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 void blk_mq_sched_assign_ioc(struct request
*rq
, struct bio
*bio
)
36 struct request_queue
*q
= rq
->q
;
37 struct io_context
*ioc
= rq_ioc(bio
);
40 spin_lock_irq(q
->queue_lock
);
41 icq
= ioc_lookup_icq(ioc
, q
);
42 spin_unlock_irq(q
->queue_lock
);
45 icq
= ioc_create_icq(ioc
, q
, GFP_ATOMIC
);
49 get_io_context(icq
->ioc
);
54 * Mark a hardware queue as needing a restart. For shared queues, maintain
55 * a count of how many hardware queues are marked for restart.
57 static void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx
*hctx
)
59 if (test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
62 set_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
);
65 void blk_mq_sched_restart(struct blk_mq_hw_ctx
*hctx
)
67 if (!test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
69 clear_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
);
71 blk_mq_run_hw_queue(hctx
, true);
75 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
76 * its queue by itself in its completion handler, so we don't need to
77 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
79 static void blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx
*hctx
)
81 struct request_queue
*q
= hctx
->queue
;
82 struct elevator_queue
*e
= q
->elevator
;
88 if (e
->type
->ops
.has_work
&& !e
->type
->ops
.has_work(hctx
))
91 if (!blk_mq_get_dispatch_budget(hctx
))
94 rq
= e
->type
->ops
.dispatch_request(hctx
);
96 blk_mq_put_dispatch_budget(hctx
);
101 * Now this rq owns the budget which has to be released
102 * if this rq won't be queued to driver via .queue_rq()
103 * in blk_mq_dispatch_rq_list().
105 list_add(&rq
->queuelist
, &rq_list
);
106 } while (blk_mq_dispatch_rq_list(q
, &rq_list
, true));
109 static struct blk_mq_ctx
*blk_mq_next_ctx(struct blk_mq_hw_ctx
*hctx
,
110 struct blk_mq_ctx
*ctx
)
112 unsigned short idx
= ctx
->index_hw
[hctx
->type
];
114 if (++idx
== hctx
->nr_ctx
)
117 return hctx
->ctxs
[idx
];
121 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
122 * its queue by itself in its completion handler, so we don't need to
123 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
125 static void blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx
*hctx
)
127 struct request_queue
*q
= hctx
->queue
;
129 struct blk_mq_ctx
*ctx
= READ_ONCE(hctx
->dispatch_from
);
134 if (!sbitmap_any_bit_set(&hctx
->ctx_map
))
137 if (!blk_mq_get_dispatch_budget(hctx
))
140 rq
= blk_mq_dequeue_from_ctx(hctx
, ctx
);
142 blk_mq_put_dispatch_budget(hctx
);
147 * Now this rq owns the budget which has to be released
148 * if this rq won't be queued to driver via .queue_rq()
149 * in blk_mq_dispatch_rq_list().
151 list_add(&rq
->queuelist
, &rq_list
);
153 /* round robin for fair dispatch */
154 ctx
= blk_mq_next_ctx(hctx
, rq
->mq_ctx
);
156 } while (blk_mq_dispatch_rq_list(q
, &rq_list
, true));
158 WRITE_ONCE(hctx
->dispatch_from
, ctx
);
161 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx
*hctx
)
163 struct request_queue
*q
= hctx
->queue
;
164 struct elevator_queue
*e
= q
->elevator
;
165 const bool has_sched_dispatch
= e
&& e
->type
->ops
.dispatch_request
;
168 /* RCU or SRCU read lock is needed before checking quiesced flag */
169 if (unlikely(blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(q
)))
175 * If we have previous entries on our dispatch list, grab them first for
176 * more fair dispatch.
178 if (!list_empty_careful(&hctx
->dispatch
)) {
179 spin_lock(&hctx
->lock
);
180 if (!list_empty(&hctx
->dispatch
))
181 list_splice_init(&hctx
->dispatch
, &rq_list
);
182 spin_unlock(&hctx
->lock
);
186 * Only ask the scheduler for requests, if we didn't have residual
187 * requests from the dispatch list. This is to avoid the case where
188 * we only ever dispatch a fraction of the requests available because
189 * of low device queue depth. Once we pull requests out of the IO
190 * scheduler, we can no longer merge or sort them. So it's best to
191 * leave them there for as long as we can. Mark the hw queue as
192 * needing a restart in that case.
