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
,
71 struct request
*rq
, struct io_context
*ioc
)
75 spin_lock_irq(q
->queue_lock
);
76 icq
= ioc_lookup_icq(ioc
, q
);
77 spin_unlock_irq(q
->queue_lock
);
80 icq
= ioc_create_icq(ioc
, q
, GFP_ATOMIC
);
86 if (!blk_mq_sched_get_rq_priv(q
, rq
)) {
87 rq
->rq_flags
|= RQF_ELVPRIV
;
88 get_io_context(icq
->ioc
);
95 static void blk_mq_sched_assign_ioc(struct request_queue
*q
,
96 struct request
*rq
, struct bio
*bio
)
98 struct io_context
*ioc
;
102 __blk_mq_sched_assign_ioc(q
, rq
, ioc
);
105 struct request
*blk_mq_sched_get_request(struct request_queue
*q
,
108 struct blk_mq_alloc_data
*data
)
110 struct elevator_queue
*e
= q
->elevator
;
111 struct blk_mq_hw_ctx
*hctx
;
112 struct blk_mq_ctx
*ctx
;
115 blk_queue_enter_live(q
);
116 ctx
= blk_mq_get_ctx(q
);
117 hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
119 blk_mq_set_alloc_data(data
, q
, data
->flags
, ctx
, hctx
);
122 data
->flags
|= BLK_MQ_REQ_INTERNAL
;
125 * Flush requests are special and go directly to the
128 if (!op_is_flush(op
) && e
->type
->ops
.mq
.get_request
) {
129 rq
= e
->type
->ops
.mq
.get_request(q
, op
, data
);
131 rq
->rq_flags
|= RQF_QUEUED
;
133 rq
= __blk_mq_alloc_request(data
, op
);
135 rq
= __blk_mq_alloc_request(data
, op
);
137 data
->hctx
->tags
->rqs
[rq
->tag
] = rq
;
141 if (!op_is_flush(op
)) {
143 if (e
&& e
->type
->icq_cache
)
144 blk_mq_sched_assign_ioc(q
, rq
, bio
);
146 data
->hctx
->queued
++;
154 void blk_mq_sched_put_request(struct request
*rq
)
156 struct request_queue
*q
= rq
->q
;
157 struct elevator_queue
*e
= q
->elevator
;
159 if (rq
->rq_flags
& RQF_ELVPRIV
) {
160 blk_mq_sched_put_rq_priv(rq
->q
, rq
);
162 put_io_context(rq
->elv
.icq
->ioc
);
167 if ((rq
->rq_flags
& RQF_QUEUED
) && e
&& e
->type
->ops
.mq
.put_request
)
168 e
->type
->ops
.mq
.put_request(rq
);
170 blk_mq_finish_request(rq
);
173 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx
*hctx
)
175 struct elevator_queue
*e
= hctx
->queue
->elevator
;
178 if (unlikely(blk_mq_hctx_stopped(hctx
)))
184 * If we have previous entries on our dispatch list, grab them first for
185 * more fair dispatch.
187 if (!list_empty_careful(&hctx
->dispatch
)) {
188 spin_lock(&hctx
->lock
);
189 if (!list_empty(&hctx
->dispatch
))
190 list_splice_init(&hctx
->dispatch
, &rq_list
);
191 spin_unlock(&hctx
->lock
);
195 * Only ask the scheduler for requests, if we didn't have residual
196 * requests from the dispatch list. This is to avoid the case where
197 * we only ever dispatch a fraction of the requests available because
198 * of low device queue depth. Once we pull requests out of the IO
199 * scheduler, we can no longer merge or sort them. So it's best to
200 * leave them there for as long as we can. Mark the hw queue as
201 * needing a restart in that case.
203 if (!list_empty(&rq_list
)) {
204 blk_mq_sched_mark_restart(hctx
);
205 blk_mq_dispatch_rq_list(hctx
, &rq_list
);
206 } else if (!e
|| !e
->type
->ops
.mq
.dispatch_request
) {
207 blk_mq_flush_busy_ctxs(hctx
, &rq_list
);
208 blk_mq_dispatch_rq_list(hctx
, &rq_list
);
213 rq
= e
->type
->ops
.mq
.dispatch_request(hctx
);
216 list_add(&rq
->queuelist
, &rq_list
);
217 } while (blk_mq_dispatch_rq_list(hctx
, &rq_list
));
221 void blk_mq_sched_move_to_dispatch(struct blk_mq_hw_ctx
*hctx
,
222 struct list_head
*rq_list
,
223 struct request
*(*get_rq
)(struct blk_mq_hw_ctx
*))
232 list_add_tail(&rq
->queuelist
, rq_list
);
235 EXPORT_SYMBOL_GPL(blk_mq_sched_move_to_dispatch
);
237 bool blk_mq_sched_try_merge(struct request_queue
*q
, struct bio
*bio
)
242 ret
= elv_merge(q
, &rq
, bio
);
243 if (ret
== ELEVATOR_BACK_MERGE
) {
244 if (!blk_mq_sched_allow_merge(q
, rq
, bio
))
246 if (bio_attempt_back_merge(q
, rq
, bio
)) {
247 if (!attempt_back_merge(q
, rq
))
248 elv_merged_request(q
, rq
, ret
);
251 } else if (ret
== ELEVATOR_FRONT_MERGE
) {
252 if (!blk_mq_sched_allow_merge(q
, rq
, bio
))
254 if (bio_attempt_front_merge(q
, rq
, bio
)) {
255 if (!attempt_front_merge(q
, rq
))
256 elv_merged_request(q
, rq
, ret
);
263 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge
);
265 bool __blk_mq_sched_bio_merge(struct request_queue
*q
, struct bio
*bio
)
267 struct elevator_queue
*e
= q
->elevator
;
269 if (e
->type
->ops
.mq
.bio_merge
) {
270 struct blk_mq_ctx
*ctx
= blk_mq_get_ctx(q
);
271 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
274 return e
->type
->ops
.mq
.bio_merge(hctx
, bio
);
280 bool blk_mq_sched_try_insert_merge(struct request_queue
*q
, struct request
*rq
)
282 return rq_mergeable(rq
) && elv_attempt_insert_merge(q
, rq
);
284 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge
);
286 void blk_mq_sched_request_inserted(struct request
*rq
)
288 trace_block_rq_insert(rq
->q
, rq
);
290 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted
);
292 bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
295 rq
->rq_flags
|= RQF_SORTED
;
300 * If we already have a real request tag, send directly to
303 spin_lock(&hctx
->lock
);
304 list_add(&rq
->queuelist
, &hctx
->dispatch
);
305 spin_unlock(&hctx
->lock
);
308 EXPORT_SYMBOL_GPL(blk_mq_sched_bypass_insert
);
310 static void blk_mq_sched_restart_hctx(struct blk_mq_hw_ctx
*hctx
)
312 if (test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
)) {
313 clear_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
);
314 if (blk_mq_hctx_has_pending(hctx
))
315 blk_mq_run_hw_queue(hctx
, true);
319 void blk_mq_sched_restart_queues(struct blk_mq_hw_ctx
*hctx
)
323 if (!(hctx
->flags
& BLK_MQ_F_TAG_SHARED
))
324 blk_mq_sched_restart_hctx(hctx
);
326 struct request_queue
*q
= hctx
->queue
;
328 if (!test_bit(QUEUE_FLAG_RESTART
, &q
->queue_flags
))
331 clear_bit(QUEUE_FLAG_RESTART
, &q
->queue_flags
);
333 queue_for_each_hw_ctx(q
, hctx
, i
)
334 blk_mq_sched_restart_hctx(hctx
);
339 * Add flush/fua to the queue. If we fail getting a driver tag, then
340 * punt to the requeue list. Requeue will re-invoke us from a context
341 * that's safe to block from.
343 static void blk_mq_sched_insert_flush(struct blk_mq_hw_ctx
*hctx
,
344 struct request
*rq
, bool can_block
)
346 if (blk_mq_get_driver_tag(rq
, &hctx
, can_block
)) {
347 blk_insert_flush(rq
);
348 blk_mq_run_hw_queue(hctx
, true);
350 blk_mq_add_to_requeue_list(rq
, true, true);
353 void blk_mq_sched_insert_request(struct request
*rq
, bool at_head
,
354 bool run_queue
, bool async
, bool can_block
)
356 struct request_queue
*q
= rq
->q
;
357 struct elevator_queue
*e
= q
->elevator
;
358 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
359 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
361 if (rq
->tag
== -1 && op_is_flush(rq
->cmd_flags
)) {
362 blk_mq_sched_insert_flush(hctx
, rq
, can_block
);
366 if (e
&& e
->type
->ops
.mq
.insert_requests
) {
369 list_add(&rq
->queuelist
, &list
);
370 e
->type
->ops
.mq
.insert_requests(hctx
, &list
, at_head
);
372 spin_lock(&ctx
->lock
);
373 __blk_mq_insert_request(hctx
, rq
, at_head
);
374 spin_unlock(&ctx
->lock
);
378 blk_mq_run_hw_queue(hctx
, async
);
381 void blk_mq_sched_insert_requests(struct request_queue
*q
,
382 struct blk_mq_ctx
*ctx
,
383 struct list_head
*list
, bool run_queue_async
)
385 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
386 struct elevator_queue
*e
= hctx
->queue
->elevator
;
388 if (e
&& e
->type
->ops
.mq
.insert_requests
)
389 e
->type
->ops
.mq
.insert_requests(hctx
, list
, false);
391 blk_mq_insert_requests(hctx
, ctx
, list
);
393 blk_mq_run_hw_queue(hctx
, run_queue_async
);
396 static void blk_mq_sched_free_tags(struct blk_mq_tag_set
*set
,
397 struct blk_mq_hw_ctx
*hctx
,
398 unsigned int hctx_idx
)
400 if (hctx
->sched_tags
) {
401 blk_mq_free_rqs(set
, hctx
->sched_tags
, hctx_idx
);
402 blk_mq_free_rq_map(hctx
->sched_tags
);
403 hctx
->sched_tags
= NULL
;
407 int blk_mq_sched_setup(struct request_queue
*q
)
409 struct blk_mq_tag_set
*set
= q
->tag_set
;
410 struct blk_mq_hw_ctx
*hctx
;
414 * Default to 256, since we don't split into sync/async like the
415 * old code did. Additionally, this is a per-hw queue depth.
417 q
->nr_requests
= 2 * BLKDEV_MAX_RQ
;
420 * We're switching to using an IO scheduler, so setup the hctx
421 * scheduler tags and switch the request map from the regular
422 * tags to scheduler tags. First allocate what we need, so we
423 * can safely fail and fallback, if needed.
426 queue_for_each_hw_ctx(q
, hctx
, i
) {
427 hctx
->sched_tags
= blk_mq_alloc_rq_map(set
, i
, q
->nr_requests
, 0);
428 if (!hctx
->sched_tags
) {
432 ret
= blk_mq_alloc_rqs(set
, hctx
->sched_tags
, i
, q
->nr_requests
);
438 * If we failed, free what we did allocate
441 queue_for_each_hw_ctx(q
, hctx
, i
) {
442 if (!hctx
->sched_tags
)
444 blk_mq_sched_free_tags(set
, hctx
, i
);
453 void blk_mq_sched_teardown(struct request_queue
*q
)
455 struct blk_mq_tag_set
*set
= q
->tag_set
;
456 struct blk_mq_hw_ctx
*hctx
;
459 queue_for_each_hw_ctx(q
, hctx
, i
)
460 blk_mq_sched_free_tags(set
, hctx
, i
);
463 int blk_mq_sched_init(struct request_queue
*q
)
467 #if defined(CONFIG_DEFAULT_SQ_NONE)
468 if (q
->nr_hw_queues
== 1)
471 #if defined(CONFIG_DEFAULT_MQ_NONE)
472 if (q
->nr_hw_queues
> 1)
476 mutex_lock(&q
->sysfs_lock
);
477 ret
= elevator_init(q
, NULL
);
478 mutex_unlock(&q
->sysfs_lock
);