactive applications automatically, using the following formula:
196 * duration = (R / r) * T, where r is the peak rate of the device, and
197 * R and T are two reference parameters.
198 * In particular, R is the peak rate of the reference device (see below),
199 * and T is a reference time: given the systems that are likely to be
200 * installed on the reference device according to its speed class, T is
201 * about the maximum time needed, under BFQ and while reading two files in
202 * parallel, to load typical large applications on these systems.
203 * In practice, the slower/faster the device at hand is, the more/less it
204 * takes to load applications with respect to the reference device.
205 * Accordingly, the longer/shorter BFQ grants weight raising to interactive
208 * BFQ uses four different reference pairs (R, T), depending on:
209 * . whether the device is rotational or non-rotational;
210 * . whether the device is slow, such as old or portable HDDs, as well as
211 * SD cards, or fast, such as newer HDDs and SSDs.
213 * The device's speed class is dynamically (re)detected in
214 * bfq_update_peak_rate() every time the estimated peak rate is updated.
216 * In the following definitions, R_slow[0]/R_fast[0] and
217 * T_slow[0]/T_fast[0] are the reference values for a slow/fast
218 * rotational device, whereas R_slow[1]/R_fast[1] and
219 * T_slow[1]/T_fast[1] are the reference values for a slow/fast
220 * non-rotational device. Finally, device_speed_thresh are the
221 * thresholds used to switch between speed classes. The reference
222 * rates are not the actual peak rates of the devices used as a
223 * reference, but slightly lower values. The reason for using these
224 * slightly lower values is that the peak-rate estimator tends to
225 * yield slightly lower values than the actual peak rate (it can yield
226 * the actual peak rate only if there is only one process doing I/O,
227 * and the process does sequential I/O).
229 * Both the reference peak rates and the thresholds are measured in
230 * sectors/usec, left-shifted by BFQ_RATE_SHIFT.
232 static int R_slow
[2] = {1000, 10700};
233 static int R_fast
[2] = {14000, 33000};
235 * To improve readability, a conversion function is used to initialize the
236 * following arrays, which entails that they can be initialized only in a
239 static int T_slow
[2];
240 static int T_fast
[2];
241 static int device_speed_thresh
[2];
243 #define RQ_BIC(rq) icq_to_bic((rq)->elv.priv[0])
244 #define RQ_BFQQ(rq) ((rq)->elv.priv[1])
246 struct bfq_queue
*bic_to_bfqq(struct bfq_io_cq
*bic
, bool is_sync
)
248 return bic
->bfqq
[is_sync
];
251 void bic_set_bfqq(struct bfq_io_cq
*bic
, struct bfq_queue
*bfqq
, bool is_sync
)
253 bic
->bfqq
[is_sync
] = bfqq
;
256 struct bfq_data
*bic_to_bfqd(struct bfq_io_cq
*bic
)
258 return bic
->icq
.q
->elevator
->elevator_data
;
262 * icq_to_bic - convert iocontext queue structure to bfq_io_cq.
263 * @icq: the iocontext queue.
265 static struct bfq_io_cq
*icq_to_bic(struct io_cq
*icq
)
267 /* bic->icq is the first member, %NULL will convert to %NULL */
268 return container_of(icq
, struct bfq_io_cq
, icq
);
272 * bfq_bic_lookup - search into @ioc a bic associated to @bfqd.
273 * @bfqd: the lookup key.
274 * @ioc: the io_context of the process doing I/O.
275 * @q: the request queue.
277 static struct bfq_io_cq
*bfq_bic_lookup(struct bfq_data
*bfqd
,
278 struct io_context
*ioc
,
279 struct request_queue
*q
)
283 struct bfq_io_cq
*icq
;
285 spin_lock_irqsave(q
->queue_lock
, flags
);
286 icq
= icq_to_bic(ioc_lookup_icq(ioc
, q
));
287 spin_unlock_irqrestore(q
->queue_lock
, flags
);
296 * Scheduler run of queue, if there are requests pending and no one in the
297 * driver that will restart queueing.
299 void bfq_schedule_dispatch(struct bfq_data
*bfqd
)
301 if (bfqd
->queued
!= 0) {
302 bfq_log(bfqd
, "schedule dispatch");
303 blk_mq_run_hw_queues(bfqd
->queue
, true);
307 #define bfq_class_idle(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
308 #define bfq_class_rt(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_RT)
310 #define bfq_sample_valid(samples) ((samples) > 80)
313 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
314 * We choose the request that is closesr to the head right now. Distance
315 * behind the head is penalized and only allowed to a certain extent.
317 static struct request
*bfq_choose_req(struct bfq_data
*bfqd
,
322 sector_t s1
, s2
, d1
= 0, d2
= 0;
323 unsigned long back_max
;
324 #define BFQ_RQ1_WRAP 0x01 /* request 1 wraps */
325 #define BFQ_RQ2_WRAP 0x02 /* request 2 wraps */
326 unsigned int wrap
= 0; /* bit mask: requests behind the disk head? */
328 if (!rq1
|| rq1
== rq2
)
333 if (rq_is_sync(rq1
) && !rq_is_sync(rq2
))
335 else if (rq_is_sync(rq2
) && !rq_is_sync(rq1
))
337 if ((rq1
->cmd_flags
& REQ_META
) && !(rq2
->cmd_flags
& REQ_META
))
339 else if ((rq2
->cmd_flags
& REQ_META
) && !(rq1
->cmd_flags
& REQ_META
))
342 s1
= blk_rq_pos(rq1
);
343 s2
= blk_rq_pos(rq2
);
346 * By definition, 1KiB is 2 sectors.
348 back_max
= bfqd
->bfq_back_max
* 2;
351 * Strict one way elevator _except_ in the case where we allow
352 * short backward seeks which are biased as twice the cost of a
353 * similar forward seek.
357 else if (s1
+ back_max
>= last
)
358 d1
= (last
- s1
) * bfqd
->bfq_back_penalty
;
360 wrap
|= BFQ_RQ1_WRAP
;
364 else if (s2
+ back_max
>= last
)
365 d2
= (last
- s2
) * bfqd
->bfq_back_penalty
;
367 wrap
|= BFQ_RQ2_WRAP
;
369 /* Found required data */
372 * By doing switch() on the bit mask "wrap" we avoid having to
373 * check two variables for all permutations: --> faster!
376 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
391 case BFQ_RQ1_WRAP
|BFQ_RQ2_WRAP
: /* both rqs wrapped */
394 * Since both rqs are wrapped,
395 * start with the one that's further behind head
396 * (--> only *one* back seek required),
397 * since back seek takes more time than forward.
406 static struct bfq_queue
*
407 bfq_rq_pos_tree_lookup(struct bfq_data
*bfqd
, struct rb_root
*root
,
408 sector_t sector
, struct rb_node
**ret_parent
,
409 struct rb_node
***rb_link
)
411 struct rb_node
**p
, *parent
;
412 struct bfq_queue
*bfqq
= NULL
;
420 bfqq
= rb_entry(parent
, struct bfq_queue
, pos_node
);
423 * Sort strictly based on sector. Smallest to the left,
424 * largest to the right.
426 if (sector
> blk_rq_pos(bfqq
->next_rq
))
428 else if (sector
< blk_rq_pos(bfqq
->next_rq
))
436 *ret_parent
= parent
;
440 bfq_log(bfqd
, "rq_pos_tree_lookup %llu: returning %d",
441 (unsigned long long)sector
,
442 bfqq
? bfqq
->pid
: 0);
447 void bfq_pos_tree_add_move(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
449 struct rb_node
**p
, *parent
;
450 struct bfq_queue
*__bfqq
;
452 if (bfqq
->pos_root
) {
453 rb_erase(&bfqq
->pos_node
, bfqq
->pos_root
);
454 bfqq
->pos_root
= NULL
;
457 if (bfq_class_idle(bfqq
))
462 bfqq
->pos_root
= &bfq_bfqq_to_bfqg(bfqq
)->rq_pos_tree
;
463 __bfqq
= bfq_rq_pos_tree_lookup(bfqd
, bfqq
->pos_root
,
464 blk_rq_pos(bfqq
->next_rq
), &parent
, &p
);
466 rb_link_node(&bfqq
->pos_node
, parent
, p
);
467 rb_insert_color(&bfqq
->pos_node
, bfqq
->pos_root
);
469 bfqq
->pos_root
= NULL
;
473 * Tell whether there are active queues or groups with differentiated weights.
475 static bool bfq_differentiated_weights(struct bfq_data
*bfqd
)
478 * For weights to differ, at least one of the trees must contain
479 * at least two nodes.
481 return (!RB_EMPTY_ROOT(&bfqd
->queue_weights_tree
) &&
482 (bfqd
->queue_weights_tree
.rb_node
->rb_left
||
483 bfqd
->queue_weights_tree
.rb_node
->rb_right
)
484 #ifdef CONFIG_BFQ_GROUP_IOSCHED
486 (!RB_EMPTY_ROOT(&bfqd
->group_weights_tree
) &&
487 (bfqd
->group_weights_tree
.rb_node
->rb_left
||
488 bfqd
->group_weights_tree
.rb_node
->rb_right
)
494 * The following function returns true if every queue must receive the
495 * same share of the throughput (this condition is used when deciding
496 * whether idling may be disabled, see the comments in the function
497 * bfq_bfqq_may_idle()).
499 * Such a scenario occurs when:
500 * 1) all active queues have the same weight,
501 * 2) all active groups at the same level in the groups tree have the same
503 * 3) all active groups at the same level in the groups tree have the same
504 * number of children.
506 * Unfortunately, keeping the necessary state for evaluating exactly the
507 * above symmetry conditions would be quite complex and time-consuming.
508 * Therefore this function evaluates, instead, the following stronger
509 * sub-conditions, for which it is much easier to maintain the needed
511 * 1) all active queues have the same weight,
512 * 2) all active groups have the same weight,
513 * 3) all active groups have at most one active child each.
514 * In particular, the last two conditions are always true if hierarchical
515 * support and the cgroups interface are not enabled, thus no state needs
516 * to be maintained in this case.
518 static bool bfq_symmetric_scenario(struct bfq_data
*bfqd
)
520 return !bfq_differentiated_weights(bfqd
);
524 * If the weight-counter tree passed as input contains no counter for
525 * the weight of the input entity, then add that counter; otherwise just
526 * increment the existing counter.
528 * Note that weight-counter trees contain few nodes in mostly symmetric
529 * scenarios. For example, if all queues have the same weight, then the
530 * weight-counter tree for the queues may contain at most one node.
531 * This holds even if low_latency is on, because weight-raised queues
532 * are not inserted in the tree.
533 * In most scenarios, the rate at which nodes are created/destroyed
536 void bfq_weights_tree_add(struct bfq_data
*bfqd
, struct bfq_entity
*entity
,
537 struct rb_root
*root
)
539 struct rb_node
**new = &(root
->rb_node
), *parent
= NULL
;
542 * Do not insert if the entity is already associated with a
543 * counter, which happens if:
544 * 1) the entity is associated with a queue,
545 * 2) a request arrival has caused the queue to become both
546 * non-weight-raised, and hence change its weight, and
547 * backlogged; in this respect, each of the two events
548 * causes an invocation of this function,
549 * 3) this is the invocation of this function caused by the
550 * second event. This second invocation is actually useless,
551 * and we handle this fact by exiting immediately. More
552 * efficient or clearer solutions might possibly be adopted.
554 if (entity
->weight_counter
)
558 struct bfq_weight_counter
*__counter
= container_of(*new,
559 struct bfq_weight_counter
,
563 if (entity
->weight
== __counter
->weight
) {
564 entity
->weight_counter
= __counter
;
567 if (entity
->weight
< __counter
->weight
)
568 new = &((*new)->rb_left
);
570 new = &((*new)->rb_right
);
573 entity
->weight_counter
= kzalloc(sizeof(struct bfq_weight_counter
),
577 * In the unlucky event of an allocation failure, we just
578 * exit. This will cause the weight of entity to not be
579 * considered in bfq_differentiated_weights, which, in its
580 * turn, causes the scenario to be deemed wrongly symmetric in
581 * case entity's weight would have been the only weight making
582 * the scenario asymmetric. On the bright side, no unbalance
583 * will however occur when entity becomes inactive again (the
584 * invocation of this function is triggered by an activation
585 * of entity). In fact, bfq_weights_tree_remove does nothing
586 * if !entity->weight_counter.
588 if (unlikely(!entity
->weight_counter
))
591 entity
->weight_counter
->weight
= entity
->weight
;
592 rb_link_node(&entity
->weight_counter
->weights_node
, parent
, new);
593 rb_insert_color(&entity
->weight_counter
->weights_node
, root
);
596 entity
->weight_counter
->num_active
++;
600 * Decrement the weight counter associated with the entity, and, if the
601 * counter reaches 0, remove the counter from the tree.
602 * See the comments to the function bfq_weights_tree_add() for considerations
605 void bfq_weights_tree_remove(struct bfq_data
*bfqd
, struct bfq_entity
*entity
,
606 struct rb_root
*root
)
608 if (!entity
->weight_counter
)
611 entity
->weight_counter
->num_active
--;
612 if (entity
->weight_counter
->num_active
> 0)
613 goto reset_entity_pointer
;
615 rb_erase(&entity
->weight_counter
->weights_node
, root
);
616 kfree(entity
->weight_counter
);
618 reset_entity_pointer
:
619 entity
->weight_counter
= NULL
;
623 * Return expired entry, or NULL to just start from scratch in rbtree.
625 static struct request
*bfq_check_fifo(struct bfq_queue
*bfqq
,
626 struct request
*last
)
630 if (bfq_bfqq_fifo_expire(bfqq
))
633 bfq_mark_bfqq_fifo_expire(bfqq
);
635 rq
= rq_entry_fifo(bfqq
->fifo
.next
);
637 if (rq
== last
|| ktime_get_ns() < rq
->fifo_time
)
640 bfq_log_bfqq(bfqq
->bfqd
, bfqq
, "check_fifo: returned %p", rq
);
644 static struct request
*bfq_find_next_rq(struct bfq_data
*bfqd
,
645 struct bfq_queue
*bfqq
,
646 struct request
*last
)
648 struct rb_node
*rbnext
= rb_next(&last
->rb_node
);
649 struct rb_node
*rbprev
= rb_prev(&last
->rb_node
);
650 struct request
*next
, *prev
= NULL
;
652 /* Follow expired path, else get first next available. */
653 next
= bfq_check_fifo(bfqq
, last
);
658 prev
= rb_entry_rq(rbprev
);
661 next
= rb_entry_rq(rbnext
);
663 rbnext
= rb_first(&bfqq
->sort_list
);
664 if (rbnext
&& rbnext
!= &last
->rb_node
)
665 next
= rb_entry_rq(rbnext
);
668 return bfq_choose_req(bfqd
, next
, prev
, blk_rq_pos(last
));
671 /* see the definition of bfq_async_charge_factor for details */
672 static unsigned long bfq_serv_to_charge(struct request
*rq
,
673 struct bfq_queue
*bfqq
)
675 if (bfq_bfqq_sync(bfqq
) || bfqq
->wr_coeff
> 1)
676 return blk_rq_sectors(rq
);
679 * If there are no weight-raised queues, then amplify service
680 * by just the async charge factor; otherwise amplify service
681 * by twice the async charge factor, to further reduce latency
682 * for weight-raised queues.
684 if (bfqq
->bfqd
->wr_busy_queues
== 0)
685 return blk_rq_sectors(rq
) * bfq_async_charge_factor
;
687 return blk_rq_sectors(rq
) * 2 * bfq_async_charge_factor
;
691 * bfq_updated_next_req - update the queue after a new next_rq selection.
692 * @bfqd: the device data the queue belongs to.
693 * @bfqq: the queue to update.
695 * If the first request of a queue changes we make sure that the queue
696 * has enough budget to serve at least its first request (if the
697 * request has grown). We do this because if the queue has not enough
698 * budget for its first request, it has to go through two dispatch
699 * rounds to actually get it dispatched.
701 static void bfq_updated_next_req(struct bfq_data
*bfqd
,
702 struct bfq_queue
*bfqq
)
704 struct bfq_entity
*entity
= &bfqq
->entity
;
705 struct request
*next_rq
= bfqq
->next_rq
;
706 unsigned long new_budget
;
711 if (bfqq
== bfqd
->in_service_queue
)
713 * In order not to break guarantees, budgets cannot be
714 * changed after an entity has been selected.
718 new_budget
= max_t(unsigned long, bfqq
->max_budget
,
719 bfq_serv_to_charge(next_rq
, bfqq
));
720 if (entity
->budget
!= new_budget
) {
721 entity
->budget
= new_budget
;
722 bfq_log_bfqq(bfqd
, bfqq
, "updated next rq: new budget %lu",
724 bfq_requeue_bfqq(bfqd
, bfqq
, false);
728 static unsigned int bfq_wr_duration(struct bfq_data
*bfqd
)
732 if (bfqd
->bfq_wr_max_time
> 0)
733 return bfqd
->bfq_wr_max_time
;
736 do_div(dur
, bfqd
->peak_rate
);
739 * Limit duration between 3 and 13 seconds. Tests show that
740 * higher values than 13 seconds often yield the opposite of
741 * the desired result, i.e., worsen responsiveness by letting
742 * non-interactive and non-soft-real-time applications
743 * preserve weight raising for a too long time interval.
745 * On the other end, lower values than 3 seconds make it
746 * difficult for most interactive tasks to complete their jobs
747 * before weight-raising finishes.
749 if (dur
> msecs_to_jiffies(13000))
750 dur
= msecs_to_jiffies(13000);
751 else if (dur
< msecs_to_jiffies(3000))
752 dur
= msecs_to_jiffies(3000);
757 /* switch back from soft real-time to interactive weight raising */
758 static void switch_back_to_interactive_wr(struct bfq_queue
*bfqq
,
759 struct bfq_data
*bfqd
)
761 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
;
762 bfqq
->wr_cur_max_time
= bfq_wr_duration(bfqd
);
763 bfqq
->last_wr_start_finish
= bfqq
->wr_start_at_switch_to_srt
;
767 bfq_bfqq_resume_state(struct bfq_queue
*bfqq
, struct bfq_data
*bfqd
,
768 struct bfq_io_cq
*bic
, bool bfq_already_existing
)
770 unsigned int old_wr_coeff
= bfqq
->wr_coeff
;
771 bool busy
= bfq_already_existing
&& bfq_bfqq_busy(bfqq
);
773 if (bic
->saved_has_short_ttime
)
774 bfq_mark_bfqq_has_short_ttime(bfqq
);
776 bfq_clear_bfqq_has_short_ttime(bfqq
);
778 if (bic
->saved_IO_bound
)
779 bfq_mark_bfqq_IO_bound(bfqq
);
781 bfq_clear_bfqq_IO_bound(bfqq
);
783 bfqq
->ttime
= bic
->saved_ttime
;
784 bfqq
->wr_coeff
= bic
->saved_wr_coeff
;
785 bfqq
->wr_start_at_switch_to_srt
= bic
->saved_wr_start_at_switch_to_srt
;
786 bfqq
->last_wr_start_finish
= bic
->saved_last_wr_start_finish
;
787 bfqq
->wr_cur_max_time
= bic
->saved_wr_cur_max_time
;
789 if (bfqq
->wr_coeff
> 1 && (bfq_bfqq_in_large_burst(bfqq
) ||
790 time_is_before_jiffies(bfqq
->last_wr_start_finish
+
791 bfqq
->wr_cur_max_time
))) {
792 if (bfqq
->wr_cur_max_time
== bfqd
->bfq_wr_rt_max_time
&&
793 !bfq_bfqq_in_large_burst(bfqq
) &&
794 time_is_after_eq_jiffies(bfqq
->wr_start_at_switch_to_srt
+
795 bfq_wr_duration(bfqd
))) {
796 switch_back_to_interactive_wr(bfqq
, bfqd
);
799 bfq_log_bfqq(bfqq
->bfqd
, bfqq
,
800 "resume state: switching off wr");
804 /* make sure weight will be updated, however we got here */
805 bfqq
->entity
.prio_changed
= 1;
810 if (old_wr_coeff
== 1 && bfqq
->wr_coeff
> 1)
811 bfqd
->wr_busy_queues
++;
812 else if (old_wr_coeff
> 1 && bfqq
->wr_coeff
== 1)
813 bfqd
->wr_busy_queues
--;
816 static int bfqq_process_refs(struct bfq_queue
*bfqq
)
818 return bfqq
->ref
- bfqq
->allocated
- bfqq
->entity
.on_st
;
821 /* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */
822 static void bfq_reset_burst_list(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
824 struct bfq_queue
*item
;
825 struct hlist_node
*n
;
827 hlist_for_each_entry_safe(item
, n
, &bfqd
->burst_list
, burst_list_node
)
828 hlist_del_init(&item
->burst_list_node
);
829 hlist_add_head(&bfqq
->burst_list_node
, &bfqd
->burst_list
);
830 bfqd
->burst_size
= 1;
831 bfqd
->burst_parent_entity
= bfqq
->entity
.parent
;
834 /* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */
835 static void bfq_add_to_burst(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
837 /* Increment burst size to take into account also bfqq */
840 if (bfqd
->burst_size
== bfqd
->bfq_large_burst_thresh
) {
841 struct bfq_queue
*pos
, *bfqq_item
;
842 struct hlist_node
*n
;
845 * Enough queues have been activated shortly after each
846 * other to consider this burst as large.
848 bfqd
->large_burst
= true;
851 * We can now mark all queues in the burst list as
852 * belonging to a large burst.
854 hlist_for_each_entry(bfqq_item
, &bfqd
->burst_list
,
856 bfq_mark_bfqq_in_large_burst(bfqq_item
);
857 bfq_mark_bfqq_in_large_burst(bfqq
);
860 * From now on, and until the current burst finishes, any
861 * new queue being activated shortly after the last queue
862 * was inserted in the burst can be immediately marked as
863 * belonging to a large burst. So the burst list is not
864 * needed any more. Remove it.
866 hlist_for_each_entry_safe(pos
, n
, &bfqd
->burst_list
,
868 hlist_del_init(&pos
->burst_list_node
);
870 * Burst not yet large: add bfqq to the burst list. Do
871 * not increment the ref counter for bfqq, because bfqq
872 * is removed from the burst list before freeing bfqq
875 hlist_add_head(&bfqq
->burst_list_node
, &bfqd
->burst_list
);
879 * If many queues belonging to the same group happen to be created
880 * shortly after each other, then the processes associated with these
881 * queues have typically a common goal. In particular, bursts of queue
882 * creations are usually caused by services or applications that spawn
883 * many parallel threads/processes. Examples are systemd during boot,
884 * or git grep. To help these processes get their job done as soon as
885 * possible, it is usually better to not grant either weight-raising
886 * or device idling to their queues.
888 * In this comment we describe, firstly, the reasons why this fact
889 * holds, and, secondly, the next function, which implements the main
890 * steps needed to properly mark these queues so that they can then be
891 * treated in a different way.
893 * The above services or applications benefit mostly from a high
894 * throughput: the quicker the requests of the activated queues are
895 * cumulatively served, the sooner the target job of these queues gets
896 * completed. As a consequence, weight-raising any of these queues,
897 * which also implies idling the device for it, is almost always
898 * counterproductive. In most cases it just lowers throughput.
900 * On the other hand, a burst of queue creations may be caused also by
901 * the start of an application that does not consist of a lot of
902 * parallel I/O-bound threads. In fact, with a complex application,
903 * several short processes may need to be executed to start-up the
904 * application. In this respect, to start an application as quickly as
905 * possible, the best thing to do is in any case to privilege the I/O
906 * related to the application with respect to all other
907 * I/O. Therefore, the best strategy to start as quickly as possible
908 * an application that causes a burst of queue creations is to
909 * weight-raise all the queues created during the burst. This is the
910 * exact opposite of the best strategy for the other type of bursts.
