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 * BFQ is described in [1], where also a reference to the initial, more
60 * theoretical paper on BFQ can be found. The interested reader can find
61 * in the latter paper full details on the main algorithm, as well as
62 * formulas of the guarantees and formal proofs of all the properties.
63 * With respect to the version of BFQ presented in these papers, this
64 * implementation adds a few more heuristics, such as the one that
65 * guarantees a low latency to soft real-time applications, and a
66 * hierarchical extension based on H-WF2Q+.
68 * B-WF2Q+ is based on WF2Q+, which is described in [2], together with
69 * H-WF2Q+, while the augmented tree used here to implement B-WF2Q+
70 * with O(log N) complexity derives from the one introduced with EEVDF
73 * [1] P. Valente, A. Avanzini, "Evolution of the BFQ Storage I/O
74 * Scheduler", Proceedings of the First Workshop on Mobile System
75 * Technologies (MST-2015), May 2015.
76 * http://algogroup.unimore.it/people/paolo/disk_sched/mst-2015.pdf
78 * [2] Jon C.R. Bennett and H. Zhang, "Hierarchical Packet Fair Queueing
79 * Algorithms", IEEE/ACM Transactions on Networking, 5(5):675-689,
82 * http://www.cs.cmu.edu/~hzhang/papers/TON-97-Oct.ps.gz
84 * [3] I. Stoica and H. Abdel-Wahab, "Earliest Eligible Virtual Deadline
85 * First: A Flexible and Accurate Mechanism for Proportional Share
86 * Resource Allocation", technical report.
88 * http://www.cs.berkeley.edu/~istoica/papers/eevdf-tr-95.pdf
90 #include <linux/module.h>
91 #include <linux/slab.h>
92 #include <linux/blkdev.h>
93 #include <linux/cgroup.h>
94 #include <linux/elevator.h>
95 #include <linux/ktime.h>
96 #include <linux/rbtree.h>
97 #include <linux/ioprio.h>
98 #include <linux/sbitmap.h>
99 #include <linux/delay.h>
103 #include "blk-mq-tag.h"
104 #include "blk-mq-sched.h"
105 #include "bfq-iosched.h"
107 #define BFQ_BFQQ_FNS(name) \
108 void bfq_mark_bfqq_##name(struct bfq_queue *bfqq) \
110 __set_bit(BFQQF_##name, &(bfqq)->flags); \
112 void bfq_clear_bfqq_##name(struct bfq_queue *bfqq) \
114 __clear_bit(BFQQF_##name, &(bfqq)->flags); \
116 int bfq_bfqq_##name(const struct bfq_queue *bfqq) \
118 return test_bit(BFQQF_##name, &(bfqq)->flags); \
121 BFQ_BFQQ_FNS(just_created
);
123 BFQ_BFQQ_FNS(wait_request
);
124 BFQ_BFQQ_FNS(non_blocking_wait_rq
);
125 BFQ_BFQQ_FNS(fifo_expire
);
126 BFQ_BFQQ_FNS(idle_window
);
128 BFQ_BFQQ_FNS(IO_bound
);
129 BFQ_BFQQ_FNS(in_large_burst
);
131 BFQ_BFQQ_FNS(split_coop
);
132 BFQ_BFQQ_FNS(softrt_update
);
133 #undef BFQ_BFQQ_FNS \
135 /* Expiration time of sync (0) and async (1) requests, in ns. */
136 static const u64 bfq_fifo_expire
[2] = { NSEC_PER_SEC
/ 4, NSEC_PER_SEC
/ 8 };
138 /* Maximum backwards seek (magic number lifted from CFQ), in KiB. */
139 static const int bfq_back_max
= 16 * 1024;
141 /* Penalty of a backwards seek, in number of sectors. */
142 static const int bfq_back_penalty
= 2;
144 /* Idling period duration, in ns. */
145 static u64 bfq_slice_idle
= NSEC_PER_SEC
/ 125;
147 /* Minimum number of assigned budgets for which stats are safe to compute. */
148 static const int bfq_stats_min_budgets
= 194;
150 /* Default maximum budget values, in sectors and number of requests. */
151 static const int bfq_default_max_budget
= 16 * 1024;
154 * Async to sync throughput distribution is controlled as follows:
155 * when an async request is served, the entity is charged the number
156 * of sectors of the request, multiplied by the factor below
158 static const int bfq_async_charge_factor
= 10;
160 /* Default timeout values, in jiffies, approximating CFQ defaults. */
161 const int bfq_timeout
= HZ
/ 8;
163 static struct kmem_cache
*bfq_pool
;
165 /* Below this threshold (in ns), we consider thinktime immediate. */
166 #define BFQ_MIN_TT (2 * NSEC_PER_MSEC)
168 /* hw_tag detection: parallel requests threshold and min samples needed. */
169 #define BFQ_HW_QUEUE_THRESHOLD 4
170 #define BFQ_HW_QUEUE_SAMPLES 32
172 #define BFQQ_SEEK_THR (sector_t)(8 * 100)
173 #define BFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
174 #define BFQQ_CLOSE_THR (sector_t)(8 * 1024)
175 #define BFQQ_SEEKY(bfqq) (hweight32(bfqq->seek_history) > 32/8)
177 /* Min number of samples required to perform peak-rate update */
178 #define BFQ_RATE_MIN_SAMPLES 32
179 /* Min observation time interval required to perform a peak-rate update (ns) */
180 #define BFQ_RATE_MIN_INTERVAL (300*NSEC_PER_MSEC)
181 /* Target observation time interval for a peak-rate update (ns) */
182 #define BFQ_RATE_REF_INTERVAL NSEC_PER_SEC
184 /* Shift used for peak rate fixed precision calculations. */
185 #define BFQ_RATE_SHIFT 16
188 * By default, BFQ computes the duration of the weight raising for
189 * interactive applications automatically, using the following formula:
190 * duration = (R / r) * T, where r is the peak rate of the device, and
191 * R and T are two reference parameters.
192 * In particular, R is the peak rate of the reference device (see below),
193 * and T is a reference time: given the systems that are likely to be
194 * installed on the reference device according to its speed class, T is
195 * about the maximum time needed, under BFQ and while reading two files in
196 * parallel, to load typical large applications on these systems.
197 * In practice, the slower/faster the device at hand is, the more/less it
198 * takes to load applications with respect to the reference device.
199 * Accordingly, the longer/shorter BFQ grants weight raising to interactive
202 * BFQ uses four different reference pairs (R, T), depending on:
203 * . whether the device is rotational or non-rotational;
204 * . whether the device is slow, such as old or portable HDDs, as well as
205 * SD cards, or fast, such as newer HDDs and SSDs.
207 * The device's speed class is dynamically (re)detected in
208 * bfq_update_peak_rate() every time the estimated peak rate is updated.
210 * In the following definitions, R_slow[0]/R_fast[0] and
211 * T_slow[0]/T_fast[0] are the reference values for a slow/fast
212 * rotational device, whereas R_slow[1]/R_fast[1] and
213 * T_slow[1]/T_fast[1] are the reference values for a slow/fast
214 * non-rotational device. Finally, device_speed_thresh are the
215 * thresholds used to switch between speed classes. The reference
216 * rates are not the actual peak rates of the devices used as a
217 * reference, but slightly lower values. The reason for using these
218 * slightly lower values is that the peak-rate estimator tends to
219 * yield slightly lower values than the actual peak rate (it can yield
220 * the actual peak rate only if there is only one process doing I/O,
221 * and the process does sequential I/O).
223 * Both the reference peak rates and the thresholds are measured in
224 * sectors/usec, left-shifted by BFQ_RATE_SHIFT.
226 static int R_slow
[2] = {1000, 10700};
227 static int R_fast
[2] = {14000, 33000};
229 * To improve readability, a conversion function is used to initialize the
230 * following arrays, which entails that they can be initialized only in a
233 static int T_slow
[2];
234 static int T_fast
[2];
235 static int device_speed_thresh
[2];
237 #define RQ_BIC(rq) ((struct bfq_io_cq *) (rq)->elv.priv[0])
238 #define RQ_BFQQ(rq) ((rq)->elv.priv[1])
240 struct bfq_queue
*bic_to_bfqq(struct bfq_io_cq
*bic
, bool is_sync
)
242 return bic
->bfqq
[is_sync
];
245 void bic_set_bfqq(struct bfq_io_cq
*bic
, struct bfq_queue
*bfqq
, bool is_sync
)
247 bic
->bfqq
[is_sync
] = bfqq
;
250 struct bfq_data
*bic_to_bfqd(struct bfq_io_cq
*bic
)
252 return bic
->icq
.q
->elevator
->elevator_data
;
256 * icq_to_bic - convert iocontext queue structure to bfq_io_cq.
257 * @icq: the iocontext queue.
259 static struct bfq_io_cq
*icq_to_bic(struct io_cq
*icq
)
261 /* bic->icq is the first member, %NULL will convert to %NULL */
262 return container_of(icq
, struct bfq_io_cq
, icq
);
266 * bfq_bic_lookup - search into @ioc a bic associated to @bfqd.
267 * @bfqd: the lookup key.
268 * @ioc: the io_context of the process doing I/O.
269 * @q: the request queue.
271 static struct bfq_io_cq
*bfq_bic_lookup(struct bfq_data
*bfqd
,
272 struct io_context
*ioc
,
273 struct request_queue
*q
)
277 struct bfq_io_cq
*icq
;
279 spin_lock_irqsave(q
->queue_lock
, flags
);
280 icq
= icq_to_bic(ioc_lookup_icq(ioc
, q
));
281 spin_unlock_irqrestore(q
->queue_lock
, flags
);
290 * Scheduler run of queue, if there are requests pending and no one in the
291 * driver that will restart queueing.
293 void bfq_schedule_dispatch(struct bfq_data
*bfqd
)
295 if (bfqd
->queued
!= 0) {
296 bfq_log(bfqd
, "schedule dispatch");
297 blk_mq_run_hw_queues(bfqd
->queue
, true);
301 #define bfq_class_idle(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
302 #define bfq_class_rt(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_RT)
304 #define bfq_sample_valid(samples) ((samples) > 80)
307 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
308 * We choose the request that is closesr to the head right now. Distance
309 * behind the head is penalized and only allowed to a certain extent.
311 static struct request
*bfq_choose_req(struct bfq_data
*bfqd
,
316 sector_t s1
, s2
, d1
= 0, d2
= 0;
317 unsigned long back_max
;
318 #define BFQ_RQ1_WRAP 0x01 /* request 1 wraps */
319 #define BFQ_RQ2_WRAP 0x02 /* request 2 wraps */
320 unsigned int wrap
= 0; /* bit mask: requests behind the disk head? */
322 if (!rq1
|| rq1
== rq2
)
327 if (rq_is_sync(rq1
) && !rq_is_sync(rq2
))
329 else if (rq_is_sync(rq2
) && !rq_is_sync(rq1
))
331 if ((rq1
->cmd_flags
& REQ_META
) && !(rq2
->cmd_flags
& REQ_META
))
333 else if ((rq2
->cmd_flags
& REQ_META
) && !(rq1
->cmd_flags
& REQ_META
))
336 s1
= blk_rq_pos(rq1
);
337 s2
= blk_rq_pos(rq2
);
340 * By definition, 1KiB is 2 sectors.
342 back_max
= bfqd
->bfq_back_max
* 2;
345 * Strict one way elevator _except_ in the case where we allow
346 * short backward seeks which are biased as twice the cost of a
347 * similar forward seek.
351 else if (s1
+ back_max
>= last
)
352 d1
= (last
- s1
) * bfqd
->bfq_back_penalty
;
354 wrap
|= BFQ_RQ1_WRAP
;
358 else if (s2
+ back_max
>= last
)
359 d2
= (last
- s2
) * bfqd
->bfq_back_penalty
;
361 wrap
|= BFQ_RQ2_WRAP
;
363 /* Found required data */
366 * By doing switch() on the bit mask "wrap" we avoid having to
367 * check two variables for all permutations: --> faster!
370 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
385 case BFQ_RQ1_WRAP
|BFQ_RQ2_WRAP
: /* both rqs wrapped */
388 * Since both rqs are wrapped,
389 * start with the one that's further behind head
390 * (--> only *one* back seek required),
391 * since back seek takes more time than forward.
400 static struct bfq_queue
*
401 bfq_rq_pos_tree_lookup(struct bfq_data
*bfqd
, struct rb_root
*root
,
402 sector_t sector
, struct rb_node
**ret_parent
,
403 struct rb_node
***rb_link
)
405 struct rb_node
**p
, *parent
;
406 struct bfq_queue
*bfqq
= NULL
;
414 bfqq
= rb_entry(parent
, struct bfq_queue
, pos_node
);
417 * Sort strictly based on sector. Smallest to the left,
418 * largest to the right.
420 if (sector
> blk_rq_pos(bfqq
->next_rq
))
422 else if (sector
< blk_rq_pos(bfqq
->next_rq
))
430 *ret_parent
= parent
;
434 bfq_log(bfqd
, "rq_pos_tree_lookup %llu: returning %d",
435 (unsigned long long)sector
,
436 bfqq
? bfqq
->pid
: 0);
441 void bfq_pos_tree_add_move(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
443 struct rb_node
**p
, *parent
;
444 struct bfq_queue
*__bfqq
;
446 if (bfqq
->pos_root
) {
447 rb_erase(&bfqq
->pos_node
, bfqq
->pos_root
);
448 bfqq
->pos_root
= NULL
;
451 if (bfq_class_idle(bfqq
))
456 bfqq
->pos_root
= &bfq_bfqq_to_bfqg(bfqq
)->rq_pos_tree
;
457 __bfqq
= bfq_rq_pos_tree_lookup(bfqd
, bfqq
->pos_root
,
458 blk_rq_pos(bfqq
->next_rq
), &parent
, &p
);
460 rb_link_node(&bfqq
->pos_node
, parent
, p
);
461 rb_insert_color(&bfqq
->pos_node
, bfqq
->pos_root
);
463 bfqq
->pos_root
= NULL
;
467 * Tell whether there are active queues or groups with differentiated weights.
469 static bool bfq_differentiated_weights(struct bfq_data
*bfqd
)
472 * For weights to differ, at least one of the trees must contain
473 * at least two nodes.
475 return (!RB_EMPTY_ROOT(&bfqd
->queue_weights_tree
) &&
476 (bfqd
->queue_weights_tree
.rb_node
->rb_left
||
477 bfqd
->queue_weights_tree
.rb_node
->rb_right
)
478 #ifdef CONFIG_BFQ_GROUP_IOSCHED
480 (!RB_EMPTY_ROOT(&bfqd
->group_weights_tree
) &&
481 (bfqd
->group_weights_tree
.rb_node
->rb_left
||
482 bfqd
->group_weights_tree
.rb_node
->rb_right
)
488 * The following function returns true if every queue must receive the
489 * same share of the throughput (this condition is used when deciding
490 * whether idling may be disabled, see the comments in the function
491 * bfq_bfqq_may_idle()).
493 * Such a scenario occurs when:
494 * 1) all active queues have the same weight,
495 * 2) all active groups at the same level in the groups tree have the same
497 * 3) all active groups at the same level in the groups tree have the same
498 * number of children.
500 * Unfortunately, keeping the necessary state for evaluating exactly the
501 * above symmetry conditions would be quite complex and time-consuming.
502 * Therefore this function evaluates, instead, the following stronger
503 * sub-conditions, for which it is much easier to maintain the needed
505 * 1) all active queues have the same weight,
506 * 2) all active groups have the same weight,
507 * 3) all active groups have at most one active child each.
508 * In particular, the last two conditions are always true if hierarchical
509 * support and the cgroups interface are not enabled, thus no state needs
510 * to be maintained in this case.
512 static bool bfq_symmetric_scenario(struct bfq_data
*bfqd
)
514 return !bfq_differentiated_weights(bfqd
);
518 * If the weight-counter tree passed as input contains no counter for
519 * the weight of the input entity, then add that counter; otherwise just
520 * increment the existing counter.
522 * Note that weight-counter trees contain few nodes in mostly symmetric
523 * scenarios. For example, if all queues have the same weight, then the
524 * weight-counter tree for the queues may contain at most one node.
525 * This holds even if low_latency is on, because weight-raised queues
526 * are not inserted in the tree.
527 * In most scenarios, the rate at which nodes are created/destroyed
530 void bfq_weights_tree_add(struct bfq_data
*bfqd
, struct bfq_entity
*entity
,
531 struct rb_root
*root
)
533 struct rb_node
**new = &(root
->rb_node
), *parent
= NULL
;
536 * Do not insert if the entity is already associated with a
537 * counter, which happens if:
538 * 1) the entity is associated with a queue,
539 * 2) a request arrival has caused the queue to become both
540 * non-weight-raised, and hence change its weight, and
541 * backlogged; in this respect, each of the two events
542 * causes an invocation of this function,
543 * 3) this is the invocation of this function caused by the
544 * second event. This second invocation is actually useless,
545 * and we handle this fact by exiting immediately. More
546 * efficient or clearer solutions might possibly be adopted.
548 if (entity
->weight_counter
)
552 struct bfq_weight_counter
*__counter
= container_of(*new,
553 struct bfq_weight_counter
,
557 if (entity
->weight
== __counter
->weight
) {
558 entity
->weight_counter
= __counter
;
561 if (entity
->weight
< __counter
->weight
)
562 new = &((*new)->rb_left
);
564 new = &((*new)->rb_right
);
567 entity
->weight_counter
= kzalloc(sizeof(struct bfq_weight_counter
),
571 * In the unlucky event of an allocation failure, we just
572 * exit. This will cause the weight of entity to not be
573 * considered in bfq_differentiated_weights, which, in its
574 * turn, causes the scenario to be deemed wrongly symmetric in
575 * case entity's weight would have been the only weight making
576 * the scenario asymmetric. On the bright side, no unbalance
577 * will however occur when entity becomes inactive again (the
578 * invocation of this function is triggered by an activation
579 * of entity). In fact, bfq_weights_tree_remove does nothing
580 * if !entity->weight_counter.
582 if (unlikely(!entity
->weight_counter
))
585 entity
->weight_counter
->weight
= entity
->weight
;
586 rb_link_node(&entity
->weight_counter
->weights_node
, parent
, new);
587 rb_insert_color(&entity
->weight_counter
->weights_node
, root
);
590 entity
->weight_counter
->num_active
++;
594 * Decrement the weight counter associated with the entity, and, if the
595 * counter reaches 0, remove the counter from the tree.
596 * See the comments to the function bfq_weights_tree_add() for considerations
599 void bfq_weights_tree_remove(struct bfq_data
*bfqd
, struct bfq_entity
*entity
,
600 struct rb_root
*root
)
602 if (!entity
->weight_counter
)
605 entity
->weight_counter
->num_active
--;
606 if (entity
->weight_counter
->num_active
> 0)
607 goto reset_entity_pointer
;
609 rb_erase(&entity
->weight_counter
->weights_node
, root
);
610 kfree(entity
->weight_counter
);
612 reset_entity_pointer
:
613 entity
->weight_counter
= NULL
;
617 * Return expired entry, or NULL to just start from scratch in rbtree.
619 static struct request
*bfq_check_fifo(struct bfq_queue
*bfqq
,
620 struct request
*last
)
624 if (bfq_bfqq_fifo_expire(bfqq
))
627 bfq_mark_bfqq_fifo_expire(bfqq
);
629 rq
= rq_entry_fifo(bfqq
->fifo
.next
);
631 if (rq
== last
|| ktime_get_ns() < rq
->fifo_time
)
634 bfq_log_bfqq(bfqq
->bfqd
, bfqq
, "check_fifo: returned %p", rq
);
638 static struct request
*bfq_find_next_rq(struct bfq_data
*bfqd
,
639 struct bfq_queue
*bfqq
,
640 struct request
*last
)
642 struct rb_node
*rbnext
= rb_next(&last
->rb_node
);
643 struct rb_node
*rbprev
= rb_prev(&last
->rb_node
);
644 struct request
*next
, *prev
= NULL
;
646 /* Follow expired path, else get first next available. */
647 next
= bfq_check_fifo(bfqq
, last
);
652 prev
= rb_entry_rq(rbprev
);
655 next
= rb_entry_rq(rbnext
);
657 rbnext
= rb_first(&bfqq
->sort_list
);
658 if (rbnext
&& rbnext
!= &last
->rb_node
)
659 next
= rb_entry_rq(rbnext
);
662 return bfq_choose_req(bfqd
, next
, prev
, blk_rq_pos(last
));
665 /* see the definition of bfq_async_charge_factor for details */
666 static unsigned long bfq_serv_to_charge(struct request
*rq
,
667 struct bfq_queue
*bfqq
)
669 if (bfq_bfqq_sync(bfqq
) || bfqq
->wr_coeff
> 1)
670 return blk_rq_sectors(rq
);
673 * If there are no weight-raised queues, then amplify service
674 * by just the async charge factor; otherwise amplify service
675 * by twice the async charge factor, to further reduce latency
676 * for weight-raised queues.
678 if (bfqq
->bfqd
->wr_busy_queues
== 0)
679 return blk_rq_sectors(rq
) * bfq_async_charge_factor
;
681 return blk_rq_sectors(rq
) * 2 * bfq_async_charge_factor
;
685 * bfq_updated_next_req - update the queue after a new next_rq selection.
686 * @bfqd: the device data the queue belongs to.
687 * @bfqq: the queue to update.
689 * If the first request of a queue changes we make sure that the queue
690 * has enough budget to serve at least its first request (if the
691 * request has grown). We do this because if the queue has not enough
692 * budget for its first request, it has to go through two dispatch
693 * rounds to actually get it dispatched.
695 static void bfq_updated_next_req(struct bfq_data
*bfqd
,
696 struct bfq_queue
*bfqq
)
698 struct bfq_entity
*entity
= &bfqq
->entity
;
699 struct request
*next_rq
= bfqq
->next_rq
;
700 unsigned long new_budget
;
705 if (bfqq
== bfqd
->in_service_queue
)
707 * In order not to break guarantees, budgets cannot be
708 * changed after an entity has been selected.
712 new_budget
= max_t(unsigned long, bfqq
->max_budget
,
713 bfq_serv_to_charge(next_rq
, bfqq
));
714 if (entity
->budget
!= new_budget
) {
715 entity
->budget
= new_budget
;
716 bfq_log_bfqq(bfqd
, bfqq
, "updated next rq: new budget %lu",
718 bfq_requeue_bfqq(bfqd
, bfqq
);
723 bfq_bfqq_resume_state(struct bfq_queue
*bfqq
, struct bfq_io_cq
*bic
)
725 if (bic
->saved_idle_window
)
726 bfq_mark_bfqq_idle_window(bfqq
);
728 bfq_clear_bfqq_idle_window(bfqq
);
730 if (bic
->saved_IO_bound
)
731 bfq_mark_bfqq_IO_bound(bfqq
);
733 bfq_clear_bfqq_IO_bound(bfqq
);
735 bfqq
->ttime
= bic
->saved_ttime
;
736 bfqq
->wr_coeff
= bic
->saved_wr_coeff
;
737 bfqq
->wr_start_at_switch_to_srt
= bic
->saved_wr_start_at_switch_to_srt
;
738 bfqq
->last_wr_start_finish
= bic
->saved_last_wr_start_finish
;
739 bfqq
->wr_cur_max_time
= bic
->saved_wr_cur_max_time
;
741 if (bfqq
->wr_coeff
> 1 && (bfq_bfqq_in_large_burst(bfqq
) ||
742 time_is_before_jiffies(bfqq
->last_wr_start_finish
+
743 bfqq
->wr_cur_max_time
))) {
744 bfq_log_bfqq(bfqq
->bfqd
, bfqq
,
745 "resume state: switching off wr");
750 /* make sure weight will be updated, however we got here */
751 bfqq
->entity
.prio_changed
= 1;
754 static int bfqq_process_refs(struct bfq_queue
*bfqq
)
756 return bfqq
->ref
- bfqq
->allocated
- bfqq
->entity
.on_st
;
759 /* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */
760 static void bfq_reset_burst_list(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
762 struct bfq_queue
*item
;
763 struct hlist_node
*n
;
765 hlist_for_each_entry_safe(item
, n
, &bfqd
->burst_list
, burst_list_node
)
766 hlist_del_init(&item
->burst_list_node
);
767 hlist_add_head(&bfqq
->burst_list_node
, &bfqd
->burst_list
);
768 bfqd
->burst_size
= 1;
769 bfqd
->burst_parent_entity
= bfqq
->entity
.parent
;
772 /* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */
773 static void bfq_add_to_burst(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
775 /* Increment burst size to take into account also bfqq */
778 if (bfqd
->burst_size
== bfqd
->bfq_large_burst_thresh
) {
779 struct bfq_queue
*pos
, *bfqq_item
;
780 struct hlist_node
*n
;
783 * Enough queues have been activated shortly after each
784 * other to consider this burst as large.
786 bfqd
->large_burst
= true;
789 * We can now mark all queues in the burst list as
790 * belonging to a large burst.
792 hlist_for_each_entry(bfqq_item
, &bfqd
->burst_list
,
794 bfq_mark_bfqq_in_large_burst(bfqq_item
);
795 bfq_mark_bfqq_in_large_burst(bfqq
);
798 * From now on, and until the current burst finishes, any
799 * new queue being activated shortly after the last queue
800 * was inserted in the burst can be immediately marked as
801 * belonging to a large burst. So the burst list is not
802 * needed any more. Remove it.
804 hlist_for_each_entry_safe(pos
, n
, &bfqd
->burst_list
,
806 hlist_del_init(&pos
->burst_list_node
);
808 * Burst not yet large: add bfqq to the burst list. Do
809 * not increment the ref counter for bfqq, because bfqq
810 * is removed from the burst list before freeing bfqq
813 hlist_add_head(&bfqq
->burst_list_node
, &bfqd
->burst_list
);
817 * If many queues belonging to the same group happen to be created
818 * shortly after each other, then the processes associated with these
819 * queues have typically a common goal. In particular, bursts of queue
820 * creations are usually caused by services or applications that spawn
821 * many parallel threads/processes. Examples are systemd during boot,
822 * or git grep. To help these processes get their job done as soon as
823 * possible, it is usually better to not grant either weight-raising
824 * or device idling to their queues.
826 * In this comment we describe, firstly, the reasons why this fact
827 * holds, and, secondly, the next function, which implements the main
828 * steps needed to properly mark these queues so that they can then be
829 * treated in a different way.
831 * The above services or applications benefit mostly from a high
832 * throughput: the quicker the requests of the activated queues are
833 * cumulatively served, the sooner the target job of these queues gets
834 * completed. As a consequence, weight-raising any of these queues,
835 * which also implies idling the device for it, is almost always
836 * counterproductive. In most cases it just lowers throughput.
838 * On the other hand, a burst of queue creations may be caused also by
839 * the start of an application that does not consist of a lot of
840 * parallel I/O-bound threads. In fact, with a complex application,
841 * several short processes may need to be executed to start-up the
842 * application. In this respect, to start an application as quickly as
843 * possible, the best thing to do is in any case to privilege the I/O
844 * related to the application with respect to all other
845 * I/O. Therefore, the best strategy to start as quickly as possible
846 * an application that causes a burst of queue creations is to
847 * weight-raise all the queues created during the burst. This is the
848 * exact opposite of the best strategy for the other type of bursts.
850 * In the end, to take the best action for each of the two cases, the
851 * two types of bursts need to be distinguished. Fortunately, this
852 * seems relatively easy, by looking at the sizes of the bursts. In
853 * particular, we found a threshold such that only bursts with a
854 * larger size than that threshold are apparently caused by
855 * services or commands such as systemd or git grep. For brevity,
856 * hereafter we call just 'large' these bursts. BFQ *does not*
857 * weight-raise queues whose creation occurs in a large burst. In
858 * addition, for each of these queues BFQ performs or does not perform
859 * idling depending on which choice boosts the throughput more. The
860 * exact choice depends on the device and request pattern at
863 * Unfortunately, false positives may occur while an interactive task
864 * is starting (e.g., an application is being started). The
865 * consequence is that the queues associated with the task do not
866 * enjoy weight raising as expected. Fortunately these false positives
867 * are very rare. They typically occur if some service happens to
868 * start doing I/O exactly when the interactive task starts.
870 * Turning back to the next function, it implements all the steps
871 * needed to detect the occurrence of a large burst and to properly
872 * mark all the queues belonging to it (so that they can then be
873 * treated in a different way). This goal is achieved by maintaining a
874 * "burst list" that holds, temporarily, the queues that belong to the
875 * burst in progress. The list is then used to mark these queues as
876 * belonging to a large burst if the burst does become large. The main
877 * steps are the following.
879 * . when the very first queue is created, the queue is inserted into the
880 * list (as it could be the first queue in a possible burst)
882 * . if the current burst has not yet become large, and a queue Q that does
883 * not yet belong to the burst is activated shortly after the last time
884 * at which a new queue entered the burst list, then the function appends
885 * Q to the burst list
887 * . if, as a consequence of the previous step, the burst size reaches
888 * the large-burst threshold, then
890 * . all the queues in the burst list are marked as belonging to a
893 * . the burst list is deleted; in fact, the burst list already served
894 * its purpose (keeping temporarily track of the queues in a burst,
895 * so as to be able to mark them as belonging to a large burst in the
896 * previous sub-step), and now is not needed any more
898 * . the device enters a large-burst mode
900 * . if a queue Q that does not belong to the burst is created while
901 * the device is in large-burst mode and shortly after the last time
902 * at which a queue either entered the burst list or was marked as
903 * belonging to the current large burst, then Q is immediately marked
904 * as belonging to a large burst.
906 * . if a queue Q that does not belong to the burst is created a while
907 * later, i.e., not shortly after, than the last time at which a queue
908 * either entered the burst list or was marked as belonging to the
909 * current large burst, then the current burst is deemed as finished and:
911 * . the large-burst mode is reset if set
913 * . the burst list is emptied
915 * . Q is inserted in the burst list, as Q may be the first queue
916 * in a possible new burst (then the burst list contains just Q
919 static void bfq_handle_burst(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
922 * If bfqq is already in the burst list or is part of a large
923 * burst, or finally has just been split, then there is
924 * nothing else to do.
926 if (!hlist_unhashed(&bfqq
->burst_list_node
) ||
927 bfq_bfqq_in_large_burst(bfqq
) ||
928 time_is_after_eq_jiffies(bfqq
->split_time
+
929 msecs_to_jiffies(10)))
933 * If bfqq's creation happens late enough, or bfqq belongs to
934 * a different group than the burst group, then the current
935 * burst is finished, and related data structures must be
938 * In this respect, consider the special case where bfqq is
939 * the very first queue created after BFQ is selected for this
940 * device. In this case, last_ins_in_burst and
941 * burst_parent_entity are not yet significant when we get
942 * here. But it is easy to verify that, whether or not the
943 * following condition is true, bfqq will end up being
944 * inserted into the burst list. In particular the list will
945 * happen to contain only bfqq. And this is exactly what has
946 * to happen, as bfqq may be the first queue of the first
949 if (time_is_before_jiffies(bfqd
->last_ins_in_burst
+
950 bfqd
->bfq_burst_interval
) ||
951 bfqq
->entity
.parent
!= bfqd
->burst_parent_entity
) {
952 bfqd
->large_burst
= false;
953 bfq_reset_burst_list(bfqd
, bfqq
);
958 * If we get here, then bfqq is being activated shortly after the
959 * last queue. So, if the current burst is also large, we can mark
960 * bfqq as belonging to this large burst immediately.
962 if (bfqd
->large_burst
) {
963 bfq_mark_bfqq_in_large_burst(bfqq
);
968 * If we get here, then a large-burst state has not yet been
969 * reached, but bfqq is being activated shortly after the last
970 * queue. Then we add bfqq to the burst.
972 bfq_add_to_burst(bfqd
, bfqq
);
975 * At this point, bfqq either has been added to the current
976 * burst or has caused the current burst to terminate and a
977 * possible new burst to start. In particular, in the second
978 * case, bfqq has become the first queue in the possible new
979 * burst. In both cases last_ins_in_burst needs to be moved
982 bfqd
->last_ins_in_burst
= jiffies
;
985 static int bfq_bfqq_budget_left(struct bfq_queue
*bfqq
)
987 struct bfq_entity
*entity
= &bfqq
->entity
;
989 return entity
->budget
- entity
->service
;
993 * If enough samples have been computed, return the current max budget
994 * stored in bfqd, which is dynamically updated according to the
995 * estimated disk peak rate; otherwise return the default max budget
997 static int bfq_max_budget(struct bfq_data
*bfqd
)
999 if (bfqd
->budgets_assigned
< bfq_stats_min_budgets
)
1000 return bfq_default_max_budget
;
1002 return bfqd
->bfq_max_budget
;
1006 * Return min budget, which is a fraction of the current or default
1007 * max budget (trying with 1/32)
1009 static int bfq_min_budget(struct bfq_data
*bfqd
)
1011 if (bfqd
->budgets_assigned
< bfq_stats_min_budgets
)
1012 return bfq_default_max_budget
/ 32;
1014 return bfqd
->bfq_max_budget
/ 32;
1018 * The next function, invoked after the input queue bfqq switches from
1019 * idle to busy, updates the budget of bfqq. The function also tells
1020 * whether the in-service queue should be expired, by returning
1021 * true. The purpose of expiring the in-service queue is to give bfqq
1022 * the chance to possibly preempt the in-service queue, and the reason
1023 * for preempting the in-service queue is to achieve one of the two
1026 * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has
1027 * expired because it has remained idle. In particular, bfqq may have
1028 * expired for one of the following two reasons:
1030 * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling
1031 * and did not make it to issue a new request before its last
1032 * request was served;
1034 * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue
1035 * a new request before the expiration of the idling-time.
1037 * Even if bfqq has expired for one of the above reasons, the process
1038 * associated with the queue may be however issuing requests greedily,
1039 * and thus be sensitive to the bandwidth it receives (bfqq may have
1040 * remained idle for other reasons: CPU high load, bfqq not enjoying
1041 * idling, I/O throttling somewhere in the path from the process to
1042 * the I/O scheduler, ...). But if, after every expiration for one of
1043 * the above two reasons, bfqq has to wait for the service of at least
1044 * one full budget of another queue before being served again, then
1045 * bfqq is likely to get a much lower bandwidth or resource time than
1046 * its reserved ones. To address this issue, two countermeasures need
1049 * First, the budget and the timestamps of bfqq need to be updated in
1050 * a special way on bfqq reactivation: they need to be updated as if
1051 * bfqq did not remain idle and did not expire. In fact, if they are
1052 * computed as if bfqq expired and remained idle until reactivation,
1053 * then the process associated with bfqq is treated as if, instead of
1054 * being greedy, it stopped issuing requests when bfqq remained idle,
1055 * and restarts issuing requests only on this reactivation. In other
1056 * words, the scheduler does not help the process recover the "service
1057 * hole" between bfqq expiration and reactivation. As a consequence,
1058 * the process receives a lower bandwidth than its reserved one. In
1059 * contrast, to recover this hole, the budget must be updated as if
1060 * bfqq was not expired at all before this reactivation, i.e., it must
1061 * be set to the value of the remaining budget when bfqq was
1062 * expired. Along the same line, timestamps need to be assigned the
1063 * value they had the last time bfqq was selected for service, i.e.,
1064 * before last expiration. Thus timestamps need to be back-shifted
1065 * with respect to their normal computation (see [1] for more details
1066 * on this tricky aspect).
1068 * Secondly, to allow the process to recover the hole, the in-service
1069 * queue must be expired too, to give bfqq the chance to preempt it
1070 * immediately. In fact, if bfqq has to wait for a full budget of the
1071 * in-service queue to be completed, then it may become impossible to
1072 * let the process recover the hole, even if the back-shifted
1073 * timestamps of bfqq are lower than those of the in-service queue. If
1074 * this happens for most or all of the holes, then the process may not
1075 * receive its reserved bandwidth. In this respect, it is worth noting
1076 * that, being the service of outstanding requests unpreemptible, a
1077 * little fraction of the holes may however be unrecoverable, thereby
1078 * causing a little loss of bandwidth.
1080 * The last important point is detecting whether bfqq does need this
1081 * bandwidth recovery. In this respect, the next function deems the
1082 * process associated with bfqq greedy, and thus allows it to recover
1083 * the hole, if: 1) the process is waiting for the arrival of a new
1084 * request (which implies that bfqq expired for one of the above two
1085 * reasons), and 2) such a request has arrived soon. The first
1086 * condition is controlled through the flag non_blocking_wait_rq,
1087 * while the second through the flag arrived_in_time. If both
1088 * conditions hold, then the function computes the budget in the
1089 * above-described special way, and signals that the in-service queue
1090 * should be expired. Timestamp back-shifting is done later in
1091 * __bfq_activate_entity.
1093 * 2. Reduce latency. Even if timestamps are not backshifted to let
1094 * the process associated with bfqq recover a service hole, bfqq may
1095 * however happen to have, after being (re)activated, a lower finish
1096 * timestamp than the in-service queue. That is, the next budget of
1097 * bfqq may have to be completed before the one of the in-service
1098 * queue. If this is the case, then preempting the in-service queue
1099 * allows this goal to be achieved, apart from the unpreemptible,
1100 * outstanding requests mentioned above.
1102 * Unfortunately, regardless of which of the above two goals one wants
1103 * to achieve, service trees need first to be updated to know whether
1104 * the in-service queue must be preempted. To have service trees
1105 * correctly updated, the in-service queue must be expired and
1106 * rescheduled, and bfqq must be scheduled too. This is one of the
1107 * most costly operations (in future versions, the scheduling
1108 * mechanism may be re-designed in such a way to make it possible to
1109 * know whether preemption is needed without needing to update service
1110 * trees). In addition, queue preemptions almost always cause random
1111 * I/O, and thus loss of throughput. Because of these facts, the next
1112 * function adopts the following simple scheme to avoid both costly
1113 * operations and too frequent preemptions: it requests the expiration
1114 * of the in-service queue (unconditionally) only for queues that need
1115 * to recover a hole, or that either are weight-raised or deserve to
1118 static bool bfq_bfqq_update_budg_for_activation(struct bfq_data
*bfqd
,
1119 struct bfq_queue
*bfqq
,
1120 bool arrived_in_time
,
1121 bool wr_or_deserves_wr
)
1123 struct bfq_entity
*entity
= &bfqq
->entity
;
1125 if (bfq_bfqq_non_blocking_wait_rq(bfqq
) && arrived_in_time
) {
1127 * We do not clear the flag non_blocking_wait_rq here, as
1128 * the latter is used in bfq_activate_bfqq to signal
1129 * that timestamps need to be back-shifted (and is
1130 * cleared right after).
1134 * In next assignment we rely on that either
1135 * entity->service or entity->budget are not updated
1136 * on expiration if bfqq is empty (see
1137 * __bfq_bfqq_recalc_budget). Thus both quantities
1138 * remain unchanged after such an expiration, and the
1139 * following statement therefore assigns to
1140 * entity->budget the remaining budget on such an
1141 * expiration. For clarity, entity->service is not
1142 * updated on expiration in any case, and, in normal
1143 * operation, is reset only when bfqq is selected for
1144 * service (see bfq_get_next_queue).
1146 entity
->budget
= min_t(unsigned long,
1147 bfq_bfqq_budget_left(bfqq
),
1153 entity
->budget
= max_t(unsigned long, bfqq
->max_budget
,
1154 bfq_serv_to_charge(bfqq
->next_rq
, bfqq
));
1155 bfq_clear_bfqq_non_blocking_wait_rq(bfqq
);
1156 return wr_or_deserves_wr
;
1159 static unsigned int bfq_wr_duration(struct bfq_data
*bfqd
)
1163 if (bfqd
->bfq_wr_max_time
> 0)
1164 return bfqd
->bfq_wr_max_time
;
1166 dur
= bfqd
->RT_prod
;
1167 do_div(dur
, bfqd
->peak_rate
);
1170 * Limit duration between 3 and 13 seconds. Tests show that
1171 * higher values than 13 seconds often yield the opposite of
1172 * the desired result, i.e., worsen responsiveness by letting
1173 * non-interactive and non-soft-real-time applications
1174 * preserve weight raising for a too long time interval.
1176 * On the other end, lower values than 3 seconds make it
1177 * difficult for most interactive tasks to complete their jobs
1178 * before weight-raising finishes.
1180 if (dur
> msecs_to_jiffies(13000))
1181 dur
= msecs_to_jiffies(13000);
1182 else if (dur
< msecs_to_jiffies(3000))
1183 dur
= msecs_to_jiffies(3000);
1188 static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data
*bfqd
,
1189 struct bfq_queue
*bfqq
,
1190 unsigned int old_wr_coeff
,
1191 bool wr_or_deserves_wr
,
1196 if (old_wr_coeff
== 1 && wr_or_deserves_wr
) {
1197 /* start a weight-raising period */
1199 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
;
1200 bfqq
->wr_cur_max_time
= bfq_wr_duration(bfqd
);
1202 bfqq
->wr_start_at_switch_to_srt
= jiffies
;
1203 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
*
1204 BFQ_SOFTRT_WEIGHT_FACTOR
;
1205 bfqq
->wr_cur_max_time
=
1206 bfqd
->bfq_wr_rt_max_time
;
1210 * If needed, further reduce budget to make sure it is
1211 * close to bfqq's backlog, so as to reduce the
1212 * scheduling-error component due to a too large
1213 * budget. Do not care about throughput consequences,
1214 * but only about latency. Finally, do not assign a
1215 * too small budget either, to avoid increasing
1216 * latency by causing too frequent expirations.
1218 bfqq
->entity
.budget
= min_t(unsigned long,
1219 bfqq
->entity
.budget
,
1220 2 * bfq_min_budget(bfqd
));
1221 } else if (old_wr_coeff
> 1) {
1222 if (interactive
) { /* update wr coeff and duration */
1223 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
;
1224 bfqq
->wr_cur_max_time
= bfq_wr_duration(bfqd
);
1225 } else if (in_burst
)
1229 * The application is now or still meeting the
1230 * requirements for being deemed soft rt. We
1231 * can then correctly and safely (re)charge
1232 * the weight-raising duration for the
1233 * application with the weight-raising
1234 * duration for soft rt applications.
1236 * In particular, doing this recharge now, i.e.,
1237 * before the weight-raising period for the
1238 * application finishes, reduces the probability
1239 * of the following negative scenario:
1240 * 1) the weight of a soft rt application is
1241 * raised at startup (as for any newly
1242 * created application),
1243 * 2) since the application is not interactive,
1244 * at a certain time weight-raising is
1245 * stopped for the application,
1246 * 3) at that time the application happens to
1247 * still have pending requests, and hence
1248 * is destined to not have a chance to be
1249 * deemed soft rt before these requests are
1250 * completed (see the comments to the
1251 * function bfq_bfqq_softrt_next_start()
1252 * for details on soft rt detection),
1253 * 4) these pending requests experience a high
1254 * latency because the application is not
1255 * weight-raised while they are pending.
1257 if (bfqq
->wr_cur_max_time
!=
1258 bfqd
->bfq_wr_rt_max_time
) {
1259 bfqq
->wr_start_at_switch_to_srt
=
1260 bfqq
->last_wr_start_finish
;
1262 bfqq
->wr_cur_max_time
=
1263 bfqd
->bfq_wr_rt_max_time
;
1264 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
*
1265 BFQ_SOFTRT_WEIGHT_FACTOR
;
1267 bfqq
->last_wr_start_finish
= jiffies
;
1272 static bool bfq_bfqq_idle_for_long_time(struct bfq_data
*bfqd
,
1273 struct bfq_queue
*bfqq
)
1275 return bfqq
->dispatched
== 0 &&
1276 time_is_before_jiffies(
1277 bfqq
->budget_timeout
+
1278 bfqd
->bfq_wr_min_idle_time
);
1281 static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data
*bfqd
,
1282 struct bfq_queue
*bfqq
,
1287 bool soft_rt
, in_burst
, wr_or_deserves_wr
,
1288 bfqq_wants_to_preempt
,
1289 idle_for_long_time
= bfq_bfqq_idle_for_long_time(bfqd
, bfqq
),
1291 * See the comments on
1292 * bfq_bfqq_update_budg_for_activation for
1293 * details on the usage of the next variable.
1295 arrived_in_time
= ktime_get_ns() <=
1296 bfqq
->ttime
.last_end_request
+
1297 bfqd
->bfq_slice_idle
* 3;
1299 bfqg_stats_update_io_add(bfqq_group(RQ_BFQQ(rq
)), bfqq
, rq
->cmd_flags
);
1302 * bfqq deserves to be weight-raised if:
1304 * - it does not belong to a large burst,
1305 * - it has been idle for enough time or is soft real-time,
1306 * - is linked to a bfq_io_cq (it is not shared in any sense).
1308 in_burst
= bfq_bfqq_in_large_burst(bfqq
);
1309 soft_rt
= bfqd
->bfq_wr_max_softrt_rate
> 0 &&
1311 time_is_before_jiffies(bfqq
->soft_rt_next_start
);
1312 *interactive
= !in_burst
&& idle_for_long_time
;
1313 wr_or_deserves_wr
= bfqd
->low_latency
&&
1314 (bfqq
->wr_coeff
> 1 ||
1315 (bfq_bfqq_sync(bfqq
) &&
1316 bfqq
->bic
&& (*interactive
|| soft_rt
)));
1319 * Using the last flag, update budget and check whether bfqq
1320 * may want to preempt the in-service queue.
1322 bfqq_wants_to_preempt
=
1323 bfq_bfqq_update_budg_for_activation(bfqd
, bfqq
,
1328 * If bfqq happened to be activated in a burst, but has been
1329 * idle for much more than an interactive queue, then we
1330 * assume that, in the overall I/O initiated in the burst, the
1331 * I/O associated with bfqq is finished. So bfqq does not need
1332 * to be treated as a queue belonging to a burst
1333 * anymore. Accordingly, we reset bfqq's in_large_burst flag
1334 * if set, and remove bfqq from the burst list if it's
1335 * there. We do not decrement burst_size, because the fact
1336 * that bfqq does not need to belong to the burst list any
1337 * more does not invalidate the fact that bfqq was created in
1340 if (likely(!bfq_bfqq_just_created(bfqq
)) &&
1341 idle_for_long_time
&&
1342 time_is_before_jiffies(
1343 bfqq
->budget_timeout
+
1344 msecs_to_jiffies(10000))) {
1345 hlist_del_init(&bfqq
->burst_list_node
);
1346 bfq_clear_bfqq_in_large_burst(bfqq
);
1349 bfq_clear_bfqq_just_created(bfqq
);
1352 if (!bfq_bfqq_IO_bound(bfqq
)) {
1353 if (arrived_in_time
) {
1354 bfqq
->requests_within_timer
++;
1355 if (bfqq
->requests_within_timer
>=
1356 bfqd
->bfq_requests_within_timer
)
1357 bfq_mark_bfqq_IO_bound(bfqq
);
1359 bfqq
->requests_within_timer
= 0;
1362 if (bfqd
->low_latency
) {
1363 if (unlikely(time_is_after_jiffies(bfqq
->split_time
)))
1366 jiffies
- bfqd
->bfq_wr_min_idle_time
- 1;
1368 if (time_is_before_jiffies(bfqq
->split_time
+
1369 bfqd
->bfq_wr_min_idle_time
)) {
1370 bfq_update_bfqq_wr_on_rq_arrival(bfqd
, bfqq
,
1377 if (old_wr_coeff
!= bfqq
->wr_coeff
)
1378 bfqq
->entity
.prio_changed
= 1;
1382 bfqq
->last_idle_bklogged
= jiffies
;
1383 bfqq
->service_from_backlogged
= 0;
1384 bfq_clear_bfqq_softrt_update(bfqq
);
1386 bfq_add_bfqq_busy(bfqd
, bfqq
);
1389 * Expire in-service queue only if preemption may be needed
1390 * for guarantees. In this respect, the function
1391 * next_queue_may_preempt just checks a simple, necessary
1392 * condition, and not a sufficient condition based on
1393 * timestamps. In fact, for the latter condition to be
1394 * evaluated, timestamps would need first to be updated, and
1395 * this operation is quite costly (see the comments on the
1396 * function bfq_bfqq_update_budg_for_activation).
1398 if (bfqd
->in_service_queue
&& bfqq_wants_to_preempt
&&
1399 bfqd
->in_service_queue
->wr_coeff
< bfqq
->wr_coeff
&&
1400 next_queue_may_preempt(bfqd
))
1401 bfq_bfqq_expire(bfqd
, bfqd
->in_service_queue
,
1402 false, BFQQE_PREEMPTED
);
1405 static void bfq_add_request(struct request
*rq
)
1407 struct bfq_queue
*bfqq
= RQ_BFQQ(rq
);
1408 struct bfq_data
*bfqd
= bfqq
->bfqd
;
1409 struct request
*next_rq
, *prev
;
1410 unsigned int old_wr_coeff
= bfqq
->wr_coeff
;
1411 bool interactive
= false;
1413 bfq_log_bfqq(bfqd
, bfqq
, "add_request %d", rq_is_sync(rq
));
1414 bfqq
->queued
[rq_is_sync(rq
)]++;
1417 elv_rb_add(&bfqq
->sort_list
, rq
);
1420 * Check if this request is a better next-serve candidate.
1422 prev
= bfqq
->next_rq
;
1423 next_rq
= bfq_choose_req(bfqd
, bfqq
->next_rq
, rq
, bfqd
->last_position
);
1424 bfqq
->next_rq
= next_rq
;
1427 * Adjust priority tree position, if next_rq changes.
1429 if (prev
!= bfqq
->next_rq
)
1430 bfq_pos_tree_add_move(bfqd
, bfqq
);
1432 if (!bfq_bfqq_busy(bfqq
)) /* switching to busy ... */
1433 bfq_bfqq_handle_idle_busy_switch(bfqd
, bfqq
, old_wr_coeff
,
1436 if (bfqd
->low_latency
&& old_wr_coeff
== 1 && !rq_is_sync(rq
) &&
1437 time_is_before_jiffies(
1438 bfqq
->last_wr_start_finish
+
1439 bfqd
->bfq_wr_min_inter_arr_async
)) {
1440 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
;
1441 bfqq
->wr_cur_max_time
= bfq_wr_duration(bfqd
);
1443 bfqd
->wr_busy_queues
++;
1444 bfqq
->entity
.prio_changed
= 1;
1446 if (prev
!= bfqq
->next_rq
)
1447 bfq_updated_next_req(bfqd
, bfqq
);
1451 * Assign jiffies to last_wr_start_finish in the following
1454 * . if bfqq is not going to be weight-raised, because, for
1455 * non weight-raised queues, last_wr_start_finish stores the
1456 * arrival time of the last request; as of now, this piece
1457 * of information is used only for deciding whether to
1458 * weight-raise async queues
1460 * . if bfqq is not weight-raised, because, if bfqq is now
1461 * switching to weight-raised, then last_wr_start_finish
1462 * stores the time when weight-raising starts
1464 * . if bfqq is interactive, because, regardless of whether
1465 * bfqq is currently weight-raised, the weight-raising
1466 * period must start or restart (this case is considered
1467 * separately because it is not detected by the above
1468 * conditions, if bfqq is already weight-raised)
1470 * last_wr_start_finish has to be updated also if bfqq is soft
1471 * real-time, because the weight-raising period is constantly
1472 * restarted on idle-to-busy transitions for these queues, but
1473 * this is already done in bfq_bfqq_handle_idle_busy_switch if
1476 if (bfqd
->low_latency
&&
1477 (old_wr_coeff
== 1 || bfqq
->wr_coeff
== 1 || interactive
))
1478 bfqq
->last_wr_start_finish
= jiffies
;
1481 static struct request
*bfq_find_rq_fmerge(struct bfq_data
*bfqd
,
1483 struct request_queue
*q
)
1485 struct bfq_queue
*bfqq
= bfqd
->bio_bfqq
;
1489 return elv_rb_find(&bfqq
->sort_list
, bio_end_sector(bio
));
1494 static sector_t
get_sdist(sector_t last_pos
, struct request
*rq
)
1497 return abs(blk_rq_pos(rq
) - last_pos
);
1502 #if 0 /* Still not clear if we can do without next two functions */
1503 static void bfq_activate_request(struct request_queue
*q
, struct request
*rq
)
1505 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
1507 bfqd
->rq_in_driver
++;
1510 static void bfq_deactivate_request(struct request_queue
*q
, struct request
*rq
)
1512 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
1514 bfqd
->rq_in_driver
--;
1518 static void bfq_remove_request(struct request_queue
*q
,
1521 struct bfq_queue
*bfqq
= RQ_BFQQ(rq
);
1522 struct bfq_data
*bfqd
= bfqq
->bfqd
;
1523 const int sync
= rq_is_sync(rq
);
1525 if (bfqq
->next_rq
== rq
) {
1526 bfqq
->next_rq
= bfq_find_next_rq(bfqd
, bfqq
, rq
);
1527 bfq_updated_next_req(bfqd
, bfqq
);
1530 if (rq
->queuelist
.prev
!= &rq
->queuelist
)
1531 list_del_init(&rq
->queuelist
);
1532 bfqq
->queued
[sync
]--;
1534 elv_rb_del(&bfqq
->sort_list
, rq
);
1536 elv_rqhash_del(q
, rq
);
1537 if (q
->last_merge
== rq
)
1538 q
->last_merge
= NULL
;
1540 if (RB_EMPTY_ROOT(&bfqq
->sort_list
)) {
1541 bfqq
->next_rq
= NULL
;
1543 if (bfq_bfqq_busy(bfqq
) && bfqq
!= bfqd
->in_service_queue
) {
1544 bfq_del_bfqq_busy(bfqd
, bfqq
, false);
1546 * bfqq emptied. In normal operation, when
1547 * bfqq is empty, bfqq->entity.service and
1548 * bfqq->entity.budget must contain,
1549 * respectively, the service received and the
1550 * budget used last time bfqq emptied. These
1551 * facts do not hold in this case, as at least
1552 * this last removal occurred while bfqq is
1553 * not in service. To avoid inconsistencies,
1554 * reset both bfqq->entity.service and
1555 * bfqq->entity.budget, if bfqq has still a
1556 * process that may issue I/O requests to it.
1558 bfqq
->entity
.budget
= bfqq
->entity
.service
= 0;
1562 * Remove queue from request-position tree as it is empty.
1564 if (bfqq
->pos_root
) {
1565 rb_erase(&bfqq
->pos_node
, bfqq
->pos_root
);
1566 bfqq
->pos_root
= NULL
;
1570 if (rq
->cmd_flags
& REQ_META
)
1571 bfqq
->meta_pending
--;
1573 bfqg_stats_update_io_remove(bfqq_group(bfqq
), rq
->cmd_flags
);
1576 static bool bfq_bio_merge(struct blk_mq_hw_ctx
*hctx
, struct bio
*bio
)
1578 struct request_queue
*q
= hctx
->queue
;
1579 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
1580 struct request
*free
= NULL
;
1582 * bfq_bic_lookup grabs the queue_lock: invoke it now and
1583 * store its return value for later use, to avoid nesting
1584 * queue_lock inside the bfqd->lock. We assume that the bic
1585 * returned by bfq_bic_lookup does not go away before
1586 * bfqd->lock is taken.
1588 struct bfq_io_cq
*bic
= bfq_bic_lookup(bfqd
, current
->io_context
, q
);
1591 spin_lock_irq(&bfqd
->lock
);
1594 bfqd
->bio_bfqq
= bic_to_bfqq(bic
, op_is_sync(bio
->bi_opf
));
1596 bfqd
->bio_bfqq
= NULL
;
1597 bfqd
->bio_bic
= bic
;
1599 ret
= blk_mq_sched_try_merge(q
, bio
, &free
);
1602 blk_mq_free_request(free
);
1603 spin_unlock_irq(&bfqd
->lock
);
1608 static int bfq_request_merge(struct request_queue
*q
, struct request
**req
,
1611 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
1612 struct request
*__rq
;
1614 __rq
= bfq_find_rq_fmerge(bfqd
, bio
, q
);
1615 if (__rq
&& elv_bio_merge_ok(__rq
, bio
)) {
1617 return ELEVATOR_FRONT_MERGE
;
1620 return ELEVATOR_NO_MERGE
;
1623 static void bfq_request_merged(struct request_queue
*q
, struct request
*req
,
1624 enum elv_merge type
)
1626 if (type
== ELEVATOR_FRONT_MERGE
&&
1627 rb_prev(&req
->rb_node
) &&
1629 blk_rq_pos(container_of(rb_prev(&req
->rb_node
),
1630 struct request
, rb_node
))) {
1631 struct bfq_queue
*bfqq
= RQ_BFQQ(req
);
1632 struct bfq_data
*bfqd
= bfqq
->bfqd
;
1633 struct request
*prev
, *next_rq
;
1635 /* Reposition request in its sort_list */
1636 elv_rb_del(&bfqq
->sort_list
, req
);
1637 elv_rb_add(&bfqq
->sort_list
, req
);
1639 /* Choose next request to be served for bfqq */
1640 prev
= bfqq
->next_rq
;
1641 next_rq
= bfq_choose_req(bfqd
, bfqq
->next_rq
, req
,
1642 bfqd
->last_position
);
1643 bfqq
->next_rq
= next_rq
;
1645 * If next_rq changes, update both the queue's budget to
1646 * fit the new request and the queue's position in its
1649 if (prev
!= bfqq
->next_rq
) {
1650 bfq_updated_next_req(bfqd
, bfqq
);
1651 bfq_pos_tree_add_move(bfqd
, bfqq
);
1656 static void bfq_requests_merged(struct request_queue
*q
, struct request
*rq
,
1657 struct request
*next
)
1659 struct bfq_queue
*bfqq
= RQ_BFQQ(rq
), *next_bfqq
= RQ_BFQQ(next
);
1661 if (!RB_EMPTY_NODE(&rq
->rb_node
))
1663 spin_lock_irq(&bfqq
->bfqd
->lock
);
1666 * If next and rq belong to the same bfq_queue and next is older
1667 * than rq, then reposition rq in the fifo (by substituting next
1668 * with rq). Otherwise, if next and rq belong to different
1669 * bfq_queues, never reposition rq: in fact, we would have to
1670 * reposition it with respect to next's position in its own fifo,
1671 * which would most certainly be too expensive with respect to
1674 if (bfqq
== next_bfqq
&&
1675 !list_empty(&rq
->queuelist
) && !list_empty(&next
->queuelist
) &&
1676 next
->fifo_time
< rq
->fifo_time
) {
1677 list_del_init(&rq
->queuelist
);
1678 list_replace_init(&next
->queuelist
, &rq
->queuelist
);
1679 rq
->fifo_time
= next
->fifo_time
;
1682 if (bfqq
->next_rq
== next
)
1685 bfq_remove_request(q
, next
);
1687 spin_unlock_irq(&bfqq
->bfqd
->lock
);
1689 bfqg_stats_update_io_merged(bfqq_group(bfqq
), next
->cmd_flags
);
1692 /* Must be called with bfqq != NULL */
1693 static void bfq_bfqq_end_wr(struct bfq_queue
*bfqq
)
1695 if (bfq_bfqq_busy(bfqq
))
1696 bfqq
->bfqd
->wr_busy_queues
--;
1698 bfqq
->wr_cur_max_time
= 0;
1699 bfqq
->last_wr_start_finish
= jiffies
;
1701 * Trigger a weight change on the next invocation of
1702 * __bfq_entity_update_weight_prio.
1704 bfqq
->entity
.prio_changed
= 1;
1707 void bfq_end_wr_async_queues(struct bfq_data
*bfqd
,
1708 struct bfq_group
*bfqg
)
1712 for (i
= 0; i
< 2; i
++)
1713 for (j
= 0; j
< IOPRIO_BE_NR
; j
++)
1714 if (bfqg
->async_bfqq
[i
][j
])
1715 bfq_bfqq_end_wr(bfqg
->async_bfqq
[i
][j
]);
1716 if (bfqg
->async_idle_bfqq
)
1717 bfq_bfqq_end_wr(bfqg
->async_idle_bfqq
);
1720 static void bfq_end_wr(struct bfq_data
*bfqd
)
1722 struct bfq_queue
*bfqq
;
1724 spin_lock_irq(&bfqd
->lock
);
1726 list_for_each_entry(bfqq
, &bfqd
->active_list
, bfqq_list
)
1727 bfq_bfqq_end_wr(bfqq
);
1728 list_for_each_entry(bfqq
, &bfqd
->idle_list
, bfqq_list
)
1729 bfq_bfqq_end_wr(bfqq
);
1730 bfq_end_wr_async(bfqd
);
1732 spin_unlock_irq(&bfqd
->lock
);
1735 static sector_t
bfq_io_struct_pos(void *io_struct
, bool request
)
1738 return blk_rq_pos(io_struct
);
1740 return ((struct bio
*)io_struct
)->bi_iter
.bi_sector
;
1743 static int bfq_rq_close_to_sector(void *io_struct
, bool request
,
1746 return abs(bfq_io_struct_pos(io_struct
, request
) - sector
) <=
1750 static struct bfq_queue
*bfqq_find_close(struct bfq_data
*bfqd
,
1751 struct bfq_queue
*bfqq
,
1754 struct rb_root
*root
= &bfq_bfqq_to_bfqg(bfqq
)->rq_pos_tree
;
1755 struct rb_node
*parent
, *node
;
1756 struct bfq_queue
*__bfqq
;
1758 if (RB_EMPTY_ROOT(root
))
1762 * First, if we find a request starting at the end of the last
1763 * request, choose it.
1765 __bfqq
= bfq_rq_pos_tree_lookup(bfqd
, root
, sector
, &parent
, NULL
);
1770 * If the exact sector wasn't found, the parent of the NULL leaf
1771 * will contain the closest sector (rq_pos_tree sorted by
1772 * next_request position).
1774 __bfqq
= rb_entry(parent
, struct bfq_queue
, pos_node
);
1775 if (bfq_rq_close_to_sector(__bfqq
->next_rq
, true, sector
))
1778 if (blk_rq_pos(__bfqq
->next_rq
) < sector
)
1779 node
= rb_next(&__bfqq
->pos_node
);
1781 node
= rb_prev(&__bfqq
->pos_node
);
1785 __bfqq
= rb_entry(node
, struct bfq_queue
, pos_node
);
1786 if (bfq_rq_close_to_sector(__bfqq
->next_rq
, true, sector
))
1792 static struct bfq_queue
*bfq_find_close_cooperator(struct bfq_data
*bfqd
,
1793 struct bfq_queue
*cur_bfqq
,
1796 struct bfq_queue
*bfqq
;
1799 * We shall notice if some of the queues are cooperating,
1800 * e.g., working closely on the same area of the device. In
1801 * that case, we can group them together and: 1) don't waste
1802 * time idling, and 2) serve the union of their requests in
1803 * the best possible order for throughput.
1805 bfqq
= bfqq_find_close(bfqd
, cur_bfqq
, sector
);
1806 if (!bfqq
|| bfqq
== cur_bfqq
)
1812 static struct bfq_queue
*
1813 bfq_setup_merge(struct bfq_queue
*bfqq
, struct bfq_queue
*new_bfqq
)
1815 int process_refs
, new_process_refs
;
1816 struct bfq_queue
*__bfqq
;
1819 * If there are no process references on the new_bfqq, then it is
1820 * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain
1821 * may have dropped their last reference (not just their last process
1824 if (!bfqq_process_refs(new_bfqq
))
1827 /* Avoid a circular list and skip interim queue merges. */
1828 while ((__bfqq
= new_bfqq
->new_bfqq
)) {
1834 process_refs
= bfqq_process_refs(bfqq
);
1835 new_process_refs
= bfqq_process_refs(new_bfqq
);
1837 * If the process for the bfqq has gone away, there is no
1838 * sense in merging the queues.
1840 if (process_refs
== 0 || new_process_refs
== 0)
1843 bfq_log_bfqq(bfqq
->bfqd
, bfqq
, "scheduling merge with queue %d",
1847 * Merging is just a redirection: the requests of the process
1848 * owning one of the two queues are redirected to the other queue.
1849 * The latter queue, in its turn, is set as shared if this is the
1850 * first time that the requests of some process are redirected to
1853 * We redirect bfqq to new_bfqq and not the opposite, because
1854 * we are in the context of the process owning bfqq, thus we
1855 * have the io_cq of this process. So we can immediately
1856 * configure this io_cq to redirect the requests of the
1857 * process to new_bfqq. In contrast, the io_cq of new_bfqq is
1858 * not available any more (new_bfqq->bic == NULL).
1860 * Anyway, even in case new_bfqq coincides with the in-service
1861 * queue, redirecting requests the in-service queue is the
1862 * best option, as we feed the in-service queue with new
1863 * requests close to the last request served and, by doing so,
1864 * are likely to increase the throughput.
1866 bfqq
->new_bfqq
= new_bfqq
;
1867 new_bfqq
->ref
+= process_refs
;
1871 static bool bfq_may_be_close_cooperator(struct bfq_queue
*bfqq
,
1872 struct bfq_queue
*new_bfqq
)
1874 if (bfq_class_idle(bfqq
) || bfq_class_idle(new_bfqq
) ||
1875 (bfqq
->ioprio_class
!= new_bfqq
->ioprio_class
))
1879 * If either of the queues has already been detected as seeky,
1880 * then merging it with the other queue is unlikely to lead to
1883 if (BFQQ_SEEKY(bfqq
) || BFQQ_SEEKY(new_bfqq
))
1887 * Interleaved I/O is known to be done by (some) applications
1888 * only for reads, so it does not make sense to merge async
1891 if (!bfq_bfqq_sync(bfqq
) || !bfq_bfqq_sync(new_bfqq
))
1898 * If this function returns true, then bfqq cannot be merged. The idea
1899 * is that true cooperation happens very early after processes start
1900 * to do I/O. Usually, late cooperations are just accidental false
1901 * positives. In case bfqq is weight-raised, such false positives
1902 * would evidently degrade latency guarantees for bfqq.
1904 static bool wr_from_too_long(struct bfq_queue
*bfqq
)
1906 return bfqq
->wr_coeff
> 1 &&
1907 time_is_before_jiffies(bfqq
->last_wr_start_finish
+
1908 msecs_to_jiffies(100));
1912 * Attempt to schedule a merge of bfqq with the currently in-service
1913 * queue or with a close queue among the scheduled queues. Return
1914 * NULL if no merge was scheduled, a pointer to the shared bfq_queue
1915 * structure otherwise.
1917 * The OOM queue is not allowed to participate to cooperation: in fact, since
1918 * the requests temporarily redirected to the OOM queue could be redirected
1919 * again to dedicated queues at any time, the state needed to correctly
1920 * handle merging with the OOM queue would be quite complex and expensive
1921 * to maintain. Besides, in such a critical condition as an out of memory,
1922 * the benefits of queue merging may be little relevant, or even negligible.
1924 * Weight-raised queues can be merged only if their weight-raising
1925 * period has just started. In fact cooperating processes are usually
1926 * started together. Thus, with this filter we avoid false positives
1927 * that would jeopardize low-latency guarantees.
1929 * WARNING: queue merging may impair fairness among non-weight raised
1930 * queues, for at least two reasons: 1) the original weight of a
1931 * merged queue may change during the merged state, 2) even being the
1932 * weight the same, a merged queue may be bloated with many more
1933 * requests than the ones produced by its originally-associated
1936 static struct bfq_queue
*
1937 bfq_setup_cooperator(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
1938 void *io_struct
, bool request
)
1940 struct bfq_queue
*in_service_bfqq
, *new_bfqq
;
1943 return bfqq
->new_bfqq
;
1946 wr_from_too_long(bfqq
) ||
1947 unlikely(bfqq
== &bfqd
->oom_bfqq
))
1950 /* If there is only one backlogged queue, don't search. */
1951 if (bfqd
->busy_queues
== 1)
1954 in_service_bfqq
= bfqd
->in_service_queue
;
1956 if (!in_service_bfqq
|| in_service_bfqq
== bfqq
1957 || wr_from_too_long(in_service_bfqq
) ||
1958 unlikely(in_service_bfqq
== &bfqd
->oom_bfqq
))
1959 goto check_scheduled
;
1961 if (bfq_rq_close_to_sector(io_struct
, request
, bfqd
->last_position
) &&
1962 bfqq
->entity
.parent
== in_service_bfqq
->entity
.parent
&&
1963 bfq_may_be_close_cooperator(bfqq
, in_service_bfqq
)) {
1964 new_bfqq
= bfq_setup_merge(bfqq
, in_service_bfqq
);
1969 * Check whether there is a cooperator among currently scheduled
1970 * queues. The only thing we need is that the bio/request is not
1971 * NULL, as we need it to establish whether a cooperator exists.
1974 new_bfqq
= bfq_find_close_cooperator(bfqd
, bfqq
,
1975 bfq_io_struct_pos(io_struct
, request
));
1977 if (new_bfqq
&& !wr_from_too_long(new_bfqq
) &&
1978 likely(new_bfqq
!= &bfqd
->oom_bfqq
) &&
1979 bfq_may_be_close_cooperator(bfqq
, new_bfqq
))
1980 return bfq_setup_merge(bfqq
, new_bfqq
);
1985 static void bfq_bfqq_save_state(struct bfq_queue
*bfqq
)
1987 struct bfq_io_cq
*bic
= bfqq
->bic
;
1990 * If !bfqq->bic, the queue is already shared or its requests
1991 * have already been redirected to a shared queue; both idle window
1992 * and weight raising state have already been saved. Do nothing.
1997 bic
->saved_ttime
= bfqq
->ttime
;
1998 bic
->saved_idle_window
= bfq_bfqq_idle_window(bfqq
);
1999 bic
->saved_IO_bound
= bfq_bfqq_IO_bound(bfqq
);
2000 bic
->saved_in_large_burst
= bfq_bfqq_in_large_burst(bfqq
);
2001 bic
->was_in_burst_list
= !hlist_unhashed(&bfqq
->burst_list_node
);
2002 bic
->saved_wr_coeff
= bfqq
->wr_coeff
;
2003 bic
->saved_wr_start_at_switch_to_srt
= bfqq
->wr_start_at_switch_to_srt
;
2004 bic
->saved_last_wr_start_finish
= bfqq
->last_wr_start_finish
;
2005 bic
->saved_wr_cur_max_time
= bfqq
->wr_cur_max_time
;
2009 bfq_merge_bfqqs(struct bfq_data
*bfqd
, struct bfq_io_cq
*bic
,
2010 struct bfq_queue
*bfqq
, struct bfq_queue
*new_bfqq
)
2012 bfq_log_bfqq(bfqd
, bfqq
, "merging with queue %lu",
2013 (unsigned long)new_bfqq
->pid
);
2014 /* Save weight raising and idle window of the merged queues */
2015 bfq_bfqq_save_state(bfqq
);
2016 bfq_bfqq_save_state(new_bfqq
);
2017 if (bfq_bfqq_IO_bound(bfqq
))
2018 bfq_mark_bfqq_IO_bound(new_bfqq
);
2019 bfq_clear_bfqq_IO_bound(bfqq
);
2022 * If bfqq is weight-raised, then let new_bfqq inherit
2023 * weight-raising. To reduce false positives, neglect the case
2024 * where bfqq has just been created, but has not yet made it
2025 * to be weight-raised (which may happen because EQM may merge
2026 * bfqq even before bfq_add_request is executed for the first
2027 * time for bfqq). Handling this case would however be very
2028 * easy, thanks to the flag just_created.
2030 if (new_bfqq
->wr_coeff
== 1 && bfqq
->wr_coeff
> 1) {
2031 new_bfqq
->wr_coeff
= bfqq
->wr_coeff
;
2032 new_bfqq
->wr_cur_max_time
= bfqq
->wr_cur_max_time
;
2033 new_bfqq
->last_wr_start_finish
= bfqq
->last_wr_start_finish
;
2034 new_bfqq
->wr_start_at_switch_to_srt
=
2035 bfqq
->wr_start_at_switch_to_srt
;
2036 if (bfq_bfqq_busy(new_bfqq
))
2037 bfqd
->wr_busy_queues
++;
2038 new_bfqq
->entity
.prio_changed
= 1;
2041 if (bfqq
->wr_coeff
> 1) { /* bfqq has given its wr to new_bfqq */
2043 bfqq
->entity
.prio_changed
= 1;
2044 if (bfq_bfqq_busy(bfqq
))
2045 bfqd
->wr_busy_queues
--;
2048 bfq_log_bfqq(bfqd
, new_bfqq
, "merge_bfqqs: wr_busy %d",
2049 bfqd
->wr_busy_queues
);
2052 * Merge queues (that is, let bic redirect its requests to new_bfqq)
2054 bic_set_bfqq(bic
, new_bfqq
, 1);
2055 bfq_mark_bfqq_coop(new_bfqq
);
2057 * new_bfqq now belongs to at least two bics (it is a shared queue):
2058 * set new_bfqq->bic to NULL. bfqq either:
2059 * - does not belong to any bic any more, and hence bfqq->bic must
2060 * be set to NULL, or
2061 * - is a queue whose owning bics have already been redirected to a
2062 * different queue, hence the queue is destined to not belong to
2063 * any bic soon and bfqq->bic is already NULL (therefore the next
2064 * assignment causes no harm).
2066 new_bfqq
->bic
= NULL
;
2068 /* release process reference to bfqq */
2069 bfq_put_queue(bfqq
);
2072 static bool bfq_allow_bio_merge(struct request_queue
*q
, struct request
*rq
,
2075 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
2076 bool is_sync
= op_is_sync(bio
->bi_opf
);
2077 struct bfq_queue
*bfqq
= bfqd
->bio_bfqq
, *new_bfqq
;
2080 * Disallow merge of a sync bio into an async request.
2082 if (is_sync
&& !rq_is_sync(rq
))
2086 * Lookup the bfqq that this bio will be queued with. Allow
2087 * merge only if rq is queued there.
2093 * We take advantage of this function to perform an early merge
2094 * of the queues of possible cooperating processes.
2096 new_bfqq
= bfq_setup_cooperator(bfqd
, bfqq
, bio
, false);
2099 * bic still points to bfqq, then it has not yet been
2100 * redirected to some other bfq_queue, and a queue
2101 * merge beween bfqq and new_bfqq can be safely
2102 * fulfillled, i.e., bic can be redirected to new_bfqq
2103 * and bfqq can be put.
2105 bfq_merge_bfqqs(bfqd
, bfqd
->bio_bic
, bfqq
,
2108 * If we get here, bio will be queued into new_queue,
2109 * so use new_bfqq to decide whether bio and rq can be
2115 * Change also bqfd->bio_bfqq, as
2116 * bfqd->bio_bic now points to new_bfqq, and
2117 * this function may be invoked again (and then may
2118 * use again bqfd->bio_bfqq).
2120 bfqd
->bio_bfqq
= bfqq
;
2123 return bfqq
== RQ_BFQQ(rq
);
2127 * Set the maximum time for the in-service queue to consume its
2128 * budget. This prevents seeky processes from lowering the throughput.
2129 * In practice, a time-slice service scheme is used with seeky
2132 static void bfq_set_budget_timeout(struct bfq_data
*bfqd
,
2133 struct bfq_queue
*bfqq
)
2135 unsigned int timeout_coeff
;
2137 if (bfqq
->wr_cur_max_time
== bfqd
->bfq_wr_rt_max_time
)
2140 timeout_coeff
= bfqq
->entity
.weight
/ bfqq
->entity
.orig_weight
;
2142 bfqd
->last_budget_start
= ktime_get();
2144 bfqq
->budget_timeout
= jiffies
+
2145 bfqd
->bfq_timeout
* timeout_coeff
;
2148 static void __bfq_set_in_service_queue(struct bfq_data
*bfqd
,
2149 struct bfq_queue
*bfqq
)
2152 bfqg_stats_update_avg_queue_size(bfqq_group(bfqq
));
2153 bfq_clear_bfqq_fifo_expire(bfqq
);
2155 bfqd
->budgets_assigned
= (bfqd
->budgets_assigned
* 7 + 256) / 8;
2157 if (time_is_before_jiffies(bfqq
->last_wr_start_finish
) &&
2158 bfqq
->wr_coeff
> 1 &&
2159 bfqq
->wr_cur_max_time
== bfqd
->bfq_wr_rt_max_time
&&
2160 time_is_before_jiffies(bfqq
->budget_timeout
)) {
2162 * For soft real-time queues, move the start
2163 * of the weight-raising period forward by the
2164 * time the queue has not received any
2165 * service. Otherwise, a relatively long
2166 * service delay is likely to cause the
2167 * weight-raising period of the queue to end,
2168 * because of the short duration of the
2169 * weight-raising period of a soft real-time
2170 * queue. It is worth noting that this move
2171 * is not so dangerous for the other queues,
2172 * because soft real-time queues are not
2175 * To not add a further variable, we use the
2176 * overloaded field budget_timeout to
2177 * determine for how long the queue has not
2178 * received service, i.e., how much time has
2179 * elapsed since the queue expired. However,
2180 * this is a little imprecise, because
2181 * budget_timeout is set to jiffies if bfqq
2182 * not only expires, but also remains with no
2185 if (time_after(bfqq
->budget_timeout
,
2186 bfqq
->last_wr_start_finish
))
2187 bfqq
->last_wr_start_finish
+=
2188 jiffies
- bfqq
->budget_timeout
;
2190 bfqq
->last_wr_start_finish
= jiffies
;
2193 bfq_set_budget_timeout(bfqd
, bfqq
);
2194 bfq_log_bfqq(bfqd
, bfqq
,
2195 "set_in_service_queue, cur-budget = %d",
2196 bfqq
->entity
.budget
);
2199 bfqd
->in_service_queue
= bfqq
;
2203 * Get and set a new queue for service.
2205 static struct bfq_queue
*bfq_set_in_service_queue(struct bfq_data
*bfqd
)
2207 struct bfq_queue
*bfqq
= bfq_get_next_queue(bfqd
);
2209 __bfq_set_in_service_queue(bfqd
, bfqq
);
2213 static void bfq_arm_slice_timer(struct bfq_data
*bfqd
)
2215 struct bfq_queue
*bfqq
= bfqd
->in_service_queue
;
2218 bfq_mark_bfqq_wait_request(bfqq
);
2221 * We don't want to idle for seeks, but we do want to allow
2222 * fair distribution of slice time for a process doing back-to-back
2223 * seeks. So allow a little bit of time for him to submit a new rq.
2225 sl
= bfqd
->bfq_slice_idle
;
2227 * Unless the queue is being weight-raised or the scenario is
2228 * asymmetric, grant only minimum idle time if the queue
2229 * is seeky. A long idling is preserved for a weight-raised
2230 * queue, or, more in general, in an asymmetric scenario,
2231 * because a long idling is needed for guaranteeing to a queue
2232 * its reserved share of the throughput (in particular, it is
2233 * needed if the queue has a higher weight than some other
2236 if (BFQQ_SEEKY(bfqq
) && bfqq
->wr_coeff
== 1 &&
2237 bfq_symmetric_scenario(bfqd
))
2238 sl
= min_t(u64
, sl
, BFQ_MIN_TT
);
2240 bfqd
->last_idling_start
= ktime_get();
2241 hrtimer_start(&bfqd
->idle_slice_timer
, ns_to_ktime(sl
),
2243 bfqg_stats_set_start_idle_time(bfqq_group(bfqq
));
2247 * In autotuning mode, max_budget is dynamically recomputed as the
2248 * amount of sectors transferred in timeout at the estimated peak
2249 * rate. This enables BFQ to utilize a full timeslice with a full
2250 * budget, even if the in-service queue is served at peak rate. And
2251 * this maximises throughput with sequential workloads.
2253 static unsigned long bfq_calc_max_budget(struct bfq_data
*bfqd
)
2255 return (u64
)bfqd
->peak_rate
* USEC_PER_MSEC
*
2256 jiffies_to_msecs(bfqd
->bfq_timeout
)>>BFQ_RATE_SHIFT
;
2260 * Update parameters related to throughput and responsiveness, as a
2261 * function of the estimated peak rate. See comments on
2262 * bfq_calc_max_budget(), and on T_slow and T_fast arrays.
2264 static void update_thr_responsiveness_params(struct bfq_data
*bfqd
)
2266 int dev_type
= blk_queue_nonrot(bfqd
->queue
);
2268 if (bfqd
->bfq_user_max_budget
== 0)
2269 bfqd
->bfq_max_budget
=
2270 bfq_calc_max_budget(bfqd
);
2272 if (bfqd
->device_speed
== BFQ_BFQD_FAST
&&
2273 bfqd
->peak_rate
< device_speed_thresh
[dev_type
]) {
2274 bfqd
->device_speed
= BFQ_BFQD_SLOW
;
2275 bfqd
->RT_prod
= R_slow
[dev_type
] *
2277 } else if (bfqd
->device_speed
== BFQ_BFQD_SLOW
&&
2278 bfqd
->peak_rate
> device_speed_thresh
[dev_type
]) {
2279 bfqd
->device_speed
= BFQ_BFQD_FAST
;
2280 bfqd
->RT_prod
= R_fast
[dev_type
] *
2285 "dev_type %s dev_speed_class = %s (%llu sects/sec), thresh %llu setcs/sec",
2286 dev_type
== 0 ? "ROT" : "NONROT",
2287 bfqd
->device_speed
== BFQ_BFQD_FAST
? "FAST" : "SLOW",
2288 bfqd
->device_speed
== BFQ_BFQD_FAST
?
2289 (USEC_PER_SEC
*(u64
)R_fast
[dev_type
])>>BFQ_RATE_SHIFT
:
2290 (USEC_PER_SEC
*(u64
)R_slow
[dev_type
])>>BFQ_RATE_SHIFT
,
2291 (USEC_PER_SEC
*(u64
)device_speed_thresh
[dev_type
])>>
2295 static void bfq_reset_rate_computation(struct bfq_data
*bfqd
,
2298 if (rq
!= NULL
) { /* new rq dispatch now, reset accordingly */
2299 bfqd
->last_dispatch
= bfqd
->first_dispatch
= ktime_get_ns();
2300 bfqd
->peak_rate_samples
= 1;
2301 bfqd
->sequential_samples
= 0;
2302 bfqd
->tot_sectors_dispatched
= bfqd
->last_rq_max_size
=
2304 } else /* no new rq dispatched, just reset the number of samples */
2305 bfqd
->peak_rate_samples
= 0; /* full re-init on next disp. */
2308 "reset_rate_computation at end, sample %u/%u tot_sects %llu",
2309 bfqd
->peak_rate_samples
, bfqd
->sequential_samples
,
2310 bfqd
->tot_sectors_dispatched
);
2313 static void bfq_update_rate_reset(struct bfq_data
*bfqd
, struct request
*rq
)
2315 u32 rate
, weight
, divisor
;
2318 * For the convergence property to hold (see comments on
2319 * bfq_update_peak_rate()) and for the assessment to be
2320 * reliable, a minimum number of samples must be present, and
2321 * a minimum amount of time must have elapsed. If not so, do
2322 * not compute new rate. Just reset parameters, to get ready
2323 * for a new evaluation attempt.
2325 if (bfqd
->peak_rate_samples
< BFQ_RATE_MIN_SAMPLES
||
2326 bfqd
->delta_from_first
< BFQ_RATE_MIN_INTERVAL
)
2327 goto reset_computation
;
2330 * If a new request completion has occurred after last
2331 * dispatch, then, to approximate the rate at which requests
2332 * have been served by the device, it is more precise to
2333 * extend the observation interval to the last completion.
2335 bfqd
->delta_from_first
=
2336 max_t(u64
, bfqd
->delta_from_first
,
2337 bfqd
->last_completion
- bfqd
->first_dispatch
);
2340 * Rate computed in sects/usec, and not sects/nsec, for
2343 rate
= div64_ul(bfqd
->tot_sectors_dispatched
<<BFQ_RATE_SHIFT
,
2344 div_u64(bfqd
->delta_from_first
, NSEC_PER_USEC
));
2347 * Peak rate not updated if:
2348 * - the percentage of sequential dispatches is below 3/4 of the
2349 * total, and rate is below the current estimated peak rate
2350 * - rate is unreasonably high (> 20M sectors/sec)
2352 if ((bfqd
->sequential_samples
< (3 * bfqd
->peak_rate_samples
)>>2 &&
2353 rate
<= bfqd
->peak_rate
) ||
2354 rate
> 20<<BFQ_RATE_SHIFT
)
2355 goto reset_computation
;
2358 * We have to update the peak rate, at last! To this purpose,
2359 * we use a low-pass filter. We compute the smoothing constant
2360 * of the filter as a function of the 'weight' of the new
2363 * As can be seen in next formulas, we define this weight as a
2364 * quantity proportional to how sequential the workload is,
2365 * and to how long the observation time interval is.
2367 * The weight runs from 0 to 8. The maximum value of the
2368 * weight, 8, yields the minimum value for the smoothing
2369 * constant. At this minimum value for the smoothing constant,
2370 * the measured rate contributes for half of the next value of
2371 * the estimated peak rate.
2373 * So, the first step is to compute the weight as a function
2374 * of how sequential the workload is. Note that the weight
2375 * cannot reach 9, because bfqd->sequential_samples cannot
2376 * become equal to bfqd->peak_rate_samples, which, in its
2377 * turn, holds true because bfqd->sequential_samples is not
2378 * incremented for the first sample.
2380 weight
= (9 * bfqd
->sequential_samples
) / bfqd
->peak_rate_samples
;
2383 * Second step: further refine the weight as a function of the
2384 * duration of the observation interval.
2386 weight
= min_t(u32
, 8,
2387 div_u64(weight
* bfqd
->delta_from_first
,
2388 BFQ_RATE_REF_INTERVAL
));
2391 * Divisor ranging from 10, for minimum weight, to 2, for
2394 divisor
= 10 - weight
;
2397 * Finally, update peak rate:
2399 * peak_rate = peak_rate * (divisor-1) / divisor + rate / divisor
2401 bfqd
->peak_rate
*= divisor
-1;
2402 bfqd
->peak_rate
/= divisor
;
2403 rate
/= divisor
; /* smoothing constant alpha = 1/divisor */
2405 bfqd
->peak_rate
+= rate
;
2406 update_thr_responsiveness_params(bfqd
);
2409 bfq_reset_rate_computation(bfqd
, rq
);
2413 * Update the read/write peak rate (the main quantity used for
2414 * auto-tuning, see update_thr_responsiveness_params()).
2416 * It is not trivial to estimate the peak rate (correctly): because of
2417 * the presence of sw and hw queues between the scheduler and the
2418 * device components that finally serve I/O requests, it is hard to
2419 * say exactly when a given dispatched request is served inside the
2420 * device, and for how long. As a consequence, it is hard to know
2421 * precisely at what rate a given set of requests is actually served
2424 * On the opposite end, the dispatch time of any request is trivially
2425 * available, and, from this piece of information, the "dispatch rate"
2426 * of requests can be immediately computed. So, the idea in the next
2427 * function is to use what is known, namely request dispatch times
2428 * (plus, when useful, request completion times), to estimate what is
2429 * unknown, namely in-device request service rate.
2431 * The main issue is that, because of the above facts, the rate at
2432 * which a certain set of requests is dispatched over a certain time
2433 * interval can vary greatly with respect to the rate at which the
2434 * same requests are then served. But, since the size of any
2435 * intermediate queue is limited, and the service scheme is lossless
2436 * (no request is silently dropped), the following obvious convergence
2437 * property holds: the number of requests dispatched MUST become
2438 * closer and closer to the number of requests completed as the
2439 * observation interval grows. This is the key property used in
2440 * the next function to estimate the peak service rate as a function
2441 * of the observed dispatch rate. The function assumes to be invoked
2442 * on every request dispatch.
2444 static void bfq_update_peak_rate(struct bfq_data
*bfqd
, struct request
*rq
)
2446 u64 now_ns
= ktime_get_ns();
2448 if (bfqd
->peak_rate_samples
== 0) { /* first dispatch */
2449 bfq_log(bfqd
, "update_peak_rate: goto reset, samples %d",
2450 bfqd
->peak_rate_samples
);
2451 bfq_reset_rate_computation(bfqd
, rq
);
2452 goto update_last_values
; /* will add one sample */
2456 * Device idle for very long: the observation interval lasting
2457 * up to this dispatch cannot be a valid observation interval
2458 * for computing a new peak rate (similarly to the late-
2459 * completion event in bfq_completed_request()). Go to
2460 * update_rate_and_reset to have the following three steps
2462 * - close the observation interval at the last (previous)
2463 * request dispatch or completion
2464 * - compute rate, if possible, for that observation interval
2465 * - start a new observation interval with this dispatch
2467 if (now_ns
- bfqd
->last_dispatch
> 100*NSEC_PER_MSEC
&&
2468 bfqd
->rq_in_driver
== 0)
2469 goto update_rate_and_reset
;
2471 /* Update sampling information */
2472 bfqd
->peak_rate_samples
++;
2474 if ((bfqd
->rq_in_driver
> 0 ||
2475 now_ns
- bfqd
->last_completion
< BFQ_MIN_TT
)
2476 && get_sdist(bfqd
->last_position
, rq
) < BFQQ_SEEK_THR
)
2477 bfqd
->sequential_samples
++;
2479 bfqd
->tot_sectors_dispatched
+= blk_rq_sectors(rq
);
2481 /* Reset max observed rq size every 32 dispatches */
2482 if (likely(bfqd
->peak_rate_samples
% 32))
2483 bfqd
->last_rq_max_size
=
2484 max_t(u32
, blk_rq_sectors(rq
), bfqd
->last_rq_max_size
);
2486 bfqd
->last_rq_max_size
= blk_rq_sectors(rq
);
2488 bfqd
->delta_from_first
= now_ns
- bfqd
->first_dispatch
;
2490 /* Target observation interval not yet reached, go on sampling */
2491 if (bfqd
->delta_from_first
< BFQ_RATE_REF_INTERVAL
)
2492 goto update_last_values
;
2494 update_rate_and_reset
:
2495 bfq_update_rate_reset(bfqd
, rq
);
2497 bfqd
->last_position
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
2498 bfqd
->last_dispatch
= now_ns
;
2502 * Remove request from internal lists.
2504 static void bfq_dispatch_remove(struct request_queue
*q
, struct request
*rq
)
2506 struct bfq_queue
*bfqq
= RQ_BFQQ(rq
);
2509 * For consistency, the next instruction should have been
2510 * executed after removing the request from the queue and
2511 * dispatching it. We execute instead this instruction before
2512 * bfq_remove_request() (and hence introduce a temporary
2513 * inconsistency), for efficiency. In fact, should this
2514 * dispatch occur for a non in-service bfqq, this anticipated
2515 * increment prevents two counters related to bfqq->dispatched
2516 * from risking to be, first, uselessly decremented, and then
2517 * incremented again when the (new) value of bfqq->dispatched
2518 * happens to be taken into account.
2521 bfq_update_peak_rate(q
->elevator
->elevator_data
, rq
);
2523 bfq_remove_request(q
, rq
);
2526 static void __bfq_bfqq_expire(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
2529 * If this bfqq is shared between multiple processes, check
2530 * to make sure that those processes are still issuing I/Os
2531 * within the mean seek distance. If not, it may be time to
2532 * break the queues apart again.
2534 if (bfq_bfqq_coop(bfqq
) && BFQQ_SEEKY(bfqq
))
2535 bfq_mark_bfqq_split_coop(bfqq
);
2537 if (RB_EMPTY_ROOT(&bfqq
->sort_list
)) {
2538 if (bfqq
->dispatched
== 0)
2540 * Overloading budget_timeout field to store
2541 * the time at which the queue remains with no
2542 * backlog and no outstanding request; used by
2543 * the weight-raising mechanism.
2545 bfqq
->budget_timeout
= jiffies
;
2547 bfq_del_bfqq_busy(bfqd
, bfqq
, true);
2549 bfq_requeue_bfqq(bfqd
, bfqq
);
2551 * Resort priority tree of potential close cooperators.
2553 bfq_pos_tree_add_move(bfqd
, bfqq
);
2557 * All in-service entities must have been properly deactivated
2558 * or requeued before executing the next function, which
2559 * resets all in-service entites as no more in service.
2561 __bfq_bfqd_reset_in_service(bfqd
);
2565 * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior.
2566 * @bfqd: device data.
2567 * @bfqq: queue to update.
2568 * @reason: reason for expiration.
2570 * Handle the feedback on @bfqq budget at queue expiration.
2571 * See the body for detailed comments.
2573 static void __bfq_bfqq_recalc_budget(struct bfq_data
*bfqd
,
2574 struct bfq_queue
*bfqq
,
2575 enum bfqq_expiration reason
)
2577 struct request
*next_rq
;
2578 int budget
, min_budget
;
2580 min_budget
= bfq_min_budget(bfqd
);
2582 if (bfqq
->wr_coeff
== 1)
2583 budget
= bfqq
->max_budget
;
2585 * Use a constant, low budget for weight-raised queues,
2586 * to help achieve a low latency. Keep it slightly higher
2587 * than the minimum possible budget, to cause a little
2588 * bit fewer expirations.
2590 budget
= 2 * min_budget
;
2592 bfq_log_bfqq(bfqd
, bfqq
, "recalc_budg: last budg %d, budg left %d",
2593 bfqq
->entity
.budget
, bfq_bfqq_budget_left(bfqq
));
2594 bfq_log_bfqq(bfqd
, bfqq
, "recalc_budg: last max_budg %d, min budg %d",
2595 budget
, bfq_min_budget(bfqd
));
2596 bfq_log_bfqq(bfqd
, bfqq
, "recalc_budg: sync %d, seeky %d",
2597 bfq_bfqq_sync(bfqq
), BFQQ_SEEKY(bfqd
->in_service_queue
));
2599 if (bfq_bfqq_sync(bfqq
) && bfqq
->wr_coeff
== 1) {
2602 * Caveat: in all the following cases we trade latency
2605 case BFQQE_TOO_IDLE
:
2607 * This is the only case where we may reduce
2608 * the budget: if there is no request of the
2609 * process still waiting for completion, then
2610 * we assume (tentatively) that the timer has
2611 * expired because the batch of requests of
2612 * the process could have been served with a
2613 * smaller budget. Hence, betting that
2614 * process will behave in the same way when it
2615 * becomes backlogged again, we reduce its
2616 * next budget. As long as we guess right,
2617 * this budget cut reduces the latency
2618 * experienced by the process.
2620 * However, if there are still outstanding
2621 * requests, then the process may have not yet
2622 * issued its next request just because it is
2623 * still waiting for the completion of some of
2624 * the still outstanding ones. So in this
2625 * subcase we do not reduce its budget, on the
2626 * contrary we increase it to possibly boost
2627 * the throughput, as discussed in the
2628 * comments to the BUDGET_TIMEOUT case.
2630 if (bfqq
->dispatched
> 0) /* still outstanding reqs */
2631 budget
= min(budget
* 2, bfqd
->bfq_max_budget
);
2633 if (budget
> 5 * min_budget
)
2634 budget
-= 4 * min_budget
;
2636 budget
= min_budget
;
2639 case BFQQE_BUDGET_TIMEOUT
:
2641 * We double the budget here because it gives
2642 * the chance to boost the throughput if this
2643 * is not a seeky process (and has bumped into
2644 * this timeout because of, e.g., ZBR).
2646 budget
= min(budget
* 2, bfqd
->bfq_max_budget
);
2648 case BFQQE_BUDGET_EXHAUSTED
:
2650 * The process still has backlog, and did not
2651 * let either the budget timeout or the disk
2652 * idling timeout expire. Hence it is not
2653 * seeky, has a short thinktime and may be
2654 * happy with a higher budget too. So
2655 * definitely increase the budget of this good
2656 * candidate to boost the disk throughput.
2658 budget
= min(budget
* 4, bfqd
->bfq_max_budget
);
2660 case BFQQE_NO_MORE_REQUESTS
:
2662 * For queues that expire for this reason, it
2663 * is particularly important to keep the
2664 * budget close to the actual service they
2665 * need. Doing so reduces the timestamp
2666 * misalignment problem described in the
2667 * comments in the body of
2668 * __bfq_activate_entity. In fact, suppose
2669 * that a queue systematically expires for
2670 * BFQQE_NO_MORE_REQUESTS and presents a
2671 * new request in time to enjoy timestamp
2672 * back-shifting. The larger the budget of the
2673 * queue is with respect to the service the
2674 * queue actually requests in each service
2675 * slot, the more times the queue can be
2676 * reactivated with the same virtual finish
2677 * time. It follows that, even if this finish
2678 * time is pushed to the system virtual time
2679 * to reduce the consequent timestamp
2680 * misalignment, the queue unjustly enjoys for
2681 * many re-activations a lower finish time
2682 * than all newly activated queues.
2684 * The service needed by bfqq is measured
2685 * quite precisely by bfqq->entity.service.
2686 * Since bfqq does not enjoy device idling,
2687 * bfqq->entity.service is equal to the number
2688 * of sectors that the process associated with
2689 * bfqq requested to read/write before waiting
2690 * for request completions, or blocking for
2693 budget
= max_t(int, bfqq
->entity
.service
, min_budget
);
2698 } else if (!bfq_bfqq_sync(bfqq
)) {
2700 * Async queues get always the maximum possible
2701 * budget, as for them we do not care about latency
2702 * (in addition, their ability to dispatch is limited
2703 * by the charging factor).
2705 budget
= bfqd
->bfq_max_budget
;
2708 bfqq
->max_budget
= budget
;
2710 if (bfqd
->budgets_assigned
>= bfq_stats_min_budgets
&&
2711 !bfqd
->bfq_user_max_budget
)
2712 bfqq
->max_budget
= min(bfqq
->max_budget
, bfqd
->bfq_max_budget
);
2715 * If there is still backlog, then assign a new budget, making
2716 * sure that it is large enough for the next request. Since
2717 * the finish time of bfqq must be kept in sync with the
2718 * budget, be sure to call __bfq_bfqq_expire() *after* this
2721 * If there is no backlog, then no need to update the budget;
2722 * it will be updated on the arrival of a new request.
2724 next_rq
= bfqq
->next_rq
;
2726 bfqq
->entity
.budget
= max_t(unsigned long, bfqq
->max_budget
,
2727 bfq_serv_to_charge(next_rq
, bfqq
));
2729 bfq_log_bfqq(bfqd
, bfqq
, "head sect: %u, new budget %d",
2730 next_rq
? blk_rq_sectors(next_rq
) : 0,
2731 bfqq
->entity
.budget
);
2735 * Return true if the process associated with bfqq is "slow". The slow
2736 * flag is used, in addition to the budget timeout, to reduce the
2737 * amount of service provided to seeky processes, and thus reduce
2738 * their chances to lower the throughput. More details in the comments
2739 * on the function bfq_bfqq_expire().
2741 * An important observation is in order: as discussed in the comments
2742 * on the function bfq_update_peak_rate(), with devices with internal
2743 * queues, it is hard if ever possible to know when and for how long
2744 * an I/O request is processed by the device (apart from the trivial
2745 * I/O pattern where a new request is dispatched only after the
2746 * previous one has been completed). This makes it hard to evaluate
2747 * the real rate at which the I/O requests of each bfq_queue are
2748 * served. In fact, for an I/O scheduler like BFQ, serving a
2749 * bfq_queue means just dispatching its requests during its service
2750 * slot (i.e., until the budget of the queue is exhausted, or the
2751 * queue remains idle, or, finally, a timeout fires). But, during the
2752 * service slot of a bfq_queue, around 100 ms at most, the device may
2753 * be even still processing requests of bfq_queues served in previous
2754 * service slots. On the opposite end, the requests of the in-service
2755 * bfq_queue may be completed after the service slot of the queue
2758 * Anyway, unless more sophisticated solutions are used
2759 * (where possible), the sum of the sizes of the requests dispatched
2760 * during the service slot of a bfq_queue is probably the only
2761 * approximation available for the service received by the bfq_queue
2762 * during its service slot. And this sum is the quantity used in this
2763 * function to evaluate the I/O speed of a process.
2765 static bool bfq_bfqq_is_slow(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
2766 bool compensate
, enum bfqq_expiration reason
,
2767 unsigned long *delta_ms
)
2769 ktime_t delta_ktime
;
2771 bool slow
= BFQQ_SEEKY(bfqq
); /* if delta too short, use seekyness */
2773 if (!bfq_bfqq_sync(bfqq
))
2777 delta_ktime
= bfqd
->last_idling_start
;
2779 delta_ktime
= ktime_get();
2780 delta_ktime
= ktime_sub(delta_ktime
, bfqd
->last_budget_start
);
2781 delta_usecs
= ktime_to_us(delta_ktime
);
2783 /* don't use too short time intervals */
2784 if (delta_usecs
< 1000) {
2785 if (blk_queue_nonrot(bfqd
->queue
))
2787 * give same worst-case guarantees as idling
2790 *delta_ms
= BFQ_MIN_TT
/ NSEC_PER_MSEC
;
2791 else /* charge at least one seek */
2792 *delta_ms
= bfq_slice_idle
/ NSEC_PER_MSEC
;
2797 *delta_ms
= delta_usecs
/ USEC_PER_MSEC
;
2800 * Use only long (> 20ms) intervals to filter out excessive
2801 * spikes in service rate estimation.
2803 if (delta_usecs
> 20000) {
2805 * Caveat for rotational devices: processes doing I/O
2806 * in the slower disk zones tend to be slow(er) even
2807 * if not seeky. In this respect, the estimated peak
2808 * rate is likely to be an average over the disk
2809 * surface. Accordingly, to not be too harsh with
2810 * unlucky processes, a process is deemed slow only if
2811 * its rate has been lower than half of the estimated
2814 slow
= bfqq
->entity
.service
< bfqd
->bfq_max_budget
/ 2;
2817 bfq_log_bfqq(bfqd
, bfqq
, "bfq_bfqq_is_slow: slow %d", slow
);
2823 * To be deemed as soft real-time, an application must meet two
2824 * requirements. First, the application must not require an average
2825 * bandwidth higher than the approximate bandwidth required to playback or
2826 * record a compressed high-definition video.
2827 * The next function is invoked on the completion of the last request of a
2828 * batch, to compute the next-start time instant, soft_rt_next_start, such
2829 * that, if the next request of the application does not arrive before
2830 * soft_rt_next_start, then the above requirement on the bandwidth is met.
2832 * The second requirement is that the request pattern of the application is
2833 * isochronous, i.e., that, after issuing a request or a batch of requests,
2834 * the application stops issuing new requests until all its pending requests
2835 * have been completed. After that, the application may issue a new batch,
2837 * For this reason the next function is invoked to compute
2838 * soft_rt_next_start only for applications that meet this requirement,
2839 * whereas soft_rt_next_start is set to infinity for applications that do
2842 * Unfortunately, even a greedy application may happen to behave in an
2843 * isochronous way if the CPU load is high. In fact, the application may
2844 * stop issuing requests while the CPUs are busy serving other processes,
2845 * then restart, then stop again for a while, and so on. In addition, if
2846 * the disk achieves a low enough throughput with the request pattern
2847 * issued by the application (e.g., because the request pattern is random
2848 * and/or the device is slow), then the application may meet the above
2849 * bandwidth requirement too. To prevent such a greedy application to be
2850 * deemed as soft real-time, a further rule is used in the computation of
2851 * soft_rt_next_start: soft_rt_next_start must be higher than the current
2852 * time plus the maximum time for which the arrival of a request is waited
2853 * for when a sync queue becomes idle, namely bfqd->bfq_slice_idle.
2854 * This filters out greedy applications, as the latter issue instead their
2855 * next request as soon as possible after the last one has been completed
2856 * (in contrast, when a batch of requests is completed, a soft real-time
2857 * application spends some time processing data).
2859 * Unfortunately, the last filter may easily generate false positives if
2860 * only bfqd->bfq_slice_idle is used as a reference time interval and one
2861 * or both the following cases occur:
2862 * 1) HZ is so low that the duration of a jiffy is comparable to or higher
2863 * than bfqd->bfq_slice_idle. This happens, e.g., on slow devices with
2865 * 2) jiffies, instead of increasing at a constant rate, may stop increasing
2866 * for a while, then suddenly 'jump' by several units to recover the lost
2867 * increments. This seems to happen, e.g., inside virtual machines.
2868 * To address this issue, we do not use as a reference time interval just
2869 * bfqd->bfq_slice_idle, but bfqd->bfq_slice_idle plus a few jiffies. In
2870 * particular we add the minimum number of jiffies for which the filter
2871 * seems to be quite precise also in embedded systems and KVM/QEMU virtual
2874 static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data
*bfqd
,
2875 struct bfq_queue
*bfqq
)
2877 return max(bfqq
->last_idle_bklogged
+
2878 HZ
* bfqq
->service_from_backlogged
/
2879 bfqd
->bfq_wr_max_softrt_rate
,
2880 jiffies
+ nsecs_to_jiffies(bfqq
->bfqd
->bfq_slice_idle
) + 4);
2884 * Return the farthest future time instant according to jiffies
2887 static unsigned long bfq_greatest_from_now(void)
2889 return jiffies
+ MAX_JIFFY_OFFSET
;
2893 * Return the farthest past time instant according to jiffies
2896 static unsigned long bfq_smallest_from_now(void)
2898 return jiffies
- MAX_JIFFY_OFFSET
;
2902 * bfq_bfqq_expire - expire a queue.
2903 * @bfqd: device owning the queue.
2904 * @bfqq: the queue to expire.
2905 * @compensate: if true, compensate for the time spent idling.
2906 * @reason: the reason causing the expiration.
2908 * If the process associated with bfqq does slow I/O (e.g., because it
2909 * issues random requests), we charge bfqq with the time it has been
2910 * in service instead of the service it has received (see
2911 * bfq_bfqq_charge_time for details on how this goal is achieved). As
2912 * a consequence, bfqq will typically get higher timestamps upon
2913 * reactivation, and hence it will be rescheduled as if it had
2914 * received more service than what it has actually received. In the
2915 * end, bfqq receives less service in proportion to how slowly its
2916 * associated process consumes its budgets (and hence how seriously it
2917 * tends to lower the throughput). In addition, this time-charging
2918 * strategy guarantees time fairness among slow processes. In
2919 * contrast, if the process associated with bfqq is not slow, we
2920 * charge bfqq exactly with the service it has received.
2922 * Charging time to the first type of queues and the exact service to
2923 * the other has the effect of using the WF2Q+ policy to schedule the
2924 * former on a timeslice basis, without violating service domain
2925 * guarantees among the latter.
2927 void bfq_bfqq_expire(struct bfq_data
*bfqd
,
2928 struct bfq_queue
*bfqq
,
2930 enum bfqq_expiration reason
)
2933 unsigned long delta
= 0;
2934 struct bfq_entity
*entity
= &bfqq
->entity
;
2938 * Check whether the process is slow (see bfq_bfqq_is_slow).
2940 slow
= bfq_bfqq_is_slow(bfqd
, bfqq
, compensate
, reason
, &delta
);
2943 * Increase service_from_backlogged before next statement,
2944 * because the possible next invocation of
2945 * bfq_bfqq_charge_time would likely inflate
2946 * entity->service. In contrast, service_from_backlogged must
2947 * contain real service, to enable the soft real-time
2948 * heuristic to correctly compute the bandwidth consumed by
2951 bfqq
->service_from_backlogged
+= entity
->service
;
2954 * As above explained, charge slow (typically seeky) and
2955 * timed-out queues with the time and not the service
2956 * received, to favor sequential workloads.
2958 * Processes doing I/O in the slower disk zones will tend to
2959 * be slow(er) even if not seeky. Therefore, since the
2960 * estimated peak rate is actually an average over the disk
2961 * surface, these processes may timeout just for bad luck. To
2962 * avoid punishing them, do not charge time to processes that
2963 * succeeded in consuming at least 2/3 of their budget. This
2964 * allows BFQ to preserve enough elasticity to still perform
2965 * bandwidth, and not time, distribution with little unlucky
2966 * or quasi-sequential processes.
2968 if (bfqq
->wr_coeff
== 1 &&
2970 (reason
== BFQQE_BUDGET_TIMEOUT
&&
2971 bfq_bfqq_budget_left(bfqq
) >= entity
->budget
/ 3)))
2972 bfq_bfqq_charge_time(bfqd
, bfqq
, delta
);
2974 if (reason
== BFQQE_TOO_IDLE
&&
2975 entity
->service
<= 2 * entity
->budget
/ 10)
2976 bfq_clear_bfqq_IO_bound(bfqq
);
2978 if (bfqd
->low_latency
&& bfqq
->wr_coeff
== 1)
2979 bfqq
->last_wr_start_finish
= jiffies
;
2981 if (bfqd
->low_latency
&& bfqd
->bfq_wr_max_softrt_rate
> 0 &&
2982 RB_EMPTY_ROOT(&bfqq
->sort_list
)) {
2984 * If we get here, and there are no outstanding
2985 * requests, then the request pattern is isochronous
2986 * (see the comments on the function
2987 * bfq_bfqq_softrt_next_start()). Thus we can compute
2988 * soft_rt_next_start. If, instead, the queue still
2989 * has outstanding requests, then we have to wait for
2990 * the completion of all the outstanding requests to
2991 * discover whether the request pattern is actually
2994 if (bfqq
->dispatched
== 0)
2995 bfqq
->soft_rt_next_start
=
2996 bfq_bfqq_softrt_next_start(bfqd
, bfqq
);
2999 * The application is still waiting for the
3000 * completion of one or more requests:
3001 * prevent it from possibly being incorrectly
3002 * deemed as soft real-time by setting its
3003 * soft_rt_next_start to infinity. In fact,
3004 * without this assignment, the application
3005 * would be incorrectly deemed as soft
3007 * 1) it issued a new request before the
3008 * completion of all its in-flight
3010 * 2) at that time, its soft_rt_next_start
3011 * happened to be in the past.
3013 bfqq
->soft_rt_next_start
=
3014 bfq_greatest_from_now();
3016 * Schedule an update of soft_rt_next_start to when
3017 * the task may be discovered to be isochronous.
3019 bfq_mark_bfqq_softrt_update(bfqq
);
3023 bfq_log_bfqq(bfqd
, bfqq
,
3024 "expire (%d, slow %d, num_disp %d, idle_win %d)", reason
,
3025 slow
, bfqq
->dispatched
, bfq_bfqq_idle_window(bfqq
));
3028 * Increase, decrease or leave budget unchanged according to
3031 __bfq_bfqq_recalc_budget(bfqd
, bfqq
, reason
);
3033 __bfq_bfqq_expire(bfqd
, bfqq
);
3035 /* mark bfqq as waiting a request only if a bic still points to it */
3036 if (ref
> 1 && !bfq_bfqq_busy(bfqq
) &&
3037 reason
!= BFQQE_BUDGET_TIMEOUT
&&
3038 reason
!= BFQQE_BUDGET_EXHAUSTED
)
3039 bfq_mark_bfqq_non_blocking_wait_rq(bfqq
);
3043 * Budget timeout is not implemented through a dedicated timer, but
3044 * just checked on request arrivals and completions, as well as on
3045 * idle timer expirations.
3047 static bool bfq_bfqq_budget_timeout(struct bfq_queue
*bfqq
)
3049 return time_is_before_eq_jiffies(bfqq
->budget_timeout
);
3053 * If we expire a queue that is actively waiting (i.e., with the
3054 * device idled) for the arrival of a new request, then we may incur
3055 * the timestamp misalignment problem described in the body of the
3056 * function __bfq_activate_entity. Hence we return true only if this
3057 * condition does not hold, or if the queue is slow enough to deserve
3058 * only to be kicked off for preserving a high throughput.
3060 static bool bfq_may_expire_for_budg_timeout(struct bfq_queue
*bfqq
)
3062 bfq_log_bfqq(bfqq
->bfqd
, bfqq
,
3063 "may_budget_timeout: wait_request %d left %d timeout %d",
3064 bfq_bfqq_wait_request(bfqq
),
3065 bfq_bfqq_budget_left(bfqq
) >= bfqq
->entity
.budget
/ 3,
3066 bfq_bfqq_budget_timeout(bfqq
));
3068 return (!bfq_bfqq_wait_request(bfqq
) ||
3069 bfq_bfqq_budget_left(bfqq
) >= bfqq
->entity
.budget
/ 3)
3071 bfq_bfqq_budget_timeout(bfqq
);
3075 * For a queue that becomes empty, device idling is allowed only if
3076 * this function returns true for the queue. As a consequence, since
3077 * device idling plays a critical role in both throughput boosting and
3078 * service guarantees, the return value of this function plays a
3079 * critical role in both these aspects as well.
3081 * In a nutshell, this function returns true only if idling is
3082 * beneficial for throughput or, even if detrimental for throughput,
3083 * idling is however necessary to preserve service guarantees (low
3084 * latency, desired throughput distribution, ...). In particular, on
3085 * NCQ-capable devices, this function tries to return false, so as to
3086 * help keep the drives' internal queues full, whenever this helps the
3087 * device boost the throughput without causing any service-guarantee
3090 * In more detail, the return value of this function is obtained by,
3091 * first, computing a number of boolean variables that take into
3092 * account throughput and service-guarantee issues, and, then,
3093 * combining these variables in a logical expression. Most of the
3094 * issues taken into account are not trivial. We discuss these issues
3095 * individually while introducing the variables.
3097 static bool bfq_bfqq_may_idle(struct bfq_queue
*bfqq
)
3099 struct bfq_data
*bfqd
= bfqq
->bfqd
;
3100 bool idling_boosts_thr
, idling_boosts_thr_without_issues
,
3101 idling_needed_for_service_guarantees
,
3102 asymmetric_scenario
;
3104 if (bfqd
->strict_guarantees
)
3108 * The next variable takes into account the cases where idling
3109 * boosts the throughput.
3111 * The value of the variable is computed considering, first, that
3112 * idling is virtually always beneficial for the throughput if:
3113 * (a) the device is not NCQ-capable, or
3114 * (b) regardless of the presence of NCQ, the device is rotational
3115 * and the request pattern for bfqq is I/O-bound and sequential.
3117 * Secondly, and in contrast to the above item (b), idling an
3118 * NCQ-capable flash-based device would not boost the
3119 * throughput even with sequential I/O; rather it would lower
3120 * the throughput in proportion to how fast the device
3121 * is. Accordingly, the next variable is true if any of the
3122 * above conditions (a) and (b) is true, and, in particular,
3123 * happens to be false if bfqd is an NCQ-capable flash-based
3126 idling_boosts_thr
= !bfqd
->hw_tag
||
3127 (!blk_queue_nonrot(bfqd
->queue
) && bfq_bfqq_IO_bound(bfqq
) &&
3128 bfq_bfqq_idle_window(bfqq
));
3131 * The value of the next variable,
3132 * idling_boosts_thr_without_issues, is equal to that of
3133 * idling_boosts_thr, unless a special case holds. In this
3134 * special case, described below, idling may cause problems to
3135 * weight-raised queues.
3137 * When the request pool is saturated (e.g., in the presence
3138 * of write hogs), if the processes associated with
3139 * non-weight-raised queues ask for requests at a lower rate,
3140 * then processes associated with weight-raised queues have a
3141 * higher probability to get a request from the pool
3142 * immediately (or at least soon) when they need one. Thus
3143 * they have a higher probability to actually get a fraction
3144 * of the device throughput proportional to their high
3145 * weight. This is especially true with NCQ-capable drives,
3146 * which enqueue several requests in advance, and further
3147 * reorder internally-queued requests.
3149 * For this reason, we force to false the value of
3150 * idling_boosts_thr_without_issues if there are weight-raised
3151 * busy queues. In this case, and if bfqq is not weight-raised,
3152 * this guarantees that the device is not idled for bfqq (if,
3153 * instead, bfqq is weight-raised, then idling will be
3154 * guaranteed by another variable, see below). Combined with
3155 * the timestamping rules of BFQ (see [1] for details), this
3156 * behavior causes bfqq, and hence any sync non-weight-raised
3157 * queue, to get a lower number of requests served, and thus
3158 * to ask for a lower number of requests from the request
3159 * pool, before the busy weight-raised queues get served
3160 * again. This often mitigates starvation problems in the
3161 * presence of heavy write workloads and NCQ, thereby
3162 * guaranteeing a higher application and system responsiveness
3163 * in these hostile scenarios.
3165 idling_boosts_thr_without_issues
= idling_boosts_thr
&&
3166 bfqd
->wr_busy_queues
== 0;
3169 * There is then a case where idling must be performed not
3170 * for throughput concerns, but to preserve service
3173 * To introduce this case, we can note that allowing the drive
3174 * to enqueue more than one request at a time, and hence
3175 * delegating de facto final scheduling decisions to the
3176 * drive's internal scheduler, entails loss of control on the
3177 * actual request service order. In particular, the critical
3178 * situation is when requests from different processes happen
3179 * to be present, at the same time, in the internal queue(s)
3180 * of the drive. In such a situation, the drive, by deciding
3181 * the service order of the internally-queued requests, does
3182 * determine also the actual throughput distribution among
3183 * these processes. But the drive typically has no notion or
3184 * concern about per-process throughput distribution, and
3185 * makes its decisions only on a per-request basis. Therefore,
3186 * the service distribution enforced by the drive's internal
3187 * scheduler is likely to coincide with the desired
3188 * device-throughput distribution only in a completely
3189 * symmetric scenario where:
3190 * (i) each of these processes must get the same throughput as
3192 * (ii) all these processes have the same I/O pattern
3193 (either sequential or random).
3194 * In fact, in such a scenario, the drive will tend to treat
3195 * the requests of each of these processes in about the same
3196 * way as the requests of the others, and thus to provide
3197 * each of these processes with about the same throughput
3198 * (which is exactly the desired throughput distribution). In
3199 * contrast, in any asymmetric scenario, device idling is
3200 * certainly needed to guarantee that bfqq receives its
3201 * assigned fraction of the device throughput (see [1] for
3204 * We address this issue by controlling, actually, only the
3205 * symmetry sub-condition (i), i.e., provided that
3206 * sub-condition (i) holds, idling is not performed,
3207 * regardless of whether sub-condition (ii) holds. In other
3208 * words, only if sub-condition (i) holds, then idling is
3209 * allowed, and the device tends to be prevented from queueing
3210 * many requests, possibly of several processes. The reason
3211 * for not controlling also sub-condition (ii) is that we
3212 * exploit preemption to preserve guarantees in case of
3213 * symmetric scenarios, even if (ii) does not hold, as
3214 * explained in the next two paragraphs.
3216 * Even if a queue, say Q, is expired when it remains idle, Q
3217 * can still preempt the new in-service queue if the next
3218 * request of Q arrives soon (see the comments on
3219 * bfq_bfqq_update_budg_for_activation). If all queues and
3220 * groups have the same weight, this form of preemption,
3221 * combined with the hole-recovery heuristic described in the
3222 * comments on function bfq_bfqq_update_budg_for_activation,
3223 * are enough to preserve a correct bandwidth distribution in
3224 * the mid term, even without idling. In fact, even if not
3225 * idling allows the internal queues of the device to contain
3226 * many requests, and thus to reorder requests, we can rather
3227 * safely assume that the internal scheduler still preserves a
3228 * minimum of mid-term fairness. The motivation for using
3229 * preemption instead of idling is that, by not idling,
3230 * service guarantees are preserved without minimally
3231 * sacrificing throughput. In other words, both a high
3232 * throughput and its desired distribution are obtained.
3234 * More precisely, this preemption-based, idleless approach
3235 * provides fairness in terms of IOPS, and not sectors per
3236 * second. This can be seen with a simple example. Suppose
3237 * that there are two queues with the same weight, but that
3238 * the first queue receives requests of 8 sectors, while the
3239 * second queue receives requests of 1024 sectors. In
3240 * addition, suppose that each of the two queues contains at
3241 * most one request at a time, which implies that each queue
3242 * always remains idle after it is served. Finally, after
3243 * remaining idle, each queue receives very quickly a new
3244 * request. It follows that the two queues are served
3245 * alternatively, preempting each other if needed. This
3246 * implies that, although both queues have the same weight,
3247 * the queue with large requests receives a service that is
3248 * 1024/8 times as high as the service received by the other
3251 * On the other hand, device idling is performed, and thus
3252 * pure sector-domain guarantees are provided, for the
3253 * following queues, which are likely to need stronger
3254 * throughput guarantees: weight-raised queues, and queues
3255 * with a higher weight than other queues. When such queues
3256 * are active, sub-condition (i) is false, which triggers
3259 * According to the above considerations, the next variable is
3260 * true (only) if sub-condition (i) holds. To compute the
3261 * value of this variable, we not only use the return value of
3262 * the function bfq_symmetric_scenario(), but also check
3263 * whether bfqq is being weight-raised, because
3264 * bfq_symmetric_scenario() does not take into account also
3265 * weight-raised queues (see comments on
3266 * bfq_weights_tree_add()).
3268 * As a side note, it is worth considering that the above
3269 * device-idling countermeasures may however fail in the
3270 * following unlucky scenario: if idling is (correctly)
3271 * disabled in a time period during which all symmetry
3272 * sub-conditions hold, and hence the device is allowed to
3273 * enqueue many requests, but at some later point in time some
3274 * sub-condition stops to hold, then it may become impossible
3275 * to let requests be served in the desired order until all
3276 * the requests already queued in the device have been served.
3278 asymmetric_scenario
= bfqq
->wr_coeff
> 1 ||
3279 !bfq_symmetric_scenario(bfqd
);
3282 * Finally, there is a case where maximizing throughput is the
3283 * best choice even if it may cause unfairness toward
3284 * bfqq. Such a case is when bfqq became active in a burst of
3285 * queue activations. Queues that became active during a large
3286 * burst benefit only from throughput, as discussed in the
3287 * comments on bfq_handle_burst. Thus, if bfqq became active
3288 * in a burst and not idling the device maximizes throughput,
3289 * then the device must no be idled, because not idling the
3290 * device provides bfqq and all other queues in the burst with
3291 * maximum benefit. Combining this and the above case, we can
3292 * now establish when idling is actually needed to preserve
3293 * service guarantees.
3295 idling_needed_for_service_guarantees
=
3296 asymmetric_scenario
&& !bfq_bfqq_in_large_burst(bfqq
);
3299 * We have now all the components we need to compute the return
3300 * value of the function, which is true only if both the following
3302 * 1) bfqq is sync, because idling make sense only for sync queues;
3303 * 2) idling either boosts the throughput (without issues), or
3304 * is necessary to preserve service guarantees.
3306 return bfq_bfqq_sync(bfqq
) &&
3307 (idling_boosts_thr_without_issues
||
3308 idling_needed_for_service_guarantees
);
3312 * If the in-service queue is empty but the function bfq_bfqq_may_idle
3313 * returns true, then:
3314 * 1) the queue must remain in service and cannot be expired, and
3315 * 2) the device must be idled to wait for the possible arrival of a new
3316 * request for the queue.
3317 * See the comments on the function bfq_bfqq_may_idle for the reasons
3318 * why performing device idling is the best choice to boost the throughput
3319 * and preserve service guarantees when bfq_bfqq_may_idle itself
3322 static bool bfq_bfqq_must_idle(struct bfq_queue
*bfqq
)
3324 struct bfq_data
*bfqd
= bfqq
->bfqd
;
3326 return RB_EMPTY_ROOT(&bfqq
->sort_list
) && bfqd
->bfq_slice_idle
!= 0 &&
3327 bfq_bfqq_may_idle(bfqq
);
3331 * Select a queue for service. If we have a current queue in service,
3332 * check whether to continue servicing it, or retrieve and set a new one.
3334 static struct bfq_queue
*bfq_select_queue(struct bfq_data
*bfqd
)
3336 struct bfq_queue
*bfqq
;
3337 struct request
*next_rq
;
3338 enum bfqq_expiration reason
= BFQQE_BUDGET_TIMEOUT
;
3340 bfqq
= bfqd
->in_service_queue
;
3344 bfq_log_bfqq(bfqd
, bfqq
, "select_queue: already in-service queue");
3346 if (bfq_may_expire_for_budg_timeout(bfqq
) &&
3347 !bfq_bfqq_wait_request(bfqq
) &&
3348 !bfq_bfqq_must_idle(bfqq
))
3353 * This loop is rarely executed more than once. Even when it
3354 * happens, it is much more convenient to re-execute this loop
3355 * than to return NULL and trigger a new dispatch to get a
3358 next_rq
= bfqq
->next_rq
;
3360 * If bfqq has requests queued and it has enough budget left to
3361 * serve them, keep the queue, otherwise expire it.
3364 if (bfq_serv_to_charge(next_rq
, bfqq
) >
3365 bfq_bfqq_budget_left(bfqq
)) {
3367 * Expire the queue for budget exhaustion,
3368 * which makes sure that the next budget is
3369 * enough to serve the next request, even if
3370 * it comes from the fifo expired path.
3372 reason
= BFQQE_BUDGET_EXHAUSTED
;
3376 * The idle timer may be pending because we may
3377 * not disable disk idling even when a new request
3380 if (bfq_bfqq_wait_request(bfqq
)) {
3382 * If we get here: 1) at least a new request
3383 * has arrived but we have not disabled the
3384 * timer because the request was too small,
3385 * 2) then the block layer has unplugged
3386 * the device, causing the dispatch to be
3389 * Since the device is unplugged, now the
3390 * requests are probably large enough to
3391 * provide a reasonable throughput.
3392 * So we disable idling.
3394 bfq_clear_bfqq_wait_request(bfqq
);
3395 hrtimer_try_to_cancel(&bfqd
->idle_slice_timer
);
3396 bfqg_stats_update_idle_time(bfqq_group(bfqq
));
3403 * No requests pending. However, if the in-service queue is idling
3404 * for a new request, or has requests waiting for a completion and
3405 * may idle after their completion, then keep it anyway.
3407 if (bfq_bfqq_wait_request(bfqq
) ||
3408 (bfqq
->dispatched
!= 0 && bfq_bfqq_may_idle(bfqq
))) {
3413 reason
= BFQQE_NO_MORE_REQUESTS
;
3415 bfq_bfqq_expire(bfqd
, bfqq
, false, reason
);
3417 bfqq
= bfq_set_in_service_queue(bfqd
);
3419 bfq_log_bfqq(bfqd
, bfqq
, "select_queue: checking new queue");
3424 bfq_log_bfqq(bfqd
, bfqq
, "select_queue: returned this queue");
3426 bfq_log(bfqd
, "select_queue: no queue returned");
3431 static void bfq_update_wr_data(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
3433 struct bfq_entity
*entity
= &bfqq
->entity
;
3435 if (bfqq
->wr_coeff
> 1) { /* queue is being weight-raised */
3436 bfq_log_bfqq(bfqd
, bfqq
,
3437 "raising period dur %u/%u msec, old coeff %u, w %d(%d)",
3438 jiffies_to_msecs(jiffies
- bfqq
->last_wr_start_finish
),
3439 jiffies_to_msecs(bfqq
->wr_cur_max_time
),
3441 bfqq
->entity
.weight
, bfqq
->entity
.orig_weight
);
3443 if (entity
->prio_changed
)
3444 bfq_log_bfqq(bfqd
, bfqq
, "WARN: pending prio change");
3447 * If the queue was activated in a burst, or too much
3448 * time has elapsed from the beginning of this
3449 * weight-raising period, then end weight raising.
3451 if (bfq_bfqq_in_large_burst(bfqq
))
3452 bfq_bfqq_end_wr(bfqq
);
3453 else if (time_is_before_jiffies(bfqq
->last_wr_start_finish
+
3454 bfqq
->wr_cur_max_time
)) {
3455 if (bfqq
->wr_cur_max_time
!= bfqd
->bfq_wr_rt_max_time
||
3456 time_is_before_jiffies(bfqq
->wr_start_at_switch_to_srt
+
3457 bfq_wr_duration(bfqd
)))
3458 bfq_bfqq_end_wr(bfqq
);
3460 /* switch back to interactive wr */
3461 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
;
3462 bfqq
->wr_cur_max_time
= bfq_wr_duration(bfqd
);
3463 bfqq
->last_wr_start_finish
=
3464 bfqq
->wr_start_at_switch_to_srt
;
3465 bfqq
->entity
.prio_changed
= 1;
3469 /* Update weight both if it must be raised and if it must be lowered */
3470 if ((entity
->weight
> entity
->orig_weight
) != (bfqq
->wr_coeff
> 1))
3471 __bfq_entity_update_weight_prio(
3472 bfq_entity_service_tree(entity
),
3477 * Dispatch next request from bfqq.
3479 static struct request
*bfq_dispatch_rq_from_bfqq(struct bfq_data
*bfqd
,
3480 struct bfq_queue
*bfqq
)
3482 struct request
*rq
= bfqq
->next_rq
;
3483 unsigned long service_to_charge
;
3485 service_to_charge
= bfq_serv_to_charge(rq
, bfqq
);
3487 bfq_bfqq_served(bfqq
, service_to_charge
);
3489 bfq_dispatch_remove(bfqd
->queue
, rq
);
3492 * If weight raising has to terminate for bfqq, then next
3493 * function causes an immediate update of bfqq's weight,
3494 * without waiting for next activation. As a consequence, on
3495 * expiration, bfqq will be timestamped as if has never been
3496 * weight-raised during this service slot, even if it has
3497 * received part or even most of the service as a
3498 * weight-raised queue. This inflates bfqq's timestamps, which
3499 * is beneficial, as bfqq is then more willing to leave the
3500 * device immediately to possible other weight-raised queues.
3502 bfq_update_wr_data(bfqd
, bfqq
);
3505 * Expire bfqq, pretending that its budget expired, if bfqq
3506 * belongs to CLASS_IDLE and other queues are waiting for
3509 if (bfqd
->busy_queues
> 1 && bfq_class_idle(bfqq
))
3515 bfq_bfqq_expire(bfqd
, bfqq
, false, BFQQE_BUDGET_EXHAUSTED
);
3519 static bool bfq_has_work(struct blk_mq_hw_ctx
*hctx
)
3521 struct bfq_data
*bfqd
= hctx
->queue
->elevator
->elevator_data
;
3524 * Avoiding lock: a race on bfqd->busy_queues should cause at
3525 * most a call to dispatch for nothing
3527 return !list_empty_careful(&bfqd
->dispatch
) ||
3528 bfqd
->busy_queues
> 0;
3531 static struct request
*__bfq_dispatch_request(struct blk_mq_hw_ctx
*hctx
)
3533 struct bfq_data
*bfqd
= hctx
->queue
->elevator
->elevator_data
;
3534 struct request
*rq
= NULL
;
3535 struct bfq_queue
*bfqq
= NULL
;
3537 if (!list_empty(&bfqd
->dispatch
)) {
3538 rq
= list_first_entry(&bfqd
->dispatch
, struct request
,
3540 list_del_init(&rq
->queuelist
);
3546 * Increment counters here, because this
3547 * dispatch does not follow the standard
3548 * dispatch flow (where counters are
3553 goto inc_in_driver_start_rq
;
3557 * We exploit the put_rq_private hook to decrement
3558 * rq_in_driver, but put_rq_private will not be
3559 * invoked on this request. So, to avoid unbalance,
3560 * just start this request, without incrementing
3561 * rq_in_driver. As a negative consequence,
3562 * rq_in_driver is deceptively lower than it should be
3563 * while this request is in service. This may cause
3564 * bfq_schedule_dispatch to be invoked uselessly.
3566 * As for implementing an exact solution, the
3567 * put_request hook, if defined, is probably invoked
3568 * also on this request. So, by exploiting this hook,
3569 * we could 1) increment rq_in_driver here, and 2)
3570 * decrement it in put_request. Such a solution would
3571 * let the value of the counter be always accurate,
3572 * but it would entail using an extra interface
3573 * function. This cost seems higher than the benefit,
3574 * being the frequency of non-elevator-private
3575 * requests very low.
3580 bfq_log(bfqd
, "dispatch requests: %d busy queues", bfqd
->busy_queues
);
3582 if (bfqd
->busy_queues
== 0)
3586 * Force device to serve one request at a time if
3587 * strict_guarantees is true. Forcing this service scheme is
3588 * currently the ONLY way to guarantee that the request
3589 * service order enforced by the scheduler is respected by a
3590 * queueing device. Otherwise the device is free even to make
3591 * some unlucky request wait for as long as the device
3594 * Of course, serving one request at at time may cause loss of
3597 if (bfqd
->strict_guarantees
&& bfqd
->rq_in_driver
> 0)
3600 bfqq
= bfq_select_queue(bfqd
);
3604 rq
= bfq_dispatch_rq_from_bfqq(bfqd
, bfqq
);
3607 inc_in_driver_start_rq
:
3608 bfqd
->rq_in_driver
++;
3610 rq
->rq_flags
|= RQF_STARTED
;
3616 static struct request
*bfq_dispatch_request(struct blk_mq_hw_ctx
*hctx
)
3618 struct bfq_data
*bfqd
= hctx
->queue
->elevator
->elevator_data
;
3621 spin_lock_irq(&bfqd
->lock
);
3623 rq
= __bfq_dispatch_request(hctx
);
3624 spin_unlock_irq(&bfqd
->lock
);
3630 * Task holds one reference to the queue, dropped when task exits. Each rq
3631 * in-flight on this queue also holds a reference, dropped when rq is freed.
3633 * Scheduler lock must be held here. Recall not to use bfqq after calling
3634 * this function on it.
3636 void bfq_put_queue(struct bfq_queue
*bfqq
)
3638 #ifdef CONFIG_BFQ_GROUP_IOSCHED
3639 struct bfq_group
*bfqg
= bfqq_group(bfqq
);
3643 bfq_log_bfqq(bfqq
->bfqd
, bfqq
, "put_queue: %p %d",
3650 if (bfq_bfqq_sync(bfqq
))
3652 * The fact that this queue is being destroyed does not
3653 * invalidate the fact that this queue may have been
3654 * activated during the current burst. As a consequence,
3655 * although the queue does not exist anymore, and hence
3656 * needs to be removed from the burst list if there,
3657 * the burst size has not to be decremented.
3659 hlist_del_init(&bfqq
->burst_list_node
);
3661 kmem_cache_free(bfq_pool
, bfqq
);
3662 #ifdef CONFIG_BFQ_GROUP_IOSCHED
3667 static void bfq_put_cooperator(struct bfq_queue
*bfqq
)
3669 struct bfq_queue
*__bfqq
, *next
;
3672 * If this queue was scheduled to merge with another queue, be
3673 * sure to drop the reference taken on that queue (and others in
3674 * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs.
3676 __bfqq
= bfqq
->new_bfqq
;
3680 next
= __bfqq
->new_bfqq
;
3681 bfq_put_queue(__bfqq
);
3686 static void bfq_exit_bfqq(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
3688 if (bfqq
== bfqd
->in_service_queue
) {
3689 __bfq_bfqq_expire(bfqd
, bfqq
);
3690 bfq_schedule_dispatch(bfqd
);
3693 bfq_log_bfqq(bfqd
, bfqq
, "exit_bfqq: %p, %d", bfqq
, bfqq
->ref
);
3695 bfq_put_cooperator(bfqq
);
3697 bfq_put_queue(bfqq
); /* release process reference */
3700 static void bfq_exit_icq_bfqq(struct bfq_io_cq
*bic
, bool is_sync
)
3702 struct bfq_queue
*bfqq
= bic_to_bfqq(bic
, is_sync
);
3703 struct bfq_data
*bfqd
;
3706 bfqd
= bfqq
->bfqd
; /* NULL if scheduler already exited */
3709 unsigned long flags
;
3711 spin_lock_irqsave(&bfqd
->lock
, flags
);
3712 bfq_exit_bfqq(bfqd
, bfqq
);
3713 bic_set_bfqq(bic
, NULL
, is_sync
);
3714 spin_unlock_irqrestore(&bfqd
->lock
, flags
);
3718 static void bfq_exit_icq(struct io_cq
*icq
)
3720 struct bfq_io_cq
*bic
= icq_to_bic(icq
);
3722 bfq_exit_icq_bfqq(bic
, true);
3723 bfq_exit_icq_bfqq(bic
, false);
3727 * Update the entity prio values; note that the new values will not
3728 * be used until the next (re)activation.
3731 bfq_set_next_ioprio_data(struct bfq_queue
*bfqq
, struct bfq_io_cq
*bic
)
3733 struct task_struct
*tsk
= current
;
3735 struct bfq_data
*bfqd
= bfqq
->bfqd
;
3740 ioprio_class
= IOPRIO_PRIO_CLASS(bic
->ioprio
);
3741 switch (ioprio_class
) {
3743 dev_err(bfqq
->bfqd
->queue
->backing_dev_info
->dev
,
3744 "bfq: bad prio class %d\n", ioprio_class
);
3745 case IOPRIO_CLASS_NONE
:
3747 * No prio set, inherit CPU scheduling settings.
3749 bfqq
->new_ioprio
= task_nice_ioprio(tsk
);
3750 bfqq
->new_ioprio_class
= task_nice_ioclass(tsk
);
3752 case IOPRIO_CLASS_RT
:
3753 bfqq
->new_ioprio
= IOPRIO_PRIO_DATA(bic
->ioprio
);
3754 bfqq
->new_ioprio_class
= IOPRIO_CLASS_RT
;
3756 case IOPRIO_CLASS_BE
:
3757 bfqq
->new_ioprio
= IOPRIO_PRIO_DATA(bic
->ioprio
);
3758 bfqq
->new_ioprio_class
= IOPRIO_CLASS_BE
;
3760 case IOPRIO_CLASS_IDLE
:
3761 bfqq
->new_ioprio_class
= IOPRIO_CLASS_IDLE
;
3762 bfqq
->new_ioprio
= 7;
3763 bfq_clear_bfqq_idle_window(bfqq
);
3767 if (bfqq
->new_ioprio
>= IOPRIO_BE_NR
) {
3768 pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n",
3770 bfqq
->new_ioprio
= IOPRIO_BE_NR
;
3773 bfqq
->entity
.new_weight
= bfq_ioprio_to_weight(bfqq
->new_ioprio
);
3774 bfqq
->entity
.prio_changed
= 1;
3777 static struct bfq_queue
*bfq_get_queue(struct bfq_data
*bfqd
,
3778 struct bio
*bio
, bool is_sync
,
3779 struct bfq_io_cq
*bic
);
3781 static void bfq_check_ioprio_change(struct bfq_io_cq
*bic
, struct bio
*bio
)
3783 struct bfq_data
*bfqd
= bic_to_bfqd(bic
);
3784 struct bfq_queue
*bfqq
;
3785 int ioprio
= bic
->icq
.ioc
->ioprio
;
3788 * This condition may trigger on a newly created bic, be sure to
3789 * drop the lock before returning.
3791 if (unlikely(!bfqd
) || likely(bic
->ioprio
== ioprio
))
3794 bic
->ioprio
= ioprio
;
3796 bfqq
= bic_to_bfqq(bic
, false);
3798 /* release process reference on this queue */
3799 bfq_put_queue(bfqq
);
3800 bfqq
= bfq_get_queue(bfqd
, bio
, BLK_RW_ASYNC
, bic
);
3801 bic_set_bfqq(bic
, bfqq
, false);
3804 bfqq
= bic_to_bfqq(bic
, true);
3806 bfq_set_next_ioprio_data(bfqq
, bic
);
3809 static void bfq_init_bfqq(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
3810 struct bfq_io_cq
*bic
, pid_t pid
, int is_sync
)
3812 RB_CLEAR_NODE(&bfqq
->entity
.rb_node
);
3813 INIT_LIST_HEAD(&bfqq
->fifo
);
3814 INIT_HLIST_NODE(&bfqq
->burst_list_node
);
3820 bfq_set_next_ioprio_data(bfqq
, bic
);
3823 if (!bfq_class_idle(bfqq
))
3824 bfq_mark_bfqq_idle_window(bfqq
);
3825 bfq_mark_bfqq_sync(bfqq
);
3826 bfq_mark_bfqq_just_created(bfqq
);
3828 bfq_clear_bfqq_sync(bfqq
);
3830 /* set end request to minus infinity from now */
3831 bfqq
->ttime
.last_end_request
= ktime_get_ns() + 1;
3833 bfq_mark_bfqq_IO_bound(bfqq
);
3837 /* Tentative initial value to trade off between thr and lat */
3838 bfqq
->max_budget
= (2 * bfq_max_budget(bfqd
)) / 3;
3839 bfqq
->budget_timeout
= bfq_smallest_from_now();
3842 bfqq
->last_wr_start_finish
= jiffies
;
3843 bfqq
->wr_start_at_switch_to_srt
= bfq_smallest_from_now();
3844 bfqq
->split_time
= bfq_smallest_from_now();
3847 * Set to the value for which bfqq will not be deemed as
3848 * soft rt when it becomes backlogged.
3850 bfqq
->soft_rt_next_start
= bfq_greatest_from_now();
3852 /* first request is almost certainly seeky */
3853 bfqq
->seek_history
= 1;
3856 static struct bfq_queue
**bfq_async_queue_prio(struct bfq_data
*bfqd
,
3857 struct bfq_group
*bfqg
,
3858 int ioprio_class
, int ioprio
)
3860 switch (ioprio_class
) {
3861 case IOPRIO_CLASS_RT
:
3862 return &bfqg
->async_bfqq
[0][ioprio
];
3863 case IOPRIO_CLASS_NONE
:
3864 ioprio
= IOPRIO_NORM
;
3866 case IOPRIO_CLASS_BE
:
3867 return &bfqg
->async_bfqq
[1][ioprio
];
3868 case IOPRIO_CLASS_IDLE
:
3869 return &bfqg
->async_idle_bfqq
;
3875 static struct bfq_queue
*bfq_get_queue(struct bfq_data
*bfqd
,
3876 struct bio
*bio
, bool is_sync
,
3877 struct bfq_io_cq
*bic
)
3879 const int ioprio
= IOPRIO_PRIO_DATA(bic
->ioprio
);
3880 const int ioprio_class
= IOPRIO_PRIO_CLASS(bic
->ioprio
);
3881 struct bfq_queue
**async_bfqq
= NULL
;
3882 struct bfq_queue
*bfqq
;
3883 struct bfq_group
*bfqg
;
3887 bfqg
= bfq_find_set_group(bfqd
, bio_blkcg(bio
));
3889 bfqq
= &bfqd
->oom_bfqq
;
3894 async_bfqq
= bfq_async_queue_prio(bfqd
, bfqg
, ioprio_class
,
3901 bfqq
= kmem_cache_alloc_node(bfq_pool
,
3902 GFP_NOWAIT
| __GFP_ZERO
| __GFP_NOWARN
,
3906 bfq_init_bfqq(bfqd
, bfqq
, bic
, current
->pid
,
3908 bfq_init_entity(&bfqq
->entity
, bfqg
);
3909 bfq_log_bfqq(bfqd
, bfqq
, "allocated");
3911 bfqq
= &bfqd
->oom_bfqq
;
3912 bfq_log_bfqq(bfqd
, bfqq
, "using oom bfqq");
3917 * Pin the queue now that it's allocated, scheduler exit will
3922 * Extra group reference, w.r.t. sync
3923 * queue. This extra reference is removed
3924 * only if bfqq->bfqg disappears, to
3925 * guarantee that this queue is not freed
3926 * until its group goes away.
3928 bfq_log_bfqq(bfqd
, bfqq
, "get_queue, bfqq not in async: %p, %d",
3934 bfqq
->ref
++; /* get a process reference to this queue */
3935 bfq_log_bfqq(bfqd
, bfqq
, "get_queue, at end: %p, %d", bfqq
, bfqq
->ref
);
3940 static void bfq_update_io_thinktime(struct bfq_data
*bfqd
,
3941 struct bfq_queue
*bfqq
)
3943 struct bfq_ttime
*ttime
= &bfqq
->ttime
;
3944 u64 elapsed
= ktime_get_ns() - bfqq
->ttime
.last_end_request
;
3946 elapsed
= min_t(u64
, elapsed
, 2ULL * bfqd
->bfq_slice_idle
);
3948 ttime
->ttime_samples
= (7*bfqq
->ttime
.ttime_samples
+ 256) / 8;
3949 ttime
->ttime_total
= div_u64(7*ttime
->ttime_total
+ 256*elapsed
, 8);
3950 ttime
->ttime_mean
= div64_ul(ttime
->ttime_total
+ 128,
3951 ttime
->ttime_samples
);
3955 bfq_update_io_seektime(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
3958 bfqq
->seek_history
<<= 1;
3959 bfqq
->seek_history
|=
3960 get_sdist(bfqq
->last_request_pos
, rq
) > BFQQ_SEEK_THR
&&
3961 (!blk_queue_nonrot(bfqd
->queue
) ||
3962 blk_rq_sectors(rq
) < BFQQ_SECT_THR_NONROT
);
3966 * Disable idle window if the process thinks too long or seeks so much that
3967 * it doesn't matter.
3969 static void bfq_update_idle_window(struct bfq_data
*bfqd
,
3970 struct bfq_queue
*bfqq
,
3971 struct bfq_io_cq
*bic
)
3975 /* Don't idle for async or idle io prio class. */
3976 if (!bfq_bfqq_sync(bfqq
) || bfq_class_idle(bfqq
))
3979 /* Idle window just restored, statistics are meaningless. */
3980 if (time_is_after_eq_jiffies(bfqq
->split_time
+
3981 bfqd
->bfq_wr_min_idle_time
))
3984 enable_idle
= bfq_bfqq_idle_window(bfqq
);
3986 if (atomic_read(&bic
->icq
.ioc
->active_ref
) == 0 ||
3987 bfqd
->bfq_slice_idle
== 0 ||
3988 (bfqd
->hw_tag
&& BFQQ_SEEKY(bfqq
) &&
3989 bfqq
->wr_coeff
== 1))
3991 else if (bfq_sample_valid(bfqq
->ttime
.ttime_samples
)) {
3992 if (bfqq
->ttime
.ttime_mean
> bfqd
->bfq_slice_idle
&&
3993 bfqq
->wr_coeff
== 1)
3998 bfq_log_bfqq(bfqd
, bfqq
, "update_idle_window: enable_idle %d",
4002 bfq_mark_bfqq_idle_window(bfqq
);
4004 bfq_clear_bfqq_idle_window(bfqq
);
4008 * Called when a new fs request (rq) is added to bfqq. Check if there's
4009 * something we should do about it.
4011 static void bfq_rq_enqueued(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
4014 struct bfq_io_cq
*bic
= RQ_BIC(rq
);
4016 if (rq
->cmd_flags
& REQ_META
)
4017 bfqq
->meta_pending
++;
4019 bfq_update_io_thinktime(bfqd
, bfqq
);
4020 bfq_update_io_seektime(bfqd
, bfqq
, rq
);
4021 if (bfqq
->entity
.service
> bfq_max_budget(bfqd
) / 8 ||
4023 bfq_update_idle_window(bfqd
, bfqq
, bic
);
4025 bfq_log_bfqq(bfqd
, bfqq
,
4026 "rq_enqueued: idle_window=%d (seeky %d)",
4027 bfq_bfqq_idle_window(bfqq
), BFQQ_SEEKY(bfqq
));
4029 bfqq
->last_request_pos
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
4031 if (bfqq
== bfqd
->in_service_queue
&& bfq_bfqq_wait_request(bfqq
)) {
4032 bool small_req
= bfqq
->queued
[rq_is_sync(rq
)] == 1 &&
4033 blk_rq_sectors(rq
) < 32;
4034 bool budget_timeout
= bfq_bfqq_budget_timeout(bfqq
);
4037 * There is just this request queued: if the request
4038 * is small and the queue is not to be expired, then
4041 * In this way, if the device is being idled to wait
4042 * for a new request from the in-service queue, we
4043 * avoid unplugging the device and committing the
4044 * device to serve just a small request. On the
4045 * contrary, we wait for the block layer to decide
4046 * when to unplug the device: hopefully, new requests
4047 * will be merged to this one quickly, then the device
4048 * will be unplugged and larger requests will be
4051 if (small_req
&& !budget_timeout
)
4055 * A large enough request arrived, or the queue is to
4056 * be expired: in both cases disk idling is to be
4057 * stopped, so clear wait_request flag and reset
4060 bfq_clear_bfqq_wait_request(bfqq
);
4061 hrtimer_try_to_cancel(&bfqd
->idle_slice_timer
);
4062 bfqg_stats_update_idle_time(bfqq_group(bfqq
));
4065 * The queue is not empty, because a new request just
4066 * arrived. Hence we can safely expire the queue, in
4067 * case of budget timeout, without risking that the
4068 * timestamps of the queue are not updated correctly.
4069 * See [1] for more details.
4072 bfq_bfqq_expire(bfqd
, bfqq
, false,
4073 BFQQE_BUDGET_TIMEOUT
);
4077 static void __bfq_insert_request(struct bfq_data
*bfqd
, struct request
*rq
)
4079 struct bfq_queue
*bfqq
= RQ_BFQQ(rq
),
4080 *new_bfqq
= bfq_setup_cooperator(bfqd
, bfqq
, rq
, true);
4083 if (bic_to_bfqq(RQ_BIC(rq
), 1) != bfqq
)
4084 new_bfqq
= bic_to_bfqq(RQ_BIC(rq
), 1);
4086 * Release the request's reference to the old bfqq
4087 * and make sure one is taken to the shared queue.
4089 new_bfqq
->allocated
++;
4092 bfq_clear_bfqq_just_created(bfqq
);
4094 * If the bic associated with the process
4095 * issuing this request still points to bfqq
4096 * (and thus has not been already redirected
4097 * to new_bfqq or even some other bfq_queue),
4098 * then complete the merge and redirect it to
4101 if (bic_to_bfqq(RQ_BIC(rq
), 1) == bfqq
)
4102 bfq_merge_bfqqs(bfqd
, RQ_BIC(rq
),
4105 * rq is about to be enqueued into new_bfqq,
4106 * release rq reference on bfqq
4108 bfq_put_queue(bfqq
);
4109 rq
->elv
.priv
[1] = new_bfqq
;
4113 bfq_add_request(rq
);
4115 rq
->fifo_time
= ktime_get_ns() + bfqd
->bfq_fifo_expire
[rq_is_sync(rq
)];
4116 list_add_tail(&rq
->queuelist
, &bfqq
->fifo
);
4118 bfq_rq_enqueued(bfqd
, bfqq
, rq
);
4121 static void bfq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
4124 struct request_queue
*q
= hctx
->queue
;
4125 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
4127 spin_lock_irq(&bfqd
->lock
);
4128 if (blk_mq_sched_try_insert_merge(q
, rq
)) {
4129 spin_unlock_irq(&bfqd
->lock
);
4133 spin_unlock_irq(&bfqd
->lock
);
4135 blk_mq_sched_request_inserted(rq
);
4137 spin_lock_irq(&bfqd
->lock
);
4138 if (at_head
|| blk_rq_is_passthrough(rq
)) {
4140 list_add(&rq
->queuelist
, &bfqd
->dispatch
);
4142 list_add_tail(&rq
->queuelist
, &bfqd
->dispatch
);
4144 __bfq_insert_request(bfqd
, rq
);
4146 if (rq_mergeable(rq
)) {
4147 elv_rqhash_add(q
, rq
);
4153 spin_unlock_irq(&bfqd
->lock
);
4156 static void bfq_insert_requests(struct blk_mq_hw_ctx
*hctx
,
4157 struct list_head
*list
, bool at_head
)
4159 while (!list_empty(list
)) {
4162 rq
= list_first_entry(list
, struct request
, queuelist
);
4163 list_del_init(&rq
->queuelist
);
4164 bfq_insert_request(hctx
, rq
, at_head
);
4168 static void bfq_update_hw_tag(struct bfq_data
*bfqd
)
4170 bfqd
->max_rq_in_driver
= max_t(int, bfqd
->max_rq_in_driver
,
4171 bfqd
->rq_in_driver
);
4173 if (bfqd
->hw_tag
== 1)
4177 * This sample is valid if the number of outstanding requests
4178 * is large enough to allow a queueing behavior. Note that the
4179 * sum is not exact, as it's not taking into account deactivated
4182 if (bfqd
->rq_in_driver
+ bfqd
->queued
< BFQ_HW_QUEUE_THRESHOLD
)
4185 if (bfqd
->hw_tag_samples
++ < BFQ_HW_QUEUE_SAMPLES
)
4188 bfqd
->hw_tag
= bfqd
->max_rq_in_driver
> BFQ_HW_QUEUE_THRESHOLD
;
4189 bfqd
->max_rq_in_driver
= 0;
4190 bfqd
->hw_tag_samples
= 0;
4193 static void bfq_completed_request(struct bfq_queue
*bfqq
, struct bfq_data
*bfqd
)
4198 bfq_update_hw_tag(bfqd
);
4200 bfqd
->rq_in_driver
--;
4203 if (!bfqq
->dispatched
&& !bfq_bfqq_busy(bfqq
)) {
4205 * Set budget_timeout (which we overload to store the
4206 * time at which the queue remains with no backlog and
4207 * no outstanding request; used by the weight-raising
4210 bfqq
->budget_timeout
= jiffies
;
4212 bfq_weights_tree_remove(bfqd
, &bfqq
->entity
,
4213 &bfqd
->queue_weights_tree
);
4216 now_ns
= ktime_get_ns();
4218 bfqq
->ttime
.last_end_request
= now_ns
;
4221 * Using us instead of ns, to get a reasonable precision in
4222 * computing rate in next check.
4224 delta_us
= div_u64(now_ns
- bfqd
->last_completion
, NSEC_PER_USEC
);
4227 * If the request took rather long to complete, and, according
4228 * to the maximum request size recorded, this completion latency
4229 * implies that the request was certainly served at a very low
4230 * rate (less than 1M sectors/sec), then the whole observation
4231 * interval that lasts up to this time instant cannot be a
4232 * valid time interval for computing a new peak rate. Invoke
4233 * bfq_update_rate_reset to have the following three steps
4235 * - close the observation interval at the last (previous)
4236 * request dispatch or completion
4237 * - compute rate, if possible, for that observation interval
4238 * - reset to zero samples, which will trigger a proper
4239 * re-initialization of the observation interval on next
4242 if (delta_us
> BFQ_MIN_TT
/NSEC_PER_USEC
&&
4243 (bfqd
->last_rq_max_size
<<BFQ_RATE_SHIFT
)/delta_us
<
4244 1UL<<(BFQ_RATE_SHIFT
- 10))
4245 bfq_update_rate_reset(bfqd
, NULL
);
4246 bfqd
->last_completion
= now_ns
;
4249 * If we are waiting to discover whether the request pattern
4250 * of the task associated with the queue is actually
4251 * isochronous, and both requisites for this condition to hold
4252 * are now satisfied, then compute soft_rt_next_start (see the
4253 * comments on the function bfq_bfqq_softrt_next_start()). We
4254 * schedule this delayed check when bfqq expires, if it still
4255 * has in-flight requests.
4257 if (bfq_bfqq_softrt_update(bfqq
) && bfqq
->dispatched
== 0 &&
4258 RB_EMPTY_ROOT(&bfqq
->sort_list
))
4259 bfqq
->soft_rt_next_start
=
4260 bfq_bfqq_softrt_next_start(bfqd
, bfqq
);
4263 * If this is the in-service queue, check if it needs to be expired,
4264 * or if we want to idle in case it has no pending requests.
4266 if (bfqd
->in_service_queue
== bfqq
) {
4267 if (bfqq
->dispatched
== 0 && bfq_bfqq_must_idle(bfqq
)) {
4268 bfq_arm_slice_timer(bfqd
);
4270 } else if (bfq_may_expire_for_budg_timeout(bfqq
))
4271 bfq_bfqq_expire(bfqd
, bfqq
, false,
4272 BFQQE_BUDGET_TIMEOUT
);
4273 else if (RB_EMPTY_ROOT(&bfqq
->sort_list
) &&
4274 (bfqq
->dispatched
== 0 ||
4275 !bfq_bfqq_may_idle(bfqq
)))
4276 bfq_bfqq_expire(bfqd
, bfqq
, false,
4277 BFQQE_NO_MORE_REQUESTS
);
4281 static void bfq_put_rq_priv_body(struct bfq_queue
*bfqq
)
4285 bfq_put_queue(bfqq
);
4288 static void bfq_put_rq_private(struct request_queue
*q
, struct request
*rq
)
4290 struct bfq_queue
*bfqq
= RQ_BFQQ(rq
);
4291 struct bfq_data
*bfqd
= bfqq
->bfqd
;
4293 if (rq
->rq_flags
& RQF_STARTED
)
4294 bfqg_stats_update_completion(bfqq_group(bfqq
),
4295 rq_start_time_ns(rq
),
4296 rq_io_start_time_ns(rq
),
4299 if (likely(rq
->rq_flags
& RQF_STARTED
)) {
4300 unsigned long flags
;
4302 spin_lock_irqsave(&bfqd
->lock
, flags
);
4304 bfq_completed_request(bfqq
, bfqd
);
4305 bfq_put_rq_priv_body(bfqq
);
4307 spin_unlock_irqrestore(&bfqd
->lock
, flags
);
4310 * Request rq may be still/already in the scheduler,
4311 * in which case we need to remove it. And we cannot
4312 * defer such a check and removal, to avoid
4313 * inconsistencies in the time interval from the end
4314 * of this function to the start of the deferred work.
4315 * This situation seems to occur only in process
4316 * context, as a consequence of a merge. In the
4317 * current version of the code, this implies that the
4321 if (!RB_EMPTY_NODE(&rq
->rb_node
))
4322 bfq_remove_request(q
, rq
);
4323 bfq_put_rq_priv_body(bfqq
);
4326 rq
->elv
.priv
[0] = NULL
;
4327 rq
->elv
.priv
[1] = NULL
;
4331 * Returns NULL if a new bfqq should be allocated, or the old bfqq if this
4332 * was the last process referring to that bfqq.
4334 static struct bfq_queue
*
4335 bfq_split_bfqq(struct bfq_io_cq
*bic
, struct bfq_queue
*bfqq
)
4337 bfq_log_bfqq(bfqq
->bfqd
, bfqq
, "splitting queue");
4339 if (bfqq_process_refs(bfqq
) == 1) {
4340 bfqq
->pid
= current
->pid
;
4341 bfq_clear_bfqq_coop(bfqq
);
4342 bfq_clear_bfqq_split_coop(bfqq
);
4346 bic_set_bfqq(bic
, NULL
, 1);
4348 bfq_put_cooperator(bfqq
);
4350 bfq_put_queue(bfqq
);
4354 static struct bfq_queue
*bfq_get_bfqq_handle_split(struct bfq_data
*bfqd
,
4355 struct bfq_io_cq
*bic
,
4357 bool split
, bool is_sync
,
4360 struct bfq_queue
*bfqq
= bic_to_bfqq(bic
, is_sync
);
4362 if (likely(bfqq
&& bfqq
!= &bfqd
->oom_bfqq
))
4369 bfq_put_queue(bfqq
);
4370 bfqq
= bfq_get_queue(bfqd
, bio
, is_sync
, bic
);
4372 bic_set_bfqq(bic
, bfqq
, is_sync
);
4373 if (split
&& is_sync
) {
4374 if ((bic
->was_in_burst_list
&& bfqd
->large_burst
) ||
4375 bic
->saved_in_large_burst
)
4376 bfq_mark_bfqq_in_large_burst(bfqq
);
4378 bfq_clear_bfqq_in_large_burst(bfqq
);
4379 if (bic
->was_in_burst_list
)
4380 hlist_add_head(&bfqq
->burst_list_node
,
4383 bfqq
->split_time
= jiffies
;
4390 * Allocate bfq data structures associated with this request.
4392 static int bfq_get_rq_private(struct request_queue
*q
, struct request
*rq
,
4395 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
4396 struct bfq_io_cq
*bic
= icq_to_bic(rq
->elv
.icq
);
4397 const int is_sync
= rq_is_sync(rq
);
4398 struct bfq_queue
*bfqq
;
4399 bool new_queue
= false;
4402 spin_lock_irq(&bfqd
->lock
);
4407 bfq_check_ioprio_change(bic
, bio
);
4409 bfq_bic_update_cgroup(bic
, bio
);
4411 bfqq
= bfq_get_bfqq_handle_split(bfqd
, bic
, bio
, false, is_sync
,
4414 if (likely(!new_queue
)) {
4415 /* If the queue was seeky for too long, break it apart. */
4416 if (bfq_bfqq_coop(bfqq
) && bfq_bfqq_split_coop(bfqq
)) {
4417 bfq_log_bfqq(bfqd
, bfqq
, "breaking apart bfqq");
4419 /* Update bic before losing reference to bfqq */
4420 if (bfq_bfqq_in_large_burst(bfqq
))
4421 bic
->saved_in_large_burst
= true;
4423 bfqq
= bfq_split_bfqq(bic
, bfqq
);
4427 bfqq
= bfq_get_bfqq_handle_split(bfqd
, bic
, bio
,
4435 bfq_log_bfqq(bfqd
, bfqq
, "get_request %p: bfqq %p, %d",
4436 rq
, bfqq
, bfqq
->ref
);
4438 rq
->elv
.priv
[0] = bic
;
4439 rq
->elv
.priv
[1] = bfqq
;
4442 * If a bfq_queue has only one process reference, it is owned
4443 * by only this bic: we can then set bfqq->bic = bic. in
4444 * addition, if the queue has also just been split, we have to
4447 if (likely(bfqq
!= &bfqd
->oom_bfqq
) && bfqq_process_refs(bfqq
) == 1) {
4451 * The queue has just been split from a shared
4452 * queue: restore the idle window and the
4453 * possible weight raising period.
4455 bfq_bfqq_resume_state(bfqq
, bic
);
4459 if (unlikely(bfq_bfqq_just_created(bfqq
)))
4460 bfq_handle_burst(bfqd
, bfqq
);
4462 spin_unlock_irq(&bfqd
->lock
);
4467 spin_unlock_irq(&bfqd
->lock
);
4472 static void bfq_idle_slice_timer_body(struct bfq_queue
*bfqq
)
4474 struct bfq_data
*bfqd
= bfqq
->bfqd
;
4475 enum bfqq_expiration reason
;
4476 unsigned long flags
;
4478 spin_lock_irqsave(&bfqd
->lock
, flags
);
4479 bfq_clear_bfqq_wait_request(bfqq
);
4481 if (bfqq
!= bfqd
->in_service_queue
) {
4482 spin_unlock_irqrestore(&bfqd
->lock
, flags
);
4486 if (bfq_bfqq_budget_timeout(bfqq
))
4488 * Also here the queue can be safely expired
4489 * for budget timeout without wasting
4492 reason
= BFQQE_BUDGET_TIMEOUT
;
4493 else if (bfqq
->queued
[0] == 0 && bfqq
->queued
[1] == 0)
4495 * The queue may not be empty upon timer expiration,
4496 * because we may not disable the timer when the
4497 * first request of the in-service queue arrives
4498 * during disk idling.
4500 reason
= BFQQE_TOO_IDLE
;
4502 goto schedule_dispatch
;
4504 bfq_bfqq_expire(bfqd
, bfqq
, true, reason
);
4507 spin_unlock_irqrestore(&bfqd
->lock
, flags
);
4508 bfq_schedule_dispatch(bfqd
);
4512 * Handler of the expiration of the timer running if the in-service queue
4513 * is idling inside its time slice.
4515 static enum hrtimer_restart
bfq_idle_slice_timer(struct hrtimer
*timer
)
4517 struct bfq_data
*bfqd
= container_of(timer
, struct bfq_data
,
4519 struct bfq_queue
*bfqq
= bfqd
->in_service_queue
;
4522 * Theoretical race here: the in-service queue can be NULL or
4523 * different from the queue that was idling if a new request
4524 * arrives for the current queue and there is a full dispatch
4525 * cycle that changes the in-service queue. This can hardly
4526 * happen, but in the worst case we just expire a queue too
4530 bfq_idle_slice_timer_body(bfqq
);
4532 return HRTIMER_NORESTART
;
4535 static void __bfq_put_async_bfqq(struct bfq_data
*bfqd
,
4536 struct bfq_queue
**bfqq_ptr
)
4538 struct bfq_queue
*bfqq
= *bfqq_ptr
;
4540 bfq_log(bfqd
, "put_async_bfqq: %p", bfqq
);
4542 bfq_bfqq_move(bfqd
, bfqq
, bfqd
->root_group
);
4544 bfq_log_bfqq(bfqd
, bfqq
, "put_async_bfqq: putting %p, %d",
4546 bfq_put_queue(bfqq
);
4552 * Release all the bfqg references to its async queues. If we are
4553 * deallocating the group these queues may still contain requests, so
4554 * we reparent them to the root cgroup (i.e., the only one that will
4555 * exist for sure until all the requests on a device are gone).
4557 void bfq_put_async_queues(struct bfq_data
*bfqd
, struct bfq_group
*bfqg
)
4561 for (i
= 0; i
< 2; i
++)
4562 for (j
= 0; j
< IOPRIO_BE_NR
; j
++)
4563 __bfq_put_async_bfqq(bfqd
, &bfqg
->async_bfqq
[i
][j
]);
4565 __bfq_put_async_bfqq(bfqd
, &bfqg
->async_idle_bfqq
);
4568 static void bfq_exit_queue(struct elevator_queue
*e
)
4570 struct bfq_data
*bfqd
= e
->elevator_data
;
4571 struct bfq_queue
*bfqq
, *n
;
4573 hrtimer_cancel(&bfqd
->idle_slice_timer
);
4575 spin_lock_irq(&bfqd
->lock
);
4576 list_for_each_entry_safe(bfqq
, n
, &bfqd
->idle_list
, bfqq_list
)
4577 bfq_deactivate_bfqq(bfqd
, bfqq
, false, false);
4578 spin_unlock_irq(&bfqd
->lock
);
4580 hrtimer_cancel(&bfqd
->idle_slice_timer
);
4582 #ifdef CONFIG_BFQ_GROUP_IOSCHED
4583 blkcg_deactivate_policy(bfqd
->queue
, &blkcg_policy_bfq
);
4585 spin_lock_irq(&bfqd
->lock
);
4586 bfq_put_async_queues(bfqd
, bfqd
->root_group
);
4587 kfree(bfqd
->root_group
);
4588 spin_unlock_irq(&bfqd
->lock
);
4594 static void bfq_init_root_group(struct bfq_group
*root_group
,
4595 struct bfq_data
*bfqd
)
4599 #ifdef CONFIG_BFQ_GROUP_IOSCHED
4600 root_group
->entity
.parent
= NULL
;
4601 root_group
->my_entity
= NULL
;
4602 root_group
->bfqd
= bfqd
;
4604 root_group
->rq_pos_tree
= RB_ROOT
;
4605 for (i
= 0; i
< BFQ_IOPRIO_CLASSES
; i
++)
4606 root_group
->sched_data
.service_tree
[i
] = BFQ_SERVICE_TREE_INIT
;
4607 root_group
->sched_data
.bfq_class_idle_last_service
= jiffies
;
4610 static int bfq_init_queue(struct request_queue
*q
, struct elevator_type
*e
)
4612 struct bfq_data
*bfqd
;
4613 struct elevator_queue
*eq
;
4615 eq
= elevator_alloc(q
, e
);
4619 bfqd
= kzalloc_node(sizeof(*bfqd
), GFP_KERNEL
, q
->node
);
4621 kobject_put(&eq
->kobj
);
4624 eq
->elevator_data
= bfqd
;
4626 spin_lock_irq(q
->queue_lock
);
4628 spin_unlock_irq(q
->queue_lock
);
4631 * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues.
4632 * Grab a permanent reference to it, so that the normal code flow
4633 * will not attempt to free it.
4635 bfq_init_bfqq(bfqd
, &bfqd
->oom_bfqq
, NULL
, 1, 0);
4636 bfqd
->oom_bfqq
.ref
++;
4637 bfqd
->oom_bfqq
.new_ioprio
= BFQ_DEFAULT_QUEUE_IOPRIO
;
4638 bfqd
->oom_bfqq
.new_ioprio_class
= IOPRIO_CLASS_BE
;
4639 bfqd
->oom_bfqq
.entity
.new_weight
=
4640 bfq_ioprio_to_weight(bfqd
->oom_bfqq
.new_ioprio
);
4642 /* oom_bfqq does not participate to bursts */
4643 bfq_clear_bfqq_just_created(&bfqd
->oom_bfqq
);
4646 * Trigger weight initialization, according to ioprio, at the
4647 * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio
4648 * class won't be changed any more.
4650 bfqd
->oom_bfqq
.entity
.prio_changed
= 1;
4654 INIT_LIST_HEAD(&bfqd
->dispatch
);
4656 hrtimer_init(&bfqd
->idle_slice_timer
, CLOCK_MONOTONIC
,
4658 bfqd
->idle_slice_timer
.function
= bfq_idle_slice_timer
;
4660 bfqd
->queue_weights_tree
= RB_ROOT
;
4661 bfqd
->group_weights_tree
= RB_ROOT
;
4663 INIT_LIST_HEAD(&bfqd
->active_list
);
4664 INIT_LIST_HEAD(&bfqd
->idle_list
);
4665 INIT_HLIST_HEAD(&bfqd
->burst_list
);
4669 bfqd
->bfq_max_budget
= bfq_default_max_budget
;
4671 bfqd
->bfq_fifo_expire
[0] = bfq_fifo_expire
[0];
4672 bfqd
->bfq_fifo_expire
[1] = bfq_fifo_expire
[1];
4673 bfqd
->bfq_back_max
= bfq_back_max
;
4674 bfqd
->bfq_back_penalty
= bfq_back_penalty
;
4675 bfqd
->bfq_slice_idle
= bfq_slice_idle
;
4676 bfqd
->bfq_timeout
= bfq_timeout
;
4678 bfqd
->bfq_requests_within_timer
= 120;
4680 bfqd
->bfq_large_burst_thresh
= 8;
4681 bfqd
->bfq_burst_interval
= msecs_to_jiffies(180);
4683 bfqd
->low_latency
= true;
4686 * Trade-off between responsiveness and fairness.
4688 bfqd
->bfq_wr_coeff
= 30;
4689 bfqd
->bfq_wr_rt_max_time
= msecs_to_jiffies(300);
4690 bfqd
->bfq_wr_max_time
= 0;
4691 bfqd
->bfq_wr_min_idle_time
= msecs_to_jiffies(2000);
4692 bfqd
->bfq_wr_min_inter_arr_async
= msecs_to_jiffies(500);
4693 bfqd
->bfq_wr_max_softrt_rate
= 7000; /*
4694 * Approximate rate required
4695 * to playback or record a
4696 * high-definition compressed
4699 bfqd
->wr_busy_queues
= 0;
4702 * Begin by assuming, optimistically, that the device is a
4703 * high-speed one, and that its peak rate is equal to 2/3 of
4704 * the highest reference rate.
4706 bfqd
->RT_prod
= R_fast
[blk_queue_nonrot(bfqd
->queue
)] *
4707 T_fast
[blk_queue_nonrot(bfqd
->queue
)];
4708 bfqd
->peak_rate
= R_fast
[blk_queue_nonrot(bfqd
->queue
)] * 2 / 3;
4709 bfqd
->device_speed
= BFQ_BFQD_FAST
;
4711 spin_lock_init(&bfqd
->lock
);
4714 * The invocation of the next bfq_create_group_hierarchy
4715 * function is the head of a chain of function calls
4716 * (bfq_create_group_hierarchy->blkcg_activate_policy->
4717 * blk_mq_freeze_queue) that may lead to the invocation of the
4718 * has_work hook function. For this reason,
4719 * bfq_create_group_hierarchy is invoked only after all
4720 * scheduler data has been initialized, apart from the fields
4721 * that can be initialized only after invoking
4722 * bfq_create_group_hierarchy. This, in particular, enables
4723 * has_work to correctly return false. Of course, to avoid
4724 * other inconsistencies, the blk-mq stack must then refrain
4725 * from invoking further scheduler hooks before this init
4726 * function is finished.
4728 bfqd
->root_group
= bfq_create_group_hierarchy(bfqd
, q
->node
);
4729 if (!bfqd
->root_group
)
4731 bfq_init_root_group(bfqd
->root_group
, bfqd
);
4732 bfq_init_entity(&bfqd
->oom_bfqq
.entity
, bfqd
->root_group
);
4739 kobject_put(&eq
->kobj
);
4743 static void bfq_slab_kill(void)
4745 kmem_cache_destroy(bfq_pool
);
4748 static int __init
bfq_slab_setup(void)
4750 bfq_pool
= KMEM_CACHE(bfq_queue
, 0);
4756 static ssize_t
bfq_var_show(unsigned int var
, char *page
)
4758 return sprintf(page
, "%u\n", var
);
4761 static ssize_t
bfq_var_store(unsigned long *var
, const char *page
,
4764 unsigned long new_val
;
4765 int ret
= kstrtoul(page
, 10, &new_val
);
4773 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
4774 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
4776 struct bfq_data *bfqd = e->elevator_data; \
4777 u64 __data = __VAR; \
4779 __data = jiffies_to_msecs(__data); \
4780 else if (__CONV == 2) \
4781 __data = div_u64(__data, NSEC_PER_MSEC); \
4782 return bfq_var_show(__data, (page)); \
4784 SHOW_FUNCTION(bfq_fifo_expire_sync_show
, bfqd
->bfq_fifo_expire
[1], 2);
4785 SHOW_FUNCTION(bfq_fifo_expire_async_show
, bfqd
->bfq_fifo_expire
[0], 2);
4786 SHOW_FUNCTION(bfq_back_seek_max_show
, bfqd
->bfq_back_max
, 0);
4787 SHOW_FUNCTION(bfq_back_seek_penalty_show
, bfqd
->bfq_back_penalty
, 0);
4788 SHOW_FUNCTION(bfq_slice_idle_show
, bfqd
->bfq_slice_idle
, 2);
4789 SHOW_FUNCTION(bfq_max_budget_show
, bfqd
->bfq_user_max_budget
, 0);
4790 SHOW_FUNCTION(bfq_timeout_sync_show
, bfqd
->bfq_timeout
, 1);
4791 SHOW_FUNCTION(bfq_strict_guarantees_show
, bfqd
->strict_guarantees
, 0);
4792 SHOW_FUNCTION(bfq_low_latency_show
, bfqd
->low_latency
, 0);
4793 #undef SHOW_FUNCTION
4795 #define USEC_SHOW_FUNCTION(__FUNC, __VAR) \
4796 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
4798 struct bfq_data *bfqd = e->elevator_data; \
4799 u64 __data = __VAR; \
4800 __data = div_u64(__data, NSEC_PER_USEC); \
4801 return bfq_var_show(__data, (page)); \
4803 USEC_SHOW_FUNCTION(bfq_slice_idle_us_show
, bfqd
->bfq_slice_idle
);
4804 #undef USEC_SHOW_FUNCTION
4806 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
4808 __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4810 struct bfq_data *bfqd = e->elevator_data; \
4811 unsigned long uninitialized_var(__data); \
4812 int ret = bfq_var_store(&__data, (page), count); \
4813 if (__data < (MIN)) \
4815 else if (__data > (MAX)) \
4818 *(__PTR) = msecs_to_jiffies(__data); \
4819 else if (__CONV == 2) \
4820 *(__PTR) = (u64)__data * NSEC_PER_MSEC; \
4822 *(__PTR) = __data; \
4825 STORE_FUNCTION(bfq_fifo_expire_sync_store
, &bfqd
->bfq_fifo_expire
[1], 1,
4827 STORE_FUNCTION(bfq_fifo_expire_async_store
, &bfqd
->bfq_fifo_expire
[0], 1,
4829 STORE_FUNCTION(bfq_back_seek_max_store
, &bfqd
->bfq_back_max
, 0, INT_MAX
, 0);
4830 STORE_FUNCTION(bfq_back_seek_penalty_store
, &bfqd
->bfq_back_penalty
, 1,
4832 STORE_FUNCTION(bfq_slice_idle_store
, &bfqd
->bfq_slice_idle
, 0, INT_MAX
, 2);
4833 #undef STORE_FUNCTION
4835 #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
4836 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\
4838 struct bfq_data *bfqd = e->elevator_data; \
4839 unsigned long uninitialized_var(__data); \
4840 int ret = bfq_var_store(&__data, (page), count); \
4841 if (__data < (MIN)) \
4843 else if (__data > (MAX)) \
4845 *(__PTR) = (u64)__data * NSEC_PER_USEC; \
4848 USEC_STORE_FUNCTION(bfq_slice_idle_us_store
, &bfqd
->bfq_slice_idle
, 0,
4850 #undef USEC_STORE_FUNCTION
4852 static ssize_t
bfq_max_budget_store(struct elevator_queue
*e
,
4853 const char *page
, size_t count
)
4855 struct bfq_data
*bfqd
= e
->elevator_data
;
4856 unsigned long uninitialized_var(__data
);
4857 int ret
= bfq_var_store(&__data
, (page
), count
);
4860 bfqd
->bfq_max_budget
= bfq_calc_max_budget(bfqd
);
4862 if (__data
> INT_MAX
)
4864 bfqd
->bfq_max_budget
= __data
;
4867 bfqd
->bfq_user_max_budget
= __data
;
4873 * Leaving this name to preserve name compatibility with cfq
4874 * parameters, but this timeout is used for both sync and async.
4876 static ssize_t
bfq_timeout_sync_store(struct elevator_queue
*e
,
4877 const char *page
, size_t count
)
4879 struct bfq_data
*bfqd
= e
->elevator_data
;
4880 unsigned long uninitialized_var(__data
);
4881 int ret
= bfq_var_store(&__data
, (page
), count
);
4885 else if (__data
> INT_MAX
)
4888 bfqd
->bfq_timeout
= msecs_to_jiffies(__data
);
4889 if (bfqd
->bfq_user_max_budget
== 0)
4890 bfqd
->bfq_max_budget
= bfq_calc_max_budget(bfqd
);
4895 static ssize_t
bfq_strict_guarantees_store(struct elevator_queue
*e
,
4896 const char *page
, size_t count
)
4898 struct bfq_data
*bfqd
= e
->elevator_data
;
4899 unsigned long uninitialized_var(__data
);
4900 int ret
= bfq_var_store(&__data
, (page
), count
);
4904 if (!bfqd
->strict_guarantees
&& __data
== 1
4905 && bfqd
->bfq_slice_idle
< 8 * NSEC_PER_MSEC
)
4906 bfqd
->bfq_slice_idle
= 8 * NSEC_PER_MSEC
;
4908 bfqd
->strict_guarantees
= __data
;
4913 static ssize_t
bfq_low_latency_store(struct elevator_queue
*e
,
4914 const char *page
, size_t count
)
4916 struct bfq_data
*bfqd
= e
->elevator_data
;
4917 unsigned long uninitialized_var(__data
);
4918 int ret
= bfq_var_store(&__data
, (page
), count
);
4922 if (__data
== 0 && bfqd
->low_latency
!= 0)
4924 bfqd
->low_latency
= __data
;
4929 #define BFQ_ATTR(name) \
4930 __ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store)
4932 static struct elv_fs_entry bfq_attrs
[] = {
4933 BFQ_ATTR(fifo_expire_sync
),
4934 BFQ_ATTR(fifo_expire_async
),
4935 BFQ_ATTR(back_seek_max
),
4936 BFQ_ATTR(back_seek_penalty
),
4937 BFQ_ATTR(slice_idle
),
4938 BFQ_ATTR(slice_idle_us
),
4939 BFQ_ATTR(max_budget
),
4940 BFQ_ATTR(timeout_sync
),
4941 BFQ_ATTR(strict_guarantees
),
4942 BFQ_ATTR(low_latency
),
4946 static struct elevator_type iosched_bfq_mq
= {
4948 .get_rq_priv
= bfq_get_rq_private
,
4949 .put_rq_priv
= bfq_put_rq_private
,
4950 .exit_icq
= bfq_exit_icq
,
4951 .insert_requests
= bfq_insert_requests
,
4952 .dispatch_request
= bfq_dispatch_request
,
4953 .next_request
= elv_rb_latter_request
,
4954 .former_request
= elv_rb_former_request
,
4955 .allow_merge
= bfq_allow_bio_merge
,
4956 .bio_merge
= bfq_bio_merge
,
4957 .request_merge
= bfq_request_merge
,
4958 .requests_merged
= bfq_requests_merged
,
4959 .request_merged
= bfq_request_merged
,
4960 .has_work
= bfq_has_work
,
4961 .init_sched
= bfq_init_queue
,
4962 .exit_sched
= bfq_exit_queue
,
4966 .icq_size
= sizeof(struct bfq_io_cq
),
4967 .icq_align
= __alignof__(struct bfq_io_cq
),
4968 .elevator_attrs
= bfq_attrs
,
4969 .elevator_name
= "bfq",
4970 .elevator_owner
= THIS_MODULE
,
4973 static int __init
bfq_init(void)
4977 #ifdef CONFIG_BFQ_GROUP_IOSCHED
4978 ret
= blkcg_policy_register(&blkcg_policy_bfq
);
4984 if (bfq_slab_setup())
4988 * Times to load large popular applications for the typical
4989 * systems installed on the reference devices (see the
4990 * comments before the definitions of the next two
4991 * arrays). Actually, we use slightly slower values, as the
4992 * estimated peak rate tends to be smaller than the actual
4993 * peak rate. The reason for this last fact is that estimates
4994 * are computed over much shorter time intervals than the long
4995 * intervals typically used for benchmarking. Why? First, to
4996 * adapt more quickly to variations. Second, because an I/O
4997 * scheduler cannot rely on a peak-rate-evaluation workload to
4998 * be run for a long time.
5000 T_slow
[0] = msecs_to_jiffies(3500); /* actually 4 sec */
5001 T_slow
[1] = msecs_to_jiffies(6000); /* actually 6.5 sec */
5002 T_fast
[0] = msecs_to_jiffies(7000); /* actually 8 sec */
5003 T_fast
[1] = msecs_to_jiffies(2500); /* actually 3 sec */
5006 * Thresholds that determine the switch between speed classes
5007 * (see the comments before the definition of the array
5008 * device_speed_thresh). These thresholds are biased towards
5009 * transitions to the fast class. This is safer than the
5010 * opposite bias. In fact, a wrong transition to the slow
5011 * class results in short weight-raising periods, because the
5012 * speed of the device then tends to be higher that the
5013 * reference peak rate. On the opposite end, a wrong
5014 * transition to the fast class tends to increase
5015 * weight-raising periods, because of the opposite reason.
5017 device_speed_thresh
[0] = (4 * R_slow
[0]) / 3;
5018 device_speed_thresh
[1] = (4 * R_slow
[1]) / 3;
5020 ret
= elv_register(&iosched_bfq_mq
);
5027 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5028 blkcg_policy_unregister(&blkcg_policy_bfq
);
5033 static void __exit
bfq_exit(void)
5035 elv_unregister(&iosched_bfq_mq
);
5036 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5037 blkcg_policy_unregister(&blkcg_policy_bfq
);
5042 module_init(bfq_init
);
5043 module_exit(bfq_exit
);
5045 MODULE_AUTHOR("Paolo Valente");
5046 MODULE_LICENSE("GPL");
5047 MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler");