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);
729 bfq_bfqq_resume_state(struct bfq_queue
*bfqq
, struct bfq_data
*bfqd
,
730 struct bfq_io_cq
*bic
, bool bfq_already_existing
)
732 unsigned int old_wr_coeff
= bfqq
->wr_coeff
;
733 bool busy
= bfq_already_existing
&& bfq_bfqq_busy(bfqq
);
735 if (bic
->saved_has_short_ttime
)
736 bfq_mark_bfqq_has_short_ttime(bfqq
);
738 bfq_clear_bfqq_has_short_ttime(bfqq
);
740 if (bic
->saved_IO_bound
)
741 bfq_mark_bfqq_IO_bound(bfqq
);
743 bfq_clear_bfqq_IO_bound(bfqq
);
745 bfqq
->ttime
= bic
->saved_ttime
;
746 bfqq
->wr_coeff
= bic
->saved_wr_coeff
;
747 bfqq
->wr_start_at_switch_to_srt
= bic
->saved_wr_start_at_switch_to_srt
;
748 bfqq
->last_wr_start_finish
= bic
->saved_last_wr_start_finish
;
749 bfqq
->wr_cur_max_time
= bic
->saved_wr_cur_max_time
;
751 if (bfqq
->wr_coeff
> 1 && (bfq_bfqq_in_large_burst(bfqq
) ||
752 time_is_before_jiffies(bfqq
->last_wr_start_finish
+
753 bfqq
->wr_cur_max_time
))) {
754 bfq_log_bfqq(bfqq
->bfqd
, bfqq
,
755 "resume state: switching off wr");
760 /* make sure weight will be updated, however we got here */
761 bfqq
->entity
.prio_changed
= 1;
766 if (old_wr_coeff
== 1 && bfqq
->wr_coeff
> 1)
767 bfqd
->wr_busy_queues
++;
768 else if (old_wr_coeff
> 1 && bfqq
->wr_coeff
== 1)
769 bfqd
->wr_busy_queues
--;
772 static int bfqq_process_refs(struct bfq_queue
*bfqq
)
774 return bfqq
->ref
- bfqq
->allocated
- bfqq
->entity
.on_st
;
777 /* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */
778 static void bfq_reset_burst_list(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
780 struct bfq_queue
*item
;
781 struct hlist_node
*n
;
783 hlist_for_each_entry_safe(item
, n
, &bfqd
->burst_list
, burst_list_node
)
784 hlist_del_init(&item
->burst_list_node
);
785 hlist_add_head(&bfqq
->burst_list_node
, &bfqd
->burst_list
);
786 bfqd
->burst_size
= 1;
787 bfqd
->burst_parent_entity
= bfqq
->entity
.parent
;
790 /* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */
791 static void bfq_add_to_burst(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
793 /* Increment burst size to take into account also bfqq */
796 if (bfqd
->burst_size
== bfqd
->bfq_large_burst_thresh
) {
797 struct bfq_queue
*pos
, *bfqq_item
;
798 struct hlist_node
*n
;
801 * Enough queues have been activated shortly after each
802 * other to consider this burst as large.
804 bfqd
->large_burst
= true;
807 * We can now mark all queues in the burst list as
808 * belonging to a large burst.
810 hlist_for_each_entry(bfqq_item
, &bfqd
->burst_list
,
812 bfq_mark_bfqq_in_large_burst(bfqq_item
);
813 bfq_mark_bfqq_in_large_burst(bfqq
);
816 * From now on, and until the current burst finishes, any
817 * new queue being activated shortly after the last queue
818 * was inserted in the burst can be immediately marked as
819 * belonging to a large burst. So the burst list is not
820 * needed any more. Remove it.
822 hlist_for_each_entry_safe(pos
, n
, &bfqd
->burst_list
,
824 hlist_del_init(&pos
->burst_list_node
);
826 * Burst not yet large: add bfqq to the burst list. Do
827 * not increment the ref counter for bfqq, because bfqq
828 * is removed from the burst list before freeing bfqq
831 hlist_add_head(&bfqq
->burst_list_node
, &bfqd
->burst_list
);
835 * If many queues belonging to the same group happen to be created
836 * shortly after each other, then the processes associated with these
837 * queues have typically a common goal. In particular, bursts of queue
838 * creations are usually caused by services or applications that spawn
839 * many parallel threads/processes. Examples are systemd during boot,
840 * or git grep. To help these processes get their job done as soon as
841 * possible, it is usually better to not grant either weight-raising
842 * or device idling to their queues.
844 * In this comment we describe, firstly, the reasons why this fact
845 * holds, and, secondly, the next function, which implements the main
846 * steps needed to properly mark these queues so that they can then be
847 * treated in a different way.
849 * The above services or applications benefit mostly from a high
850 * throughput: the quicker the requests of the activated queues are
851 * cumulatively served, the sooner the target job of these queues gets
852 * completed. As a consequence, weight-raising any of these queues,
853 * which also implies idling the device for it, is almost always
854 * counterproductive. In most cases it just lowers throughput.
856 * On the other hand, a burst of queue creations may be caused also by
857 * the start of an application that does not consist of a lot of
858 * parallel I/O-bound threads. In fact, with a complex application,
859 * several short processes may need to be executed to start-up the
860 * application. In this respect, to start an application as quickly as
861 * possible, the best thing to do is in any case to privilege the I/O
862 * related to the application with respect to all other
863 * I/O. Therefore, the best strategy to start as quickly as possible
864 * an application that causes a burst of queue creations is to
865 * weight-raise all the queues created during the burst. This is the
866 * exact opposite of the best strategy for the other type of bursts.
868 * In the end, to take the best action for each of the two cases, the
869 * two types of bursts need to be distinguished. Fortunately, this
870 * seems relatively easy, by looking at the sizes of the bursts. In
871 * particular, we found a threshold such that only bursts with a
872 * larger size than that threshold are apparently caused by
873 * services or commands such as systemd or git grep. For brevity,
874 * hereafter we call just 'large' these bursts. BFQ *does not*
875 * weight-raise queues whose creation occurs in a large burst. In
876 * addition, for each of these queues BFQ performs or does not perform
877 * idling depending on which choice boosts the throughput more. The
878 * exact choice depends on the device and request pattern at
881 * Unfortunately, false positives may occur while an interactive task
882 * is starting (e.g., an application is being started). The
883 * consequence is that the queues associated with the task do not
884 * enjoy weight raising as expected. Fortunately these false positives
885 * are very rare. They typically occur if some service happens to
886 * start doing I/O exactly when the interactive task starts.
888 * Turning back to the next function, it implements all the steps
889 * needed to detect the occurrence of a large burst and to properly
890 * mark all the queues belonging to it (so that they can then be
891 * treated in a different way). This goal is achieved by maintaining a
892 * "burst list" that holds, temporarily, the queues that belong to the
893 * burst in progress. The list is then used to mark these queues as
894 * belonging to a large burst if the burst does become large. The main
895 * steps are the following.
897 * . when the very first queue is created, the queue is inserted into the
898 * list (as it could be the first queue in a possible burst)
900 * . if the current burst has not yet become large, and a queue Q that does
901 * not yet belong to the burst is activated shortly after the last time
902 * at which a new queue entered the burst list, then the function appends
903 * Q to the burst list
905 * . if, as a consequence of the previous step, the burst size reaches
906 * the large-burst threshold, then
908 * . all the queues in the burst list are marked as belonging to a
911 * . the burst list is deleted; in fact, the burst list already served
912 * its purpose (keeping temporarily track of the queues in a burst,
913 * so as to be able to mark them as belonging to a large burst in the
914 * previous sub-step), and now is not needed any more
916 * . the device enters a large-burst mode
918 * . if a queue Q that does not belong to the burst is created while
919 * the device is in large-burst mode and shortly after the last time
920 * at which a queue either entered the burst list or was marked as
921 * belonging to the current large burst, then Q is immediately marked
922 * as belonging to a large burst.
924 * . if a queue Q that does not belong to the burst is created a while
925 * later, i.e., not shortly after, than the last time at which a queue
926 * either entered the burst list or was marked as belonging to the
927 * current large burst, then the current burst is deemed as finished and:
929 * . the large-burst mode is reset if set
931 * . the burst list is emptied
933 * . Q is inserted in the burst list, as Q may be the first queue
934 * in a possible new burst (then the burst list contains just Q
937 static void bfq_handle_burst(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
940 * If bfqq is already in the burst list or is part of a large
941 * burst, or finally has just been split, then there is
942 * nothing else to do.
944 if (!hlist_unhashed(&bfqq
->burst_list_node
) ||
945 bfq_bfqq_in_large_burst(bfqq
) ||
946 time_is_after_eq_jiffies(bfqq
->split_time
+
947 msecs_to_jiffies(10)))
951 * If bfqq's creation happens late enough, or bfqq belongs to
952 * a different group than the burst group, then the current
953 * burst is finished, and related data structures must be
956 * In this respect, consider the special case where bfqq is
957 * the very first queue created after BFQ is selected for this
958 * device. In this case, last_ins_in_burst and
959 * burst_parent_entity are not yet significant when we get
960 * here. But it is easy to verify that, whether or not the
961 * following condition is true, bfqq will end up being
962 * inserted into the burst list. In particular the list will
963 * happen to contain only bfqq. And this is exactly what has
964 * to happen, as bfqq may be the first queue of the first
967 if (time_is_before_jiffies(bfqd
->last_ins_in_burst
+
968 bfqd
->bfq_burst_interval
) ||
969 bfqq
->entity
.parent
!= bfqd
->burst_parent_entity
) {
970 bfqd
->large_burst
= false;
971 bfq_reset_burst_list(bfqd
, bfqq
);
976 * If we get here, then bfqq is being activated shortly after the
977 * last queue. So, if the current burst is also large, we can mark
978 * bfqq as belonging to this large burst immediately.
980 if (bfqd
->large_burst
) {
981 bfq_mark_bfqq_in_large_burst(bfqq
);
986 * If we get here, then a large-burst state has not yet been
987 * reached, but bfqq is being activated shortly after the last
988 * queue. Then we add bfqq to the burst.
990 bfq_add_to_burst(bfqd
, bfqq
);
993 * At this point, bfqq either has been added to the current
994 * burst or has caused the current burst to terminate and a
995 * possible new burst to start. In particular, in the second
996 * case, bfqq has become the first queue in the possible new
997 * burst. In both cases last_ins_in_burst needs to be moved
1000 bfqd
->last_ins_in_burst
= jiffies
;
1003 static int bfq_bfqq_budget_left(struct bfq_queue
*bfqq
)
1005 struct bfq_entity
*entity
= &bfqq
->entity
;
1007 return entity
->budget
- entity
->service
;
1011 * If enough samples have been computed, return the current max budget
1012 * stored in bfqd, which is dynamically updated according to the
1013 * estimated disk peak rate; otherwise return the default max budget
1015 static int bfq_max_budget(struct bfq_data
*bfqd
)
1017 if (bfqd
->budgets_assigned
< bfq_stats_min_budgets
)
1018 return bfq_default_max_budget
;
1020 return bfqd
->bfq_max_budget
;
1024 * Return min budget, which is a fraction of the current or default
1025 * max budget (trying with 1/32)
1027 static int bfq_min_budget(struct bfq_data
*bfqd
)
1029 if (bfqd
->budgets_assigned
< bfq_stats_min_budgets
)
1030 return bfq_default_max_budget
/ 32;
1032 return bfqd
->bfq_max_budget
/ 32;
1036 * The next function, invoked after the input queue bfqq switches from
1037 * idle to busy, updates the budget of bfqq. The function also tells
1038 * whether the in-service queue should be expired, by returning
1039 * true. The purpose of expiring the in-service queue is to give bfqq
1040 * the chance to possibly preempt the in-service queue, and the reason
1041 * for preempting the in-service queue is to achieve one of the two
1044 * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has
1045 * expired because it has remained idle. In particular, bfqq may have
1046 * expired for one of the following two reasons:
1048 * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling
1049 * and did not make it to issue a new request before its last
1050 * request was served;
1052 * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue
1053 * a new request before the expiration of the idling-time.
1055 * Even if bfqq has expired for one of the above reasons, the process
1056 * associated with the queue may be however issuing requests greedily,
1057 * and thus be sensitive to the bandwidth it receives (bfqq may have
1058 * remained idle for other reasons: CPU high load, bfqq not enjoying
1059 * idling, I/O throttling somewhere in the path from the process to
1060 * the I/O scheduler, ...). But if, after every expiration for one of
1061 * the above two reasons, bfqq has to wait for the service of at least
1062 * one full budget of another queue before being served again, then
1063 * bfqq is likely to get a much lower bandwidth or resource time than
1064 * its reserved ones. To address this issue, two countermeasures need
1067 * First, the budget and the timestamps of bfqq need to be updated in
1068 * a special way on bfqq reactivation: they need to be updated as if
1069 * bfqq did not remain idle and did not expire. In fact, if they are
1070 * computed as if bfqq expired and remained idle until reactivation,
1071 * then the process associated with bfqq is treated as if, instead of
1072 * being greedy, it stopped issuing requests when bfqq remained idle,
1073 * and restarts issuing requests only on this reactivation. In other
1074 * words, the scheduler does not help the process recover the "service
1075 * hole" between bfqq expiration and reactivation. As a consequence,
1076 * the process receives a lower bandwidth than its reserved one. In
1077 * contrast, to recover this hole, the budget must be updated as if
1078 * bfqq was not expired at all before this reactivation, i.e., it must
1079 * be set to the value of the remaining budget when bfqq was
1080 * expired. Along the same line, timestamps need to be assigned the
1081 * value they had the last time bfqq was selected for service, i.e.,
1082 * before last expiration. Thus timestamps need to be back-shifted
1083 * with respect to their normal computation (see [1] for more details
1084 * on this tricky aspect).
1086 * Secondly, to allow the process to recover the hole, the in-service
1087 * queue must be expired too, to give bfqq the chance to preempt it
1088 * immediately. In fact, if bfqq has to wait for a full budget of the
1089 * in-service queue to be completed, then it may become impossible to
1090 * let the process recover the hole, even if the back-shifted
1091 * timestamps of bfqq are lower than those of the in-service queue. If
1092 * this happens for most or all of the holes, then the process may not
1093 * receive its reserved bandwidth. In this respect, it is worth noting
1094 * that, being the service of outstanding requests unpreemptible, a
1095 * little fraction of the holes may however be unrecoverable, thereby
1096 * causing a little loss of bandwidth.
1098 * The last important point is detecting whether bfqq does need this
1099 * bandwidth recovery. In this respect, the next function deems the
1100 * process associated with bfqq greedy, and thus allows it to recover
1101 * the hole, if: 1) the process is waiting for the arrival of a new
1102 * request (which implies that bfqq expired for one of the above two
1103 * reasons), and 2) such a request has arrived soon. The first
1104 * condition is controlled through the flag non_blocking_wait_rq,
1105 * while the second through the flag arrived_in_time. If both
1106 * conditions hold, then the function computes the budget in the
1107 * above-described special way, and signals that the in-service queue
1108 * should be expired. Timestamp back-shifting is done later in
1109 * __bfq_activate_entity.
1111 * 2. Reduce latency. Even if timestamps are not backshifted to let
1112 * the process associated with bfqq recover a service hole, bfqq may
1113 * however happen to have, after being (re)activated, a lower finish
1114 * timestamp than the in-service queue. That is, the next budget of
1115 * bfqq may have to be completed before the one of the in-service
1116 * queue. If this is the case, then preempting the in-service queue
1117 * allows this goal to be achieved, apart from the unpreemptible,
1118 * outstanding requests mentioned above.
1120 * Unfortunately, regardless of which of the above two goals one wants
1121 * to achieve, service trees need first to be updated to know whether
1122 * the in-service queue must be preempted. To have service trees
1123 * correctly updated, the in-service queue must be expired and
1124 * rescheduled, and bfqq must be scheduled too. This is one of the
1125 * most costly operations (in future versions, the scheduling
1126 * mechanism may be re-designed in such a way to make it possible to
1127 * know whether preemption is needed without needing to update service
1128 * trees). In addition, queue preemptions almost always cause random
1129 * I/O, and thus loss of throughput. Because of these facts, the next
1130 * function adopts the following simple scheme to avoid both costly
1131 * operations and too frequent preemptions: it requests the expiration
1132 * of the in-service queue (unconditionally) only for queues that need
1133 * to recover a hole, or that either are weight-raised or deserve to
1136 static bool bfq_bfqq_update_budg_for_activation(struct bfq_data
*bfqd
,
1137 struct bfq_queue
*bfqq
,
1138 bool arrived_in_time
,
1139 bool wr_or_deserves_wr
)
1141 struct bfq_entity
*entity
= &bfqq
->entity
;
1143 if (bfq_bfqq_non_blocking_wait_rq(bfqq
) && arrived_in_time
) {
1145 * We do not clear the flag non_blocking_wait_rq here, as
1146 * the latter is used in bfq_activate_bfqq to signal
1147 * that timestamps need to be back-shifted (and is
1148 * cleared right after).
1152 * In next assignment we rely on that either
1153 * entity->service or entity->budget are not updated
1154 * on expiration if bfqq is empty (see
1155 * __bfq_bfqq_recalc_budget). Thus both quantities
1156 * remain unchanged after such an expiration, and the
1157 * following statement therefore assigns to
1158 * entity->budget the remaining budget on such an
1159 * expiration. For clarity, entity->service is not
1160 * updated on expiration in any case, and, in normal
1161 * operation, is reset only when bfqq is selected for
1162 * service (see bfq_get_next_queue).
1164 entity
->budget
= min_t(unsigned long,
1165 bfq_bfqq_budget_left(bfqq
),
1171 entity
->budget
= max_t(unsigned long, bfqq
->max_budget
,
1172 bfq_serv_to_charge(bfqq
->next_rq
, bfqq
));
1173 bfq_clear_bfqq_non_blocking_wait_rq(bfqq
);
1174 return wr_or_deserves_wr
;
1177 static unsigned int bfq_wr_duration(struct bfq_data
*bfqd
)
1181 if (bfqd
->bfq_wr_max_time
> 0)
1182 return bfqd
->bfq_wr_max_time
;
1184 dur
= bfqd
->RT_prod
;
1185 do_div(dur
, bfqd
->peak_rate
);
1188 * Limit duration between 3 and 13 seconds. Tests show that
1189 * higher values than 13 seconds often yield the opposite of
1190 * the desired result, i.e., worsen responsiveness by letting
1191 * non-interactive and non-soft-real-time applications
1192 * preserve weight raising for a too long time interval.
1194 * On the other end, lower values than 3 seconds make it
1195 * difficult for most interactive tasks to complete their jobs
1196 * before weight-raising finishes.
1198 if (dur
> msecs_to_jiffies(13000))
1199 dur
= msecs_to_jiffies(13000);
1200 else if (dur
< msecs_to_jiffies(3000))
1201 dur
= msecs_to_jiffies(3000);
1206 static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data
*bfqd
,
1207 struct bfq_queue
*bfqq
,
1208 unsigned int old_wr_coeff
,
1209 bool wr_or_deserves_wr
,
1214 if (old_wr_coeff
== 1 && wr_or_deserves_wr
) {
1215 /* start a weight-raising period */
1217 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
;
1218 bfqq
->wr_cur_max_time
= bfq_wr_duration(bfqd
);
1220 bfqq
->wr_start_at_switch_to_srt
= jiffies
;
1221 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
*
1222 BFQ_SOFTRT_WEIGHT_FACTOR
;
1223 bfqq
->wr_cur_max_time
=
1224 bfqd
->bfq_wr_rt_max_time
;
1228 * If needed, further reduce budget to make sure it is
1229 * close to bfqq's backlog, so as to reduce the
1230 * scheduling-error component due to a too large
1231 * budget. Do not care about throughput consequences,
1232 * but only about latency. Finally, do not assign a
1233 * too small budget either, to avoid increasing
1234 * latency by causing too frequent expirations.
1236 bfqq
->entity
.budget
= min_t(unsigned long,
1237 bfqq
->entity
.budget
,
1238 2 * bfq_min_budget(bfqd
));
1239 } else if (old_wr_coeff
> 1) {
1240 if (interactive
) { /* update wr coeff and duration */
1241 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
;
1242 bfqq
->wr_cur_max_time
= bfq_wr_duration(bfqd
);
1243 } else if (in_burst
)
1247 * The application is now or still meeting the
1248 * requirements for being deemed soft rt. We
1249 * can then correctly and safely (re)charge
1250 * the weight-raising duration for the
1251 * application with the weight-raising
1252 * duration for soft rt applications.