912 * In the end, to take the best action for each of the two cases, the
913 * two types of bursts need to be distinguished. Fortunately, this
914 * seems relatively easy, by looking at the sizes of the bursts. In
915 * particular, we found a threshold such that only bursts with a
916 * larger size than that threshold are apparently caused by
917 * services or commands such as systemd or git grep. For brevity,
918 * hereafter we call just 'large' these bursts. BFQ *does not*
919 * weight-raise queues whose creation occurs in a large burst. In
920 * addition, for each of these queues BFQ performs or does not perform
921 * idling depending on which choice boosts the throughput more. The
922 * exact choice depends on the device and request pattern at
925 * Unfortunately, false positives may occur while an interactive task
926 * is starting (e.g., an application is being started). The
927 * consequence is that the queues associated with the task do not
928 * enjoy weight raising as expected. Fortunately these false positives
929 * are very rare. They typically occur if some service happens to
930 * start doing I/O exactly when the interactive task starts.
932 * Turning back to the next function, it implements all the steps
933 * needed to detect the occurrence of a large burst and to properly
934 * mark all the queues belonging to it (so that they can then be
935 * treated in a different way). This goal is achieved by maintaining a
936 * "burst list" that holds, temporarily, the queues that belong to the
937 * burst in progress. The list is then used to mark these queues as
938 * belonging to a large burst if the burst does become large. The main
939 * steps are the following.
941 * . when the very first queue is created, the queue is inserted into the
942 * list (as it could be the first queue in a possible burst)
944 * . if the current burst has not yet become large, and a queue Q that does
945 * not yet belong to the burst is activated shortly after the last time
946 * at which a new queue entered the burst list, then the function appends
947 * Q to the burst list
949 * . if, as a consequence of the previous step, the burst size reaches
950 * the large-burst threshold, then
952 * . all the queues in the burst list are marked as belonging to a
955 * . the burst list is deleted; in fact, the burst list already served
956 * its purpose (keeping temporarily track of the queues in a burst,
957 * so as to be able to mark them as belonging to a large burst in the
958 * previous sub-step), and now is not needed any more
960 * . the device enters a large-burst mode
962 * . if a queue Q that does not belong to the burst is created while
963 * the device is in large-burst mode and shortly after the last time
964 * at which a queue either entered the burst list or was marked as
965 * belonging to the current large burst, then Q is immediately marked
966 * as belonging to a large burst.
968 * . if a queue Q that does not belong to the burst is created a while
969 * later, i.e., not shortly after, than the last time at which a queue
970 * either entered the burst list or was marked as belonging to the
971 * current large burst, then the current burst is deemed as finished and:
973 * . the large-burst mode is reset if set
975 * . the burst list is emptied
977 * . Q is inserted in the burst list, as Q may be the first queue
978 * in a possible new burst (then the burst list contains just Q
981 static void bfq_handle_burst(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
984 * If bfqq is already in the burst list or is part of a large
985 * burst, or finally has just been split, then there is
986 * nothing else to do.
988 if (!hlist_unhashed(&bfqq
->burst_list_node
) ||
989 bfq_bfqq_in_large_burst(bfqq
) ||
990 time_is_after_eq_jiffies(bfqq
->split_time
+
991 msecs_to_jiffies(10)))
995 * If bfqq's creation happens late enough, or bfqq belongs to
996 * a different group than the burst group, then the current
997 * burst is finished, and related data structures must be
1000 * In this respect, consider the special case where bfqq is
1001 * the very first queue created after BFQ is selected for this
1002 * device. In this case, last_ins_in_burst and
1003 * burst_parent_entity are not yet significant when we get
1004 * here. But it is easy to verify that, whether or not the
1005 * following condition is true, bfqq will end up being
1006 * inserted into the burst list. In particular the list will
1007 * happen to contain only bfqq. And this is exactly what has
1008 * to happen, as bfqq may be the first queue of the first
1011 if (time_is_before_jiffies(bfqd
->last_ins_in_burst
+
1012 bfqd
->bfq_burst_interval
) ||
1013 bfqq
->entity
.parent
!= bfqd
->burst_parent_entity
) {
1014 bfqd
->large_burst
= false;
1015 bfq_reset_burst_list(bfqd
, bfqq
);
1020 * If we get here, then bfqq is being activated shortly after the
1021 * last queue. So, if the current burst is also large, we can mark
1022 * bfqq as belonging to this large burst immediately.
1024 if (bfqd
->large_burst
) {
1025 bfq_mark_bfqq_in_large_burst(bfqq
);
1030 * If we get here, then a large-burst state has not yet been
1031 * reached, but bfqq is being activated shortly after the last
1032 * queue. Then we add bfqq to the burst.
1034 bfq_add_to_burst(bfqd
, bfqq
);
1037 * At this point, bfqq either has been added to the current
1038 * burst or has caused the current burst to terminate and a
1039 * possible new burst to start. In particular, in the second
1040 * case, bfqq has become the first queue in the possible new
1041 * burst. In both cases last_ins_in_burst needs to be moved
1044 bfqd
->last_ins_in_burst
= jiffies
;
1047 static int bfq_bfqq_budget_left(struct bfq_queue
*bfqq
)
1049 struct bfq_entity
*entity
= &bfqq
->entity
;
1051 return entity
->budget
- entity
->service
;
1055 * If enough samples have been computed, return the current max budget
1056 * stored in bfqd, which is dynamically updated according to the
1057 * estimated disk peak rate; otherwise return the default max budget
1059 static int bfq_max_budget(struct bfq_data
*bfqd
)
1061 if (bfqd
->budgets_assigned
< bfq_stats_min_budgets
)
1062 return bfq_default_max_budget
;
1064 return bfqd
->bfq_max_budget
;
1068 * Return min budget, which is a fraction of the current or default
1069 * max budget (trying with 1/32)
1071 static int bfq_min_budget(struct bfq_data
*bfqd
)
1073 if (bfqd
->budgets_assigned
< bfq_stats_min_budgets
)
1074 return bfq_default_max_budget
/ 32;
1076 return bfqd
->bfq_max_budget
/ 32;
1080 * The next function, invoked after the input queue bfqq switches from
1081 * idle to busy, updates the budget of bfqq. The function also tells
1082 * whether the in-service queue should be expired, by returning
1083 * true. The purpose of expiring the in-service queue is to give bfqq
1084 * the chance to possibly preempt the in-service queue, and the reason
1085 * for preempting the in-service queue is to achieve one of the two
1088 * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has
1089 * expired because it has remained idle. In particular, bfqq may have
1090 * expired for one of the following two reasons:
1092 * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling
1093 * and did not make it to issue a new request before its last
1094 * request was served;
1096 * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue
1097 * a new request before the expiration of the idling-time.
1099 * Even if bfqq has expired for one of the above reasons, the process
1100 * associated with the queue may be however issuing requests greedily,
1101 * and thus be sensitive to the bandwidth it receives (bfqq may have
1102 * remained idle for other reasons: CPU high load, bfqq not enjoying
1103 * idling, I/O throttling somewhere in the path from the process to
1104 * the I/O scheduler, ...). But if, after every expiration for one of
1105 * the above two reasons, bfqq has to wait for the service of at least
1106 * one full budget of another queue before being served again, then
1107 * bfqq is likely to get a much lower bandwidth or resource time than
1108 * its reserved ones. To address this issue, two countermeasures need
1111 * First, the budget and the timestamps of bfqq need to be updated in
1112 * a special way on bfqq reactivation: they need to be updated as if
1113 * bfqq did not remain idle and did not expire. In fact, if they are
1114 * computed as if bfqq expired and remained idle until reactivation,
1115 * then the process associated with bfqq is treated as if, instead of
1116 * being greedy, it stopped issuing requests when bfqq remained idle,
1117 * and restarts issuing requests only on this reactivation. In other
1118 * words, the scheduler does not help the process recover the "service
1119 * hole" between bfqq expiration and reactivation. As a consequence,
1120 * the process receives a lower bandwidth than its reserved one. In
1121 * contrast, to recover this hole, the budget must be updated as if
1122 * bfqq was not expired at all before this reactivation, i.e., it must
1123 * be set to the value of the remaining budget when bfqq was
1124 * expired. Along the same line, timestamps need to be assigned the
1125 * value they had the last time bfqq was selected for service, i.e.,
1126 * before last expiration. Thus timestamps need to be back-shifted
1127 * with respect to their normal computation (see [1] for more details
1128 * on this tricky aspect).
1130 * Secondly, to allow the process to recover the hole, the in-service
1131 * queue must be expired too, to give bfqq the chance to preempt it
1132 * immediately. In fact, if bfqq has to wait for a full budget of the
1133 * in-service queue to be completed, then it may become impossible to
1134 * let the process recover the hole, even if the back-shifted
1135 * timestamps of bfqq are lower than those of the in-service queue. If
1136 * this happens for most or all of the holes, then the process may not
1137 * receive its reserved bandwidth. In this respect, it is worth noting
1138 * that, being the service of outstanding requests unpreemptible, a
1139 * little fraction of the holes may however be unrecoverable, thereby
1140 * causing a little loss of bandwidth.
1142 * The last important point is detecting whether bfqq does need this
1143 * bandwidth recovery. In this respect, the next function deems the
1144 * process associated with bfqq greedy, and thus allows it to recover
1145 * the hole, if: 1) the process is waiting for the arrival of a new
1146 * request (which implies that bfqq expired for one of the above two
1147 * reasons), and 2) such a request has arrived soon. The first
1148 * condition is controlled through the flag non_blocking_wait_rq,
1149 * while the second through the flag arrived_in_time. If both
1150 * conditions hold, then the function computes the budget in the
1151 * above-described special way, and signals that the in-service queue
1152 * should be expired. Timestamp back-shifting is done later in
1153 * __bfq_activate_entity.
1155 * 2. Reduce latency. Even if timestamps are not backshifted to let
1156 * the process associated with bfqq recover a service hole, bfqq may
1157 * however happen to have, after being (re)activated, a lower finish
1158 * timestamp than the in-service queue. That is, the next budget of
1159 * bfqq may have to be completed before the one of the in-service
1160 * queue. If this is the case, then preempting the in-service queue
1161 * allows this goal to be achieved, apart from the unpreemptible,
1162 * outstanding requests mentioned above.
1164 * Unfortunately, regardless of which of the above two goals one wants
1165 * to achieve, service trees need first to be updated to know whether
1166 * the in-service queue must be preempted. To have service trees
1167 * correctly updated, the in-service queue must be expired and
1168 * rescheduled, and bfqq must be scheduled too. This is one of the
1169 * most costly operations (in future versions, the scheduling
1170 * mechanism may be re-designed in such a way to make it possible to
1171 * know whether preemption is needed without needing to update service
1172 * trees). In addition, queue preemptions almost always cause random
1173 * I/O, and thus loss of throughput. Because of these facts, the next
1174 * function adopts the following simple scheme to avoid both costly
1175 * operations and too frequent preemptions: it requests the expiration
1176 * of the in-service queue (unconditionally) only for queues that need
1177 * to recover a hole, or that either are weight-raised or deserve to
1180 static bool bfq_bfqq_update_budg_for_activation(struct bfq_data
*bfqd
,
1181 struct bfq_queue
*bfqq
,
1182 bool arrived_in_time
,
1183 bool wr_or_deserves_wr
)
1185 struct bfq_entity
*entity
= &bfqq
->entity
;
1187 if (bfq_bfqq_non_blocking_wait_rq(bfqq
) && arrived_in_time
) {
1189 * We do not clear the flag non_blocking_wait_rq here, as
1190 * the latter is used in bfq_activate_bfqq to signal
1191 * that timestamps need to be back-shifted (and is
1192 * cleared right after).
1196 * In next assignment we rely on that either
1197 * entity->service or entity->budget are not updated
1198 * on expiration if bfqq is empty (see
1199 * __bfq_bfqq_recalc_budget). Thus both quantities
1200 * remain unchanged after such an expiration, and the
1201 * following statement therefore assigns to
1202 * entity->budget the remaining budget on such an
1203 * expiration. For clarity, entity->service is not
1204 * updated on expiration in any case, and, in normal
1205 * operation, is reset only when bfqq is selected for
1206 * service (see bfq_get_next_queue).
1208 entity
->budget
= min_t(unsigned long,
1209 bfq_bfqq_budget_left(bfqq
),
1215 entity
->budget
= max_t(unsigned long, bfqq
->max_budget
,
1216 bfq_serv_to_charge(bfqq
->next_rq
, bfqq
));
1217 bfq_clear_bfqq_non_blocking_wait_rq(bfqq
);
1218 return wr_or_deserves_wr
;
1222 * Return the farthest future time instant according to jiffies
1225 static unsigned long bfq_greatest_from_now(void)
1227 return jiffies
+ MAX_JIFFY_OFFSET
;
1231 * Return the farthest past time instant according to jiffies
1234 static unsigned long bfq_smallest_from_now(void)
1236 return jiffies
- MAX_JIFFY_OFFSET
;
1239 static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data
*bfqd
,
1240 struct bfq_queue
*bfqq
,
1241 unsigned int old_wr_coeff
,
1242 bool wr_or_deserves_wr
,
1247 if (old_wr_coeff
== 1 && wr_or_deserves_wr
) {
1248 /* start a weight-raising period */
1250 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
;
1251 bfqq
->wr_cur_max_time
= bfq_wr_duration(bfqd
);
1254 * No interactive weight raising in progress
1255 * here: assign minus infinity to
1256 * wr_start_at_switch_to_srt, to make sure
1257 * that, at the end of the soft-real-time
1258 * weight raising periods that is starting
1259 * now, no interactive weight-raising period
1260 * may be wrongly considered as still in
1261 * progress (and thus actually started by
1264 bfqq
->wr_start_at_switch_to_srt
=
1265 bfq_smallest_from_now();
1266 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
*
1267 BFQ_SOFTRT_WEIGHT_FACTOR
;
1268 bfqq
->wr_cur_max_time
=
1269 bfqd
->bfq_wr_rt_max_time
;
1273 * If needed, further reduce budget to make sure it is
1274 * close to bfqq's backlog, so as to reduce the
1275 * scheduling-error component due to a too large
1276 * budget. Do not care about throughput consequences,
1277 * but only about latency. Finally, do not assign a
1278 * too small budget either, to avoid increasing
1279 * latency by causing too frequent expirations.
1281 bfqq
->entity
.budget
= min_t(unsigned long,
1282 bfqq
->entity
.budget
,
1283 2 * bfq_min_budget(bfqd
));
1284 } else if (old_wr_coeff
> 1) {
1285 if (interactive
) { /* update wr coeff and duration */
1286 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
;
1287 bfqq
->wr_cur_max_time
= bfq_wr_duration(bfqd
);
1288 } else if (in_burst
)
1292 * The application is now or still meeting the
1293 * requirements for being deemed soft rt. We
1294 * can then correctly and safely (re)charge
1295 * the weight-raising duration for the
1296 * application with the weight-raising
1297 * duration for soft rt applications.
1299 * In particular, doing this recharge now, i.e.,
1300 * before the weight-raising period for the
1301 * application finishes, reduces the probability
1302 * of the following negative scenario:
1303 * 1) the weight of a soft rt application is
1304 * raised at startup (as for any newly
1305 * created application),
1306 * 2) since the application is not interactive,
1307 * at a certain time weight-raising is
1308 * stopped for the application,
1309 * 3) at that time the application happens to
1310 * still have pending requests, and hence
1311 * is destined to not have a chance to be
1312 * deemed soft rt before these requests are
1313 * completed (see the comments to the
1314 * function bfq_bfqq_softrt_next_start()
1315 * for details on soft rt detection),
1316 * 4) these pending requests experience a high
1317 * latency because the application is not
1318 * weight-raised while they are pending.
1320 if (bfqq
->wr_cur_max_time
!=
1321 bfqd
->bfq_wr_rt_max_time
) {
1322 bfqq
->wr_start_at_switch_to_srt
=
1323 bfqq
->last_wr_start_finish
;
1325 bfqq
->wr_cur_max_time
=
1326 bfqd
->bfq_wr_rt_max_time
;
1327 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
*
1328 BFQ_SOFTRT_WEIGHT_FACTOR
;
1330 bfqq
->last_wr_start_finish
= jiffies
;
1335 static bool bfq_bfqq_idle_for_long_time(struct bfq_data
*bfqd
,
1336 struct bfq_queue
*bfqq
)
1338 return bfqq
->dispatched
== 0 &&
1339 time_is_before_jiffies(
1340 bfqq
->budget_timeout
+
1341 bfqd
->bfq_wr_min_idle_time
);
1344 static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data
*bfqd
,
1345 struct bfq_queue
*bfqq
,
1350 bool soft_rt
, in_burst
, wr_or_deserves_wr
,
1351 bfqq_wants_to_preempt
,
1352 idle_for_long_time
= bfq_bfqq_idle_for_long_time(bfqd
, bfqq
),
1354 * See the comments on
1355 * bfq_bfqq_update_budg_for_activation for
1356 * details on the usage of the next variable.
1358 arrived_in_time
= ktime_get_ns() <=
1359 bfqq
->ttime
.last_end_request
+
1360 bfqd
->bfq_slice_idle
* 3;
1364 * bfqq deserves to be weight-raised if:
1366 * - it does not belong to a large burst,
1367 * - it has been idle for enough time or is soft real-time,
1368 * - is linked to a bfq_io_cq (it is not shared in any sense).
1370 in_burst
= bfq_bfqq_in_large_burst(bfqq
);
1371 soft_rt
= bfqd
->bfq_wr_max_softrt_rate
> 0 &&
1373 time_is_before_jiffies(bfqq
->soft_rt_next_start
);
1374 *interactive
= !in_burst
&& idle_for_long_time
;
1375 wr_or_deserves_wr
= bfqd
->low_latency
&&
1376 (bfqq
->wr_coeff
> 1 ||
1377 (bfq_bfqq_sync(bfqq
) &&
1378 bfqq
->bic
&& (*interactive
|| soft_rt
)));
1381 * Using the last flag, update budget and check whether bfqq
1382 * may want to preempt the in-service queue.
1384 bfqq_wants_to_preempt
=
1385 bfq_bfqq_update_budg_for_activation(bfqd
, bfqq
,
1390 * If bfqq happened to be activated in a burst, but has been
1391 * idle for much more than an interactive queue, then we
1392 * assume that, in the overall I/O initiated in the burst, the
1393 * I/O associated with bfqq is finished. So bfqq does not need
1394 * to be treated as a queue belonging to a burst
1395 * anymore. Accordingly, we reset bfqq's in_large_burst flag
1396 * if set, and remove bfqq from the burst list if it's
1397 * there. We do not decrement burst_size, because the fact
1398 * that bfqq does not need to belong to the burst list any
1399 * more does not invalidate the fact that bfqq was created in
1402 if (likely(!bfq_bfqq_just_created(bfqq
)) &&
1403 idle_for_long_time
&&
1404 time_is_before_jiffies(
1405 bfqq
->budget_timeout
+
1406 msecs_to_jiffies(10000))) {
1407 hlist_del_init(&bfqq
->burst_list_node
);
1408 bfq_clear_bfqq_in_large_burst(bfqq
);
1411 bfq_clear_bfqq_just_created(bfqq
);
1414 if (!bfq_bfqq_IO_bound(bfqq
)) {
1415 if (arrived_in_time
) {
1416 bfqq
->requests_within_timer
++;
1417 if (bfqq
->requests_within_timer
>=
1418 bfqd
->bfq_requests_within_timer
)
1419 bfq_mark_bfqq_IO_bound(bfqq
);
1421 bfqq
->requests_within_timer
= 0;
1424 if (bfqd
->low_latency
) {
1425 if (unlikely(time_is_after_jiffies(bfqq
->split_time
)))
1428 jiffies
- bfqd
->bfq_wr_min_idle_time
- 1;
1430 if (time_is_before_jiffies(bfqq
->split_time
+
1431 bfqd
->bfq_wr_min_idle_time
)) {
1432 bfq_update_bfqq_wr_on_rq_arrival(bfqd
, bfqq
,
1439 if (old_wr_coeff
!= bfqq
->wr_coeff
)
1440 bfqq
->entity
.prio_changed
= 1;
1444 bfqq
->last_idle_bklogged
= jiffies
;
1445 bfqq
->service_from_backlogged
= 0;
1446 bfq_clear_bfqq_softrt_update(bfqq
);
1448 bfq_add_bfqq_busy(bfqd
, bfqq
);
1451 * Expire in-service queue only if preemption may be needed
1452 * for guarantees. In this respect, the function
1453 * next_queue_may_preempt just checks a simple, necessary
1454 * condition, and not a sufficient condition based on
1455 * timestamps. In fact, for the latter condition to be
1456 * evaluated, timestamps would need first to be updated, and
1457 * this operation is quite costly (see the comments on the
1458 * function bfq_bfqq_update_budg_for_activation).
1460 if (bfqd
->in_service_queue
&& bfqq_wants_to_preempt
&&
1461 bfqd
->in_service_queue
->wr_coeff
< bfqq
->wr_coeff
&&
1462 next_queue_may_preempt(bfqd
))
1463 bfq_bfqq_expire(bfqd
, bfqd
->in_service_queue
,
1464 false, BFQQE_PREEMPTED
);
1467 static void bfq_add_request(struct request
*rq
)
1469 struct bfq_queue
*bfqq
= RQ_BFQQ(rq
);
1470 struct bfq_data
*bfqd
= bfqq
->bfqd
;
1471 struct request
*next_rq
, *prev
;
1472 unsigned int old_wr_coeff
= bfqq
->wr_coeff
;
1473 bool interactive
= false;
1475 bfq_log_bfqq(bfqd
, bfqq
, "add_request %d", rq_is_sync(rq
));
1476 bfqq
->queued
[rq_is_sync(rq
)]++;
1479 elv_rb_add(&bfqq
->sort_list
, rq
);
1482 * Check if this request is a better next-serve candidate.
1484 prev
= bfqq
->next_rq
;
1485 next_rq
= bfq_choose_req(bfqd
, bfqq
->next_rq
, rq
, bfqd
->last_position
);
1486 bfqq
->next_rq
= next_rq
;
1489 * Adjust priority tree position, if next_rq changes.
