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-sched.h"
15 #include "blk-mq-tag.h"
18 void blk_mq_sched_free_hctx_data(struct request_queue
*q
,
19 void (*exit
)(struct blk_mq_hw_ctx
*))
21 struct blk_mq_hw_ctx
*hctx
;
24 queue_for_each_hw_ctx(q
, hctx
, i
) {
25 if (exit
&& hctx
->sched_data
)
27 kfree(hctx
->sched_data
);
28 hctx
->sched_data
= NULL
;
31 EXPORT_SYMBOL_GPL(blk_mq_sched_free_hctx_data
);
33 int blk_mq_sched_init_hctx_data(struct request_queue
*q
, size_t size
,
34 int (*init
)(struct blk_mq_hw_ctx
*),
35 void (*exit
)(struct blk_mq_hw_ctx
*))
37 struct blk_mq_hw_ctx
*hctx
;
41 queue_for_each_hw_ctx(q
, hctx
, i
) {
42 hctx
->sched_data
= kmalloc_node(size
, GFP_KERNEL
, hctx
->numa_node
);
43 if (!hctx
->sched_data
) {
52 * We don't want to give exit() a partially
53 * initialized sched_data. init() must clean up
56 kfree(hctx
->sched_data
);
57 hctx
->sched_data
= NULL
;
65 blk_mq_sched_free_hctx_data(q
, exit
);
68 EXPORT_SYMBOL_GPL(blk_mq_sched_init_hctx_data
);
70 static void __blk_mq_sched_assign_ioc(struct request_queue
*q
,
73 struct io_context
*ioc
)
77 spin_lock_irq(q
->queue_lock
);
78 icq
= ioc_lookup_icq(ioc
, q
);
79 spin_unlock_irq(q
->queue_lock
);
82 icq
= ioc_create_icq(ioc
, q
, GFP_ATOMIC
);
88 if (!blk_mq_sched_get_rq_priv(q
, rq
, bio
)) {
89 rq
->rq_flags
|= RQF_ELVPRIV
;
90 get_io_context(icq
->ioc
);
97 static void blk_mq_sched_assign_ioc(struct request_queue
*q
,
98 struct request
*rq
, struct bio
*bio
)
100 struct io_context
*ioc
;
104 __blk_mq_sched_assign_ioc(q
, rq
, bio
, ioc
);
107 struct request
*blk_mq_sched_get_request(struct request_queue
*q
,
110 struct blk_mq_alloc_data
*data
)
112 struct elevator_queue
*e
= q
->elevator
;
115 blk_queue_enter_live(q
);
117 if (likely(!data
->ctx
))
118 data
->ctx
= blk_mq_get_ctx(q
);
119 if (likely(!data
->hctx
))
120 data
->hctx
= blk_mq_map_queue(q
, data
->ctx
->cpu
);
123 data
->flags
|= BLK_MQ_REQ_INTERNAL
;
126 * Flush requests are special and go directly to the
129 if (!op_is_flush(op
) && e
->type
->ops
.mq
.get_request
) {
130 rq
= e
->type
->ops
.mq
.get_request(q
, op
, data
);
132 rq
->rq_flags
|= RQF_QUEUED
;
134 rq
= __blk_mq_alloc_request(data
, op
);
136 rq
= __blk_mq_alloc_request(data
, op
);
140 if (!op_is_flush(op
)) {
142 if (e
&& e
->type
->icq_cache
)
143 blk_mq_sched_assign_ioc(q
, rq
, bio
);
145 data
->hctx
->queued
++;
153 void blk_mq_sched_put_request(struct request
*rq
)
155 struct request_queue
*q
= rq
->q
;
156 struct elevator_queue
*e
= q
->elevator
;
158 if (rq
->rq_flags
& RQF_ELVPRIV
) {
159 blk_mq_sched_put_rq_priv(rq
->q
, rq
);
161 put_io_context(rq
->elv
.icq
->ioc
);
166 if ((rq
->rq_flags
& RQF_QUEUED
) && e
&& e
->type
->ops
.mq
.put_request
)
167 e
->type
->ops
.mq
.put_request(rq
);
169 blk_mq_finish_request(rq
);
172 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx
*hctx
)
174 struct request_queue
*q
= hctx
->queue
;
175 struct elevator_queue
*e
= q
->elevator
;
176 const bool has_sched_dispatch
= e
&& e
->type
->ops
.mq
.dispatch_request
;
177 bool did_work
= false;
180 if (unlikely(blk_mq_hctx_stopped(hctx
)))
186 * If we have previous entries on our dispatch list, grab them first for
187 * more fair dispatch.
189 if (!list_empty_careful(&hctx
->dispatch
)) {
190 spin_lock(&hctx
->lock
);
191 if (!list_empty(&hctx
->dispatch
))
192 list_splice_init(&hctx
->dispatch
, &rq_list
);
193 spin_unlock(&hctx
->lock
);
197 * Only ask the scheduler for requests, if we didn't have residual
198 * requests from the dispatch list. This is to avoid the case where
199 * we only ever dispatch a fraction of the requests available because
200 * of low device queue depth. Once we pull requests out of the IO
201 * scheduler, we can no longer merge or sort them. So it's best to
202 * leave them there for as long as we can. Mark the hw queue as
203 * needing a restart in that case.
205 if (!list_empty(&rq_list
)) {
206 blk_mq_sched_mark_restart_hctx(hctx
);
207 did_work
= blk_mq_dispatch_rq_list(q
, &rq_list
);
208 } else if (!has_sched_dispatch
) {
209 blk_mq_flush_busy_ctxs(hctx
, &rq_list
);
210 blk_mq_dispatch_rq_list(q
, &rq_list
);
214 * We want to dispatch from the scheduler if we had no work left
215 * on the dispatch list, OR if we did have work but weren't able
218 if (!did_work
&& has_sched_dispatch
) {
222 rq
= e
->type
->ops
.mq
.dispatch_request(hctx
);
225 list_add(&rq
->queuelist
, &rq_list
);
226 } while (blk_mq_dispatch_rq_list(q
, &rq_list
));
230 void blk_mq_sched_move_to_dispatch(struct blk_mq_hw_ctx
*hctx
,
231 struct list_head
*rq_list
,
232 struct request
*(*get_rq
)(struct blk_mq_hw_ctx
*))
241 list_add_tail(&rq
->queuelist
, rq_list
);
244 EXPORT_SYMBOL_GPL(blk_mq_sched_move_to_dispatch
);
246 bool blk_mq_sched_try_merge(struct request_queue
*q
, struct bio
*bio
,
247 struct request
**merged_request
)
251 switch (elv_merge(q
, &rq
, bio
)) {
252 case ELEVATOR_BACK_MERGE
:
253 if (!blk_mq_sched_allow_merge(q
, rq
, bio
))
255 if (!bio_attempt_back_merge(q
, rq
, bio
))
257 *merged_request
= attempt_back_merge(q
, rq
);
258 if (!*merged_request
)
259 elv_merged_request(q
, rq
, ELEVATOR_BACK_MERGE
);
261 case ELEVATOR_FRONT_MERGE
:
262 if (!blk_mq_sched_allow_merge(q
, rq
, bio
))
264 if (!bio_attempt_front_merge(q
, rq
, bio
))
266 *merged_request
= attempt_front_merge(q
, rq
);
267 if (!*merged_request
)
268 elv_merged_request(q
, rq
, ELEVATOR_FRONT_MERGE
);
274 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge
);
276 bool __blk_mq_sched_bio_merge(struct request_queue
*q
, struct bio
*bio
)
278 struct elevator_queue
*e
= q
->elevator
;
280 if (e
->type
->ops
.mq
.bio_merge
) {
281 struct blk_mq_ctx
*ctx
= blk_mq_get_ctx(q
);
282 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
285 return e
->type
->ops
.mq
.bio_merge(hctx
, bio
);
291 bool blk_mq_sched_try_insert_merge(struct request_queue
*q
, struct request
*rq
)
293 return rq_mergeable(rq
) && elv_attempt_insert_merge(q
, rq
);
295 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge
);
297 void blk_mq_sched_request_inserted(struct request
*rq
)
299 trace_block_rq_insert(rq
->q
, rq
);
301 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted
);
303 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx
*hctx
,
307 rq
->rq_flags
|= RQF_SORTED
;
312 * If we already have a real request tag, send directly to
315 spin_lock(&hctx
->lock
);
316 list_add(&rq
->queuelist
, &hctx
->dispatch
);
317 spin_unlock(&hctx
->lock
);
321 static void blk_mq_sched_restart_hctx(struct blk_mq_hw_ctx
*hctx
)
323 if (test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
)) {
324 clear_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
);
325 if (blk_mq_hctx_has_pending(hctx
))
326 blk_mq_run_hw_queue(hctx
, true);
330 void blk_mq_sched_restart_queues(struct blk_mq_hw_ctx
*hctx
)
332 struct request_queue
*q
= hctx
->queue
;
335 if (test_bit(QUEUE_FLAG_RESTART
, &q
->queue_flags
)) {
336 if (test_and_clear_bit(QUEUE_FLAG_RESTART
, &q
->queue_flags
)) {
337 queue_for_each_hw_ctx(q
, hctx
, i
)
338 blk_mq_sched_restart_hctx(hctx
);
341 blk_mq_sched_restart_hctx(hctx
);
346 * Add flush/fua to the queue. If we fail getting a driver tag, then
347 * punt to the requeue list. Requeue will re-invoke us from a context
348 * that's safe to block from.
350 static void blk_mq_sched_insert_flush(struct blk_mq_hw_ctx
*hctx
,
351 struct request
*rq
, bool can_block
)
353 if (blk_mq_get_driver_tag(rq
, &hctx
, can_block
)) {
354 blk_insert_flush(rq
);
355 blk_mq_run_hw_queue(hctx
, true);
357 blk_mq_add_to_requeue_list(rq
, false, true);
360 void blk_mq_sched_insert_request(struct request
*rq
, bool at_head
,
361 bool run_queue
, bool async
, bool can_block
)
363 struct request_queue
*q
= rq
->q
;
364 struct elevator_queue
*e
= q
->elevator
;
365 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
366 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
368 if (rq
->tag
== -1 && op_is_flush(rq
->cmd_flags
)) {
369 blk_mq_sched_insert_flush(hctx
, rq
, can_block
);
373 if (e
&& blk_mq_sched_bypass_insert(hctx
, rq
))
376 if (e
&& e
->type
->ops
.mq
.insert_requests
) {
379 list_add(&rq
->queuelist
, &list
);
380 e
->type
->ops
.mq
.insert_requests(hctx
, &list
, at_head
);
382 spin_lock(&ctx
->lock
);
383 __blk_mq_insert_request(hctx
, rq
, at_head
);
384 spin_unlock(&ctx
->lock
);
389 blk_mq_run_hw_queue(hctx
, async
);
392 void blk_mq_sched_insert_requests(struct request_queue
*q
,
393 struct blk_mq_ctx
*ctx
,
394 struct list_head
*list
, bool run_queue_async
)
396 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
397 struct elevator_queue
*e
= hctx
->queue
->elevator
;
400 struct request
*rq
, *next
;
403 * We bypass requests that already have a driver tag assigned,
404 * which should only be flushes. Flushes are only ever inserted
405 * as single requests, so we shouldn't ever hit the
406 * WARN_ON_ONCE() below (but let's handle it just in case).
408 list_for_each_entry_safe(rq
, next
, list
, queuelist
) {
409 if (WARN_ON_ONCE(rq
->tag
!= -1)) {
410 list_del_init(&rq
->queuelist
);
411 blk_mq_sched_bypass_insert(hctx
, rq
);
416 if (e
&& e
->type
->ops
.mq
.insert_requests
)
417 e
->type
->ops
.mq
.insert_requests(hctx
, list
, false);
419 blk_mq_insert_requests(hctx
, ctx
, list
);
421 blk_mq_run_hw_queue(hctx
, run_queue_async
);
424 static void blk_mq_sched_free_tags(struct blk_mq_tag_set
*set
,
425 struct blk_mq_hw_ctx
*hctx
,
426 unsigned int hctx_idx
)
428 if (hctx
->sched_tags
) {
429 blk_mq_free_rqs(set
, hctx
->sched_tags
, hctx_idx
);
430 blk_mq_free_rq_map(hctx
->sched_tags
);
431 hctx
->sched_tags
= NULL
;
435 static int blk_mq_sched_alloc_tags(struct request_queue
*q
,
436 struct blk_mq_hw_ctx
*hctx
,
437 unsigned int hctx_idx
)
439 struct blk_mq_tag_set
*set
= q
->tag_set
;
442 hctx
->sched_tags
= blk_mq_alloc_rq_map(set
, hctx_idx
, q
->nr_requests
,
444 if (!hctx
->sched_tags
)
447 ret
= blk_mq_alloc_rqs(set
, hctx
->sched_tags
, hctx_idx
, q
->nr_requests
);
449 blk_mq_sched_free_tags(set
, hctx
, hctx_idx
);
454 static void blk_mq_sched_tags_teardown(struct request_queue
*q
)
456 struct blk_mq_tag_set
*set
= q
->tag_set
;
457 struct blk_mq_hw_ctx
*hctx
;
460 queue_for_each_hw_ctx(q
, hctx
, i
)
461 blk_mq_sched_free_tags(set
, hctx
, i
);
464 int blk_mq_sched_init_hctx(struct request_queue
*q
, struct blk_mq_hw_ctx
*hctx
,
465 unsigned int hctx_idx
)
467 struct elevator_queue
*e
= q
->elevator
;
472 return blk_mq_sched_alloc_tags(q
, hctx
, hctx_idx
);
475 void blk_mq_sched_exit_hctx(struct request_queue
*q
, struct blk_mq_hw_ctx
*hctx
,
476 unsigned int hctx_idx
)
478 struct elevator_queue
*e
= q
->elevator
;
483 blk_mq_sched_free_tags(q
->tag_set
, hctx
, hctx_idx
);
486 int blk_mq_init_sched(struct request_queue
*q
, struct elevator_type
*e
)
488 struct blk_mq_hw_ctx
*hctx
;
498 * Default to 256, since we don't split into sync/async like the
499 * old code did. Additionally, this is a per-hw queue depth.
501 q
->nr_requests
= 2 * BLKDEV_MAX_RQ
;
503 queue_for_each_hw_ctx(q
, hctx
, i
) {
504 ret
= blk_mq_sched_alloc_tags(q
, hctx
, i
);
509 ret
= e
->ops
.mq
.init_sched(q
, e
);
516 blk_mq_sched_tags_teardown(q
);
521 void blk_mq_exit_sched(struct request_queue
*q
, struct elevator_queue
*e
)
523 if (e
->type
->ops
.mq
.exit_sched
)
524 e
->type
->ops
.mq
.exit_sched(e
);
525 blk_mq_sched_tags_teardown(q
);
529 int blk_mq_sched_init(struct request_queue
*q
)
533 mutex_lock(&q
->sysfs_lock
);
534 ret
= elevator_init(q
, NULL
);
535 mutex_unlock(&q
->sysfs_lock
);