1254 * In particular, doing this recharge now, i.e.,
1255 * before the weight-raising period for the
1256 * application finishes, reduces the probability
1257 * of the following negative scenario:
1258 * 1) the weight of a soft rt application is
1259 * raised at startup (as for any newly
1260 * created application),
1261 * 2) since the application is not interactive,
1262 * at a certain time weight-raising is
1263 * stopped for the application,
1264 * 3) at that time the application happens to
1265 * still have pending requests, and hence
1266 * is destined to not have a chance to be
1267 * deemed soft rt before these requests are
1268 * completed (see the comments to the
1269 * function bfq_bfqq_softrt_next_start()
1270 * for details on soft rt detection),
1271 * 4) these pending requests experience a high
1272 * latency because the application is not
1273 * weight-raised while they are pending.
1275 if (bfqq
->wr_cur_max_time
!=
1276 bfqd
->bfq_wr_rt_max_time
) {
1277 bfqq
->wr_start_at_switch_to_srt
=
1278 bfqq
->last_wr_start_finish
;
1280 bfqq
->wr_cur_max_time
=
1281 bfqd
->bfq_wr_rt_max_time
;
1282 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
*
1283 BFQ_SOFTRT_WEIGHT_FACTOR
;
1285 bfqq
->last_wr_start_finish
= jiffies
;
1290 static bool bfq_bfqq_idle_for_long_time(struct bfq_data
*bfqd
,
1291 struct bfq_queue
*bfqq
)
1293 return bfqq
->dispatched
== 0 &&
1294 time_is_before_jiffies(
1295 bfqq
->budget_timeout
+
1296 bfqd
->bfq_wr_min_idle_time
);
1299 static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data
*bfqd
,
1300 struct bfq_queue
*bfqq
,
1305 bool soft_rt
, in_burst
, wr_or_deserves_wr
,
1306 bfqq_wants_to_preempt
,
1307 idle_for_long_time
= bfq_bfqq_idle_for_long_time(bfqd
, bfqq
),
1309 * See the comments on
1310 * bfq_bfqq_update_budg_for_activation for
1311 * details on the usage of the next variable.
1313 arrived_in_time
= ktime_get_ns() <=
1314 bfqq
->ttime
.last_end_request
+
1315 bfqd
->bfq_slice_idle
* 3;
1317 bfqg_stats_update_io_add(bfqq_group(RQ_BFQQ(rq
)), bfqq
, rq
->cmd_flags
);
1320 * bfqq deserves to be weight-raised if:
1322 * - it does not belong to a large burst,
1323 * - it has been idle for enough time or is soft real-time,
1324 * - is linked to a bfq_io_cq (it is not shared in any sense).
1326 in_burst
= bfq_bfqq_in_large_burst(bfqq
);
1327 soft_rt
= bfqd
->bfq_wr_max_softrt_rate
> 0 &&
1329 time_is_before_jiffies(bfqq
->soft_rt_next_start
);
1330 *interactive
= !in_burst
&& idle_for_long_time
;
1331 wr_or_deserves_wr
= bfqd
->low_latency
&&
1332 (bfqq
->wr_coeff
> 1 ||
1333 (bfq_bfqq_sync(bfqq
) &&
1334 bfqq
->bic
&& (*interactive
|| soft_rt
)));
1337 * Using the last flag, update budget and check whether bfqq
1338 * may want to preempt the in-service queue.
1340 bfqq_wants_to_preempt
=
1341 bfq_bfqq_update_budg_for_activation(bfqd
, bfqq
,
1346 * If bfqq happened to be activated in a burst, but has been
1347 * idle for much more than an interactive queue, then we
1348 * assume that, in the overall I/O initiated in the burst, the
1349 * I/O associated with bfqq is finished. So bfqq does not need
1350 * to be treated as a queue belonging to a burst
1351 * anymore. Accordingly, we reset bfqq's in_large_burst flag
1352 * if set, and remove bfqq from the burst list if it's
1353 * there. We do not decrement burst_size, because the fact
1354 * that bfqq does not need to belong to the burst list any
1355 * more does not invalidate the fact that bfqq was created in
1358 if (likely(!bfq_bfqq_just_created(bfqq
)) &&
1359 idle_for_long_time
&&
1360 time_is_before_jiffies(
1361 bfqq
->budget_timeout
+
1362 msecs_to_jiffies(10000))) {
1363 hlist_del_init(&bfqq
->burst_list_node
);
1364 bfq_clear_bfqq_in_large_burst(bfqq
);
1367 bfq_clear_bfqq_just_created(bfqq
);
1370 if (!bfq_bfqq_IO_bound(bfqq
)) {
1371 if (arrived_in_time
) {
1372 bfqq
->requests_within_timer
++;
1373 if (bfqq
->requests_within_timer
>=
1374 bfqd
->bfq_requests_within_timer
)
1375 bfq_mark_bfqq_IO_bound(bfqq
);
1377 bfqq
->requests_within_timer
= 0;
1380 if (bfqd
->low_latency
) {
1381 if (unlikely(time_is_after_jiffies(bfqq
->split_time
)))
1384 jiffies
- bfqd
->bfq_wr_min_idle_time
- 1;
1386 if (time_is_before_jiffies(bfqq
->split_time
+
1387 bfqd
->bfq_wr_min_idle_time
)) {
1388 bfq_update_bfqq_wr_on_rq_arrival(bfqd
, bfqq
,
1395 if (old_wr_coeff
!= bfqq
->wr_coeff
)
1396 bfqq
->entity
.prio_changed
= 1;
1400 bfqq
->last_idle_bklogged
= jiffies
;
1401 bfqq
->service_from_backlogged
= 0;
1402 bfq_clear_bfqq_softrt_update(bfqq
);
1404 bfq_add_bfqq_busy(bfqd
, bfqq
);
1407 * Expire in-service queue only if preemption may be needed
1408 * for guarantees. In this respect, the function
1409 * next_queue_may_preempt just checks a simple, necessary
1410 * condition, and not a sufficient condition based on
1411 * timestamps. In fact, for the latter condition to be
1412 * evaluated, timestamps would need first to be updated, and
1413 * this operation is quite costly (see the comments on the
1414 * function bfq_bfqq_update_budg_for_activation).
1416 if (bfqd
->in_service_queue
&& bfqq_wants_to_preempt
&&
1417 bfqd
->in_service_queue
->wr_coeff
< bfqq
->wr_coeff
&&
1418 next_queue_may_preempt(bfqd
))
1419 bfq_bfqq_expire(bfqd
, bfqd
->in_service_queue
,
1420 false, BFQQE_PREEMPTED
);
1423 static void bfq_add_request(struct request
*rq
)
1425 struct bfq_queue
*bfqq
= RQ_BFQQ(rq
);
1426 struct bfq_data
*bfqd
= bfqq
->bfqd
;
1427 struct request
*next_rq
, *prev
;
1428 unsigned int old_wr_coeff
= bfqq
->wr_coeff
;
1429 bool interactive
= false;
1431 bfq_log_bfqq(bfqd
, bfqq
, "add_request %d", rq_is_sync(rq
));
1432 bfqq
->queued
[rq_is_sync(rq
)]++;
1435 elv_rb_add(&bfqq
->sort_list
, rq
);
1438 * Check if this request is a better next-serve candidate.
1440 prev
= bfqq
->next_rq
;
1441 next_rq
= bfq_choose_req(bfqd
, bfqq
->next_rq
, rq
, bfqd
->last_position
);
1442 bfqq
->next_rq
= next_rq
;
1445 * Adjust priority tree position, if next_rq changes.
1447 if (prev
!= bfqq
->next_rq
)
1448 bfq_pos_tree_add_move(bfqd
, bfqq
);
1450 if (!bfq_bfqq_busy(bfqq
)) /* switching to busy ... */
1451 bfq_bfqq_handle_idle_busy_switch(bfqd
, bfqq
, old_wr_coeff
,
1454 if (bfqd
->low_latency
&& old_wr_coeff
== 1 && !rq_is_sync(rq
) &&
1455 time_is_before_jiffies(
1456 bfqq
->last_wr_start_finish
+
1457 bfqd
->bfq_wr_min_inter_arr_async
)) {
1458 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
;
1459 bfqq
->wr_cur_max_time
= bfq_wr_duration(bfqd
);
1461 bfqd
->wr_busy_queues
++;
1462 bfqq
->entity
.prio_changed
= 1;
1464 if (prev
!= bfqq
->next_rq
)
1465 bfq_updated_next_req(bfqd
, bfqq
);
1469 * Assign jiffies to last_wr_start_finish in the following
1472 * . if bfqq is not going to be weight-raised, because, for
1473 * non weight-raised queues, last_wr_start_finish stores the
1474 * arrival time of the last request; as of now, this piece
1475 * of information is used only for deciding whether to
1476 * weight-raise async queues
1478 * . if bfqq is not weight-raised, because, if bfqq is now
1479 * switching to weight-raised, then last_wr_start_finish
1480 * stores the time when weight-raising starts
1482 * . if bfqq is interactive, because, regardless of whether
1483 * bfqq is currently weight-raised, the weight-raising
1484 * period must start or restart (this case is considered
1485 * separately because it is not detected by the above
1486 * conditions, if bfqq is already weight-raised)
1488 * last_wr_start_finish has to be updated also if bfqq is soft
1489 * real-time, because the weight-raising period is constantly
1490 * restarted on idle-to-busy transitions for these queues, but
1491 * this is already done in bfq_bfqq_handle_idle_busy_switch if
1494 if (bfqd
->low_latency
&&
1495 (old_wr_coeff
== 1 || bfqq
->wr_coeff
== 1 || interactive
))
1496 bfqq
->last_wr_start_finish
= jiffies
;
1499 static struct request
*bfq_find_rq_fmerge(struct bfq_data
*bfqd
,
1501 struct request_queue
*q
)
1503 struct bfq_queue
*bfqq
= bfqd
->bio_bfqq
;
1507 return elv_rb_find(&bfqq
->sort_list
, bio_end_sector(bio
));
1512 static sector_t
get_sdist(sector_t last_pos
, struct request
*rq
)
1515 return abs(blk_rq_pos(rq
) - last_pos
);
1520 #if 0 /* Still not clear if we can do without next two functions */
1521 static void bfq_activate_request(struct request_queue
*q
, struct request
*rq
)
1523 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
1525 bfqd
->rq_in_driver
++;
1528 static void bfq_deactivate_request(struct request_queue
*q
, struct request
*rq
)
1530 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
1532 bfqd
->rq_in_driver
--;
1536 static void bfq_remove_request(struct request_queue
*q
,
1539 struct bfq_queue
*bfqq
= RQ_BFQQ(rq
);
1540 struct bfq_data
*bfqd
= bfqq
->bfqd
;
1541 const int sync
= rq_is_sync(rq
);
1543 if (bfqq
->next_rq
== rq
) {
1544 bfqq
->next_rq
= bfq_find_next_rq(bfqd
, bfqq
, rq
);
1545 bfq_updated_next_req(bfqd
, bfqq
);
1548 if (rq
->queuelist
.prev
!= &rq
->queuelist
)
1549 list_del_init(&rq
->queuelist
);
1550 bfqq
->queued
[sync
]--;
1552 elv_rb_del(&bfqq
->sort_list
, rq
);
1554 elv_rqhash_del(q
, rq
);
1555 if (q
->last_merge
== rq
)
1556 q
->last_merge
= NULL
;
1558 if (RB_EMPTY_ROOT(&bfqq
->sort_list
)) {
1559 bfqq
->next_rq
= NULL
;
1561 if (bfq_bfqq_busy(bfqq
) && bfqq
!= bfqd
->in_service_queue
) {
1562 bfq_del_bfqq_busy(bfqd
, bfqq
, false);
1564 * bfqq emptied. In normal operation, when
1565 * bfqq is empty, bfqq->entity.service and
1566 * bfqq->entity.budget must contain,
1567 * respectively, the service received and the
1568 * budget used last time bfqq emptied. These
1569 * facts do not hold in this case, as at least
1570 * this last removal occurred while bfqq is
1571 * not in service. To avoid inconsistencies,
1572 * reset both bfqq->entity.service and
1573 * bfqq->entity.budget, if bfqq has still a
1574 * process that may issue I/O requests to it.
1576 bfqq
->entity
.budget
= bfqq
->entity
.service
= 0;
1580 * Remove queue from request-position tree as it is empty.
1582 if (bfqq
->pos_root
) {
1583 rb_erase(&bfqq
->pos_node
, bfqq
->pos_root
);
1584 bfqq
->pos_root
= NULL
;
1588 if (rq
->cmd_flags
& REQ_META
)
1589 bfqq
->meta_pending
--;
1591 bfqg_stats_update_io_remove(bfqq_group(bfqq
), rq
->cmd_flags
);
1594 static bool bfq_bio_merge(struct blk_mq_hw_ctx
*hctx
, struct bio
*bio
)
1596 struct request_queue
*q
= hctx
->queue
;
1597 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
1598 struct request
*free
= NULL
;
1600 * bfq_bic_lookup grabs the queue_lock: invoke it now and
1601 * store its return value for later use, to avoid nesting
1602 * queue_lock inside the bfqd->lock. We assume that the bic
1603 * returned by bfq_bic_lookup does not go away before
1604 * bfqd->lock is taken.
1606 struct bfq_io_cq
*bic
= bfq_bic_lookup(bfqd
, current
->io_context
, q
);
1609 spin_lock_irq(&bfqd
->lock
);
1612 bfqd
->bio_bfqq
= bic_to_bfqq(bic
, op_is_sync(bio
->bi_opf
));
1614 bfqd
->bio_bfqq
= NULL
;
1615 bfqd
->bio_bic
= bic
;
1617 ret
= blk_mq_sched_try_merge(q
, bio
, &free
);
1620 blk_mq_free_request(free
);
1621 spin_unlock_irq(&bfqd
->lock
);
1626 static int bfq_request_merge(struct request_queue
*q
, struct request
**req
,
1629 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
1630 struct request
*__rq
;
1632 __rq
= bfq_find_rq_fmerge(bfqd
, bio
, q
);
1633 if (__rq
&& elv_bio_merge_ok(__rq
, bio
)) {
1635 return ELEVATOR_FRONT_MERGE
;
1638 return ELEVATOR_NO_MERGE
;
1641 static void bfq_request_merged(struct request_queue
*q
, struct request
*req
,
1642 enum elv_merge type
)
1644 if (type
== ELEVATOR_FRONT_MERGE
&&
1645 rb_prev(&req
->rb_node
) &&
1647 blk_rq_pos(container_of(rb_prev(&req
->rb_node
),
1648 struct request
, rb_node
))) {
1649 struct bfq_queue
*bfqq
= RQ_BFQQ(req
);
1650 struct bfq_data
*bfqd
= bfqq
->bfqd
;
1651 struct request
*prev
, *next_rq
;
1653 /* Reposition request in its sort_list */
1654 elv_rb_del(&bfqq
->sort_list
, req
);
1655 elv_rb_add(&bfqq
->sort_list
, req
);
1657 /* Choose next request to be served for bfqq */
1658 prev
= bfqq
->next_rq
;
1659 next_rq
= bfq_choose_req(bfqd
, bfqq
->next_rq
, req
,
1660 bfqd
->last_position
);
1661 bfqq
->next_rq
= next_rq
;
1663 * If next_rq changes, update both the queue's budget to
1664 * fit the new request and the queue's position in its
1667 if (prev
!= bfqq
->next_rq
) {
1668 bfq_updated_next_req(bfqd
, bfqq
);
1669 bfq_pos_tree_add_move(bfqd
, bfqq
);
1674 static void bfq_requests_merged(struct request_queue
*q
, struct request
*rq
,
1675 struct request
*next
)
1677 struct bfq_queue
*bfqq
= RQ_BFQQ(rq
), *next_bfqq
= RQ_BFQQ(next
);
1679 if (!RB_EMPTY_NODE(&rq
->rb_node
))
1681 spin_lock_irq(&bfqq
->bfqd
->lock
);
1684 * If next and rq belong to the same bfq_queue and next is older
1685 * than rq, then reposition rq in the fifo (by substituting next
1686 * with rq). Otherwise, if next and rq belong to different
1687 * bfq_queues, never reposition rq: in fact, we would have to
1688 * reposition it with respect to next's position in its own fifo,
1689 * which would most certainly be too expensive with respect to
1692 if (bfqq
== next_bfqq
&&
1693 !list_empty(&rq
->queuelist
) && !list_empty(&next
->queuelist
) &&
1694 next
->fifo_time
< rq
->fifo_time
) {
1695 list_del_init(&rq
->queuelist
);
1696 list_replace_init(&next
->queuelist
, &rq
->queuelist
);
1697 rq
->fifo_time
= next
->fifo_time
;
1700 if (bfqq
->next_rq
== next
)
1703 bfq_remove_request(q
, next
);
1705 spin_unlock_irq(&bfqq
->bfqd
->lock
);
1707 bfqg_stats_update_io_merged(bfqq_group(bfqq
), next
->cmd_flags
);
1710 /* Must be called with bfqq != NULL */
1711 static void bfq_bfqq_end_wr(struct bfq_queue
*bfqq
)
1713 if (bfq_bfqq_busy(bfqq
))
1714 bfqq
->bfqd
->wr_busy_queues
--;
1716 bfqq
->wr_cur_max_time
= 0;
1717 bfqq
->last_wr_start_finish
= jiffies
;
1719 * Trigger a weight change on the next invocation of
1720 * __bfq_entity_update_weight_prio.
1722 bfqq
->entity
.prio_changed
= 1;
1725 void bfq_end_wr_async_queues(struct bfq_data
*bfqd
,
1726 struct bfq_group
*bfqg
)
1730 for (i
= 0; i
< 2; i
++)
1731 for (j
= 0; j
< IOPRIO_BE_NR
; j
++)
1732 if (bfqg
->async_bfqq
[i
][j
])
1733 bfq_bfqq_end_wr(bfqg
->async_bfqq
[i
][j
]);
1734 if (bfqg
->async_idle_bfqq
)
1735 bfq_bfqq_end_wr(bfqg
->async_idle_bfqq
);
1738 static void bfq_end_wr(struct bfq_data
*bfqd
)
1740 struct bfq_queue
*bfqq
;
1742 spin_lock_irq(&bfqd
->lock
);
1744 list_for_each_entry(bfqq
, &bfqd
->active_list
, bfqq_list
)
1745 bfq_bfqq_end_wr(bfqq
);
1746 list_for_each_entry(bfqq
, &bfqd
->idle_list
, bfqq_list
)
1747 bfq_bfqq_end_wr(bfqq
);
1748 bfq_end_wr_async(bfqd
);
1750 spin_unlock_irq(&bfqd
->lock
);
1753 static sector_t
bfq_io_struct_pos(void *io_struct
, bool request
)
1756 return blk_rq_pos(io_struct
);
1758 return ((struct bio
*)io_struct
)->bi_iter
.bi_sector
;
1761 static int bfq_rq_close_to_sector(void *io_struct
, bool request
,
1764 return abs(bfq_io_struct_pos(io_struct
, request
) - sector
) <=
1768 static struct bfq_queue
*bfqq_find_close(struct bfq_data
*bfqd
,
1769 struct bfq_queue
*bfqq
,
1772 struct rb_root
*root
= &bfq_bfqq_to_bfqg(bfqq
)->rq_pos_tree
;
1773 struct rb_node
*parent
, *node
;
1774 struct bfq_queue
*__bfqq
;
1776 if (RB_EMPTY_ROOT(root
))
1780 * First, if we find a request starting at the end of the last
1781 * request, choose it.
1783 __bfqq
= bfq_rq_pos_tree_lookup(bfqd
, root
, sector
, &parent
, NULL
);
1788 * If the exact sector wasn't found, the parent of the NULL leaf
1789 * will contain the closest sector (rq_pos_tree sorted by
1790 * next_request position).
1792 __bfqq
= rb_entry(parent
, struct bfq_queue
, pos_node
);
1793 if (bfq_rq_close_to_sector(__bfqq
->next_rq
, true, sector
))
1796 if (blk_rq_pos(__bfqq
->next_rq
) < sector
)
1797 node
= rb_next(&__bfqq
->pos_node
);
1799 node
= rb_prev(&__bfqq
->pos_node
);
1803 __bfqq
= rb_entry(node
, struct bfq_queue
, pos_node
);
1804 if (bfq_rq_close_to_sector(__bfqq
->next_rq
, true, sector
))
1810 static struct bfq_queue
*bfq_find_close_cooperator(struct bfq_data
*bfqd
,
1811 struct bfq_queue
*cur_bfqq
,
1814 struct bfq_queue
*bfqq
;
1817 * We shall notice if some of the queues are cooperating,
1818 * e.g., working closely on the same area of the device. In
1819 * that case, we can group them together and: 1) don't waste
1820 * time idling, and 2) serve the union of their requests in
1821 * the best possible order for throughput.