1491 if (prev
!= bfqq
->next_rq
)
1492 bfq_pos_tree_add_move(bfqd
, bfqq
);
1494 if (!bfq_bfqq_busy(bfqq
)) /* switching to busy ... */
1495 bfq_bfqq_handle_idle_busy_switch(bfqd
, bfqq
, old_wr_coeff
,
1498 if (bfqd
->low_latency
&& old_wr_coeff
== 1 && !rq_is_sync(rq
) &&
1499 time_is_before_jiffies(
1500 bfqq
->last_wr_start_finish
+
1501 bfqd
->bfq_wr_min_inter_arr_async
)) {
1502 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
;
1503 bfqq
->wr_cur_max_time
= bfq_wr_duration(bfqd
);
1505 bfqd
->wr_busy_queues
++;
1506 bfqq
->entity
.prio_changed
= 1;
1508 if (prev
!= bfqq
->next_rq
)
1509 bfq_updated_next_req(bfqd
, bfqq
);
1513 * Assign jiffies to last_wr_start_finish in the following
1516 * . if bfqq is not going to be weight-raised, because, for
1517 * non weight-raised queues, last_wr_start_finish stores the
1518 * arrival time of the last request; as of now, this piece
1519 * of information is used only for deciding whether to
1520 * weight-raise async queues
1522 * . if bfqq is not weight-raised, because, if bfqq is now
1523 * switching to weight-raised, then last_wr_start_finish
1524 * stores the time when weight-raising starts
1526 * . if bfqq is interactive, because, regardless of whether
1527 * bfqq is currently weight-raised, the weight-raising
1528 * period must start or restart (this case is considered
1529 * separately because it is not detected by the above
1530 * conditions, if bfqq is already weight-raised)
1532 * last_wr_start_finish has to be updated also if bfqq is soft
1533 * real-time, because the weight-raising period is constantly
1534 * restarted on idle-to-busy transitions for these queues, but
1535 * this is already done in bfq_bfqq_handle_idle_busy_switch if
1538 if (bfqd
->low_latency
&&
1539 (old_wr_coeff
== 1 || bfqq
->wr_coeff
== 1 || interactive
))
1540 bfqq
->last_wr_start_finish
= jiffies
;
1543 static struct request
*bfq_find_rq_fmerge(struct bfq_data
*bfqd
,
1545 struct request_queue
*q
)
1547 struct bfq_queue
*bfqq
= bfqd
->bio_bfqq
;
1551 return elv_rb_find(&bfqq
->sort_list
, bio_end_sector(bio
));
1556 static sector_t
get_sdist(sector_t last_pos
, struct request
*rq
)
1559 return abs(blk_rq_pos(rq
) - last_pos
);
1564 #if 0 /* Still not clear if we can do without next two functions */
1565 static void bfq_activate_request(struct request_queue
*q
, struct request
*rq
)
1567 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
1569 bfqd
->rq_in_driver
++;
1572 static void bfq_deactivate_request(struct request_queue
*q
, struct request
*rq
)
1574 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
1576 bfqd
->rq_in_driver
--;
1580 static void bfq_remove_request(struct request_queue
*q
,
1583 struct bfq_queue
*bfqq
= RQ_BFQQ(rq
);
1584 struct bfq_data
*bfqd
= bfqq
->bfqd
;
1585 const int sync
= rq_is_sync(rq
);
1587 if (bfqq
->next_rq
== rq
) {
1588 bfqq
->next_rq
= bfq_find_next_rq(bfqd
, bfqq
, rq
);
1589 bfq_updated_next_req(bfqd
, bfqq
);
1592 if (rq
->queuelist
.prev
!= &rq
->queuelist
)
1593 list_del_init(&rq
->queuelist
);
1594 bfqq
->queued
[sync
]--;
1596 elv_rb_del(&bfqq
->sort_list
, rq
);
1598 elv_rqhash_del(q
, rq
);
1599 if (q
->last_merge
== rq
)
1600 q
->last_merge
= NULL
;
1602 if (RB_EMPTY_ROOT(&bfqq
->sort_list
)) {
1603 bfqq
->next_rq
= NULL
;
1605 if (bfq_bfqq_busy(bfqq
) && bfqq
!= bfqd
->in_service_queue
) {
1606 bfq_del_bfqq_busy(bfqd
, bfqq
, false);
1608 * bfqq emptied. In normal operation, when
1609 * bfqq is empty, bfqq->entity.service and
1610 * bfqq->entity.budget must contain,
1611 * respectively, the service received and the
1612 * budget used last time bfqq emptied. These
1613 * facts do not hold in this case, as at least
1614 * this last removal occurred while bfqq is
1615 * not in service. To avoid inconsistencies,
1616 * reset both bfqq->entity.service and
1617 * bfqq->entity.budget, if bfqq has still a
1618 * process that may issue I/O requests to it.
1620 bfqq
->entity
.budget
= bfqq
->entity
.service
= 0;
1624 * Remove queue from request-position tree as it is empty.
1626 if (bfqq
->pos_root
) {
1627 rb_erase(&bfqq
->pos_node
, bfqq
->pos_root
);
1628 bfqq
->pos_root
= NULL
;
1632 if (rq
->cmd_flags
& REQ_META
)
1633 bfqq
->meta_pending
--;
1637 static bool bfq_bio_merge(struct blk_mq_hw_ctx
*hctx
, struct bio
*bio
)
1639 struct request_queue
*q
= hctx
->queue
;
1640 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
1641 struct request
*free
= NULL
;
1643 * bfq_bic_lookup grabs the queue_lock: invoke it now and
1644 * store its return value for later use, to avoid nesting
1645 * queue_lock inside the bfqd->lock. We assume that the bic
1646 * returned by bfq_bic_lookup does not go away before
1647 * bfqd->lock is taken.
1649 struct bfq_io_cq
*bic
= bfq_bic_lookup(bfqd
, current
->io_context
, q
);
1652 spin_lock_irq(&bfqd
->lock
);
1655 bfqd
->bio_bfqq
= bic_to_bfqq(bic
, op_is_sync(bio
->bi_opf
));
1657 bfqd
->bio_bfqq
= NULL
;
1658 bfqd
->bio_bic
= bic
;
1660 ret
= blk_mq_sched_try_merge(q
, bio
, &free
);
1663 blk_mq_free_request(free
);
1664 spin_unlock_irq(&bfqd
->lock
);
1669 static int bfq_request_merge(struct request_queue
*q
, struct request
**req
,
1672 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
1673 struct request
*__rq
;
1675 __rq
= bfq_find_rq_fmerge(bfqd
, bio
, q
);
1676 if (__rq
&& elv_bio_merge_ok(__rq
, bio
)) {
1678 return ELEVATOR_FRONT_MERGE
;
1681 return ELEVATOR_NO_MERGE
;
1684 static void bfq_request_merged(struct request_queue
*q
, struct request
*req
,
1685 enum elv_merge type
)
1687 if (type
== ELEVATOR_FRONT_MERGE
&&
1688 rb_prev(&req
->rb_node
) &&
1690 blk_rq_pos(container_of(rb_prev(&req
->rb_node
),
1691 struct request
, rb_node
))) {
1692 struct bfq_queue
*bfqq
= RQ_BFQQ(req
);
1693 struct bfq_data
*bfqd
= bfqq
->bfqd
;
1694 struct request
*prev
, *next_rq
;
1696 /* Reposition request in its sort_list */
1697 elv_rb_del(&bfqq
->sort_list
, req
);
1698 elv_rb_add(&bfqq
->sort_list
, req
);
1700 /* Choose next request to be served for bfqq */
1701 prev
= bfqq
->next_rq
;
1702 next_rq
= bfq_choose_req(bfqd
, bfqq
->next_rq
, req
,
1703 bfqd
->last_position
);
1704 bfqq
->next_rq
= next_rq
;
1706 * If next_rq changes, update both the queue's budget to
1707 * fit the new request and the queue's position in its
1710 if (prev
!= bfqq
->next_rq
) {
1711 bfq_updated_next_req(bfqd
, bfqq
);
1712 bfq_pos_tree_add_move(bfqd
, bfqq
);
1717 static void bfq_requests_merged(struct request_queue
*q
, struct request
*rq
,
1718 struct request
*next
)
1720 struct bfq_queue
*bfqq
= RQ_BFQQ(rq
), *next_bfqq
= RQ_BFQQ(next
);
1722 if (!RB_EMPTY_NODE(&rq
->rb_node
))
1726 * If next and rq belong to the same bfq_queue and next is older
1727 * than rq, then reposition rq in the fifo (by substituting next
1728 * with rq). Otherwise, if next and rq belong to different
1729 * bfq_queues, never reposition rq: in fact, we would have to
1730 * reposition it with respect to next's position in its own fifo,
1731 * which would most certainly be too expensive with respect to
1734 if (bfqq
== next_bfqq
&&
1735 !list_empty(&rq
->queuelist
) && !list_empty(&next
->queuelist
) &&
1736 next
->fifo_time
< rq
->fifo_time
) {
1737 list_del_init(&rq
->queuelist
);
1738 list_replace_init(&next
->queuelist
, &rq
->queuelist
);
1739 rq
->fifo_time
= next
->fifo_time
;
1742 if (bfqq
->next_rq
== next
)
1745 bfq_remove_request(q
, next
);
1746 bfqg_stats_update_io_remove(bfqq_group(bfqq
), next
->cmd_flags
);
1749 bfqg_stats_update_io_merged(bfqq_group(bfqq
), next
->cmd_flags
);
1752 /* Must be called with bfqq != NULL */
1753 static void bfq_bfqq_end_wr(struct bfq_queue
*bfqq
)
1755 if (bfq_bfqq_busy(bfqq
))
1756 bfqq
->bfqd
->wr_busy_queues
--;
1758 bfqq
->wr_cur_max_time
= 0;
1759 bfqq
->last_wr_start_finish
= jiffies
;
1761 * Trigger a weight change on the next invocation of
1762 * __bfq_entity_update_weight_prio.
1764 bfqq
->entity
.prio_changed
= 1;
1767 void bfq_end_wr_async_queues(struct bfq_data
*bfqd
,
1768 struct bfq_group
*bfqg
)
1772 for (i
= 0; i
< 2; i
++)
1773 for (j
= 0; j
< IOPRIO_BE_NR
; j
++)
1774 if (bfqg
->async_bfqq
[i
][j
])
1775 bfq_bfqq_end_wr(bfqg
->async_bfqq
[i
][j
]);
1776 if (bfqg
->async_idle_bfqq
)
1777 bfq_bfqq_end_wr(bfqg
->async_idle_bfqq
);
1780 static void bfq_end_wr(struct bfq_data
*bfqd
)
1782 struct bfq_queue
*bfqq
;
1784 spin_lock_irq(&bfqd
->lock
);
1786 list_for_each_entry(bfqq
, &bfqd
->active_list
, bfqq_list
)
1787 bfq_bfqq_end_wr(bfqq
);
1788 list_for_each_entry(bfqq
, &bfqd
->idle_list
, bfqq_list
)
1789 bfq_bfqq_end_wr(bfqq
);
1790 bfq_end_wr_async(bfqd
);
1792 spin_unlock_irq(&bfqd
->lock
);
1795 static sector_t
bfq_io_struct_pos(void *io_struct
, bool request
)
1798 return blk_rq_pos(io_struct
);
1800 return ((struct bio
*)io_struct
)->bi_iter
.bi_sector
;
1803 static int bfq_rq_close_to_sector(void *io_struct
, bool request
,
1806 return abs(bfq_io_struct_pos(io_struct
, request
) - sector
) <=
1810 static struct bfq_queue
*bfqq_find_close(struct bfq_data
*bfqd
,
1811 struct bfq_queue
*bfqq
,
1814 struct rb_root
*root
= &bfq_bfqq_to_bfqg(bfqq
)->rq_pos_tree
;
1815 struct rb_node
*parent
, *node
;
1816 struct bfq_queue
*__bfqq
;
1818 if (RB_EMPTY_ROOT(root
))
1822 * First, if we find a request starting at the end of the last
1823 * request, choose it.
1825 __bfqq
= bfq_rq_pos_tree_lookup(bfqd
, root
, sector
, &parent
, NULL
);
1830 * If the exact sector wasn't found, the parent of the NULL leaf
1831 * will contain the closest sector (rq_pos_tree sorted by
1832 * next_request position).
1834 __bfqq
= rb_entry(parent
, struct bfq_queue
, pos_node
);
1835 if (bfq_rq_close_to_sector(__bfqq
->next_rq
, true, sector
))
1838 if (blk_rq_pos(__bfqq
->next_rq
) < sector
)
1839 node
= rb_next(&__bfqq
->pos_node
);
1841 node
= rb_prev(&__bfqq
->pos_node
);
1845 __bfqq
= rb_entry(node
, struct bfq_queue
, pos_node
);
1846 if (bfq_rq_close_to_sector(__bfqq
->next_rq
, true, sector
))
1852 static struct bfq_queue
*bfq_find_close_cooperator(struct bfq_data
*bfqd
,
1853 struct bfq_queue
*cur_bfqq
,
1856 struct bfq_queue
*bfqq
;
1859 * We shall notice if some of the queues are cooperating,
1860 * e.g., working closely on the same area of the device. In
1861 * that case, we can group them together and: 1) don't waste
1862 * time idling, and 2) serve the union of their requests in
1863 * the best possible order for throughput.
1865 bfqq
= bfqq_find_close(bfqd
, cur_bfqq
, sector
);
1866 if (!bfqq
|| bfqq
== cur_bfqq
)
1872 static struct bfq_queue
*
1873 bfq_setup_merge(struct bfq_queue
*bfqq
, struct bfq_queue
*new_bfqq
)
1875 int process_refs
, new_process_refs
;
1876 struct bfq_queue
*__bfqq
;
1879 * If there are no process references on the new_bfqq, then it is
1880 * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain
1881 * may have dropped their last reference (not just their last process
1884 if (!bfqq_process_refs(new_bfqq
))
1887 /* Avoid a circular list and skip interim queue merges. */
1888 while ((__bfqq
= new_bfqq
->new_bfqq
)) {
1894 process_refs
= bfqq_process_refs(bfqq
);
1895 new_process_refs
= bfqq_process_refs(new_bfqq
);
1897 * If the process for the bfqq has gone away, there is no
1898 * sense in merging the queues.
1900 if (process_refs
== 0 || new_process_refs
== 0)
1903 bfq_log_bfqq(bfqq
->bfqd
, bfqq
, "scheduling merge with queue %d",
1907 * Merging is just a redirection: the requests of the process
1908 * owning one of the two queues are redirected to the other queue.
1909 * The latter queue, in its turn, is set as shared if this is the
1910 * first time that the requests of some process are redirected to
1913 * We redirect bfqq to new_bfqq and not the opposite, because
1914 * we are in the context of the process owning bfqq, thus we
1915 * have the io_cq of this process. So we can immediately
1916 * configure this io_cq to redirect the requests of the
1917 * process to new_bfqq. In contrast, the io_cq of new_bfqq is
1918 * not available any more (new_bfqq->bic == NULL).
1920 * Anyway, even in case new_bfqq coincides with the in-service
1921 * queue, redirecting requests the in-service queue is the
1922 * best option, as we feed the in-service queue with new
1923 * requests close to the last request served and, by doing so,
1924 * are likely to increase the throughput.
1926 bfqq
->new_bfqq
= new_bfqq
;
1927 new_bfqq
->ref
+= process_refs
;
1931 static bool bfq_may_be_close_cooperator(struct bfq_queue
*bfqq
,
1932 struct bfq_queue
*new_bfqq
)
1934 if (bfq_class_idle(bfqq
) || bfq_class_idle(new_bfqq
) ||
1935 (bfqq
->ioprio_class
!= new_bfqq
->ioprio_class
))
1939 * If either of the queues has already been detected as seeky,
1940 * then merging it with the other queue is unlikely to lead to
1943 if (BFQQ_SEEKY(bfqq
) || BFQQ_SEEKY(new_bfqq
))
1947 * Interleaved I/O is known to be done by (some) applications
1948 * only for reads, so it does not make sense to merge async
1951 if (!bfq_bfqq_sync(bfqq
) || !bfq_bfqq_sync(new_bfqq
))
1958 * If this function returns true, then bfqq cannot be merged. The idea
1959 * is that true cooperation happens very early after processes start
1960 * to do I/O. Usually, late cooperations are just accidental false
1961 * positives. In case bfqq is weight-raised, such false positives
1962 * would evidently degrade latency guarantees for bfqq.
1964 static bool wr_from_too_long(struct bfq_queue
*bfqq
)
1966 return bfqq
->wr_coeff
> 1 &&
1967 time_is_before_jiffies(bfqq
->last_wr_start_finish
+
1968 msecs_to_jiffies(100));
1972 * Attempt to schedule a merge of bfqq with the currently in-service
1973 * queue or with a close queue among the scheduled queues. Return
1974 * NULL if no merge was scheduled, a pointer to the shared bfq_queue
1975 * structure otherwise.
1977 * The OOM queue is not allowed to participate to cooperation: in fact, since
1978 * the requests temporarily redirected to the OOM queue could be redirected
1979 * again to dedicated queues at any time, the state needed to correctly
1980 * handle merging with the OOM queue would be quite complex and expensive
1981 * to maintain. Besides, in such a critical condition as an out of memory,
1982 * the benefits of queue merging may be little relevant, or even negligible.
1984 * Weight-raised queues can be merged only if their weight-raising
1985 * period has just started. In fact cooperating processes are usually
1986 * started together. Thus, with this filter we avoid false positives
1987 * that would jeopardize low-latency guarantees.
1989 * WARNING: queue merging may impair fairness among non-weight raised
1990 * queues, for at least two reasons: 1) the original weight of a
1991 * merged queue may change during the merged state, 2) even being the
1992 * weight the same, a merged queue may be bloated with many more
1993 * requests than the ones produced by its originally-associated
1996 static struct bfq_queue
*
1997 bfq_setup_cooperator(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
1998 void *io_struct
, bool request
)
2000 struct bfq_queue
*in_service_bfqq
, *new_bfqq
;
2003 return bfqq
->new_bfqq
;
2006 wr_from_too_long(bfqq
) ||
2007 unlikely(bfqq
== &bfqd
->oom_bfqq
))
2010 /* If there is only one backlogged queue, don't search. */
2011 if (bfqd
->busy_queues
== 1)
2014 in_service_bfqq
= bfqd
->in_service_queue
;
2016 if (!in_service_bfqq
|| in_service_bfqq
== bfqq
2017 || wr_from_too_long(in_service_bfqq
) ||
2018 unlikely(in_service_bfqq
== &bfqd
->oom_bfqq
))
2019 goto check_scheduled
;
2021 if (bfq_rq_close_to_sector(io_struct
, request
, bfqd
->last_position
) &&
2022 bfqq
->entity
.parent
== in_service_bfqq
->entity
.parent
&&
2023 bfq_may_be_close_cooperator(bfqq
, in_service_bfqq
)) {
2024 new_bfqq
= bfq_setup_merge(bfqq
, in_service_bfqq
);
2029 * Check whether there is a cooperator among currently scheduled
2030 * queues. The only thing we need is that the bio/request is not
2031 * NULL, as we need it to establish whether a cooperator exists.
2034 new_bfqq
= bfq_find_close_cooperator(bfqd
, bfqq
,
2035 bfq_io_struct_pos(io_struct
, request
));
2037 if (new_bfqq
&& !wr_from_too_long(new_bfqq
) &&
2038 likely(new_bfqq
!= &bfqd
->oom_bfqq
) &&
2039 bfq_may_be_close_cooperator(bfqq
, new_bfqq
))
2040 return bfq_setup_merge(bfqq
, new_bfqq
);
2045 static void bfq_bfqq_save_state(struct bfq_queue
*bfqq
)
2047 struct bfq_io_cq
*bic
= bfqq
->bic
;
2050 * If !bfqq->bic, the queue is already shared or its requests
2051 * have already been redirected to a shared queue; both idle window
2052 * and weight raising state have already been saved. Do nothing.
2057 bic
->saved_ttime
= bfqq
->ttime
;
2058 bic
->saved_has_short_ttime
= bfq_bfqq_has_short_ttime(bfqq
);
2059 bic
->saved_IO_bound
= bfq_bfqq_IO_bound(bfqq
);
2060 bic
->saved_in_large_burst
= bfq_bfqq_in_large_burst(bfqq
);
2061 bic
->was_in_burst_list
= !hlist_unhashed(&bfqq
->burst_list_node
);
2062 if (unlikely(bfq_bfqq_just_created(bfqq
) &&
2063 !bfq_bfqq_in_large_burst(bfqq
))) {
2065 * bfqq being merged right after being created: bfqq
2066 * would have deserved interactive weight raising, but
2067 * did not make it to be set in a weight-raised state,
2068 * because of this early merge. Store directly the
2069 * weight-raising state that would have been assigned
2070 * to bfqq, so that to avoid that bfqq unjustly fails
2071 * to enjoy weight raising if split soon.
2073 bic
->saved_wr_coeff
= bfqq
->bfqd
->bfq_wr_coeff
;
2074 bic
->saved_wr_cur_max_time
= bfq_wr_duration(bfqq
->bfqd
);
2075 bic
->saved_last_wr_start_finish
= jiffies
;
2077 bic
->saved_wr_coeff
= bfqq
->wr_coeff
;
2078 bic
->saved_wr_start_at_switch_to_srt
=
2079 bfqq
->wr_start_at_switch_to_srt
;
2080 bic
->saved_last_wr_start_finish
= bfqq
->last_wr_start_finish
;
2081 bic
->saved_wr_cur_max_time
= bfqq
->wr_cur_max_time
;
2086 bfq_merge_bfqqs(struct bfq_data
*bfqd
, struct bfq_io_cq
*bic
,
2087 struct bfq_queue
*bfqq
, struct bfq_queue
*new_bfqq
)
2089 bfq_log_bfqq(bfqd
, bfqq
, "merging with queue %lu",
2090 (unsigned long)new_bfqq
->pid
);
2091 /* Save weight raising and idle window of the merged queues */
2092 bfq_bfqq_save_state(bfqq
);
2093 bfq_bfqq_save_state(new_bfqq
);
2094 if (bfq_bfqq_IO_bound(bfqq
))
2095 bfq_mark_bfqq_IO_bound(new_bfqq
);
2096 bfq_clear_bfqq_IO_bound(bfqq
);
2099 * If bfqq is weight-raised, then let new_bfqq inherit
2100 * weight-raising. To reduce false positives, neglect the case
2101 * where bfqq has just been created, but has not yet made it
2102 * to be weight-raised (which may happen because EQM may merge
2103 * bfqq even before bfq_add_request is executed for the first
2104 * time for bfqq). Handling this case would however be very
2105 * easy, thanks to the flag just_created.
2107 if (new_bfqq
->wr_coeff
== 1 && bfqq
->wr_coeff
> 1) {
2108 new_bfqq
->wr_coeff
= bfqq
->wr_coeff
;
2109 new_bfqq
->wr_cur_max_time
= bfqq
->wr_cur_max_time
;
2110 new_bfqq
->last_wr_start_finish
= bfqq
->last_wr_start_finish
;
2111 new_bfqq
->wr_start_at_switch_to_srt
=
2112 bfqq
->wr_start_at_switch_to_srt
;
2113 if (bfq_bfqq_busy(new_bfqq
))
2114 bfqd
->wr_busy_queues
++;
2115 new_bfqq
->entity
.prio_changed
= 1;
2118 if (bfqq
->wr_coeff
> 1) { /* bfqq has given its wr to new_bfqq */
2120 bfqq
->entity
.prio_changed
= 1;
2121 if (bfq_bfqq_busy(bfqq
))
2122 bfqd
->wr_busy_queues
--;
2125 bfq_log_bfqq(bfqd
, new_bfqq
, "merge_bfqqs: wr_busy %d",
2126 bfqd
->wr_busy_queues
);
2129 * Merge queues (that is, let bic redirect its requests to new_bfqq)
2131 bic_set_bfqq(bic
, new_bfqq
, 1);
2132 bfq_mark_bfqq_coop(new_bfqq
);
2134 * new_bfqq now belongs to at least two bics (it is a shared queue):
2135 * set new_bfqq->bic to NULL. bfqq either:
2136 * - does not belong to any bic any more, and hence bfqq->bic must
2137 * be set to NULL, or
2138 * - is a queue whose owning bics have already been redirected to a
2139 * different queue, hence the queue is destined to not belong to
2140 * any bic soon and bfqq->bic is already NULL (therefore the next
2141 * assignment causes no harm).
2143 new_bfqq
->bic
= NULL
;
2145 /* release process reference to bfqq */
2146 bfq_put_queue(bfqq
);
2149 static bool bfq_allow_bio_merge(struct request_queue
*q
, struct request
*rq
,
2152 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
2153 bool is_sync
= op_is_sync(bio
->bi_opf
);
2154 struct bfq_queue
*bfqq
= bfqd
->bio_bfqq
, *new_bfqq
;
2157 * Disallow merge of a sync bio into an async request.
2159 if (is_sync
&& !rq_is_sync(rq
))
2163 * Lookup the bfqq that this bio will be queued with. Allow
2164 * merge only if rq is queued there.
2170 * We take advantage of this function to perform an early merge
2171 * of the queues of possible cooperating processes.
