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 if (hctx
->flags
& BLK_MQ_F_TAG_SHARED
) {
63 struct request_queue
*q
= hctx
->queue
;
65 if (!test_and_set_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
66 atomic_inc(&q
->shared_hctx_restart
);
68 set_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
);
71 void blk_mq_sched_restart(struct blk_mq_hw_ctx
*hctx
)
73 if (!test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
76 clear_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
);
78 if (blk_mq_hctx_has_pending(hctx
)) {
79 blk_mq_run_hw_queue(hctx
, true);
85 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
86 * its queue by itself in its completion handler, so we don't need to
87 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
89 static void blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx
*hctx
)
91 struct request_queue
*q
= hctx
->queue
;
92 struct elevator_queue
*e
= q
->elevator
;
98 if (e
->type
->ops
.mq
.has_work
&&
99 !e
->type
->ops
.mq
.has_work(hctx
))
102 if (!blk_mq_get_dispatch_budget(hctx
))
105 rq
= e
->type
->ops
.mq
.dispatch_request(hctx
);
107 blk_mq_put_dispatch_budget(hctx
);
112 * Now this rq owns the budget which has to be released
113 * if this rq won't be queued to driver via .queue_rq()
114 * in blk_mq_dispatch_rq_list().
116 list_add(&rq
->queuelist
, &rq_list
);
117 } while (blk_mq_dispatch_rq_list(q
, &rq_list
, true));
120 static struct blk_mq_ctx
*blk_mq_next_ctx(struct blk_mq_hw_ctx
*hctx
,
121 struct blk_mq_ctx
*ctx
)
123 unsigned idx
= ctx
->index_hw
;
125 if (++idx
== hctx
->nr_ctx
)
128 return hctx
->ctxs
[idx
];
132 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
133 * its queue by itself in its completion handler, so we don't need to
134 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
136 static void blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx
*hctx
)
138 struct request_queue
*q
= hctx
->queue
;
140 struct blk_mq_ctx
*ctx
= READ_ONCE(hctx
->dispatch_from
);
145 if (!sbitmap_any_bit_set(&hctx
->ctx_map
))
148 if (!blk_mq_get_dispatch_budget(hctx
))
151 rq
= blk_mq_dequeue_from_ctx(hctx
, ctx
);
153 blk_mq_put_dispatch_budget(hctx
);
158 * Now this rq owns the budget which has to be released
159 * if this rq won't be queued to driver via .queue_rq()
160 * in blk_mq_dispatch_rq_list().
162 list_add(&rq
->queuelist
, &rq_list
);
164 /* round robin for fair dispatch */
165 ctx
= blk_mq_next_ctx(hctx
, rq
->mq_ctx
);
167 } while (blk_mq_dispatch_rq_list(q
, &rq_list
, true));
169 WRITE_ONCE(hctx
->dispatch_from
, ctx
);
172 /* return true if hw queue need to be run again */
173 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx
*hctx
)
175 struct request_queue
*q
= hctx
->queue
;
176 struct elevator_queue
*e
= q
->elevator
;
177 const bool has_sched_dispatch
= e
&& e
->type
->ops
.mq
.dispatch_request
;
180 /* RCU or SRCU read lock is needed before checking quiesced flag */
181 if (unlikely(blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(q
)))
187 * If we have previous entries on our dispatch list, grab them first for
188 * more fair dispatch.
190 if (!list_empty_careful(&hctx
->dispatch
)) {
191 spin_lock(&hctx
->lock
);
192 if (!list_empty(&hctx
->dispatch
))
193 list_splice_init(&hctx
->dispatch
, &rq_list
);
194 spin_unlock(&hctx
->lock
);
198 * Only ask the scheduler for requests, if we didn't have residual
199 * requests from the dispatch list. This is to avoid the case where
200 * we only ever dispatch a fraction of the requests available because
201 * of low device queue depth. Once we pull requests out of the IO
202 * scheduler, we can no longer merge or sort them. So it's best to
203 * leave them there for as long as we can. Mark the hw queue as
204 * needing a restart in that case.
206 * We want to dispatch from the scheduler if there was nothing
207 * on the dispatch list or we were able to dispatch from the
210 if (!list_empty(&rq_list
)) {
211 blk_mq_sched_mark_restart_hctx(hctx
);
212 if (blk_mq_dispatch_rq_list(q
, &rq_list
, false)) {
213 if (has_sched_dispatch
)
214 blk_mq_do_dispatch_sched(hctx
);
216 blk_mq_do_dispatch_ctx(hctx
);
218 } else if (has_sched_dispatch
) {
219 blk_mq_do_dispatch_sched(hctx
);
220 } else if (q
->mq_ops
->get_budget
) {
222 * If we need to get budget before queuing request, we
223 * dequeue request one by one from sw queue for avoiding
224 * to mess up I/O merge when dispatch runs out of resource.
226 * TODO: get more budgets, and dequeue more requests in
229 blk_mq_do_dispatch_ctx(hctx
);
231 blk_mq_flush_busy_ctxs(hctx
, &rq_list
);
232 blk_mq_dispatch_rq_list(q
, &rq_list
, false);
236 bool blk_mq_sched_try_merge(struct request_queue
*q
, struct bio
*bio
,
237 struct request
**merged_request
)
241 switch (elv_merge(q
, &rq
, bio
)) {
242 case ELEVATOR_BACK_MERGE
:
243 if (!blk_mq_sched_allow_merge(q
, rq
, bio
))
245 if (!bio_attempt_back_merge(q
, rq
, bio
))
247 *merged_request
= attempt_back_merge(q
, rq
);
248 if (!*merged_request
)
249 elv_merged_request(q
, rq
, ELEVATOR_BACK_MERGE
);
251 case ELEVATOR_FRONT_MERGE
:
252 if (!blk_mq_sched_allow_merge(q
, rq
, bio
))
254 if (!bio_attempt_front_merge(q
, rq
, bio
))
256 *merged_request
= attempt_front_merge(q
, rq
);
257 if (!*merged_request
)
258 elv_merged_request(q
, rq
, ELEVATOR_FRONT_MERGE
);
264 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge
);
267 * Reverse check our software queue for entries that we could potentially
268 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
269 * too much time checking for merges.
271 static bool blk_mq_attempt_merge(struct request_queue
*q
,
272 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
277 lockdep_assert_held(&ctx
->lock
);
279 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
285 if (!blk_rq_merge_ok(rq
, bio
))
288 switch (blk_try_merge(rq
, bio
)) {
289 case ELEVATOR_BACK_MERGE
:
290 if (blk_mq_sched_allow_merge(q
, rq
, bio
))
291 merged
= bio_attempt_back_merge(q
, rq
, bio
);
293 case ELEVATOR_FRONT_MERGE
:
294 if (blk_mq_sched_allow_merge(q
, rq
, bio
))
295 merged
= bio_attempt_front_merge(q
, rq
, bio
);
297 case ELEVATOR_DISCARD_MERGE
:
298 merged
= bio_attempt_discard_merge(q
, rq
, bio
);
312 bool __blk_mq_sched_bio_merge(struct request_queue
*q
, struct bio
*bio
)
314 struct elevator_queue
*e
= q
->elevator
;
315 struct blk_mq_ctx
*ctx
= blk_mq_get_ctx(q
);
316 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
319 if (e
&& e
->type
->ops
.mq
.bio_merge
) {
321 return e
->type
->ops
.mq
.bio_merge(hctx
, bio
);
324 if (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) {
325 /* default per sw-queue merge */
326 spin_lock(&ctx
->lock
);
327 ret
= blk_mq_attempt_merge(q
, ctx
, bio
);
328 spin_unlock(&ctx
->lock
);
335 bool blk_mq_sched_try_insert_merge(struct request_queue
*q
, struct request
*rq
)
337 return rq_mergeable(rq
) && elv_attempt_insert_merge(q
, rq
);
339 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge
);
341 void blk_mq_sched_request_inserted(struct request
*rq
)
343 trace_block_rq_insert(rq
->q
, rq
);
345 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted
);
347 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx
*hctx
,
351 rq
->rq_flags
|= RQF_SORTED
;
356 * If we already have a real request tag, send directly to
359 spin_lock(&hctx
->lock
);
360 list_add(&rq
->queuelist
, &hctx
->dispatch
);
361 spin_unlock(&hctx
->lock
);
366 * Add flush/fua to the queue. If we fail getting a driver tag, then
367 * punt to the requeue list. Requeue will re-invoke us from a context
368 * that's safe to block from.
370 static void blk_mq_sched_insert_flush(struct blk_mq_hw_ctx
*hctx
,
371 struct request
*rq
, bool can_block
)
373 if (blk_mq_get_driver_tag(rq
, &hctx
, can_block
)) {
374 blk_insert_flush(rq
);
375 blk_mq_run_hw_queue(hctx
, true);
377 blk_mq_add_to_requeue_list(rq
, false, true);
380 void blk_mq_sched_insert_request(struct request
*rq
, bool at_head
,
381 bool run_queue
, bool async
, bool can_block
)
383 struct request_queue
*q
= rq
->q
;
384 struct elevator_queue
*e
= q
->elevator
;
385 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
386 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
388 if (rq
->tag
== -1 && op_is_flush(rq
->cmd_flags
)) {
389 blk_mq_sched_insert_flush(hctx
, rq
, can_block
);
393 if (e
&& blk_mq_sched_bypass_insert(hctx
, rq
))
396 if (e
&& e
->type
->ops
.mq
.insert_requests
) {
399 list_add(&rq
->queuelist
, &list
);
400 e
->type
->ops
.mq
.insert_requests(hctx
, &list
, at_head
);
402 spin_lock(&ctx
->lock
);
403 __blk_mq_insert_request(hctx
, rq
, at_head
);
404 spin_unlock(&ctx
->lock
);
409 blk_mq_run_hw_queue(hctx
, async
);
412 void blk_mq_sched_insert_requests(struct request_queue
*q
,
413 struct blk_mq_ctx
*ctx
,
414 struct list_head
*list
, bool run_queue_async
)
416 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
417 struct elevator_queue
*e
= hctx
->queue
->elevator
;
420 struct request
*rq
, *next
;
423 * We bypass requests that already have a driver tag assigned,
424 * which should only be flushes. Flushes are only ever inserted
425 * as single requests, so we shouldn't ever hit the
426 * WARN_ON_ONCE() below (but let's handle it just in case).
428 list_for_each_entry_safe(rq
, next
, list
, queuelist
) {
429 if (WARN_ON_ONCE(rq
->tag
!= -1)) {
430 list_del_init(&rq
->queuelist
);
431 blk_mq_sched_bypass_insert(hctx
, rq
);
436 if (e
&& e
->type
->ops
.mq
.insert_requests
)
437 e
->type
->ops
.mq
.insert_requests(hctx
, list
, false);
439 blk_mq_insert_requests(hctx
, ctx
, list
);
441 blk_mq_run_hw_queue(hctx
, run_queue_async
);
444 static void blk_mq_sched_free_tags(struct blk_mq_tag_set
*set
,
445 struct blk_mq_hw_ctx
*hctx
,
446 unsigned int hctx_idx
)
448 if (hctx
->sched_tags
) {
449 blk_mq_free_rqs(set
, hctx
->sched_tags
, hctx_idx
);
450 blk_mq_free_rq_map(hctx
->sched_tags
);
451 hctx
->sched_tags
= NULL
;
455 static int blk_mq_sched_alloc_tags(struct request_queue
*q
,
456 struct blk_mq_hw_ctx
*hctx
,
457 unsigned int hctx_idx
)
459 struct blk_mq_tag_set
*set
= q
->tag_set
;
462 hctx
->sched_tags
= blk_mq_alloc_rq_map(set
, hctx_idx
, q
->nr_requests
,
464 if (!hctx
->sched_tags
)
467 ret
= blk_mq_alloc_rqs(set
, hctx
->sched_tags
, hctx_idx
, q
->nr_requests
);
469 blk_mq_sched_free_tags(set
, hctx
, hctx_idx
);
474 static void blk_mq_sched_tags_teardown(struct request_queue
*q
)
476 struct blk_mq_tag_set
*set
= q
->tag_set
;
477 struct blk_mq_hw_ctx
*hctx
;
480 queue_for_each_hw_ctx(q
, hctx
, i
)
481 blk_mq_sched_free_tags(set
, hctx
, i
);
484 int blk_mq_sched_init_hctx(struct request_queue
*q
, struct blk_mq_hw_ctx
*hctx
,
485 unsigned int hctx_idx
)
487 struct elevator_queue
*e
= q
->elevator
;
493 ret
= blk_mq_sched_alloc_tags(q
, hctx
, hctx_idx
);
497 if (e
->type
->ops
.mq
.init_hctx
) {
498 ret
= e
->type
->ops
.mq
.init_hctx(hctx
, hctx_idx
);
500 blk_mq_sched_free_tags(q
->tag_set
, hctx
, hctx_idx
);
505 blk_mq_debugfs_register_sched_hctx(q
, hctx
);
510 void blk_mq_sched_exit_hctx(struct request_queue
*q
, struct blk_mq_hw_ctx
*hctx
,
511 unsigned int hctx_idx
)
513 struct elevator_queue
*e
= q
->elevator
;
518 blk_mq_debugfs_unregister_sched_hctx(hctx
);
520 if (e
->type
->ops
.mq
.exit_hctx
&& hctx
->sched_data
) {
521 e
->type
->ops
.mq
.exit_hctx(hctx
, hctx_idx
);
522 hctx
->sched_data
= NULL
;
525 blk_mq_sched_free_tags(q
->tag_set
, hctx
, hctx_idx
);
528 int blk_mq_init_sched(struct request_queue
*q
, struct elevator_type
*e
)
530 struct blk_mq_hw_ctx
*hctx
;
531 struct elevator_queue
*eq
;
541 * Default to double of smaller one between hw queue_depth and 128,
542 * since we don't split into sync/async like the old code did.
543 * Additionally, this is a per-hw queue depth.
545 q
->nr_requests
= 2 * min_t(unsigned int, q
->tag_set
->queue_depth
,
548 queue_for_each_hw_ctx(q
, hctx
, i
) {
549 ret
= blk_mq_sched_alloc_tags(q
, hctx
, i
);
554 ret
= e
->ops
.mq
.init_sched(q
, e
);
558 blk_mq_debugfs_register_sched(q
);
560 queue_for_each_hw_ctx(q
, hctx
, i
) {
561 if (e
->ops
.mq
.init_hctx
) {
562 ret
= e
->ops
.mq
.init_hctx(hctx
, i
);
565 blk_mq_exit_sched(q
, eq
);
566 kobject_put(&eq
->kobj
);
570 blk_mq_debugfs_register_sched_hctx(q
, hctx
);
576 blk_mq_sched_tags_teardown(q
);
581 void blk_mq_exit_sched(struct request_queue
*q
, struct elevator_queue
*e
)
583 struct blk_mq_hw_ctx
*hctx
;
586 queue_for_each_hw_ctx(q
, hctx
, i
) {
587 blk_mq_debugfs_unregister_sched_hctx(hctx
);
588 if (e
->type
->ops
.mq
.exit_hctx
&& hctx
->sched_data
) {
589 e
->type
->ops
.mq
.exit_hctx(hctx
, i
);
590 hctx
->sched_data
= NULL
;
593 blk_mq_debugfs_unregister_sched(q
);
594 if (e
->type
->ops
.mq
.exit_sched
)
595 e
->type
->ops
.mq
.exit_sched(e
);
596 blk_mq_sched_tags_teardown(q
);
600 int blk_mq_sched_init(struct request_queue
*q
)
604 mutex_lock(&q
->sysfs_lock
);
605 ret
= elevator_init(q
, NULL
);
606 mutex_unlock(&q
->sysfs_lock
);