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
);
72 * Mark a hardware queue as needing a restart. For shared queues, maintain
73 * a count of how many hardware queues are marked for restart.
75 static void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx
*hctx
)
77 if (test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
80 if (hctx
->flags
& BLK_MQ_F_TAG_SHARED
) {
81 struct request_queue
*q
= hctx
->queue
;
83 if (!test_and_set_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
84 atomic_inc(&q
->shared_hctx_restart
);
86 set_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
);
89 static bool blk_mq_sched_restart_hctx(struct blk_mq_hw_ctx
*hctx
)
91 if (!test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
94 if (hctx
->flags
& BLK_MQ_F_TAG_SHARED
) {
95 struct request_queue
*q
= hctx
->queue
;
97 if (test_and_clear_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
98 atomic_dec(&q
->shared_hctx_restart
);
100 clear_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
);
102 if (blk_mq_hctx_has_pending(hctx
)) {
103 blk_mq_run_hw_queue(hctx
, true);
110 struct request
*blk_mq_sched_get_request(struct request_queue
*q
,
113 struct blk_mq_alloc_data
*data
)
115 struct elevator_queue
*e
= q
->elevator
;
118 blk_queue_enter_live(q
);
120 if (likely(!data
->ctx
))
121 data
->ctx
= blk_mq_get_ctx(q
);
122 if (likely(!data
->hctx
))
123 data
->hctx
= blk_mq_map_queue(q
, data
->ctx
->cpu
);
126 data
->flags
|= BLK_MQ_REQ_INTERNAL
;
129 * Flush requests are special and go directly to the
132 if (!op_is_flush(op
) && e
->type
->ops
.mq
.get_request
) {
133 rq
= e
->type
->ops
.mq
.get_request(q
, op
, data
);
135 rq
->rq_flags
|= RQF_QUEUED
;
137 rq
= __blk_mq_alloc_request(data
, op
);
139 rq
= __blk_mq_alloc_request(data
, op
);
143 if (!op_is_flush(op
)) {
145 if (e
&& e
->type
->icq_cache
)
146 blk_mq_sched_assign_ioc(q
, rq
, bio
);
148 data
->hctx
->queued
++;
156 void blk_mq_sched_put_request(struct request
*rq
)
158 struct request_queue
*q
= rq
->q
;
159 struct elevator_queue
*e
= q
->elevator
;
161 if (rq
->rq_flags
& RQF_ELVPRIV
) {
162 blk_mq_sched_put_rq_priv(rq
->q
, rq
);
164 put_io_context(rq
->elv
.icq
->ioc
);
169 if ((rq
->rq_flags
& RQF_QUEUED
) && e
&& e
->type
->ops
.mq
.put_request
)
170 e
->type
->ops
.mq
.put_request(rq
);
172 blk_mq_finish_request(rq
);
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
;
180 bool did_work
= false;
183 if (unlikely(blk_mq_hctx_stopped(hctx
)))
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 if (!list_empty(&rq_list
)) {
209 blk_mq_sched_mark_restart_hctx(hctx
);
210 did_work
= blk_mq_dispatch_rq_list(q
, &rq_list
);
211 } else if (!has_sched_dispatch
) {
212 blk_mq_flush_busy_ctxs(hctx
, &rq_list
);
213 blk_mq_dispatch_rq_list(q
, &rq_list
);
217 * We want to dispatch from the scheduler if we had no work left
218 * on the dispatch list, OR if we did have work but weren't able
221 if (!did_work
&& has_sched_dispatch
) {
225 rq
= e
->type
->ops
.mq
.dispatch_request(hctx
);
228 list_add(&rq
->queuelist
, &rq_list
);
229 } while (blk_mq_dispatch_rq_list(q
, &rq_list
));
233 bool blk_mq_sched_try_merge(struct request_queue
*q
, struct bio
*bio
,
234 struct request
**merged_request
)
238 switch (elv_merge(q
, &rq
, bio
)) {
239 case ELEVATOR_BACK_MERGE
:
240 if (!blk_mq_sched_allow_merge(q
, rq
, bio
))
242 if (!bio_attempt_back_merge(q
, rq