2173 new_bfqq
= bfq_setup_cooperator(bfqd
, bfqq
, bio
, false);
2176 * bic still points to bfqq, then it has not yet been
2177 * redirected to some other bfq_queue, and a queue
2178 * merge beween bfqq and new_bfqq can be safely
2179 * fulfillled, i.e., bic can be redirected to new_bfqq
2180 * and bfqq can be put.
2182 bfq_merge_bfqqs(bfqd
, bfqd
->bio_bic
, bfqq
,
2185 * If we get here, bio will be queued into new_queue,
2186 * so use new_bfqq to decide whether bio and rq can be
2192 * Change also bqfd->bio_bfqq, as
2193 * bfqd->bio_bic now points to new_bfqq, and
2194 * this function may be invoked again (and then may
2195 * use again bqfd->bio_bfqq).
2197 bfqd
->bio_bfqq
= bfqq
;
2200 return bfqq
== RQ_BFQQ(rq
);
2204 * Set the maximum time for the in-service queue to consume its
2205 * budget. This prevents seeky processes from lowering the throughput.
2206 * In practice, a time-slice service scheme is used with seeky
2209 static void bfq_set_budget_timeout(struct bfq_data
*bfqd
,
2210 struct bfq_queue
*bfqq
)
2212 unsigned int timeout_coeff
;
2214 if (bfqq
->wr_cur_max_time
== bfqd
->bfq_wr_rt_max_time
)
2217 timeout_coeff
= bfqq
->entity
.weight
/ bfqq
->entity
.orig_weight
;
2219 bfqd
->last_budget_start
= ktime_get();
2221 bfqq
->budget_timeout
= jiffies
+
2222 bfqd
->bfq_timeout
* timeout_coeff
;
2225 static void __bfq_set_in_service_queue(struct bfq_data
*bfqd
,
2226 struct bfq_queue
*bfqq
)
2229 bfq_clear_bfqq_fifo_expire(bfqq
);
2231 bfqd
->budgets_assigned
= (bfqd
->budgets_assigned
* 7 + 256) / 8;
2233 if (time_is_before_jiffies(bfqq
->last_wr_start_finish
) &&
2234 bfqq
->wr_coeff
> 1 &&
2235 bfqq
->wr_cur_max_time
== bfqd
->bfq_wr_rt_max_time
&&
2236 time_is_before_jiffies(bfqq
->budget_timeout
)) {
2238 * For soft real-time queues, move the start
2239 * of the weight-raising period forward by the
2240 * time the queue has not received any
2241 * service. Otherwise, a relatively long
2242 * service delay is likely to cause the
2243 * weight-raising period of the queue to end,
2244 * because of the short duration of the
2245 * weight-raising period of a soft real-time
2246 * queue. It is worth noting that this move
2247 * is not so dangerous for the other queues,
2248 * because soft real-time queues are not
2251 * To not add a further variable, we use the
2252 * overloaded field budget_timeout to
2253 * determine for how long the queue has not
2254 * received service, i.e., how much time has
2255 * elapsed since the queue expired. However,
2256 * this is a little imprecise, because
2257 * budget_timeout is set to jiffies if bfqq
2258 * not only expires, but also remains with no
2261 if (time_after(bfqq
->budget_timeout
,
2262 bfqq
->last_wr_start_finish
))
2263 bfqq
->last_wr_start_finish
+=
2264 jiffies
- bfqq
->budget_timeout
;
2266 bfqq
->last_wr_start_finish
= jiffies
;
2269 bfq_set_budget_timeout(bfqd
, bfqq
);
2270 bfq_log_bfqq(bfqd
, bfqq
,
2271 "set_in_service_queue, cur-budget = %d",
2272 bfqq
->entity
.budget
);
2275 bfqd
->in_service_queue
= bfqq
;
2279 * Get and set a new queue for service.
2281 static struct bfq_queue
*bfq_set_in_service_queue(struct bfq_data
*bfqd
)
2283 struct bfq_queue
*bfqq
= bfq_get_next_queue(bfqd
);
2285 __bfq_set_in_service_queue(bfqd
, bfqq
);
2289 static void bfq_arm_slice_timer(struct bfq_data
*bfqd
)
2291 struct bfq_queue
*bfqq
= bfqd
->in_service_queue
;
2294 bfq_mark_bfqq_wait_request(bfqq
);
2297 * We don't want to idle for seeks, but we do want to allow
2298 * fair distribution of slice time for a process doing back-to-back
2299 * seeks. So allow a little bit of time for him to submit a new rq.
2301 sl
= bfqd
->bfq_slice_idle
;
2303 * Unless the queue is being weight-raised or the scenario is
2304 * asymmetric, grant only minimum idle time if the queue
2305 * is seeky. A long idling is preserved for a weight-raised
2306 * queue, or, more in general, in an asymmetric scenario,
2307 * because a long idling is needed for guaranteeing to a queue
2308 * its reserved share of the throughput (in particular, it is
2309 * needed if the queue has a higher weight than some other
2312 if (BFQQ_SEEKY(bfqq
) && bfqq
->wr_coeff
== 1 &&
2313 bfq_symmetric_scenario(bfqd
))
2314 sl
= min_t(u64
, sl
, BFQ_MIN_TT
);
2316 bfqd
->last_idling_start
= ktime_get();
2317 hrtimer_start(&bfqd
->idle_slice_timer
, ns_to_ktime(sl
),
2319 bfqg_stats_set_start_idle_time(bfqq_group(bfqq
));
2323 * In autotuning mode, max_budget is dynamically recomputed as the
2324 * amount of sectors transferred in timeout at the estimated peak
2325 * rate. This enables BFQ to utilize a full timeslice with a full
2326 * budget, even if the in-service queue is served at peak rate. And
2327 * this maximises throughput with sequential workloads.
2329 static unsigned long bfq_calc_max_budget(struct bfq_data
*bfqd
)
2331 return (u64
)bfqd
->peak_rate
* USEC_PER_MSEC
*
2332 jiffies_to_msecs(bfqd
->bfq_timeout
)>>BFQ_RATE_SHIFT
;
2336 * Update parameters related to throughput and responsiveness, as a
2337 * function of the estimated peak rate. See comments on
2338 * bfq_calc_max_budget(), and on T_slow and T_fast arrays.
2340 static void update_thr_responsiveness_params(struct bfq_data
*bfqd
)
2342 int dev_type
= blk_queue_nonrot(bfqd
->queue
);
2344 if (bfqd
->bfq_user_max_budget
== 0)
2345 bfqd
->bfq_max_budget
=
2346 bfq_calc_max_budget(bfqd
);
2348 if (bfqd
->device_speed
== BFQ_BFQD_FAST
&&
2349 bfqd
->peak_rate
< device_speed_thresh
[dev_type
]) {
2350 bfqd
->device_speed
= BFQ_BFQD_SLOW
;
2351 bfqd
->RT_prod
= R_slow
[dev_type
] *
2353 } else if (bfqd
->device_speed
== BFQ_BFQD_SLOW
&&
2354 bfqd
->peak_rate
> device_speed_thresh
[dev_type
]) {
2355 bfqd
->device_speed
= BFQ_BFQD_FAST
;
2356 bfqd
->RT_prod
= R_fast
[dev_type
] *
2361 "dev_type %s dev_speed_class = %s (%llu sects/sec), thresh %llu setcs/sec",
2362 dev_type
== 0 ? "ROT" : "NONROT",
2363 bfqd
->device_speed
== BFQ_BFQD_FAST
? "FAST" : "SLOW",
2364 bfqd
->device_speed
== BFQ_BFQD_FAST
?
2365 (USEC_PER_SEC
*(u64
)R_fast
[dev_type
])>>BFQ_RATE_SHIFT
:
2366 (USEC_PER_SEC
*(u64
)R_slow
[dev_type
])>>BFQ_RATE_SHIFT
,
2367 (USEC_PER_SEC
*(u64
)device_speed_thresh
[dev_type
])>>
2371 static void bfq_reset_rate_computation(struct bfq_data
*bfqd
,
2374 if (rq
!= NULL
) { /* new rq dispatch now, reset accordingly */
2375 bfqd
->last_dispatch
= bfqd
->first_dispatch
= ktime_get_ns();
2376 bfqd
->peak_rate_samples
= 1;
2377 bfqd
->sequential_samples
= 0;
2378 bfqd
->tot_sectors_dispatched
= bfqd
->last_rq_max_size
=
2380 } else /* no new rq dispatched, just reset the number of samples */
2381 bfqd
->peak_rate_samples
= 0; /* full re-init on next disp. */
2384 "reset_rate_computation at end, sample %u/%u tot_sects %llu",
2385 bfqd
->peak_rate_samples
, bfqd
->sequential_samples
,
2386 bfqd
->tot_sectors_dispatched
);
2389 static void bfq_update_rate_reset(struct bfq_data
*bfqd
, struct request
*rq
)
2391 u32 rate
, weight
, divisor
;
2394 * For the convergence property to hold (see comments on
2395 * bfq_update_peak_rate()) and for the assessment to be
2396 * reliable, a minimum number of samples must be present, and
2397 * a minimum amount of time must have elapsed. If not so, do
2398 * not compute new rate. Just reset parameters, to get ready
2399 * for a new evaluation attempt.
2401 if (bfqd
->peak_rate_samples
< BFQ_RATE_MIN_SAMPLES
||
2402 bfqd
->delta_from_first
< BFQ_RATE_MIN_INTERVAL
)
2403 goto reset_computation
;
2406 * If a new request completion has occurred after last
2407 * dispatch, then, to approximate the rate at which requests
2408 * have been served by the device, it is more precise to
2409 * extend the observation interval to the last completion.
2411 bfqd
->delta_from_first
=
2412 max_t(u64
, bfqd
->delta_from_first
,
2413 bfqd
->last_completion
- bfqd
->first_dispatch
);
2416 * Rate computed in sects/usec, and not sects/nsec, for
2419 rate
= div64_ul(bfqd
->tot_sectors_dispatched
<<BFQ_RATE_SHIFT
,
2420 div_u64(bfqd
->delta_from_first
, NSEC_PER_USEC
));
2423 * Peak rate not updated if:
2424 * - the percentage of sequential dispatches is below 3/4 of the
2425 * total, and rate is below the current estimated peak rate
2426 * - rate is unreasonably high (> 20M sectors/sec)
2428 if ((bfqd
->sequential_samples
< (3 * bfqd
->peak_rate_samples
)>>2 &&
2429 rate
<= bfqd
->peak_rate
) ||
2430 rate
> 20<<BFQ_RATE_SHIFT
)
2431 goto reset_computation
;
2434 * We have to update the peak rate, at last! To this purpose,
2435 * we use a low-pass filter. We compute the smoothing constant
2436 * of the filter as a function of the 'weight' of the new
2439 * As can be seen in next formulas, we define this weight as a
2440 * quantity proportional to how sequential the workload is,
2441 * and to how long the observation time interval is.
2443 * The weight runs from 0 to 8. The maximum value of the
2444 * weight, 8, yields the minimum value for the smoothing
2445 * constant. At this minimum value for the smoothing constant,
2446 * the measured rate contributes for half of the next value of
2447 * the estimated peak rate.
2449 * So, the first step is to compute the weight as a function
2450 * of how sequential the workload is. Note that the weight
2451 * cannot reach 9, because bfqd->sequential_samples cannot
2452 * become equal to bfqd->peak_rate_samples, which, in its
2453 * turn, holds true because bfqd->sequential_samples is not
2454 * incremented for the first sample.
2456 weight
= (9 * bfqd
->sequential_samples
) / bfqd
->peak_rate_samples
;
2459 * Second step: further refine the weight as a function of the
2460 * duration of the observation interval.
2462 weight
= min_t(u32
, 8,
2463 div_u64(weight
* bfqd
->delta_from_first
,
2464 BFQ_RATE_REF_INTERVAL
));
2467 * Divisor ranging from 10, for minimum weight, to 2, for
2470 divisor
= 10 - weight
;
2473 * Finally, update peak rate:
2475 * peak_rate = peak_rate * (divisor-1) / divisor + rate / divisor
2477 bfqd
->peak_rate
*= divisor
-1;
2478 bfqd
->peak_rate
/= divisor
;
2479 rate
/= divisor
; /* smoothing constant alpha = 1/divisor */
2481 bfqd
->peak_rate
+= rate
;
2482 update_thr_responsiveness_params(bfqd
);
2485 bfq_reset_rate_computation(bfqd
, rq
);
2489 * Update the read/write peak rate (the main quantity used for
2490 * auto-tuning, see update_thr_responsiveness_params()).
2492 * It is not trivial to estimate the peak rate (correctly): because of
2493 * the presence of sw and hw queues between the scheduler and the
2494 * device components that finally serve I/O requests, it is hard to
2495 * say exactly when a given dispatched request is served inside the
2496 * device, and for how long. As a consequence, it is hard to know
2497 * precisely at what rate a given set of requests is actually served
2500 * On the opposite end, the dispatch time of any request is trivially
2501 * available, and, from this piece of information, the "dispatch rate"
2502 * of requests can be immediately computed. So, the idea in the next
2503 * function is to use what is known, namely request dispatch times
2504 * (plus, when useful, request completion times), to estimate what is
2505 * unknown, namely in-device request service rate.
2507 * The main issue is that, because of the above facts, the rate at
2508 * which a certain set of requests is dispatched over a certain time
2509 * interval can vary greatly with respect to the rate at which the
2510 * same requests are then served. But, since the size of any
2511 * intermediate queue is limited, and the service scheme is lossless
2512 * (no request is silently dropped), the following obvious convergence
2513 * property holds: the number of requests dispatched MUST become
2514 * closer and closer to the number of requests completed as the
2515 * observation interval grows. This is the key property used in
2516 * the next function to estimate the peak service rate as a function
2517 * of the observed dispatch rate. The function assumes to be invoked
2518 * on every request dispatch.
2520 static void bfq_update_peak_rate(struct bfq_data
*bfqd
, struct request
*rq
)
2522 u64 now_ns
= ktime_get_ns();
2524 if (bfqd
->peak_rate_samples
== 0) { /* first dispatch */
2525 bfq_log(bfqd
, "update_peak_rate: goto reset, samples %d",
2526 bfqd
->peak_rate_samples
);
2527 bfq_reset_rate_computation(bfqd
, rq
);
2528 goto update_last_values
; /* will add one sample */
2532 * Device idle for very long: the observation interval lasting
2533 * up to this dispatch cannot be a valid observation interval
2534 * for computing a new peak rate (similarly to the late-
2535 * completion event in bfq_completed_request()). Go to
2536 * update_rate_and_reset to have the following three steps
2538 * - close the observation interval at the last (previous)
2539 * request dispatch or completion
2540 * - compute rate, if possible, for that observation interval
2541 * - start a new observation interval with this dispatch
2543 if (now_ns
- bfqd
->last_dispatch
> 100*NSEC_PER_MSEC
&&
2544 bfqd
->rq_in_driver
== 0)
2545 goto update_rate_and_reset
;
2547 /* Update sampling information */
2548 bfqd
->peak_rate_samples
++;
2550 if ((bfqd
->rq_in_driver
> 0 ||
2551 now_ns
- bfqd
->last_completion
< BFQ_MIN_TT
)
2552 && get_sdist(bfqd
->last_position
, rq
) < BFQQ_SEEK_THR
)
2553 bfqd
->sequential_samples
++;
2555 bfqd
->tot_sectors_dispatched
+= blk_rq_sectors(rq
);
2557 /* Reset max observed rq size every 32 dispatches */
2558 if (likely(bfqd
->peak_rate_samples
% 32))
2559 bfqd
->last_rq_max_size
=
2560 max_t(u32
, blk_rq_sectors(rq
), bfqd
->last_rq_max_size
);
2562 bfqd
->last_rq_max_size
= blk_rq_sectors(rq
);
2564 bfqd
->delta_from_first
= now_ns
- bfqd
->first_dispatch
;
2566 /* Target observation interval not yet reached, go on sampling */
2567 if (bfqd
->delta_from_first
< BFQ_RATE_REF_INTERVAL
)
2568 goto update_last_values
;
2570 update_rate_and_reset
:
2571 bfq_update_rate_reset(bfqd
, rq
);
2573 bfqd
->last_position
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
2574 bfqd
->last_dispatch
= now_ns
;
2578 * Remove request from internal lists.
2580 static void bfq_dispatch_remove(struct request_queue
*q
, struct request
*rq
)
2582 struct bfq_queue
*bfqq
= RQ_BFQQ(rq
);
2585 * For consistency, the next instruction should have been
2586 * executed after removing the request from the queue and
2587 * dispatching it. We execute instead this instruction before
2588 * bfq_remove_request() (and hence introduce a temporary
2589 * inconsistency), for efficiency. In fact, should this
2590 * dispatch occur for a non in-service bfqq, this anticipated
2591 * increment prevents two counters related to bfqq->dispatched
2592 * from risking to be, first, uselessly decremented, and then
2593 * incremented again when the (new) value of bfqq->dispatched
2594 * happens to be taken into account.
2597 bfq_update_peak_rate(q
->elevator
->elevator_data
, rq
);
2599 bfq_remove_request(q
, rq
);
2602 static void __bfq_bfqq_expire(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
2605 * If this bfqq is shared between multiple processes, check
2606 * to make sure that those processes are still issuing I/Os
2607 * within the mean seek distance. If not, it may be time to
2608 * break the queues apart again.
2610 if (bfq_bfqq_coop(bfqq
) && BFQQ_SEEKY(bfqq
))
2611 bfq_mark_bfqq_split_coop(bfqq
);
2613 if (RB_EMPTY_ROOT(&bfqq
->sort_list
)) {
2614 if (bfqq
->dispatched
== 0)
2616 * Overloading budget_timeout field to store
2617 * the time at which the queue remains with no
2618 * backlog and no outstanding request; used by
2619 * the weight-raising mechanism.
2621 bfqq
->budget_timeout
= jiffies
;
2623 bfq_del_bfqq_busy(bfqd
, bfqq
, true);
2625 bfq_requeue_bfqq(bfqd
, bfqq
, true);
2627 * Resort priority tree of potential close cooperators.
2629 bfq_pos_tree_add_move(bfqd
, bfqq
);
2633 * All in-service entities must have been properly deactivated
2634 * or requeued before executing the next function, which
2635 * resets all in-service entites as no more in service.
2637 __bfq_bfqd_reset_in_service(bfqd
);
2641 * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior.
2642 * @bfqd: device data.
2643 * @bfqq: queue to update.
2644 * @reason: reason for expiration.
2646 * Handle the feedback on @bfqq budget at queue expiration.
2647 * See the body for detailed comments.
2649 static void __bfq_bfqq_recalc_budget(struct bfq_data
*bfqd
,
2650 struct bfq_queue
*bfqq
,
2651 enum bfqq_expiration reason
)
2653 struct request
*next_rq
;
2654 int budget
, min_budget
;
2656 min_budget
= bfq_min_budget(bfqd
);
2658 if (bfqq
->wr_coeff
== 1)
2659 budget
= bfqq
->max_budget
;
2661 * Use a constant, low budget for weight-raised queues,
2662 * to help achieve a low latency. Keep it slightly higher
2663 * than the minimum possible budget, to cause a little
2664 * bit fewer expirations.
2666 budget
= 2 * min_budget
;
2668 bfq_log_bfqq(bfqd
, bfqq
, "recalc_budg: last budg %d, budg left %d",
2669 bfqq
->entity
.budget
, bfq_bfqq_budget_left(bfqq
));
2670 bfq_log_bfqq(bfqd
, bfqq
, "recalc_budg: last max_budg %d, min budg %d",
2671 budget
, bfq_min_budget(bfqd
));
2672 bfq_log_bfqq(bfqd
, bfqq
, "recalc_budg: sync %d, seeky %d",
2673 bfq_bfqq_sync(bfqq
), BFQQ_SEEKY(bfqd
->in_service_queue
));
2675 if (bfq_bfqq_sync(bfqq
) && bfqq
->wr_coeff
== 1) {
2678 * Caveat: in all the following cases we trade latency
2681 case BFQQE_TOO_IDLE
:
2683 * This is the only case where we may reduce
2684 * the budget: if there is no request of the
2685 * process still waiting for completion, then
2686 * we assume (tentatively) that the timer has
2687 * expired because the batch of requests of
2688 * the process could have been served with a
2689 * smaller budget. Hence, betting that
2690 * process will behave in the same way when it
2691 * becomes backlogged again, we reduce its
2692 * next budget. As long as we guess right,
2693 * this budget cut reduces the latency
2694 * experienced by the process.
2696 * However, if there are still outstanding
2697 * requests, then the process may have not yet
2698 * issued its next request just because it is
2699 * still waiting for the completion of some of
2700 * the still outstanding ones. So in this
2701 * subcase we do not reduce its budget, on the
2702 * contrary we increase it to possibly boost
2703 * the throughput, as discussed in the
2704 * comments to the BUDGET_TIMEOUT case.
2706 if (bfqq
->dispatched
> 0) /* still outstanding reqs */
2707 budget
= min(budget
* 2, bfqd
->bfq_max_budget
);
2709 if (budget
> 5 * min_budget
)
2710 budget
-= 4 * min_budget
;
2712 budget
= min_budget
;
2715 case BFQQE_BUDGET_TIMEOUT
:
2717 * We double the budget here because it gives
2718 * the chance to boost the throughput if this
2719 * is not a seeky process (and has bumped into
2720 * this timeout because of, e.g., ZBR).
2722 budget
= min(budget
* 2, bfqd
->bfq_max_budget
);
2724 case BFQQE_BUDGET_EXHAUSTED
:
2726 * The process still has backlog, and did not
2727 * let either the budget timeout or the disk
2728 * idling timeout expire. Hence it is not
2729 * seeky, has a short thinktime and may be
2730 * happy with a higher budget too. So
2731 * definitely increase the budget of this good
2732 * candidate to boost the disk throughput.
2734 budget
= min(budget
* 4, bfqd
->bfq_max_budget
);
2736 case BFQQE_NO_MORE_REQUESTS
:
2738 * For queues that expire for this reason, it
2739 * is particularly important to keep the
2740 * budget close to the actual service they
2741 * need. Doing so reduces the timestamp
2742 * misalignment problem described in the
2743 * comments in the body of
2744 * __bfq_activate_entity. In fact, suppose
2745 * that a queue systematically expires for
2746 * BFQQE_NO_MORE_REQUESTS and presents a
2747 * new request in time to enjoy timestamp
2748 * back-shifting. The larger the budget of the
2749 * queue is with respect to the service the
2750 * queue actually requests in each service
2751 * slot, the more times the queue can be
2752 * reactivated with the same virtual finish
2753 * time. It follows that, even if this finish
2754 * time is pushed to the system virtual time
2755 * to reduce the consequent timestamp
2756 * misalignment, the queue unjustly enjoys for
2757 * many re-activations a lower finish time
2758 * than all newly activated queues.
2760 * The service needed by bfqq is measured
2761 * quite precisely by bfqq->entity.service.
2762 * Since bfqq does not enjoy device idling,
2763 * bfqq->entity.service is equal to the number
2764 * of sectors that the process associated with
2765 * bfqq requested to read/write before waiting
2766 * for request completions, or blocking for
2769 budget
= max_t(int, bfqq
->entity
.service
, min_budget
);
2774 } else if (!bfq_bfqq_sync(bfqq
)) {
2776 * Async queues get always the maximum possible
2777 * budget, as for them we do not care about latency
2778 * (in addition, their ability to dispatch is limited
2779 * by the charging factor).
2781 budget
= bfqd
->bfq_max_budget
;
2784 bfqq
->max_budget
= budget
;
2786 if (bfqd
->budgets_assigned
>= bfq_stats_min_budgets
&&
2787 !bfqd
->bfq_user_max_budget
)
2788 bfqq
->max_budget
= min(bfqq
->max_budget
, bfqd
->bfq_max_budget
);
2791 * If there is still backlog, then assign a new budget, making
2792 * sure that it is large enough for the next request. Since
2793 * the finish time of bfqq must be kept in sync with the
2794 * budget, be sure to call __bfq_bfqq_expire() *after* this
2797 * If there is no backlog, then no need to update the budget;
2798 * it will be updated on the arrival of a new request.
