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
;
113 struct blk_mq_hw_ctx
*hctx
;
114 struct blk_mq_ctx
*ctx
;
117 blk_queue_enter_live(q
);
118 ctx
= blk_mq_get_ctx(q
);
119 hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
121 blk_mq_set_alloc_data(data
, q
, data
->flags
, ctx
, hctx
);
124 data
->flags
|= BLK_MQ_REQ_INTERNAL
;
127 * Flush requests are special and go directly to the
130 if (!op_is_flush(op
) && e
->type
->ops
.mq
.get_request
) {
131 rq
= e
->type
->ops
.mq
.get_request(q
, op
, data
);
133 rq
->rq_flags
|= RQF_QUEUED
;
135 rq
= __blk_mq_alloc_request(data
, op
);
137 rq
= __blk_mq_alloc_request(data
, op
);
139 data
->hctx
->tags
->rqs
[rq
->tag
] = rq
;
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 elevator_queue
*e
= hctx
->queue
->elevator
;
178 const bool has_sched_dispatch
= e
&& e
->type
->ops
.mq
.dispatch_request
;
179 bool did_work
= false;
182 if (unlikely(blk_mq_hctx_stopped(hctx
)))
188 * If we have previous entries on our dispatch list, grab them first for
189 * more fair dispatch.
191 if (!list_empty_careful(&hctx
->dispatch
)) {
192 spin_lock(&hctx
->lock
);
193 if (!list_empty(&hctx
->dispatch
))
194 list_splice_init(&hctx
->dispatch
, &rq_list
);
195 spin_unlock(&hctx
->lock
);
199 * Only ask the scheduler for requests, if we didn't have residual
200 * requests from the dispatch list. This is to avoid the case where
201 * we only ever dispatch a fraction of the requests available because
202 * of low device queue depth. Once we pull requests out of the IO
203 * scheduler, we can no longer merge or sort them. So it's best to
204 * leave them there for as long as we can. Mark the hw queue as
205 * needing a restart in that case.
207 if (!list_empty(&rq_list
)) {
208 blk_mq_sched_mark_restart_hctx(hctx
);
209 did_work
= blk_mq_dispatch_rq_list(hctx
, &rq_list
);
210 } else if (!has_sched_dispatch
) {
211 blk_mq_flush_busy_ctxs(hctx
, &rq_list
);
212 blk_mq_dispatch_rq_list(hctx
, &rq_list
);
216 * We want to dispatch from the scheduler if we had no work left
217 * on the dispatch list, OR if we did have work but weren't able
220 if (!did_work
&& has_sched_dispatch
) {
224 rq
= e
->type
->ops
.mq
.dispatch_request(hctx
);
227 list_add(&rq
->queuelist
, &rq_list
);
228 } while (blk_mq_dispatch_rq_list(hctx
, &rq_list
));
232 void blk_mq_sched_move_to_dispatch(struct blk_mq_hw_ctx
*hctx
,
233 struct list_head
*rq_list
,
234 struct request
*(*get_rq
)(struct blk_mq_hw_ctx
*))
243 list_add_tail(&rq
->queuelist
, rq_list
);
246 EXPORT_SYMBOL_GPL(blk_mq_sched_move_to_dispatch
);
248 bool blk_mq_sched_try_merge(struct request_queue
*q
, struct bio
*bio
,
249 struct request
**merged_request
)
253 switch (elv_merge(q
, &rq
, bio
)) {
254 case ELEVATOR_BACK_MERGE
:
255 if (!blk_mq_sched_allow_merge(q
, rq
, bio
))
257 if (!bio_attempt_back_merge(q
, rq
, bio
))
259 *merged_request
= attempt_back_merge(q
, rq
);
260 if (!*merged_request
)
261 elv_merged_request(q
, rq
, ELEVATOR_BACK_MERGE
);
263 case ELEVATOR_FRONT_MERGE
:
264 if (!blk_mq_sched_allow_merge(q
, rq
, bio
))
266 if (!bio_attempt_front_merge(q
, rq
, bio
))
268 *merged_request
= attempt_front_merge(q
, rq
);
269 if (!*merged_request
)
270 elv_merged_request(q
, rq
, ELEVATOR_FRONT_MERGE
);
276 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge
);
278 bool __blk_mq_sched_bio_merge(struct request_queue
*q
, struct bio
*bio
)
280 struct elevator_queue
*e
= q
->elevator
;
282 if (e
->type
->ops
.mq
.bio_merge
) {
283 struct blk_mq_ctx
*ctx
= blk_mq_get_ctx(q
);
284 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
287 return e
->type
->ops
.mq
.bio_merge(hctx
, bio
);
293 bool blk_mq_sched_try_insert_merge(struct request_queue
*q
, struct request
*rq
)
295 return rq_mergeable(rq
) && elv_attempt_insert_merge(q
, rq
);
297 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge
);
299 void blk_mq_sched_request_inserted(struct request
*rq
)
301 trace_block_rq_insert(rq
->q
, rq
);
303 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted
);
305 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx
*hctx
,
309 rq
->rq_flags
|= RQF_SORTED
;
314 * If we already have a real request tag, send directly to
317 spin_lock(&hctx
->lock
);
318 list_add(&rq
->queuelist
, &hctx
->dispatch
);
319 spin_unlock(&hctx
->lock
);
323 static void blk_mq_sched_restart_hctx(struct blk_mq_hw_ctx
*hctx
)
325 if (test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
)) {
326 clear_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
);
327 if (blk_mq_hctx_has_pending(hctx
))
328 blk_mq_run_hw_queue(hctx
, true);
332 void blk_mq_sched_restart_queues(struct blk_mq_hw_ctx
*hctx
)
334 struct request_queue
*q
= hctx
->queue
;
337 if (test_bit(QUEUE_FLAG_RESTART
, &q
->queue_flags
)) {
338 if (test_and_clear_bit(QUEUE_FLAG_RESTART
, &q
->queue_flags
)) {
339 queue_for_each_hw_ctx(q
, hctx
, i
)
340 blk_mq_sched_restart_hctx(hctx
);
343 blk_mq_sched_restart_hctx(hctx
);
348 * Add flush/fua to the queue. If we fail getting a driver tag, then
349 * punt to the requeue list. Requeue will re-invoke us from a context
350 * that's safe to block from.
352 static void blk_mq_sched_insert_flush(struct blk_mq_hw_ctx
*hctx
,
353 struct request
*rq
, bool can_block
)
355 if (blk_mq_get_driver_tag(rq
, &hctx
, can_block
)) {
356 blk_insert_flush(rq
);
357 blk_mq_run_hw_queue(hctx
, true);
359 blk_mq_add_to_requeue_list(rq
, false, true);
362 void blk_mq_sched_insert_request(struct request
*rq
, bool at_head
,
363 bool run_queue
, bool async
, bool can_block
)
365 struct request_queue
*q
= rq
->q
;
366 struct elevator_queue
*e
= q
->elevator
;
367 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
368 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
370 if (rq
->tag
== -1 && op_is_flush(rq
->cmd_flags
)) {
371 blk_mq_sched_insert_flush(hctx
, rq
, can_block
);
375 if (e
&& blk_mq_sched_bypass_insert(hctx
, rq
))
378 if (e
&& e
->type
->ops
.mq
.insert_requests
) {
381 list_add(&rq
->queuelist
, &list
);
382 e
->type
->ops
.mq
.insert_requests(hctx
, &list
, at_head
);
384 spin_lock(&ctx
->lock
);
385 __blk_mq_insert_request(hctx
, rq
, at_head
);
386 spin_unlock(&ctx
->lock
);
391 blk_mq_run_hw_queue(hctx
, async
);
394 void blk_mq_sched_insert_requests(struct request_queue
*q
,
395 struct blk_mq_ctx
*ctx
,
396 struct list_head
*list
, bool run_queue_async
)
398 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
399 struct elevator_queue
*e
= hctx
->queue
->elevator
;
402 struct request
*rq
, *next
;
405 * We bypass requests that already have a driver tag assigned,
406 * which should only be flushes. Flushes are only ever inserted
407 * as single requests, so we shouldn't ever hit the
408 * WARN_ON_ONCE() below (but let's handle it just in case).
410 list_for_each_entry_safe(rq
, next
, list
, queuelist
) {
411 if (WARN_ON_ONCE(rq
->tag
!= -1)) {
412 list_del_init(&rq
->queuelist
);
413 blk_mq_sched_bypass_insert(hctx
, rq
);
418 if (e
&& e
->type
->ops
.mq
.insert_requests
)
419 e
->type
->ops
.mq
.insert_requests(hctx
, list
, false);
421 blk_mq_insert_requests(hctx
, ctx
, list
);
423 blk_mq_run_hw_queue(hctx
, run_queue_async
);
426 static void blk_mq_sched_free_tags(struct blk_mq_tag_set
*set
,
427 struct blk_mq_hw_ctx
*hctx
,
428 unsigned int hctx_idx
)
430 if (hctx
->sched_tags
) {
431 blk_mq_free_rqs(set
, hctx
->sched_tags
, hctx_idx
);
432 blk_mq_free_rq_map(hctx
->sched_tags
);
433 hctx
->sched_tags
= NULL
;
437 int blk_mq_sched_setup(struct request_queue
*q
)
439 struct blk_mq_tag_set
*set
= q
->tag_set
;
440 struct blk_mq_hw_ctx
*hctx
;
444 * Default to 256, since we don't split into sync/async like the
445 * old code did. Additionally, this is a per-hw queue depth.
447 q
->nr_requests
= 2 * BLKDEV_MAX_RQ
;
450 * We're switching to using an IO scheduler, so setup the hctx
451 * scheduler tags and switch the request map from the regular
452 * tags to scheduler tags. First allocate what we need, so we
453 * can safely fail and fallback, if needed.
456 queue_for_each_hw_ctx(q
, hctx
, i
) {
457 hctx
->sched_tags
= blk_mq_alloc_rq_map(set
, i
, q
->nr_requests
, 0);
458 if (!hctx
->sched_tags
) {
462 ret
= blk_mq_alloc_rqs(set
, hctx
->sched_tags
, i
, q
->nr_requests
);
468 * If we failed, free what we did allocate
471 queue_for_each_hw_ctx(q
, hctx
, i
) {
472 if (!hctx
->sched_tags
)
474 blk_mq_sched_free_tags(set
, hctx
, i
);
483 void blk_mq_sched_teardown(struct request_queue
*q
)
485 struct blk_mq_tag_set
*set
= q
->tag_set
;
486 struct blk_mq_hw_ctx
*hctx
;
489 queue_for_each_hw_ctx(q
, hctx
, i
)
490 blk_mq_sched_free_tags(set
, hctx
, i
);
493 int blk_mq_sched_init(struct request_queue
*q
)
497 mutex_lock(&q
->sysfs_lock
);
498 ret
= elevator_init(q
, NULL
);
499 mutex_unlock(&q
->sysfs_lock
);