, bio
))
244 *merged_request
= attempt_back_merge(q
, rq
);
245 if (!*merged_request
)
246 elv_merged_request(q
, rq
, ELEVATOR_BACK_MERGE
);
248 case ELEVATOR_FRONT_MERGE
:
249 if (!blk_mq_sched_allow_merge(q
, rq
, bio
))
251 if (!bio_attempt_front_merge(q
, rq
, bio
))
253 *merged_request
= attempt_front_merge(q
, rq
);
254 if (!*merged_request
)
255 elv_merged_request(q
, rq
, ELEVATOR_FRONT_MERGE
);
261 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge
);
263 bool __blk_mq_sched_bio_merge(struct request_queue
*q
, struct bio
*bio
)
265 struct elevator_queue
*e
= q
->elevator
;
267 if (e
->type
->ops
.mq
.bio_merge
) {
268 struct blk_mq_ctx
*ctx
= blk_mq_get_ctx(q
);
269 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
272 return e
->type
->ops
.mq
.bio_merge(hctx
, bio
);
278 bool blk_mq_sched_try_insert_merge(struct request_queue
*q
, struct request
*rq
)
280 return rq_mergeable(rq
) && elv_attempt_insert_merge(q
, rq
);
282 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge
);
284 void blk_mq_sched_request_inserted(struct request
*rq
)
286 trace_block_rq_insert(rq
->q
, rq
);
288 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted
);
290 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx
*hctx
,
294 rq
->rq_flags
|= RQF_SORTED
;
299 * If we already have a real request tag, send directly to
302 spin_lock(&hctx
->lock
);
303 list_add(&rq
->queuelist
, &hctx
->dispatch
);
304 spin_unlock(&hctx
->lock
);
309 * list_for_each_entry_rcu_rr - iterate in a round-robin fashion over rcu list
311 * @skip: the list element that will not be examined. Iteration starts at
313 * @head: head of the list to examine. This list must have at least one
314 * element, namely @skip.
315 * @member: name of the list_head structure within typeof(*pos).
317 #define list_for_each_entry_rcu_rr(pos, skip, head, member) \
318 for ((pos) = (skip); \
319 (pos = (pos)->member.next != (head) ? list_entry_rcu( \
320 (pos)->member.next, typeof(*pos), member) : \
321 list_entry_rcu((pos)->member.next->next, typeof(*pos), member)), \
325 * Called after a driver tag has been freed to check whether a hctx needs to
326 * be restarted. Restarts @hctx if its tag set is not shared. Restarts hardware
327 * queues in a round-robin fashion if the tag set of @hctx is shared with other
330 void blk_mq_sched_restart(struct blk_mq_hw_ctx
*const hctx
)
332 struct blk_mq_tags
*const tags
= hctx
->tags
;
333 struct blk_mq_tag_set
*const set
= hctx
->queue
->tag_set
;
334 struct request_queue
*const queue
= hctx
->queue
, *q
;
335 struct blk_mq_hw_ctx
*hctx2
;
338 if (set
->flags
& BLK_MQ_F_TAG_SHARED
) {
340 * If this is 0, then we know that no hardware queues
341 * have RESTART marked. We're done.
343 if (!atomic_read(&queue
->shared_hctx_restart
))
347 list_for_each_entry_rcu_rr(q
, queue
, &set
->tag_list
,
349 queue_for_each_hw_ctx(q
, hctx2
, i
)
350 if (hctx2
->tags
== tags
&&
351 blk_mq_sched_restart_hctx(hctx2
))
354 j
= hctx
->queue_num
+ 1;
355 for (i
= 0; i
< queue
->nr_hw_queues
; i
++, j
++) {
356 if (j
== queue
->nr_hw_queues
)
358 hctx2
= queue
->queue_hw_ctx
[j
];
359 if (hctx2
->tags
== tags
&&
360 blk_mq_sched_restart_hctx(hctx2
))
366 blk_mq_sched_restart_hctx(hctx
);
371 * Add flush/fua to the queue. If we fail getting a driver tag, then
372 * punt to the requeue list. Requeue will re-invoke us from a context
373 * that's safe to block from.
