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 static bool blk_mq_sched_restart_hctx(struct blk_mq_hw_ctx
*hctx
)
73 if (!test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
76 if (hctx
->flags
& BLK_MQ_F_TAG_SHARED
) {
77 struct request_queue
*q
= hctx
->queue
;
79 if (test_and_clear_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
80 atomic_dec(&q
->shared_hctx_restart
);
82 clear_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
);
84 return blk_mq_run_hw_queue(hctx
, true);
88 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
89 * its queue by itself in its completion handler, so we don't need to
90 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
92 static void blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx
*hctx
)
94 struct request_queue
*q
= hctx
->queue
;
95 struct elevator_queue
*e
= q
->elevator
;
101 if (e
->type
->ops
.mq
.has_work
&&
102 !e
->type
->ops
.mq
.has_work(hctx
))
105 if (!blk_mq_get_dispatch_budget(hctx
))
108 rq
= e
->type
->ops
.mq
.dispatch_request(hctx
);
110 blk_mq_put_dispatch_budget(hctx
);
115 * Now this rq owns the budget which has to be released
116 * if this rq won't be queued to driver via .queue_rq()
117 * in blk_mq_dispatch_rq_list().
119 list_add(&rq
->queuelist
, &rq_list
);
120 } while (blk_mq_dispatch_rq_list(q
, &rq_list
, true));
123 static struct blk_mq_ctx
*blk_mq_next_ctx(struct blk_mq_hw_ctx
*hctx
,
124 struct blk_mq_ctx
*ctx
)
126 unsigned idx
= ctx
->index_hw
;
128 if (++idx
== hctx
->nr_ctx
)
131 return hctx
->ctxs
[idx
];
135 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
136 * its queue by itself in its completion handler, so we don't need to
137 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
139 static void blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx
*hctx
)
141 struct request_queue
*q
= hctx
->queue
;
143 struct blk_mq_ctx
*ctx
= READ_ONCE(hctx
->dispatch_from
);
148 if (!sbitmap_any_bit_set(&hctx
->ctx_map
))
151 if (!blk_mq_get_dispatch_budget(hctx
))
154 rq
= blk_mq_dequeue_from_ctx(hctx
, ctx
);
156 blk_mq_put_dispatch_budget(hctx
);
161 * Now this rq owns the budget which has to be released
162 * if this rq won't be queued to driver via .queue_rq()
163 * in blk_mq_dispatch_rq_list().
165 list_add(&rq
->queuelist
, &rq_list
);
167 /* round robin for fair dispatch */
168 ctx
= blk_mq_next_ctx(hctx
, rq
->mq_ctx
);
170 } while (blk_mq_dispatch_rq_list(q
, &rq_list
, true));
172 WRITE_ONCE(hctx
->dispatch_from
, ctx
);
175 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx
*hctx
)
177 struct request_queue
*q
= hctx
->queue
;
178 struct elevator_queue
*e
= q
->elevator
;
179 const bool has_sched_dispatch
= e
&& e
->type
->ops
.mq
.dispatch_request
;
182 /* RCU or SRCU read lock is needed before checking quiesced flag */
183 if (unlikely(blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(q
)))
189 * If we have previous entries on our dispatch list, grab them first for
190 * more fair dispatch.
192 if (!list_empty_careful(&hctx
->dispatch
)) {
193 spin_lock(&hctx
->lock
);
194 if (!list_empty(&hctx
->dispatch
))
195 list_splice_init(&hctx
->dispatch
, &rq_list
);
196 spin_unlock(&hctx
->lock
);
200 * Only ask the scheduler for requests, if we didn't have residual
201 * requests from the dispatch list. This is to avoid the case where
202 * we only ever dispatch a fraction of the requests available because
203 * of low device queue depth. Once we pull requests out of the IO
204 * scheduler, we can no longer merge or sort them. So it's best to
205 * leave them there for as long as we can. Mark the hw queue as
206 * needing a restart in that case.
208 * We want to dispatch from the scheduler if there was nothing
209 * on the dispatch list or we were able to dispatch from the
212 if (!list_empty(&rq_list
)) {
213 blk_mq_sched_mark_restart_hctx(hctx
);
214 if (blk_mq_dispatch_rq_list(q
, &rq_list
, false)) {
215 if (has_sched_dispatch
)
216 blk_mq_do_dispatch_sched(hctx
);
218 blk_mq_do_dispatch_ctx(hctx
);
220 } else if (has_sched_dispatch
) {
221 blk_mq_do_dispatch_sched(hctx
);
222 } else if (q
->mq_ops
->get_budget
) {
224 * If we need to get budget before queuing request, we
225 * dequeue request one by one from sw queue for avoiding
226 * to mess up I/O merge when dispatch runs out of resource.
228 * TODO: get more budgets, and dequeue more requests in
231 blk_mq_do_dispatch_ctx(hctx
);
233 blk_mq_flush_busy_ctxs(hctx
, &rq_list
);
234 blk_mq_dispatch_rq_list(q
, &rq_list
, false);
238 bool blk_mq_sched_try_merge(struct request_queue
*q
, struct bio
*bio
,
239 struct request
**merged_request
)
243 switch (elv_merge(q
, &rq
, bio
)) {
244 case ELEVATOR_BACK_MERGE
:
245 if (!blk_mq_sched_allow_merge(q
, rq
, bio
))
247 if (!bio_attempt_back_merge(q
, rq
, bio
))
249 *merged_request
= attempt_back_merge(q
, rq
);
250 if (!*merged_request
)
251 elv_merged_request(q
, rq
, ELEVATOR_BACK_MERGE
);
253 case ELEVATOR_FRONT_MERGE
:
254 if (!blk_mq_sched_allow_merge(q
, rq
, bio
))
256 if (!bio_attempt_front_merge(q
, rq
, bio
))
258 *merged_request
= attempt_front_merge(q
, rq
);
259 if (!*merged_request
)
260 elv_merged_request(q
, rq
, ELEVATOR_FRONT_MERGE
);
262 case ELEVATOR_DISCARD_MERGE
:
263 return bio_attempt_discard_merge(q
, rq
, bio
);
268 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge
);
271 * Reverse check our software queue for entries that we could potentially
272 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
273 * too much time checking for merges.
275 static bool blk_mq_attempt_merge(struct request_queue
*q
,
276 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
281 lockdep_assert_held(&ctx
->lock
);
283 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
289 if (!blk_rq_merge_ok(rq
, bio
))
292 switch (blk_try_merge(rq
, bio
)) {
293 case ELEVATOR_BACK_MERGE
:
294 if (blk_mq_sched_allow_merge(q
, rq
, bio
))
295 merged
= bio_attempt_back_merge(q
, rq
, bio
);
297 case ELEVATOR_FRONT_MERGE
:
298 if (blk_mq_sched_allow_merge(q
, rq
, bio
))
299 merged
= bio_attempt_front_merge(q
, rq
, bio
);
301 case ELEVATOR_DISCARD_MERGE
:
302 merged
= bio_attempt_discard_merge(q
, rq
, bio
);
316 bool __blk_mq_sched_bio_merge(struct request_queue
*q
, struct bio
*bio
)
318 struct elevator_queue
*e
= q
->elevator
;
319 struct blk_mq_ctx
*ctx
= blk_mq_get_ctx(q
);
320 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
323 if (e
&& e
->type
->ops
.mq
.bio_merge
) {
325 return e
->type
->ops
.mq
.bio_merge(hctx
, bio
);
328 if (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) {
329 /* default per sw-queue merge */
330 spin_lock(&ctx
->lock
);
331 ret
= blk_mq_attempt_merge(q
, ctx
, bio
);
332 spin_unlock(&ctx
->lock
);
339 bool blk_mq_sched_try_insert_merge(struct request_queue
*q
, struct request
*rq
)
341 return rq_mergeable(rq
) && elv_attempt_insert_merge(q
, rq
);
343 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge
);
345 void blk_mq_sched_request_inserted(struct request
*rq
)
347 trace_block_rq_insert(rq
->q
, rq
);
349 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted
);
351 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx
*hctx
,
355 /* dispatch flush rq directly */
356 if (rq
->rq_flags
& RQF_FLUSH_SEQ
) {
357 spin_lock(&hctx
->lock
);
358 list_add(&rq
->queuelist
, &hctx
->dispatch
);
359 spin_unlock(&hctx
->lock
);
364 rq
->rq_flags
|= RQF_SORTED
;
370 * list_for_each_entry_rcu_rr - iterate in a round-robin fashion over rcu list
372 * @skip: the list element that will not be examined. Iteration starts at
374 * @head: head of the list to examine. This list must have at least one
375 * element, namely @skip.
376 * @member: name of the list_head structure within typeof(*pos).
378 #define list_for_each_entry_rcu_rr(pos, skip, head, member) \
379 for ((pos) = (skip); \
380 (pos = (pos)->member.next != (head) ? list_entry_rcu( \
381 (pos)->member.next, typeof(*pos), member) : \
382 list_entry_rcu((pos)->member.next->next, typeof(*pos), member)), \
386 * Called after a driver tag has been freed to check whether a hctx needs to
387 * be restarted. Restarts @hctx if its tag set is not shared. Restarts hardware
388 * queues in a round-robin fashion if the tag set of @hctx is shared with other
391 void blk_mq_sched_restart(struct blk_mq_hw_ctx
*const hctx
)
393 struct blk_mq_tags
*const tags
= hctx
->tags
;
394 struct blk_mq_tag_set
*const set
= hctx
->queue
->tag_set
;
395 struct request_queue
*const queue
= hctx
->queue
, *q
;
396 struct blk_mq_hw_ctx
*hctx2
;
399 if (set
->flags
& BLK_MQ_F_TAG_SHARED
) {
401 * If this is 0, then we know that no hardware queues
402 * have RESTART marked. We're done.
404 if (!atomic_read(&queue
->shared_hctx_restart
))
408 list_for_each_entry_rcu_rr(q
, queue
, &set
->tag_list
,
410 queue_for_each_hw_ctx(q
, hctx2
, i
)
411 if (hctx2
->tags
== tags
&&
412 blk_mq_sched_restart_hctx(hctx2
))
415 j
= hctx
->queue_num
+ 1;
416 for (i
= 0; i
< queue
->nr_hw_queues
; i
++, j
++) {
417 if (j
== queue
->nr_hw_queues
)
419 hctx2
= queue
->queue_hw_ctx
[j
];
420 if (hctx2
->tags
== tags
&&
421 blk_mq_sched_restart_hctx(hctx2
))
427 blk_mq_sched_restart_hctx(hctx
);
431 void blk_mq_sched_insert_request(struct request
*rq
, bool at_head
,
432 bool run_queue
, bool async
)
434 struct request_queue
*q
= rq
->q
;
435 struct elevator_queue
*e
= q
->elevator
;
436 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
437 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
439 /* flush rq in flush machinery need to be dispatched directly */
440 if (!(rq
->rq_flags
& RQF_FLUSH_SEQ
) && op_is_flush(rq
->cmd_flags
)) {
441 blk_insert_flush(rq
);
445 WARN_ON(e
&& (rq
->tag
!= -1));
447 if (blk_mq_sched_bypass_insert(hctx
, !!e
, rq
))
450 if (e
&& e
->type
->ops
.mq
.insert_requests
) {
453 list_add(&rq
->queuelist
, &list
);
454 e
->type
->ops
.mq
.insert_requests(hctx
, &list
, at_head
);
456 spin_lock(&ctx
->lock
);
457 __blk_mq_insert_request(hctx
, rq
, at_head
);
458 spin_unlock(&ctx
->lock
);
463 blk_mq_run_hw_queue(hctx
, async
);
466 void blk_mq_sched_insert_requests(struct request_queue
*q
,
467 struct blk_mq_ctx
*ctx
,
468 struct list_head
*list
, bool run_queue_async
)
470 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
471 struct elevator_queue
*e
= hctx
->queue
->elevator
;
473 if (e
&& e
->type
->ops
.mq
.insert_requests
)
474 e
->type
->ops
.mq
.insert_requests(hctx
, list
, false);
476 blk_mq_insert_requests(hctx
, ctx
, list
);
478 blk_mq_run_hw_queue(hctx
, run_queue_async
);
481 static void blk_mq_sched_free_tags(struct blk_mq_tag_set
*set
,
482 struct blk_mq_hw_ctx
*hctx
,
483 unsigned int hctx_idx
)
485 if (hctx
->sched_tags
) {
486 blk_mq_free_rqs(set
, hctx
->sched_tags
, hctx_idx
);
487 blk_mq_free_rq_map(hctx
->sched_tags
);
488 hctx
->sched_tags
= NULL
;
492 static int blk_mq_sched_alloc_tags(struct request_queue
*q
,
493 struct blk_mq_hw_ctx
*hctx
,
494 unsigned int hctx_idx
)
496 struct blk_mq_tag_set
*set
= q
->tag_set
;
499 hctx
->sched_tags
= blk_mq_alloc_rq_map(set
, hctx_idx
, q
->nr_requests
,
501 if (!hctx
->sched_tags
)
504 ret
= blk_mq_alloc_rqs(set
, hctx
->sched_tags
, hctx_idx
, q
->nr_requests
);
506 blk_mq_sched_free_tags(set
, hctx
, hctx_idx
);
511 static void blk_mq_sched_tags_teardown(struct request_queue
*q
)
513 struct blk_mq_tag_set
*set
= q
->tag_set
;
514 struct blk_mq_hw_ctx
*hctx
;
517 queue_for_each_hw_ctx(q
, hctx
, i
)
518 blk_mq_sched_free_tags(set
, hctx
, i
);
521 int blk_mq_sched_init_hctx(struct request_queue
*q
, struct blk_mq_hw_ctx
*hctx
,
522 unsigned int hctx_idx
)
524 struct elevator_queue
*e
= q
->elevator
;
530 ret
= blk_mq_sched_alloc_tags(q
, hctx
, hctx_idx
);
534 if (e
->type
->ops
.mq
.init_hctx
) {
535 ret
= e
->type
->ops
.mq
.init_hctx(hctx
, hctx_idx
);
537 blk_mq_sched_free_tags(q
->tag_set
, hctx
, hctx_idx
);
542 blk_mq_debugfs_register_sched_hctx(q
, hctx
);
547 void blk_mq_sched_exit_hctx(struct request_queue
*q
, struct blk_mq_hw_ctx
*hctx
,
548 unsigned int hctx_idx
)
550 struct elevator_queue
*e
= q
->elevator
;
555 blk_mq_debugfs_unregister_sched_hctx(hctx
);
557 if (e
->type
->ops
.mq
.exit_hctx
&& hctx
->sched_data
) {
558 e
->type
->ops
.mq
.exit_hctx(hctx
, hctx_idx
);
559 hctx
->sched_data
= NULL
;
562 blk_mq_sched_free_tags(q
->tag_set
, hctx
, hctx_idx
);
565 int blk_mq_init_sched(struct request_queue
*q
, struct elevator_type
*e
)
567 struct blk_mq_hw_ctx
*hctx
;
568 struct elevator_queue
*eq
;
578 * Default to double of smaller one between hw queue_depth and 128,
579 * since we don't split into sync/async like the old code did.
580 * Additionally, this is a per-hw queue depth.
582 q
->nr_requests
= 2 * min_t(unsigned int, q
->tag_set
->queue_depth
,
585 queue_for_each_hw_ctx(q
, hctx
, i
) {
586 ret
= blk_mq_sched_alloc_tags(q
, hctx
, i
);
591 ret
= e
->ops
.mq
.init_sched(q
, e
);
595 blk_mq_debugfs_register_sched(q
);
597 queue_for_each_hw_ctx(q
, hctx
, i
) {
598 if (e
->ops
.mq
.init_hctx
) {
599 ret
= e
->ops
.mq
.init_hctx(hctx
, i
);
602 blk_mq_exit_sched(q
, eq
);
603 kobject_put(&eq
->kobj
);
607 blk_mq_debugfs_register_sched_hctx(q
, hctx
);
613 blk_mq_sched_tags_teardown(q
);
618 void blk_mq_exit_sched(struct request_queue
*q
, struct elevator_queue
*e
)
620 struct blk_mq_hw_ctx
*hctx
;
623 queue_for_each_hw_ctx(q
, hctx
, i
) {
624 blk_mq_debugfs_unregister_sched_hctx(hctx
);
625 if (e
->type
->ops
.mq
.exit_hctx
&& hctx
->sched_data
) {
626 e
->type
->ops
.mq
.exit_hctx(hctx
, i
);
627 hctx
->sched_data
= NULL
;
630 blk_mq_debugfs_unregister_sched(q
);
631 if (e
->type
->ops
.mq
.exit_sched
)
632 e
->type
->ops
.mq
.exit_sched(e
);
633 blk_mq_sched_tags_teardown(q
);
637 int blk_mq_sched_init(struct request_queue
*q
)
641 mutex_lock(&q
->sysfs_lock
);
642 ret
= elevator_init(q
, NULL
);
643 mutex_unlock(&q
->sysfs_lock
);