2800 next_rq
= bfqq
->next_rq
;
2802 bfqq
->entity
.budget
= max_t(unsigned long, bfqq
->max_budget
,
2803 bfq_serv_to_charge(next_rq
, bfqq
));
2805 bfq_log_bfqq(bfqd
, bfqq
, "head sect: %u, new budget %d",
2806 next_rq
? blk_rq_sectors(next_rq
) : 0,
2807 bfqq
->entity
.budget
);
2811 * Return true if the process associated with bfqq is "slow". The slow
2812 * flag is used, in addition to the budget timeout, to reduce the
2813 * amount of service provided to seeky processes, and thus reduce
2814 * their chances to lower the throughput. More details in the comments
2815 * on the function bfq_bfqq_expire().
2817 * An important observation is in order: as discussed in the comments
2818 * on the function bfq_update_peak_rate(), with devices with internal
2819 * queues, it is hard if ever possible to know when and for how long
2820 * an I/O request is processed by the device (apart from the trivial
2821 * I/O pattern where a new request is dispatched only after the
2822 * previous one has been completed). This makes it hard to evaluate
2823 * the real rate at which the I/O requests of each bfq_queue are
2824 * served. In fact, for an I/O scheduler like BFQ, serving a
2825 * bfq_queue means just dispatching its requests during its service
2826 * slot (i.e., until the budget of the queue is exhausted, or the
2827 * queue remains idle, or, finally, a timeout fires). But, during the
2828 * service slot of a bfq_queue, around 100 ms at most, the device may
2829 * be even still processing requests of bfq_queues served in previous
2830 * service slots. On the opposite end, the requests of the in-service
2831 * bfq_queue may be completed after the service slot of the queue
2834 * Anyway, unless more sophisticated solutions are used
2835 * (where possible), the sum of the sizes of the requests dispatched
2836 * during the service slot of a bfq_queue is probably the only
2837 * approximation available for the service received by the bfq_queue
2838 * during its service slot. And this sum is the quantity used in this
2839 * function to evaluate the I/O speed of a process.
2841 static bool bfq_bfqq_is_slow(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
2842 bool compensate
, enum bfqq_expiration reason
,
2843 unsigned long *delta_ms
)
2845 ktime_t delta_ktime
;
2847 bool slow
= BFQQ_SEEKY(bfqq
); /* if delta too short, use seekyness */
2849 if (!bfq_bfqq_sync(bfqq
))
2853 delta_ktime
= bfqd
->last_idling_start
;
2855 delta_ktime
= ktime_get();
2856 delta_ktime
= ktime_sub(delta_ktime
, bfqd
->last_budget_start
);
2857 delta_usecs
= ktime_to_us(delta_ktime
);
2859 /* don't use too short time intervals */
2860 if (delta_usecs
< 1000) {
2861 if (blk_queue_nonrot(bfqd
->queue
))
2863 * give same worst-case guarantees as idling
2866 *delta_ms
= BFQ_MIN_TT
/ NSEC_PER_MSEC
;
2867 else /* charge at least one seek */
2868 *delta_ms
= bfq_slice_idle
/ NSEC_PER_MSEC
;
2873 *delta_ms
= delta_usecs
/ USEC_PER_MSEC
;
2876 * Use only long (> 20ms) intervals to filter out excessive
2877 * spikes in service rate estimation.
2879 if (delta_usecs
> 20000) {
2881 * Caveat for rotational devices: processes doing I/O
2882 * in the slower disk zones tend to be slow(er) even
2883 * if not seeky. In this respect, the estimated peak
2884 * rate is likely to be an average over the disk
2885 * surface. Accordingly, to not be too harsh with
2886 * unlucky processes, a process is deemed slow only if
2887 * its rate has been lower than half of the estimated
2890 slow
= bfqq
->entity
.service
< bfqd
->bfq_max_budget
/ 2;
2893 bfq_log_bfqq(bfqd
, bfqq
, "bfq_bfqq_is_slow: slow %d", slow
);
2899 * To be deemed as soft real-time, an application must meet two
2900 * requirements. First, the application must not require an average
2901 * bandwidth higher than the approximate bandwidth required to playback or
2902 * record a compressed high-definition video.
2903 * The next function is invoked on the completion of the last request of a
2904 * batch, to compute the next-start time instant, soft_rt_next_start, such
2905 * that, if the next request of the application does not arrive before
2906 * soft_rt_next_start, then the above requirement on the bandwidth is met.
2908 * The second requirement is that the request pattern of the application is
2909 * isochronous, i.e., that, after issuing a request or a batch of requests,
2910 * the application stops issuing new requests until all its pending requests
2911 * have been completed. After that, the application may issue a new batch,
2913 * For this reason the next function is invoked to compute
2914 * soft_rt_next_start only for applications that meet this requirement,
2915 * whereas soft_rt_next_start is set to infinity for applications that do
2918 * Unfortunately, even a greedy application may happen to behave in an
2919 * isochronous way if the CPU load is high. In fact, the application may
2920 * stop issuing requests while the CPUs are busy serving other processes,
2921 * then restart, then stop again for a while, and so on. In addition, if
2922 * the disk achieves a low enough throughput with the request pattern
2923 * issued by the application (e.g., because the request pattern is random
2924 * and/or the device is slow), then the application may meet the above
2925 * bandwidth requirement too. To prevent such a greedy application to be
2926 * deemed as soft real-time, a further rule is used in the computation of
2927 * soft_rt_next_start: soft_rt_next_start must be higher than the current
2928 * time plus the maximum time for which the arrival of a request is waited
2929 * for when a sync queue becomes idle, namely bfqd->bfq_slice_idle.
2930 * This filters out greedy applications, as the latter issue instead their
2931 * next request as soon as possible after the last one has been completed
2932 * (in contrast, when a batch of requests is completed, a soft real-time
2933 * application spends some time processing data).
2935 * Unfortunately, the last filter may easily generate false positives if
2936 * only bfqd->bfq_slice_idle is used as a reference time interval and one
2937 * or both the following cases occur:
2938 * 1) HZ is so low that the duration of a jiffy is comparable to or higher
2939 * than bfqd->bfq_slice_idle. This happens, e.g., on slow devices with
2941 * 2) jiffies, instead of increasing at a constant rate, may stop increasing
2942 * for a while, then suddenly 'jump' by several units to recover the lost
2943 * increments. This seems to happen, e.g., inside virtual machines.
2944 * To address this issue, we do not use as a reference time interval just
2945 * bfqd->bfq_slice_idle, but bfqd->bfq_slice_idle plus a few jiffies. In
2946 * particular we add the minimum number of jiffies for which the filter
2947 * seems to be quite precise also in embedded systems and KVM/QEMU virtual
2950 static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data
*bfqd
,
2951 struct bfq_queue
*bfqq
)
2953 return max(bfqq
->last_idle_bklogged
+
2954 HZ
* bfqq
->service_from_backlogged
/
2955 bfqd
->bfq_wr_max_softrt_rate
,
2956 jiffies
+ nsecs_to_jiffies(bfqq
->bfqd
->bfq_slice_idle
) + 4);
2960 * bfq_bfqq_expire - expire a queue.
2961 * @bfqd: device owning the queue.
2962 * @bfqq: the queue to expire.
2963 * @compensate: if true, compensate for the time spent idling.
2964 * @reason: the reason causing the expiration.
2966 * If the process associated with bfqq does slow I/O (e.g., because it
2967 * issues random requests), we charge bfqq with the time it has been
2968 * in service instead of the service it has received (see
2969 * bfq_bfqq_charge_time for details on how this goal is achieved). As
2970 * a consequence, bfqq will typically get higher timestamps upon
2971 * reactivation, and hence it will be rescheduled as if it had
2972 * received more service than what it has actually received. In the
2973 * end, bfqq receives less service in proportion to how slowly its
2974 * associated process consumes its budgets (and hence how seriously it
2975 * tends to lower the throughput). In addition, this time-charging
2976 * strategy guarantees time fairness among slow processes. In
2977 * contrast, if the process associated with bfqq is not slow, we
2978 * charge bfqq exactly with the service it has received.
2980 * Charging time to the first type of queues and the exact service to
2981 * the other has the effect of using the WF2Q+ policy to schedule the
2982 * former on a timeslice basis, without violating service domain
2983 * guarantees among the latter.
2985 void bfq_bfqq_expire(struct bfq_data
*bfqd
,
2986 struct bfq_queue
*bfqq
,
2988 enum bfqq_expiration reason
)
2991 unsigned long delta
= 0;
2992 struct bfq_entity
*entity
= &bfqq
->entity
;
2996 * Check whether the process is slow (see bfq_bfqq_is_slow).
2998 slow
= bfq_bfqq_is_slow(bfqd
, bfqq
, compensate
, reason
, &delta
);
3001 * Increase service_from_backlogged before next statement,
3002 * because the possible next invocation of
3003 * bfq_bfqq_charge_time would likely inflate
3004 * entity->service. In contrast, service_from_backlogged must
3005 * contain real service, to enable the soft real-time
3006 * heuristic to correctly compute the bandwidth consumed by
3009 bfqq
->service_from_backlogged
+= entity
->service
;
3012 * As above explained, charge slow (typically seeky) and
3013 * timed-out queues with the time and not the service
3014 * received, to favor sequential workloads.
3016 * Processes doing I/O in the slower disk zones will tend to
3017 * be slow(er) even if not seeky. Therefore, since the
3018 * estimated peak rate is actually an average over the disk
3019 * surface, these processes may timeout just for bad luck. To
3020 * avoid punishing them, do not charge time to processes that
3021 * succeeded in consuming at least 2/3 of their budget. This
3022 * allows BFQ to preserve enough elasticity to still perform
3023 * bandwidth, and not time, distribution with little unlucky
3024 * or quasi-sequential processes.
3026 if (bfqq
->wr_coeff
== 1 &&
3028 (reason
== BFQQE_BUDGET_TIMEOUT
&&
3029 bfq_bfqq_budget_left(bfqq
) >= entity
->budget
/ 3)))
3030 bfq_bfqq_charge_time(bfqd
, bfqq
, delta
);
3032 if (reason
== BFQQE_TOO_IDLE
&&
3033 entity
->service
<= 2 * entity
->budget
/ 10)
3034 bfq_clear_bfqq_IO_bound(bfqq
);
3036 if (bfqd
->low_latency
&& bfqq
->wr_coeff
== 1)
3037 bfqq
->last_wr_start_finish
= jiffies
;
3039 if (bfqd
->low_latency
&& bfqd
->bfq_wr_max_softrt_rate
> 0 &&
3040 RB_EMPTY_ROOT(&bfqq
->sort_list
)) {
3042 * If we get here, and there are no outstanding
3043 * requests, then the request pattern is isochronous
3044 * (see the comments on the function
3045 * bfq_bfqq_softrt_next_start()). Thus we can compute
3046 * soft_rt_next_start. If, instead, the queue still
3047 * has outstanding requests, then we have to wait for
3048 * the completion of all the outstanding requests to
3049 * discover whether the request pattern is actually
3052 if (bfqq
->dispatched
== 0)
3053 bfqq
->soft_rt_next_start
=
3054 bfq_bfqq_softrt_next_start(bfqd
, bfqq
);
3057 * The application is still waiting for the
3058 * completion of one or more requests:
3059 * prevent it from possibly being incorrectly
3060 * deemed as soft real-time by setting its
3061 * soft_rt_next_start to infinity. In fact,
3062 * without this assignment, the application
3063 * would be incorrectly deemed as soft
3065 * 1) it issued a new request before the
3066 * completion of all its in-flight
3068 * 2) at that time, its soft_rt_next_start
3069 * happened to be in the past.
3071 bfqq
->soft_rt_next_start
=
3072 bfq_greatest_from_now();
3074 * Schedule an update of soft_rt_next_start to when
3075 * the task may be discovered to be isochronous.
3077 bfq_mark_bfqq_softrt_update(bfqq
);
3081 bfq_log_bfqq(bfqd
, bfqq
,
3082 "expire (%d, slow %d, num_disp %d, short_ttime %d)", reason
,
3083 slow
, bfqq
->dispatched
, bfq_bfqq_has_short_ttime(bfqq
));
3086 * Increase, decrease or leave budget unchanged according to
3089 __bfq_bfqq_recalc_budget(bfqd
, bfqq
, reason
);
3091 __bfq_bfqq_expire(bfqd
, bfqq
);
3093 /* mark bfqq as waiting a request only if a bic still points to it */
3094 if (ref
> 1 && !bfq_bfqq_busy(bfqq
) &&
3095 reason
!= BFQQE_BUDGET_TIMEOUT
&&
3096 reason
!= BFQQE_BUDGET_EXHAUSTED
)
3097 bfq_mark_bfqq_non_blocking_wait_rq(bfqq
);
3101 * Budget timeout is not implemented through a dedicated timer, but
3102 * just checked on request arrivals and completions, as well as on
3103 * idle timer expirations.
3105 static bool bfq_bfqq_budget_timeout(struct bfq_queue
*bfqq
)
3107 return time_is_before_eq_jiffies(bfqq
->budget_timeout
);
3111 * If we expire a queue that is actively waiting (i.e., with the
3112 * device idled) for the arrival of a new request, then we may incur
3113 * the timestamp misalignment problem described in the body of the
3114 * function __bfq_activate_entity. Hence we return true only if this
3115 * condition does not hold, or if the queue is slow enough to deserve
3116 * only to be kicked off for preserving a high throughput.
3118 static bool bfq_may_expire_for_budg_timeout(struct bfq_queue
*bfqq
)
3120 bfq_log_bfqq(bfqq
->bfqd
, bfqq
,
3121 "may_budget_timeout: wait_request %d left %d timeout %d",
3122 bfq_bfqq_wait_request(bfqq
),
3123 bfq_bfqq_budget_left(bfqq
) >= bfqq
->entity
.budget
/ 3,
3124 bfq_bfqq_budget_timeout(bfqq
));
3126 return (!bfq_bfqq_wait_request(bfqq
) ||
3127 bfq_bfqq_budget_left(bfqq
) >= bfqq
->entity
.budget
/ 3)
3129 bfq_bfqq_budget_timeout(bfqq
);
3133 * For a queue that becomes empty, device idling is allowed only if
3134 * this function returns true for the queue. As a consequence, since
3135 * device idling plays a critical role in both throughput boosting and
3136 * service guarantees, the return value of this function plays a
3137 * critical role in both these aspects as well.
3139 * In a nutshell, this function returns true only if idling is
3140 * beneficial for throughput or, even if detrimental for throughput,
3141 * idling is however necessary to preserve service guarantees (low
3142 * latency, desired throughput distribution, ...). In particular, on
3143 * NCQ-capable devices, this function tries to return false, so as to
3144 * help keep the drives' internal queues full, whenever this helps the
3145 * device boost the throughput without causing any service-guarantee
3148 * In more detail, the return value of this function is obtained by,
3149 * first, computing a number of boolean variables that take into
3150 * account throughput and service-guarantee issues, and, then,
3151 * combining these variables in a logical expression. Most of the
3152 * issues taken into account are not trivial. We discuss these issues
3153 * individually while introducing the variables.
3155 static bool bfq_bfqq_may_idle(struct bfq_queue
*bfqq
)
3157 struct bfq_data
*bfqd
= bfqq
->bfqd
;
3158 bool rot_without_queueing
=
3159 !blk_queue_nonrot(bfqd
->queue
) && !bfqd
->hw_tag
,
3160 bfqq_sequential_and_IO_bound
,
3161 idling_boosts_thr
, idling_boosts_thr_without_issues
,
3162 idling_needed_for_service_guarantees
,
3163 asymmetric_scenario
;
3165 if (bfqd
->strict_guarantees
)
3169 * Idling is performed only if slice_idle > 0. In addition, we
3172 * (b) bfqq is in the idle io prio class: in this case we do
3173 * not idle because we want to minimize the bandwidth that
3174 * queues in this class can steal to higher-priority queues
3176 if (bfqd
->bfq_slice_idle
== 0 || !bfq_bfqq_sync(bfqq
) ||
3177 bfq_class_idle(bfqq
))
3180 bfqq_sequential_and_IO_bound
= !BFQQ_SEEKY(bfqq
) &&
3181 bfq_bfqq_IO_bound(bfqq
) && bfq_bfqq_has_short_ttime(bfqq
);
3184 * The next variable takes into account the cases where idling
3185 * boosts the throughput.
3187 * The value of the variable is computed considering, first, that
3188 * idling is virtually always beneficial for the throughput if:
3189 * (a) the device is not NCQ-capable and rotational, or
3190 * (b) regardless of the presence of NCQ, the device is rotational and
3191 * the request pattern for bfqq is I/O-bound and sequential, or
3192 * (c) regardless of whether it is rotational, the device is
3193 * not NCQ-capable and the request pattern for bfqq is
3194 * I/O-bound and sequential.
3196 * Secondly, and in contrast to the above item (b), idling an
3197 * NCQ-capable flash-based device would not boost the
3198 * throughput even with sequential I/O; rather it would lower
3199 * the throughput in proportion to how fast the device
3200 * is. Accordingly, the next variable is true if any of the
3201 * above conditions (a), (b) or (c) is true, and, in
3202 * particular, happens to be false if bfqd is an NCQ-capable
3203 * flash-based device.
3205 idling_boosts_thr
= rot_without_queueing
||
3206 ((!blk_queue_nonrot(bfqd
->queue
) || !bfqd
->hw_tag
) &&
3207 bfqq_sequential_and_IO_bound
);
3210 * The value of the next variable,
3211 * idling_boosts_thr_without_issues, is equal to that of
3212 * idling_boosts_thr, unless a special case holds. In this
3213 * special case, described below, idling may cause problems to
3214 * weight-raised queues.
3216 * When the request pool is saturated (e.g., in the presence
3217 * of write hogs), if the processes associated with
3218 * non-weight-raised queues ask for requests at a lower rate,
3219 * then processes associated with weight-raised queues have a
3220 * higher probability to get a request from the pool
3221 * immediately (or at least soon) when they need one. Thus
3222 * they have a higher probability to actually get a fraction
3223 * of the device throughput proportional to their high
3224 * weight. This is especially true with NCQ-capable drives,
3225 * which enqueue several requests in advance, and further
3226 * reorder internally-queued requests.
3228 * For this reason, we force to false the value of
3229 * idling_boosts_thr_without_issues if there are weight-raised
3230 * busy queues. In this case, and if bfqq is not weight-raised,
3231 * this guarantees that the device is not idled for bfqq (if,
3232 * instead, bfqq is weight-raised, then idling will be
3233 * guaranteed by another variable, see below). Combined with
3234 * the timestamping rules of BFQ (see [1] for details), this
3235 * behavior causes bfqq, and hence any sync non-weight-raised
3236 * queue, to get a lower number of requests served, and thus
3237 * to ask for a lower number of requests from the request
3238 * pool, before the busy weight-raised queues get served
3239 * again. This often mitigates starvation problems in the
3240 * presence of heavy write workloads and NCQ, thereby
3241 * guaranteeing a higher application and system responsiveness
3242 * in these hostile scenarios.
3244 idling_boosts_thr_without_issues
= idling_boosts_thr
&&
3245 bfqd
->wr_busy_queues
== 0;
3248 * There is then a case where idling must be performed not
3249 * for throughput concerns, but to preserve service
3252 * To introduce this case, we can note that allowing the drive
3253 * to enqueue more than one request at a time, and hence
3254 * delegating de facto final scheduling decisions to the
3255 * drive's internal scheduler, entails loss of control on the
3256 * actual request service order. In particular, the critical
3257 * situation is when requests from different processes happen
3258 * to be present, at the same time, in the internal queue(s)
3259 * of the drive. In such a situation, the drive, by deciding
3260 * the service order of the internally-queued requests, does
3261 * determine also the actual throughput distribution among
3262 * these processes. But the drive typically has no notion or
3263 * concern about per-process throughput distribution, and
3264 * makes its decisions only on a per-request basis. Therefore,
3265 * the service distribution enforced by the drive's internal
3266 * scheduler is likely to coincide with the desired
3267 * device-throughput distribution only in a completely
3268 * symmetric scenario where:
3269 * (i) each of these processes must get the same throughput as
3271 * (ii) all these processes have the same I/O pattern
3272 (either sequential or random).
3273 * In fact, in such a scenario, the drive will tend to treat
3274 * the requests of each of these processes in about the same
3275 * way as the requests of the others, and thus to provide
3276 * each of these processes with about the same throughput
3277 * (which is exactly the desired throughput distribution). In
3278 * contrast, in any asymmetric scenario, device idling is
3279 * certainly needed to guarantee that bfqq receives its
3280 * assigned fraction of the device throughput (see [1] for
3283 * We address this issue by controlling, actually, only the
3284 * symmetry sub-condition (i), i.e., provided that
3285 * sub-condition (i) holds, idling is not performed,
3286 * regardless of whether sub-condition (ii) holds. In other
3287 * words, only if sub-condition (i) holds, then idling is
3288 * allowed, and the device tends to be prevented from queueing
3289 * many requests, possibly of several processes. The reason
3290 * for not controlling also sub-condition (ii) is that we
3291 * exploit preemption to preserve guarantees in case of
3292 * symmetric scenarios, even if (ii) does not hold, as
3293 * explained in the next two paragraphs.
3295 * Even if a queue, say Q, is expired when it remains idle, Q
3296 * can still preempt the new in-service queue if the next
3297 * request of Q arrives soon (see the comments on
3298 * bfq_bfqq_update_budg_for_activation). If all queues and
3299 * groups have the same weight, this form of preemption,
3300 * combined with the hole-recovery heuristic described in the
3301 * comments on function bfq_bfqq_update_budg_for_activation,
3302 * are enough to preserve a correct bandwidth distribution in
3303 * the mid term, even without idling. In fact, even if not
3304 * idling allows the internal queues of the device to contain
3305 * many requests, and thus to reorder requests, we can rather
3306 * safely assume that the internal scheduler still preserves a
3307 * minimum of mid-term fairness. The motivation for using
3308 * preemption instead of idling is that, by not idling,
3309 * service guarantees are preserved without minimally
3310 * sacrificing throughput. In other words, both a high
3311 * throughput and its desired distribution are obtained.
3313 * More precisely, this preemption-based, idleless approach
3314 * provides fairness in terms of IOPS, and not sectors per
3315 * second. This can be seen with a simple example. Suppose
3316 * that there are two queues with the same weight, but that
3317 * the first queue receives requests of 8 sectors, while the
3318 * second queue receives requests of 1024 sectors. In
3319 * addition, suppose that each of the two queues contains at
3320 * most one request at a time, which implies that each queue
3321 * always remains idle after it is served. Finally, after
3322 * remaining idle, each queue receives very quickly a new
3323 * request. It follows that the two queues are served
3324 * alternatively, preempting each other if needed. This
3325 * implies that, although both queues have the same weight,
3326 * the queue with large requests receives a service that is
3327 * 1024/8 times as high as the service received by the other
3330 * On the other hand, device idling is performed, and thus
3331 * pure sector-domain guarantees are provided, for the
3332 * following queues, which are likely to need stronger
3333 * throughput guarantees: weight-raised queues, and queues
3334 * with a higher weight than other queues. When such queues
3335 * are active, sub-condition (i) is false, which triggers
3338 * According to the above considerations, the next variable is
3339 * true (only) if sub-condition (i) holds. To compute the
3340 * value of this variable, we not only use the return value of
3341 * the function bfq_symmetric_scenario(), but also check
3342 * whether bfqq is being weight-raised, because
3343 * bfq_symmetric_scenario() does not take into account also
3344 * weight-raised queues (see comments on
3345 * bfq_weights_tree_add()).
3347 * As a side note, it is worth considering that the above
3348 * device-idling countermeasures may however fail in the
3349 * following unlucky scenario: if idling is (correctly)
3350 * disabled in a time period during which all symmetry
3351 * sub-conditions hold, and hence the device is allowed to
3352 * enqueue many requests, but at some later point in time some
3353 * sub-condition stops to hold, then it may become impossible
3354 * to let requests be served in the desired order until all
3355 * the requests already queued in the device have been served.
3357 asymmetric_scenario
= bfqq
->wr_coeff
> 1 ||
3358 !bfq_symmetric_scenario(bfqd
);
3361 * Finally, there is a case where maximizing throughput is the
3362 * best choice even if it may cause unfairness toward
3363 * bfqq. Such a case is when bfqq became active in a burst of
3364 * queue activations. Queues that became active during a large
3365 * burst benefit only from throughput, as discussed in the
3366 * comments on bfq_handle_burst. Thus, if bfqq became active
3367 * in a burst and not idling the device maximizes throughput,
3368 * then the device must no be idled, because not idling the
3369 * device provides bfqq and all other queues in the burst with
3370 * maximum benefit. Combining this and the above case, we can
3371 * now establish when idling is actually needed to preserve
3372 * service guarantees.
3374 idling_needed_for_service_guarantees
=
3375 asymmetric_scenario
&& !bfq_bfqq_in_large_burst(bfqq
);
3378 * We have now all the components we need to compute the
3379 * return value of the function, which is true only if idling
3380 * either boosts the throughput (without issues), or is
3381 * necessary to preserve service guarantees.
3383 return idling_boosts_thr_without_issues
||
3384 idling_needed_for_service_guarantees
;
3388 * If the in-service queue is empty but the function bfq_bfqq_may_idle
3389 * returns true, then:
3390 * 1) the queue must remain in service and cannot be expired, and
3391 * 2) the device must be idled to wait for the possible arrival of a new
3392 * request for the queue.
3393 * See the comments on the function bfq_bfqq_may_idle for the reasons
3394 * why performing device idling is the best choice to boost the throughput
3395 * and preserve service guarantees when bfq_bfqq_may_idle itself
3398 static bool bfq_bfqq_must_idle(struct bfq_queue
*bfqq
)
3400 return RB_EMPTY_ROOT(&bfqq
->sort_list
) && bfq_bfqq_may_idle(bfqq
);
3404 * Select a queue for service. If we have a current queue in service,
3405 * check whether to continue servicing it, or retrieve and set a new one.
3407 static struct bfq_queue
*bfq_select_queue(struct bfq_data
*bfqd
)
3409 struct bfq_queue
*bfqq
;
3410 struct request
*next_rq
;
3411 enum bfqq_expiration reason
= BFQQE_BUDGET_TIMEOUT
;
3413 bfqq
= bfqd
->in_service_queue
;
3417 bfq_log_bfqq(bfqd
, bfqq
, "select_queue: already in-service queue");
3419 if (bfq_may_expire_for_budg_timeout(bfqq
) &&
3420 !bfq_bfqq_wait_request(bfqq
) &&
3421 !bfq_bfqq_must_idle(bfqq
))
3426 * This loop is rarely executed more than once. Even when it
3427 * happens, it is much more convenient to re-execute this loop
3428 * than to return NULL and trigger a new dispatch to get a
3431 next_rq
= bfqq
->next_rq
;
3433 * If bfqq has requests queued and it has enough budget left to
3434 * serve them, keep the queue, otherwise expire it.
3437 if (bfq_serv_to_charge(next_rq
, bfqq
) >
3438 bfq_bfqq_budget_left(bfqq
)) {
3440 * Expire the queue for budget exhaustion,
3441 * which makes sure that the next budget is
3442 * enough to serve the next request, even if
3443 * it comes from the fifo expired path.
3445 reason
= BFQQE_BUDGET_EXHAUSTED
;
3449 * The idle timer may be pending because we may
3450 * not disable disk idling even when a new request
3453 if (bfq_bfqq_wait_request(bfqq
)) {
3455 * If we get here: 1) at least a new request
3456 * has arrived but we have not disabled the
3457 * timer because the request was too small,
3458 * 2) then the block layer has unplugged
3459 * the device, causing the dispatch to be
3462 * Since the device is unplugged, now the
3463 * requests are probably large enough to
3464 * provide a reasonable throughput.
3465 * So we disable idling.
3467 bfq_clear_bfqq_wait_request(bfqq
);
3468 hrtimer_try_to_cancel(&bfqd
->idle_slice_timer
);
3475 * No requests pending. However, if the in-service queue is idling
3476 * for a new request, or has requests waiting for a completion and
3477 * may idle after their completion, then keep it anyway.
3479 if (bfq_bfqq_wait_request(bfqq
) ||
3480 (bfqq
->dispatched
!= 0 && bfq_bfqq_may_idle(bfqq
))) {
3485 reason
= BFQQE_NO_MORE_REQUESTS
;
3487 bfq_bfqq_expire(bfqd
, bfqq
, false, reason
);
3489 bfqq
= bfq_set_in_service_queue(bfqd
);
3491 bfq_log_bfqq(bfqd
, bfqq
, "select_queue: checking new queue");
3496 bfq_log_bfqq(bfqd
, bfqq
, "select_queue: returned this queue");
3498 bfq_log(bfqd
, "select_queue: no queue returned");
3503 static void bfq_update_wr_data(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
3505 struct bfq_entity
*entity
= &bfqq
->entity
;
3507 if (bfqq
->wr_coeff
> 1) { /* queue is being weight-raised */
3508 bfq_log_bfqq(bfqd
, bfqq
,
3509 "raising period dur %u/%u msec, old coeff %u, w %d(%d)",
3510 jiffies_to_msecs(jiffies
- bfqq
->last_wr_start_finish
),
3511 jiffies_to_msecs(bfqq
->wr_cur_max_time
),
3513 bfqq
->entity
.weight
, bfqq
->entity
.orig_weight
);
3515 if (entity
->prio_changed
)
3516 bfq_log_bfqq(bfqd
, bfqq
, "WARN: pending prio change");
3519 * If the queue was activated in a burst, or too much
3520 * time has elapsed from the beginning of this
3521 * weight-raising period, then end weight raising.
3523 if (bfq_bfqq_in_large_burst(bfqq
))
3524 bfq_bfqq_end_wr(bfqq
);
3525 else if (time_is_before_jiffies(bfqq
->last_wr_start_finish
+
3526 bfqq
->wr_cur_max_time
)) {
3527 if (bfqq
->wr_cur_max_time
!= bfqd
->bfq_wr_rt_max_time
||
3528 time_is_before_jiffies(bfqq
->wr_start_at_switch_to_srt
+
3529 bfq_wr_duration(bfqd
)))
3530 bfq_bfqq_end_wr(bfqq
);
3532 switch_back_to_interactive_wr(bfqq
, bfqd
);
3533 bfqq
->entity
.prio_changed
= 1;
3538 * To improve latency (for this or other queues), immediately
3539 * update weight both if it must be raised and if it must be
3540 * lowered. Since, entity may be on some active tree here, and
3541 * might have a pending change of its ioprio class, invoke
3542 * next function with the last parameter unset (see the
3543 * comments on the function).
3545 if ((entity
->weight
> entity
->orig_weight
) != (bfqq
->wr_coeff
> 1))
3546 __bfq_entity_update_weight_prio(bfq_entity_service_tree(entity
),
3551 * Dispatch next request from bfqq.
3553 static struct request
*bfq_dispatch_rq_from_bfqq(struct bfq_data
*bfqd
,
3554 struct bfq_queue
*bfqq
)
3556 struct request
*rq
= bfqq
->next_rq
;
3557 unsigned long service_to_charge
;
3559 service_to_charge
= bfq_serv_to_charge(rq
, bfqq
);
3561 bfq_bfqq_served(bfqq
, service_to_charge
);
3563 bfq_dispatch_remove(bfqd
->queue
, rq
);
3566 * If weight raising has to terminate for bfqq, then next
3567 * function causes an immediate update of bfqq's weight,
3568 * without waiting for next activation. As a consequence, on
3569 * expiration, bfqq will be timestamped as if has never been
3570 * weight-raised during this service slot, even if it has
3571 * received part or even most of the service as a
3572 * weight-raised queue. This inflates bfqq's timestamps, which
3573 * is beneficial, as bfqq is then more willing to leave the
3574 * device immediately to possible other weight-raised queues.
3576 bfq_update_wr_data(bfqd
, bfqq
);
3579 * Expire bfqq, pretending that its budget expired, if bfqq
3580 * belongs to CLASS_IDLE and other queues are waiting for
3583 if (bfqd
->busy_queues
> 1 && bfq_class_idle(bfqq
))
3589 bfq_bfqq_expire(bfqd
, bfqq
, false, BFQQE_BUDGET_EXHAUSTED
);
3593 static bool bfq_has_work(struct blk_mq_hw_ctx
*hctx
)
3595 struct bfq_data
*bfqd
= hctx
->queue
->elevator
->elevator_data
;
3598 * Avoiding lock: a race on bfqd->busy_queues should cause at
3599 * most a call to dispatch for nothing
3601 return !list_empty_careful(&bfqd
->dispatch
) ||
3602 bfqd
->busy_queues
> 0;
3605 static struct request
*__bfq_dispatch_request(struct blk_mq_hw_ctx
*hctx
)
3607 struct bfq_data
*bfqd
= hctx
->queue
->elevator
->elevator_data
;
3608 struct request
*rq
= NULL
;
3609 struct bfq_queue
*bfqq
= NULL
;
3611 if (!list_empty(&bfqd
->dispatch
)) {
3612 rq
= list_first_entry(&bfqd
->dispatch
, struct request
,
3614 list_del_init(&rq
->queuelist
);
3620 * Increment counters here, because this
3621 * dispatch does not follow the standard
3622 * dispatch flow (where counters are
3627 goto inc_in_driver_start_rq
;
3631 * We exploit the bfq_finish_requeue_request hook to
3632 * decrement rq_in_driver, but
3633 * bfq_finish_requeue_request will not be invoked on
3634 * this request. So, to avoid unbalance, just start
3635 * this request, without incrementing rq_in_driver. As
3636 * a negative consequence, rq_in_driver is deceptively
3637 * lower than it should be while this request is in
3638 * service. This may cause bfq_schedule_dispatch to be
3639 * invoked uselessly.
3641 * As for implementing an exact solution, the
3642 * bfq_finish_requeue_request hook, if defined, is
3643 * probably invoked also on this request. So, by
3644 * exploiting this hook, we could 1) increment
3645 * rq_in_driver here, and 2) decrement it in
3646 * bfq_finish_requeue_request. Such a solution would
3647 * let the value of the counter be always accurate,
3648 * but it would entail using an extra interface
3649 * function. This cost seems higher than the benefit,
3650 * being the frequency of non-elevator-private
3651 * requests very low.
3656 bfq_log(bfqd
, "dispatch requests: %d busy queues", bfqd
->busy_queues
);
3658 if (bfqd
->busy_queues
== 0)
3662 * Force device to serve one request at a time if
3663 * strict_guarantees is true. Forcing this service scheme is
3664 * currently the ONLY way to guarantee that the request
3665 * service order enforced by the scheduler is respected by a
3666 * queueing device. Otherwise the device is free even to make
3667 * some unlucky request wait for as long as the device
3670 * Of course, serving one request at at time may cause loss of
3673 if (bfqd
->strict_guarantees
&& bfqd
->rq_in_driver
> 0)
3676 bfqq
= bfq_select_queue(bfqd
);
3680 rq
= bfq_dispatch_rq_from_bfqq(bfqd
, bfqq
);
3683 inc_in_driver_start_rq
:
3684 bfqd
->rq_in_driver
++;
3686 rq
->rq_flags
|= RQF_STARTED
;
3692 #if defined(CONFIG_BFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
3693 static void bfq_update_dispatch_stats(struct request_queue
*q
,
3695 struct bfq_queue
*in_serv_queue
,
3696 bool idle_timer_disabled
)
3698 struct bfq_queue
*bfqq
= rq
? RQ_BFQQ(rq
) : NULL
;
3700 if (!idle_timer_disabled
&& !bfqq
)
3704 * rq and bfqq are guaranteed to exist until this function
3705 * ends, for the following reasons. First, rq can be
3706 * dispatched to the device, and then can be completed and
3707 * freed, only after this function ends. Second, rq cannot be
3708 * merged (and thus freed because of a merge) any longer,
3709 * because it has already started. Thus rq cannot be freed
3710 * before this function ends, and, since rq has a reference to
3711 * bfqq, the same guarantee holds for bfqq too.
3713 * In addition, the following queue lock guarantees that
3714 * bfqq_group(bfqq) exists as well.
3716 spin_lock_irq(q
->queue_lock
);
3717 if (idle_timer_disabled
)
3719 * Since the idle timer has been disabled,
3720 * in_serv_queue contained some request when
3721 * __bfq_dispatch_request was invoked above, which
3722 * implies that rq was picked exactly from
3723 * in_serv_queue. Thus in_serv_queue == bfqq, and is
3724 * therefore guaranteed to exist because of the above
3727 bfqg_stats_update_idle_time(bfqq_group(in_serv_queue
));
3729 struct bfq_group
*bfqg
= bfqq_group(bfqq
);
3731 bfqg_stats_update_avg_queue_size(bfqg
);
3732 bfqg_stats_set_start_empty_time(bfqg
);
3733 bfqg_stats_update_io_remove(bfqg
, rq
->cmd_flags
);
3735 spin_unlock_irq(q
->queue_lock
);
3738 static inline void bfq_update_dispatch_stats(struct request_queue
*q
,
3740 struct bfq_queue
*in_serv_queue
,
3741 bool idle_timer_disabled
) {}
3744 static struct request
*bfq_dispatch_request(struct blk_mq_hw_ctx
*hctx
)
3746 struct bfq_data
*bfqd
= hctx
->queue
->elevator
->elevator_data
;
3748 struct bfq_queue
*in_serv_queue
;
3749 bool waiting_rq
, idle_timer_disabled
;
3751 spin_lock_irq(&bfqd
->lock
);
3753 in_serv_queue
= bfqd
->in_service_queue
;
3754 waiting_rq
= in_serv_queue
&& bfq_bfqq_wait_request(in_serv_queue
);
3756 rq
= __bfq_dispatch_request(hctx
);
3758 idle_timer_disabled
=
3759 waiting_rq
&& !bfq_bfqq_wait_request(in_serv_queue
);
3761 spin_unlock_irq(&bfqd
->lock
);
3763 bfq_update_dispatch_stats(hctx
->queue
, rq
, in_serv_queue
,
3764 idle_timer_disabled
);
3770 * Task holds one reference to the queue, dropped when task exits. Each rq
3771 * in-flight on this queue also holds a reference, dropped when rq is freed.
3773 * Scheduler lock must be held here. Recall not to use bfqq after calling
3774 * this function on it.
3776 void bfq_put_queue(struct bfq_queue
*bfqq
)
3778 #ifdef CONFIG_BFQ_GROUP_IOSCHED
3779 struct bfq_group
*bfqg
= bfqq_group(bfqq
);
3783 bfq_log_bfqq(bfqq
->bfqd
, bfqq
, "put_queue: %p %d",
3790 if (!hlist_unhashed(&bfqq
->burst_list_node
)) {
3791 hlist_del_init(&bfqq
->burst_list_node
);
3793 * Decrement also burst size after the removal, if the
3794 * process associated with bfqq is exiting, and thus
3795 * does not contribute to the burst any longer. This
3796 * decrement helps filter out false positives of large
3797 * bursts, when some short-lived process (often due to
3798 * the execution of commands by some service) happens
3799 * to start and exit while a complex application is
3800 * starting, and thus spawning several processes that
3801 * do I/O (and that *must not* be treated as a large
3802 * burst, see comments on bfq_handle_burst).
3804 * In particular, the decrement is performed only if:
3805 * 1) bfqq is not a merged queue, because, if it is,
3806 * then this free of bfqq is not triggered by the exit
3807 * of the process bfqq is associated with, but exactly
3808 * by the fact that bfqq has just been merged.
3809 * 2) burst_size is greater than 0, to handle
3810 * unbalanced decrements. Unbalanced decrements may
3811 * happen in te following case: bfqq is inserted into
3812 * the current burst list--without incrementing
3813 * bust_size--because of a split, but the current
3814 * burst list is not the burst list bfqq belonged to
3815 * (see comments on the case of a split in
3818 if (bfqq
->bic
&& bfqq
->bfqd
->burst_size
> 0)
3819 bfqq
->bfqd
->burst_size
--;
3822 kmem_cache_free(bfq_pool
, bfqq
);
3823 #ifdef CONFIG_BFQ_GROUP_IOSCHED
3824 bfqg_and_blkg_put(bfqg
);
3828 static void bfq_put_cooperator(struct bfq_queue
*bfqq
)
3830 struct bfq_queue
*__bfqq
, *next
;
3833 * If this queue was scheduled to merge with another queue, be
3834 * sure to drop the reference taken on that queue (and others in
3835 * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs.
3837 __bfqq
= bfqq
->new_bfqq
;
3841 next
= __bfqq
->new_bfqq
;
3842 bfq_put_queue(__bfqq
);
3847 static void bfq_exit_bfqq(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
3849 if (bfqq
== bfqd
->in_service_queue
) {
3850 __bfq_bfqq_expire(bfqd
, bfqq
);
3851 bfq_schedule_dispatch(bfqd
);
3854 bfq_log_bfqq(bfqd
, bfqq
, "exit_bfqq: %p, %d", bfqq
, bfqq
->ref
);
3856 bfq_put_cooperator(bfqq
);
3858 bfq_put_queue(bfqq
); /* release process reference */
3861 static void bfq_exit_icq_bfqq(struct bfq_io_cq
*bic
, bool is_sync
)
3863 struct bfq_queue
*bfqq
= bic_to_bfqq(bic
, is_sync
);
3864 struct bfq_data
*bfqd
;
3867 bfqd
= bfqq
->bfqd
; /* NULL if scheduler already exited */
3870 unsigned long flags
;
3872 spin_lock_irqsave(&bfqd
->lock
, flags
);
3873 bfq_exit_bfqq(bfqd
, bfqq
);
3874 bic_set_bfqq(bic
, NULL
, is_sync
);
3875 spin_unlock_irqrestore(&bfqd
->lock
, flags
);
3879 static void bfq_exit_icq(struct io_cq
*icq
)
3881 struct bfq_io_cq
*bic
= icq_to_bic(icq
);
3883 bfq_exit_icq_bfqq(bic
, true);
3884 bfq_exit_icq_bfqq(bic
, false);
3888 * Update the entity prio values; note that the new values will not
3889 * be used until the next (re)activation.
3892 bfq_set_next_ioprio_data(struct bfq_queue
*bfqq
, struct bfq_io_cq
*bic
)
3894 struct task_struct
*tsk
= current
;
3896 struct bfq_data
*bfqd
= bfqq
->bfqd
;
3901 ioprio_class
= IOPRIO_PRIO_CLASS(bic
->ioprio
);
3902 switch (ioprio_class
) {
3904 dev_err(bfqq
->bfqd
->queue
->backing_dev_info
->dev
,
3905 "bfq: bad prio class %d\n", ioprio_class
);
3907 case IOPRIO_CLASS_NONE
:
3909 * No prio set, inherit CPU scheduling settings.
3911 bfqq
->new_ioprio
= task_nice_ioprio(tsk
);
3912 bfqq
->new_ioprio_class
= task_nice_ioclass(tsk
);
3914 case IOPRIO_CLASS_RT
:
3915 bfqq
->new_ioprio
= IOPRIO_PRIO_DATA(bic
->ioprio
);
3916 bfqq
->new_ioprio_class
= IOPRIO_CLASS_RT
;
3918 case IOPRIO_CLASS_BE
:
3919 bfqq
->new_ioprio
= IOPRIO_PRIO_DATA(bic
->ioprio
);
3920 bfqq
->new_ioprio_class
= IOPRIO_CLASS_BE
;
3922 case IOPRIO_CLASS_IDLE
:
3923 bfqq
->new_ioprio_class
= IOPRIO_CLASS_IDLE
;
3924 bfqq
->new_ioprio
= 7;
3928 if (bfqq
->new_ioprio
>= IOPRIO_BE_NR
) {
3929 pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n",
3931 bfqq
->new_ioprio
= IOPRIO_BE_NR
;
3934 bfqq
->entity
.new_weight
= bfq_ioprio_to_weight(bfqq
->new_ioprio
);
3935 bfqq
->entity
.prio_changed
= 1;
3938 static struct bfq_queue
*bfq_get_queue(struct bfq_data
*bfqd
,
3939 struct bio
*bio
, bool is_sync
,
3940 struct bfq_io_cq
*bic
);
3942 static void bfq_check_ioprio_change(struct bfq_io_cq
*bic
, struct bio
*bio
)
3944 struct bfq_data
*bfqd
= bic_to_bfqd(bic
);
3945 struct bfq_queue
*bfqq
;
3946 int ioprio
= bic
->icq
.ioc
->ioprio
;
3949 * This condition may trigger on a newly created bic, be sure to
3950 * drop the lock before returning.
3952 if (unlikely(!bfqd
) || likely(bic
->ioprio
== ioprio
))
3955 bic
->ioprio
= ioprio
;
3957 bfqq
= bic_to_bfqq(bic
, false);
3959 /* release process reference on this queue */
3960 bfq_put_queue(bfqq
);
3961 bfqq
= bfq_get_queue(bfqd
, bio
, BLK_RW_ASYNC
, bic
);
3962 bic_set_bfqq(bic
, bfqq
, false);
3965 bfqq
= bic_to_bfqq(bic
, true);
3967 bfq_set_next_ioprio_data(bfqq
, bic
);
3970 static void bfq_init_bfqq(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
3971 struct bfq_io_cq
*bic
, pid_t pid
, int is_sync
)
3973 RB_CLEAR_NODE(&bfqq
->entity
.rb_node
);
3974 INIT_LIST_HEAD(&bfqq
->fifo
);
3975 INIT_HLIST_NODE(&bfqq
->burst_list_node
);
3981 bfq_set_next_ioprio_data(bfqq
, bic
);
3985 * No need to mark as has_short_ttime if in
3986 * idle_class, because no device idling is performed
3987 * for queues in idle class
3989 if (!bfq_class_idle(bfqq
))
3990 /* tentatively mark as has_short_ttime */
3991 bfq_mark_bfqq_has_short_ttime(bfqq
);
3992 bfq_mark_bfqq_sync(bfqq
);
3993 bfq_mark_bfqq_just_created(bfqq
);
3995 bfq_clear_bfqq_sync(bfqq
);
3997 /* set end request to minus infinity from now */
3998 bfqq
->ttime
.last_end_request
= ktime_get_ns() + 1;
4000 bfq_mark_bfqq_IO_bound(bfqq
);
4004 /* Tentative initial value to trade off between thr and lat */
4005 bfqq
->max_budget
= (2 * bfq_max_budget(bfqd
)) / 3;
4006 bfqq
->budget_timeout
= bfq_smallest_from_now();
4009 bfqq
->last_wr_start_finish
= jiffies
;
4010 bfqq
->wr_start_at_switch_to_srt
= bfq_smallest_from_now();
4011 bfqq
->split_time
= bfq_smallest_from_now();
4014 * Set to the value for which bfqq will not be deemed as
4015 * soft rt when it becomes backlogged.
4017 bfqq
->soft_rt_next_start
= bfq_greatest_from_now();
4019 /* first request is almost certainly seeky */
4020 bfqq
->seek_history
= 1;
4023 static struct bfq_queue
**bfq_async_queue_prio(struct bfq_data
*bfqd
,
4024 struct bfq_group
*bfqg
,
4025 int ioprio_class
, int ioprio
)
4027 switch (ioprio_class
) {
4028 case IOPRIO_CLASS_RT
:
4029 return &bfqg
->async_bfqq
[0][ioprio
];
4030 case IOPRIO_CLASS_NONE
:
4031 ioprio
= IOPRIO_NORM
;
4033 case IOPRIO_CLASS_BE
:
4034 return &bfqg
->async_bfqq
[1][ioprio
];
4035 case IOPRIO_CLASS_IDLE
:
4036 return &bfqg
->async_idle_bfqq
;
4042 static struct bfq_queue
*bfq_get_queue(struct bfq_data
*bfqd
,
4043 struct bio
*bio
, bool is_sync
,
4044 struct bfq_io_cq
*bic
)
4046 const int ioprio
= IOPRIO_PRIO_DATA(bic
->ioprio
);
4047 const int ioprio_class
= IOPRIO_PRIO_CLASS(bic
->ioprio
);
4048 struct bfq_queue
**async_bfqq
= NULL
;
4049 struct bfq_queue
*bfqq
;
4050 struct bfq_group
*bfqg
;
4054 bfqg
= bfq_find_set_group(bfqd
, bio_blkcg(bio
));
4056 bfqq
= &bfqd
->oom_bfqq
;
4061 async_bfqq
= bfq_async_queue_prio(bfqd
, bfqg
, ioprio_class
,
4068 bfqq
= kmem_cache_alloc_node(bfq_pool
,
4069 GFP_NOWAIT
| __GFP_ZERO
| __GFP_NOWARN
,
4073 bfq_init_bfqq(bfqd
, bfqq
, bic
, current
->pid
,
4075 bfq_init_entity(&bfqq
->entity
, bfqg
);
4076 bfq_log_bfqq(bfqd
, bfqq
, "allocated");
4078 bfqq
= &bfqd
->oom_bfqq
;
4079 bfq_log_bfqq(bfqd
, bfqq
, "using oom bfqq");
4084 * Pin the queue now that it's allocated, scheduler exit will
4089 * Extra group reference, w.r.t. sync
4090 * queue. This extra reference is removed
4091 * only if bfqq->bfqg disappears, to
4092 * guarantee that this queue is not freed
4093 * until its group goes away.
4095 bfq_log_bfqq(bfqd
, bfqq
, "get_queue, bfqq not in async: %p, %d",
4101 bfqq
->ref
++; /* get a process reference to this queue */
4102 bfq_log_bfqq(bfqd
, bfqq
, "get_queue, at end: %p, %d", bfqq
, bfqq
->ref
);
4107 static void bfq_update_io_thinktime(struct bfq_data
*bfqd
,
4108 struct bfq_queue
*bfqq
)
4110 struct bfq_ttime
*ttime
= &bfqq
->ttime
;
4111 u64 elapsed
= ktime_get_ns() - bfqq
->ttime
.last_end_request
;
4113 elapsed
= min_t(u64
, elapsed
, 2ULL * bfqd
->bfq_slice_idle
);
4115 ttime
->ttime_samples
= (7*bfqq
->ttime
.ttime_samples
+ 256) / 8;
4116 ttime
->ttime_total
= div_u64(7*ttime
->ttime_total
+ 256*elapsed
, 8);
4117 ttime
->ttime_mean
= div64_ul(ttime
->ttime_total
+ 128,
4118 ttime
->ttime_samples
);
4122 bfq_update_io_seektime(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
4125 bfqq
->seek_history
<<= 1;
4126 bfqq
->seek_history
|=
4127 get_sdist(bfqq
->last_request_pos
, rq
) > BFQQ_SEEK_THR
&&
4128 (!blk_queue_nonrot(bfqd
->queue
) ||
4129 blk_rq_sectors(rq
) < BFQQ_SECT_THR_NONROT
);
4132 static void bfq_update_has_short_ttime(struct bfq_data
*bfqd
,
4133 struct bfq_queue
*bfqq
,
4134 struct bfq_io_cq
*bic
)
4136 bool has_short_ttime
= true;
4139 * No need to update has_short_ttime if bfqq is async or in
4140 * idle io prio class, or if bfq_slice_idle is zero, because
4141 * no device idling is performed for bfqq in this case.
4143 if (!bfq_bfqq_sync(bfqq
) || bfq_class_idle(bfqq
) ||
4144 bfqd
->bfq_slice_idle
== 0)
4147 /* Idle window just restored, statistics are meaningless. */
4148 if (time_is_after_eq_jiffies(bfqq
->split_time
+
4149 bfqd
->bfq_wr_min_idle_time
))
4152 /* Think time is infinite if no process is linked to
4153 * bfqq. Otherwise check average think time to
4154 * decide whether to mark as has_short_ttime
4156 if (atomic_read(&bic
->icq
.ioc
->active_ref
) == 0 ||
4157 (bfq_sample_valid(bfqq
->ttime
.ttime_samples
) &&
4158 bfqq
->ttime
.ttime_mean
> bfqd
->bfq_slice_idle
))
4159 has_short_ttime
= false;
4161 bfq_log_bfqq(bfqd
, bfqq
, "update_has_short_ttime: has_short_ttime %d",
4164 if (has_short_ttime
)
4165 bfq_mark_bfqq_has_short_ttime(bfqq
);
4167 bfq_clear_bfqq_has_short_ttime(bfqq
);
4171 * Called when a new fs request (rq) is added to bfqq. Check if there's
4172 * something we should do about it.
4174 static void bfq_rq_enqueued(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
4177 struct bfq_io_cq
*bic
= RQ_BIC(rq
);
4179 if (rq
->cmd_flags
& REQ_META
)
4180 bfqq
->meta_pending
++;
4182 bfq_update_io_thinktime(bfqd
, bfqq
);
4183 bfq_update_has_short_ttime(bfqd
, bfqq
, bic
);
4184 bfq_update_io_seektime(bfqd
, bfqq
, rq
);
4186 bfq_log_bfqq(bfqd
, bfqq
,
4187 "rq_enqueued: has_short_ttime=%d (seeky %d)",
4188 bfq_bfqq_has_short_ttime(bfqq
), BFQQ_SEEKY(bfqq
));
4190 bfqq
->last_request_pos
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
4192 if (bfqq
== bfqd
->in_service_queue
&& bfq_bfqq_wait_request(bfqq
)) {
4193 bool small_req
= bfqq
->queued
[rq_is_sync(rq
)] == 1 &&
4194 blk_rq_sectors(rq
) < 32;
4195 bool budget_timeout
= bfq_bfqq_budget_timeout(bfqq
);
4198 * There is just this request queued: if the request
4199 * is small and the queue is not to be expired, then
4202 * In this way, if the device is being idled to wait
4203 * for a new request from the in-service queue, we
4204 * avoid unplugging the device and committing the
4205 * device to serve just a small request. On the
4206 * contrary, we wait for the block layer to decide
4207 * when to unplug the device: hopefully, new requests
4208 * will be merged to this one quickly, then the device
4209 * will be unplugged and larger requests will be
4212 if (small_req
&& !budget_timeout
)
4216 * A large enough request arrived, or the queue is to
4217 * be expired: in both cases disk idling is to be
4218 * stopped, so clear wait_request flag and reset
4221 bfq_clear_bfqq_wait_request(bfqq
);
4222 hrtimer_try_to_cancel(&bfqd
->idle_slice_timer
);
4225 * The queue is not empty, because a new request just
4226 * arrived. Hence we can safely expire the queue, in
4227 * case of budget timeout, without risking that the
4228 * timestamps of the queue are not updated correctly.
4229 * See [1] for more details.
4232 bfq_bfqq_expire(bfqd
, bfqq
, false,
4233 BFQQE_BUDGET_TIMEOUT
);
4237 /* returns true if it causes the idle timer to be disabled */
4238 static bool __bfq_insert_request(struct bfq_data
*bfqd
, struct request
*rq
)
4240 struct bfq_queue
*bfqq
= RQ_BFQQ(rq
),
4241 *new_bfqq
= bfq_setup_cooperator(bfqd
, bfqq
, rq
, true);
4242 bool waiting
, idle_timer_disabled
= false;
4245 if (bic_to_bfqq(RQ_BIC(rq
), 1) != bfqq
)
4246 new_bfqq
= bic_to_bfqq(RQ_BIC(rq
), 1);
4248 * Release the request's reference to the old bfqq
4249 * and make sure one is taken to the shared queue.
4251 new_bfqq
->allocated
++;
4255 * If the bic associated with the process
4256 * issuing this request still points to bfqq
4257 * (and thus has not been already redirected
4258 * to new_bfqq or even some other bfq_queue),
4259 * then complete the merge and redirect it to
4262 if (bic_to_bfqq(RQ_BIC(rq
), 1) == bfqq
)
4263 bfq_merge_bfqqs(bfqd
, RQ_BIC(rq
),
4266 bfq_clear_bfqq_just_created(bfqq
);
4268 * rq is about to be enqueued into new_bfqq,
4269 * release rq reference on bfqq
4271 bfq_put_queue(bfqq
);
4272 rq
->elv
.priv
[1] = new_bfqq
;
4276 waiting
= bfqq
&& bfq_bfqq_wait_request(bfqq
);
4277 bfq_add_request(rq
);
4278 idle_timer_disabled
= waiting
&& !bfq_bfqq_wait_request(bfqq
);
4280 rq
->fifo_time
= ktime_get_ns() + bfqd
->bfq_fifo_expire
[rq_is_sync(rq
)];
4281 list_add_tail(&rq
->queuelist
, &bfqq
->fifo
);
4283 bfq_rq_enqueued(bfqd
, bfqq
, rq
);
4285 return idle_timer_disabled
;
4288 #if defined(CONFIG_BFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
4289 static void bfq_update_insert_stats(struct request_queue
*q
,
4290 struct bfq_queue
*bfqq
,
4291 bool idle_timer_disabled
,
4292 unsigned int cmd_flags
)
4298 * bfqq still exists, because it can disappear only after
4299 * either it is merged with another queue, or the process it
4300 * is associated with exits. But both actions must be taken by
4301 * the same process currently executing this flow of
4304 * In addition, the following queue lock guarantees that
4305 * bfqq_group(bfqq) exists as well.
4307 spin_lock_irq(q
->queue_lock
);
4308 bfqg_stats_update_io_add(bfqq_group(bfqq
), bfqq
, cmd_flags
);
4309 if (idle_timer_disabled
)
4310 bfqg_stats_update_idle_time(bfqq_group(bfqq
));
4311 spin_unlock_irq(q
->queue_lock
);
4314 static inline void bfq_update_insert_stats(struct request_queue
*q
,
4315 struct bfq_queue
*bfqq
,
4316 bool idle_timer_disabled
,
4317 unsigned int cmd_flags
) {}
4320 static void bfq_prepare_request(struct request
*rq
, struct bio
*bio
);
4322 static void bfq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
4325 struct request_queue
*q
= hctx
->queue
;
4326 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
4327 struct bfq_queue
*bfqq
= RQ_BFQQ(rq
);
4328 bool idle_timer_disabled
= false;
4329 unsigned int cmd_flags
;
4331 spin_lock_irq(&bfqd
->lock
);
4332 if (blk_mq_sched_try_insert_merge(q
, rq
)) {
4333 spin_unlock_irq(&bfqd
->lock
);
4337 spin_unlock_irq(&bfqd
->lock
);
4339 blk_mq_sched_request_inserted(rq
);
4341 spin_lock_irq(&bfqd
->lock
);
4342 if (at_head
|| blk_rq_is_passthrough(rq
)) {
4344 list_add(&rq
->queuelist
, &bfqd
->dispatch
);
4346 list_add_tail(&rq
->queuelist
, &bfqd
->dispatch
);
4348 if (WARN_ON_ONCE(!bfqq
)) {
4350 * This should never happen. Most likely rq is
4351 * a requeued regular request, being
4352 * re-inserted without being first
4353 * re-prepared. Do a prepare, to avoid
4356 bfq_prepare_request(rq
, rq
->bio
);
4360 idle_timer_disabled
= __bfq_insert_request(bfqd
, rq
);
4362 * Update bfqq, because, if a queue merge has occurred
4363 * in __bfq_insert_request, then rq has been
4364 * redirected into a new queue.
4368 if (rq_mergeable(rq
)) {
4369 elv_rqhash_add(q
, rq
);
4376 * Cache cmd_flags before releasing scheduler lock, because rq
4377 * may disappear afterwards (for example, because of a request
4380 cmd_flags
= rq
->cmd_flags
;
4382 spin_unlock_irq(&bfqd
->lock
);
4384 bfq_update_insert_stats(q
, bfqq
, idle_timer_disabled
,
4388 static void bfq_insert_requests(struct blk_mq_hw_ctx
*hctx
,
4389 struct list_head
*list
, bool at_head
)
4391 while (!list_empty(list
)) {
4394 rq
= list_first_entry(list
, struct request
, queuelist
);
4395 list_del_init(&rq
->queuelist
);
4396 bfq_insert_request(hctx
, rq
, at_head
);
4400 static void bfq_update_hw_tag(struct bfq_data
*bfqd
)
4402 bfqd
->max_rq_in_driver
= max_t(int, bfqd
->max_rq_in_driver
,
4403 bfqd
->rq_in_driver
);
4405 if (bfqd
->hw_tag
== 1)
4409 * This sample is valid if the number of outstanding requests
4410 * is large enough to allow a queueing behavior. Note that the
4411 * sum is not exact, as it's not taking into account deactivated
4414 if (bfqd
->rq_in_driver
+ bfqd
->queued
< BFQ_HW_QUEUE_THRESHOLD
)
4417 if (bfqd
->hw_tag_samples
++ < BFQ_HW_QUEUE_SAMPLES
)
4420 bfqd
->hw_tag
= bfqd
->max_rq_in_driver
> BFQ_HW_QUEUE_THRESHOLD
;
4421 bfqd
->max_rq_in_driver
= 0;
4422 bfqd
->hw_tag_samples
= 0;
4425 static void bfq_completed_request(struct bfq_queue
*bfqq
, struct bfq_data
*bfqd
)
4430 bfq_update_hw_tag(bfqd
);
4432 bfqd
->rq_in_driver
--;
4435 if (!bfqq
->dispatched
&& !bfq_bfqq_busy(bfqq
)) {
4437 * Set budget_timeout (which we overload to store the
4438 * time at which the queue remains with no backlog and
4439 * no outstanding request; used by the weight-raising
4442 bfqq
->budget_timeout
= jiffies
;
4444 bfq_weights_tree_remove(bfqd
, &bfqq
->entity
,
4445 &bfqd
->queue_weights_tree
);
4448 now_ns
= ktime_get_ns();
4450 bfqq
->ttime
.last_end_request
= now_ns
;
4453 * Using us instead of ns, to get a reasonable precision in
4454 * computing rate in next check.
4456 delta_us
= div_u64(now_ns
- bfqd
->last_completion
, NSEC_PER_USEC
);
4459 * If the request took rather long to complete, and, according
4460 * to the maximum request size recorded, this completion latency
4461 * implies that the request was certainly served at a very low
4462 * rate (less than 1M sectors/sec), then the whole observation
4463 * interval that lasts up to this time instant cannot be a
4464 * valid time interval for computing a new peak rate. Invoke
4465 * bfq_update_rate_reset to have the following three steps
4467 * - close the observation interval at the last (previous)
4468 * request dispatch or completion
4469 * - compute rate, if possible, for that observation interval
4470 * - reset to zero samples, which will trigger a proper
4471 * re-initialization of the observation interval on next
4474 if (delta_us
> BFQ_MIN_TT
/NSEC_PER_USEC
&&
4475 (bfqd
->last_rq_max_size
<<BFQ_RATE_SHIFT
)/delta_us
<
4476 1UL<<(BFQ_RATE_SHIFT
- 10))
4477 bfq_update_rate_reset(bfqd
, NULL
);
4478 bfqd
->last_completion
= now_ns
;
4481 * If we are waiting to discover whether the request pattern
4482 * of the task associated with the queue is actually
4483 * isochronous, and both requisites for this condition to hold
4484 * are now satisfied, then compute soft_rt_next_start (see the
4485 * comments on the function bfq_bfqq_softrt_next_start()). We
4486 * schedule this delayed check when bfqq expires, if it still
4487 * has in-flight requests.
4489 if (bfq_bfqq_softrt_update(bfqq
) && bfqq
->dispatched
== 0 &&
4490 RB_EMPTY_ROOT(&bfqq
->sort_list
))
4491 bfqq
->soft_rt_next_start
=
4492 bfq_bfqq_softrt_next_start(bfqd
, bfqq
);
4495 * If this is the in-service queue, check if it needs to be expired,
4496 * or if we want to idle in case it has no pending requests.
4498 if (bfqd
->in_service_queue
== bfqq
) {
4499 if (bfqq
->dispatched
== 0 && bfq_bfqq_must_idle(bfqq
)) {
4500 bfq_arm_slice_timer(bfqd
);
4502 } else if (bfq_may_expire_for_budg_timeout(bfqq
))
4503 bfq_bfqq_expire(bfqd
, bfqq
, false,
4504 BFQQE_BUDGET_TIMEOUT
);
4505 else if (RB_EMPTY_ROOT(&bfqq
->sort_list
) &&
4506 (bfqq
->dispatched
== 0 ||
4507 !bfq_bfqq_may_idle(bfqq
)))
4508 bfq_bfqq_expire(bfqd
, bfqq
, false,
4509 BFQQE_NO_MORE_REQUESTS
);
4512 if (!bfqd
->rq_in_driver
)
4513 bfq_schedule_dispatch(bfqd
);
4516 static void bfq_finish_requeue_request_body(struct bfq_queue
*bfqq
)
4520 bfq_put_queue(bfqq
);
4524 * Handle either a requeue or a finish for rq. The things to do are
4525 * the same in both cases: all references to rq are to be dropped. In
4526 * particular, rq is considered completed from the point of view of
4529 static void bfq_finish_requeue_request(struct request
*rq
)
4531 struct bfq_queue
*bfqq
= RQ_BFQQ(rq
);
4532 struct bfq_data
*bfqd
;
4535 * Requeue and finish hooks are invoked in blk-mq without
4536 * checking whether the involved request is actually still
4537 * referenced in the scheduler. To handle this fact, the
4538 * following two checks make this function exit in case of
4539 * spurious invocations, for which there is nothing to do.
4541 * First, check whether rq has nothing to do with an elevator.
4543 if (unlikely(!(rq
->rq_flags
& RQF_ELVPRIV
)))
4547 * rq either is not associated with any icq, or is an already
4548 * requeued request that has not (yet) been re-inserted into
4551 if (!rq
->elv
.icq
|| !bfqq
)
4556 if (rq
->rq_flags
& RQF_STARTED
)
4557 bfqg_stats_update_completion(bfqq_group(bfqq
),
4558 rq_start_time_ns(rq
),
4559 rq_io_start_time_ns(rq
),
4562 if (likely(rq
->rq_flags
& RQF_STARTED
)) {
4563 unsigned long flags
;
4565 spin_lock_irqsave(&bfqd
->lock
, flags
);
4567 bfq_completed_request(bfqq
, bfqd
);
4568 bfq_finish_requeue_request_body(bfqq
);
4570 spin_unlock_irqrestore(&bfqd
->lock
, flags
);
4573 * Request rq may be still/already in the scheduler,
4574 * in which case we need to remove it (this should
4575 * never happen in case of requeue). And we cannot
4576 * defer such a check and removal, to avoid
4577 * inconsistencies in the time interval from the end
4578 * of this function to the start of the deferred work.
4579 * This situation seems to occur only in process
4580 * context, as a consequence of a merge. In the
4581 * current version of the code, this implies that the
4585 if (!RB_EMPTY_NODE(&rq
->rb_node
)) {
4586 bfq_remove_request(rq
->q
, rq
);
4587 bfqg_stats_update_io_remove(bfqq_group(bfqq
),
4590 bfq_finish_requeue_request_body(bfqq
);
4594 * Reset private fields. In case of a requeue, this allows
4595 * this function to correctly do nothing if it is spuriously
4596 * invoked again on this same request (see the check at the
4597 * beginning of the function). Probably, a better general
4598 * design would be to prevent blk-mq from invoking the requeue
4599 * or finish hooks of an elevator, for a request that is not
4600 * referred by that elevator.
4602 * Resetting the following fields would break the
4603 * request-insertion logic if rq is re-inserted into a bfq
4604 * internal queue, without a re-preparation. Here we assume
4605 * that re-insertions of requeued requests, without
4606 * re-preparation, can happen only for pass_through or at_head
4607 * requests (which are not re-inserted into bfq internal
4610 rq
->elv
.priv
[0] = NULL
;
4611 rq
->elv
.priv
[1] = NULL
;
4615 * Returns NULL if a new bfqq should be allocated, or the old bfqq if this
4616 * was the last process referring to that bfqq.
4618 static struct bfq_queue
*
4619 bfq_split_bfqq(struct bfq_io_cq
*bic
, struct bfq_queue
*bfqq
)
4621 bfq_log_bfqq(bfqq
->bfqd
, bfqq
, "splitting queue");
4623 if (bfqq_process_refs(bfqq
) == 1) {
4624 bfqq
->pid
= current
->pid
;
4625 bfq_clear_bfqq_coop(bfqq
);
4626 bfq_clear_bfqq_split_coop(bfqq
);
4630 bic_set_bfqq(bic
, NULL
, 1);
4632 bfq_put_cooperator(bfqq
);
4634 bfq_put_queue(bfqq
);
4638 static struct bfq_queue
*bfq_get_bfqq_handle_split(struct bfq_data
*bfqd
,
4639 struct bfq_io_cq
*bic
,
4641 bool split
, bool is_sync
,
4644 struct bfq_queue
*bfqq
= bic_to_bfqq(bic
, is_sync
);
4646 if (likely(bfqq
&& bfqq
!= &bfqd
->oom_bfqq
))
4653 bfq_put_queue(bfqq
);
4654 bfqq
= bfq_get_queue(bfqd
, bio
, is_sync
, bic
);
4656 bic_set_bfqq(bic
, bfqq
, is_sync
);
4657 if (split
&& is_sync
) {
4658 if ((bic
->was_in_burst_list
&& bfqd
->large_burst
) ||
4659 bic
->saved_in_large_burst
)
4660 bfq_mark_bfqq_in_large_burst(bfqq
);
4662 bfq_clear_bfqq_in_large_burst(bfqq
);
4663 if (bic
->was_in_burst_list
)
4665 * If bfqq was in the current
4666 * burst list before being
4667 * merged, then we have to add
4668 * it back. And we do not need
4669 * to increase burst_size, as
4670 * we did not decrement
4671 * burst_size when we removed
4672 * bfqq from the burst list as
4673 * a consequence of a merge
4675 * bfq_put_queue). In this
4676 * respect, it would be rather
4677 * costly to know whether the
4678 * current burst list is still
4679 * the same burst list from
4680 * which bfqq was removed on
4681 * the merge. To avoid this
4682 * cost, if bfqq was in a
4683 * burst list, then we add
4684 * bfqq to the current burst
4685 * list without any further
4686 * check. This can cause
4687 * inappropriate insertions,
4688 * but rarely enough to not
4689 * harm the detection of large
4690 * bursts significantly.
4692 hlist_add_head(&bfqq
->burst_list_node
,
4695 bfqq
->split_time
= jiffies
;
4702 * Allocate bfq data structures associated with this request.
4704 static void bfq_prepare_request(struct request
*rq
, struct bio
*bio
)
4706 struct request_queue
*q
= rq
->q
;
4707 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
4708 struct bfq_io_cq
*bic
;
4709 const int is_sync
= rq_is_sync(rq
);
4710 struct bfq_queue
*bfqq
;
4711 bool new_queue
= false;
4712 bool bfqq_already_existing
= false, split
= false;
4715 * Even if we don't have an icq attached, we should still clear
4716 * the scheduler pointers, as they might point to previously
4717 * allocated bic/bfqq structs.