1823 bfqq
= bfqq_find_close(bfqd
, cur_bfqq
, sector
);
1824 if (!bfqq
|| bfqq
== cur_bfqq
)
1830 static struct bfq_queue
*
1831 bfq_setup_merge(struct bfq_queue
*bfqq
, struct bfq_queue
*new_bfqq
)
1833 int process_refs
, new_process_refs
;
1834 struct bfq_queue
*__bfqq
;
1837 * If there are no process references on the new_bfqq, then it is
1838 * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain
1839 * may have dropped their last reference (not just their last process
1842 if (!bfqq_process_refs(new_bfqq
))
1845 /* Avoid a circular list and skip interim queue merges. */
1846 while ((__bfqq
= new_bfqq
->new_bfqq
)) {
1852 process_refs
= bfqq_process_refs(bfqq
);
1853 new_process_refs
= bfqq_process_refs(new_bfqq
);
1855 * If the process for the bfqq has gone away, there is no
1856 * sense in merging the queues.
1858 if (process_refs
== 0 || new_process_refs
== 0)
1861 bfq_log_bfqq(bfqq
->bfqd
, bfqq
, "scheduling merge with queue %d",
1865 * Merging is just a redirection: the requests of the process
1866 * owning one of the two queues are redirected to the other queue.
1867 * The latter queue, in its turn, is set as shared if this is the
1868 * first time that the requests of some process are redirected to
1871 * We redirect bfqq to new_bfqq and not the opposite, because
1872 * we are in the context of the process owning bfqq, thus we
1873 * have the io_cq of this process. So we can immediately
1874 * configure this io_cq to redirect the requests of the
1875 * process to new_bfqq. In contrast, the io_cq of new_bfqq is
1876 * not available any more (new_bfqq->bic == NULL).
1878 * Anyway, even in case new_bfqq coincides with the in-service
1879 * queue, redirecting requests the in-service queue is the
1880 * best option, as we feed the in-service queue with new
1881 * requests close to the last request served and, by doing so,
1882 * are likely to increase the throughput.
1884 bfqq
->new_bfqq
= new_bfqq
;
1885 new_bfqq
->ref
+= process_refs
;
1889 static bool bfq_may_be_close_cooperator(struct bfq_queue
*bfqq
,
1890 struct bfq_queue
*new_bfqq
)
1892 if (bfq_class_idle(bfqq
) || bfq_class_idle(new_bfqq
) ||
1893 (bfqq
->ioprio_class
!= new_bfqq
->ioprio_class
))
1897 * If either of the queues has already been detected as seeky,
1898 * then merging it with the other queue is unlikely to lead to
1901 if (BFQQ_SEEKY(bfqq
) || BFQQ_SEEKY(new_bfqq
))
1905 * Interleaved I/O is known to be done by (some) applications
1906 * only for reads, so it does not make sense to merge async
1909 if (!bfq_bfqq_sync(bfqq
) || !bfq_bfqq_sync(new_bfqq
))
1916 * If this function returns true, then bfqq cannot be merged. The idea
1917 * is that true cooperation happens very early after processes start
1918 * to do I/O. Usually, late cooperations are just accidental false
1919 * positives. In case bfqq is weight-raised, such false positives
1920 * would evidently degrade latency guarantees for bfqq.
1922 static bool wr_from_too_long(struct bfq_queue
*bfqq
)
1924 return bfqq
->wr_coeff
> 1 &&
1925 time_is_before_jiffies(bfqq
->last_wr_start_finish
+
1926 msecs_to_jiffies(100));
1930 * Attempt to schedule a merge of bfqq with the currently in-service
1931 * queue or with a close queue among the scheduled queues. Return
1932 * NULL if no merge was scheduled, a pointer to the shared bfq_queue
1933 * structure otherwise.
1935 * The OOM queue is not allowed to participate to cooperation: in fact, since
1936 * the requests temporarily redirected to the OOM queue could be redirected
1937 * again to dedicated queues at any time, the state needed to correctly
1938 * handle merging with the OOM queue would be quite complex and expensive
1939 * to maintain. Besides, in such a critical condition as an out of memory,
1940 * the benefits of queue merging may be little relevant, or even negligible.
1942 * Weight-raised queues can be merged only if their weight-raising
1943 * period has just started. In fact cooperating processes are usually
1944 * started together. Thus, with this filter we avoid false positives
1945 * that would jeopardize low-latency guarantees.
1947 * WARNING: queue merging may impair fairness among non-weight raised
1948 * queues, for at least two reasons: 1) the original weight of a
1949 * merged queue may change during the merged state, 2) even being the
1950 * weight the same, a merged queue may be bloated with many more
1951 * requests than the ones produced by its originally-associated
1954 static struct bfq_queue
*
1955 bfq_setup_cooperator(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
1956 void *io_struct
, bool request
)
1958 struct bfq_queue
*in_service_bfqq
, *new_bfqq
;
1961 return bfqq
->new_bfqq
;
1964 wr_from_too_long(bfqq
) ||
1965 unlikely(bfqq
== &bfqd
->oom_bfqq
))
1968 /* If there is only one backlogged queue, don't search. */
1969 if (bfqd
->busy_queues
== 1)
1972 in_service_bfqq
= bfqd
->in_service_queue
;
1974 if (!in_service_bfqq
|| in_service_bfqq
== bfqq
1975 || wr_from_too_long(in_service_bfqq
) ||
1976 unlikely(in_service_bfqq
== &bfqd
->oom_bfqq
))
1977 goto check_scheduled
;
1979 if (bfq_rq_close_to_sector(io_struct
, request
, bfqd
->last_position
) &&
1980 bfqq
->entity
.parent
== in_service_bfqq
->entity
.parent
&&
1981 bfq_may_be_close_cooperator(bfqq
, in_service_bfqq
)) {
1982 new_bfqq
= bfq_setup_merge(bfqq
, in_service_bfqq
);
1987 * Check whether there is a cooperator among currently scheduled
1988 * queues. The only thing we need is that the bio/request is not
1989 * NULL, as we need it to establish whether a cooperator exists.
1992 new_bfqq
= bfq_find_close_cooperator(bfqd
, bfqq
,
1993 bfq_io_struct_pos(io_struct
, request
));
1995 if (new_bfqq
&& !wr_from_too_long(new_bfqq
) &&
1996 likely(new_bfqq
!= &bfqd
->oom_bfqq
) &&
1997 bfq_may_be_close_cooperator(bfqq
, new_bfqq
))
1998 return bfq_setup_merge(bfqq
, new_bfqq
);
2003 static void bfq_bfqq_save_state(struct bfq_queue
*bfqq
)
2005 struct bfq_io_cq
*bic
= bfqq
->bic
;
2008 * If !bfqq->bic, the queue is already shared or its requests
2009 * have already been redirected to a shared queue; both idle window
2010 * and weight raising state have already been saved. Do nothing.
2015 bic
->saved_ttime
= bfqq
->ttime
;
2016 bic
->saved_has_short_ttime
= bfq_bfqq_has_short_ttime(bfqq
);
2017 bic
->saved_IO_bound
= bfq_bfqq_IO_bound(bfqq
);
2018 bic
->saved_in_large_burst
= bfq_bfqq_in_large_burst(bfqq
);
2019 bic
->was_in_burst_list
= !hlist_unhashed(&bfqq
->burst_list_node
);
2020 bic
->saved_wr_coeff
= bfqq
->wr_coeff
;
2021 bic
->saved_wr_start_at_switch_to_srt
= bfqq
->wr_start_at_switch_to_srt
;
2022 bic
->saved_last_wr_start_finish
= bfqq
->last_wr_start_finish
;
2023 bic
->saved_wr_cur_max_time
= bfqq
->wr_cur_max_time
;
2027 bfq_merge_bfqqs(struct bfq_data
*bfqd
, struct bfq_io_cq
*bic
,
2028 struct bfq_queue
*bfqq
, struct bfq_queue
*new_bfqq
)
2030 bfq_log_bfqq(bfqd
, bfqq
, "merging with queue %lu",
2031 (unsigned long)new_bfqq
->pid
);
2032 /* Save weight raising and idle window of the merged queues */
2033 bfq_bfqq_save_state(bfqq
);
2034 bfq_bfqq_save_state(new_bfqq
);
2035 if (bfq_bfqq_IO_bound(bfqq
))
2036 bfq_mark_bfqq_IO_bound(new_bfqq
);
2037 bfq_clear_bfqq_IO_bound(bfqq
);
2040 * If bfqq is weight-raised, then let new_bfqq inherit
2041 * weight-raising. To reduce false positives, neglect the case
2042 * where bfqq has just been created, but has not yet made it
2043 * to be weight-raised (which may happen because EQM may merge
2044 * bfqq even before bfq_add_request is executed for the first
2045 * time for bfqq). Handling this case would however be very
2046 * easy, thanks to the flag just_created.
2048 if (new_bfqq
->wr_coeff
== 1 && bfqq
->wr_coeff
> 1) {
2049 new_bfqq
->wr_coeff
= bfqq
->wr_coeff
;
2050 new_bfqq
->wr_cur_max_time
= bfqq
->wr_cur_max_time
;
2051 new_bfqq
->last_wr_start_finish
= bfqq
->last_wr_start_finish
;
2052 new_bfqq
->wr_start_at_switch_to_srt
=
2053 bfqq
->wr_start_at_switch_to_srt
;
2054 if (bfq_bfqq_busy(new_bfqq
))
2055 bfqd
->wr_busy_queues
++;
2056 new_bfqq
->entity
.prio_changed
= 1;
2059 if (bfqq
->wr_coeff
> 1) { /* bfqq has given its wr to new_bfqq */
2061 bfqq
->entity
.prio_changed
= 1;
2062 if (bfq_bfqq_busy(bfqq
))
2063 bfqd
->wr_busy_queues
--;
2066 bfq_log_bfqq(bfqd
, new_bfqq
, "merge_bfqqs: wr_busy %d",
2067 bfqd
->wr_busy_queues
);
2070 * Merge queues (that is, let bic redirect its requests to new_bfqq)
2072 bic_set_bfqq(bic
, new_bfqq
, 1);
2073 bfq_mark_bfqq_coop(new_bfqq
);
2075 * new_bfqq now belongs to at least two bics (it is a shared queue):
2076 * set new_bfqq->bic to NULL. bfqq either:
2077 * - does not belong to any bic any more, and hence bfqq->bic must
2078 * be set to NULL, or
2079 * - is a queue whose owning bics have already been redirected to a
2080 * different queue, hence the queue is destined to not belong to
2081 * any bic soon and bfqq->bic is already NULL (therefore the next
2082 * assignment causes no harm).
2084 new_bfqq
->bic
= NULL
;
2086 /* release process reference to bfqq */
2087 bfq_put_queue(bfqq
);
2090 static bool bfq_allow_bio_merge(struct request_queue
*q
, struct request
*rq
,
2093 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
2094 bool is_sync
= op_is_sync(bio
->bi_opf
);
2095 struct bfq_queue
*bfqq
= bfqd
->bio_bfqq
, *new_bfqq
;
2098 * Disallow merge of a sync bio into an async request.
2100 if (is_sync
&& !rq_is_sync(rq
))
2104 * Lookup the bfqq that this bio will be queued with. Allow
2105 * merge only if rq is queued there.
2111 * We take advantage of this function to perform an early merge
2112 * of the queues of possible cooperating processes.
2114 new_bfqq
= bfq_setup_cooperator(bfqd
, bfqq
, bio
, false);
2117 * bic still points to bfqq, then it has not yet been
2118 * redirected to some other bfq_queue, and a queue
2119 * merge beween bfqq and new_bfqq can be safely
2120 * fulfillled, i.e., bic can be redirected to new_bfqq
2121 * and bfqq can be put.
2123 bfq_merge_bfqqs(bfqd
, bfqd
->bio_bic
, bfqq
,
2126 * If we get here, bio will be queued into new_queue,
2127 * so use new_bfqq to decide whether bio and rq can be
2133 * Change also bqfd->bio_bfqq, as
2134 * bfqd->bio_bic now points to new_bfqq, and
2135 * this function may be invoked again (and then may
2136 * use again bqfd->bio_bfqq).
2138 bfqd
->bio_bfqq
= bfqq
;
2141 return bfqq
== RQ_BFQQ(rq
);
2145 * Set the maximum time for the in-service queue to consume its
2146 * budget. This prevents seeky processes from lowering the throughput.
2147 * In practice, a time-slice service scheme is used with seeky
2150 static void bfq_set_budget_timeout(struct bfq_data
*bfqd
,
2151 struct bfq_queue
*bfqq
)
2153 unsigned int timeout_coeff
;
2155 if (bfqq
->wr_cur_max_time
== bfqd
->bfq_wr_rt_max_time
)
2158 timeout_coeff
= bfqq
->entity
.weight
/ bfqq
->entity
.orig_weight
;
2160 bfqd
->last_budget_start
= ktime_get();
2162 bfqq
->budget_timeout
= jiffies
+
2163 bfqd
->bfq_timeout
* timeout_coeff
;
2166 static void __bfq_set_in_service_queue(struct bfq_data
*bfqd
,
2167 struct bfq_queue
*bfqq
)
2170 bfqg_stats_update_avg_queue_size(bfqq_group(bfqq
));
2171 bfq_clear_bfqq_fifo_expire(bfqq
);
2173 bfqd
->budgets_assigned
= (bfqd
->budgets_assigned
* 7 + 256) / 8;
2175 if (time_is_before_jiffies(bfqq
->last_wr_start_finish
) &&
2176 bfqq
->wr_coeff
> 1 &&
2177 bfqq
->wr_cur_max_time
== bfqd
->bfq_wr_rt_max_time
&&
2178 time_is_before_jiffies(bfqq
->budget_timeout
)) {
2180 * For soft real-time queues, move the start
2181 * of the weight-raising period forward by the
2182 * time the queue has not received any
2183 * service. Otherwise, a relatively long
2184 * service delay is likely to cause the
2185 * weight-raising period of the queue to end,
2186 * because of the short duration of the
2187 * weight-raising period of a soft real-time
2188 * queue. It is worth noting that this move
2189 * is not so dangerous for the other queues,
2190 * because soft real-time queues are not
2193 * To not add a further variable, we use the
2194 * overloaded field budget_timeout to
2195 * determine for how long the queue has not
2196 * received service, i.e., how much time has
2197 * elapsed since the queue expired. However,
2198 * this is a little imprecise, because
2199 * budget_timeout is set to jiffies if bfqq
2200 * not only expires, but also remains with no
2203 if (time_after(bfqq
->budget_timeout
,
2204 bfqq
->last_wr_start_finish
))
2205 bfqq
->last_wr_start_finish
+=
2206 jiffies
- bfqq
->budget_timeout
;
2208 bfqq
->last_wr_start_finish
= jiffies
;
2211 bfq_set_budget_timeout(bfqd
, bfqq
);
2212 bfq_log_bfqq(bfqd
, bfqq
,
2213 "set_in_service_queue, cur-budget = %d",
2214 bfqq
->entity
.budget
);
2217 bfqd
->in_service_queue
= bfqq
;
2221 * Get and set a new queue for service.
2223 static struct bfq_queue
*bfq_set_in_service_queue(struct bfq_data
*bfqd
)
2225 struct bfq_queue
*bfqq
= bfq_get_next_queue(bfqd
);
2227 __bfq_set_in_service_queue(bfqd
, bfqq
);
2231 static void bfq_arm_slice_timer(struct bfq_data
*bfqd
)
2233 struct bfq_queue
*bfqq
= bfqd
->in_service_queue
;
2236 bfq_mark_bfqq_wait_request(bfqq
);
2239 * We don't want to idle for seeks, but we do want to allow
2240 * fair distribution of slice time for a process doing back-to-back
2241 * seeks. So allow a little bit of time for him to submit a new rq.
2243 sl
= bfqd
->bfq_slice_idle
;
2245 * Unless the queue is being weight-raised or the scenario is
2246 * asymmetric, grant only minimum idle time if the queue
2247 * is seeky. A long idling is preserved for a weight-raised
2248 * queue, or, more in general, in an asymmetric scenario,
2249 * because a long idling is needed for guaranteeing to a queue
2250 * its reserved share of the throughput (in particular, it is
2251 * needed if the queue has a higher weight than some other
2254 if (BFQQ_SEEKY(bfqq
) && bfqq
->wr_coeff
== 1 &&
2255 bfq_symmetric_scenario(bfqd
))
2256 sl
= min_t(u64
, sl
, BFQ_MIN_TT
);
2258 bfqd
->last_idling_start
= ktime_get();
2259 hrtimer_start(&bfqd
->idle_slice_timer
, ns_to_ktime(sl
),
2261 bfqg_stats_set_start_idle_time(bfqq_group(bfqq
));
2265 * In autotuning mode, max_budget is dynamically recomputed as the
2266 * amount of sectors transferred in timeout at the estimated peak
2267 * rate. This enables BFQ to utilize a full timeslice with a full
2268 * budget, even if the in-service queue is served at peak rate. And
2269 * this maximises throughput with sequential workloads.
2271 static unsigned long bfq_calc_max_budget(struct bfq_data
*bfqd
)
2273 return (u64
)bfqd
->peak_rate
* USEC_PER_MSEC
*
2274 jiffies_to_msecs(bfqd
->bfq_timeout
)>>BFQ_RATE_SHIFT
;
2278 * Update parameters related to throughput and responsiveness, as a
2279 * function of the estimated peak rate. See comments on
2280 * bfq_calc_max_budget(), and on T_slow and T_fast arrays.
2282 static void update_thr_responsiveness_params(struct bfq_data
*bfqd
)
2284 int dev_type
= blk_queue_nonrot(bfqd
->queue
);
2286 if (bfqd
->bfq_user_max_budget
== 0)
2287 bfqd
->bfq_max_budget
=
2288 bfq_calc_max_budget(bfqd
);
2290 if (bfqd
->device_speed
== BFQ_BFQD_FAST
&&
2291 bfqd
->peak_rate
< device_speed_thresh
[dev_type
]) {
2292 bfqd
->device_speed
= BFQ_BFQD_SLOW
;
2293 bfqd
->RT_prod
= R_slow
[dev_type
] *
2295 } else if (bfqd
->device_speed
== BFQ_BFQD_SLOW
&&
2296 bfqd
->peak_rate
> device_speed_thresh
[dev_type
]) {
2297 bfqd
->device_speed
= BFQ_BFQD_FAST
;
2298 bfqd
->RT_prod
= R_fast
[dev_type
] *
2303 "dev_type %s dev_speed_class = %s (%llu sects/sec), thresh %llu setcs/sec",
2304 dev_type
== 0 ? "ROT" : "NONROT",
2305 bfqd
->device_speed
== BFQ_BFQD_FAST
? "FAST" : "SLOW",
2306 bfqd
->device_speed
== BFQ_BFQD_FAST
?
2307 (USEC_PER_SEC
*(u64
)R_fast
[dev_type
])>>BFQ_RATE_SHIFT
:
2308 (USEC_PER_SEC
*(u64
)R_slow
[dev_type
])>>BFQ_RATE_SHIFT
,
2309 (USEC_PER_SEC
*(u64
)device_speed_thresh
[dev_type
])>>
2313 static void bfq_reset_rate_computation(struct bfq_data
*bfqd
,
2316 if (rq
!= NULL
) { /* new rq dispatch now, reset accordingly */
2317 bfqd
->last_dispatch
= bfqd
->first_dispatch
= ktime_get_ns();
2318 bfqd
->peak_rate_samples
= 1;
2319 bfqd
->sequential_samples
= 0;
2320 bfqd
->tot_sectors_dispatched
= bfqd
->last_rq_max_size
=
2322 } else /* no new rq dispatched, just reset the number of samples */
2323 bfqd
->peak_rate_samples
= 0; /* full re-init on next disp. */
2326 "reset_rate_computation at end, sample %u/%u tot_sects %llu",
2327 bfqd
->peak_rate_samples
, bfqd
->sequential_samples
,
2328 bfqd
->tot_sectors_dispatched
);
2331 static void bfq_update_rate_reset(struct bfq_data
*bfqd
, struct request
*rq
)
2333 u32 rate
, weight
, divisor
;
2336 * For the convergence property to hold (see comments on
2337 * bfq_update_peak_rate()) and for the assessment to be
2338 * reliable, a minimum number of samples must be present, and
2339 * a minimum amount of time must have elapsed. If not so, do
2340 * not compute new rate. Just reset parameters, to get ready
2341 * for a new evaluation attempt.