375 static void blk_mq_sched_insert_flush(struct blk_mq_hw_ctx
*hctx
,
376 struct request
*rq
, bool can_block
)
378 if (blk_mq_get_driver_tag(rq
, &hctx
, can_block
)) {
379 blk_insert_flush(rq
);
380 blk_mq_run_hw_queue(hctx
, true);
382 blk_mq_add_to_requeue_list(rq
, false, true);
385 void blk_mq_sched_insert_request(struct request
*rq
, bool at_head
,
386 bool run_queue
, bool async
, bool can_block
)
388 struct request_queue
*q
= rq
->q
;
389 struct elevator_queue
*e
= q
->elevator
;
390 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
391 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
393 if (rq
->tag
== -1 && op_is_flush(rq
->cmd_flags
)) {
394 blk_mq_sched_insert_flush(hctx
, rq
, can_block
);
398 if (e
&& blk_mq_sched_bypass_insert(hctx
, rq
))
401 if (e
&& e
->type
->ops
.mq
.insert_requests
) {
404 list_add(&rq
->queuelist
, &list
);
405 e
->type
->ops
.mq
.insert_requests(hctx
, &list
, at_head
);
407 spin_lock(&ctx
->lock
);
408 __blk_mq_insert_request(hctx
, rq
, at_head
);
409 spin_unlock(&ctx
->lock
);
414 blk_mq_run_hw_queue(hctx
, async
);
417 void blk_mq_sched_insert_requests(struct request_queue
*q
,
418 struct blk_mq_ctx
*ctx
,
419 struct list_head
*list
, bool run_queue_async
)
421 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
422 struct elevator_queue
*e
= hctx
->queue
->elevator
;
425 struct request
*rq
, *next
;
428 * We bypass requests that already have a driver tag assigned,
429 * which should only be flushes. Flushes are only ever inserted
430 * as single requests, so we shouldn't ever hit the
431 * WARN_ON_ONCE() below (but let's handle it just in case).
433 list_for_each_entry_safe(rq
, next
, list
, queuelist
) {
434 if (WARN_ON_ONCE(rq
->tag
!= -1)) {
435 list_del_init(&rq
->queuelist
);
436 blk_mq_sched_bypass_insert(hctx
, rq
);
441 if (e
&& e
->type
->ops
.mq
.insert_requests
)
442 e
->type
->ops
.mq
.insert_requests(hctx
, list
, false);
444 blk_mq_insert_requests(hctx
, ctx
, list
);
446 blk_mq_run_hw_queue(hctx
, run_queue_async
);
449 static void blk_mq_sched_free_tags(struct blk_mq_tag_set
*set
,
450 struct blk_mq_hw_ctx
*hctx
,
451 unsigned int hctx_idx
)
453 if (hctx
->sched_tags
) {
454 blk_mq_free_rqs(set
, hctx
->sched_tags
, hctx_idx
);
455 blk_mq_free_rq_map(hctx
->sched_tags
);
456 hctx
->sched_tags
= NULL
;
460 static int blk_mq_sched_alloc_tags(struct request_queue
*q
,
461 struct blk_mq_hw_ctx
*hctx
,
462 unsigned int hctx_idx
)
464 struct blk_mq_tag_set
*set
= q
->tag_set
;
467 hctx
->sched_tags
= blk_mq_alloc_rq_map(set
, hctx_idx
, q
->nr_requests
,
469 if (!hctx
->sched_tags
)
472 ret
= blk_mq_alloc_rqs(set
, hctx
->sched_tags
, hctx_idx
, q
->nr_requests
);
474 blk_mq_sched_free_tags(set
, hctx
, hctx_idx
);
479 static void blk_mq_sched_tags_teardown(struct request_queue
*q
)
481 struct blk_mq_tag_set
*set
= q
->tag_set
;
482 struct blk_mq_hw_ctx
*hctx
;
485 queue_for_each_hw_ctx(q
, hctx
, i
)
486 blk_mq_sched_free_tags(set
, hctx
, i
);
489 int blk_mq_sched_init_hctx(struct request_queue
*q
, struct blk_mq_hw_ctx
*hctx
,
490 unsigned int hctx_idx
)
492 struct elevator_queue
*e
= q
->elevator
;
498 ret
= blk_mq_sched_alloc_tags(q
, hctx
, hctx_idx
);
502 if (e
->type
->ops
.mq
.init_hctx
) {
503 ret
= e
->type
->ops
.mq
.init_hctx(hctx
, hctx_idx
);
505 blk_mq_sched_free_tags(q
->tag_set
, hctx
, hctx_idx
);
510 blk_mq_debugfs_register_sched_hctx(q
, hctx
);
515 void blk_mq_sched_exit_hctx(struct request_queue
*q
, struct blk_mq_hw_ctx
*hctx
,
516 unsigned int hctx_idx
)
518 struct elevator_queue
*e
= q
->elevator
;
523 blk_mq_debugfs_unregister_sched_hctx(hctx
);
525 if (e
->type
->ops
.mq
.exit_hctx
&& hctx
->sched_data
) {
526 e
->type
->ops
.mq
.exit_hctx(hctx
, hctx_idx
);
527 hctx
->sched_data
= NULL
;
530 blk_mq_sched_free_tags(q
->tag_set
, hctx
, hctx_idx
);
533 int blk_mq_init_sched(struct request_queue
*q
, struct elevator_type
*e
)
535 struct blk_mq_hw_ctx
*hctx
;
536 struct elevator_queue
*eq
;
546 * Default to 256, since we don't split into sync/async like the
547 * old code did. Additionally, this is a per-hw queue depth.
549 q
->nr_requests
= 2 * BLKDEV_MAX_RQ
;
551 queue_for_each_hw_ctx(q
, hctx
, i
) {
552 ret
= blk_mq_sched_alloc_tags(q
, hctx
, i
);
557 ret
= e
->ops
.mq
.init_sched(q
, e
);
561 blk_mq_debugfs_register_sched(q
);
563 queue_for_each_hw_ctx(q
, hctx
, i
) {
564 if (e
->ops
.mq
.init_hctx
) {
565 ret
= e
->ops
.mq
.init_hctx(hctx
, i
);
568 blk_mq_exit_sched(q
, eq
);
569 kobject_put(&eq
->kobj
);
573 blk_mq_debugfs_register_sched_hctx(q
, hctx
);
579 blk_mq_sched_tags_teardown(q
);
584 void blk_mq_exit_sched(struct request_queue
*q
, struct elevator_queue
*e
)
586 struct blk_mq_hw_ctx
*hctx
;
589 queue_for_each_hw_ctx(q
, hctx
, i
) {
590 blk_mq_debugfs_unregister_sched_hctx(hctx
);
591 if (e
->type
->ops
.mq
.exit_hctx
&& hctx
->sched_data
) {
592 e
->type
->ops
.mq
.exit_hctx(hctx
, i
);
593 hctx
->sched_data
= NULL
;
596 blk_mq_debugfs_unregister_sched(q
);
597 if (e
->type
->ops
.mq
.exit_sched
)
598 e
->type
->ops
.mq
.exit_sched(e
);
599 blk_mq_sched_tags_teardown(q
);
603 int blk_mq_sched_init(struct request_queue
*q
)
607 mutex_lock(&q
->sysfs_lock
);
608 ret
= elevator_init(q
, NULL
);
609 mutex_unlock(&q
->sysfs_lock
);