4720 rq
->elv
.priv
[0] = rq
->elv
.priv
[1] = NULL
;
4724 bic
= icq_to_bic(rq
->elv
.icq
);
4726 spin_lock_irq(&bfqd
->lock
);
4728 bfq_check_ioprio_change(bic
, bio
);
4730 bfq_bic_update_cgroup(bic
, bio
);
4732 bfqq
= bfq_get_bfqq_handle_split(bfqd
, bic
, bio
, false, is_sync
,
4735 if (likely(!new_queue
)) {
4736 /* If the queue was seeky for too long, break it apart. */
4737 if (bfq_bfqq_coop(bfqq
) && bfq_bfqq_split_coop(bfqq
)) {
4738 bfq_log_bfqq(bfqd
, bfqq
, "breaking apart bfqq");
4740 /* Update bic before losing reference to bfqq */
4741 if (bfq_bfqq_in_large_burst(bfqq
))
4742 bic
->saved_in_large_burst
= true;
4744 bfqq
= bfq_split_bfqq(bic
, bfqq
);
4748 bfqq
= bfq_get_bfqq_handle_split(bfqd
, bic
, bio
,
4752 bfqq_already_existing
= true;
4758 bfq_log_bfqq(bfqd
, bfqq
, "get_request %p: bfqq %p, %d",
4759 rq
, bfqq
, bfqq
->ref
);
4761 rq
->elv
.priv
[0] = bic
;
4762 rq
->elv
.priv
[1] = bfqq
;
4765 * If a bfq_queue has only one process reference, it is owned
4766 * by only this bic: we can then set bfqq->bic = bic. in
4767 * addition, if the queue has also just been split, we have to
4770 if (likely(bfqq
!= &bfqd
->oom_bfqq
) && bfqq_process_refs(bfqq
) == 1) {
4774 * The queue has just been split from a shared
4775 * queue: restore the idle window and the
4776 * possible weight raising period.
4778 bfq_bfqq_resume_state(bfqq
, bfqd
, bic
,
4779 bfqq_already_existing
);
4783 if (unlikely(bfq_bfqq_just_created(bfqq
)))
4784 bfq_handle_burst(bfqd
, bfqq
);
4786 spin_unlock_irq(&bfqd
->lock
);
4789 static void bfq_idle_slice_timer_body(struct bfq_queue
*bfqq
)
4791 struct bfq_data
*bfqd
= bfqq
->bfqd
;
4792 enum bfqq_expiration reason
;
4793 unsigned long flags
;
4795 spin_lock_irqsave(&bfqd
->lock
, flags
);
4796 bfq_clear_bfqq_wait_request(bfqq
);
4798 if (bfqq
!= bfqd
->in_service_queue
) {
4799 spin_unlock_irqrestore(&bfqd
->lock
, flags
);
4803 if (bfq_bfqq_budget_timeout(bfqq
))
4805 * Also here the queue can be safely expired
4806 * for budget timeout without wasting
4809 reason
= BFQQE_BUDGET_TIMEOUT
;
4810 else if (bfqq
->queued
[0] == 0 && bfqq
->queued
[1] == 0)
4812 * The queue may not be empty upon timer expiration,
4813 * because we may not disable the timer when the
4814 * first request of the in-service queue arrives
4815 * during disk idling.
4817 reason
= BFQQE_TOO_IDLE
;
4819 goto schedule_dispatch
;
4821 bfq_bfqq_expire(bfqd
, bfqq
, true, reason
);
4824 spin_unlock_irqrestore(&bfqd
->lock
, flags
);
4825 bfq_schedule_dispatch(bfqd
);
4829 * Handler of the expiration of the timer running if the in-service queue
4830 * is idling inside its time slice.
4832 static enum hrtimer_restart
bfq_idle_slice_timer(struct hrtimer
*timer
)
4834 struct bfq_data
*bfqd
= container_of(timer
, struct bfq_data
,
4836 struct bfq_queue
*bfqq
= bfqd
->in_service_queue
;
4839 * Theoretical race here: the in-service queue can be NULL or
4840 * different from the queue that was idling if a new request
4841 * arrives for the current queue and there is a full dispatch
4842 * cycle that changes the in-service queue. This can hardly
4843 * happen, but in the worst case we just expire a queue too
4847 bfq_idle_slice_timer_body(bfqq
);
4849 return HRTIMER_NORESTART
;
4852 static void __bfq_put_async_bfqq(struct bfq_data
*bfqd
,
4853 struct bfq_queue
**bfqq_ptr
)
4855 struct bfq_queue
*bfqq
= *bfqq_ptr
;
4857 bfq_log(bfqd
, "put_async_bfqq: %p", bfqq
);
4859 bfq_bfqq_move(bfqd
, bfqq
, bfqd
->root_group
);
4861 bfq_log_bfqq(bfqd
, bfqq
, "put_async_bfqq: putting %p, %d",
4863 bfq_put_queue(bfqq
);
4869 * Release all the bfqg references to its async queues. If we are
4870 * deallocating the group these queues may still contain requests, so
4871 * we reparent them to the root cgroup (i.e., the only one that will
4872 * exist for sure until all the requests on a device are gone).
4874 void bfq_put_async_queues(struct bfq_data
*bfqd
, struct bfq_group
*bfqg
)
4878 for (i
= 0; i
< 2; i
++)
4879 for (j
= 0; j
< IOPRIO_BE_NR
; j
++)
4880 __bfq_put_async_bfqq(bfqd
, &bfqg
->async_bfqq
[i
][j
]);
4882 __bfq_put_async_bfqq(bfqd
, &bfqg
->async_idle_bfqq
);
4885 static void bfq_exit_queue(struct elevator_queue
*e
)
4887 struct bfq_data
*bfqd
= e
->elevator_data
;
4888 struct bfq_queue
*bfqq
, *n
;
4890 hrtimer_cancel(&bfqd
->idle_slice_timer
);
4892 spin_lock_irq(&bfqd
->lock
);
4893 list_for_each_entry_safe(bfqq
, n
, &bfqd
->idle_list
, bfqq_list
)
4894 bfq_deactivate_bfqq(bfqd
, bfqq
, false, false);
4895 spin_unlock_irq(&bfqd
->lock
);
4897 hrtimer_cancel(&bfqd
->idle_slice_timer
);
4899 #ifdef CONFIG_BFQ_GROUP_IOSCHED
4900 blkcg_deactivate_policy(bfqd
->queue
, &blkcg_policy_bfq
);
4902 spin_lock_irq(&bfqd
->lock
);
4903 bfq_put_async_queues(bfqd
, bfqd
->root_group
);
4904 kfree(bfqd
->root_group
);
4905 spin_unlock_irq(&bfqd
->lock
);
4911 static void bfq_init_root_group(struct bfq_group
*root_group
,
4912 struct bfq_data
*bfqd
)
4916 #ifdef CONFIG_BFQ_GROUP_IOSCHED
4917 root_group
->entity
.parent
= NULL
;
4918 root_group
->my_entity
= NULL
;
4919 root_group
->bfqd
= bfqd
;
4921 root_group
->rq_pos_tree
= RB_ROOT
;
4922 for (i
= 0; i
< BFQ_IOPRIO_CLASSES
; i
++)
4923 root_group
->sched_data
.service_tree
[i
] = BFQ_SERVICE_TREE_INIT
;
4924 root_group
->sched_data
.bfq_class_idle_last_service
= jiffies
;
4927 static int bfq_init_queue(struct request_queue
*q
, struct elevator_type
*e
)
4929 struct bfq_data
*bfqd
;
4930 struct elevator_queue
*eq
;
4932 eq
= elevator_alloc(q
, e
);
4936 bfqd
= kzalloc_node(sizeof(*bfqd
), GFP_KERNEL
, q
->node
);
4938 kobject_put(&eq
->kobj
);
4941 eq
->elevator_data
= bfqd
;
4943 spin_lock_irq(q
->queue_lock
);
4945 spin_unlock_irq(q
->queue_lock
);
4948 * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues.
4949 * Grab a permanent reference to it, so that the normal code flow
4950 * will not attempt to free it.
4952 bfq_init_bfqq(bfqd
, &bfqd
->oom_bfqq
, NULL
, 1, 0);
4953 bfqd
->oom_bfqq
.ref
++;
4954 bfqd
->oom_bfqq
.new_ioprio
= BFQ_DEFAULT_QUEUE_IOPRIO
;
4955 bfqd
->oom_bfqq
.new_ioprio_class
= IOPRIO_CLASS_BE
;
4956 bfqd
->oom_bfqq
.entity
.new_weight
=
4957 bfq_ioprio_to_weight(bfqd
->oom_bfqq
.new_ioprio
);
4959 /* oom_bfqq does not participate to bursts */
4960 bfq_clear_bfqq_just_created(&bfqd
->oom_bfqq
);
4963 * Trigger weight initialization, according to ioprio, at the
4964 * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio
4965 * class won't be changed any more.
4967 bfqd
->oom_bfqq
.entity
.prio_changed
= 1;
4971 INIT_LIST_HEAD(&bfqd
->dispatch
);
4973 hrtimer_init(&bfqd
->idle_slice_timer
, CLOCK_MONOTONIC
,
4975 bfqd
->idle_slice_timer
.function
= bfq_idle_slice_timer
;
4977 bfqd
->queue_weights_tree
= RB_ROOT
;
4978 bfqd
->group_weights_tree
= RB_ROOT
;
4980 INIT_LIST_HEAD(&bfqd
->active_list
);
4981 INIT_LIST_HEAD(&bfqd
->idle_list
);
4982 INIT_HLIST_HEAD(&bfqd
->burst_list
);
4986 bfqd
->bfq_max_budget
= bfq_default_max_budget
;
4988 bfqd
->bfq_fifo_expire
[0] = bfq_fifo_expire
[0];
4989 bfqd
->bfq_fifo_expire
[1] = bfq_fifo_expire
[1];
4990 bfqd
->bfq_back_max
= bfq_back_max
;
4991 bfqd
->bfq_back_penalty
= bfq_back_penalty
;
4992 bfqd
->bfq_slice_idle
= bfq_slice_idle
;
4993 bfqd
->bfq_timeout
= bfq_timeout
;
4995 bfqd
->bfq_requests_within_timer
= 120;
4997 bfqd
->bfq_large_burst_thresh
= 8;
4998 bfqd
->bfq_burst_interval
= msecs_to_jiffies(180);
5000 bfqd
->low_latency
= true;
5003 * Trade-off between responsiveness and fairness.
5005 bfqd
->bfq_wr_coeff
= 30;
5006 bfqd
->bfq_wr_rt_max_time
= msecs_to_jiffies(300);
5007 bfqd
->bfq_wr_max_time
= 0;
5008 bfqd
->bfq_wr_min_idle_time
= msecs_to_jiffies(2000);
5009 bfqd
->bfq_wr_min_inter_arr_async
= msecs_to_jiffies(500);
5010 bfqd
->bfq_wr_max_softrt_rate
= 7000; /*
5011 * Approximate rate required
5012 * to playback or record a
5013 * high-definition compressed
5016 bfqd
->wr_busy_queues
= 0;
5019 * Begin by assuming, optimistically, that the device is a
5020 * high-speed one, and that its peak rate is equal to 2/3 of
5021 * the highest reference rate.
5023 bfqd
->RT_prod
= R_fast
[blk_queue_nonrot(bfqd
->queue
)] *
5024 T_fast
[blk_queue_nonrot(bfqd
->queue
)];
5025 bfqd
->peak_rate
= R_fast
[blk_queue_nonrot(bfqd
->queue
)] * 2 / 3;
5026 bfqd
->device_speed
= BFQ_BFQD_FAST
;
5028 spin_lock_init(&bfqd
->lock
);
5031 * The invocation of the next bfq_create_group_hierarchy
5032 * function is the head of a chain of function calls
5033 * (bfq_create_group_hierarchy->blkcg_activate_policy->
5034 * blk_mq_freeze_queue) that may lead to the invocation of the
5035 * has_work hook function. For this reason,
5036 * bfq_create_group_hierarchy is invoked only after all
5037 * scheduler data has been initialized, apart from the fields
5038 * that can be initialized only after invoking
5039 * bfq_create_group_hierarchy. This, in particular, enables
5040 * has_work to correctly return false. Of course, to avoid
5041 * other inconsistencies, the blk-mq stack must then refrain
5042 * from invoking further scheduler hooks before this init
5043 * function is finished.
5045 bfqd
->root_group
= bfq_create_group_hierarchy(bfqd
, q
->node
);
5046 if (!bfqd
->root_group
)
5048 bfq_init_root_group(bfqd
->root_group
, bfqd
);
5049 bfq_init_entity(&bfqd
->oom_bfqq
.entity
, bfqd
->root_group
);
5051 wbt_disable_default(q
);
5056 kobject_put(&eq
->kobj
);
5060 static void bfq_slab_kill(void)
5062 kmem_cache_destroy(bfq_pool
);
5065 static int __init
bfq_slab_setup(void)
5067 bfq_pool
= KMEM_CACHE(bfq_queue
, 0);
5073 static ssize_t
bfq_var_show(unsigned int var
, char *page
)
5075 return sprintf(page
, "%u\n", var
);
5078 static int bfq_var_store(unsigned long *var
, const char *page
)
5080 unsigned long new_val
;
5081 int ret
= kstrtoul(page
, 10, &new_val
);
5089 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
5090 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
5092 struct bfq_data *bfqd = e->elevator_data; \
5093 u64 __data = __VAR; \
5095 __data = jiffies_to_msecs(__data); \
5096 else if (__CONV == 2) \
5097 __data = div_u64(__data, NSEC_PER_MSEC); \
5098 return bfq_var_show(__data, (page)); \
5100 SHOW_FUNCTION(bfq_fifo_expire_sync_show
, bfqd
->bfq_fifo_expire
[1], 2);
5101 SHOW_FUNCTION(bfq_fifo_expire_async_show
, bfqd
->bfq_fifo_expire
[0], 2);
5102 SHOW_FUNCTION(bfq_back_seek_max_show
, bfqd
->bfq_back_max
, 0);
5103 SHOW_FUNCTION(bfq_back_seek_penalty_show
, bfqd
->bfq_back_penalty
, 0);
5104 SHOW_FUNCTION(bfq_slice_idle_show
, bfqd
->bfq_slice_idle
, 2);
5105 SHOW_FUNCTION(bfq_max_budget_show
, bfqd
->bfq_user_max_budget
, 0);
5106 SHOW_FUNCTION(bfq_timeout_sync_show
, bfqd
->bfq_timeout
, 1);
5107 SHOW_FUNCTION(bfq_strict_guarantees_show
, bfqd
->strict_guarantees
, 0);
5108 SHOW_FUNCTION(bfq_low_latency_show
, bfqd
->low_latency
, 0);
5109 #undef SHOW_FUNCTION
5111 #define USEC_SHOW_FUNCTION(__FUNC, __VAR) \
5112 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
5114 struct bfq_data *bfqd = e->elevator_data; \
5115 u64 __data = __VAR; \
5116 __data = div_u64(__data, NSEC_PER_USEC); \
5117 return bfq_var_show(__data, (page)); \
5119 USEC_SHOW_FUNCTION(bfq_slice_idle_us_show
, bfqd
->bfq_slice_idle
);
5120 #undef USEC_SHOW_FUNCTION
5122 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
5124 __FUNC(struct elevator_queue *e, const char *page, size_t count) \
5126 struct bfq_data *bfqd = e->elevator_data; \
5127 unsigned long __data, __min = (MIN), __max = (MAX); \
5130 ret = bfq_var_store(&__data, (page)); \
5133 if (__data < __min) \
5135 else if (__data > __max) \
5138 *(__PTR) = msecs_to_jiffies(__data); \
5139 else if (__CONV == 2) \
5140 *(__PTR) = (u64)__data * NSEC_PER_MSEC; \
5142 *(__PTR) = __data; \
5145 STORE_FUNCTION(bfq_fifo_expire_sync_store
, &bfqd
->bfq_fifo_expire
[1], 1,
5147 STORE_FUNCTION(bfq_fifo_expire_async_store
, &bfqd
->bfq_fifo_expire
[0], 1,
5149 STORE_FUNCTION(bfq_back_seek_max_store
, &bfqd
->bfq_back_max
, 0, INT_MAX
, 0);
5150 STORE_FUNCTION(bfq_back_seek_penalty_store
, &bfqd
->bfq_back_penalty
, 1,
5152 STORE_FUNCTION(bfq_slice_idle_store
, &bfqd
->bfq_slice_idle
, 0, INT_MAX
, 2);
5153 #undef STORE_FUNCTION
5155 #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
5156 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\
5158 struct bfq_data *bfqd = e->elevator_data; \
5159 unsigned long __data, __min = (MIN), __max = (MAX); \
5162 ret = bfq_var_store(&__data, (page)); \
5165 if (__data < __min) \
5167 else if (__data > __max) \
5169 *(__PTR) = (u64)__data * NSEC_PER_USEC; \
5172 USEC_STORE_FUNCTION(bfq_slice_idle_us_store
, &bfqd
->bfq_slice_idle
, 0,
5174 #undef USEC_STORE_FUNCTION
5176 static ssize_t
bfq_max_budget_store(struct elevator_queue
*e
,
5177 const char *page
, size_t count
)
5179 struct bfq_data
*bfqd
= e
->elevator_data
;
5180 unsigned long __data
;
5183 ret
= bfq_var_store(&__data
, (page
));
5188 bfqd
->bfq_max_budget
= bfq_calc_max_budget(bfqd
);
5190 if (__data
> INT_MAX
)
5192 bfqd
->bfq_max_budget
= __data
;
5195 bfqd
->bfq_user_max_budget
= __data
;
5201 * Leaving this name to preserve name compatibility with cfq
5202 * parameters, but this timeout is used for both sync and async.
5204 static ssize_t
bfq_timeout_sync_store(struct elevator_queue
*e
,
5205 const char *page
, size_t count
)
5207 struct bfq_data
*bfqd
= e
->elevator_data
;
5208 unsigned long __data
;
5211 ret
= bfq_var_store(&__data
, (page
));
5217 else if (__data
> INT_MAX
)
5220 bfqd
->bfq_timeout
= msecs_to_jiffies(__data
);
5221 if (bfqd
->bfq_user_max_budget
== 0)
5222 bfqd
->bfq_max_budget
= bfq_calc_max_budget(bfqd
);
5227 static ssize_t
bfq_strict_guarantees_store(struct elevator_queue
*e
,
5228 const char *page
, size_t count
)
5230 struct bfq_data
*bfqd
= e
->elevator_data
;
5231 unsigned long __data
;
5234 ret
= bfq_var_store(&__data
, (page
));
5240 if (!bfqd
->strict_guarantees
&& __data
== 1
5241 && bfqd
->bfq_slice_idle
< 8 * NSEC_PER_MSEC
)
5242 bfqd
->bfq_slice_idle
= 8 * NSEC_PER_MSEC
;
5244 bfqd
->strict_guarantees
= __data
;
5249 static ssize_t
bfq_low_latency_store(struct elevator_queue
*e
,
5250 const char *page
, size_t count
)
5252 struct bfq_data
*bfqd
= e
->elevator_data
;
5253 unsigned long __data
;
5256 ret
= bfq_var_store(&__data
, (page
));
5262 if (__data
== 0 && bfqd
->low_latency
!= 0)
5264 bfqd
->low_latency
= __data
;
5269 #define BFQ_ATTR(name) \
5270 __ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store)
5272 static struct elv_fs_entry bfq_attrs
[] = {
5273 BFQ_ATTR(fifo_expire_sync
),
5274 BFQ_ATTR(fifo_expire_async
),
5275 BFQ_ATTR(back_seek_max
),
5276 BFQ_ATTR(back_seek_penalty
),
5277 BFQ_ATTR(slice_idle
),
5278 BFQ_ATTR(slice_idle_us
),
5279 BFQ_ATTR(max_budget
),
5280 BFQ_ATTR(timeout_sync
),
5281 BFQ_ATTR(strict_guarantees
),
5282 BFQ_ATTR(low_latency
),
5286 static struct elevator_type iosched_bfq_mq
= {
5288 .prepare_request
= bfq_prepare_request
,
5289 .requeue_request
= bfq_finish_requeue_request
,
5290 .finish_request
= bfq_finish_requeue_request
,
5291 .exit_icq
= bfq_exit_icq
,
5292 .insert_requests
= bfq_insert_requests
,
5293 .dispatch_request
= bfq_dispatch_request
,
5294 .next_request
= elv_rb_latter_request
,
5295 .former_request
= elv_rb_former_request
,
5296 .allow_merge
= bfq_allow_bio_merge
,
5297 .bio_merge
= bfq_bio_merge
,
5298 .request_merge
= bfq_request_merge
,
5299 .requests_merged
= bfq_requests_merged
,
5300 .request_merged
= bfq_request_merged
,
5301 .has_work
= bfq_has_work
,
5302 .init_sched
= bfq_init_queue
,
5303 .exit_sched
= bfq_exit_queue
,
5307 .icq_size
= sizeof(struct bfq_io_cq
),
5308 .icq_align
= __alignof__(struct bfq_io_cq
),
5309 .elevator_attrs
= bfq_attrs
,
5310 .elevator_name
= "bfq",
5311 .elevator_owner
= THIS_MODULE
,
5313 MODULE_ALIAS("bfq-iosched");
5315 static int __init
bfq_init(void)
5319 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5320 ret
= blkcg_policy_register(&blkcg_policy_bfq
);
5326 if (bfq_slab_setup())
5330 * Times to load large popular applications for the typical
5331 * systems installed on the reference devices (see the
5332 * comments before the definitions of the next two
5333 * arrays). Actually, we use slightly slower values, as the
5334 * estimated peak rate tends to be smaller than the actual
5335 * peak rate. The reason for this last fact is that estimates
5336 * are computed over much shorter time intervals than the long
5337 * intervals typically used for benchmarking. Why? First, to
5338 * adapt more quickly to variations. Second, because an I/O
5339 * scheduler cannot rely on a peak-rate-evaluation workload to
5340 * be run for a long time.
5342 T_slow
[0] = msecs_to_jiffies(3500); /* actually 4 sec */
5343 T_slow
[1] = msecs_to_jiffies(6000); /* actually 6.5 sec */
5344 T_fast
[0] = msecs_to_jiffies(7000); /* actually 8 sec */
5345 T_fast
[1] = msecs_to_jiffies(2500); /* actually 3 sec */
5348 * Thresholds that determine the switch between speed classes
5349 * (see the comments before the definition of the array
5350 * device_speed_thresh). These thresholds are biased towards
5351 * transitions to the fast class. This is safer than the
5352 * opposite bias. In fact, a wrong transition to the slow
5353 * class results in short weight-raising periods, because the
5354 * speed of the device then tends to be higher that the
5355 * reference peak rate. On the opposite end, a wrong
5356 * transition to the fast class tends to increase
5357 * weight-raising periods, because of the opposite reason.
5359 device_speed_thresh
[0] = (4 * R_slow
[0]) / 3;
5360 device_speed_thresh
[1] = (4 * R_slow
[1]) / 3;
5362 ret
= elv_register(&iosched_bfq_mq
);
5371 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5372 blkcg_policy_unregister(&blkcg_policy_bfq
);
5377 static void __exit
bfq_exit(void)
5379 elv_unregister(&iosched_bfq_mq
);
5380 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5381 blkcg_policy_unregister(&blkcg_policy_bfq
);
5386 module_init(bfq_init
);
5387 module_exit(bfq_exit
);
5389 MODULE_AUTHOR("Paolo Valente");
5390 MODULE_LICENSE("GPL");
5391 MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler");