2 * blk-mq scheduling framework
4 * Copyright (C) 2016 Jens Axboe
6 #include <linux/kernel.h>
7 #include <linux/module.h>
8 #include <linux/blk-mq.h>
10 #include <trace/events/block.h>
14 #include "blk-mq-debugfs.h"
15 #include "blk-mq-sched.h"
16 #include "blk-mq-tag.h"
19 void blk_mq_sched_free_hctx_data(struct request_queue
*q
,
20 void (*exit
)(struct blk_mq_hw_ctx
*))
22 struct blk_mq_hw_ctx
*hctx
;
25 queue_for_each_hw_ctx(q
, hctx
, i
) {
26 if (exit
&& hctx
->sched_data
)
28 kfree(hctx
->sched_data
);
29 hctx
->sched_data
= NULL
;
32 EXPORT_SYMBOL_GPL(blk_mq_sched_free_hctx_data
);
34 static void __blk_mq_sched_assign_ioc(struct request_queue
*q
,
37 struct io_context
*ioc
)
41 spin_lock_irq(q
->queue_lock
);
42 icq
= ioc_lookup_icq(ioc
, q
);
43 spin_unlock_irq(q
->queue_lock
);
46 icq
= ioc_create_icq(ioc
, q
, GFP_ATOMIC
);
52 if (!blk_mq_sched_get_rq_priv(q
, rq
, bio
)) {
53 rq
->rq_flags
|= RQF_ELVPRIV
;
54 get_io_context(icq
->ioc
);
61 static void blk_mq_sched_assign_ioc(struct request_queue
*q
,
62 struct request
*rq
, struct bio
*bio
)
64 struct io_context
*ioc
;
68 __blk_mq_sched_assign_ioc(q
, rq
, bio
, ioc
);
71 struct request
*blk_mq_sched_get_request(struct request_queue
*q
,
74 struct blk_mq_alloc_data
*data
)
76 struct elevator_queue
*e
= q
->elevator
;
79 blk_queue_enter_live(q
);
81 if (likely(!data
->ctx
))
82 data
->ctx
= blk_mq_get_ctx(q
);
83 if (likely(!data
->hctx
))
84 data
->hctx
= blk_mq_map_queue(q
, data
->ctx
->cpu
);
87 data
->flags
|= BLK_MQ_REQ_INTERNAL
;
90 * Flush requests are special and go directly to the
93 if (!op_is_flush(op
) && e
->type
->ops
.mq
.get_request
) {
94 rq
= e
->type
->ops
.mq
.get_request(q
, op
, data
);
96 rq
->rq_flags
|= RQF_QUEUED
;
98 rq
= __blk_mq_alloc_request(data
, op
);
100 rq
= __blk_mq_alloc_request(data
, op
);
104 if (!op_is_flush(op
)) {
106 if (e
&& e
->type
->icq_cache
)
107 blk_mq_sched_assign_ioc(q
, rq
, bio
);
109 data
->hctx
->queued
++;
117 void blk_mq_sched_put_request(struct request
*rq
)
119 struct request_queue
*q
= rq
->q
;
120 struct elevator_queue
*e
= q
->elevator
;
122 if (rq
->rq_flags
& RQF_ELVPRIV
) {
123 blk_mq_sched_put_rq_priv(rq
->q
, rq
);
125 put_io_context(rq
->elv
.icq
->ioc
);
130 if ((rq
->rq_flags
& RQF_QUEUED
) && e
&& e
->type
->ops
.mq
.put_request
)
131 e
->type
->ops
.mq
.put_request(rq
);
133 blk_mq_finish_request(rq
);
136 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx
*hctx
)
138 struct request_queue
*q
= hctx
->queue
;
139 struct elevator_queue
*e
= q
->elevator
;
140 const bool has_sched_dispatch
= e
&& e
->type
->ops
.mq
.dispatch_request
;
141 bool did_work
= false;
144 if (unlikely(blk_mq_hctx_stopped(hctx
)))
150 * If we have previous entries on our dispatch list, grab them first for
151 * more fair dispatch.
153 if (!list_empty_careful(&hctx
->dispatch
)) {
154 spin_lock(&hctx
->lock
);
155 if (!list_empty(&hctx
->dispatch
))
156 list_splice_init(&hctx
->dispatch
, &rq_list
);
157 spin_unlock(&hctx
->lock
);
161 * Only ask the scheduler for requests, if we didn't have residual
162 * requests from the dispatch list. This is to avoid the case where
163 * we only ever dispatch a fraction of the requests available because
164 * of low device queue depth. Once we pull requests out of the IO
165 * scheduler, we can no longer merge or sort them. So it's best to
166 * leave them there for as long as we can. Mark the hw queue as
167 * needing a restart in that case.
169 if (!list_empty(&rq_list
)) {
170 blk_mq_sched_mark_restart_hctx(hctx
);
171 did_work
= blk_mq_dispatch_rq_list(q
, &rq_list
);
172 } else if (!has_sched_dispatch
) {
173 blk_mq_flush_busy_ctxs(hctx
, &rq_list
);
174 blk_mq_dispatch_rq_list(q
, &rq_list
);
178 * We want to dispatch from the scheduler if we had no work left
179 * on the dispatch list, OR if we did have work but weren't able
182 if (!did_work
&& has_sched_dispatch
) {
186 rq
= e
->type
->ops
.mq
.dispatch_request(hctx
);
189 list_add(&rq
->queuelist
, &rq_list
);
190 } while (blk_mq_dispatch_rq_list(q
, &rq_list
));
194 bool blk_mq_sched_try_merge(struct request_queue
*q
, struct bio
*bio
,
195 struct request
**merged_request
)
199 switch (elv_merge(q
, &rq
, bio
)) {
200 case ELEVATOR_BACK_MERGE
:
201 if (!blk_mq_sched_allow_merge(q
, rq
, bio
))
203 if (!bio_attempt_back_merge(q
, rq
, bio
))
205 *merged_request
= attempt_back_merge(q
, rq
);
206 if (!*merged_request
)
207 elv_merged_request(q
, rq
, ELEVATOR_BACK_MERGE
);
209 case ELEVATOR_FRONT_MERGE
:
210 if (!blk_mq_sched_allow_merge(q
, rq
, bio
))
212 if (!bio_attempt_front_merge(q
, rq
, bio
))
214 *merged_request
= attempt_front_merge(q
, rq
);
215 if (!*merged_request
)
216 elv_merged_request(q
, rq
, ELEVATOR_FRONT_MERGE
);
222 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge
);
225 * Reverse check our software queue for entries that we could potentially
226 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
227 * too much time checking for merges.
229 static bool blk_mq_attempt_merge(struct request_queue
*q
,
230 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
235 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
241 if (!blk_rq_merge_ok(rq
, bio
))
244 switch (blk_try_merge(rq
, bio
)) {
245 case ELEVATOR_BACK_MERGE
:
246 if (blk_mq_sched_allow_merge(q
, rq
, bio
))
247 merged
= bio_attempt_back_merge(q
, rq
, bio
);
249 case ELEVATOR_FRONT_MERGE
:
250 if (blk_mq_sched_allow_merge(q
, rq
, bio
))
251 merged
= bio_attempt_front_merge(q
, rq
, bio
);
253 case ELEVATOR_DISCARD_MERGE
:
254 merged
= bio_attempt_discard_merge(q
, rq
, bio
);
268 bool __blk_mq_sched_bio_merge(struct request_queue
*q
, struct bio
*bio
)
270 struct elevator_queue
*e
= q
->elevator
;
271 struct blk_mq_ctx
*ctx
= blk_mq_get_ctx(q
);
272 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
275 if (e
&& e
->type
->ops
.mq
.bio_merge
) {
277 return e
->type
->ops
.mq
.bio_merge(hctx
, bio
);
280 if (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) {
281 /* default per sw-queue merge */
282 spin_lock(&ctx
->lock
);
283 ret
= blk_mq_attempt_merge(q
, ctx
, bio
);
284 spin_unlock(&ctx
->lock
);
291 bool blk_mq_sched_try_insert_merge(struct request_queue
*q
, struct request
*rq
)
293 return rq_mergeable(rq
) && elv_attempt_insert_merge(q
, rq
);
295 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge
);
297 void blk_mq_sched_request_inserted(struct request
*rq
)
299 trace_block_rq_insert(rq
->q
, rq
);
301 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted
);
303 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx
*hctx
,
307 rq
->rq_flags
|= RQF_SORTED
;
312 * If we already have a real request tag, send directly to
315 spin_lock(&hctx
->lock
);
316 list_add(&rq
->queuelist
, &hctx
->dispatch
);
317 spin_unlock(&hctx
->lock
);
321 static bool blk_mq_sched_restart_hctx(struct blk_mq_hw_ctx
*hctx
)
323 if (test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
)) {
324 clear_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
);
325 if (blk_mq_hctx_has_pending(hctx
)) {
326 blk_mq_run_hw_queue(hctx
, true);
334 * list_for_each_entry_rcu_rr - iterate in a round-robin fashion over rcu list
336 * @skip: the list element that will not be examined. Iteration starts at
338 * @head: head of the list to examine. This list must have at least one
339 * element, namely @skip.
340 * @member: name of the list_head structure within typeof(*pos).
342 #define list_for_each_entry_rcu_rr(pos, skip, head, member) \
343 for ((pos) = (skip); \
344 (pos = (pos)->member.next != (head) ? list_entry_rcu( \
345 (pos)->member.next, typeof(*pos), member) : \
346 list_entry_rcu((pos)->member.next->next, typeof(*pos), member)), \
350 * Called after a driver tag has been freed to check whether a hctx needs to
351 * be restarted. Restarts @hctx if its tag set is not shared. Restarts hardware
352 * queues in a round-robin fashion if the tag set of @hctx is shared with other
355 void blk_mq_sched_restart(struct blk_mq_hw_ctx
*const hctx
)
357 struct blk_mq_tags
*const tags
= hctx
->tags
;
358 struct blk_mq_tag_set
*const set
= hctx
->queue
->tag_set
;
359 struct request_queue
*const queue
= hctx
->queue
, *q
;
360 struct blk_mq_hw_ctx
*hctx2
;
363 if (set
->flags
& BLK_MQ_F_TAG_SHARED
) {
365 list_for_each_entry_rcu_rr(q
, queue
, &set
->tag_list
,
367 queue_for_each_hw_ctx(q
, hctx2
, i
)
368 if (hctx2
->tags
== tags
&&
369 blk_mq_sched_restart_hctx(hctx2
))
372 j
= hctx
->queue_num
+ 1;
373 for (i
= 0; i
< queue
->nr_hw_queues
; i
++, j
++) {
374 if (j
== queue
->nr_hw_queues
)
376 hctx2
= queue
->queue_hw_ctx
[j
];
377 if (hctx2
->tags
== tags
&&
378 blk_mq_sched_restart_hctx(hctx2
))
384 blk_mq_sched_restart_hctx(hctx
);
389 * Add flush/fua to the queue. If we fail getting a driver tag, then
390 * punt to the requeue list. Requeue will re-invoke us from a context
391 * that's safe to block from.
393 static void blk_mq_sched_insert_flush(struct blk_mq_hw_ctx
*hctx
,
394 struct request
*rq
, bool can_block
)
396 if (blk_mq_get_driver_tag(rq
, &hctx
, can_block
)) {
397 blk_insert_flush(rq
);
398 blk_mq_run_hw_queue(hctx
, true);
400 blk_mq_add_to_requeue_list(rq
, false, true);
403 void blk_mq_sched_insert_request(struct request
*rq
, bool at_head
,
404 bool run_queue
, bool async
, bool can_block
)
406 struct request_queue
*q
= rq
->q
;
407 struct elevator_queue
*e
= q
->elevator
;
408 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
409 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
411 if (rq
->tag
== -1 && op_is_flush(rq
->cmd_flags
)) {
412 blk_mq_sched_insert_flush(hctx
, rq
, can_block
);
416 if (e
&& blk_mq_sched_bypass_insert(hctx
, rq
))
419 if (e
&& e
->type
->ops
.mq
.insert_requests
) {
422 list_add(&rq
->queuelist
, &list
);
423 e
->type
->ops
.mq
.insert_requests(hctx
, &list
, at_head
);
425 spin_lock(&ctx
->lock
);
426 __blk_mq_insert_request(hctx
, rq
, at_head
);
427 spin_unlock(&ctx
->lock
);
432 blk_mq_run_hw_queue(hctx
, async
);
435 void blk_mq_sched_insert_requests(struct request_queue
*q
,
436 struct blk_mq_ctx
*ctx
,
437 struct list_head
*list
, bool run_queue_async
)
439 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
440 struct elevator_queue
*e
= hctx
->queue
->elevator
;
443 struct request
*rq
, *next
;
446 * We bypass requests that already have a driver tag assigned,
447 * which should only be flushes. Flushes are only ever inserted
448 * as single requests, so we shouldn't ever hit the
449 * WARN_ON_ONCE() below (but let's handle it just in case).
451 list_for_each_entry_safe(rq
, next
, list
, queuelist
) {
452 if (WARN_ON_ONCE(rq
->tag
!= -1)) {
453 list_del_init(&rq
->queuelist
);
454 blk_mq_sched_bypass_insert(hctx
, rq
);
459 if (e
&& e
->type
->ops
.mq
.insert_requests
)
460 e
->type
->ops
.mq
.insert_requests(hctx
, list
, false);
462 blk_mq_insert_requests(hctx
, ctx
, list
);
464 blk_mq_run_hw_queue(hctx
, run_queue_async
);
467 static void blk_mq_sched_free_tags(struct blk_mq_tag_set
*set
,
468 struct blk_mq_hw_ctx
*hctx
,
469 unsigned int hctx_idx
)
471 if (hctx
->sched_tags
) {
472 blk_mq_free_rqs(set
, hctx
->sched_tags
, hctx_idx
);
473 blk_mq_free_rq_map(hctx
->sched_tags
);
474 hctx
->sched_tags
= NULL
;
478 static int blk_mq_sched_alloc_tags(struct request_queue
*q
,
479 struct blk_mq_hw_ctx
*hctx
,
480 unsigned int hctx_idx
)
482 struct blk_mq_tag_set
*set
= q
->tag_set
;
485 hctx
->sched_tags
= blk_mq_alloc_rq_map(set
, hctx_idx
, q
->nr_requests
,
487 if (!hctx
->sched_tags
)
490 ret
= blk_mq_alloc_rqs(set
, hctx
->sched_tags
, hctx_idx
, q
->nr_requests
);
492 blk_mq_sched_free_tags(set
, hctx
, hctx_idx
);
497 static void blk_mq_sched_tags_teardown(struct request_queue
*q
)
499 struct blk_mq_tag_set
*set
= q
->tag_set
;
500 struct blk_mq_hw_ctx
*hctx
;
503 queue_for_each_hw_ctx(q
, hctx
, i
)
504 blk_mq_sched_free_tags(set
, hctx
, i
);
507 int blk_mq_sched_init_hctx(struct request_queue
*q
, struct blk_mq_hw_ctx
*hctx
,
508 unsigned int hctx_idx
)
510 struct elevator_queue
*e
= q
->elevator
;
516 ret
= blk_mq_sched_alloc_tags(q
, hctx
, hctx_idx
);
520 if (e
->type
->ops
.mq
.init_hctx
) {
521 ret
= e
->type
->ops
.mq
.init_hctx(hctx
, hctx_idx
);
523 blk_mq_sched_free_tags(q
->tag_set
, hctx
, hctx_idx
);
528 blk_mq_debugfs_register_sched_hctx(q
, hctx
);
533 void blk_mq_sched_exit_hctx(struct request_queue
*q
, struct blk_mq_hw_ctx
*hctx
,
534 unsigned int hctx_idx
)
536 struct elevator_queue
*e
= q
->elevator
;
541 blk_mq_debugfs_unregister_sched_hctx(hctx
);
543 if (e
->type
->ops
.mq
.exit_hctx
&& hctx
->sched_data
) {
544 e
->type
->ops
.mq
.exit_hctx(hctx
, hctx_idx
);
545 hctx
->sched_data
= NULL
;
548 blk_mq_sched_free_tags(q
->tag_set
, hctx
, hctx_idx
);
551 int blk_mq_init_sched(struct request_queue
*q
, struct elevator_type
*e
)
553 struct blk_mq_hw_ctx
*hctx
;
554 struct elevator_queue
*eq
;
564 * Default to 256, since we don't split into sync/async like the
565 * old code did. Additionally, this is a per-hw queue depth.
567 q
->nr_requests
= 2 * BLKDEV_MAX_RQ
;
569 queue_for_each_hw_ctx(q
, hctx
, i
) {
570 ret
= blk_mq_sched_alloc_tags(q
, hctx
, i
);
575 ret
= e
->ops
.mq
.init_sched(q
, e
);
579 blk_mq_debugfs_register_sched(q
);
581 queue_for_each_hw_ctx(q
, hctx
, i
) {
582 if (e
->ops
.mq
.init_hctx
) {
583 ret
= e
->ops
.mq
.init_hctx(hctx
, i
);
586 blk_mq_exit_sched(q
, eq
);
587 kobject_put(&eq
->kobj
);
591 blk_mq_debugfs_register_sched_hctx(q
, hctx
);
597 blk_mq_sched_tags_teardown(q
);
602 void blk_mq_exit_sched(struct request_queue
*q
, struct elevator_queue
*e
)
604 struct blk_mq_hw_ctx
*hctx
;
607 queue_for_each_hw_ctx(q
, hctx
, i
) {
608 blk_mq_debugfs_unregister_sched_hctx(hctx
);
609 if (e
->type
->ops
.mq
.exit_hctx
&& hctx
->sched_data
) {
610 e
->type
->ops
.mq
.exit_hctx(hctx
, i
);
611 hctx
->sched_data
= NULL
;
614 blk_mq_debugfs_unregister_sched(q
);
615 if (e
->type
->ops
.mq
.exit_sched
)
616 e
->type
->ops
.mq
.exit_sched(e
);
617 blk_mq_sched_tags_teardown(q
);
621 int blk_mq_sched_init(struct request_queue
*q
)
625 mutex_lock(&q
->sysfs_lock
);
626 ret
= elevator_init(q
, NULL
);
627 mutex_unlock(&q
->sysfs_lock
);