2343 if (bfqd
->peak_rate_samples
< BFQ_RATE_MIN_SAMPLES
||
2344 bfqd
->delta_from_first
< BFQ_RATE_MIN_INTERVAL
)
2345 goto reset_computation
;
2348 * If a new request completion has occurred after last
2349 * dispatch, then, to approximate the rate at which requests
2350 * have been served by the device, it is more precise to
2351 * extend the observation interval to the last completion.
2353 bfqd
->delta_from_first
=
2354 max_t(u64
, bfqd
->delta_from_first
,
2355 bfqd
->last_completion
- bfqd
->first_dispatch
);
2358 * Rate computed in sects/usec, and not sects/nsec, for
2361 rate
= div64_ul(bfqd
->tot_sectors_dispatched
<<BFQ_RATE_SHIFT
,
2362 div_u64(bfqd
->delta_from_first
, NSEC_PER_USEC
));
2365 * Peak rate not updated if:
2366 * - the percentage of sequential dispatches is below 3/4 of the
2367 * total, and rate is below the current estimated peak rate
2368 * - rate is unreasonably high (> 20M sectors/sec)
2370 if ((bfqd
->sequential_samples
< (3 * bfqd
->peak_rate_samples
)>>2 &&
2371 rate
<= bfqd
->peak_rate
) ||
2372 rate
> 20<<BFQ_RATE_SHIFT
)
2373 goto reset_computation
;
2376 * We have to update the peak rate, at last! To this purpose,
2377 * we use a low-pass filter. We compute the smoothing constant
2378 * of the filter as a function of the 'weight' of the new
2381 * As can be seen in next formulas, we define this weight as a
2382 * quantity proportional to how sequential the workload is,
2383 * and to how long the observation time interval is.
2385 * The weight runs from 0 to 8. The maximum value of the
2386 * weight, 8, yields the minimum value for the smoothing
2387 * constant. At this minimum value for the smoothing constant,
2388 * the measured rate contributes for half of the next value of
2389 * the estimated peak rate.
2391 * So, the first step is to compute the weight as a function
2392 * of how sequential the workload is. Note that the weight
2393 * cannot reach 9, because bfqd->sequential_samples cannot
2394 * become equal to bfqd->peak_rate_samples, which, in its
2395 * turn, holds true because bfqd->sequential_samples is not
2396 * incremented for the first sample.
2398 weight
= (9 * bfqd
->sequential_samples
) / bfqd
->peak_rate_samples
;
2401 * Second step: further refine the weight as a function of the
2402 * duration of the observation interval.
2404 weight
= min_t(u32
, 8,
2405 div_u64(weight
* bfqd
->delta_from_first
,
2406 BFQ_RATE_REF_INTERVAL
));
2409 * Divisor ranging from 10, for minimum weight, to 2, for
2412 divisor
= 10 - weight
;
2415 * Finally, update peak rate:
2417 * peak_rate = peak_rate * (divisor-1) / divisor + rate / divisor
2419 bfqd
->peak_rate
*= divisor
-1;
2420 bfqd
->peak_rate
/= divisor
;
2421 rate
/= divisor
; /* smoothing constant alpha = 1/divisor */
2423 bfqd
->peak_rate
+= rate
;
2424 update_thr_responsiveness_params(bfqd
);
2427 bfq_reset_rate_computation(bfqd
, rq
);
2431 * Update the read/write peak rate (the main quantity used for
2432 * auto-tuning, see update_thr_responsiveness_params()).
2434 * It is not trivial to estimate the peak rate (correctly): because of
2435 * the presence of sw and hw queues between the scheduler and the
2436 * device components that finally serve I/O requests, it is hard to
2437 * say exactly when a given dispatched request is served inside the
2438 * device, and for how long. As a consequence, it is hard to know
2439 * precisely at what rate a given set of requests is actually served
2442 * On the opposite end, the dispatch time of any request is trivially
2443 * available, and, from this piece of information, the "dispatch rate"
2444 * of requests can be immediately computed. So, the idea in the next
2445 * function is to use what is known, namely request dispatch times
2446 * (plus, when useful, request completion times), to estimate what is
2447 * unknown, namely in-device request service rate.
2449 * The main issue is that, because of the above facts, the rate at
2450 * which a certain set of requests is dispatched over a certain time
2451 * interval can vary greatly with respect to the rate at which the
2452 * same requests are then served. But, since the size of any
2453 * intermediate queue is limited, and the service scheme is lossless
2454 * (no request is silently dropped), the following obvious convergence
2455 * property holds: the number of requests dispatched MUST become
2456 * closer and closer to the number of requests completed as the
2457 * observation interval grows. This is the key property used in
2458 * the next function to estimate the peak service rate as a function
2459 * of the observed dispatch rate. The function assumes to be invoked
2460 * on every request dispatch.
2462 static void bfq_update_peak_rate(struct bfq_data
*bfqd
, struct request
*rq
)
2464 u64 now_ns
= ktime_get_ns();
2466 if (bfqd
->peak_rate_samples
== 0) { /* first dispatch */
2467 bfq_log(bfqd
, "update_peak_rate: goto reset, samples %d",
2468 bfqd
->peak_rate_samples
);
2469 bfq_reset_rate_computation(bfqd
, rq
);
2470 goto update_last_values
; /* will add one sample */
2474 * Device idle for very long: the observation interval lasting
2475 * up to this dispatch cannot be a valid observation interval
2476 * for computing a new peak rate (similarly to the late-
2477 * completion event in bfq_completed_request()). Go to
2478 * update_rate_and_reset to have the following three steps
2480 * - close the observation interval at the last (previous)
2481 * request dispatch or completion
2482 * - compute rate, if possible, for that observation interval
2483 * - start a new observation interval with this dispatch
2485 if (now_ns
- bfqd
->last_dispatch
> 100*NSEC_PER_MSEC
&&
2486 bfqd
->rq_in_driver
== 0)
2487 goto update_rate_and_reset
;
2489 /* Update sampling information */
2490 bfqd
->peak_rate_samples
++;
2492 if ((bfqd
->rq_in_driver
> 0 ||
2493 now_ns
- bfqd
->last_completion
< BFQ_MIN_TT
)
2494 && get_sdist(bfqd
->last_position
, rq
) < BFQQ_SEEK_THR
)
2495 bfqd
->sequential_samples
++;
2497 bfqd
->tot_sectors_dispatched
+= blk_rq_sectors(rq
);
2499 /* Reset max observed rq size every 32 dispatches */
2500 if (likely(bfqd
->peak_rate_samples
% 32))
2501 bfqd
->last_rq_max_size
=
2502 max_t(u32
, blk_rq_sectors(rq
), bfqd
->last_rq_max_size
);
2504 bfqd
->last_rq_max_size
= blk_rq_sectors(rq
);
2506 bfqd
->delta_from_first
= now_ns
- bfqd
->first_dispatch
;
2508 /* Target observation interval not yet reached, go on sampling */
2509 if (bfqd
->delta_from_first
< BFQ_RATE_REF_INTERVAL
)
2510 goto update_last_values
;
2512 update_rate_and_reset
:
2513 bfq_update_rate_reset(bfqd
, rq
);
2515 bfqd
->last_position
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
2516 bfqd
->last_dispatch
= now_ns
;
2520 * Remove request from internal lists.
2522 static void bfq_dispatch_remove(struct request_queue
*q
, struct request
*rq
)
2524 struct bfq_queue
*bfqq
= RQ_BFQQ(rq
);
2527 * For consistency, the next instruction should have been
2528 * executed after removing the request from the queue and
2529 * dispatching it. We execute instead this instruction before
2530 * bfq_remove_request() (and hence introduce a temporary
2531 * inconsistency), for efficiency. In fact, should this
2532 * dispatch occur for a non in-service bfqq, this anticipated
2533 * increment prevents two counters related to bfqq->dispatched
2534 * from risking to be, first, uselessly decremented, and then
2535 * incremented again when the (new) value of bfqq->dispatched
2536 * happens to be taken into account.
2539 bfq_update_peak_rate(q
->elevator
->elevator_data
, rq
);
2541 bfq_remove_request(q
, rq
);
2544 static void __bfq_bfqq_expire(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
2547 * If this bfqq is shared between multiple processes, check
2548 * to make sure that those processes are still issuing I/Os
2549 * within the mean seek distance. If not, it may be time to
2550 * break the queues apart again.
2552 if (bfq_bfqq_coop(bfqq
) && BFQQ_SEEKY(bfqq
))
2553 bfq_mark_bfqq_split_coop(bfqq
);
2555 if (RB_EMPTY_ROOT(&bfqq
->sort_list
)) {
2556 if (bfqq
->dispatched
== 0)
2558 * Overloading budget_timeout field to store
2559 * the time at which the queue remains with no
2560 * backlog and no outstanding request; used by
2561 * the weight-raising mechanism.
2563 bfqq
->budget_timeout
= jiffies
;
2565 bfq_del_bfqq_busy(bfqd
, bfqq
, true);
2567 bfq_requeue_bfqq(bfqd
, bfqq
, true);
2569 * Resort priority tree of potential close cooperators.
2571 bfq_pos_tree_add_move(bfqd
, bfqq
);
2575 * All in-service entities must have been properly deactivated
2576 * or requeued before executing the next function, which
2577 * resets all in-service entites as no more in service.
2579 __bfq_bfqd_reset_in_service(bfqd
);
2583 * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior.
2584 * @bfqd: device data.
2585 * @bfqq: queue to update.
2586 * @reason: reason for expiration.
2588 * Handle the feedback on @bfqq budget at queue expiration.
2589 * See the body for detailed comments.
2591 static void __bfq_bfqq_recalc_budget(struct bfq_data
*bfqd
,
2592 struct bfq_queue
*bfqq
,
2593 enum bfqq_expiration reason
)
2595 struct request
*next_rq
;
2596 int budget
, min_budget
;
2598 min_budget
= bfq_min_budget(bfqd
);
2600 if (bfqq
->wr_coeff
== 1)
2601 budget
= bfqq
->max_budget
;
2603 * Use a constant, low budget for weight-raised queues,
2604 * to help achieve a low latency. Keep it slightly higher
2605 * than the minimum possible budget, to cause a little
2606 * bit fewer expirations.
2608 budget
= 2 * min_budget
;
2610 bfq_log_bfqq(bfqd
, bfqq
, "recalc_budg: last budg %d, budg left %d",
2611 bfqq
->entity
.budget
, bfq_bfqq_budget_left(bfqq
));
2612 bfq_log_bfqq(bfqd
, bfqq
, "recalc_budg: last max_budg %d, min budg %d",
2613 budget
, bfq_min_budget(bfqd
));
2614 bfq_log_bfqq(bfqd
, bfqq
, "recalc_budg: sync %d, seeky %d",
2615 bfq_bfqq_sync(bfqq
), BFQQ_SEEKY(bfqd
->in_service_queue
));
2617 if (bfq_bfqq_sync(bfqq
) && bfqq
->wr_coeff
== 1) {
2620 * Caveat: in all the following cases we trade latency
2623 case BFQQE_TOO_IDLE
:
2625 * This is the only case where we may reduce
2626 * the budget: if there is no request of the
2627 * process still waiting for completion, then
2628 * we assume (tentatively) that the timer has
2629 * expired because the batch of requests of
2630 * the process could have been served with a
2631 * smaller budget. Hence, betting that
2632 * process will behave in the same way when it
2633 * becomes backlogged again, we reduce its
2634 * next budget. As long as we guess right,
2635 * this budget cut reduces the latency
2636 * experienced by the process.
2638 * However, if there are still outstanding
2639 * requests, then the process may have not yet
2640 * issued its next request just because it is
2641 * still waiting for the completion of some of
2642 * the still outstanding ones. So in this
2643 * subcase we do not reduce its budget, on the
2644 * contrary we increase it to possibly boost
2645 * the throughput, as discussed in the
2646 * comments to the BUDGET_TIMEOUT case.
2648 if (bfqq
->dispatched
> 0) /* still outstanding reqs */
2649 budget
= min(budget
* 2, bfqd
->bfq_max_budget
);
2651 if (budget
> 5 * min_budget
)
2652 budget
-= 4 * min_budget
;
2654 budget
= min_budget
;
2657 case BFQQE_BUDGET_TIMEOUT
:
2659 * We double the budget here because it gives
2660 * the chance to boost the throughput if this
2661 * is not a seeky process (and has bumped into
2662 * this timeout because of, e.g., ZBR).
2664 budget
= min(budget
* 2, bfqd
->bfq_max_budget
);
2666 case BFQQE_BUDGET_EXHAUSTED
:
2668 * The process still has backlog, and did not
2669 * let either the budget timeout or the disk
2670 * idling timeout expire. Hence it is not
2671 * seeky, has a short thinktime and may be
2672 * happy with a higher budget too. So
2673 * definitely increase the budget of this good
2674 * candidate to boost the disk throughput.
2676 budget
= min(budget
* 4, bfqd
->bfq_max_budget
);
2678 case BFQQE_NO_MORE_REQUESTS
:
2680 * For queues that expire for this reason, it
2681 * is particularly important to keep the
2682 * budget close to the actual service they
2683 * need. Doing so reduces the timestamp
2684 * misalignment problem described in the
2685 * comments in the body of
2686 * __bfq_activate_entity. In fact, suppose
2687 * that a queue systematically expires for
2688 * BFQQE_NO_MORE_REQUESTS and presents a
2689 * new request in time to enjoy timestamp
2690 * back-shifting. The larger the budget of the
2691 * queue is with respect to the service the
2692 * queue actually requests in each service
2693 * slot, the more times the queue can be
2694 * reactivated with the same virtual finish
2695 * time. It follows that, even if this finish
2696 * time is pushed to the system virtual time
2697 * to reduce the consequent timestamp
2698 * misalignment, the queue unjustly enjoys for
2699 * many re-activations a lower finish time
2700 * than all newly activated queues.
2702 * The service needed by bfqq is measured
2703 * quite precisely by bfqq->entity.service.
2704 * Since bfqq does not enjoy device idling,
2705 * bfqq->entity.service is equal to the number
2706 * of sectors that the process associated with
2707 * bfqq requested to read/write before waiting
2708 * for request completions, or blocking for
2711 budget
= max_t(int, bfqq
->entity
.service
, min_budget
);
2716 } else if (!bfq_bfqq_sync(bfqq
)) {
2718 * Async queues get always the maximum possible
2719 * budget, as for them we do not care about latency
2720 * (in addition, their ability to dispatch is limited
2721 * by the charging factor).
2723 budget
= bfqd
->bfq_max_budget
;
2726 bfqq
->max_budget
= budget
;
2728 if (bfqd
->budgets_assigned
>= bfq_stats_min_budgets
&&
2729 !bfqd
->bfq_user_max_budget
)
2730 bfqq
->max_budget
= min(bfqq
->max_budget
, bfqd
->bfq_max_budget
);
2733 * If there is still backlog, then assign a new budget, making
2734 * sure that it is large enough for the next request. Since
2735 * the finish time of bfqq must be kept in sync with the
2736 * budget, be sure to call __bfq_bfqq_expire() *after* this
2739 * If there is no backlog, then no need to update the budget;
2740 * it will be updated on the arrival of a new request.
2742 next_rq
= bfqq
->next_rq
;
2744 bfqq
->entity
.budget
= max_t(unsigned long, bfqq
->max_budget
,
2745 bfq_serv_to_charge(next_rq
, bfqq
));
2747 bfq_log_bfqq(bfqd
, bfqq
, "head sect: %u, new budget %d",
2748 next_rq
? blk_rq_sectors(next_rq
) : 0,
2749 bfqq
->entity
.budget
);
2753 * Return true if the process associated with bfqq is "slow". The slow
2754 * flag is used, in addition to the budget timeout, to reduce the
2755 * amount of service provided to seeky processes, and thus reduce
2756 * their chances to lower the throughput. More details in the comments
2757 * on the function bfq_bfqq_expire().
2759 * An important observation is in order: as discussed in the comments
2760 * on the function bfq_update_peak_rate(), with devices with internal
2761 * queues, it is hard if ever possible to know when and for how long
2762 * an I/O request is processed by the device (apart from the trivial
2763 * I/O pattern where a new request is dispatched only after the
2764 * previous one has been completed). This makes it hard to evaluate
2765 * the real rate at which the I/O requests of each bfq_queue are
2766 * served. In fact, for an I/O scheduler like BFQ, serving a
2767 * bfq_queue means just dispatching its requests during its service
2768 * slot (i.e., until the budget of the queue is exhausted, or the
2769 * queue remains idle, or, finally, a timeout fires). But, during the
2770 * service slot of a bfq_queue, around 100 ms at most, the device may
2771 * be even still processing requests of bfq_queues served in previous
2772 * service slots. On the opposite end, the requests of the in-service
2773 * bfq_queue may be completed after the service slot of the queue
2776 * Anyway, unless more sophisticated solutions are used
2777 * (where possible), the sum of the sizes of the requests dispatched
2778 * during the service slot of a bfq_queue is probably the only
2779 * approximation available for the service received by the bfq_queue
2780 * during its service slot. And this sum is the quantity used in this
2781 * function to evaluate the I/O speed of a process.
2783 static bool bfq_bfqq_is_slow(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
2784 bool compensate
, enum bfqq_expiration reason
,
2785 unsigned long *delta_ms
)
2787 ktime_t delta_ktime
;
2789 bool slow
= BFQQ_SEEKY(bfqq
); /* if delta too short, use seekyness */
2791 if (!bfq_bfqq_sync(bfqq
))
2795 delta_ktime
= bfqd
->last_idling_start
;
2797 delta_ktime
= ktime_get();
2798 delta_ktime
= ktime_sub(delta_ktime
, bfqd
->last_budget_start
);
2799 delta_usecs
= ktime_to_us(delta_ktime
);
2801 /* don't use too short time intervals */
2802 if (delta_usecs
< 1000) {
2803 if (blk_queue_nonrot(bfqd
->queue
))
2805 * give same worst-case guarantees as idling
2808 *delta_ms
= BFQ_MIN_TT
/ NSEC_PER_MSEC
;
2809 else /* charge at least one seek */
2810 *delta_ms
= bfq_slice_idle
/ NSEC_PER_MSEC
;
2815 *delta_ms
= delta_usecs
/ USEC_PER_MSEC
;
2818 * Use only long (> 20ms) intervals to filter out excessive
2819 * spikes in service rate estimation.
2821 if (delta_usecs
> 20000) {
2823 * Caveat for rotational devices: processes doing I/O
2824 * in the slower disk zones tend to be slow(er) even
2825 * if not seeky. In this respect, the estimated peak
2826 * rate is likely to be an average over the disk
2827 * surface. Accordingly, to not be too harsh with
2828 * unlucky processes, a process is deemed slow only if
2829 * its rate has been lower than half of the estimated
2832 slow
= bfqq
->entity
.service
< bfqd
->bfq_max_budget
/ 2;
2835 bfq_log_bfqq(bfqd
, bfqq
, "bfq_bfqq_is_slow: slow %d", slow
);
2841 * To be deemed as soft real-time, an application must meet two
2842 * requirements. First, the application must not require an average
2843 * bandwidth higher than the approximate bandwidth required to playback or
2844 * record a compressed high-definition video.
2845 * The next function is invoked on the completion of the last request of a
2846 * batch, to compute the next-start time instant, soft_rt_next_start, such
2847 * that, if the next request of the application does not arrive before
2848 * soft_rt_next_start, then the above requirement on the bandwidth is met.
2850 * The second requirement is that the request pattern of the application is
2851 * isochronous, i.e., that, after issuing a request or a batch of requests,
2852 * the application stops issuing new requests until all its pending requests
2853 * have been completed. After that, the application may issue a new batch,
2855 * For this reason the next function is invoked to compute
2856 * soft_rt_next_start only for applications that meet this requirement,
2857 * whereas soft_rt_next_start is set to infinity for applications that do
2860 * Unfortunately, even a greedy application may happen to behave in an
2861 * isochronous way if the CPU load is high. In fact, the application may
2862 * stop issuing requests while the CPUs are busy serving other processes,
2863 * then restart, then stop again for a while, and so on. In addition, if
2864 * the disk achieves a low enough throughput with the request pattern
2865 * issued by the application (e.g., because the request pattern is random
2866 * and/or the device is slow), then the application may meet the above
2867 * bandwidth requirement too. To prevent such a greedy application to be
2868 * deemed as soft real-time, a further rule is used in the computation of
2869 * soft_rt_next_start: soft_rt_next_start must be higher than the current
2870 * time plus the maximum time for which the arrival of a request is waited
2871 * for when a sync queue becomes idle, namely bfqd->bfq_slice_idle.
