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
)
36 struct request_queue
*q
= rq
->q
;
37 struct io_context
*ioc
;
41 * May not have an IO context if it's a passthrough request
43 ioc
= current
->io_context
;
47 spin_lock_irq(&q
->queue_lock
);
48 icq
= ioc_lookup_icq(ioc
, q
);
49 spin_unlock_irq(&q
->queue_lock
);
52 icq
= ioc_create_icq(ioc
, q
, GFP_ATOMIC
);
56 get_io_context(icq
->ioc
);
61 * Mark a hardware queue as needing a restart. For shared queues, maintain
62 * a count of how many hardware queues are marked for restart.
64 void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx
*hctx
)
66 if (test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
69 set_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
);
71 EXPORT_SYMBOL_GPL(blk_mq_sched_mark_restart_hctx
);
73 void blk_mq_sched_restart(struct blk_mq_hw_ctx
*hctx
)
75 if (!test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
77 clear_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
);
79 blk_mq_run_hw_queue(hctx
, true);
83 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
84 * its queue by itself in its completion handler, so we don't need to
85 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
87 static void blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx
*hctx
)
89 struct request_queue
*q
= hctx
->queue
;
90 struct elevator_queue
*e
= q
->elevator
;
96 if (e
->type
->ops
.has_work
&& !e
->type
->ops
.has_work(hctx
))
99 if (!blk_mq_get_dispatch_budget(hctx
))
102 rq
= e
->type
->ops
.dispatch_request(hctx
);
104 blk_mq_put_dispatch_budget(hctx
);
109 * Now this rq owns the budget which has to be released
110 * if this rq won't be queued to driver via .queue_rq()
111 * in blk_mq_dispatch_rq_list().
113 list_add(&rq
->queuelist
, &rq_list
);
114 } while (blk_mq_dispatch_rq_list(q
, &rq_list
, true));
117 static struct blk_mq_ctx
*blk_mq_next_ctx(struct blk_mq_hw_ctx
*hctx
,
118 struct blk_mq_ctx
*ctx
)
120 unsigned short idx
= ctx
->index_hw
[hctx
->type
];
122 if (++idx
== hctx
->nr_ctx
)
125 return hctx
->ctxs
[idx
];
129 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
130 * its queue by itself in its completion handler, so we don't need to
131 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
133 static void blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx
*hctx
)
135 struct request_queue
*q
= hctx
->queue
;
137 struct blk_mq_ctx
*ctx
= READ_ONCE(hctx
->dispatch_from
);
142 if (!sbitmap_any_bit_set(&hctx
->ctx_map
))
145 if (!blk_mq_get_dispatch_budget(hctx
))
148 rq
= blk_mq_dequeue_from_ctx(hctx
, ctx
);
150 blk_mq_put_dispatch_budget(hctx
);
155 * Now this rq owns the budget which has to be released
156 * if this rq won't be queued to driver via .queue_rq()
157 * in blk_mq_dispatch_rq_list().
159 list_add(&rq
->queuelist
, &rq_list
);
161 /* round robin for fair dispatch */
162 ctx
= blk_mq_next_ctx(hctx
, rq
->mq_ctx
);
164 } while (blk_mq_dispatch_rq_list(q
, &rq_list
, true));
166 WRITE_ONCE(hctx
->dispatch_from
, ctx
);
169 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx
*hctx
)
171 struct request_queue
*q
= hctx
->queue
;
172 struct elevator_queue
*e
= q
->elevator
;
173 const bool has_sched_dispatch
= e
&& e
->type
->ops
.dispatch_request
;
176 /* RCU or SRCU read lock is needed before checking quiesced flag */
177 if (unlikely(blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(q
)))
183 * If we have previous entries on our dispatch list, grab them first for
184 * more fair dispatch.
186 if (!list_empty_careful(&hctx
->dispatch
)) {
187 spin_lock(&hctx
->lock
);
188 if (!list_empty(&hctx
->dispatch
))
189 list_splice_init(&hctx
->dispatch
, &rq_list
);
190 spin_unlock(&hctx
->lock
);
194 * Only ask the scheduler for requests, if we didn't have residual
195 * requests from the dispatch list. This is to avoid the case where
196 * we only ever dispatch a fraction of the requests available because
197 * of low device queue depth. Once we pull requests out of the IO
198 * scheduler, we can no longer merge or sort them. So it's best to
199 * leave them there for as long as we can. Mark the hw queue as
200 * needing a restart in that case.
202 * We want to dispatch from the scheduler if there was nothing
203 * on the dispatch list or we were able to dispatch from the
206 if (!list_empty(&rq_list
)) {
207 blk_mq_sched_mark_restart_hctx(hctx
);
208 if (blk_mq_dispatch_rq_list(q
, &rq_list
, false)) {
209 if (has_sched_dispatch
)
210 blk_mq_do_dispatch_sched(hctx
);
212 blk_mq_do_dispatch_ctx(hctx
);
214 } else if (has_sched_dispatch
) {
215 blk_mq_do_dispatch_sched(hctx
);
216 } else if (hctx
->dispatch_busy
) {
217 /* dequeue request one by one from sw queue if queue is busy */
218 blk_mq_do_dispatch_ctx(hctx
);
220 blk_mq_flush_busy_ctxs(hctx
, &rq_list
);
221 blk_mq_dispatch_rq_list(q
, &rq_list
, false);
225 bool blk_mq_sched_try_merge(struct request_queue
*q
, struct bio
*bio
,
226 struct request
**merged_request
)
230 switch (elv_merge(q
, &rq
, bio
)) {
231 case ELEVATOR_BACK_MERGE
:
232 if (!blk_mq_sched_allow_merge(q
, rq
, bio
))
234 if (!bio_attempt_back_merge(q
, rq
, bio
))
236 *merged_request
= attempt_back_merge(q
, rq
);
237 if (!*merged_request
)
238 elv_merged_request(q
, rq
, ELEVATOR_BACK_MERGE
);
240 case ELEVATOR_FRONT_MERGE
:
241 if (!blk_mq_sched_allow_merge(q
, rq
, bio
))
243 if (!bio_attempt_front_merge(q
, rq
, bio
))
245 *merged_request
= attempt_front_merge(q
, rq
);
246 if (!*merged_request
)
247 elv_merged_request(q
, rq
, ELEVATOR_FRONT_MERGE
);
249 case ELEVATOR_DISCARD_MERGE
:
250 return bio_attempt_discard_merge(q
, rq
, bio
);
255 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge
);
258 * Iterate list of requests and see if we can merge this bio with any
261 bool blk_mq_bio_list_merge(struct request_queue
*q
, struct list_head
*list
,
267 list_for_each_entry_reverse(rq
, list
, queuelist
) {
273 if (!blk_rq_merge_ok(rq
, bio
))
276 switch (blk_try_merge(rq
, bio
)) {
277 case ELEVATOR_BACK_MERGE
:
278 if (blk_mq_sched_allow_merge(q
, rq
, bio
))
279 merged
= bio_attempt_back_merge(q
, rq
, bio
);
281 case ELEVATOR_FRONT_MERGE
:
282 if (blk_mq_sched_allow_merge(q
, rq
, bio
))
283 merged
= bio_attempt_front_merge(q
, rq
, bio
);
285 case ELEVATOR_DISCARD_MERGE
:
286 merged
= bio_attempt_discard_merge(q
, rq
, bio
);
297 EXPORT_SYMBOL_GPL(blk_mq_bio_list_merge
);
300 * Reverse check our software queue for entries that we could potentially
301 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
302 * too much time checking for merges.
304 static bool blk_mq_attempt_merge(struct request_queue
*q
,
305 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
307 lockdep_assert_held(&ctx
->lock
);
309 if (blk_mq_bio_list_merge(q
, &ctx
->rq_list
, bio
)) {
317 bool __blk_mq_sched_bio_merge(struct request_queue
*q
, struct bio
*bio
)
319 struct elevator_queue
*e
= q
->elevator
;
320 struct blk_mq_ctx
*ctx
= blk_mq_get_ctx(q
);
321 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, bio
->bi_opf
, ctx
->cpu
);
324 if (e
&& e
->type
->ops
.bio_merge
) {
326 return e
->type
->ops
.bio_merge(hctx
, bio
);
329 if ((hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
330 !list_empty_careful(&ctx
->rq_list
)) {
331 /* default per sw-queue merge */
332 spin_lock(&ctx
->lock
);
333 ret
= blk_mq_attempt_merge(q
, ctx
, bio
);
334 spin_unlock(&ctx
->lock
);
341 bool blk_mq_sched_try_insert_merge(struct request_queue
*q
, struct request
*rq
)
343 return rq_mergeable(rq
) && elv_attempt_insert_merge(q
, rq
);
345 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge
);
347 void blk_mq_sched_request_inserted(struct request
*rq
)
349 trace_block_rq_insert(rq
->q
, rq
);
351 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted
);
353 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx
*hctx
,
357 /* dispatch flush rq directly */
358 if (rq
->rq_flags
& RQF_FLUSH_SEQ
) {
359 spin_lock(&hctx
->lock
);
360 list_add(&rq
->queuelist
, &hctx
->dispatch
);
361 spin_unlock(&hctx
->lock
);
366 rq
->rq_flags
|= RQF_SORTED
;
371 void blk_mq_sched_insert_request(struct request
*rq
, bool at_head
,
372 bool run_queue
, bool async
)
374 struct request_queue
*q
= rq
->q
;
375 struct elevator_queue
*e
= q
->elevator
;
376 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
377 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
379 /* flush rq in flush machinery need to be dispatched directly */
380 if (!(rq
->rq_flags
& RQF_FLUSH_SEQ
) && op_is_flush(rq
->cmd_flags
)) {
381 blk_insert_flush(rq
);
385 WARN_ON(e
&& (rq
->tag
!= -1));
387 if (blk_mq_sched_bypass_insert(hctx
, !!e
, rq
))
390 if (e
&& e
->type
->ops
.insert_requests
) {
393 list_add(&rq
->queuelist
, &list
);
394 e
->type
->ops
.insert_requests(hctx
, &list
, at_head
);
396 spin_lock(&ctx
->lock
);
397 __blk_mq_insert_request(hctx
, rq
, at_head
);
398 spin_unlock(&ctx
->lock
);
403 blk_mq_run_hw_queue(hctx
, async
);
406 void blk_mq_sched_insert_requests(struct blk_mq_hw_ctx
*hctx
,
407 struct blk_mq_ctx
*ctx
,
408 struct list_head
*list
, bool run_queue_async
)
410 struct elevator_queue
*e
;
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
);
424 blk_mq_insert_requests(hctx
, ctx
, list
);
427 blk_mq_run_hw_queue(hctx
, run_queue_async
);
430 static void blk_mq_sched_free_tags(struct blk_mq_tag_set
*set
,
431 struct blk_mq_hw_ctx
*hctx
,
432 unsigned int hctx_idx
)
434 if (hctx
->sched_tags
) {
435 blk_mq_free_rqs(set
, hctx
->sched_tags
, hctx_idx
);
436 blk_mq_free_rq_map(hctx
->sched_tags
);
437 hctx
->sched_tags
= NULL
;
441 static int blk_mq_sched_alloc_tags(struct request_queue
*q
,
442 struct blk_mq_hw_ctx
*hctx
,
443 unsigned int hctx_idx
)
445 struct blk_mq_tag_set
*set
= q
->tag_set
;
448 hctx
->sched_tags
= blk_mq_alloc_rq_map(set
, hctx_idx
, q
->nr_requests
,
450 if (!hctx
->sched_tags
)
453 ret
= blk_mq_alloc_rqs(set
, hctx
->sched_tags
, hctx_idx
, q
->nr_requests
);
455 blk_mq_sched_free_tags(set
, hctx
, hctx_idx
);
460 static void blk_mq_sched_tags_teardown(struct request_queue
*q
)
462 struct blk_mq_tag_set
*set
= q
->tag_set
;
463 struct blk_mq_hw_ctx
*hctx
;
466 queue_for_each_hw_ctx(q
, hctx
, i
)
467 blk_mq_sched_free_tags(set
, hctx
, i
);
470 int blk_mq_init_sched(struct request_queue
*q
, struct elevator_type
*e
)
472 struct blk_mq_hw_ctx
*hctx
;
473 struct elevator_queue
*eq
;
479 q
->nr_requests
= q
->tag_set
->queue_depth
;
484 * Default to double of smaller one between hw queue_depth and 128,
485 * since we don't split into sync/async like the old code did.
486 * Additionally, this is a per-hw queue depth.
488 q
->nr_requests
= 2 * min_t(unsigned int, q
->tag_set
->queue_depth
,
491 queue_for_each_hw_ctx(q
, hctx
, i
) {
492 ret
= blk_mq_sched_alloc_tags(q
, hctx
, i
);
497 ret
= e
->ops
.init_sched(q
, e
);
501 blk_mq_debugfs_register_sched(q
);
503 queue_for_each_hw_ctx(q
, hctx
, i
) {
504 if (e
->ops
.init_hctx
) {
505 ret
= e
->ops
.init_hctx(hctx
, i
);
508 blk_mq_exit_sched(q
, eq
);
509 kobject_put(&eq
->kobj
);
513 blk_mq_debugfs_register_sched_hctx(q
, hctx
);
519 blk_mq_sched_tags_teardown(q
);
524 void blk_mq_exit_sched(struct request_queue
*q
, struct elevator_queue
*e
)
526 struct blk_mq_hw_ctx
*hctx
;
529 queue_for_each_hw_ctx(q
, hctx
, i
) {
530 blk_mq_debugfs_unregister_sched_hctx(hctx
);
531 if (e
->type
->ops
.exit_hctx
&& hctx
->sched_data
) {
532 e
->type
->ops
.exit_hctx(hctx
, i
);
533 hctx
->sched_data
= NULL
;
536 blk_mq_debugfs_unregister_sched(q
);
537 if (e
->type
->ops
.exit_sched
)
538 e
->type
->ops
.exit_sched(e
);
539 blk_mq_sched_tags_teardown(q
);