194 * We want to dispatch from the scheduler if there was nothing
195 * on the dispatch list or we were able to dispatch from the
198 if (!list_empty(&rq_list
)) {
199 blk_mq_sched_mark_restart_hctx(hctx
);
200 if (blk_mq_dispatch_rq_list(q
, &rq_list
, false)) {
201 if (has_sched_dispatch
)
202 blk_mq_do_dispatch_sched(hctx
);
204 blk_mq_do_dispatch_ctx(hctx
);
206 } else if (has_sched_dispatch
) {
207 blk_mq_do_dispatch_sched(hctx
);
208 } else if (hctx
->dispatch_busy
) {
209 /* dequeue request one by one from sw queue if queue is busy */
210 blk_mq_do_dispatch_ctx(hctx
);
212 blk_mq_flush_busy_ctxs(hctx
, &rq_list
);
213 blk_mq_dispatch_rq_list(q
, &rq_list
, false);
217 bool blk_mq_sched_try_merge(struct request_queue
*q
, struct bio
*bio
,
218 struct request
**merged_request
)
222 switch (elv_merge(q
, &rq
, bio
)) {
223 case ELEVATOR_BACK_MERGE
:
224 if (!blk_mq_sched_allow_merge(q
, rq
, bio
))
226 if (!bio_attempt_back_merge(q
, rq
, bio
))
228 *merged_request
= attempt_back_merge(q
, rq
);
229 if (!*merged_request
)
230 elv_merged_request(q
, rq
, ELEVATOR_BACK_MERGE
);
232 case ELEVATOR_FRONT_MERGE
:
233 if (!blk_mq_sched_allow_merge(q
, rq
, bio
))
235 if (!bio_attempt_front_merge(q
, rq
, bio
))
237 *merged_request
= attempt_front_merge(q
, rq
);
238 if (!*merged_request
)
239 elv_merged_request(q
, rq
, ELEVATOR_FRONT_MERGE
);
241 case ELEVATOR_DISCARD_MERGE
:
242 return bio_attempt_discard_merge(q
, rq
, bio
);
247 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge
);
250 * Iterate list of requests and see if we can merge this bio with any
253 bool blk_mq_bio_list_merge(struct request_queue
*q
, struct list_head
*list
,
259 list_for_each_entry_reverse(rq
, list
, queuelist
) {
265 if (!blk_rq_merge_ok(rq
, bio
))
268 switch (blk_try_merge(rq
, bio
)) {
269 case ELEVATOR_BACK_MERGE
:
270 if (blk_mq_sched_allow_merge(q
, rq
, bio
))
271 merged
= bio_attempt_back_merge(q
, rq
, bio
);
273 case ELEVATOR_FRONT_MERGE
:
274 if (blk_mq_sched_allow_merge(q
, rq
, bio
))
275 merged
= bio_attempt_front_merge(q
, rq
, bio
);
277 case ELEVATOR_DISCARD_MERGE
:
278 merged
= bio_attempt_discard_merge(q
, rq
, bio
);
289 EXPORT_SYMBOL_GPL(blk_mq_bio_list_merge
);
292 * Reverse check our software queue for entries that we could potentially
293 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
294 * too much time checking for merges.
296 static bool blk_mq_attempt_merge(struct request_queue
*q
,
297 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
299 lockdep_assert_held(&ctx
->lock
);
301 if (blk_mq_bio_list_merge(q
, &ctx
->rq_list
, bio
)) {
309 bool __blk_mq_sched_bio_merge(struct request_queue
*q
, struct bio
*bio
)
311 struct elevator_queue
*e
= q
->elevator
;
312 struct blk_mq_ctx
*ctx
= blk_mq_get_ctx(q
);
313 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, bio
->bi_opf
, ctx
->cpu
);
316 if (e
&& e
->type
->ops
.bio_merge
) {
318 return e
->type
->ops
.bio_merge(hctx
, bio
);
321 if ((hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
322 !list_empty_careful(&ctx
->rq_list
)) {
323 /* default per sw-queue merge */
324 spin_lock(&ctx
->lock
);
325 ret
= blk_mq_attempt_merge(q
, ctx
, bio
);
326 spin_unlock(&ctx
->lock
);
333 bool blk_mq_sched_try_insert_merge(struct request_queue
*q
, struct request
*rq
)
335 return rq_mergeable(rq
) && elv_attempt_insert_merge(q
, rq
);
337 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge
);
339 void blk_mq_sched_request_inserted(struct request
*rq
)
341 trace_block_rq_insert(rq
->q
, rq
);
343 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted
);
345 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx
*hctx
,
349 /* dispatch flush rq directly */
350 if (rq
->rq_flags
& RQF_FLUSH_SEQ
) {
351 spin_lock(&hctx
->lock
);
352 list_add(&rq
->queuelist
, &hctx
->dispatch
);
353 spin_unlock(&hctx
->lock
);
358 rq
->rq_flags
|= RQF_SORTED
;
363 void blk_mq_sched_insert_request(struct request
*rq
, bool at_head
,
364 bool run_queue
, bool async
)
366 struct request_queue
*q
= rq
->q
;
367 struct elevator_queue
*e
= q
->elevator
;
368 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
369 struct blk_mq_hw_ctx
*hctx
;
371 hctx
= blk_mq_map_queue(q
, rq
->cmd_flags
, ctx
->cpu
);
373 /* flush rq in flush machinery need to be dispatched directly */
374 if (!(rq
->rq_flags
& RQF_FLUSH_SEQ
) && op_is_flush(rq
->cmd_flags
)) {
375 blk_insert_flush(rq
);
379 WARN_ON(e
&& (rq
->tag
!= -1));
381 if (blk_mq_sched_bypass_insert(hctx
, !!e
, rq
))
384 if (e
&& e
->type
->ops
.insert_requests
) {
387 list_add(&rq
->queuelist
, &list
);
388 e
->type
->ops
.insert_requests(hctx
, &list
, at_head
);
390 spin_lock(&ctx
->lock
);
391 __blk_mq_insert_request(hctx
, rq
, at_head
);
392 spin_unlock(&ctx
->lock
);
397 blk_mq_run_hw_queue(hctx
, async
);
400 void blk_mq_sched_insert_requests(struct request_queue
*q
,
401 struct blk_mq_ctx
*ctx
,
402 struct list_head
*list
, bool run_queue_async
)
404 struct blk_mq_hw_ctx
*hctx
;
405 struct elevator_queue
*e
;
408 /* For list inserts, requests better be on the same hw queue */
409 rq
= list_first_entry(list
, struct request
, queuelist
);
410 hctx
= blk_mq_map_queue(q
, rq
->cmd_flags
, ctx
->cpu
);
412 e
= hctx
->queue
->elevator
;
413 if (e
&& e
->type
->ops
.insert_requests
)
414 e
->type
->ops
.insert_requests(hctx
, list
, false);
417 * try to issue requests directly if the hw queue isn't
418 * busy in case of 'none' scheduler, and this way may save
419 * us one extra enqueue & dequeue to sw queue.
421 if (!hctx
->dispatch_busy
&& !e
&& !run_queue_async
) {
422 blk_mq_try_issue_list_directly(hctx
, list
);
423 if (list_empty(list
))
426 blk_mq_insert_requests(hctx
, ctx
, list
);
429 blk_mq_run_hw_queue(hctx
, run_queue_async
);
432 static void blk_mq_sched_free_tags(struct blk_mq_tag_set
*set
,
433 struct blk_mq_hw_ctx
*hctx
,
434 unsigned int hctx_idx
)
436 if (hctx
->sched_tags
) {
437 blk_mq_free_rqs(set
, hctx
->sched_tags
, hctx_idx
);
438 blk_mq_free_rq_map(hctx
->sched_tags
);
439 hctx
->sched_tags
= NULL
;
443 static int blk_mq_sched_alloc_tags(struct request_queue
*q
,
444 struct blk_mq_hw_ctx
*hctx
,
445 unsigned int hctx_idx
)
447 struct blk_mq_tag_set
*set
= q
->tag_set
;
450 hctx
->sched_tags
= blk_mq_alloc_rq_map(set
, hctx_idx
, q
->nr_requests
,
452 if (!hctx
->sched_tags
)
455 ret
= blk_mq_alloc_rqs(set
, hctx
->sched_tags
, hctx_idx
, q
->nr_requests
);
457 blk_mq_sched_free_tags(set
, hctx
, hctx_idx
);
462 static void blk_mq_sched_tags_teardown(struct request_queue
*q
)
464 struct blk_mq_tag_set
*set
= q
->tag_set
;
465 struct blk_mq_hw_ctx
*hctx
;
468 queue_for_each_hw_ctx(q
, hctx
, i
)
469 blk_mq_sched_free_tags(set
, hctx
, i
);
472 int blk_mq_init_sched(struct request_queue
*q
, struct elevator_type
*e
)
474 struct blk_mq_hw_ctx
*hctx
;
475 struct elevator_queue
*eq
;
481 q
->nr_requests
= q
->tag_set
->queue_depth
;
486 * Default to double of smaller one between hw queue_depth and 128,
487 * since we don't split into sync/async like the old code did.
488 * Additionally, this is a per-hw queue depth.
490 q
->nr_requests
= 2 * min_t(unsigned int, q
->tag_set
->queue_depth
,
493 queue_for_each_hw_ctx(q
, hctx
, i
) {
494 ret
= blk_mq_sched_alloc_tags(q
, hctx
, i
);
499 ret
= e
->ops
.init_sched(q
, e
);
503 blk_mq_debugfs_register_sched(q
);
505 queue_for_each_hw_ctx(q
, hctx
, i
) {
506 if (e
->ops
.init_hctx
) {
507 ret
= e
->ops
.init_hctx(hctx
, i
);
510 blk_mq_exit_sched(q
, eq
);
511 kobject_put(&eq
->kobj
);
515 blk_mq_debugfs_register_sched_hctx(q
, hctx
);
521 blk_mq_sched_tags_teardown(q
);
526 void blk_mq_exit_sched(struct request_queue
*q
, struct elevator_queue
*e
)
528 struct blk_mq_hw_ctx
*hctx
;
531 queue_for_each_hw_ctx(q
, hctx
, i
) {
532 blk_mq_debugfs_unregister_sched_hctx(hctx
);
533 if (e
->type
->ops
.exit_hctx
&& hctx
->sched_data
) {
534 e
->type
->ops
.exit_hctx(hctx
, i
);
535 hctx
->sched_data
= NULL
;
538 blk_mq_debugfs_unregister_sched(q
);
539 if (e
->type
->ops
.exit_sched
)
540 e
->type
->ops
.exit_sched(e
);
541 blk_mq_sched_tags_teardown(q
);