2872 * This filters out greedy applications, as the latter issue instead their
2873 * next request as soon as possible after the last one has been completed
2874 * (in contrast, when a batch of requests is completed, a soft real-time
2875 * application spends some time processing data).
2877 * Unfortunately, the last filter may easily generate false positives if
2878 * only bfqd->bfq_slice_idle is used as a reference time interval and one
2879 * or both the following cases occur:
2880 * 1) HZ is so low that the duration of a jiffy is comparable to or higher
2881 * than bfqd->bfq_slice_idle. This happens, e.g., on slow devices with
2883 * 2) jiffies, instead of increasing at a constant rate, may stop increasing
2884 * for a while, then suddenly 'jump' by several units to recover the lost
2885 * increments. This seems to happen, e.g., inside virtual machines.
2886 * To address this issue, we do not use as a reference time interval just
2887 * bfqd->bfq_slice_idle, but bfqd->bfq_slice_idle plus a few jiffies. In
2888 * particular we add the minimum number of jiffies for which the filter
2889 * seems to be quite precise also in embedded systems and KVM/QEMU virtual
2892 static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data
*bfqd
,
2893 struct bfq_queue
*bfqq
)
2895 return max(bfqq
->last_idle_bklogged
+
2896 HZ
* bfqq
->service_from_backlogged
/
2897 bfqd
->bfq_wr_max_softrt_rate
,
2898 jiffies
+ nsecs_to_jiffies(bfqq
->bfqd
->bfq_slice_idle
) + 4);
2902 * Return the farthest future time instant according to jiffies
2905 static unsigned long bfq_greatest_from_now(void)
2907 return jiffies
+ MAX_JIFFY_OFFSET
;
2911 * Return the farthest past time instant according to jiffies
2914 static unsigned long bfq_smallest_from_now(void)
2916 return jiffies
- MAX_JIFFY_OFFSET
;
2920 * bfq_bfqq_expire - expire a queue.
2921 * @bfqd: device owning the queue.
2922 * @bfqq: the queue to expire.
2923 * @compensate: if true, compensate for the time spent idling.
2924 * @reason: the reason causing the expiration.
2926 * If the process associated with bfqq does slow I/O (e.g., because it
2927 * issues random requests), we charge bfqq with the time it has been
2928 * in service instead of the service it has received (see
2929 * bfq_bfqq_charge_time for details on how this goal is achieved). As
2930 * a consequence, bfqq will typically get higher timestamps upon
2931 * reactivation, and hence it will be rescheduled as if it had
2932 * received more service than what it has actually received. In the
2933 * end, bfqq receives less service in proportion to how slowly its
2934 * associated process consumes its budgets (and hence how seriously it
2935 * tends to lower the throughput). In addition, this time-charging
2936 * strategy guarantees time fairness among slow processes. In
2937 * contrast, if the process associated with bfqq is not slow, we
2938 * charge bfqq exactly with the service it has received.
2940 * Charging time to the first type of queues and the exact service to
2941 * the other has the effect of using the WF2Q+ policy to schedule the
2942 * former on a timeslice basis, without violating service domain
2943 * guarantees among the latter.
2945 void bfq_bfqq_expire(struct bfq_data
*bfqd
,
2946 struct bfq_queue
*bfqq
,
2948 enum bfqq_expiration reason
)
2951 unsigned long delta
= 0;
2952 struct bfq_entity
*entity
= &bfqq
->entity
;
2956 * Check whether the process is slow (see bfq_bfqq_is_slow).
2958 slow
= bfq_bfqq_is_slow(bfqd
, bfqq
, compensate
, reason
, &delta
);
2961 * Increase service_from_backlogged before next statement,
2962 * because the possible next invocation of
2963 * bfq_bfqq_charge_time would likely inflate
2964 * entity->service. In contrast, service_from_backlogged must
2965 * contain real service, to enable the soft real-time
2966 * heuristic to correctly compute the bandwidth consumed by
2969 bfqq
->service_from_backlogged
+= entity
->service
;
2972 * As above explained, charge slow (typically seeky) and
2973 * timed-out queues with the time and not the service
2974 * received, to favor sequential workloads.
2976 * Processes doing I/O in the slower disk zones will tend to
2977 * be slow(er) even if not seeky. Therefore, since the
2978 * estimated peak rate is actually an average over the disk
2979 * surface, these processes may timeout just for bad luck. To
2980 * avoid punishing them, do not charge time to processes that
2981 * succeeded in consuming at least 2/3 of their budget. This
2982 * allows BFQ to preserve enough elasticity to still perform
2983 * bandwidth, and not time, distribution with little unlucky
2984 * or quasi-sequential processes.
2986 if (bfqq
->wr_coeff
== 1 &&
2988 (reason
== BFQQE_BUDGET_TIMEOUT
&&
2989 bfq_bfqq_budget_left(bfqq
) >= entity
->budget
/ 3)))
2990 bfq_bfqq_charge_time(bfqd
, bfqq
, delta
);
2992 if (reason
== BFQQE_TOO_IDLE
&&
2993 entity
->service
<= 2 * entity
->budget
/ 10)
2994 bfq_clear_bfqq_IO_bound(bfqq
);
2996 if (bfqd
->low_latency
&& bfqq
->wr_coeff
== 1)
2997 bfqq
->last_wr_start_finish
= jiffies
;
2999 if (bfqd
->low_latency
&& bfqd
->bfq_wr_max_softrt_rate
> 0 &&
3000 RB_EMPTY_ROOT(&bfqq
->sort_list
)) {
3002 * If we get here, and there are no outstanding
3003 * requests, then the request pattern is isochronous
3004 * (see the comments on the function
3005 * bfq_bfqq_softrt_next_start()). Thus we can compute
3006 * soft_rt_next_start. If, instead, the queue still
3007 * has outstanding requests, then we have to wait for
3008 * the completion of all the outstanding requests to
3009 * discover whether the request pattern is actually
3012 if (bfqq
->dispatched
== 0)
3013 bfqq
->soft_rt_next_start
=
3014 bfq_bfqq_softrt_next_start(bfqd
, bfqq
);
3017 * The application is still waiting for the
3018 * completion of one or more requests:
3019 * prevent it from possibly being incorrectly
3020 * deemed as soft real-time by setting its
3021 * soft_rt_next_start to infinity. In fact,
3022 * without this assignment, the application
3023 * would be incorrectly deemed as soft
3025 * 1) it issued a new request before the
3026 * completion of all its in-flight
3028 * 2) at that time, its soft_rt_next_start
3029 * happened to be in the past.
3031 bfqq
->soft_rt_next_start
=
3032 bfq_greatest_from_now();
3034 * Schedule an update of soft_rt_next_start to when
3035 * the task may be discovered to be isochronous.
3037 bfq_mark_bfqq_softrt_update(bfqq
);
3041 bfq_log_bfqq(bfqd
, bfqq
,
3042 "expire (%d, slow %d, num_disp %d, short_ttime %d)", reason
,
3043 slow
, bfqq
->dispatched
, bfq_bfqq_has_short_ttime(bfqq
));
3046 * Increase, decrease or leave budget unchanged according to
3049 __bfq_bfqq_recalc_budget(bfqd
, bfqq
, reason
);
3051 __bfq_bfqq_expire(bfqd
, bfqq
);
3053 /* mark bfqq as waiting a request only if a bic still points to it */
3054 if (ref
> 1 && !bfq_bfqq_busy(bfqq
) &&
3055 reason
!= BFQQE_BUDGET_TIMEOUT
&&
3056 reason
!= BFQQE_BUDGET_EXHAUSTED
)
3057 bfq_mark_bfqq_non_blocking_wait_rq(bfqq
);
3061 * Budget timeout is not implemented through a dedicated timer, but
3062 * just checked on request arrivals and completions, as well as on
3063 * idle timer expirations.
3065 static bool bfq_bfqq_budget_timeout(struct bfq_queue
*bfqq
)
3067 return time_is_before_eq_jiffies(bfqq
->budget_timeout
);
3071 * If we expire a queue that is actively waiting (i.e., with the
3072 * device idled) for the arrival of a new request, then we may incur
3073 * the timestamp misalignment problem described in the body of the
3074 * function __bfq_activate_entity. Hence we return true only if this
3075 * condition does not hold, or if the queue is slow enough to deserve
3076 * only to be kicked off for preserving a high throughput.
3078 static bool bfq_may_expire_for_budg_timeout(struct bfq_queue
*bfqq
)
3080 bfq_log_bfqq(bfqq
->bfqd
, bfqq
,
3081 "may_budget_timeout: wait_request %d left %d timeout %d",
3082 bfq_bfqq_wait_request(bfqq
),
3083 bfq_bfqq_budget_left(bfqq
) >= bfqq
->entity
.budget
/ 3,
3084 bfq_bfqq_budget_timeout(bfqq
));
3086 return (!bfq_bfqq_wait_request(bfqq
) ||
3087 bfq_bfqq_budget_left(bfqq
) >= bfqq
->entity
.budget
/ 3)
3089 bfq_bfqq_budget_timeout(bfqq
);
3093 * For a queue that becomes empty, device idling is allowed only if
3094 * this function returns true for the queue. As a consequence, since
3095 * device idling plays a critical role in both throughput boosting and
3096 * service guarantees, the return value of this function plays a
3097 * critical role in both these aspects as well.
3099 * In a nutshell, this function returns true only if idling is
3100 * beneficial for throughput or, even if detrimental for throughput,
3101 * idling is however necessary to preserve service guarantees (low
3102 * latency, desired throughput distribution, ...). In particular, on
3103 * NCQ-capable devices, this function tries to return false, so as to
3104 * help keep the drives' internal queues full, whenever this helps the
3105 * device boost the throughput without causing any service-guarantee
3108 * In more detail, the return value of this function is obtained by,
3109 * first, computing a number of boolean variables that take into
3110 * account throughput and service-guarantee issues, and, then,
3111 * combining these variables in a logical expression. Most of the
3112 * issues taken into account are not trivial. We discuss these issues
3113 * individually while introducing the variables.
3115 static bool bfq_bfqq_may_idle(struct bfq_queue
*bfqq
)
3117 struct bfq_data
*bfqd
= bfqq
->bfqd
;
3118 bool rot_without_queueing
=
3119 !blk_queue_nonrot(bfqd
->queue
) && !bfqd
->hw_tag
,
3120 bfqq_sequential_and_IO_bound
,
3121 idling_boosts_thr
, idling_boosts_thr_without_issues
,
3122 idling_needed_for_service_guarantees
,
3123 asymmetric_scenario
;
3125 if (bfqd
->strict_guarantees
)
3129 * Idling is performed only if slice_idle > 0. In addition, we
3132 * (b) bfqq is in the idle io prio class: in this case we do
3133 * not idle because we want to minimize the bandwidth that
3134 * queues in this class can steal to higher-priority queues
3136 if (bfqd
->bfq_slice_idle
== 0 || !bfq_bfqq_sync(bfqq
) ||
3137 bfq_class_idle(bfqq
))
3140 bfqq_sequential_and_IO_bound
= !BFQQ_SEEKY(bfqq
) &&
3141 bfq_bfqq_IO_bound(bfqq
) && bfq_bfqq_has_short_ttime(bfqq
);
3144 * The next variable takes into account the cases where idling
3145 * boosts the throughput.
3147 * The value of the variable is computed considering, first, that
3148 * idling is virtually always beneficial for the throughput if:
3149 * (a) the device is not NCQ-capable and rotational, or
3150 * (b) regardless of the presence of NCQ, the device is rotational and
3151 * the request pattern for bfqq is I/O-bound and sequential, or
3152 * (c) regardless of whether it is rotational, the device is
3153 * not NCQ-capable and the request pattern for bfqq is
3154 * I/O-bound and sequential.
3156 * Secondly, and in contrast to the above item (b), idling an
3157 * NCQ-capable flash-based device would not boost the
3158 * throughput even with sequential I/O; rather it would lower
3159 * the throughput in proportion to how fast the device
3160 * is. Accordingly, the next variable is true if any of the
3161 * above conditions (a), (b) or (c) is true, and, in
3162 * particular, happens to be false if bfqd is an NCQ-capable
3163 * flash-based device.
3165 idling_boosts_thr
= rot_without_queueing
||
3166 ((!blk_queue_nonrot(bfqd
->queue
) || !bfqd
->hw_tag
) &&
3167 bfqq_sequential_and_IO_bound
);
3170 * The value of the next variable,
3171 * idling_boosts_thr_without_issues, is equal to that of
3172 * idling_boosts_thr, unless a special case holds. In this
3173 * special case, described below, idling may cause problems to
3174 * weight-raised queues.
3176 * When the request pool is saturated (e.g., in the presence
3177 * of write hogs), if the processes associated with
3178 * non-weight-raised queues ask for requests at a lower rate,
3179 * then processes associated with weight-raised queues have a
3180 * higher probability to get a request from the pool
3181 * immediately (or at least soon) when they need one. Thus
3182 * they have a higher probability to actually get a fraction
3183 * of the device throughput proportional to their high
3184 * weight. This is especially true with NCQ-capable drives,
3185 * which enqueue several requests in advance, and further
3186 * reorder internally-queued requests.
3188 * For this reason, we force to false the value of
3189 * idling_boosts_thr_without_issues if there are weight-raised
3190 * busy queues. In this case, and if bfqq is not weight-raised,
3191 * this guarantees that the device is not idled for bfqq (if,
3192 * instead, bfqq is weight-raised, then idling will be
3193 * guaranteed by another variable, see below). Combined with
3194 * the timestamping rules of BFQ (see [1] for details), this
3195 * behavior causes bfqq, and hence any sync non-weight-raised
3196 * queue, to get a lower number of requests served, and thus
3197 * to ask for a lower number of requests from the request
3198 * pool, before the busy weight-raised queues get served
3199 * again. This often mitigates starvation problems in the
3200 * presence of heavy write workloads and NCQ, thereby
3201 * guaranteeing a higher application and system responsiveness
3202 * in these hostile scenarios.
3204 idling_boosts_thr_without_issues
= idling_boosts_thr
&&
3205 bfqd
->wr_busy_queues
== 0;
3208 * There is then a case where idling must be performed not
3209 * for throughput concerns, but to preserve service
3212 * To introduce this case, we can note that allowing the drive
3213 * to enqueue more than one request at a time, and hence
3214 * delegating de facto final scheduling decisions to the
3215 * drive's internal scheduler, entails loss of control on the
3216 * actual request service order. In particular, the critical
3217 * situation is when requests from different processes happen
3218 * to be present, at the same time, in the internal queue(s)
3219 * of the drive. In such a situation, the drive, by deciding
3220 * the service order of the internally-queued requests, does
3221 * determine also the actual throughput distribution among
3222 * these processes. But the drive typically has no notion or
3223 * concern about per-process throughput distribution, and
3224 * makes its decisions only on a per-request basis. Therefore,
3225 * the service distribution enforced by the drive's internal
3226 * scheduler is likely to coincide with the desired
3227 * device-throughput distribution only in a completely
3228 * symmetric scenario where:
3229 * (i) each of these processes must get the same throughput as
3231 * (ii) all these processes have the same I/O pattern
3232 (either sequential or random).
3233 * In fact, in such a scenario, the drive will tend to treat
3234 * the requests of each of these processes in about the same
3235 * way as the requests of the others, and thus to provide
3236 * each of these processes with about the same throughput
3237 * (which is exactly the desired throughput distribution). In
3238 * contrast, in any asymmetric scenario, device idling is
3239 * certainly needed to guarantee that bfqq receives its
3240 * assigned fraction of the device throughput (see [1] for
3243 * We address this issue by controlling, actually, only the
3244 * symmetry sub-condition (i), i.e., provided that
3245 * sub-condition (i) holds, idling is not performed,
3246 * regardless of whether sub-condition (ii) holds. In other
3247 * words, only if sub-condition (i) holds, then idling is
3248 * allowed, and the device tends to be prevented from queueing
3249 * many requests, possibly of several processes. The reason
3250 * for not controlling also sub-condition (ii) is that we
3251 * exploit preemption to preserve guarantees in case of
3252 * symmetric scenarios, even if (ii) does not hold, as
3253 * explained in the next two paragraphs.
3255 * Even if a queue, say Q, is expired when it remains idle, Q
3256 * can still preempt the new in-service queue if the next
3257 * request of Q arrives soon (see the comments on
3258 * bfq_bfqq_update_budg_for_activation). If all queues and
3259 * groups have the same weight, this form of preemption,
3260 * combined with the hole-recovery heuristic described in the
3261 * comments on function bfq_bfqq_update_budg_for_activation,
3262 * are enough to preserve a correct bandwidth distribution in
3263 * the mid term, even without idling. In fact, even if not
3264 * idling allows the internal queues of the device to contain
3265 * many requests, and thus to reorder requests, we can rather
3266 * safely assume that the internal scheduler still preserves a
3267 * minimum of mid-term fairness. The motivation for using
3268 * preemption instead of idling is that, by not idling,
3269 * service guarantees are preserved without minimally
3270 * sacrificing throughput. In other words, both a high
3271 * throughput and its desired distribution are obtained.
3273 * More precisely, this preemption-based, idleless approach
3274 * provides fairness in terms of IOPS, and not sectors per
3275 * second. This can be seen with a simple example. Suppose
3276 * that there are two queues with the same weight, but that
3277 * the first queue receives requests of 8 sectors, while the
3278 * second queue receives requests of 1024 sectors. In
3279 * addition, suppose that each of the two queues contains at
3280 * most one request at a time, which implies that each queue
3281 * always remains idle after it is served. Finally, after
3282 * remaining idle, each queue receives very quickly a new
3283 * request. It follows that the two queues are served
3284 * alternatively, preempting each other if needed. This
3285 * implies that, although both queues have the same weight,
3286 * the queue with large requests receives a service that is
3287 * 1024/8 times as high as the service received by the other
3290 * On the other hand, device idling is performed, and thus
3291 * pure sector-domain guarantees are provided, for the
3292 * following queues, which are likely to need stronger
3293 * throughput guarantees: weight-raised queues, and queues
3294 * with a higher weight than other queues. When such queues
3295 * are active, sub-condition (i) is false, which triggers
3298 * According to the above considerations, the next variable is
3299 * true (only) if sub-condition (i) holds. To compute the
3300 * value of this variable, we not only use the return value of
3301 * the function bfq_symmetric_scenario(), but also check
3302 * whether bfqq is being weight-raised, because
3303 * bfq_symmetric_scenario() does not take into account also
3304 * weight-raised queues (see comments on
3305 * bfq_weights_tree_add()).
3307 * As a side note, it is worth considering that the above
3308 * device-idling countermeasures may however fail in the
3309 * following unlucky scenario: if idling is (correctly)
3310 * disabled in a time period during which all symmetry
3311 * sub-conditions hold, and hence the device is allowed to
3312 * enqueue many requests, but at some later point in time some
3313 * sub-condition stops to hold, then it may become impossible
3314 * to let requests be served in the desired order until all
3315 * the requests already queued in the device have been served.
3317 asymmetric_scenario
= bfqq
->wr_coeff
> 1 ||
3318 !bfq_symmetric_scenario(bfqd
);
3321 * Finally, there is a case where maximizing throughput is the
3322 * best choice even if it may cause unfairness toward
3323 * bfqq. Such a case is when bfqq became active in a burst of
3324 * queue activations. Queues that became active during a large
3325 * burst benefit only from throughput, as discussed in the
3326 * comments on bfq_handle_burst. Thus, if bfqq became active
3327 * in a burst and not idling the device maximizes throughput,
3328 * then the device must no be idled, because not idling the
3329 * device provides bfqq and all other queues in the burst with
3330 * maximum benefit. Combining this and the above case, we can
3331 * now establish when idling is actually needed to preserve
3332 * service guarantees.
3334 idling_needed_for_service_guarantees
=
3335 asymmetric_scenario
&& !bfq_bfqq_in_large_burst(bfqq
);
3338 * We have now all the components we need to compute the
3339 * return value of the function, which is true only if idling
3340 * either boosts the throughput (without issues), or is
3341 * necessary to preserve service guarantees.
3343 return idling_boosts_thr_without_issues
||
3344 idling_needed_for_service_guarantees
;
3348 * If the in-service queue is empty but the function bfq_bfqq_may_idle
3349 * returns true, then:
3350 * 1) the queue must remain in service and cannot be expired, and
3351 * 2) the device must be idled to wait for the possible arrival of a new
3352 * request for the queue.
3353 * See the comments on the function bfq_bfqq_may_idle for the reasons
3354 * why performing device idling is the best choice to boost the throughput
3355 * and preserve service guarantees when bfq_bfqq_may_idle itself
3358 static bool bfq_bfqq_must_idle(struct bfq_queue
*bfqq
)
3360 return RB_EMPTY_ROOT(&bfqq
->sort_list
) && bfq_bfqq_may_idle(bfqq
);
3364 * Select a queue for service. If we have a current queue in service,
3365 * check whether to continue servicing it, or retrieve and set a new one.
3367 static struct bfq_queue
*bfq_select_queue(struct bfq_data
*bfqd
)
3369 struct bfq_queue
*bfqq
;
3370 struct request
*next_rq
;
3371 enum bfqq_expiration reason
= BFQQE_BUDGET_TIMEOUT
;
3373 bfqq
= bfqd
->in_service_queue
;
3377 bfq_log_bfqq(bfqd
, bfqq
, "select_queue: already in-service queue");
3379 if (bfq_may_expire_for_budg_timeout(bfqq
) &&
3380 !bfq_bfqq_wait_request(bfqq
) &&
3381 !bfq_bfqq_must_idle(bfqq
))
3386 * This loop is rarely executed more than once. Even when it
3387 * happens, it is much more convenient to re-execute this loop
3388 * than to return NULL and trigger a new dispatch to get a
3391 next_rq
= bfqq
->next_rq
;
3393 * If bfqq has requests queued and it has enough budget left to
3394 * serve them, keep the queue, otherwise expire it.
3397 if (bfq_serv_to_charge(next_rq
, bfqq
) >
3398 bfq_bfqq_budget_left(bfqq
)) {
3400 * Expire the queue for budget exhaustion,
3401 * which makes sure that the next budget is
3402 * enough to serve the next request, even if
3403 * it comes from the fifo expired path.
3405 reason
= BFQQE_BUDGET_EXHAUSTED
;
3409 * The idle timer may be pending because we may
3410 * not disable disk idling even when a new request
3413 if (bfq_bfqq_wait_request(bfqq
)) {
3415 * If we get here: 1) at least a new request
3416 * has arrived but we have not disabled the
3417 * timer because the request was too small,
3418 * 2) then the block layer has unplugged
3419 * the device, causing the dispatch to be
3422 * Since the device is unplugged, now the
3423 * requests are probably large enough to
3424 * provide a reasonable throughput.
3425 * So we disable idling.
3427 bfq_clear_bfqq_wait_request(bfqq
);
3428 hrtimer_try_to_cancel(&bfqd
->idle_slice_timer
);
3429 bfqg_stats_update_idle_time(bfqq_group(bfqq
));
3436 * No requests pending. However, if the in-service queue is idling
3437 * for a new request, or has requests waiting for a completion and
3438 * may idle after their completion, then keep it anyway.
3440 if (bfq_bfqq_wait_request(bfqq
) ||
3441 (bfqq
->dispatched
!= 0 && bfq_bfqq_may_idle(bfqq
))) {
3446 reason
= BFQQE_NO_MORE_REQUESTS
;
3448 bfq_bfqq_expire(bfqd
, bfqq
, false, reason
);
3450 bfqq
= bfq_set_in_service_queue(bfqd
);
3452 bfq_log_bfqq(bfqd
, bfqq
, "select_queue: checking new queue");
3457 bfq_log_bfqq(bfqd
, bfqq
, "select_queue: returned this queue");
3459 bfq_log(bfqd
, "select_queue: no queue returned");
3464 static void bfq_update_wr_data(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
3466 struct bfq_entity
*entity
= &bfqq
->entity
;
3468 if (bfqq
->wr_coeff
> 1) { /* queue is being weight-raised */
3469 bfq_log_bfqq(bfqd
, bfqq
,
3470 "raising period dur %u/%u msec, old coeff %u, w %d(%d)",
3471 jiffies_to_msecs(jiffies
- bfqq
->last_wr_start_finish
),
3472 jiffies_to_msecs(bfqq
->wr_cur_max_time
),
3474 bfqq
->entity
.weight
, bfqq
->entity
.orig_weight
);
3476 if (entity
->prio_changed
)
3477 bfq_log_bfqq(bfqd
, bfqq
, "WARN: pending prio change");
3480 * If the queue was activated in a burst, or too much
3481 * time has elapsed from the beginning of this
3482 * weight-raising period, then end weight raising.
3484 if (bfq_bfqq_in_large_burst(bfqq
))
3485 bfq_bfqq_end_wr(bfqq
);
3486 else if (time_is_before_jiffies(bfqq
->last_wr_start_finish
+
3487 bfqq
->wr_cur_max_time
)) {
3488 if (bfqq
->wr_cur_max_time
!= bfqd
->bfq_wr_rt_max_time
||
3489 time_is_before_jiffies(bfqq
->wr_start_at_switch_to_srt
+
3490 bfq_wr_duration(bfqd
)))
3491 bfq_bfqq_end_wr(bfqq
);
3493 /* switch back to interactive wr */
3494 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
;
3495 bfqq
->wr_cur_max_time
= bfq_wr_duration(bfqd
);
3496 bfqq
->last_wr_start_finish
=
3497 bfqq
->wr_start_at_switch_to_srt
;
3498 bfqq
->entity
.prio_changed
= 1;
3503 * To improve latency (for this or other queues), immediately
3504 * update weight both if it must be raised and if it must be
3505 * lowered. Since, entity may be on some active tree here, and
3506 * might have a pending change of its ioprio class, invoke
3507 * next function with the last parameter unset (see the
3508 * comments on the function).
3510 if ((entity
->weight
> entity
->orig_weight
) != (bfqq
->wr_coeff
> 1))
3511 __bfq_entity_update_weight_prio(bfq_entity_service_tree(entity
),
3516 * Dispatch next request from bfqq.
3518 static struct request
*bfq_dispatch_rq_from_bfqq(struct bfq_data
*bfqd
,
3519 struct bfq_queue
*bfqq
)
3521 struct request
*rq
= bfqq
->next_rq
;
3522 unsigned long service_to_charge
;
3524 service_to_charge
= bfq_serv_to_charge(rq
, bfqq
);
3526 bfq_bfqq_served(bfqq
, service_to_charge
);
3528 bfq_dispatch_remove(bfqd
->queue
, rq
);
3531 * If weight raising has to terminate for bfqq, then next
3532 * function causes an immediate update of bfqq's weight,
3533 * without waiting for next activation. As a consequence, on
3534 * expiration, bfqq will be timestamped as if has never been
3535 * weight-raised during this service slot, even if it has
3536 * received part or even most of the service as a
3537 * weight-raised queue. This inflates bfqq's timestamps, which
3538 * is beneficial, as bfqq is then more willing to leave the
3539 * device immediately to possible other weight-raised queues.
3541 bfq_update_wr_data(bfqd
, bfqq
);
3544 * Expire bfqq, pretending that its budget expired, if bfqq
3545 * belongs to CLASS_IDLE and other queues are waiting for
3548 if (bfqd
->busy_queues
> 1 && bfq_class_idle(bfqq
))
3554 bfq_bfqq_expire(bfqd
, bfqq
, false, BFQQE_BUDGET_EXHAUSTED
);
3558 static bool bfq_has_work(struct blk_mq_hw_ctx
*hctx
)
3560 struct bfq_data
*bfqd
= hctx
->queue
->elevator
->elevator_data
;
3563 * Avoiding lock: a race on bfqd->busy_queues should cause at
3564 * most a call to dispatch for nothing
3566 return !list_empty_careful(&bfqd
->dispatch
) ||
3567 bfqd
->busy_queues
> 0;
3570 static struct request
*__bfq_dispatch_request(struct blk_mq_hw_ctx
*hctx
)
3572 struct bfq_data
*bfqd
= hctx
->queue
->elevator
->elevator_data
;
3573 struct request
*rq
= NULL
;
3574 struct bfq_queue
*bfqq
= NULL
;
3576 if (!list_empty(&bfqd
->dispatch
)) {
3577 rq
= list_first_entry(&bfqd
->dispatch
, struct request
,
3579 list_del_init(&rq
->queuelist
);
3585 * Increment counters here, because this
3586 * dispatch does not follow the standard
3587 * dispatch flow (where counters are
3592 goto inc_in_driver_start_rq
;
3596 * We exploit the put_rq_private hook to decrement
3597 * rq_in_driver, but put_rq_private will not be
3598 * invoked on this request. So, to avoid unbalance,
3599 * just start this request, without incrementing
3600 * rq_in_driver. As a negative consequence,
3601 * rq_in_driver is deceptively lower than it should be
3602 * while this request is in service. This may cause
3603 * bfq_schedule_dispatch to be invoked uselessly.
3605 * As for implementing an exact solution, the
3606 * put_request hook, if defined, is probably invoked
3607 * also on this request. So, by exploiting this hook,
3608 * we could 1) increment rq_in_driver here, and 2)
3609 * decrement it in put_request. Such a solution would
3610 * let the value of the counter be always accurate,
3611 * but it would entail using an extra interface
3612 * function. This cost seems higher than the benefit,
3613 * being the frequency of non-elevator-private
3614 * requests very low.
3619 bfq_log(bfqd
, "dispatch requests: %d busy queues", bfqd
->busy_queues
);
3621 if (bfqd
->busy_queues
== 0)
3625 * Force device to serve one request at a time if
3626 * strict_guarantees is true. Forcing this service scheme is
3627 * currently the ONLY way to guarantee that the request
3628 * service order enforced by the scheduler is respected by a
3629 * queueing device. Otherwise the device is free even to make
3630 * some unlucky request wait for as long as the device
3633 * Of course, serving one request at at time may cause loss of
3636 if (bfqd
->strict_guarantees
&& bfqd
->rq_in_driver
> 0)
3639 bfqq
= bfq_select_queue(bfqd
);
3643 rq
= bfq_dispatch_rq_from_bfqq(bfqd
, bfqq
);
3646 inc_in_driver_start_rq
:
3647 bfqd
->rq_in_driver
++;
3649 rq
->rq_flags
|= RQF_STARTED
;
3655 static struct request
*bfq_dispatch_request(struct blk_mq_hw_ctx
*hctx
)
3657 struct bfq_data
*bfqd
= hctx
->queue
->elevator
->elevator_data
;
3660 spin_lock_irq(&bfqd
->lock
);
3662 rq
= __bfq_dispatch_request(hctx
);
3663 spin_unlock_irq(&bfqd
->lock
);
3669 * Task holds one reference to the queue, dropped when task exits. Each rq
3670 * in-flight on this queue also holds a reference, dropped when rq is freed.
3672 * Scheduler lock must be held here. Recall not to use bfqq after calling
3673 * this function on it.
3675 void bfq_put_queue(struct bfq_queue
*bfqq
)
3677 #ifdef CONFIG_BFQ_GROUP_IOSCHED
3678 struct bfq_group
*bfqg
= bfqq_group(bfqq
);
3682 bfq_log_bfqq(bfqq
->bfqd
, bfqq
, "put_queue: %p %d",
3689 if (bfq_bfqq_sync(bfqq
))
3691 * The fact that this queue is being destroyed does not
3692 * invalidate the fact that this queue may have been
3693 * activated during the current burst. As a consequence,
3694 * although the queue does not exist anymore, and hence
3695 * needs to be removed from the burst list if there,
3696 * the burst size has not to be decremented.
3698 hlist_del_init(&bfqq
->burst_list_node
);
3700 kmem_cache_free(bfq_pool
, bfqq
);
3701 #ifdef CONFIG_BFQ_GROUP_IOSCHED
3702 bfqg_and_blkg_put(bfqg
);
3706 static void bfq_put_cooperator(struct bfq_queue
*bfqq
)
3708 struct bfq_queue
*__bfqq
, *next
;
3711 * If this queue was scheduled to merge with another queue, be
3712 * sure to drop the reference taken on that queue (and others in
3713 * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs.
3715 __bfqq
= bfqq
->new_bfqq
;
3719 next
= __bfqq
->new_bfqq
;
3720 bfq_put_queue(__bfqq
);
3725 static void bfq_exit_bfqq(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
3727 if (bfqq
== bfqd
->in_service_queue
) {
3728 __bfq_bfqq_expire(bfqd
, bfqq
);
3729 bfq_schedule_dispatch(bfqd
);
3732 bfq_log_bfqq(bfqd
, bfqq
, "exit_bfqq: %p, %d", bfqq
, bfqq
->ref
);
3734 bfq_put_cooperator(bfqq
);
3736 bfq_put_queue(bfqq
); /* release process reference */
3739 static void bfq_exit_icq_bfqq(struct bfq_io_cq
*bic
, bool is_sync
)
3741 struct bfq_queue
*bfqq
= bic_to_bfqq(bic
, is_sync
);
3742 struct bfq_data
*bfqd
;
3745 bfqd
= bfqq
->bfqd
; /* NULL if scheduler already exited */
3748 unsigned long flags
;
3750 spin_lock_irqsave(&bfqd
->lock
, flags
);
3751 bfq_exit_bfqq(bfqd
, bfqq
);
3752 bic_set_bfqq(bic
, NULL
, is_sync
);
3753 spin_unlock_irqrestore(&bfqd
->lock
, flags
);
3757 static void bfq_exit_icq(struct io_cq
*icq
)
3759 struct bfq_io_cq
*bic
= icq_to_bic(icq
);
3761 bfq_exit_icq_bfqq(bic
, true);
3762 bfq_exit_icq_bfqq(bic
, false);
3766 * Update the entity prio values; note that the new values will not
3767 * be used until the next (re)activation.
3770 bfq_set_next_ioprio_data(struct bfq_queue
*bfqq
, struct bfq_io_cq
*bic
)
3772 struct task_struct
*tsk
= current
;
3774 struct bfq_data
*bfqd
= bfqq
->bfqd
;
3779 ioprio_class
= IOPRIO_PRIO_CLASS(bic
->ioprio
);
3780 switch (ioprio_class
) {
3782 dev_err(bfqq
->bfqd
->queue
->backing_dev_info
->dev
,
3783 "bfq: bad prio class %d\n", ioprio_class
);
3785 case IOPRIO_CLASS_NONE
:
3787 * No prio set, inherit CPU scheduling settings.
3789 bfqq
->new_ioprio
= task_nice_ioprio(tsk
);
3790 bfqq
->new_ioprio_class
= task_nice_ioclass(tsk
);
3792 case IOPRIO_CLASS_RT
:
3793 bfqq
->new_ioprio
= IOPRIO_PRIO_DATA(bic
->ioprio
);
3794 bfqq
->new_ioprio_class
= IOPRIO_CLASS_RT
;
3796 case IOPRIO_CLASS_BE
:
3797 bfqq
->new_ioprio
= IOPRIO_PRIO_DATA(bic
->ioprio
);
3798 bfqq
->new_ioprio_class
= IOPRIO_CLASS_BE
;
3800 case IOPRIO_CLASS_IDLE
:
3801 bfqq
->new_ioprio_class
= IOPRIO_CLASS_IDLE
;
3802 bfqq
->new_ioprio
= 7;
3806 if (bfqq
->new_ioprio
>= IOPRIO_BE_NR
) {
3807 pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n",
3809 bfqq
->new_ioprio
= IOPRIO_BE_NR
;
3812 bfqq
->entity
.new_weight
= bfq_ioprio_to_weight(bfqq
->new_ioprio
);
3813 bfqq
->entity
.prio_changed
= 1;
3816 static struct bfq_queue
*bfq_get_queue(struct bfq_data
*bfqd
,
3817 struct bio
*bio
, bool is_sync
,
3818 struct bfq_io_cq
*bic
);
3820 static void bfq_check_ioprio_change(struct bfq_io_cq
*bic
, struct bio
*bio
)
3822 struct bfq_data
*bfqd
= bic_to_bfqd(bic
);
3823 struct bfq_queue
*bfqq
;
3824 int ioprio
= bic
->icq
.ioc
->ioprio
;
3827 * This condition may trigger on a newly created bic, be sure to
3828 * drop the lock before returning.
3830 if (unlikely(!bfqd
) || likely(bic
->ioprio
== ioprio
))
3833 bic
->ioprio
= ioprio
;
3835 bfqq
= bic_to_bfqq(bic
, false);
3837 /* release process reference on this queue */
3838 bfq_put_queue(bfqq
);
3839 bfqq
= bfq_get_queue(bfqd
, bio
, BLK_RW_ASYNC
, bic
);
3840 bic_set_bfqq(bic
, bfqq
, false);
3843 bfqq
= bic_to_bfqq(bic
, true);
3845 bfq_set_next_ioprio_data(bfqq
, bic
);
3848 static void bfq_init_bfqq(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
3849 struct bfq_io_cq
*bic
, pid_t pid
, int is_sync
)
3851 RB_CLEAR_NODE(&bfqq
->entity
.rb_node
);
3852 INIT_LIST_HEAD(&bfqq
->fifo
);
3853 INIT_HLIST_NODE(&bfqq
->burst_list_node
);
3859 bfq_set_next_ioprio_data(bfqq
, bic
);
3863 * No need to mark as has_short_ttime if in
3864 * idle_class, because no device idling is performed
3865 * for queues in idle class
3867 if (!bfq_class_idle(bfqq
))
3868 /* tentatively mark as has_short_ttime */
3869 bfq_mark_bfqq_has_short_ttime(bfqq
);
3870 bfq_mark_bfqq_sync(bfqq
);
3871 bfq_mark_bfqq_just_created(bfqq
);
3873 bfq_clear_bfqq_sync(bfqq
);
3875 /* set end request to minus infinity from now */
3876 bfqq
->ttime
.last_end_request
= ktime_get_ns() + 1;
3878 bfq_mark_bfqq_IO_bound(bfqq
);
3882 /* Tentative initial value to trade off between thr and lat */
3883 bfqq
->max_budget
= (2 * bfq_max_budget(bfqd
)) / 3;
3884 bfqq
->budget_timeout
= bfq_smallest_from_now();
3887 bfqq
->last_wr_start_finish
= jiffies
;
3888 bfqq
->wr_start_at_switch_to_srt
= bfq_smallest_from_now();
3889 bfqq
->split_time
= bfq_smallest_from_now();
3892 * Set to the value for which bfqq will not be deemed as
3893 * soft rt when it becomes backlogged.
3895 bfqq
->soft_rt_next_start
= bfq_greatest_from_now();
3897 /* first request is almost certainly seeky */
3898 bfqq
->seek_history
= 1;
3901 static struct bfq_queue
**bfq_async_queue_prio(struct bfq_data
*bfqd
,
3902 struct bfq_group
*bfqg
,
3903 int ioprio_class
, int ioprio
)
3905 switch (ioprio_class
) {
3906 case IOPRIO_CLASS_RT
:
3907 return &bfqg
->async_bfqq
[0][ioprio
];
3908 case IOPRIO_CLASS_NONE
:
3909 ioprio
= IOPRIO_NORM
;
3911 case IOPRIO_CLASS_BE
:
3912 return &bfqg
->async_bfqq
[1][ioprio
];
3913 case IOPRIO_CLASS_IDLE
:
3914 return &bfqg
->async_idle_bfqq
;
3920 static struct bfq_queue
*bfq_get_queue(struct bfq_data
*bfqd
,
3921 struct bio
*bio
, bool is_sync
,
3922 struct bfq_io_cq
*bic
)
3924 const int ioprio
= IOPRIO_PRIO_DATA(bic
->ioprio
);
3925 const int ioprio_class
= IOPRIO_PRIO_CLASS(bic
->ioprio
);
3926 struct bfq_queue
**async_bfqq
= NULL
;
3927 struct bfq_queue
*bfqq
;
3928 struct bfq_group
*bfqg
;
3932 bfqg
= bfq_find_set_group(bfqd
, bio_blkcg(bio
));
3934 bfqq
= &bfqd
->oom_bfqq
;
3939 async_bfqq
= bfq_async_queue_prio(bfqd
, bfqg
, ioprio_class
,
3946 bfqq
= kmem_cache_alloc_node(bfq_pool
,
3947 GFP_NOWAIT
| __GFP_ZERO
| __GFP_NOWARN
,
3951 bfq_init_bfqq(bfqd
, bfqq
, bic
, current
->pid
,
3953 bfq_init_entity(&bfqq
->entity
, bfqg
);
3954 bfq_log_bfqq(bfqd
, bfqq
, "allocated");
3956 bfqq
= &bfqd
->oom_bfqq
;
3957 bfq_log_bfqq(bfqd
, bfqq
, "using oom bfqq");
3962 * Pin the queue now that it's allocated, scheduler exit will
3967 * Extra group reference, w.r.t. sync
3968 * queue. This extra reference is removed
3969 * only if bfqq->bfqg disappears, to
3970 * guarantee that this queue is not freed
3971 * until its group goes away.
3973 bfq_log_bfqq(bfqd
, bfqq
, "get_queue, bfqq not in async: %p, %d",
3979 bfqq
->ref
++; /* get a process reference to this queue */
3980 bfq_log_bfqq(bfqd
, bfqq
, "get_queue, at end: %p, %d", bfqq
, bfqq
->ref
);
3985 static void bfq_update_io_thinktime(struct bfq_data
*bfqd
,
3986 struct bfq_queue
*bfqq
)
3988 struct bfq_ttime
*ttime
= &bfqq
->ttime
;
3989 u64 elapsed
= ktime_get_ns() - bfqq
->ttime
.last_end_request
;
3991 elapsed
= min_t(u64
, elapsed
, 2ULL * bfqd
->bfq_slice_idle
);
3993 ttime
->ttime_samples
= (7*bfqq
->ttime
.ttime_samples
+ 256) / 8;
3994 ttime
->ttime_total
= div_u64(7*ttime
->ttime_total
+ 256*elapsed
, 8);
3995 ttime
->ttime_mean
= div64_ul(ttime
->ttime_total
+ 128,
3996 ttime
->ttime_samples
);
4000 bfq_update_io_seektime(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
4003 bfqq
->seek_history
<<= 1;
4004 bfqq
->seek_history
|=
4005 get_sdist(bfqq
->last_request_pos
, rq
) > BFQQ_SEEK_THR
&&
4006 (!blk_queue_nonrot(bfqd
->queue
) ||
4007 blk_rq_sectors(rq
) < BFQQ_SECT_THR_NONROT
);
4010 static void bfq_update_has_short_ttime(struct bfq_data
*bfqd
,
4011 struct bfq_queue
*bfqq
,
4012 struct bfq_io_cq
*bic
)
4014 bool has_short_ttime
= true;
4017 * No need to update has_short_ttime if bfqq is async or in
4018 * idle io prio class, or if bfq_slice_idle is zero, because
4019 * no device idling is performed for bfqq in this case.
4021 if (!bfq_bfqq_sync(bfqq
) || bfq_class_idle(bfqq
) ||
4022 bfqd
->bfq_slice_idle
== 0)
4025 /* Idle window just restored, statistics are meaningless. */
4026 if (time_is_after_eq_jiffies(bfqq
->split_time
+
4027 bfqd
->bfq_wr_min_idle_time
))
4030 /* Think time is infinite if no process is linked to
4031 * bfqq. Otherwise check average think time to
4032 * decide whether to mark as has_short_ttime
4034 if (atomic_read(&bic
->icq
.ioc
->active_ref
) == 0 ||
4035 (bfq_sample_valid(bfqq
->ttime
.ttime_samples
) &&
4036 bfqq
->ttime
.ttime_mean
> bfqd
->bfq_slice_idle
))
4037 has_short_ttime
= false;
4039 bfq_log_bfqq(bfqd
, bfqq
, "update_has_short_ttime: has_short_ttime %d",
4042 if (has_short_ttime
)
4043 bfq_mark_bfqq_has_short_ttime(bfqq
);
4045 bfq_clear_bfqq_has_short_ttime(bfqq
);
4049 * Called when a new fs request (rq) is added to bfqq. Check if there's
4050 * something we should do about it.
4052 static void bfq_rq_enqueued(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
4055 struct bfq_io_cq
*bic
= RQ_BIC(rq
);
4057 if (rq
->cmd_flags
& REQ_META
)
4058 bfqq
->meta_pending
++;
4060 bfq_update_io_thinktime(bfqd
, bfqq
);
4061 bfq_update_has_short_ttime(bfqd
, bfqq
, bic
);
4062 bfq_update_io_seektime(bfqd
, bfqq
, rq
);
4064 bfq_log_bfqq(bfqd
, bfqq
,
4065 "rq_enqueued: has_short_ttime=%d (seeky %d)",
4066 bfq_bfqq_has_short_ttime(bfqq
), BFQQ_SEEKY(bfqq
));
4068 bfqq
->last_request_pos
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
4070 if (bfqq
== bfqd
->in_service_queue
&& bfq_bfqq_wait_request(bfqq
)) {
4071 bool small_req
= bfqq
->queued
[rq_is_sync(rq
)] == 1 &&
4072 blk_rq_sectors(rq
) < 32;
4073 bool budget_timeout
= bfq_bfqq_budget_timeout(bfqq
);
4076 * There is just this request queued: if the request
4077 * is small and the queue is not to be expired, then
4080 * In this way, if the device is being idled to wait
4081 * for a new request from the in-service queue, we
4082 * avoid unplugging the device and committing the
4083 * device to serve just a small request. On the
4084 * contrary, we wait for the block layer to decide
4085 * when to unplug the device: hopefully, new requests
4086 * will be merged to this one quickly, then the device
4087 * will be unplugged and larger requests will be
4090 if (small_req
&& !budget_timeout
)
4094 * A large enough request arrived, or the queue is to
4095 * be expired: in both cases disk idling is to be
4096 * stopped, so clear wait_request flag and reset
4099 bfq_clear_bfqq_wait_request(bfqq
);
4100 hrtimer_try_to_cancel(&bfqd
->idle_slice_timer
);
4101 bfqg_stats_update_idle_time(bfqq_group(bfqq
));
4104 * The queue is not empty, because a new request just
4105 * arrived. Hence we can safely expire the queue, in
4106 * case of budget timeout, without risking that the
4107 * timestamps of the queue are not updated correctly.
4108 * See [1] for more details.
4111 bfq_bfqq_expire(bfqd
, bfqq
, false,
4112 BFQQE_BUDGET_TIMEOUT
);
4116 static void __bfq_insert_request(struct bfq_data
*bfqd
, struct request
*rq
)
4118 struct bfq_queue
*bfqq
= RQ_BFQQ(rq
),
4119 *new_bfqq
= bfq_setup_cooperator(bfqd
, bfqq
, rq
, true);
4122 if (bic_to_bfqq(RQ_BIC(rq
), 1) != bfqq
)
4123 new_bfqq
= bic_to_bfqq(RQ_BIC(rq
), 1);
4125 * Release the request's reference to the old bfqq
4126 * and make sure one is taken to the shared queue.
4128 new_bfqq
->allocated
++;
4131 bfq_clear_bfqq_just_created(bfqq
);
4133 * If the bic associated with the process
4134 * issuing this request still points to bfqq
4135 * (and thus has not been already redirected
4136 * to new_bfqq or even some other bfq_queue),
4137 * then complete the merge and redirect it to
4140 if (bic_to_bfqq(RQ_BIC(rq
), 1) == bfqq
)
4141 bfq_merge_bfqqs(bfqd
, RQ_BIC(rq
),
4144 * rq is about to be enqueued into new_bfqq,
4145 * release rq reference on bfqq
4147 bfq_put_queue(bfqq
);
4148 rq
->elv
.priv
[1] = new_bfqq
;
4152 bfq_add_request(rq
);
4154 rq
->fifo_time
= ktime_get_ns() + bfqd
->bfq_fifo_expire
[rq_is_sync(rq
)];
4155 list_add_tail(&rq
->queuelist
, &bfqq
->fifo
);
4157 bfq_rq_enqueued(bfqd
, bfqq
, rq
);
4160 static void bfq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
4163 struct request_queue
*q
= hctx
->queue
;
4164 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
4166 spin_lock_irq(&bfqd
->lock
);
4167 if (blk_mq_sched_try_insert_merge(q
, rq
)) {
4168 spin_unlock_irq(&bfqd
->lock
);
4172 spin_unlock_irq(&bfqd
->lock
);
4174 blk_mq_sched_request_inserted(rq
);
4176 spin_lock_irq(&bfqd
->lock
);
4177 if (at_head
|| blk_rq_is_passthrough(rq
)) {
4179 list_add(&rq
->queuelist
, &bfqd
->dispatch
);
4181 list_add_tail(&rq
->queuelist
, &bfqd
->dispatch
);
4183 __bfq_insert_request(bfqd
, rq
);
4185 if (rq_mergeable(rq
)) {
4186 elv_rqhash_add(q
, rq
);
4192 spin_unlock_irq(&bfqd
->lock
);
4195 static void bfq_insert_requests(struct blk_mq_hw_ctx
*hctx
,
4196 struct list_head
*list
, bool at_head
)
4198 while (!list_empty(list
)) {
4201 rq
= list_first_entry(list
, struct request
, queuelist
);
4202 list_del_init(&rq
->queuelist
);
4203 bfq_insert_request(hctx
, rq
, at_head
);
4207 static void bfq_update_hw_tag(struct bfq_data
*bfqd
)
4209 bfqd
->max_rq_in_driver
= max_t(int, bfqd
->max_rq_in_driver
,
4210 bfqd
->rq_in_driver
);
4212 if (bfqd
->hw_tag
== 1)
4216 * This sample is valid if the number of outstanding requests
4217 * is large enough to allow a queueing behavior. Note that the
4218 * sum is not exact, as it's not taking into account deactivated
4221 if (bfqd
->rq_in_driver
+ bfqd
->queued
< BFQ_HW_QUEUE_THRESHOLD
)
4224 if (bfqd
->hw_tag_samples
++ < BFQ_HW_QUEUE_SAMPLES
)
4227 bfqd
->hw_tag
= bfqd
->max_rq_in_driver
> BFQ_HW_QUEUE_THRESHOLD
;
4228 bfqd
->max_rq_in_driver
= 0;
4229 bfqd
->hw_tag_samples
= 0;
4232 static void bfq_completed_request(struct bfq_queue
*bfqq
, struct bfq_data
*bfqd
)
4237 bfq_update_hw_tag(bfqd
);
4239 bfqd
->rq_in_driver
--;
4242 if (!bfqq
->dispatched
&& !bfq_bfqq_busy(bfqq
)) {
4244 * Set budget_timeout (which we overload to store the
4245 * time at which the queue remains with no backlog and
4246 * no outstanding request; used by the weight-raising
4249 bfqq
->budget_timeout
= jiffies
;
4251 bfq_weights_tree_remove(bfqd
, &bfqq
->entity
,
4252 &bfqd
->queue_weights_tree
);
4255 now_ns
= ktime_get_ns();
4257 bfqq
->ttime
.last_end_request
= now_ns
;
4260 * Using us instead of ns, to get a reasonable precision in
4261 * computing rate in next check.
4263 delta_us
= div_u64(now_ns
- bfqd
->last_completion
, NSEC_PER_USEC
);
4266 * If the request took rather long to complete, and, according
4267 * to the maximum request size recorded, this completion latency
4268 * implies that the request was certainly served at a very low
4269 * rate (less than 1M sectors/sec), then the whole observation
4270 * interval that lasts up to this time instant cannot be a
4271 * valid time interval for computing a new peak rate. Invoke
4272 * bfq_update_rate_reset to have the following three steps
4274 * - close the observation interval at the last (previous)
4275 * request dispatch or completion
4276 * - compute rate, if possible, for that observation interval
4277 * - reset to zero samples, which will trigger a proper
4278 * re-initialization of the observation interval on next
4281 if (delta_us
> BFQ_MIN_TT
/NSEC_PER_USEC
&&
4282 (bfqd
->last_rq_max_size
<<BFQ_RATE_SHIFT
)/delta_us
<
4283 1UL<<(BFQ_RATE_SHIFT
- 10))
4284 bfq_update_rate_reset(bfqd
, NULL
);
4285 bfqd
->last_completion
= now_ns
;
4288 * If we are waiting to discover whether the request pattern
4289 * of the task associated with the queue is actually
4290 * isochronous, and both requisites for this condition to hold
4291 * are now satisfied, then compute soft_rt_next_start (see the
4292 * comments on the function bfq_bfqq_softrt_next_start()). We
4293 * schedule this delayed check when bfqq expires, if it still
4294 * has in-flight requests.
4296 if (bfq_bfqq_softrt_update(bfqq
) && bfqq
->dispatched
== 0 &&
4297 RB_EMPTY_ROOT(&bfqq
->sort_list
))
4298 bfqq
->soft_rt_next_start
=
4299 bfq_bfqq_softrt_next_start(bfqd
, bfqq
);
4302 * If this is the in-service queue, check if it needs to be expired,
4303 * or if we want to idle in case it has no pending requests.
4305 if (bfqd
->in_service_queue
== bfqq
) {
4306 if (bfqq
->dispatched
== 0 && bfq_bfqq_must_idle(bfqq
)) {
4307 bfq_arm_slice_timer(bfqd
);
4309 } else if (bfq_may_expire_for_budg_timeout(bfqq
))
4310 bfq_bfqq_expire(bfqd
, bfqq
, false,
4311 BFQQE_BUDGET_TIMEOUT
);
4312 else if (RB_EMPTY_ROOT(&bfqq
->sort_list
) &&
4313 (bfqq
->dispatched
== 0 ||
4314 !bfq_bfqq_may_idle(bfqq
)))
4315 bfq_bfqq_expire(bfqd
, bfqq
, false,
4316 BFQQE_NO_MORE_REQUESTS
);
4319 if (!bfqd
->rq_in_driver
)
4320 bfq_schedule_dispatch(bfqd
);
4323 static void bfq_put_rq_priv_body(struct bfq_queue
*bfqq
)
4327 bfq_put_queue(bfqq
);
4330 static void bfq_finish_request(struct request
*rq
)
4332 struct bfq_queue
*bfqq
;
4333 struct bfq_data
*bfqd
;
4341 if (rq
->rq_flags
& RQF_STARTED
)
4342 bfqg_stats_update_completion(bfqq_group(bfqq
),
4343 rq_start_time_ns(rq
),
4344 rq_io_start_time_ns(rq
),
4347 if (likely(rq
->rq_flags
& RQF_STARTED
)) {
4348 unsigned long flags
;
4350 spin_lock_irqsave(&bfqd
->lock
, flags
);
4352 bfq_completed_request(bfqq
, bfqd
);
4353 bfq_put_rq_priv_body(bfqq
);
4355 spin_unlock_irqrestore(&bfqd
->lock
, flags
);
4358 * Request rq may be still/already in the scheduler,
4359 * in which case we need to remove it. And we cannot
4360 * defer such a check and removal, to avoid
4361 * inconsistencies in the time interval from the end
4362 * of this function to the start of the deferred work.
4363 * This situation seems to occur only in process
4364 * context, as a consequence of a merge. In the
4365 * current version of the code, this implies that the
4369 if (!RB_EMPTY_NODE(&rq
->rb_node
))
4370 bfq_remove_request(rq
->q
, rq
);
4371 bfq_put_rq_priv_body(bfqq
);
4374 rq
->elv
.priv
[0] = NULL
;
4375 rq
->elv
.priv
[1] = NULL
;
4379 * Returns NULL if a new bfqq should be allocated, or the old bfqq if this
4380 * was the last process referring to that bfqq.
4382 static struct bfq_queue
*
4383 bfq_split_bfqq(struct bfq_io_cq
*bic
, struct bfq_queue
*bfqq
)
4385 bfq_log_bfqq(bfqq
->bfqd
, bfqq
, "splitting queue");
4387 if (bfqq_process_refs(bfqq
) == 1) {
4388 bfqq
->pid
= current
->pid
;
4389 bfq_clear_bfqq_coop(bfqq
);
4390 bfq_clear_bfqq_split_coop(bfqq
);
4394 bic_set_bfqq(bic
, NULL
, 1);
4396 bfq_put_cooperator(bfqq
);
4398 bfq_put_queue(bfqq
);
4402 static struct bfq_queue
*bfq_get_bfqq_handle_split(struct bfq_data
*bfqd
,
4403 struct bfq_io_cq
*bic
,
4405 bool split
, bool is_sync
,
4408 struct bfq_queue
*bfqq
= bic_to_bfqq(bic
, is_sync
);
4410 if (likely(bfqq
&& bfqq
!= &bfqd
->oom_bfqq
))
4417 bfq_put_queue(bfqq
);
4418 bfqq
= bfq_get_queue(bfqd
, bio
, is_sync
, bic
);
4420 bic_set_bfqq(bic
, bfqq
, is_sync
);
4421 if (split
&& is_sync
) {
4422 if ((bic
->was_in_burst_list
&& bfqd
->large_burst
) ||
4423 bic
->saved_in_large_burst
)
4424 bfq_mark_bfqq_in_large_burst(bfqq
);
4426 bfq_clear_bfqq_in_large_burst(bfqq
);
4427 if (bic
->was_in_burst_list
)
4428 hlist_add_head(&bfqq
->burst_list_node
,
4431 bfqq
->split_time
= jiffies
;
4438 * Allocate bfq data structures associated with this request.
4440 static void bfq_prepare_request(struct request
*rq
, struct bio
*bio
)
4442 struct request_queue
*q
= rq
->q
;
4443 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
4444 struct bfq_io_cq
*bic
;
4445 const int is_sync
= rq_is_sync(rq
);
4446 struct bfq_queue
*bfqq
;
4447 bool new_queue
= false;
4448 bool bfqq_already_existing
= false, split
= false;
4451 * Even if we don't have an icq attached, we should still clear
4452 * the scheduler pointers, as they might point to previously
4453 * allocated bic/bfqq structs.
4456 rq
->elv
.priv
[0] = rq
->elv
.priv
[1] = NULL
;
4460 bic
= icq_to_bic(rq
->elv
.icq
);
4462 spin_lock_irq(&bfqd
->lock
);
4464 bfq_check_ioprio_change(bic
, bio
);
4466 bfq_bic_update_cgroup(bic
, bio
);
4468 bfqq
= bfq_get_bfqq_handle_split(bfqd
, bic
, bio
, false, is_sync
,
4471 if (likely(!new_queue
)) {
4472 /* If the queue was seeky for too long, break it apart. */
4473 if (bfq_bfqq_coop(bfqq
) && bfq_bfqq_split_coop(bfqq
)) {
4474 bfq_log_bfqq(bfqd
, bfqq
, "breaking apart bfqq");
4476 /* Update bic before losing reference to bfqq */
4477 if (bfq_bfqq_in_large_burst(bfqq
))
4478 bic
->saved_in_large_burst
= true;
4480 bfqq
= bfq_split_bfqq(bic
, bfqq
);
4484 bfqq
= bfq_get_bfqq_handle_split(bfqd
, bic
, bio
,
4488 bfqq_already_existing
= true;
4494 bfq_log_bfqq(bfqd
, bfqq
, "get_request %p: bfqq %p, %d",
4495 rq
, bfqq
, bfqq
->ref
);
4497 rq
->elv
.priv
[0] = bic
;
4498 rq
->elv
.priv
[1] = bfqq
;
4501 * If a bfq_queue has only one process reference, it is owned
4502 * by only this bic: we can then set bfqq->bic = bic. in
4503 * addition, if the queue has also just been split, we have to
4506 if (likely(bfqq
!= &bfqd
->oom_bfqq
) && bfqq_process_refs(bfqq
) == 1) {
4510 * The queue has just been split from a shared
4511 * queue: restore the idle window and the
4512 * possible weight raising period.
4514 bfq_bfqq_resume_state(bfqq
, bfqd
, bic
,
4515 bfqq_already_existing
);
4519 if (unlikely(bfq_bfqq_just_created(bfqq
)))
4520 bfq_handle_burst(bfqd
, bfqq
);
4522 spin_unlock_irq(&bfqd
->lock
);
4525 static void bfq_idle_slice_timer_body(struct bfq_queue
*bfqq
)
4527 struct bfq_data
*bfqd
= bfqq
->bfqd
;
4528 enum bfqq_expiration reason
;
4529 unsigned long flags
;
4531 spin_lock_irqsave(&bfqd
->lock
, flags
);
4532 bfq_clear_bfqq_wait_request(bfqq
);
4534 if (bfqq
!= bfqd
->in_service_queue
) {
4535 spin_unlock_irqrestore(&bfqd
->lock
, flags
);
4539 if (bfq_bfqq_budget_timeout(bfqq
))
4541 * Also here the queue can be safely expired
4542 * for budget timeout without wasting
4545 reason
= BFQQE_BUDGET_TIMEOUT
;
4546 else if (bfqq
->queued
[0] == 0 && bfqq
->queued
[1] == 0)
4548 * The queue may not be empty upon timer expiration,
4549 * because we may not disable the timer when the
4550 * first request of the in-service queue arrives
4551 * during disk idling.
4553 reason
= BFQQE_TOO_IDLE
;
4555 goto schedule_dispatch
;
4557 bfq_bfqq_expire(bfqd
, bfqq
, true, reason
);
4560 spin_unlock_irqrestore(&bfqd
->lock
, flags
);
4561 bfq_schedule_dispatch(bfqd
);
4565 * Handler of the expiration of the timer running if the in-service queue
4566 * is idling inside its time slice.
4568 static enum hrtimer_restart
bfq_idle_slice_timer(struct hrtimer
*timer
)
4570 struct bfq_data
*bfqd
= container_of(timer
, struct bfq_data
,
4572 struct bfq_queue
*bfqq
= bfqd
->in_service_queue
;
4575 * Theoretical race here: the in-service queue can be NULL or
4576 * different from the queue that was idling if a new request
4577 * arrives for the current queue and there is a full dispatch
4578 * cycle that changes the in-service queue. This can hardly
4579 * happen, but in the worst case we just expire a queue too
4583 bfq_idle_slice_timer_body(bfqq
);
4585 return HRTIMER_NORESTART
;
4588 static void __bfq_put_async_bfqq(struct bfq_data
*bfqd
,
4589 struct bfq_queue
**bfqq_ptr
)
4591 struct bfq_queue
*bfqq
= *bfqq_ptr
;
4593 bfq_log(bfqd
, "put_async_bfqq: %p", bfqq
);
4595 bfq_bfqq_move(bfqd
, bfqq
, bfqd
->root_group
);
4597 bfq_log_bfqq(bfqd
, bfqq
, "put_async_bfqq: putting %p, %d",
4599 bfq_put_queue(bfqq
);
4605 * Release all the bfqg references to its async queues. If we are
4606 * deallocating the group these queues may still contain requests, so
4607 * we reparent them to the root cgroup (i.e., the only one that will
4608 * exist for sure until all the requests on a device are gone).
4610 void bfq_put_async_queues(struct bfq_data
*bfqd
, struct bfq_group
*bfqg
)
4614 for (i
= 0; i
< 2; i
++)
4615 for (j
= 0; j
< IOPRIO_BE_NR
; j
++)
4616 __bfq_put_async_bfqq(bfqd
, &bfqg
->async_bfqq
[i
][j
]);
4618 __bfq_put_async_bfqq(bfqd
, &bfqg
->async_idle_bfqq
);
4621 static void bfq_exit_queue(struct elevator_queue
*e
)
4623 struct bfq_data
*bfqd
= e
->elevator_data
;
4624 struct bfq_queue
*bfqq
, *n
;
4626 hrtimer_cancel(&bfqd
->idle_slice_timer
);
4628 spin_lock_irq(&bfqd
->lock
);
4629 list_for_each_entry_safe(bfqq
, n
, &bfqd
->idle_list
, bfqq_list
)
4630 bfq_deactivate_bfqq(bfqd
, bfqq
, false, false);
4631 spin_unlock_irq(&bfqd
->lock
);
4633 hrtimer_cancel(&bfqd
->idle_slice_timer
);
4635 #ifdef CONFIG_BFQ_GROUP_IOSCHED
4636 blkcg_deactivate_policy(bfqd
->queue
, &blkcg_policy_bfq
);
4638 spin_lock_irq(&bfqd
->lock
);
4639 bfq_put_async_queues(bfqd
, bfqd
->root_group
);
4640 kfree(bfqd
->root_group
);
4641 spin_unlock_irq(&bfqd
->lock
);
4647 static void bfq_init_root_group(struct bfq_group
*root_group
,
4648 struct bfq_data
*bfqd
)
4652 #ifdef CONFIG_BFQ_GROUP_IOSCHED
4653 root_group
->entity
.parent
= NULL
;
4654 root_group
->my_entity
= NULL
;
4655 root_group
->bfqd
= bfqd
;
4657 root_group
->rq_pos_tree
= RB_ROOT
;
4658 for (i
= 0; i
< BFQ_IOPRIO_CLASSES
; i
++)
4659 root_group
->sched_data
.service_tree
[i
] = BFQ_SERVICE_TREE_INIT
;
4660 root_group
->sched_data
.bfq_class_idle_last_service
= jiffies
;
4663 static int bfq_init_queue(struct request_queue
*q
, struct elevator_type
*e
)
4665 struct bfq_data
*bfqd
;
4666 struct elevator_queue
*eq
;
4668 eq
= elevator_alloc(q
, e
);
4672 bfqd
= kzalloc_node(sizeof(*bfqd
), GFP_KERNEL
, q
->node
);
4674 kobject_put(&eq
->kobj
);
4677 eq
->elevator_data
= bfqd
;
4679 spin_lock_irq(q
->queue_lock
);
4681 spin_unlock_irq(q
->queue_lock
);
4684 * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues.
4685 * Grab a permanent reference to it, so that the normal code flow
4686 * will not attempt to free it.
4688 bfq_init_bfqq(bfqd
, &bfqd
->oom_bfqq
, NULL
, 1, 0);
4689 bfqd
->oom_bfqq
.ref
++;
4690 bfqd
->oom_bfqq
.new_ioprio
= BFQ_DEFAULT_QUEUE_IOPRIO
;
4691 bfqd
->oom_bfqq
.new_ioprio_class
= IOPRIO_CLASS_BE
;
4692 bfqd
->oom_bfqq
.entity
.new_weight
=
4693 bfq_ioprio_to_weight(bfqd
->oom_bfqq
.new_ioprio
);
4695 /* oom_bfqq does not participate to bursts */
4696 bfq_clear_bfqq_just_created(&bfqd
->oom_bfqq
);
4699 * Trigger weight initialization, according to ioprio, at the
4700 * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio
4701 * class won't be changed any more.
4703 bfqd
->oom_bfqq
.entity
.prio_changed
= 1;
4707 INIT_LIST_HEAD(&bfqd
->dispatch
);
4709 hrtimer_init(&bfqd
->idle_slice_timer
, CLOCK_MONOTONIC
,
4711 bfqd
->idle_slice_timer
.function
= bfq_idle_slice_timer
;
4713 bfqd
->queue_weights_tree
= RB_ROOT
;
4714 bfqd
->group_weights_tree
= RB_ROOT
;
4716 INIT_LIST_HEAD(&bfqd
->active_list
);
4717 INIT_LIST_HEAD(&bfqd
->idle_list
);
4718 INIT_HLIST_HEAD(&bfqd
->burst_list
);
4722 bfqd
->bfq_max_budget
= bfq_default_max_budget
;
4724 bfqd
->bfq_fifo_expire
[0] = bfq_fifo_expire
[0];
4725 bfqd
->bfq_fifo_expire
[1] = bfq_fifo_expire
[1];
4726 bfqd
->bfq_back_max
= bfq_back_max
;
4727 bfqd
->bfq_back_penalty
= bfq_back_penalty
;
4728 bfqd
->bfq_slice_idle
= bfq_slice_idle
;
4729 bfqd
->bfq_timeout
= bfq_timeout
;
4731 bfqd
->bfq_requests_within_timer
= 120;
4733 bfqd
->bfq_large_burst_thresh
= 8;
4734 bfqd
->bfq_burst_interval
= msecs_to_jiffies(180);
4736 bfqd
->low_latency
= true;
4739 * Trade-off between responsiveness and fairness.
4741 bfqd
->bfq_wr_coeff
= 30;
4742 bfqd
->bfq_wr_rt_max_time
= msecs_to_jiffies(300);
4743 bfqd
->bfq_wr_max_time
= 0;
4744 bfqd
->bfq_wr_min_idle_time
= msecs_to_jiffies(2000);
4745 bfqd
->bfq_wr_min_inter_arr_async
= msecs_to_jiffies(500);
4746 bfqd
->bfq_wr_max_softrt_rate
= 7000; /*
4747 * Approximate rate required
4748 * to playback or record a
4749 * high-definition compressed
4752 bfqd
->wr_busy_queues
= 0;
4755 * Begin by assuming, optimistically, that the device is a
4756 * high-speed one, and that its peak rate is equal to 2/3 of
4757 * the highest reference rate.
4759 bfqd
->RT_prod
= R_fast
[blk_queue_nonrot(bfqd
->queue
)] *
4760 T_fast
[blk_queue_nonrot(bfqd
->queue
)];
4761 bfqd
->peak_rate
= R_fast
[blk_queue_nonrot(bfqd
->queue
)] * 2 / 3;
4762 bfqd
->device_speed
= BFQ_BFQD_FAST
;
4764 spin_lock_init(&bfqd
->lock
);
4767 * The invocation of the next bfq_create_group_hierarchy
4768 * function is the head of a chain of function calls
4769 * (bfq_create_group_hierarchy->blkcg_activate_policy->
4770 * blk_mq_freeze_queue) that may lead to the invocation of the
4771 * has_work hook function. For this reason,
4772 * bfq_create_group_hierarchy is invoked only after all
4773 * scheduler data has been initialized, apart from the fields
4774 * that can be initialized only after invoking
4775 * bfq_create_group_hierarchy. This, in particular, enables
4776 * has_work to correctly return false. Of course, to avoid
4777 * other inconsistencies, the blk-mq stack must then refrain
4778 * from invoking further scheduler hooks before this init
4779 * function is finished.
4781 bfqd
->root_group
= bfq_create_group_hierarchy(bfqd
, q
->node
);
4782 if (!bfqd
->root_group
)
4784 bfq_init_root_group(bfqd
->root_group
, bfqd
);
4785 bfq_init_entity(&bfqd
->oom_bfqq
.entity
, bfqd
->root_group
);
4787 wbt_disable_default(q
);
4792 kobject_put(&eq
->kobj
);
4796 static void bfq_slab_kill(void)
4798 kmem_cache_destroy(bfq_pool
);
4801 static int __init
bfq_slab_setup(void)
4803 bfq_pool
= KMEM_CACHE(bfq_queue
, 0);
4809 static ssize_t
bfq_var_show(unsigned int var
, char *page
)
4811 return sprintf(page
, "%u\n", var
);
4814 static int bfq_var_store(unsigned long *var
, const char *page
)
4816 unsigned long new_val
;
4817 int ret
= kstrtoul(page
, 10, &new_val
);
4825 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
4826 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
4828 struct bfq_data *bfqd = e->elevator_data; \
4829 u64 __data = __VAR; \
4831 __data = jiffies_to_msecs(__data); \
4832 else if (__CONV == 2) \
4833 __data = div_u64(__data, NSEC_PER_MSEC); \
4834 return bfq_var_show(__data, (page)); \
4836 SHOW_FUNCTION(bfq_fifo_expire_sync_show
, bfqd
->bfq_fifo_expire
[1], 2);
4837 SHOW_FUNCTION(bfq_fifo_expire_async_show
, bfqd
->bfq_fifo_expire
[0], 2);
4838 SHOW_FUNCTION(bfq_back_seek_max_show
, bfqd
->bfq_back_max
, 0);
4839 SHOW_FUNCTION(bfq_back_seek_penalty_show
, bfqd
->bfq_back_penalty
, 0);
4840 SHOW_FUNCTION(bfq_slice_idle_show
, bfqd
->bfq_slice_idle
, 2);
4841 SHOW_FUNCTION(bfq_max_budget_show
, bfqd
->bfq_user_max_budget
, 0);
4842 SHOW_FUNCTION(bfq_timeout_sync_show
, bfqd
->bfq_timeout
, 1);
4843 SHOW_FUNCTION(bfq_strict_guarantees_show
, bfqd
->strict_guarantees
, 0);
4844 SHOW_FUNCTION(bfq_low_latency_show
, bfqd
->low_latency
, 0);
4845 #undef SHOW_FUNCTION
4847 #define USEC_SHOW_FUNCTION(__FUNC, __VAR) \
4848 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
4850 struct bfq_data *bfqd = e->elevator_data; \
4851 u64 __data = __VAR; \
4852 __data = div_u64(__data, NSEC_PER_USEC); \
4853 return bfq_var_show(__data, (page)); \
4855 USEC_SHOW_FUNCTION(bfq_slice_idle_us_show
, bfqd
->bfq_slice_idle
);
4856 #undef USEC_SHOW_FUNCTION
4858 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
4860 __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4862 struct bfq_data *bfqd = e->elevator_data; \
4863 unsigned long __data, __min = (MIN), __max = (MAX); \
4866 ret = bfq_var_store(&__data, (page)); \
4869 if (__data < __min) \
4871 else if (__data > __max) \
4874 *(__PTR) = msecs_to_jiffies(__data); \
4875 else if (__CONV == 2) \
4876 *(__PTR) = (u64)__data * NSEC_PER_MSEC; \
4878 *(__PTR) = __data; \
4881 STORE_FUNCTION(bfq_fifo_expire_sync_store
, &bfqd
->bfq_fifo_expire
[1], 1,
4883 STORE_FUNCTION(bfq_fifo_expire_async_store
, &bfqd
->bfq_fifo_expire
[0], 1,
4885 STORE_FUNCTION(bfq_back_seek_max_store
, &bfqd
->bfq_back_max
, 0, INT_MAX
, 0);
4886 STORE_FUNCTION(bfq_back_seek_penalty_store
, &bfqd
->bfq_back_penalty
, 1,
4888 STORE_FUNCTION(bfq_slice_idle_store
, &bfqd
->bfq_slice_idle
, 0, INT_MAX
, 2);
4889 #undef STORE_FUNCTION
4891 #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
4892 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\
4894 struct bfq_data *bfqd = e->elevator_data; \
4895 unsigned long __data, __min = (MIN), __max = (MAX); \
4898 ret = bfq_var_store(&__data, (page)); \
4901 if (__data < __min) \
4903 else if (__data > __max) \
4905 *(__PTR) = (u64)__data * NSEC_PER_USEC; \
4908 USEC_STORE_FUNCTION(bfq_slice_idle_us_store
, &bfqd
->bfq_slice_idle
, 0,
4910 #undef USEC_STORE_FUNCTION
4912 static ssize_t
bfq_max_budget_store(struct elevator_queue
*e
,
4913 const char *page
, size_t count
)
4915 struct bfq_data
*bfqd
= e
->elevator_data
;
4916 unsigned long __data
;
4919 ret
= bfq_var_store(&__data
, (page
));
4924 bfqd
->bfq_max_budget
= bfq_calc_max_budget(bfqd
);
4926 if (__data
> INT_MAX
)
4928 bfqd
->bfq_max_budget
= __data
;
4931 bfqd
->bfq_user_max_budget
= __data
;
4937 * Leaving this name to preserve name compatibility with cfq
4938 * parameters, but this timeout is used for both sync and async.
4940 static ssize_t
bfq_timeout_sync_store(struct elevator_queue
*e
,
4941 const char *page
, size_t count
)
4943 struct bfq_data
*bfqd
= e
->elevator_data
;
4944 unsigned long __data
;
4947 ret
= bfq_var_store(&__data
, (page
));
4953 else if (__data
> INT_MAX
)
4956 bfqd
->bfq_timeout
= msecs_to_jiffies(__data
);
4957 if (bfqd
->bfq_user_max_budget
== 0)
4958 bfqd
->bfq_max_budget
= bfq_calc_max_budget(bfqd
);
4963 static ssize_t
bfq_strict_guarantees_store(struct elevator_queue
*e
,
4964 const char *page
, size_t count
)
4966 struct bfq_data
*bfqd
= e
->elevator_data
;
4967 unsigned long __data
;
4970 ret
= bfq_var_store(&__data
, (page
));
4976 if (!bfqd
->strict_guarantees
&& __data
== 1
4977 && bfqd
->bfq_slice_idle
< 8 * NSEC_PER_MSEC
)
4978 bfqd
->bfq_slice_idle
= 8 * NSEC_PER_MSEC
;
4980 bfqd
->strict_guarantees
= __data
;
4985 static ssize_t
bfq_low_latency_store(struct elevator_queue
*e
,
4986 const char *page
, size_t count
)
4988 struct bfq_data
*bfqd
= e
->elevator_data
;
4989 unsigned long __data
;
4992 ret
= bfq_var_store(&__data
, (page
));
4998 if (__data
== 0 && bfqd
->low_latency
!= 0)
5000 bfqd
->low_latency
= __data
;
5005 #define BFQ_ATTR(name) \
5006 __ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store)
5008 static struct elv_fs_entry bfq_attrs
[] = {
5009 BFQ_ATTR(fifo_expire_sync
),
5010 BFQ_ATTR(fifo_expire_async
),
5011 BFQ_ATTR(back_seek_max
),
5012 BFQ_ATTR(back_seek_penalty
),
5013 BFQ_ATTR(slice_idle
),
5014 BFQ_ATTR(slice_idle_us
),
5015 BFQ_ATTR(max_budget
),
5016 BFQ_ATTR(timeout_sync
),
5017 BFQ_ATTR(strict_guarantees
),
5018 BFQ_ATTR(low_latency
),
5022 static struct elevator_type iosched_bfq_mq
= {
5024 .prepare_request
= bfq_prepare_request
,
5025 .finish_request
= bfq_finish_request
,
5026 .exit_icq
= bfq_exit_icq
,
5027 .insert_requests
= bfq_insert_requests
,
5028 .dispatch_request
= bfq_dispatch_request
,
5029 .next_request
= elv_rb_latter_request
,
5030 .former_request
= elv_rb_former_request
,
5031 .allow_merge
= bfq_allow_bio_merge
,
5032 .bio_merge
= bfq_bio_merge
,
5033 .request_merge
= bfq_request_merge
,
5034 .requests_merged
= bfq_requests_merged
,
5035 .request_merged
= bfq_request_merged
,
5036 .has_work
= bfq_has_work
,
5037 .init_sched
= bfq_init_queue
,
5038 .exit_sched
= bfq_exit_queue
,
5042 .icq_size
= sizeof(struct bfq_io_cq
),
5043 .icq_align
= __alignof__(struct bfq_io_cq
),
5044 .elevator_attrs
= bfq_attrs
,
5045 .elevator_name
= "bfq",
5046 .elevator_owner
= THIS_MODULE
,
5048 MODULE_ALIAS("bfq-iosched");
5050 static int __init
bfq_init(void)
5054 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5055 ret
= blkcg_policy_register(&blkcg_policy_bfq
);
5061 if (bfq_slab_setup())
5065 * Times to load large popular applications for the typical
5066 * systems installed on the reference devices (see the
5067 * comments before the definitions of the next two
5068 * arrays). Actually, we use slightly slower values, as the
5069 * estimated peak rate tends to be smaller than the actual
5070 * peak rate. The reason for this last fact is that estimates
5071 * are computed over much shorter time intervals than the long
5072 * intervals typically used for benchmarking. Why? First, to
5073 * adapt more quickly to variations. Second, because an I/O
5074 * scheduler cannot rely on a peak-rate-evaluation workload to
5075 * be run for a long time.
5077 T_slow
[0] = msecs_to_jiffies(3500); /* actually 4 sec */
5078 T_slow
[1] = msecs_to_jiffies(6000); /* actually 6.5 sec */
5079 T_fast
[0] = msecs_to_jiffies(7000); /* actually 8 sec */
5080 T_fast
[1] = msecs_to_jiffies(2500); /* actually 3 sec */
5083 * Thresholds that determine the switch between speed classes
5084 * (see the comments before the definition of the array
5085 * device_speed_thresh). These thresholds are biased towards
5086 * transitions to the fast class. This is safer than the
5087 * opposite bias. In fact, a wrong transition to the slow
5088 * class results in short weight-raising periods, because the
5089 * speed of the device then tends to be higher that the
5090 * reference peak rate. On the opposite end, a wrong
5091 * transition to the fast class tends to increase
5092 * weight-raising periods, because of the opposite reason.
5094 device_speed_thresh
[0] = (4 * R_slow
[0]) / 3;
5095 device_speed_thresh
[1] = (4 * R_slow
[1]) / 3;
5097 ret
= elv_register(&iosched_bfq_mq
);
5106 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5107 blkcg_policy_unregister(&blkcg_policy_bfq
);
5112 static void __exit
bfq_exit(void)
5114 elv_unregister(&iosched_bfq_mq
);
5115 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5116 blkcg_policy_unregister(&blkcg_policy_bfq
);
5121 module_init(bfq_init
);
5122 module_exit(bfq_exit
);
5124 MODULE_AUTHOR("Paolo Valente");
5125 MODULE_LICENSE("GPL");
5126 MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler");