1 // SPDX-License-Identifier: GPL-2.0
3 * Block multiqueue core code
5 * Copyright (C) 2013-2014 Jens Axboe
6 * Copyright (C) 2013-2014 Christoph Hellwig
8 #include <linux/kernel.h>
9 #include <linux/module.h>
10 #include <linux/backing-dev.h>
11 #include <linux/bio.h>
12 #include <linux/blkdev.h>
13 #include <linux/blk-integrity.h>
14 #include <linux/kmemleak.h>
16 #include <linux/init.h>
17 #include <linux/slab.h>
18 #include <linux/workqueue.h>
19 #include <linux/smp.h>
20 #include <linux/interrupt.h>
21 #include <linux/llist.h>
22 #include <linux/cpu.h>
23 #include <linux/cache.h>
24 #include <linux/sched/sysctl.h>
25 #include <linux/sched/topology.h>
26 #include <linux/sched/signal.h>
27 #include <linux/delay.h>
28 #include <linux/crash_dump.h>
29 #include <linux/prefetch.h>
30 #include <linux/blk-crypto.h>
31 #include <linux/part_stat.h>
33 #include <trace/events/block.h>
35 #include <linux/t10-pi.h>
38 #include "blk-mq-debugfs.h"
41 #include "blk-mq-sched.h"
42 #include "blk-rq-qos.h"
43 #include "blk-ioprio.h"
45 static DEFINE_PER_CPU(struct llist_head
, blk_cpu_done
);
46 static DEFINE_PER_CPU(call_single_data_t
, blk_cpu_csd
);
48 static void blk_mq_insert_request(struct request
*rq
, blk_insert_t flags
);
49 static void blk_mq_request_bypass_insert(struct request
*rq
,
51 static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx
*hctx
,
52 struct list_head
*list
);
53 static int blk_hctx_poll(struct request_queue
*q
, struct blk_mq_hw_ctx
*hctx
,
54 struct io_comp_batch
*iob
, unsigned int flags
);
57 * Check if any of the ctx, dispatch list or elevator
58 * have pending work in this hardware queue.
60 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
62 return !list_empty_careful(&hctx
->dispatch
) ||
63 sbitmap_any_bit_set(&hctx
->ctx_map
) ||
64 blk_mq_sched_has_work(hctx
);
68 * Mark this ctx as having pending work in this hardware queue
70 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
71 struct blk_mq_ctx
*ctx
)
73 const int bit
= ctx
->index_hw
[hctx
->type
];
75 if (!sbitmap_test_bit(&hctx
->ctx_map
, bit
))
76 sbitmap_set_bit(&hctx
->ctx_map
, bit
);
79 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
80 struct blk_mq_ctx
*ctx
)
82 const int bit
= ctx
->index_hw
[hctx
->type
];
84 sbitmap_clear_bit(&hctx
->ctx_map
, bit
);
88 struct block_device
*part
;
89 unsigned int inflight
[2];
92 static bool blk_mq_check_inflight(struct request
*rq
, void *priv
)
94 struct mq_inflight
*mi
= priv
;
96 if (rq
->part
&& blk_do_io_stat(rq
) &&
97 (!mi
->part
->bd_partno
|| rq
->part
== mi
->part
) &&
98 blk_mq_rq_state(rq
) == MQ_RQ_IN_FLIGHT
)
99 mi
->inflight
[rq_data_dir(rq
)]++;
104 unsigned int blk_mq_in_flight(struct request_queue
*q
,
105 struct block_device
*part
)
107 struct mq_inflight mi
= { .part
= part
};
109 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight
, &mi
);
111 return mi
.inflight
[0] + mi
.inflight
[1];
114 void blk_mq_in_flight_rw(struct request_queue
*q
, struct block_device
*part
,
115 unsigned int inflight
[2])
117 struct mq_inflight mi
= { .part
= part
};
119 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight
, &mi
);
120 inflight
[0] = mi
.inflight
[0];
121 inflight
[1] = mi
.inflight
[1];
124 void blk_freeze_queue_start(struct request_queue
*q
)
126 mutex_lock(&q
->mq_freeze_lock
);
127 if (++q
->mq_freeze_depth
== 1) {
128 percpu_ref_kill(&q
->q_usage_counter
);
129 mutex_unlock(&q
->mq_freeze_lock
);
131 blk_mq_run_hw_queues(q
, false);
133 mutex_unlock(&q
->mq_freeze_lock
);
136 EXPORT_SYMBOL_GPL(blk_freeze_queue_start
);
138 void blk_mq_freeze_queue_wait(struct request_queue
*q
)
140 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
142 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait
);
144 int blk_mq_freeze_queue_wait_timeout(struct request_queue
*q
,
145 unsigned long timeout
)
147 return wait_event_timeout(q
->mq_freeze_wq
,
148 percpu_ref_is_zero(&q
->q_usage_counter
),
151 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout
);
154 * Guarantee no request is in use, so we can change any data structure of
155 * the queue afterward.
157 void blk_freeze_queue(struct request_queue
*q
)
160 * In the !blk_mq case we are only calling this to kill the
161 * q_usage_counter, otherwise this increases the freeze depth
162 * and waits for it to return to zero. For this reason there is
163 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
164 * exported to drivers as the only user for unfreeze is blk_mq.
166 blk_freeze_queue_start(q
);
167 blk_mq_freeze_queue_wait(q
);
170 void blk_mq_freeze_queue(struct request_queue
*q
)
173 * ...just an alias to keep freeze and unfreeze actions balanced
174 * in the blk_mq_* namespace
178 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
180 void __blk_mq_unfreeze_queue(struct request_queue
*q
, bool force_atomic
)
182 mutex_lock(&q
->mq_freeze_lock
);
184 q
->q_usage_counter
.data
->force_atomic
= true;
185 q
->mq_freeze_depth
--;
186 WARN_ON_ONCE(q
->mq_freeze_depth
< 0);
187 if (!q
->mq_freeze_depth
) {
188 percpu_ref_resurrect(&q
->q_usage_counter
);
189 wake_up_all(&q
->mq_freeze_wq
);
191 mutex_unlock(&q
->mq_freeze_lock
);
194 void blk_mq_unfreeze_queue(struct request_queue
*q
)
196 __blk_mq_unfreeze_queue(q
, false);
198 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
201 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
202 * mpt3sas driver such that this function can be removed.
204 void blk_mq_quiesce_queue_nowait(struct request_queue
*q
)
208 spin_lock_irqsave(&q
->queue_lock
, flags
);
209 if (!q
->quiesce_depth
++)
210 blk_queue_flag_set(QUEUE_FLAG_QUIESCED
, q
);
211 spin_unlock_irqrestore(&q
->queue_lock
, flags
);
213 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait
);
216 * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done
217 * @set: tag_set to wait on
219 * Note: it is driver's responsibility for making sure that quiesce has
220 * been started on or more of the request_queues of the tag_set. This
221 * function only waits for the quiesce on those request_queues that had
222 * the quiesce flag set using blk_mq_quiesce_queue_nowait.
224 void blk_mq_wait_quiesce_done(struct blk_mq_tag_set
*set
)
226 if (set
->flags
& BLK_MQ_F_BLOCKING
)
227 synchronize_srcu(set
->srcu
);
231 EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done
);
234 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
237 * Note: this function does not prevent that the struct request end_io()
238 * callback function is invoked. Once this function is returned, we make
239 * sure no dispatch can happen until the queue is unquiesced via
240 * blk_mq_unquiesce_queue().
242 void blk_mq_quiesce_queue(struct request_queue
*q
)
244 blk_mq_quiesce_queue_nowait(q
);
245 /* nothing to wait for non-mq queues */
247 blk_mq_wait_quiesce_done(q
->tag_set
);
249 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
252 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
255 * This function recovers queue into the state before quiescing
256 * which is done by blk_mq_quiesce_queue.
258 void blk_mq_unquiesce_queue(struct request_queue
*q
)
261 bool run_queue
= false;
263 spin_lock_irqsave(&q
->queue_lock
, flags
);
264 if (WARN_ON_ONCE(q
->quiesce_depth
<= 0)) {
266 } else if (!--q
->quiesce_depth
) {
267 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED
, q
);
270 spin_unlock_irqrestore(&q
->queue_lock
, flags
);
272 /* dispatch requests which are inserted during quiescing */
274 blk_mq_run_hw_queues(q
, true);
276 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue
);
278 void blk_mq_quiesce_tagset(struct blk_mq_tag_set
*set
)
280 struct request_queue
*q
;
282 mutex_lock(&set
->tag_list_lock
);
283 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
284 if (!blk_queue_skip_tagset_quiesce(q
))
285 blk_mq_quiesce_queue_nowait(q
);
287 blk_mq_wait_quiesce_done(set
);
288 mutex_unlock(&set
->tag_list_lock
);
290 EXPORT_SYMBOL_GPL(blk_mq_quiesce_tagset
);
292 void blk_mq_unquiesce_tagset(struct blk_mq_tag_set
*set
)
294 struct request_queue
*q
;
296 mutex_lock(&set
->tag_list_lock
);
297 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
298 if (!blk_queue_skip_tagset_quiesce(q
))
299 blk_mq_unquiesce_queue(q
);
301 mutex_unlock(&set
->tag_list_lock
);
303 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_tagset
);
305 void blk_mq_wake_waiters(struct request_queue
*q
)
307 struct blk_mq_hw_ctx
*hctx
;
310 queue_for_each_hw_ctx(q
, hctx
, i
)
311 if (blk_mq_hw_queue_mapped(hctx
))
312 blk_mq_tag_wakeup_all(hctx
->tags
, true);
315 void blk_rq_init(struct request_queue
*q
, struct request
*rq
)
317 memset(rq
, 0, sizeof(*rq
));
319 INIT_LIST_HEAD(&rq
->queuelist
);
321 rq
->__sector
= (sector_t
) -1;
322 INIT_HLIST_NODE(&rq
->hash
);
323 RB_CLEAR_NODE(&rq
->rb_node
);
324 rq
->tag
= BLK_MQ_NO_TAG
;
325 rq
->internal_tag
= BLK_MQ_NO_TAG
;
326 rq
->start_time_ns
= ktime_get_ns();
328 blk_crypto_rq_set_defaults(rq
);
330 EXPORT_SYMBOL(blk_rq_init
);
332 /* Set start and alloc time when the allocated request is actually used */
333 static inline void blk_mq_rq_time_init(struct request
*rq
, u64 alloc_time_ns
)
335 if (blk_mq_need_time_stamp(rq
))
336 rq
->start_time_ns
= ktime_get_ns();
338 rq
->start_time_ns
= 0;
340 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
341 if (blk_queue_rq_alloc_time(rq
->q
))
342 rq
->alloc_time_ns
= alloc_time_ns
?: rq
->start_time_ns
;
344 rq
->alloc_time_ns
= 0;
348 static struct request
*blk_mq_rq_ctx_init(struct blk_mq_alloc_data
*data
,
349 struct blk_mq_tags
*tags
, unsigned int tag
)
351 struct blk_mq_ctx
*ctx
= data
->ctx
;
352 struct blk_mq_hw_ctx
*hctx
= data
->hctx
;
353 struct request_queue
*q
= data
->q
;
354 struct request
*rq
= tags
->static_rqs
[tag
];
359 rq
->cmd_flags
= data
->cmd_flags
;
361 if (data
->flags
& BLK_MQ_REQ_PM
)
362 data
->rq_flags
|= RQF_PM
;
363 if (blk_queue_io_stat(q
))
364 data
->rq_flags
|= RQF_IO_STAT
;
365 rq
->rq_flags
= data
->rq_flags
;
367 if (data
->rq_flags
& RQF_SCHED_TAGS
) {
368 rq
->tag
= BLK_MQ_NO_TAG
;
369 rq
->internal_tag
= tag
;
372 rq
->internal_tag
= BLK_MQ_NO_TAG
;
377 rq
->io_start_time_ns
= 0;
378 rq
->stats_sectors
= 0;
379 rq
->nr_phys_segments
= 0;
380 #if defined(CONFIG_BLK_DEV_INTEGRITY)
381 rq
->nr_integrity_segments
= 0;
384 rq
->end_io_data
= NULL
;
386 blk_crypto_rq_set_defaults(rq
);
387 INIT_LIST_HEAD(&rq
->queuelist
);
388 /* tag was already set */
389 WRITE_ONCE(rq
->deadline
, 0);
392 if (rq
->rq_flags
& RQF_USE_SCHED
) {
393 struct elevator_queue
*e
= data
->q
->elevator
;
395 INIT_HLIST_NODE(&rq
->hash
);
396 RB_CLEAR_NODE(&rq
->rb_node
);
398 if (e
->type
->ops
.prepare_request
)
399 e
->type
->ops
.prepare_request(rq
);
405 static inline struct request
*
406 __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data
*data
)
408 unsigned int tag
, tag_offset
;
409 struct blk_mq_tags
*tags
;
411 unsigned long tag_mask
;
414 tag_mask
= blk_mq_get_tags(data
, data
->nr_tags
, &tag_offset
);
415 if (unlikely(!tag_mask
))
418 tags
= blk_mq_tags_from_data(data
);
419 for (i
= 0; tag_mask
; i
++) {
420 if (!(tag_mask
& (1UL << i
)))
422 tag
= tag_offset
+ i
;
423 prefetch(tags
->static_rqs
[tag
]);
424 tag_mask
&= ~(1UL << i
);
425 rq
= blk_mq_rq_ctx_init(data
, tags
, tag
);
426 rq_list_add(data
->cached_rq
, rq
);
429 if (!(data
->rq_flags
& RQF_SCHED_TAGS
))
430 blk_mq_add_active_requests(data
->hctx
, nr
);
431 /* caller already holds a reference, add for remainder */
432 percpu_ref_get_many(&data
->q
->q_usage_counter
, nr
- 1);
435 return rq_list_pop(data
->cached_rq
);
438 static struct request
*__blk_mq_alloc_requests(struct blk_mq_alloc_data
*data
)
440 struct request_queue
*q
= data
->q
;
441 u64 alloc_time_ns
= 0;
445 /* alloc_time includes depth and tag waits */
446 if (blk_queue_rq_alloc_time(q
))
447 alloc_time_ns
= ktime_get_ns();
449 if (data
->cmd_flags
& REQ_NOWAIT
)
450 data
->flags
|= BLK_MQ_REQ_NOWAIT
;
454 * All requests use scheduler tags when an I/O scheduler is
455 * enabled for the queue.
457 data
->rq_flags
|= RQF_SCHED_TAGS
;
460 * Flush/passthrough requests are special and go directly to the
463 if ((data
->cmd_flags
& REQ_OP_MASK
) != REQ_OP_FLUSH
&&
464 !blk_op_is_passthrough(data
->cmd_flags
)) {
465 struct elevator_mq_ops
*ops
= &q
->elevator
->type
->ops
;
467 WARN_ON_ONCE(data
->flags
& BLK_MQ_REQ_RESERVED
);
469 data
->rq_flags
|= RQF_USE_SCHED
;
470 if (ops
->limit_depth
)
471 ops
->limit_depth(data
->cmd_flags
, data
);
476 data
->ctx
= blk_mq_get_ctx(q
);
477 data
->hctx
= blk_mq_map_queue(q
, data
->cmd_flags
, data
->ctx
);
478 if (!(data
->rq_flags
& RQF_SCHED_TAGS
))
479 blk_mq_tag_busy(data
->hctx
);
481 if (data
->flags
& BLK_MQ_REQ_RESERVED
)
482 data
->rq_flags
|= RQF_RESV
;
485 * Try batched alloc if we want more than 1 tag.
487 if (data
->nr_tags
> 1) {
488 rq
= __blk_mq_alloc_requests_batch(data
);
490 blk_mq_rq_time_init(rq
, alloc_time_ns
);
497 * Waiting allocations only fail because of an inactive hctx. In that
498 * case just retry the hctx assignment and tag allocation as CPU hotplug
499 * should have migrated us to an online CPU by now.
501 tag
= blk_mq_get_tag(data
);
502 if (tag
== BLK_MQ_NO_TAG
) {
503 if (data
->flags
& BLK_MQ_REQ_NOWAIT
)
506 * Give up the CPU and sleep for a random short time to
507 * ensure that thread using a realtime scheduling class
508 * are migrated off the CPU, and thus off the hctx that
515 if (!(data
->rq_flags
& RQF_SCHED_TAGS
))
516 blk_mq_inc_active_requests(data
->hctx
);
517 rq
= blk_mq_rq_ctx_init(data
, blk_mq_tags_from_data(data
), tag
);
518 blk_mq_rq_time_init(rq
, alloc_time_ns
);
522 static struct request
*blk_mq_rq_cache_fill(struct request_queue
*q
,
523 struct blk_plug
*plug
,
525 blk_mq_req_flags_t flags
)
527 struct blk_mq_alloc_data data
= {
531 .nr_tags
= plug
->nr_ios
,
532 .cached_rq
= &plug
->cached_rq
,
536 if (blk_queue_enter(q
, flags
))
541 rq
= __blk_mq_alloc_requests(&data
);
547 static struct request
*blk_mq_alloc_cached_request(struct request_queue
*q
,
549 blk_mq_req_flags_t flags
)
551 struct blk_plug
*plug
= current
->plug
;
557 if (rq_list_empty(plug
->cached_rq
)) {
558 if (plug
->nr_ios
== 1)
560 rq
= blk_mq_rq_cache_fill(q
, plug
, opf
, flags
);
564 rq
= rq_list_peek(&plug
->cached_rq
);
565 if (!rq
|| rq
->q
!= q
)
568 if (blk_mq_get_hctx_type(opf
) != rq
->mq_hctx
->type
)
570 if (op_is_flush(rq
->cmd_flags
) != op_is_flush(opf
))
573 plug
->cached_rq
= rq_list_next(rq
);
574 blk_mq_rq_time_init(rq
, 0);
578 INIT_LIST_HEAD(&rq
->queuelist
);
582 struct request
*blk_mq_alloc_request(struct request_queue
*q
, blk_opf_t opf
,
583 blk_mq_req_flags_t flags
)
587 rq
= blk_mq_alloc_cached_request(q
, opf
, flags
);
589 struct blk_mq_alloc_data data
= {
597 ret
= blk_queue_enter(q
, flags
);
601 rq
= __blk_mq_alloc_requests(&data
);
606 rq
->__sector
= (sector_t
) -1;
607 rq
->bio
= rq
->biotail
= NULL
;
611 return ERR_PTR(-EWOULDBLOCK
);
613 EXPORT_SYMBOL(blk_mq_alloc_request
);
615 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
,
616 blk_opf_t opf
, blk_mq_req_flags_t flags
, unsigned int hctx_idx
)
618 struct blk_mq_alloc_data data
= {
624 u64 alloc_time_ns
= 0;
630 /* alloc_time includes depth and tag waits */
631 if (blk_queue_rq_alloc_time(q
))
632 alloc_time_ns
= ktime_get_ns();
635 * If the tag allocator sleeps we could get an allocation for a
636 * different hardware context. No need to complicate the low level
637 * allocator for this for the rare use case of a command tied to
640 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)) ||
641 WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_RESERVED
)))
642 return ERR_PTR(-EINVAL
);
644 if (hctx_idx
>= q
->nr_hw_queues
)
645 return ERR_PTR(-EIO
);
647 ret
= blk_queue_enter(q
, flags
);
652 * Check if the hardware context is actually mapped to anything.
653 * If not tell the caller that it should skip this queue.
656 data
.hctx
= xa_load(&q
->hctx_table
, hctx_idx
);
657 if (!blk_mq_hw_queue_mapped(data
.hctx
))
659 cpu
= cpumask_first_and(data
.hctx
->cpumask
, cpu_online_mask
);
660 if (cpu
>= nr_cpu_ids
)
662 data
.ctx
= __blk_mq_get_ctx(q
, cpu
);
665 data
.rq_flags
|= RQF_SCHED_TAGS
;
667 blk_mq_tag_busy(data
.hctx
);
669 if (flags
& BLK_MQ_REQ_RESERVED
)
670 data
.rq_flags
|= RQF_RESV
;
673 tag
= blk_mq_get_tag(&data
);
674 if (tag
== BLK_MQ_NO_TAG
)
676 if (!(data
.rq_flags
& RQF_SCHED_TAGS
))
677 blk_mq_inc_active_requests(data
.hctx
);
678 rq
= blk_mq_rq_ctx_init(&data
, blk_mq_tags_from_data(&data
), tag
);
679 blk_mq_rq_time_init(rq
, alloc_time_ns
);
681 rq
->__sector
= (sector_t
) -1;
682 rq
->bio
= rq
->biotail
= NULL
;
689 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
691 static void blk_mq_finish_request(struct request
*rq
)
693 struct request_queue
*q
= rq
->q
;
695 if (rq
->rq_flags
& RQF_USE_SCHED
) {
696 q
->elevator
->type
->ops
.finish_request(rq
);
698 * For postflush request that may need to be
699 * completed twice, we should clear this flag
700 * to avoid double finish_request() on the rq.
702 rq
->rq_flags
&= ~RQF_USE_SCHED
;
706 static void __blk_mq_free_request(struct request
*rq
)
708 struct request_queue
*q
= rq
->q
;
709 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
710 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
711 const int sched_tag
= rq
->internal_tag
;
713 blk_crypto_free_request(rq
);
714 blk_pm_mark_last_busy(rq
);
717 if (rq
->tag
!= BLK_MQ_NO_TAG
) {
718 blk_mq_dec_active_requests(hctx
);
719 blk_mq_put_tag(hctx
->tags
, ctx
, rq
->tag
);
721 if (sched_tag
!= BLK_MQ_NO_TAG
)
722 blk_mq_put_tag(hctx
->sched_tags
, ctx
, sched_tag
);
723 blk_mq_sched_restart(hctx
);
727 void blk_mq_free_request(struct request
*rq
)
729 struct request_queue
*q
= rq
->q
;
731 blk_mq_finish_request(rq
);
733 if (unlikely(laptop_mode
&& !blk_rq_is_passthrough(rq
)))
734 laptop_io_completion(q
->disk
->bdi
);
738 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
739 if (req_ref_put_and_test(rq
))
740 __blk_mq_free_request(rq
);
742 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
744 void blk_mq_free_plug_rqs(struct blk_plug
*plug
)
748 while ((rq
= rq_list_pop(&plug
->cached_rq
)) != NULL
)
749 blk_mq_free_request(rq
);
752 void blk_dump_rq_flags(struct request
*rq
, char *msg
)
754 printk(KERN_INFO
"%s: dev %s: flags=%llx\n", msg
,
755 rq
->q
->disk
? rq
->q
->disk
->disk_name
: "?",
756 (__force
unsigned long long) rq
->cmd_flags
);
758 printk(KERN_INFO
" sector %llu, nr/cnr %u/%u\n",
759 (unsigned long long)blk_rq_pos(rq
),
760 blk_rq_sectors(rq
), blk_rq_cur_sectors(rq
));
761 printk(KERN_INFO
" bio %p, biotail %p, len %u\n",
762 rq
->bio
, rq
->biotail
, blk_rq_bytes(rq
));
764 EXPORT_SYMBOL(blk_dump_rq_flags
);
766 static void req_bio_endio(struct request
*rq
, struct bio
*bio
,
767 unsigned int nbytes
, blk_status_t error
)
769 if (unlikely(error
)) {
770 bio
->bi_status
= error
;
771 } else if (req_op(rq
) == REQ_OP_ZONE_APPEND
) {
773 * Partial zone append completions cannot be supported as the
774 * BIO fragments may end up not being written sequentially.
776 if (bio
->bi_iter
.bi_size
!= nbytes
)
777 bio
->bi_status
= BLK_STS_IOERR
;
779 bio
->bi_iter
.bi_sector
= rq
->__sector
;
782 bio_advance(bio
, nbytes
);
784 if (unlikely(rq
->rq_flags
& RQF_QUIET
))
785 bio_set_flag(bio
, BIO_QUIET
);
786 /* don't actually finish bio if it's part of flush sequence */
787 if (bio
->bi_iter
.bi_size
== 0 && !(rq
->rq_flags
& RQF_FLUSH_SEQ
))
791 static void blk_account_io_completion(struct request
*req
, unsigned int bytes
)
793 if (req
->part
&& blk_do_io_stat(req
)) {
794 const int sgrp
= op_stat_group(req_op(req
));
797 part_stat_add(req
->part
, sectors
[sgrp
], bytes
>> 9);
802 static void blk_print_req_error(struct request
*req
, blk_status_t status
)
804 printk_ratelimited(KERN_ERR
805 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
806 "phys_seg %u prio class %u\n",
807 blk_status_to_str(status
),
808 req
->q
->disk
? req
->q
->disk
->disk_name
: "?",
809 blk_rq_pos(req
), (__force u32
)req_op(req
),
810 blk_op_str(req_op(req
)),
811 (__force u32
)(req
->cmd_flags
& ~REQ_OP_MASK
),
812 req
->nr_phys_segments
,
813 IOPRIO_PRIO_CLASS(req
->ioprio
));
817 * Fully end IO on a request. Does not support partial completions, or
820 static void blk_complete_request(struct request
*req
)
822 const bool is_flush
= (req
->rq_flags
& RQF_FLUSH_SEQ
) != 0;
823 int total_bytes
= blk_rq_bytes(req
);
824 struct bio
*bio
= req
->bio
;
826 trace_block_rq_complete(req
, BLK_STS_OK
, total_bytes
);
831 #ifdef CONFIG_BLK_DEV_INTEGRITY
832 if (blk_integrity_rq(req
) && req_op(req
) == REQ_OP_READ
)
833 req
->q
->integrity
.profile
->complete_fn(req
, total_bytes
);
837 * Upper layers may call blk_crypto_evict_key() anytime after the last
838 * bio_endio(). Therefore, the keyslot must be released before that.
840 blk_crypto_rq_put_keyslot(req
);
842 blk_account_io_completion(req
, total_bytes
);
845 struct bio
*next
= bio
->bi_next
;
847 /* Completion has already been traced */
848 bio_clear_flag(bio
, BIO_TRACE_COMPLETION
);
850 if (req_op(req
) == REQ_OP_ZONE_APPEND
)
851 bio
->bi_iter
.bi_sector
= req
->__sector
;
859 * Reset counters so that the request stacking driver
860 * can find how many bytes remain in the request
870 * blk_update_request - Complete multiple bytes without completing the request
871 * @req: the request being processed
872 * @error: block status code
873 * @nr_bytes: number of bytes to complete for @req
876 * Ends I/O on a number of bytes attached to @req, but doesn't complete
877 * the request structure even if @req doesn't have leftover.
878 * If @req has leftover, sets it up for the next range of segments.
880 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
881 * %false return from this function.
884 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
885 * except in the consistency check at the end of this function.
888 * %false - this request doesn't have any more data
889 * %true - this request has more data
891 bool blk_update_request(struct request
*req
, blk_status_t error
,
892 unsigned int nr_bytes
)
896 trace_block_rq_complete(req
, error
, nr_bytes
);
901 #ifdef CONFIG_BLK_DEV_INTEGRITY
902 if (blk_integrity_rq(req
) && req_op(req
) == REQ_OP_READ
&&
904 req
->q
->integrity
.profile
->complete_fn(req
, nr_bytes
);
908 * Upper layers may call blk_crypto_evict_key() anytime after the last
909 * bio_endio(). Therefore, the keyslot must be released before that.
911 if (blk_crypto_rq_has_keyslot(req
) && nr_bytes
>= blk_rq_bytes(req
))
912 __blk_crypto_rq_put_keyslot(req
);
914 if (unlikely(error
&& !blk_rq_is_passthrough(req
) &&
915 !(req
->rq_flags
& RQF_QUIET
)) &&
916 !test_bit(GD_DEAD
, &req
->q
->disk
->state
)) {
917 blk_print_req_error(req
, error
);
918 trace_block_rq_error(req
, error
, nr_bytes
);
921 blk_account_io_completion(req
, nr_bytes
);
925 struct bio
*bio
= req
->bio
;
926 unsigned bio_bytes
= min(bio
->bi_iter
.bi_size
, nr_bytes
);
928 if (bio_bytes
== bio
->bi_iter
.bi_size
)
929 req
->bio
= bio
->bi_next
;
931 /* Completion has already been traced */
932 bio_clear_flag(bio
, BIO_TRACE_COMPLETION
);
933 req_bio_endio(req
, bio
, bio_bytes
, error
);
935 total_bytes
+= bio_bytes
;
936 nr_bytes
-= bio_bytes
;
947 * Reset counters so that the request stacking driver
948 * can find how many bytes remain in the request
955 req
->__data_len
-= total_bytes
;
957 /* update sector only for requests with clear definition of sector */
958 if (!blk_rq_is_passthrough(req
))
959 req
->__sector
+= total_bytes
>> 9;
961 /* mixed attributes always follow the first bio */
962 if (req
->rq_flags
& RQF_MIXED_MERGE
) {
963 req
->cmd_flags
&= ~REQ_FAILFAST_MASK
;
964 req
->cmd_flags
|= req
->bio
->bi_opf
& REQ_FAILFAST_MASK
;
967 if (!(req
->rq_flags
& RQF_SPECIAL_PAYLOAD
)) {
969 * If total number of sectors is less than the first segment
970 * size, something has gone terribly wrong.
972 if (blk_rq_bytes(req
) < blk_rq_cur_bytes(req
)) {
973 blk_dump_rq_flags(req
, "request botched");
974 req
->__data_len
= blk_rq_cur_bytes(req
);
977 /* recalculate the number of segments */
978 req
->nr_phys_segments
= blk_recalc_rq_segments(req
);
983 EXPORT_SYMBOL_GPL(blk_update_request
);
985 static inline void blk_account_io_done(struct request
*req
, u64 now
)
987 trace_block_io_done(req
);
990 * Account IO completion. flush_rq isn't accounted as a
991 * normal IO on queueing nor completion. Accounting the
992 * containing request is enough.
994 if (blk_do_io_stat(req
) && req
->part
&&
995 !(req
->rq_flags
& RQF_FLUSH_SEQ
)) {
996 const int sgrp
= op_stat_group(req_op(req
));
999 update_io_ticks(req
->part
, jiffies
, true);
1000 part_stat_inc(req
->part
, ios
[sgrp
]);
1001 part_stat_add(req
->part
, nsecs
[sgrp
], now
- req
->start_time_ns
);
1006 static inline void blk_account_io_start(struct request
*req
)
1008 trace_block_io_start(req
);
1010 if (blk_do_io_stat(req
)) {
1012 * All non-passthrough requests are created from a bio with one
1013 * exception: when a flush command that is part of a flush sequence
1014 * generated by the state machine in blk-flush.c is cloned onto the
1015 * lower device by dm-multipath we can get here without a bio.
1018 req
->part
= req
->bio
->bi_bdev
;
1020 req
->part
= req
->q
->disk
->part0
;
1023 update_io_ticks(req
->part
, jiffies
, false);
1028 static inline void __blk_mq_end_request_acct(struct request
*rq
, u64 now
)
1030 if (rq
->rq_flags
& RQF_STATS
)
1031 blk_stat_add(rq
, now
);
1033 blk_mq_sched_completed_request(rq
, now
);
1034 blk_account_io_done(rq
, now
);
1037 inline void __blk_mq_end_request(struct request
*rq
, blk_status_t error
)
1039 if (blk_mq_need_time_stamp(rq
))
1040 __blk_mq_end_request_acct(rq
, ktime_get_ns());
1042 blk_mq_finish_request(rq
);
1045 rq_qos_done(rq
->q
, rq
);
1046 if (rq
->end_io(rq
, error
) == RQ_END_IO_FREE
)
1047 blk_mq_free_request(rq
);
1049 blk_mq_free_request(rq
);
1052 EXPORT_SYMBOL(__blk_mq_end_request
);
1054 void blk_mq_end_request(struct request
*rq
, blk_status_t error
)
1056 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
1058 __blk_mq_end_request(rq
, error
);
1060 EXPORT_SYMBOL(blk_mq_end_request
);
1062 #define TAG_COMP_BATCH 32
1064 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx
*hctx
,
1065 int *tag_array
, int nr_tags
)
1067 struct request_queue
*q
= hctx
->queue
;
1069 blk_mq_sub_active_requests(hctx
, nr_tags
);
1071 blk_mq_put_tags(hctx
->tags
, tag_array
, nr_tags
);
1072 percpu_ref_put_many(&q
->q_usage_counter
, nr_tags
);
1075 void blk_mq_end_request_batch(struct io_comp_batch
*iob
)
1077 int tags
[TAG_COMP_BATCH
], nr_tags
= 0;
1078 struct blk_mq_hw_ctx
*cur_hctx
= NULL
;
1083 now
= ktime_get_ns();
1085 while ((rq
= rq_list_pop(&iob
->req_list
)) != NULL
) {
1087 prefetch(rq
->rq_next
);
1089 blk_complete_request(rq
);
1091 __blk_mq_end_request_acct(rq
, now
);
1093 blk_mq_finish_request(rq
);
1095 rq_qos_done(rq
->q
, rq
);
1098 * If end_io handler returns NONE, then it still has
1099 * ownership of the request.
1101 if (rq
->end_io
&& rq
->end_io(rq
, 0) == RQ_END_IO_NONE
)
1104 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
1105 if (!req_ref_put_and_test(rq
))
1108 blk_crypto_free_request(rq
);
1109 blk_pm_mark_last_busy(rq
);
1111 if (nr_tags
== TAG_COMP_BATCH
|| cur_hctx
!= rq
->mq_hctx
) {
1113 blk_mq_flush_tag_batch(cur_hctx
, tags
, nr_tags
);
1115 cur_hctx
= rq
->mq_hctx
;
1117 tags
[nr_tags
++] = rq
->tag
;
1121 blk_mq_flush_tag_batch(cur_hctx
, tags
, nr_tags
);
1123 EXPORT_SYMBOL_GPL(blk_mq_end_request_batch
);
1125 static void blk_complete_reqs(struct llist_head
*list
)
1127 struct llist_node
*entry
= llist_reverse_order(llist_del_all(list
));
1128 struct request
*rq
, *next
;
1130 llist_for_each_entry_safe(rq
, next
, entry
, ipi_list
)
1131 rq
->q
->mq_ops
->complete(rq
);
1134 static __latent_entropy
void blk_done_softirq(struct softirq_action
*h
)
1136 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done
));
1139 static int blk_softirq_cpu_dead(unsigned int cpu
)
1141 blk_complete_reqs(&per_cpu(blk_cpu_done
, cpu
));
1145 static void __blk_mq_complete_request_remote(void *data
)
1147 __raise_softirq_irqoff(BLOCK_SOFTIRQ
);
1150 static inline bool blk_mq_complete_need_ipi(struct request
*rq
)
1152 int cpu
= raw_smp_processor_id();
1154 if (!IS_ENABLED(CONFIG_SMP
) ||
1155 !test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
))
1158 * With force threaded interrupts enabled, raising softirq from an SMP
1159 * function call will always result in waking the ksoftirqd thread.
1160 * This is probably worse than completing the request on a different
1163 if (force_irqthreads())
1166 /* same CPU or cache domain? Complete locally */
1167 if (cpu
== rq
->mq_ctx
->cpu
||
1168 (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
) &&
1169 cpus_share_cache(cpu
, rq
->mq_ctx
->cpu
)))
1172 /* don't try to IPI to an offline CPU */
1173 return cpu_online(rq
->mq_ctx
->cpu
);
1176 static void blk_mq_complete_send_ipi(struct request
*rq
)
1180 cpu
= rq
->mq_ctx
->cpu
;
1181 if (llist_add(&rq
->ipi_list
, &per_cpu(blk_cpu_done
, cpu
)))
1182 smp_call_function_single_async(cpu
, &per_cpu(blk_cpu_csd
, cpu
));
1185 static void blk_mq_raise_softirq(struct request
*rq
)
1187 struct llist_head
*list
;
1190 list
= this_cpu_ptr(&blk_cpu_done
);
1191 if (llist_add(&rq
->ipi_list
, list
))
1192 raise_softirq(BLOCK_SOFTIRQ
);
1196 bool blk_mq_complete_request_remote(struct request
*rq
)
1198 WRITE_ONCE(rq
->state
, MQ_RQ_COMPLETE
);
1201 * For request which hctx has only one ctx mapping,
1202 * or a polled request, always complete locally,
1203 * it's pointless to redirect the completion.
1205 if ((rq
->mq_hctx
->nr_ctx
== 1 &&
1206 rq
->mq_ctx
->cpu
== raw_smp_processor_id()) ||
1207 rq
->cmd_flags
& REQ_POLLED
)
1210 if (blk_mq_complete_need_ipi(rq
)) {
1211 blk_mq_complete_send_ipi(rq
);
1215 if (rq
->q
->nr_hw_queues
== 1) {
1216 blk_mq_raise_softirq(rq
);
1221 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote
);
1224 * blk_mq_complete_request - end I/O on a request
1225 * @rq: the request being processed
1228 * Complete a request by scheduling the ->complete_rq operation.
1230 void blk_mq_complete_request(struct request
*rq
)
1232 if (!blk_mq_complete_request_remote(rq
))
1233 rq
->q
->mq_ops
->complete(rq
);
1235 EXPORT_SYMBOL(blk_mq_complete_request
);
1238 * blk_mq_start_request - Start processing a request
1239 * @rq: Pointer to request to be started
1241 * Function used by device drivers to notify the block layer that a request
1242 * is going to be processed now, so blk layer can do proper initializations
1243 * such as starting the timeout timer.
1245 void blk_mq_start_request(struct request
*rq
)
1247 struct request_queue
*q
= rq
->q
;
1249 trace_block_rq_issue(rq
);
1251 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
1252 rq
->io_start_time_ns
= ktime_get_ns();
1253 rq
->stats_sectors
= blk_rq_sectors(rq
);
1254 rq
->rq_flags
|= RQF_STATS
;
1255 rq_qos_issue(q
, rq
);
1258 WARN_ON_ONCE(blk_mq_rq_state(rq
) != MQ_RQ_IDLE
);
1261 WRITE_ONCE(rq
->state
, MQ_RQ_IN_FLIGHT
);
1262 rq
->mq_hctx
->tags
->rqs
[rq
->tag
] = rq
;
1264 #ifdef CONFIG_BLK_DEV_INTEGRITY
1265 if (blk_integrity_rq(rq
) && req_op(rq
) == REQ_OP_WRITE
)
1266 q
->integrity
.profile
->prepare_fn(rq
);
1268 if (rq
->bio
&& rq
->bio
->bi_opf
& REQ_POLLED
)
1269 WRITE_ONCE(rq
->bio
->bi_cookie
, rq
->mq_hctx
->queue_num
);
1271 EXPORT_SYMBOL(blk_mq_start_request
);
1274 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1275 * queues. This is important for md arrays to benefit from merging
1278 static inline unsigned short blk_plug_max_rq_count(struct blk_plug
*plug
)
1280 if (plug
->multiple_queues
)
1281 return BLK_MAX_REQUEST_COUNT
* 2;
1282 return BLK_MAX_REQUEST_COUNT
;
1285 static void blk_add_rq_to_plug(struct blk_plug
*plug
, struct request
*rq
)
1287 struct request
*last
= rq_list_peek(&plug
->mq_list
);
1289 if (!plug
->rq_count
) {
1290 trace_block_plug(rq
->q
);
1291 } else if (plug
->rq_count
>= blk_plug_max_rq_count(plug
) ||
1292 (!blk_queue_nomerges(rq
->q
) &&
1293 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1294 blk_mq_flush_plug_list(plug
, false);
1296 trace_block_plug(rq
->q
);
1299 if (!plug
->multiple_queues
&& last
&& last
->q
!= rq
->q
)
1300 plug
->multiple_queues
= true;
1302 * Any request allocated from sched tags can't be issued to
1303 * ->queue_rqs() directly
1305 if (!plug
->has_elevator
&& (rq
->rq_flags
& RQF_SCHED_TAGS
))
1306 plug
->has_elevator
= true;
1308 rq_list_add(&plug
->mq_list
, rq
);
1313 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1314 * @rq: request to insert
1315 * @at_head: insert request at head or tail of queue
1318 * Insert a fully prepared request at the back of the I/O scheduler queue
1319 * for execution. Don't wait for completion.
1322 * This function will invoke @done directly if the queue is dead.
1324 void blk_execute_rq_nowait(struct request
*rq
, bool at_head
)
1326 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1328 WARN_ON(irqs_disabled());
1329 WARN_ON(!blk_rq_is_passthrough(rq
));
1331 blk_account_io_start(rq
);
1334 * As plugging can be enabled for passthrough requests on a zoned
1335 * device, directly accessing the plug instead of using blk_mq_plug()
1336 * should not have any consequences.
1338 if (current
->plug
&& !at_head
) {
1339 blk_add_rq_to_plug(current
->plug
, rq
);
1343 blk_mq_insert_request(rq
, at_head
? BLK_MQ_INSERT_AT_HEAD
: 0);
1344 blk_mq_run_hw_queue(hctx
, hctx
->flags
& BLK_MQ_F_BLOCKING
);
1346 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait
);
1348 struct blk_rq_wait
{
1349 struct completion done
;
1353 static enum rq_end_io_ret
blk_end_sync_rq(struct request
*rq
, blk_status_t ret
)
1355 struct blk_rq_wait
*wait
= rq
->end_io_data
;
1358 complete(&wait
->done
);
1359 return RQ_END_IO_NONE
;
1362 bool blk_rq_is_poll(struct request
*rq
)
1366 if (rq
->mq_hctx
->type
!= HCTX_TYPE_POLL
)
1370 EXPORT_SYMBOL_GPL(blk_rq_is_poll
);
1372 static void blk_rq_poll_completion(struct request
*rq
, struct completion
*wait
)
1375 blk_hctx_poll(rq
->q
, rq
->mq_hctx
, NULL
, 0);
1377 } while (!completion_done(wait
));
1381 * blk_execute_rq - insert a request into queue for execution
1382 * @rq: request to insert
1383 * @at_head: insert request at head or tail of queue
1386 * Insert a fully prepared request at the back of the I/O scheduler queue
1387 * for execution and wait for completion.
1388 * Return: The blk_status_t result provided to blk_mq_end_request().
1390 blk_status_t
blk_execute_rq(struct request
*rq
, bool at_head
)
1392 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1393 struct blk_rq_wait wait
= {
1394 .done
= COMPLETION_INITIALIZER_ONSTACK(wait
.done
),
1397 WARN_ON(irqs_disabled());
1398 WARN_ON(!blk_rq_is_passthrough(rq
));
1400 rq
->end_io_data
= &wait
;
1401 rq
->end_io
= blk_end_sync_rq
;
1403 blk_account_io_start(rq
);
1404 blk_mq_insert_request(rq
, at_head
? BLK_MQ_INSERT_AT_HEAD
: 0);
1405 blk_mq_run_hw_queue(hctx
, false);
1407 if (blk_rq_is_poll(rq
)) {
1408 blk_rq_poll_completion(rq
, &wait
.done
);
1411 * Prevent hang_check timer from firing at us during very long
1414 unsigned long hang_check
= sysctl_hung_task_timeout_secs
;
1417 while (!wait_for_completion_io_timeout(&wait
.done
,
1418 hang_check
* (HZ
/2)))
1421 wait_for_completion_io(&wait
.done
);
1426 EXPORT_SYMBOL(blk_execute_rq
);
1428 static void __blk_mq_requeue_request(struct request
*rq
)
1430 struct request_queue
*q
= rq
->q
;
1432 blk_mq_put_driver_tag(rq
);
1434 trace_block_rq_requeue(rq
);
1435 rq_qos_requeue(q
, rq
);
1437 if (blk_mq_request_started(rq
)) {
1438 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
1439 rq
->rq_flags
&= ~RQF_TIMED_OUT
;
1443 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
1445 struct request_queue
*q
= rq
->q
;
1446 unsigned long flags
;
1448 __blk_mq_requeue_request(rq
);
1450 /* this request will be re-inserted to io scheduler queue */
1451 blk_mq_sched_requeue_request(rq
);
1453 spin_lock_irqsave(&q
->requeue_lock
, flags
);
1454 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
1455 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
1457 if (kick_requeue_list
)
1458 blk_mq_kick_requeue_list(q
);
1460 EXPORT_SYMBOL(blk_mq_requeue_request
);
1462 static void blk_mq_requeue_work(struct work_struct
*work
)
1464 struct request_queue
*q
=
1465 container_of(work
, struct request_queue
, requeue_work
.work
);
1467 LIST_HEAD(flush_list
);
1470 spin_lock_irq(&q
->requeue_lock
);
1471 list_splice_init(&q
->requeue_list
, &rq_list
);
1472 list_splice_init(&q
->flush_list
, &flush_list
);
1473 spin_unlock_irq(&q
->requeue_lock
);
1475 while (!list_empty(&rq_list
)) {
1476 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
1478 * If RQF_DONTPREP ist set, the request has been started by the
1479 * driver already and might have driver-specific data allocated
1480 * already. Insert it into the hctx dispatch list to avoid
1481 * block layer merges for the request.
1483 if (rq
->rq_flags
& RQF_DONTPREP
) {
1484 list_del_init(&rq
->queuelist
);
1485 blk_mq_request_bypass_insert(rq
, 0);
1487 list_del_init(&rq
->queuelist
);
1488 blk_mq_insert_request(rq
, BLK_MQ_INSERT_AT_HEAD
);
1492 while (!list_empty(&flush_list
)) {
1493 rq
= list_entry(flush_list
.next
, struct request
, queuelist
);
1494 list_del_init(&rq
->queuelist
);
1495 blk_mq_insert_request(rq
, 0);
1498 blk_mq_run_hw_queues(q
, false);
1501 void blk_mq_kick_requeue_list(struct request_queue
*q
)
1503 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
, 0);
1505 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
1507 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
1508 unsigned long msecs
)
1510 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
,
1511 msecs_to_jiffies(msecs
));
1513 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
1515 static bool blk_mq_rq_inflight(struct request
*rq
, void *priv
)
1518 * If we find a request that isn't idle we know the queue is busy
1519 * as it's checked in the iter.
1520 * Return false to stop the iteration.
1522 if (blk_mq_request_started(rq
)) {
1532 bool blk_mq_queue_inflight(struct request_queue
*q
)
1536 blk_mq_queue_tag_busy_iter(q
, blk_mq_rq_inflight
, &busy
);
1539 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight
);
1541 static void blk_mq_rq_timed_out(struct request
*req
)
1543 req
->rq_flags
|= RQF_TIMED_OUT
;
1544 if (req
->q
->mq_ops
->timeout
) {
1545 enum blk_eh_timer_return ret
;
1547 ret
= req
->q
->mq_ops
->timeout(req
);
1548 if (ret
== BLK_EH_DONE
)
1550 WARN_ON_ONCE(ret
!= BLK_EH_RESET_TIMER
);
1556 struct blk_expired_data
{
1557 bool has_timedout_rq
;
1559 unsigned long timeout_start
;
1562 static bool blk_mq_req_expired(struct request
*rq
, struct blk_expired_data
*expired
)
1564 unsigned long deadline
;
1566 if (blk_mq_rq_state(rq
) != MQ_RQ_IN_FLIGHT
)
1568 if (rq
->rq_flags
& RQF_TIMED_OUT
)
1571 deadline
= READ_ONCE(rq
->deadline
);
1572 if (time_after_eq(expired
->timeout_start
, deadline
))
1575 if (expired
->next
== 0)
1576 expired
->next
= deadline
;
1577 else if (time_after(expired
->next
, deadline
))
1578 expired
->next
= deadline
;
1582 void blk_mq_put_rq_ref(struct request
*rq
)
1584 if (is_flush_rq(rq
)) {
1585 if (rq
->end_io(rq
, 0) == RQ_END_IO_FREE
)
1586 blk_mq_free_request(rq
);
1587 } else if (req_ref_put_and_test(rq
)) {
1588 __blk_mq_free_request(rq
);
1592 static bool blk_mq_check_expired(struct request
*rq
, void *priv
)
1594 struct blk_expired_data
*expired
= priv
;
1597 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1598 * be reallocated underneath the timeout handler's processing, then
1599 * the expire check is reliable. If the request is not expired, then
1600 * it was completed and reallocated as a new request after returning
1601 * from blk_mq_check_expired().
1603 if (blk_mq_req_expired(rq
, expired
)) {
1604 expired
->has_timedout_rq
= true;
1610 static bool blk_mq_handle_expired(struct request
*rq
, void *priv
)
1612 struct blk_expired_data
*expired
= priv
;
1614 if (blk_mq_req_expired(rq
, expired
))
1615 blk_mq_rq_timed_out(rq
);
1619 static void blk_mq_timeout_work(struct work_struct
*work
)
1621 struct request_queue
*q
=
1622 container_of(work
, struct request_queue
, timeout_work
);
1623 struct blk_expired_data expired
= {
1624 .timeout_start
= jiffies
,
1626 struct blk_mq_hw_ctx
*hctx
;
1629 /* A deadlock might occur if a request is stuck requiring a
1630 * timeout at the same time a queue freeze is waiting
1631 * completion, since the timeout code would not be able to
1632 * acquire the queue reference here.
1634 * That's why we don't use blk_queue_enter here; instead, we use
1635 * percpu_ref_tryget directly, because we need to be able to
1636 * obtain a reference even in the short window between the queue
1637 * starting to freeze, by dropping the first reference in
1638 * blk_freeze_queue_start, and the moment the last request is
1639 * consumed, marked by the instant q_usage_counter reaches
1642 if (!percpu_ref_tryget(&q
->q_usage_counter
))
1645 /* check if there is any timed-out request */
1646 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &expired
);
1647 if (expired
.has_timedout_rq
) {
1649 * Before walking tags, we must ensure any submit started
1650 * before the current time has finished. Since the submit
1651 * uses srcu or rcu, wait for a synchronization point to
1652 * ensure all running submits have finished
1654 blk_mq_wait_quiesce_done(q
->tag_set
);
1657 blk_mq_queue_tag_busy_iter(q
, blk_mq_handle_expired
, &expired
);
1660 if (expired
.next
!= 0) {
1661 mod_timer(&q
->timeout
, expired
.next
);
1664 * Request timeouts are handled as a forward rolling timer. If
1665 * we end up here it means that no requests are pending and
1666 * also that no request has been pending for a while. Mark
1667 * each hctx as idle.
1669 queue_for_each_hw_ctx(q
, hctx
, i
) {
1670 /* the hctx may be unmapped, so check it here */
1671 if (blk_mq_hw_queue_mapped(hctx
))
1672 blk_mq_tag_idle(hctx
);
1678 struct flush_busy_ctx_data
{
1679 struct blk_mq_hw_ctx
*hctx
;
1680 struct list_head
*list
;
1683 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
1685 struct flush_busy_ctx_data
*flush_data
= data
;
1686 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
1687 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
1688 enum hctx_type type
= hctx
->type
;
1690 spin_lock(&ctx
->lock
);
1691 list_splice_tail_init(&ctx
->rq_lists
[type
], flush_data
->list
);
1692 sbitmap_clear_bit(sb
, bitnr
);
1693 spin_unlock(&ctx
->lock
);
1698 * Process software queues that have been marked busy, splicing them
1699 * to the for-dispatch
1701 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
1703 struct flush_busy_ctx_data data
= {
1708 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
1710 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
1712 struct dispatch_rq_data
{
1713 struct blk_mq_hw_ctx
*hctx
;
1717 static bool dispatch_rq_from_ctx(struct sbitmap
*sb
, unsigned int bitnr
,
1720 struct dispatch_rq_data
*dispatch_data
= data
;
1721 struct blk_mq_hw_ctx
*hctx
= dispatch_data
->hctx
;
1722 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
1723 enum hctx_type type
= hctx
->type
;
1725 spin_lock(&ctx
->lock
);
1726 if (!list_empty(&ctx
->rq_lists
[type
])) {
1727 dispatch_data
->rq
= list_entry_rq(ctx
->rq_lists
[type
].next
);
1728 list_del_init(&dispatch_data
->rq
->queuelist
);
1729 if (list_empty(&ctx
->rq_lists
[type
]))
1730 sbitmap_clear_bit(sb
, bitnr
);
1732 spin_unlock(&ctx
->lock
);
1734 return !dispatch_data
->rq
;
1737 struct request
*blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx
*hctx
,
1738 struct blk_mq_ctx
*start
)
1740 unsigned off
= start
? start
->index_hw
[hctx
->type
] : 0;
1741 struct dispatch_rq_data data
= {
1746 __sbitmap_for_each_set(&hctx
->ctx_map
, off
,
1747 dispatch_rq_from_ctx
, &data
);
1752 bool __blk_mq_alloc_driver_tag(struct request
*rq
)
1754 struct sbitmap_queue
*bt
= &rq
->mq_hctx
->tags
->bitmap_tags
;
1755 unsigned int tag_offset
= rq
->mq_hctx
->tags
->nr_reserved_tags
;
1758 blk_mq_tag_busy(rq
->mq_hctx
);
1760 if (blk_mq_tag_is_reserved(rq
->mq_hctx
->sched_tags
, rq
->internal_tag
)) {
1761 bt
= &rq
->mq_hctx
->tags
->breserved_tags
;
1764 if (!hctx_may_queue(rq
->mq_hctx
, bt
))
1768 tag
= __sbitmap_queue_get(bt
);
1769 if (tag
== BLK_MQ_NO_TAG
)
1772 rq
->tag
= tag
+ tag_offset
;
1773 blk_mq_inc_active_requests(rq
->mq_hctx
);
1777 static int blk_mq_dispatch_wake(wait_queue_entry_t
*wait
, unsigned mode
,
1778 int flags
, void *key
)
1780 struct blk_mq_hw_ctx
*hctx
;
1782 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
1784 spin_lock(&hctx
->dispatch_wait_lock
);
1785 if (!list_empty(&wait
->entry
)) {
1786 struct sbitmap_queue
*sbq
;
1788 list_del_init(&wait
->entry
);
1789 sbq
= &hctx
->tags
->bitmap_tags
;
1790 atomic_dec(&sbq
->ws_active
);
1792 spin_unlock(&hctx
->dispatch_wait_lock
);
1794 blk_mq_run_hw_queue(hctx
, true);
1799 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1800 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1801 * restart. For both cases, take care to check the condition again after
1802 * marking us as waiting.
1804 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx
*hctx
,
1807 struct sbitmap_queue
*sbq
;
1808 struct wait_queue_head
*wq
;
1809 wait_queue_entry_t
*wait
;
1812 if (!(hctx
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
) &&
1813 !(blk_mq_is_shared_tags(hctx
->flags
))) {
1814 blk_mq_sched_mark_restart_hctx(hctx
);
1817 * It's possible that a tag was freed in the window between the
1818 * allocation failure and adding the hardware queue to the wait
1821 * Don't clear RESTART here, someone else could have set it.
1822 * At most this will cost an extra queue run.
1824 return blk_mq_get_driver_tag(rq
);
1827 wait
= &hctx
->dispatch_wait
;
1828 if (!list_empty_careful(&wait
->entry
))
1831 if (blk_mq_tag_is_reserved(rq
->mq_hctx
->sched_tags
, rq
->internal_tag
))
1832 sbq
= &hctx
->tags
->breserved_tags
;
1834 sbq
= &hctx
->tags
->bitmap_tags
;
1835 wq
= &bt_wait_ptr(sbq
, hctx
)->wait
;
1837 spin_lock_irq(&wq
->lock
);
1838 spin_lock(&hctx
->dispatch_wait_lock
);
1839 if (!list_empty(&wait
->entry
)) {
1840 spin_unlock(&hctx
->dispatch_wait_lock
);
1841 spin_unlock_irq(&wq
->lock
);
1845 atomic_inc(&sbq
->ws_active
);
1846 wait
->flags
&= ~WQ_FLAG_EXCLUSIVE
;
1847 __add_wait_queue(wq
, wait
);
1850 * It's possible that a tag was freed in the window between the
1851 * allocation failure and adding the hardware queue to the wait
1854 ret
= blk_mq_get_driver_tag(rq
);
1856 spin_unlock(&hctx
->dispatch_wait_lock
);
1857 spin_unlock_irq(&wq
->lock
);
1862 * We got a tag, remove ourselves from the wait queue to ensure
1863 * someone else gets the wakeup.
1865 list_del_init(&wait
->entry
);
1866 atomic_dec(&sbq
->ws_active
);
1867 spin_unlock(&hctx
->dispatch_wait_lock
);
1868 spin_unlock_irq(&wq
->lock
);
1873 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1874 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1876 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1877 * - EWMA is one simple way to compute running average value
1878 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1879 * - take 4 as factor for avoiding to get too small(0) result, and this
1880 * factor doesn't matter because EWMA decreases exponentially
1882 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx
*hctx
, bool busy
)
1886 ewma
= hctx
->dispatch_busy
;
1891 ewma
*= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
- 1;
1893 ewma
+= 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR
;
1894 ewma
/= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
;
1896 hctx
->dispatch_busy
= ewma
;
1899 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1901 static void blk_mq_handle_dev_resource(struct request
*rq
,
1902 struct list_head
*list
)
1904 list_add(&rq
->queuelist
, list
);
1905 __blk_mq_requeue_request(rq
);
1908 static void blk_mq_handle_zone_resource(struct request
*rq
,
1909 struct list_head
*zone_list
)
1912 * If we end up here it is because we cannot dispatch a request to a
1913 * specific zone due to LLD level zone-write locking or other zone
1914 * related resource not being available. In this case, set the request
1915 * aside in zone_list for retrying it later.
1917 list_add(&rq
->queuelist
, zone_list
);
1918 __blk_mq_requeue_request(rq
);
1921 enum prep_dispatch
{
1923 PREP_DISPATCH_NO_TAG
,
1924 PREP_DISPATCH_NO_BUDGET
,
1927 static enum prep_dispatch
blk_mq_prep_dispatch_rq(struct request
*rq
,
1930 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1931 int budget_token
= -1;
1934 budget_token
= blk_mq_get_dispatch_budget(rq
->q
);
1935 if (budget_token
< 0) {
1936 blk_mq_put_driver_tag(rq
);
1937 return PREP_DISPATCH_NO_BUDGET
;
1939 blk_mq_set_rq_budget_token(rq
, budget_token
);
1942 if (!blk_mq_get_driver_tag(rq
)) {
1944 * The initial allocation attempt failed, so we need to
1945 * rerun the hardware queue when a tag is freed. The
1946 * waitqueue takes care of that. If the queue is run
1947 * before we add this entry back on the dispatch list,
1948 * we'll re-run it below.
1950 if (!blk_mq_mark_tag_wait(hctx
, rq
)) {
1952 * All budgets not got from this function will be put
1953 * together during handling partial dispatch
1956 blk_mq_put_dispatch_budget(rq
->q
, budget_token
);
1957 return PREP_DISPATCH_NO_TAG
;
1961 return PREP_DISPATCH_OK
;
1964 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1965 static void blk_mq_release_budgets(struct request_queue
*q
,
1966 struct list_head
*list
)
1970 list_for_each_entry(rq
, list
, queuelist
) {
1971 int budget_token
= blk_mq_get_rq_budget_token(rq
);
1973 if (budget_token
>= 0)
1974 blk_mq_put_dispatch_budget(q
, budget_token
);
1979 * blk_mq_commit_rqs will notify driver using bd->last that there is no
1980 * more requests. (See comment in struct blk_mq_ops for commit_rqs for
1982 * Attention, we should explicitly call this in unusual cases:
1983 * 1) did not queue everything initially scheduled to queue
1984 * 2) the last attempt to queue a request failed
1986 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx
*hctx
, int queued
,
1989 if (hctx
->queue
->mq_ops
->commit_rqs
&& queued
) {
1990 trace_block_unplug(hctx
->queue
, queued
, !from_schedule
);
1991 hctx
->queue
->mq_ops
->commit_rqs(hctx
);
1996 * Returns true if we did some work AND can potentially do more.
1998 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
,
1999 unsigned int nr_budgets
)
2001 enum prep_dispatch prep
;
2002 struct request_queue
*q
= hctx
->queue
;
2005 blk_status_t ret
= BLK_STS_OK
;
2006 LIST_HEAD(zone_list
);
2007 bool needs_resource
= false;
2009 if (list_empty(list
))
2013 * Now process all the entries, sending them to the driver.
2017 struct blk_mq_queue_data bd
;
2019 rq
= list_first_entry(list
, struct request
, queuelist
);
2021 WARN_ON_ONCE(hctx
!= rq
->mq_hctx
);
2022 prep
= blk_mq_prep_dispatch_rq(rq
, !nr_budgets
);
2023 if (prep
!= PREP_DISPATCH_OK
)
2026 list_del_init(&rq
->queuelist
);
2029 bd
.last
= list_empty(list
);
2032 * once the request is queued to lld, no need to cover the
2037 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
2042 case BLK_STS_RESOURCE
:
2043 needs_resource
= true;
2045 case BLK_STS_DEV_RESOURCE
:
2046 blk_mq_handle_dev_resource(rq
, list
);
2048 case BLK_STS_ZONE_RESOURCE
:
2050 * Move the request to zone_list and keep going through
2051 * the dispatch list to find more requests the drive can
2054 blk_mq_handle_zone_resource(rq
, &zone_list
);
2055 needs_resource
= true;
2058 blk_mq_end_request(rq
, ret
);
2060 } while (!list_empty(list
));
2062 if (!list_empty(&zone_list
))
2063 list_splice_tail_init(&zone_list
, list
);
2065 /* If we didn't flush the entire list, we could have told the driver
2066 * there was more coming, but that turned out to be a lie.
2068 if (!list_empty(list
) || ret
!= BLK_STS_OK
)
2069 blk_mq_commit_rqs(hctx
, queued
, false);
2072 * Any items that need requeuing? Stuff them into hctx->dispatch,
2073 * that is where we will continue on next queue run.
2075 if (!list_empty(list
)) {
2077 /* For non-shared tags, the RESTART check will suffice */
2078 bool no_tag
= prep
== PREP_DISPATCH_NO_TAG
&&
2079 ((hctx
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
) ||
2080 blk_mq_is_shared_tags(hctx
->flags
));
2083 blk_mq_release_budgets(q
, list
);
2085 spin_lock(&hctx
->lock
);
2086 list_splice_tail_init(list
, &hctx
->dispatch
);
2087 spin_unlock(&hctx
->lock
);
2090 * Order adding requests to hctx->dispatch and checking
2091 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2092 * in blk_mq_sched_restart(). Avoid restart code path to
2093 * miss the new added requests to hctx->dispatch, meantime
2094 * SCHED_RESTART is observed here.
2099 * If SCHED_RESTART was set by the caller of this function and
2100 * it is no longer set that means that it was cleared by another
2101 * thread and hence that a queue rerun is needed.
2103 * If 'no_tag' is set, that means that we failed getting
2104 * a driver tag with an I/O scheduler attached. If our dispatch
2105 * waitqueue is no longer active, ensure that we run the queue
2106 * AFTER adding our entries back to the list.
2108 * If no I/O scheduler has been configured it is possible that
2109 * the hardware queue got stopped and restarted before requests
2110 * were pushed back onto the dispatch list. Rerun the queue to
2111 * avoid starvation. Notes:
2112 * - blk_mq_run_hw_queue() checks whether or not a queue has
2113 * been stopped before rerunning a queue.
2114 * - Some but not all block drivers stop a queue before
2115 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2118 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2119 * bit is set, run queue after a delay to avoid IO stalls
2120 * that could otherwise occur if the queue is idle. We'll do
2121 * similar if we couldn't get budget or couldn't lock a zone
2122 * and SCHED_RESTART is set.
2124 needs_restart
= blk_mq_sched_needs_restart(hctx
);
2125 if (prep
== PREP_DISPATCH_NO_BUDGET
)
2126 needs_resource
= true;
2127 if (!needs_restart
||
2128 (no_tag
&& list_empty_careful(&hctx
->dispatch_wait
.entry
)))
2129 blk_mq_run_hw_queue(hctx
, true);
2130 else if (needs_resource
)
2131 blk_mq_delay_run_hw_queue(hctx
, BLK_MQ_RESOURCE_DELAY
);
2133 blk_mq_update_dispatch_busy(hctx
, true);
2137 blk_mq_update_dispatch_busy(hctx
, false);
2141 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx
*hctx
)
2143 int cpu
= cpumask_first_and(hctx
->cpumask
, cpu_online_mask
);
2145 if (cpu
>= nr_cpu_ids
)
2146 cpu
= cpumask_first(hctx
->cpumask
);
2151 * It'd be great if the workqueue API had a way to pass
2152 * in a mask and had some smarts for more clever placement.
2153 * For now we just round-robin here, switching for every
2154 * BLK_MQ_CPU_WORK_BATCH queued items.
2156 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
2159 int next_cpu
= hctx
->next_cpu
;
2161 if (hctx
->queue
->nr_hw_queues
== 1)
2162 return WORK_CPU_UNBOUND
;
2164 if (--hctx
->next_cpu_batch
<= 0) {
2166 next_cpu
= cpumask_next_and(next_cpu
, hctx
->cpumask
,
2168 if (next_cpu
>= nr_cpu_ids
)
2169 next_cpu
= blk_mq_first_mapped_cpu(hctx
);
2170 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2174 * Do unbound schedule if we can't find a online CPU for this hctx,
2175 * and it should only happen in the path of handling CPU DEAD.
2177 if (!cpu_online(next_cpu
)) {
2184 * Make sure to re-select CPU next time once after CPUs
2185 * in hctx->cpumask become online again.
2187 hctx
->next_cpu
= next_cpu
;
2188 hctx
->next_cpu_batch
= 1;
2189 return WORK_CPU_UNBOUND
;
2192 hctx
->next_cpu
= next_cpu
;
2197 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2198 * @hctx: Pointer to the hardware queue to run.
2199 * @msecs: Milliseconds of delay to wait before running the queue.
2201 * Run a hardware queue asynchronously with a delay of @msecs.
2203 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
2205 if (unlikely(blk_mq_hctx_stopped(hctx
)))
2207 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
,
2208 msecs_to_jiffies(msecs
));
2210 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue
);
2213 * blk_mq_run_hw_queue - Start to run a hardware queue.
2214 * @hctx: Pointer to the hardware queue to run.
2215 * @async: If we want to run the queue asynchronously.
2217 * Check if the request queue is not in a quiesced state and if there are
2218 * pending requests to be sent. If this is true, run the queue to send requests
2221 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
2226 * We can't run the queue inline with interrupts disabled.
2228 WARN_ON_ONCE(!async
&& in_interrupt());
2230 might_sleep_if(!async
&& hctx
->flags
& BLK_MQ_F_BLOCKING
);
2233 * When queue is quiesced, we may be switching io scheduler, or
2234 * updating nr_hw_queues, or other things, and we can't run queue
2235 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2237 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2240 __blk_mq_run_dispatch_ops(hctx
->queue
, false,
2241 need_run
= !blk_queue_quiesced(hctx
->queue
) &&
2242 blk_mq_hctx_has_pending(hctx
));
2247 if (async
|| !cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
)) {
2248 blk_mq_delay_run_hw_queue(hctx
, 0);
2252 blk_mq_run_dispatch_ops(hctx
->queue
,
2253 blk_mq_sched_dispatch_requests(hctx
));
2255 EXPORT_SYMBOL(blk_mq_run_hw_queue
);
2258 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2261 static struct blk_mq_hw_ctx
*blk_mq_get_sq_hctx(struct request_queue
*q
)
2263 struct blk_mq_ctx
*ctx
= blk_mq_get_ctx(q
);
2265 * If the IO scheduler does not respect hardware queues when
2266 * dispatching, we just don't bother with multiple HW queues and
2267 * dispatch from hctx for the current CPU since running multiple queues
2268 * just causes lock contention inside the scheduler and pointless cache
2271 struct blk_mq_hw_ctx
*hctx
= ctx
->hctxs
[HCTX_TYPE_DEFAULT
];
2273 if (!blk_mq_hctx_stopped(hctx
))
2279 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2280 * @q: Pointer to the request queue to run.
2281 * @async: If we want to run the queue asynchronously.
2283 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
2285 struct blk_mq_hw_ctx
*hctx
, *sq_hctx
;
2289 if (blk_queue_sq_sched(q
))
2290 sq_hctx
= blk_mq_get_sq_hctx(q
);
2291 queue_for_each_hw_ctx(q
, hctx
, i
) {
2292 if (blk_mq_hctx_stopped(hctx
))
2295 * Dispatch from this hctx either if there's no hctx preferred
2296 * by IO scheduler or if it has requests that bypass the
2299 if (!sq_hctx
|| sq_hctx
== hctx
||
2300 !list_empty_careful(&hctx
->dispatch
))
2301 blk_mq_run_hw_queue(hctx
, async
);
2304 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
2307 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2308 * @q: Pointer to the request queue to run.
2309 * @msecs: Milliseconds of delay to wait before running the queues.
2311 void blk_mq_delay_run_hw_queues(struct request_queue
*q
, unsigned long msecs
)
2313 struct blk_mq_hw_ctx
*hctx
, *sq_hctx
;
2317 if (blk_queue_sq_sched(q
))
2318 sq_hctx
= blk_mq_get_sq_hctx(q
);
2319 queue_for_each_hw_ctx(q
, hctx
, i
) {
2320 if (blk_mq_hctx_stopped(hctx
))
2323 * If there is already a run_work pending, leave the
2324 * pending delay untouched. Otherwise, a hctx can stall
2325 * if another hctx is re-delaying the other's work
2326 * before the work executes.
2328 if (delayed_work_pending(&hctx
->run_work
))
2331 * Dispatch from this hctx either if there's no hctx preferred
2332 * by IO scheduler or if it has requests that bypass the
2335 if (!sq_hctx
|| sq_hctx
== hctx
||
2336 !list_empty_careful(&hctx
->dispatch
))
2337 blk_mq_delay_run_hw_queue(hctx
, msecs
);
2340 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues
);
2343 * This function is often used for pausing .queue_rq() by driver when
2344 * there isn't enough resource or some conditions aren't satisfied, and
2345 * BLK_STS_RESOURCE is usually returned.
2347 * We do not guarantee that dispatch can be drained or blocked
2348 * after blk_mq_stop_hw_queue() returns. Please use
2349 * blk_mq_quiesce_queue() for that requirement.
2351 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
2353 cancel_delayed_work(&hctx
->run_work
);
2355 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
2357 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
2360 * This function is often used for pausing .queue_rq() by driver when
2361 * there isn't enough resource or some conditions aren't satisfied, and
2362 * BLK_STS_RESOURCE is usually returned.
2364 * We do not guarantee that dispatch can be drained or blocked
2365 * after blk_mq_stop_hw_queues() returns. Please use
2366 * blk_mq_quiesce_queue() for that requirement.
2368 void blk_mq_stop_hw_queues(struct request_queue
*q
)
2370 struct blk_mq_hw_ctx
*hctx
;
2373 queue_for_each_hw_ctx(q
, hctx
, i
)
2374 blk_mq_stop_hw_queue(hctx
);
2376 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
2378 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
2380 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
2382 blk_mq_run_hw_queue(hctx
, hctx
->flags
& BLK_MQ_F_BLOCKING
);
2384 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
2386 void blk_mq_start_hw_queues(struct request_queue
*q
)
2388 struct blk_mq_hw_ctx
*hctx
;
2391 queue_for_each_hw_ctx(q
, hctx
, i
)
2392 blk_mq_start_hw_queue(hctx
);
2394 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
2396 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
2398 if (!blk_mq_hctx_stopped(hctx
))
2401 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
2402 blk_mq_run_hw_queue(hctx
, async
);
2404 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
2406 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
2408 struct blk_mq_hw_ctx
*hctx
;
2411 queue_for_each_hw_ctx(q
, hctx
, i
)
2412 blk_mq_start_stopped_hw_queue(hctx
, async
||
2413 (hctx
->flags
& BLK_MQ_F_BLOCKING
));
2415 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
2417 static void blk_mq_run_work_fn(struct work_struct
*work
)
2419 struct blk_mq_hw_ctx
*hctx
=
2420 container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
2422 blk_mq_run_dispatch_ops(hctx
->queue
,
2423 blk_mq_sched_dispatch_requests(hctx
));
2427 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2428 * @rq: Pointer to request to be inserted.
2429 * @flags: BLK_MQ_INSERT_*
2431 * Should only be used carefully, when the caller knows we want to
2432 * bypass a potential IO scheduler on the target device.
2434 static void blk_mq_request_bypass_insert(struct request
*rq
, blk_insert_t flags
)
2436 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
2438 spin_lock(&hctx
->lock
);
2439 if (flags
& BLK_MQ_INSERT_AT_HEAD
)
2440 list_add(&rq
->queuelist
, &hctx
->dispatch
);
2442 list_add_tail(&rq
->queuelist
, &hctx
->dispatch
);
2443 spin_unlock(&hctx
->lock
);
2446 static void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
,
2447 struct blk_mq_ctx
*ctx
, struct list_head
*list
,
2448 bool run_queue_async
)
2451 enum hctx_type type
= hctx
->type
;
2454 * Try to issue requests directly if the hw queue isn't busy to save an
2455 * extra enqueue & dequeue to the sw queue.
2457 if (!hctx
->dispatch_busy
&& !run_queue_async
) {
2458 blk_mq_run_dispatch_ops(hctx
->queue
,
2459 blk_mq_try_issue_list_directly(hctx
, list
));
2460 if (list_empty(list
))
2465 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2468 list_for_each_entry(rq
, list
, queuelist
) {
2469 BUG_ON(rq
->mq_ctx
!= ctx
);
2470 trace_block_rq_insert(rq
);
2471 if (rq
->cmd_flags
& REQ_NOWAIT
)
2472 run_queue_async
= true;
2475 spin_lock(&ctx
->lock
);
2476 list_splice_tail_init(list
, &ctx
->rq_lists
[type
]);
2477 blk_mq_hctx_mark_pending(hctx
, ctx
);
2478 spin_unlock(&ctx
->lock
);
2480 blk_mq_run_hw_queue(hctx
, run_queue_async
);
2483 static void blk_mq_insert_request(struct request
*rq
, blk_insert_t flags
)
2485 struct request_queue
*q
= rq
->q
;
2486 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
2487 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
2489 if (blk_rq_is_passthrough(rq
)) {
2491 * Passthrough request have to be added to hctx->dispatch
2492 * directly. The device may be in a situation where it can't
2493 * handle FS request, and always returns BLK_STS_RESOURCE for
2494 * them, which gets them added to hctx->dispatch.
2496 * If a passthrough request is required to unblock the queues,
2497 * and it is added to the scheduler queue, there is no chance to
2498 * dispatch it given we prioritize requests in hctx->dispatch.
2500 blk_mq_request_bypass_insert(rq
, flags
);
2501 } else if (req_op(rq
) == REQ_OP_FLUSH
) {
2503 * Firstly normal IO request is inserted to scheduler queue or
2504 * sw queue, meantime we add flush request to dispatch queue(
2505 * hctx->dispatch) directly and there is at most one in-flight
2506 * flush request for each hw queue, so it doesn't matter to add
2507 * flush request to tail or front of the dispatch queue.
2509 * Secondly in case of NCQ, flush request belongs to non-NCQ
2510 * command, and queueing it will fail when there is any
2511 * in-flight normal IO request(NCQ command). When adding flush
2512 * rq to the front of hctx->dispatch, it is easier to introduce
2513 * extra time to flush rq's latency because of S_SCHED_RESTART
2514 * compared with adding to the tail of dispatch queue, then
2515 * chance of flush merge is increased, and less flush requests
2516 * will be issued to controller. It is observed that ~10% time
2517 * is saved in blktests block/004 on disk attached to AHCI/NCQ
2518 * drive when adding flush rq to the front of hctx->dispatch.
2520 * Simply queue flush rq to the front of hctx->dispatch so that
2521 * intensive flush workloads can benefit in case of NCQ HW.
2523 blk_mq_request_bypass_insert(rq
, BLK_MQ_INSERT_AT_HEAD
);
2524 } else if (q
->elevator
) {
2527 WARN_ON_ONCE(rq
->tag
!= BLK_MQ_NO_TAG
);
2529 list_add(&rq
->queuelist
, &list
);
2530 q
->elevator
->type
->ops
.insert_requests(hctx
, &list
, flags
);
2532 trace_block_rq_insert(rq
);
2534 spin_lock(&ctx
->lock
);
2535 if (flags
& BLK_MQ_INSERT_AT_HEAD
)
2536 list_add(&rq
->queuelist
, &ctx
->rq_lists
[hctx
->type
]);
2538 list_add_tail(&rq
->queuelist
,
2539 &ctx
->rq_lists
[hctx
->type
]);
2540 blk_mq_hctx_mark_pending(hctx
, ctx
);
2541 spin_unlock(&ctx
->lock
);
2545 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
,
2546 unsigned int nr_segs
)
2550 if (bio
->bi_opf
& REQ_RAHEAD
)
2551 rq
->cmd_flags
|= REQ_FAILFAST_MASK
;
2553 rq
->__sector
= bio
->bi_iter
.bi_sector
;
2554 blk_rq_bio_prep(rq
, bio
, nr_segs
);
2556 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2557 err
= blk_crypto_rq_bio_prep(rq
, bio
, GFP_NOIO
);
2560 blk_account_io_start(rq
);
2563 static blk_status_t
__blk_mq_issue_directly(struct blk_mq_hw_ctx
*hctx
,
2564 struct request
*rq
, bool last
)
2566 struct request_queue
*q
= rq
->q
;
2567 struct blk_mq_queue_data bd
= {
2574 * For OK queue, we are done. For error, caller may kill it.
2575 * Any other error (busy), just add it to our list as we
2576 * previously would have done.
2578 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
2581 blk_mq_update_dispatch_busy(hctx
, false);
2583 case BLK_STS_RESOURCE
:
2584 case BLK_STS_DEV_RESOURCE
:
2585 blk_mq_update_dispatch_busy(hctx
, true);
2586 __blk_mq_requeue_request(rq
);
2589 blk_mq_update_dispatch_busy(hctx
, false);
2596 static bool blk_mq_get_budget_and_tag(struct request
*rq
)
2600 budget_token
= blk_mq_get_dispatch_budget(rq
->q
);
2601 if (budget_token
< 0)
2603 blk_mq_set_rq_budget_token(rq
, budget_token
);
2604 if (!blk_mq_get_driver_tag(rq
)) {
2605 blk_mq_put_dispatch_budget(rq
->q
, budget_token
);
2612 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2613 * @hctx: Pointer of the associated hardware queue.
2614 * @rq: Pointer to request to be sent.
2616 * If the device has enough resources to accept a new request now, send the
2617 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2618 * we can try send it another time in the future. Requests inserted at this
2619 * queue have higher priority.
2621 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
2626 if (blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(rq
->q
)) {
2627 blk_mq_insert_request(rq
, 0);
2631 if ((rq
->rq_flags
& RQF_USE_SCHED
) || !blk_mq_get_budget_and_tag(rq
)) {
2632 blk_mq_insert_request(rq
, 0);
2633 blk_mq_run_hw_queue(hctx
, rq
->cmd_flags
& REQ_NOWAIT
);
2637 ret
= __blk_mq_issue_directly(hctx
, rq
, true);
2641 case BLK_STS_RESOURCE
:
2642 case BLK_STS_DEV_RESOURCE
:
2643 blk_mq_request_bypass_insert(rq
, 0);
2644 blk_mq_run_hw_queue(hctx
, false);
2647 blk_mq_end_request(rq
, ret
);
2652 static blk_status_t
blk_mq_request_issue_directly(struct request
*rq
, bool last
)
2654 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
2656 if (blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(rq
->q
)) {
2657 blk_mq_insert_request(rq
, 0);
2661 if (!blk_mq_get_budget_and_tag(rq
))
2662 return BLK_STS_RESOURCE
;
2663 return __blk_mq_issue_directly(hctx
, rq
, last
);
2666 static void blk_mq_plug_issue_direct(struct blk_plug
*plug
)
2668 struct blk_mq_hw_ctx
*hctx
= NULL
;
2671 blk_status_t ret
= BLK_STS_OK
;
2673 while ((rq
= rq_list_pop(&plug
->mq_list
))) {
2674 bool last
= rq_list_empty(plug
->mq_list
);
2676 if (hctx
!= rq
->mq_hctx
) {
2678 blk_mq_commit_rqs(hctx
, queued
, false);
2684 ret
= blk_mq_request_issue_directly(rq
, last
);
2689 case BLK_STS_RESOURCE
:
2690 case BLK_STS_DEV_RESOURCE
:
2691 blk_mq_request_bypass_insert(rq
, 0);
2692 blk_mq_run_hw_queue(hctx
, false);
2695 blk_mq_end_request(rq
, ret
);
2701 if (ret
!= BLK_STS_OK
)
2702 blk_mq_commit_rqs(hctx
, queued
, false);
2705 static void __blk_mq_flush_plug_list(struct request_queue
*q
,
2706 struct blk_plug
*plug
)
2708 if (blk_queue_quiesced(q
))
2710 q
->mq_ops
->queue_rqs(&plug
->mq_list
);
2713 static void blk_mq_dispatch_plug_list(struct blk_plug
*plug
, bool from_sched
)
2715 struct blk_mq_hw_ctx
*this_hctx
= NULL
;
2716 struct blk_mq_ctx
*this_ctx
= NULL
;
2717 struct request
*requeue_list
= NULL
;
2718 struct request
**requeue_lastp
= &requeue_list
;
2719 unsigned int depth
= 0;
2720 bool is_passthrough
= false;
2724 struct request
*rq
= rq_list_pop(&plug
->mq_list
);
2727 this_hctx
= rq
->mq_hctx
;
2728 this_ctx
= rq
->mq_ctx
;
2729 is_passthrough
= blk_rq_is_passthrough(rq
);
2730 } else if (this_hctx
!= rq
->mq_hctx
|| this_ctx
!= rq
->mq_ctx
||
2731 is_passthrough
!= blk_rq_is_passthrough(rq
)) {
2732 rq_list_add_tail(&requeue_lastp
, rq
);
2735 list_add(&rq
->queuelist
, &list
);
2737 } while (!rq_list_empty(plug
->mq_list
));
2739 plug
->mq_list
= requeue_list
;
2740 trace_block_unplug(this_hctx
->queue
, depth
, !from_sched
);
2742 percpu_ref_get(&this_hctx
->queue
->q_usage_counter
);
2743 /* passthrough requests should never be issued to the I/O scheduler */
2744 if (is_passthrough
) {
2745 spin_lock(&this_hctx
->lock
);
2746 list_splice_tail_init(&list
, &this_hctx
->dispatch
);
2747 spin_unlock(&this_hctx
->lock
);
2748 blk_mq_run_hw_queue(this_hctx
, from_sched
);
2749 } else if (this_hctx
->queue
->elevator
) {
2750 this_hctx
->queue
->elevator
->type
->ops
.insert_requests(this_hctx
,
2752 blk_mq_run_hw_queue(this_hctx
, from_sched
);
2754 blk_mq_insert_requests(this_hctx
, this_ctx
, &list
, from_sched
);
2756 percpu_ref_put(&this_hctx
->queue
->q_usage_counter
);
2759 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
2764 * We may have been called recursively midway through handling
2765 * plug->mq_list via a schedule() in the driver's queue_rq() callback.
2766 * To avoid mq_list changing under our feet, clear rq_count early and
2767 * bail out specifically if rq_count is 0 rather than checking
2768 * whether the mq_list is empty.
2770 if (plug
->rq_count
== 0)
2774 if (!plug
->multiple_queues
&& !plug
->has_elevator
&& !from_schedule
) {
2775 struct request_queue
*q
;
2777 rq
= rq_list_peek(&plug
->mq_list
);
2781 * Peek first request and see if we have a ->queue_rqs() hook.
2782 * If we do, we can dispatch the whole plug list in one go. We
2783 * already know at this point that all requests belong to the
2784 * same queue, caller must ensure that's the case.
2786 if (q
->mq_ops
->queue_rqs
) {
2787 blk_mq_run_dispatch_ops(q
,
2788 __blk_mq_flush_plug_list(q
, plug
));
2789 if (rq_list_empty(plug
->mq_list
))
2793 blk_mq_run_dispatch_ops(q
,
2794 blk_mq_plug_issue_direct(plug
));
2795 if (rq_list_empty(plug
->mq_list
))
2800 blk_mq_dispatch_plug_list(plug
, from_schedule
);
2801 } while (!rq_list_empty(plug
->mq_list
));
2804 static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx
*hctx
,
2805 struct list_head
*list
)
2808 blk_status_t ret
= BLK_STS_OK
;
2810 while (!list_empty(list
)) {
2811 struct request
*rq
= list_first_entry(list
, struct request
,
2814 list_del_init(&rq
->queuelist
);
2815 ret
= blk_mq_request_issue_directly(rq
, list_empty(list
));
2820 case BLK_STS_RESOURCE
:
2821 case BLK_STS_DEV_RESOURCE
:
2822 blk_mq_request_bypass_insert(rq
, 0);
2823 if (list_empty(list
))
2824 blk_mq_run_hw_queue(hctx
, false);
2827 blk_mq_end_request(rq
, ret
);
2833 if (ret
!= BLK_STS_OK
)
2834 blk_mq_commit_rqs(hctx
, queued
, false);
2837 static bool blk_mq_attempt_bio_merge(struct request_queue
*q
,
2838 struct bio
*bio
, unsigned int nr_segs
)
2840 if (!blk_queue_nomerges(q
) && bio_mergeable(bio
)) {
2841 if (blk_attempt_plug_merge(q
, bio
, nr_segs
))
2843 if (blk_mq_sched_bio_merge(q
, bio
, nr_segs
))
2849 static struct request
*blk_mq_get_new_requests(struct request_queue
*q
,
2850 struct blk_plug
*plug
,
2854 struct blk_mq_alloc_data data
= {
2857 .cmd_flags
= bio
->bi_opf
,
2861 if (unlikely(bio_queue_enter(bio
)))
2864 if (blk_mq_attempt_bio_merge(q
, bio
, nsegs
))
2867 rq_qos_throttle(q
, bio
);
2870 data
.nr_tags
= plug
->nr_ios
;
2872 data
.cached_rq
= &plug
->cached_rq
;
2875 rq
= __blk_mq_alloc_requests(&data
);
2878 rq_qos_cleanup(q
, bio
);
2879 if (bio
->bi_opf
& REQ_NOWAIT
)
2880 bio_wouldblock_error(bio
);
2886 static inline struct request
*blk_mq_get_cached_request(struct request_queue
*q
,
2887 struct blk_plug
*plug
, struct bio
**bio
, unsigned int nsegs
)
2890 enum hctx_type type
, hctx_type
;
2894 rq
= rq_list_peek(&plug
->cached_rq
);
2895 if (!rq
|| rq
->q
!= q
)
2898 if (blk_mq_attempt_bio_merge(q
, *bio
, nsegs
)) {
2903 type
= blk_mq_get_hctx_type((*bio
)->bi_opf
);
2904 hctx_type
= rq
->mq_hctx
->type
;
2905 if (type
!= hctx_type
&&
2906 !(type
== HCTX_TYPE_READ
&& hctx_type
== HCTX_TYPE_DEFAULT
))
2908 if (op_is_flush(rq
->cmd_flags
) != op_is_flush((*bio
)->bi_opf
))
2912 * If any qos ->throttle() end up blocking, we will have flushed the
2913 * plug and hence killed the cached_rq list as well. Pop this entry
2914 * before we throttle.
2916 plug
->cached_rq
= rq_list_next(rq
);
2917 rq_qos_throttle(q
, *bio
);
2919 blk_mq_rq_time_init(rq
, 0);
2920 rq
->cmd_flags
= (*bio
)->bi_opf
;
2921 INIT_LIST_HEAD(&rq
->queuelist
);
2925 static void bio_set_ioprio(struct bio
*bio
)
2927 /* Nobody set ioprio so far? Initialize it based on task's nice value */
2928 if (IOPRIO_PRIO_CLASS(bio
->bi_ioprio
) == IOPRIO_CLASS_NONE
)
2929 bio
->bi_ioprio
= get_current_ioprio();
2930 blkcg_set_ioprio(bio
);
2934 * blk_mq_submit_bio - Create and send a request to block device.
2935 * @bio: Bio pointer.
2937 * Builds up a request structure from @q and @bio and send to the device. The
2938 * request may not be queued directly to hardware if:
2939 * * This request can be merged with another one
2940 * * We want to place request at plug queue for possible future merging
2941 * * There is an IO scheduler active at this queue
2943 * It will not queue the request if there is an error with the bio, or at the
2946 void blk_mq_submit_bio(struct bio
*bio
)
2948 struct request_queue
*q
= bdev_get_queue(bio
->bi_bdev
);
2949 struct blk_plug
*plug
= blk_mq_plug(bio
);
2950 const int is_sync
= op_is_sync(bio
->bi_opf
);
2951 struct blk_mq_hw_ctx
*hctx
;
2953 unsigned int nr_segs
= 1;
2956 bio
= blk_queue_bounce(bio
, q
);
2957 if (bio_may_exceed_limits(bio
, &q
->limits
)) {
2958 bio
= __bio_split_to_limits(bio
, &q
->limits
, &nr_segs
);
2963 if (!bio_integrity_prep(bio
))
2966 bio_set_ioprio(bio
);
2968 rq
= blk_mq_get_cached_request(q
, plug
, &bio
, nr_segs
);
2972 rq
= blk_mq_get_new_requests(q
, plug
, bio
, nr_segs
);
2977 trace_block_getrq(bio
);
2979 rq_qos_track(q
, rq
, bio
);
2981 blk_mq_bio_to_request(rq
, bio
, nr_segs
);
2983 ret
= blk_crypto_rq_get_keyslot(rq
);
2984 if (ret
!= BLK_STS_OK
) {
2985 bio
->bi_status
= ret
;
2987 blk_mq_free_request(rq
);
2991 if (op_is_flush(bio
->bi_opf
) && blk_insert_flush(rq
))
2995 blk_add_rq_to_plug(plug
, rq
);
3000 if ((rq
->rq_flags
& RQF_USE_SCHED
) ||
3001 (hctx
->dispatch_busy
&& (q
->nr_hw_queues
== 1 || !is_sync
))) {
3002 blk_mq_insert_request(rq
, 0);
3003 blk_mq_run_hw_queue(hctx
, true);
3005 blk_mq_run_dispatch_ops(q
, blk_mq_try_issue_directly(hctx
, rq
));
3009 #ifdef CONFIG_BLK_MQ_STACKING
3011 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
3012 * @rq: the request being queued
3014 blk_status_t
blk_insert_cloned_request(struct request
*rq
)
3016 struct request_queue
*q
= rq
->q
;
3017 unsigned int max_sectors
= blk_queue_get_max_sectors(q
, req_op(rq
));
3018 unsigned int max_segments
= blk_rq_get_max_segments(rq
);
3021 if (blk_rq_sectors(rq
) > max_sectors
) {
3023 * SCSI device does not have a good way to return if
3024 * Write Same/Zero is actually supported. If a device rejects
3025 * a non-read/write command (discard, write same,etc.) the
3026 * low-level device driver will set the relevant queue limit to
3027 * 0 to prevent blk-lib from issuing more of the offending
3028 * operations. Commands queued prior to the queue limit being
3029 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
3030 * errors being propagated to upper layers.
3032 if (max_sectors
== 0)
3033 return BLK_STS_NOTSUPP
;
3035 printk(KERN_ERR
"%s: over max size limit. (%u > %u)\n",
3036 __func__
, blk_rq_sectors(rq
), max_sectors
);
3037 return BLK_STS_IOERR
;
3041 * The queue settings related to segment counting may differ from the
3044 rq
->nr_phys_segments
= blk_recalc_rq_segments(rq
);
3045 if (rq
->nr_phys_segments
> max_segments
) {
3046 printk(KERN_ERR
"%s: over max segments limit. (%u > %u)\n",
3047 __func__
, rq
->nr_phys_segments
, max_segments
);
3048 return BLK_STS_IOERR
;
3051 if (q
->disk
&& should_fail_request(q
->disk
->part0
, blk_rq_bytes(rq
)))
3052 return BLK_STS_IOERR
;
3054 ret
= blk_crypto_rq_get_keyslot(rq
);
3055 if (ret
!= BLK_STS_OK
)
3058 blk_account_io_start(rq
);
3061 * Since we have a scheduler attached on the top device,
3062 * bypass a potential scheduler on the bottom device for
3065 blk_mq_run_dispatch_ops(q
,
3066 ret
= blk_mq_request_issue_directly(rq
, true));
3068 blk_account_io_done(rq
, ktime_get_ns());
3071 EXPORT_SYMBOL_GPL(blk_insert_cloned_request
);
3074 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3075 * @rq: the clone request to be cleaned up
3078 * Free all bios in @rq for a cloned request.
3080 void blk_rq_unprep_clone(struct request
*rq
)
3084 while ((bio
= rq
->bio
) != NULL
) {
3085 rq
->bio
= bio
->bi_next
;
3090 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone
);
3093 * blk_rq_prep_clone - Helper function to setup clone request
3094 * @rq: the request to be setup
3095 * @rq_src: original request to be cloned
3096 * @bs: bio_set that bios for clone are allocated from
3097 * @gfp_mask: memory allocation mask for bio
3098 * @bio_ctr: setup function to be called for each clone bio.
3099 * Returns %0 for success, non %0 for failure.
3100 * @data: private data to be passed to @bio_ctr
3103 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3104 * Also, pages which the original bios are pointing to are not copied
3105 * and the cloned bios just point same pages.
3106 * So cloned bios must be completed before original bios, which means
3107 * the caller must complete @rq before @rq_src.
3109 int blk_rq_prep_clone(struct request
*rq
, struct request
*rq_src
,
3110 struct bio_set
*bs
, gfp_t gfp_mask
,
3111 int (*bio_ctr
)(struct bio
*, struct bio
*, void *),
3114 struct bio
*bio
, *bio_src
;
3119 __rq_for_each_bio(bio_src
, rq_src
) {
3120 bio
= bio_alloc_clone(rq
->q
->disk
->part0
, bio_src
, gfp_mask
,
3125 if (bio_ctr
&& bio_ctr(bio
, bio_src
, data
))
3129 rq
->biotail
->bi_next
= bio
;
3132 rq
->bio
= rq
->biotail
= bio
;
3137 /* Copy attributes of the original request to the clone request. */
3138 rq
->__sector
= blk_rq_pos(rq_src
);
3139 rq
->__data_len
= blk_rq_bytes(rq_src
);
3140 if (rq_src
->rq_flags
& RQF_SPECIAL_PAYLOAD
) {
3141 rq
->rq_flags
|= RQF_SPECIAL_PAYLOAD
;
3142 rq
->special_vec
= rq_src
->special_vec
;
3144 rq
->nr_phys_segments
= rq_src
->nr_phys_segments
;
3145 rq
->ioprio
= rq_src
->ioprio
;
3147 if (rq
->bio
&& blk_crypto_rq_bio_prep(rq
, rq
->bio
, gfp_mask
) < 0)
3155 blk_rq_unprep_clone(rq
);
3159 EXPORT_SYMBOL_GPL(blk_rq_prep_clone
);
3160 #endif /* CONFIG_BLK_MQ_STACKING */
3163 * Steal bios from a request and add them to a bio list.
3164 * The request must not have been partially completed before.
3166 void blk_steal_bios(struct bio_list
*list
, struct request
*rq
)
3170 list
->tail
->bi_next
= rq
->bio
;
3172 list
->head
= rq
->bio
;
3173 list
->tail
= rq
->biotail
;
3181 EXPORT_SYMBOL_GPL(blk_steal_bios
);
3183 static size_t order_to_size(unsigned int order
)
3185 return (size_t)PAGE_SIZE
<< order
;
3188 /* called before freeing request pool in @tags */
3189 static void blk_mq_clear_rq_mapping(struct blk_mq_tags
*drv_tags
,
3190 struct blk_mq_tags
*tags
)
3193 unsigned long flags
;
3196 * There is no need to clear mapping if driver tags is not initialized
3197 * or the mapping belongs to the driver tags.
3199 if (!drv_tags
|| drv_tags
== tags
)
3202 list_for_each_entry(page
, &tags
->page_list
, lru
) {
3203 unsigned long start
= (unsigned long)page_address(page
);
3204 unsigned long end
= start
+ order_to_size(page
->private);
3207 for (i
= 0; i
< drv_tags
->nr_tags
; i
++) {
3208 struct request
*rq
= drv_tags
->rqs
[i
];
3209 unsigned long rq_addr
= (unsigned long)rq
;
3211 if (rq_addr
>= start
&& rq_addr
< end
) {
3212 WARN_ON_ONCE(req_ref_read(rq
) != 0);
3213 cmpxchg(&drv_tags
->rqs
[i
], rq
, NULL
);
3219 * Wait until all pending iteration is done.
3221 * Request reference is cleared and it is guaranteed to be observed
3222 * after the ->lock is released.
3224 spin_lock_irqsave(&drv_tags
->lock
, flags
);
3225 spin_unlock_irqrestore(&drv_tags
->lock
, flags
);
3228 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
3229 unsigned int hctx_idx
)
3231 struct blk_mq_tags
*drv_tags
;
3234 if (list_empty(&tags
->page_list
))
3237 if (blk_mq_is_shared_tags(set
->flags
))
3238 drv_tags
= set
->shared_tags
;
3240 drv_tags
= set
->tags
[hctx_idx
];
3242 if (tags
->static_rqs
&& set
->ops
->exit_request
) {
3245 for (i
= 0; i
< tags
->nr_tags
; i
++) {
3246 struct request
*rq
= tags
->static_rqs
[i
];
3250 set
->ops
->exit_request(set
, rq
, hctx_idx
);
3251 tags
->static_rqs
[i
] = NULL
;
3255 blk_mq_clear_rq_mapping(drv_tags
, tags
);
3257 while (!list_empty(&tags
->page_list
)) {
3258 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
3259 list_del_init(&page
->lru
);
3261 * Remove kmemleak object previously allocated in
3262 * blk_mq_alloc_rqs().
3264 kmemleak_free(page_address(page
));
3265 __free_pages(page
, page
->private);
3269 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
3273 kfree(tags
->static_rqs
);
3274 tags
->static_rqs
= NULL
;
3276 blk_mq_free_tags(tags
);
3279 static enum hctx_type
hctx_idx_to_type(struct blk_mq_tag_set
*set
,
3280 unsigned int hctx_idx
)
3284 for (i
= 0; i
< set
->nr_maps
; i
++) {
3285 unsigned int start
= set
->map
[i
].queue_offset
;
3286 unsigned int end
= start
+ set
->map
[i
].nr_queues
;
3288 if (hctx_idx
>= start
&& hctx_idx
< end
)
3292 if (i
>= set
->nr_maps
)
3293 i
= HCTX_TYPE_DEFAULT
;
3298 static int blk_mq_get_hctx_node(struct blk_mq_tag_set
*set
,
3299 unsigned int hctx_idx
)
3301 enum hctx_type type
= hctx_idx_to_type(set
, hctx_idx
);
3303 return blk_mq_hw_queue_to_node(&set
->map
[type
], hctx_idx
);
3306 static struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
3307 unsigned int hctx_idx
,
3308 unsigned int nr_tags
,
3309 unsigned int reserved_tags
)
3311 int node
= blk_mq_get_hctx_node(set
, hctx_idx
);
3312 struct blk_mq_tags
*tags
;
3314 if (node
== NUMA_NO_NODE
)
3315 node
= set
->numa_node
;
3317 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
,
3318 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
3322 tags
->rqs
= kcalloc_node(nr_tags
, sizeof(struct request
*),
3323 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
3328 tags
->static_rqs
= kcalloc_node(nr_tags
, sizeof(struct request
*),
3329 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
3331 if (!tags
->static_rqs
)
3339 blk_mq_free_tags(tags
);
3343 static int blk_mq_init_request(struct blk_mq_tag_set
*set
, struct request
*rq
,
3344 unsigned int hctx_idx
, int node
)
3348 if (set
->ops
->init_request
) {
3349 ret
= set
->ops
->init_request(set
, rq
, hctx_idx
, node
);
3354 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
3358 static int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
,
3359 struct blk_mq_tags
*tags
,
3360 unsigned int hctx_idx
, unsigned int depth
)
3362 unsigned int i
, j
, entries_per_page
, max_order
= 4;
3363 int node
= blk_mq_get_hctx_node(set
, hctx_idx
);
3364 size_t rq_size
, left
;
3366 if (node
== NUMA_NO_NODE
)
3367 node
= set
->numa_node
;
3369 INIT_LIST_HEAD(&tags
->page_list
);
3372 * rq_size is the size of the request plus driver payload, rounded
3373 * to the cacheline size
3375 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
3377 left
= rq_size
* depth
;
3379 for (i
= 0; i
< depth
; ) {
3380 int this_order
= max_order
;
3385 while (this_order
&& left
< order_to_size(this_order
- 1))
3389 page
= alloc_pages_node(node
,
3390 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
3396 if (order_to_size(this_order
) < rq_size
)
3403 page
->private = this_order
;
3404 list_add_tail(&page
->lru
, &tags
->page_list
);
3406 p
= page_address(page
);
3408 * Allow kmemleak to scan these pages as they contain pointers
3409 * to additional allocations like via ops->init_request().
3411 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
3412 entries_per_page
= order_to_size(this_order
) / rq_size
;
3413 to_do
= min(entries_per_page
, depth
- i
);
3414 left
-= to_do
* rq_size
;
3415 for (j
= 0; j
< to_do
; j
++) {
3416 struct request
*rq
= p
;
3418 tags
->static_rqs
[i
] = rq
;
3419 if (blk_mq_init_request(set
, rq
, hctx_idx
, node
)) {
3420 tags
->static_rqs
[i
] = NULL
;
3431 blk_mq_free_rqs(set
, tags
, hctx_idx
);
3435 struct rq_iter_data
{
3436 struct blk_mq_hw_ctx
*hctx
;
3440 static bool blk_mq_has_request(struct request
*rq
, void *data
)
3442 struct rq_iter_data
*iter_data
= data
;
3444 if (rq
->mq_hctx
!= iter_data
->hctx
)
3446 iter_data
->has_rq
= true;
3450 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx
*hctx
)
3452 struct blk_mq_tags
*tags
= hctx
->sched_tags
?
3453 hctx
->sched_tags
: hctx
->tags
;
3454 struct rq_iter_data data
= {
3458 blk_mq_all_tag_iter(tags
, blk_mq_has_request
, &data
);
3462 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu
,
3463 struct blk_mq_hw_ctx
*hctx
)
3465 if (cpumask_first_and(hctx
->cpumask
, cpu_online_mask
) != cpu
)
3467 if (cpumask_next_and(cpu
, hctx
->cpumask
, cpu_online_mask
) < nr_cpu_ids
)
3472 static int blk_mq_hctx_notify_offline(unsigned int cpu
, struct hlist_node
*node
)
3474 struct blk_mq_hw_ctx
*hctx
= hlist_entry_safe(node
,
3475 struct blk_mq_hw_ctx
, cpuhp_online
);
3477 if (!cpumask_test_cpu(cpu
, hctx
->cpumask
) ||
3478 !blk_mq_last_cpu_in_hctx(cpu
, hctx
))
3482 * Prevent new request from being allocated on the current hctx.
3484 * The smp_mb__after_atomic() Pairs with the implied barrier in
3485 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3486 * seen once we return from the tag allocator.
3488 set_bit(BLK_MQ_S_INACTIVE
, &hctx
->state
);
3489 smp_mb__after_atomic();
3492 * Try to grab a reference to the queue and wait for any outstanding
3493 * requests. If we could not grab a reference the queue has been
3494 * frozen and there are no requests.
3496 if (percpu_ref_tryget(&hctx
->queue
->q_usage_counter
)) {
3497 while (blk_mq_hctx_has_requests(hctx
))
3499 percpu_ref_put(&hctx
->queue
->q_usage_counter
);
3505 static int blk_mq_hctx_notify_online(unsigned int cpu
, struct hlist_node
*node
)
3507 struct blk_mq_hw_ctx
*hctx
= hlist_entry_safe(node
,
3508 struct blk_mq_hw_ctx
, cpuhp_online
);
3510 if (cpumask_test_cpu(cpu
, hctx
->cpumask
))
3511 clear_bit(BLK_MQ_S_INACTIVE
, &hctx
->state
);
3516 * 'cpu' is going away. splice any existing rq_list entries from this
3517 * software queue to the hw queue dispatch list, and ensure that it
3520 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
3522 struct blk_mq_hw_ctx
*hctx
;
3523 struct blk_mq_ctx
*ctx
;
3525 enum hctx_type type
;
3527 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
3528 if (!cpumask_test_cpu(cpu
, hctx
->cpumask
))
3531 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
3534 spin_lock(&ctx
->lock
);
3535 if (!list_empty(&ctx
->rq_lists
[type
])) {
3536 list_splice_init(&ctx
->rq_lists
[type
], &tmp
);
3537 blk_mq_hctx_clear_pending(hctx
, ctx
);
3539 spin_unlock(&ctx
->lock
);
3541 if (list_empty(&tmp
))
3544 spin_lock(&hctx
->lock
);
3545 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
3546 spin_unlock(&hctx
->lock
);
3548 blk_mq_run_hw_queue(hctx
, true);
3552 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
3554 if (!(hctx
->flags
& BLK_MQ_F_STACKING
))
3555 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE
,
3556 &hctx
->cpuhp_online
);
3557 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
3562 * Before freeing hw queue, clearing the flush request reference in
3563 * tags->rqs[] for avoiding potential UAF.
3565 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags
*tags
,
3566 unsigned int queue_depth
, struct request
*flush_rq
)
3569 unsigned long flags
;
3571 /* The hw queue may not be mapped yet */
3575 WARN_ON_ONCE(req_ref_read(flush_rq
) != 0);
3577 for (i
= 0; i
< queue_depth
; i
++)
3578 cmpxchg(&tags
->rqs
[i
], flush_rq
, NULL
);
3581 * Wait until all pending iteration is done.
3583 * Request reference is cleared and it is guaranteed to be observed
3584 * after the ->lock is released.
3586 spin_lock_irqsave(&tags
->lock
, flags
);
3587 spin_unlock_irqrestore(&tags
->lock
, flags
);
3590 /* hctx->ctxs will be freed in queue's release handler */
3591 static void blk_mq_exit_hctx(struct request_queue
*q
,
3592 struct blk_mq_tag_set
*set
,
3593 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
3595 struct request
*flush_rq
= hctx
->fq
->flush_rq
;
3597 if (blk_mq_hw_queue_mapped(hctx
))
3598 blk_mq_tag_idle(hctx
);
3600 if (blk_queue_init_done(q
))
3601 blk_mq_clear_flush_rq_mapping(set
->tags
[hctx_idx
],
3602 set
->queue_depth
, flush_rq
);
3603 if (set
->ops
->exit_request
)
3604 set
->ops
->exit_request(set
, flush_rq
, hctx_idx
);
3606 if (set
->ops
->exit_hctx
)
3607 set
->ops
->exit_hctx(hctx
, hctx_idx
);
3609 blk_mq_remove_cpuhp(hctx
);
3611 xa_erase(&q
->hctx_table
, hctx_idx
);
3613 spin_lock(&q
->unused_hctx_lock
);
3614 list_add(&hctx
->hctx_list
, &q
->unused_hctx_list
);
3615 spin_unlock(&q
->unused_hctx_lock
);
3618 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
3619 struct blk_mq_tag_set
*set
, int nr_queue
)
3621 struct blk_mq_hw_ctx
*hctx
;
3624 queue_for_each_hw_ctx(q
, hctx
, i
) {
3627 blk_mq_exit_hctx(q
, set
, hctx
, i
);
3631 static int blk_mq_init_hctx(struct request_queue
*q
,
3632 struct blk_mq_tag_set
*set
,
3633 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
3635 hctx
->queue_num
= hctx_idx
;
3637 if (!(hctx
->flags
& BLK_MQ_F_STACKING
))
3638 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE
,
3639 &hctx
->cpuhp_online
);
3640 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
3642 hctx
->tags
= set
->tags
[hctx_idx
];
3644 if (set
->ops
->init_hctx
&&
3645 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
3646 goto unregister_cpu_notifier
;
3648 if (blk_mq_init_request(set
, hctx
->fq
->flush_rq
, hctx_idx
,
3652 if (xa_insert(&q
->hctx_table
, hctx_idx
, hctx
, GFP_KERNEL
))
3658 if (set
->ops
->exit_request
)
3659 set
->ops
->exit_request(set
, hctx
->fq
->flush_rq
, hctx_idx
);
3661 if (set
->ops
->exit_hctx
)
3662 set
->ops
->exit_hctx(hctx
, hctx_idx
);
3663 unregister_cpu_notifier
:
3664 blk_mq_remove_cpuhp(hctx
);
3668 static struct blk_mq_hw_ctx
*
3669 blk_mq_alloc_hctx(struct request_queue
*q
, struct blk_mq_tag_set
*set
,
3672 struct blk_mq_hw_ctx
*hctx
;
3673 gfp_t gfp
= GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
;
3675 hctx
= kzalloc_node(sizeof(struct blk_mq_hw_ctx
), gfp
, node
);
3677 goto fail_alloc_hctx
;
3679 if (!zalloc_cpumask_var_node(&hctx
->cpumask
, gfp
, node
))
3682 atomic_set(&hctx
->nr_active
, 0);
3683 if (node
== NUMA_NO_NODE
)
3684 node
= set
->numa_node
;
3685 hctx
->numa_node
= node
;
3687 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
3688 spin_lock_init(&hctx
->lock
);
3689 INIT_LIST_HEAD(&hctx
->dispatch
);
3691 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_QUEUE_SHARED
;
3693 INIT_LIST_HEAD(&hctx
->hctx_list
);
3696 * Allocate space for all possible cpus to avoid allocation at
3699 hctx
->ctxs
= kmalloc_array_node(nr_cpu_ids
, sizeof(void *),
3704 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8),
3705 gfp
, node
, false, false))
3709 spin_lock_init(&hctx
->dispatch_wait_lock
);
3710 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
3711 INIT_LIST_HEAD(&hctx
->dispatch_wait
.entry
);
3713 hctx
->fq
= blk_alloc_flush_queue(hctx
->numa_node
, set
->cmd_size
, gfp
);
3717 blk_mq_hctx_kobj_init(hctx
);
3722 sbitmap_free(&hctx
->ctx_map
);
3726 free_cpumask_var(hctx
->cpumask
);
3733 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
3734 unsigned int nr_hw_queues
)
3736 struct blk_mq_tag_set
*set
= q
->tag_set
;
3739 for_each_possible_cpu(i
) {
3740 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
3741 struct blk_mq_hw_ctx
*hctx
;
3745 spin_lock_init(&__ctx
->lock
);
3746 for (k
= HCTX_TYPE_DEFAULT
; k
< HCTX_MAX_TYPES
; k
++)
3747 INIT_LIST_HEAD(&__ctx
->rq_lists
[k
]);
3752 * Set local node, IFF we have more than one hw queue. If
3753 * not, we remain on the home node of the device
3755 for (j
= 0; j
< set
->nr_maps
; j
++) {
3756 hctx
= blk_mq_map_queue_type(q
, j
, i
);
3757 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
3758 hctx
->numa_node
= cpu_to_node(i
);
3763 struct blk_mq_tags
*blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set
*set
,
3764 unsigned int hctx_idx
,
3767 struct blk_mq_tags
*tags
;
3770 tags
= blk_mq_alloc_rq_map(set
, hctx_idx
, depth
, set
->reserved_tags
);
3774 ret
= blk_mq_alloc_rqs(set
, tags
, hctx_idx
, depth
);
3776 blk_mq_free_rq_map(tags
);
3783 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set
*set
,
3786 if (blk_mq_is_shared_tags(set
->flags
)) {
3787 set
->tags
[hctx_idx
] = set
->shared_tags
;
3792 set
->tags
[hctx_idx
] = blk_mq_alloc_map_and_rqs(set
, hctx_idx
,
3795 return set
->tags
[hctx_idx
];
3798 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set
*set
,
3799 struct blk_mq_tags
*tags
,
3800 unsigned int hctx_idx
)
3803 blk_mq_free_rqs(set
, tags
, hctx_idx
);
3804 blk_mq_free_rq_map(tags
);
3808 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set
*set
,
3809 unsigned int hctx_idx
)
3811 if (!blk_mq_is_shared_tags(set
->flags
))
3812 blk_mq_free_map_and_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
3814 set
->tags
[hctx_idx
] = NULL
;
3817 static void blk_mq_map_swqueue(struct request_queue
*q
)
3819 unsigned int j
, hctx_idx
;
3821 struct blk_mq_hw_ctx
*hctx
;
3822 struct blk_mq_ctx
*ctx
;
3823 struct blk_mq_tag_set
*set
= q
->tag_set
;
3825 queue_for_each_hw_ctx(q
, hctx
, i
) {
3826 cpumask_clear(hctx
->cpumask
);
3828 hctx
->dispatch_from
= NULL
;
3832 * Map software to hardware queues.
3834 * If the cpu isn't present, the cpu is mapped to first hctx.
3836 for_each_possible_cpu(i
) {
3838 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
3839 for (j
= 0; j
< set
->nr_maps
; j
++) {
3840 if (!set
->map
[j
].nr_queues
) {
3841 ctx
->hctxs
[j
] = blk_mq_map_queue_type(q
,
3842 HCTX_TYPE_DEFAULT
, i
);
3845 hctx_idx
= set
->map
[j
].mq_map
[i
];
3846 /* unmapped hw queue can be remapped after CPU topo changed */
3847 if (!set
->tags
[hctx_idx
] &&
3848 !__blk_mq_alloc_map_and_rqs(set
, hctx_idx
)) {
3850 * If tags initialization fail for some hctx,
3851 * that hctx won't be brought online. In this
3852 * case, remap the current ctx to hctx[0] which
3853 * is guaranteed to always have tags allocated
3855 set
->map
[j
].mq_map
[i
] = 0;
3858 hctx
= blk_mq_map_queue_type(q
, j
, i
);
3859 ctx
->hctxs
[j
] = hctx
;
3861 * If the CPU is already set in the mask, then we've
3862 * mapped this one already. This can happen if
3863 * devices share queues across queue maps.
3865 if (cpumask_test_cpu(i
, hctx
->cpumask
))
3868 cpumask_set_cpu(i
, hctx
->cpumask
);
3870 ctx
->index_hw
[hctx
->type
] = hctx
->nr_ctx
;
3871 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
3874 * If the nr_ctx type overflows, we have exceeded the
3875 * amount of sw queues we can support.
3877 BUG_ON(!hctx
->nr_ctx
);
3880 for (; j
< HCTX_MAX_TYPES
; j
++)
3881 ctx
->hctxs
[j
] = blk_mq_map_queue_type(q
,
3882 HCTX_TYPE_DEFAULT
, i
);
3885 queue_for_each_hw_ctx(q
, hctx
, i
) {
3887 * If no software queues are mapped to this hardware queue,
3888 * disable it and free the request entries.
3890 if (!hctx
->nr_ctx
) {
3891 /* Never unmap queue 0. We need it as a
3892 * fallback in case of a new remap fails
3896 __blk_mq_free_map_and_rqs(set
, i
);
3902 hctx
->tags
= set
->tags
[i
];
3903 WARN_ON(!hctx
->tags
);
3906 * Set the map size to the number of mapped software queues.
3907 * This is more accurate and more efficient than looping
3908 * over all possibly mapped software queues.
3910 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
3913 * Initialize batch roundrobin counts
3915 hctx
->next_cpu
= blk_mq_first_mapped_cpu(hctx
);
3916 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
3921 * Caller needs to ensure that we're either frozen/quiesced, or that
3922 * the queue isn't live yet.
3924 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
3926 struct blk_mq_hw_ctx
*hctx
;
3929 queue_for_each_hw_ctx(q
, hctx
, i
) {
3931 hctx
->flags
|= BLK_MQ_F_TAG_QUEUE_SHARED
;
3933 blk_mq_tag_idle(hctx
);
3934 hctx
->flags
&= ~BLK_MQ_F_TAG_QUEUE_SHARED
;
3939 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set
*set
,
3942 struct request_queue
*q
;
3944 lockdep_assert_held(&set
->tag_list_lock
);
3946 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3947 blk_mq_freeze_queue(q
);
3948 queue_set_hctx_shared(q
, shared
);
3949 blk_mq_unfreeze_queue(q
);
3953 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
3955 struct blk_mq_tag_set
*set
= q
->tag_set
;
3957 mutex_lock(&set
->tag_list_lock
);
3958 list_del(&q
->tag_set_list
);
3959 if (list_is_singular(&set
->tag_list
)) {
3960 /* just transitioned to unshared */
3961 set
->flags
&= ~BLK_MQ_F_TAG_QUEUE_SHARED
;
3962 /* update existing queue */
3963 blk_mq_update_tag_set_shared(set
, false);
3965 mutex_unlock(&set
->tag_list_lock
);
3966 INIT_LIST_HEAD(&q
->tag_set_list
);
3969 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
3970 struct request_queue
*q
)
3972 mutex_lock(&set
->tag_list_lock
);
3975 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3977 if (!list_empty(&set
->tag_list
) &&
3978 !(set
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
)) {
3979 set
->flags
|= BLK_MQ_F_TAG_QUEUE_SHARED
;
3980 /* update existing queue */
3981 blk_mq_update_tag_set_shared(set
, true);
3983 if (set
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
)
3984 queue_set_hctx_shared(q
, true);
3985 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
3987 mutex_unlock(&set
->tag_list_lock
);
3990 /* All allocations will be freed in release handler of q->mq_kobj */
3991 static int blk_mq_alloc_ctxs(struct request_queue
*q
)
3993 struct blk_mq_ctxs
*ctxs
;
3996 ctxs
= kzalloc(sizeof(*ctxs
), GFP_KERNEL
);
4000 ctxs
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
4001 if (!ctxs
->queue_ctx
)
4004 for_each_possible_cpu(cpu
) {
4005 struct blk_mq_ctx
*ctx
= per_cpu_ptr(ctxs
->queue_ctx
, cpu
);
4009 q
->mq_kobj
= &ctxs
->kobj
;
4010 q
->queue_ctx
= ctxs
->queue_ctx
;
4019 * It is the actual release handler for mq, but we do it from
4020 * request queue's release handler for avoiding use-after-free
4021 * and headache because q->mq_kobj shouldn't have been introduced,
4022 * but we can't group ctx/kctx kobj without it.
4024 void blk_mq_release(struct request_queue
*q
)
4026 struct blk_mq_hw_ctx
*hctx
, *next
;
4029 queue_for_each_hw_ctx(q
, hctx
, i
)
4030 WARN_ON_ONCE(hctx
&& list_empty(&hctx
->hctx_list
));
4032 /* all hctx are in .unused_hctx_list now */
4033 list_for_each_entry_safe(hctx
, next
, &q
->unused_hctx_list
, hctx_list
) {
4034 list_del_init(&hctx
->hctx_list
);
4035 kobject_put(&hctx
->kobj
);
4038 xa_destroy(&q
->hctx_table
);
4041 * release .mq_kobj and sw queue's kobject now because
4042 * both share lifetime with request queue.
4044 blk_mq_sysfs_deinit(q
);
4047 static struct request_queue
*blk_mq_init_queue_data(struct blk_mq_tag_set
*set
,
4050 struct request_queue
*q
;
4053 q
= blk_alloc_queue(set
->numa_node
);
4055 return ERR_PTR(-ENOMEM
);
4056 q
->queuedata
= queuedata
;
4057 ret
= blk_mq_init_allocated_queue(set
, q
);
4060 return ERR_PTR(ret
);
4065 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
4067 return blk_mq_init_queue_data(set
, NULL
);
4069 EXPORT_SYMBOL(blk_mq_init_queue
);
4072 * blk_mq_destroy_queue - shutdown a request queue
4073 * @q: request queue to shutdown
4075 * This shuts down a request queue allocated by blk_mq_init_queue(). All future
4076 * requests will be failed with -ENODEV. The caller is responsible for dropping
4077 * the reference from blk_mq_init_queue() by calling blk_put_queue().
4079 * Context: can sleep
4081 void blk_mq_destroy_queue(struct request_queue
*q
)
4083 WARN_ON_ONCE(!queue_is_mq(q
));
4084 WARN_ON_ONCE(blk_queue_registered(q
));
4088 blk_queue_flag_set(QUEUE_FLAG_DYING
, q
);
4089 blk_queue_start_drain(q
);
4090 blk_mq_freeze_queue_wait(q
);
4093 blk_mq_cancel_work_sync(q
);
4094 blk_mq_exit_queue(q
);
4096 EXPORT_SYMBOL(blk_mq_destroy_queue
);
4098 struct gendisk
*__blk_mq_alloc_disk(struct blk_mq_tag_set
*set
, void *queuedata
,
4099 struct lock_class_key
*lkclass
)
4101 struct request_queue
*q
;
4102 struct gendisk
*disk
;
4104 q
= blk_mq_init_queue_data(set
, queuedata
);
4108 disk
= __alloc_disk_node(q
, set
->numa_node
, lkclass
);
4110 blk_mq_destroy_queue(q
);
4112 return ERR_PTR(-ENOMEM
);
4114 set_bit(GD_OWNS_QUEUE
, &disk
->state
);
4117 EXPORT_SYMBOL(__blk_mq_alloc_disk
);
4119 struct gendisk
*blk_mq_alloc_disk_for_queue(struct request_queue
*q
,
4120 struct lock_class_key
*lkclass
)
4122 struct gendisk
*disk
;
4124 if (!blk_get_queue(q
))
4126 disk
= __alloc_disk_node(q
, NUMA_NO_NODE
, lkclass
);
4131 EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue
);
4133 static struct blk_mq_hw_ctx
*blk_mq_alloc_and_init_hctx(
4134 struct blk_mq_tag_set
*set
, struct request_queue
*q
,
4135 int hctx_idx
, int node
)
4137 struct blk_mq_hw_ctx
*hctx
= NULL
, *tmp
;
4139 /* reuse dead hctx first */
4140 spin_lock(&q
->unused_hctx_lock
);
4141 list_for_each_entry(tmp
, &q
->unused_hctx_list
, hctx_list
) {
4142 if (tmp
->numa_node
== node
) {
4148 list_del_init(&hctx
->hctx_list
);
4149 spin_unlock(&q
->unused_hctx_lock
);
4152 hctx
= blk_mq_alloc_hctx(q
, set
, node
);
4156 if (blk_mq_init_hctx(q
, set
, hctx
, hctx_idx
))
4162 kobject_put(&hctx
->kobj
);
4167 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
4168 struct request_queue
*q
)
4170 struct blk_mq_hw_ctx
*hctx
;
4173 /* protect against switching io scheduler */
4174 mutex_lock(&q
->sysfs_lock
);
4175 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
4177 int node
= blk_mq_get_hctx_node(set
, i
);
4178 struct blk_mq_hw_ctx
*old_hctx
= xa_load(&q
->hctx_table
, i
);
4181 old_node
= old_hctx
->numa_node
;
4182 blk_mq_exit_hctx(q
, set
, old_hctx
, i
);
4185 if (!blk_mq_alloc_and_init_hctx(set
, q
, i
, node
)) {
4188 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4190 hctx
= blk_mq_alloc_and_init_hctx(set
, q
, i
, old_node
);
4191 WARN_ON_ONCE(!hctx
);
4195 * Increasing nr_hw_queues fails. Free the newly allocated
4196 * hctxs and keep the previous q->nr_hw_queues.
4198 if (i
!= set
->nr_hw_queues
) {
4199 j
= q
->nr_hw_queues
;
4202 q
->nr_hw_queues
= set
->nr_hw_queues
;
4205 xa_for_each_start(&q
->hctx_table
, j
, hctx
, j
)
4206 blk_mq_exit_hctx(q
, set
, hctx
, j
);
4207 mutex_unlock(&q
->sysfs_lock
);
4210 static void blk_mq_update_poll_flag(struct request_queue
*q
)
4212 struct blk_mq_tag_set
*set
= q
->tag_set
;
4214 if (set
->nr_maps
> HCTX_TYPE_POLL
&&
4215 set
->map
[HCTX_TYPE_POLL
].nr_queues
)
4216 blk_queue_flag_set(QUEUE_FLAG_POLL
, q
);
4218 blk_queue_flag_clear(QUEUE_FLAG_POLL
, q
);
4221 int blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
4222 struct request_queue
*q
)
4224 /* mark the queue as mq asap */
4225 q
->mq_ops
= set
->ops
;
4227 if (blk_mq_alloc_ctxs(q
))
4230 /* init q->mq_kobj and sw queues' kobjects */
4231 blk_mq_sysfs_init(q
);
4233 INIT_LIST_HEAD(&q
->unused_hctx_list
);
4234 spin_lock_init(&q
->unused_hctx_lock
);
4236 xa_init(&q
->hctx_table
);
4238 blk_mq_realloc_hw_ctxs(set
, q
);
4239 if (!q
->nr_hw_queues
)
4242 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
4243 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
4247 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
4248 blk_mq_update_poll_flag(q
);
4250 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
4251 INIT_LIST_HEAD(&q
->flush_list
);
4252 INIT_LIST_HEAD(&q
->requeue_list
);
4253 spin_lock_init(&q
->requeue_lock
);
4255 q
->nr_requests
= set
->queue_depth
;
4257 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
4258 blk_mq_add_queue_tag_set(set
, q
);
4259 blk_mq_map_swqueue(q
);
4268 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
4270 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4271 void blk_mq_exit_queue(struct request_queue
*q
)
4273 struct blk_mq_tag_set
*set
= q
->tag_set
;
4275 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4276 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
4277 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4278 blk_mq_del_queue_tag_set(q
);
4281 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
4285 if (blk_mq_is_shared_tags(set
->flags
)) {
4286 set
->shared_tags
= blk_mq_alloc_map_and_rqs(set
,
4289 if (!set
->shared_tags
)
4293 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
4294 if (!__blk_mq_alloc_map_and_rqs(set
, i
))
4303 __blk_mq_free_map_and_rqs(set
, i
);
4305 if (blk_mq_is_shared_tags(set
->flags
)) {
4306 blk_mq_free_map_and_rqs(set
, set
->shared_tags
,
4307 BLK_MQ_NO_HCTX_IDX
);
4314 * Allocate the request maps associated with this tag_set. Note that this
4315 * may reduce the depth asked for, if memory is tight. set->queue_depth
4316 * will be updated to reflect the allocated depth.
4318 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set
*set
)
4323 depth
= set
->queue_depth
;
4325 err
= __blk_mq_alloc_rq_maps(set
);
4329 set
->queue_depth
>>= 1;
4330 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
4334 } while (set
->queue_depth
);
4336 if (!set
->queue_depth
|| err
) {
4337 pr_err("blk-mq: failed to allocate request map\n");
4341 if (depth
!= set
->queue_depth
)
4342 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4343 depth
, set
->queue_depth
);
4348 static void blk_mq_update_queue_map(struct blk_mq_tag_set
*set
)
4351 * blk_mq_map_queues() and multiple .map_queues() implementations
4352 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4353 * number of hardware queues.
4355 if (set
->nr_maps
== 1)
4356 set
->map
[HCTX_TYPE_DEFAULT
].nr_queues
= set
->nr_hw_queues
;
4358 if (set
->ops
->map_queues
&& !is_kdump_kernel()) {
4362 * transport .map_queues is usually done in the following
4365 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4366 * mask = get_cpu_mask(queue)
4367 * for_each_cpu(cpu, mask)
4368 * set->map[x].mq_map[cpu] = queue;
4371 * When we need to remap, the table has to be cleared for
4372 * killing stale mapping since one CPU may not be mapped
4375 for (i
= 0; i
< set
->nr_maps
; i
++)
4376 blk_mq_clear_mq_map(&set
->map
[i
]);
4378 set
->ops
->map_queues(set
);
4380 BUG_ON(set
->nr_maps
> 1);
4381 blk_mq_map_queues(&set
->map
[HCTX_TYPE_DEFAULT
]);
4385 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set
*set
,
4386 int new_nr_hw_queues
)
4388 struct blk_mq_tags
**new_tags
;
4391 if (set
->nr_hw_queues
>= new_nr_hw_queues
)
4394 new_tags
= kcalloc_node(new_nr_hw_queues
, sizeof(struct blk_mq_tags
*),
4395 GFP_KERNEL
, set
->numa_node
);
4400 memcpy(new_tags
, set
->tags
, set
->nr_hw_queues
*
4401 sizeof(*set
->tags
));
4403 set
->tags
= new_tags
;
4405 for (i
= set
->nr_hw_queues
; i
< new_nr_hw_queues
; i
++) {
4406 if (!__blk_mq_alloc_map_and_rqs(set
, i
)) {
4407 while (--i
>= set
->nr_hw_queues
)
4408 __blk_mq_free_map_and_rqs(set
, i
);
4415 set
->nr_hw_queues
= new_nr_hw_queues
;
4420 * Alloc a tag set to be associated with one or more request queues.
4421 * May fail with EINVAL for various error conditions. May adjust the
4422 * requested depth down, if it's too large. In that case, the set
4423 * value will be stored in set->queue_depth.
4425 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
4429 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
4431 if (!set
->nr_hw_queues
)
4433 if (!set
->queue_depth
)
4435 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
4438 if (!set
->ops
->queue_rq
)
4441 if (!set
->ops
->get_budget
^ !set
->ops
->put_budget
)
4444 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
4445 pr_info("blk-mq: reduced tag depth to %u\n",
4447 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
4452 else if (set
->nr_maps
> HCTX_MAX_TYPES
)
4456 * If a crashdump is active, then we are potentially in a very
4457 * memory constrained environment. Limit us to 1 queue and
4458 * 64 tags to prevent using too much memory.
4460 if (is_kdump_kernel()) {
4461 set
->nr_hw_queues
= 1;
4463 set
->queue_depth
= min(64U, set
->queue_depth
);
4466 * There is no use for more h/w queues than cpus if we just have
4469 if (set
->nr_maps
== 1 && set
->nr_hw_queues
> nr_cpu_ids
)
4470 set
->nr_hw_queues
= nr_cpu_ids
;
4472 if (set
->flags
& BLK_MQ_F_BLOCKING
) {
4473 set
->srcu
= kmalloc(sizeof(*set
->srcu
), GFP_KERNEL
);
4476 ret
= init_srcu_struct(set
->srcu
);
4482 set
->tags
= kcalloc_node(set
->nr_hw_queues
,
4483 sizeof(struct blk_mq_tags
*), GFP_KERNEL
,
4486 goto out_cleanup_srcu
;
4488 for (i
= 0; i
< set
->nr_maps
; i
++) {
4489 set
->map
[i
].mq_map
= kcalloc_node(nr_cpu_ids
,
4490 sizeof(set
->map
[i
].mq_map
[0]),
4491 GFP_KERNEL
, set
->numa_node
);
4492 if (!set
->map
[i
].mq_map
)
4493 goto out_free_mq_map
;
4494 set
->map
[i
].nr_queues
= is_kdump_kernel() ? 1 : set
->nr_hw_queues
;
4497 blk_mq_update_queue_map(set
);
4499 ret
= blk_mq_alloc_set_map_and_rqs(set
);
4501 goto out_free_mq_map
;
4503 mutex_init(&set
->tag_list_lock
);
4504 INIT_LIST_HEAD(&set
->tag_list
);
4509 for (i
= 0; i
< set
->nr_maps
; i
++) {
4510 kfree(set
->map
[i
].mq_map
);
4511 set
->map
[i
].mq_map
= NULL
;
4516 if (set
->flags
& BLK_MQ_F_BLOCKING
)
4517 cleanup_srcu_struct(set
->srcu
);
4519 if (set
->flags
& BLK_MQ_F_BLOCKING
)
4523 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
4525 /* allocate and initialize a tagset for a simple single-queue device */
4526 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set
*set
,
4527 const struct blk_mq_ops
*ops
, unsigned int queue_depth
,
4528 unsigned int set_flags
)
4530 memset(set
, 0, sizeof(*set
));
4532 set
->nr_hw_queues
= 1;
4534 set
->queue_depth
= queue_depth
;
4535 set
->numa_node
= NUMA_NO_NODE
;
4536 set
->flags
= set_flags
;
4537 return blk_mq_alloc_tag_set(set
);
4539 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set
);
4541 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
4545 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
4546 __blk_mq_free_map_and_rqs(set
, i
);
4548 if (blk_mq_is_shared_tags(set
->flags
)) {
4549 blk_mq_free_map_and_rqs(set
, set
->shared_tags
,
4550 BLK_MQ_NO_HCTX_IDX
);
4553 for (j
= 0; j
< set
->nr_maps
; j
++) {
4554 kfree(set
->map
[j
].mq_map
);
4555 set
->map
[j
].mq_map
= NULL
;
4560 if (set
->flags
& BLK_MQ_F_BLOCKING
) {
4561 cleanup_srcu_struct(set
->srcu
);
4565 EXPORT_SYMBOL(blk_mq_free_tag_set
);
4567 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
4569 struct blk_mq_tag_set
*set
= q
->tag_set
;
4570 struct blk_mq_hw_ctx
*hctx
;
4577 if (q
->nr_requests
== nr
)
4580 blk_mq_freeze_queue(q
);
4581 blk_mq_quiesce_queue(q
);
4584 queue_for_each_hw_ctx(q
, hctx
, i
) {
4588 * If we're using an MQ scheduler, just update the scheduler
4589 * queue depth. This is similar to what the old code would do.
4591 if (hctx
->sched_tags
) {
4592 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
4595 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
, nr
,
4600 if (q
->elevator
&& q
->elevator
->type
->ops
.depth_updated
)
4601 q
->elevator
->type
->ops
.depth_updated(hctx
);
4604 q
->nr_requests
= nr
;
4605 if (blk_mq_is_shared_tags(set
->flags
)) {
4607 blk_mq_tag_update_sched_shared_tags(q
);
4609 blk_mq_tag_resize_shared_tags(set
, nr
);
4613 blk_mq_unquiesce_queue(q
);
4614 blk_mq_unfreeze_queue(q
);
4620 * request_queue and elevator_type pair.
4621 * It is just used by __blk_mq_update_nr_hw_queues to cache
4622 * the elevator_type associated with a request_queue.
4624 struct blk_mq_qe_pair
{
4625 struct list_head node
;
4626 struct request_queue
*q
;
4627 struct elevator_type
*type
;
4631 * Cache the elevator_type in qe pair list and switch the
4632 * io scheduler to 'none'
4634 static bool blk_mq_elv_switch_none(struct list_head
*head
,
4635 struct request_queue
*q
)
4637 struct blk_mq_qe_pair
*qe
;
4639 qe
= kmalloc(sizeof(*qe
), GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
);
4643 /* q->elevator needs protection from ->sysfs_lock */
4644 mutex_lock(&q
->sysfs_lock
);
4646 /* the check has to be done with holding sysfs_lock */
4652 INIT_LIST_HEAD(&qe
->node
);
4654 qe
->type
= q
->elevator
->type
;
4655 /* keep a reference to the elevator module as we'll switch back */
4656 __elevator_get(qe
->type
);
4657 list_add(&qe
->node
, head
);
4658 elevator_disable(q
);
4660 mutex_unlock(&q
->sysfs_lock
);
4665 static struct blk_mq_qe_pair
*blk_lookup_qe_pair(struct list_head
*head
,
4666 struct request_queue
*q
)
4668 struct blk_mq_qe_pair
*qe
;
4670 list_for_each_entry(qe
, head
, node
)
4677 static void blk_mq_elv_switch_back(struct list_head
*head
,
4678 struct request_queue
*q
)
4680 struct blk_mq_qe_pair
*qe
;
4681 struct elevator_type
*t
;
4683 qe
= blk_lookup_qe_pair(head
, q
);
4687 list_del(&qe
->node
);
4690 mutex_lock(&q
->sysfs_lock
);
4691 elevator_switch(q
, t
);
4692 /* drop the reference acquired in blk_mq_elv_switch_none */
4694 mutex_unlock(&q
->sysfs_lock
);
4697 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
,
4700 struct request_queue
*q
;
4702 int prev_nr_hw_queues
= set
->nr_hw_queues
;
4705 lockdep_assert_held(&set
->tag_list_lock
);
4707 if (set
->nr_maps
== 1 && nr_hw_queues
> nr_cpu_ids
)
4708 nr_hw_queues
= nr_cpu_ids
;
4709 if (nr_hw_queues
< 1)
4711 if (set
->nr_maps
== 1 && nr_hw_queues
== set
->nr_hw_queues
)
4714 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
4715 blk_mq_freeze_queue(q
);
4717 * Switch IO scheduler to 'none', cleaning up the data associated
4718 * with the previous scheduler. We will switch back once we are done
4719 * updating the new sw to hw queue mappings.
4721 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
4722 if (!blk_mq_elv_switch_none(&head
, q
))
4725 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
4726 blk_mq_debugfs_unregister_hctxs(q
);
4727 blk_mq_sysfs_unregister_hctxs(q
);
4730 if (blk_mq_realloc_tag_set_tags(set
, nr_hw_queues
) < 0)
4734 blk_mq_update_queue_map(set
);
4735 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
4736 blk_mq_realloc_hw_ctxs(set
, q
);
4737 blk_mq_update_poll_flag(q
);
4738 if (q
->nr_hw_queues
!= set
->nr_hw_queues
) {
4739 int i
= prev_nr_hw_queues
;
4741 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4742 nr_hw_queues
, prev_nr_hw_queues
);
4743 for (; i
< set
->nr_hw_queues
; i
++)
4744 __blk_mq_free_map_and_rqs(set
, i
);
4746 set
->nr_hw_queues
= prev_nr_hw_queues
;
4749 blk_mq_map_swqueue(q
);
4753 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
4754 blk_mq_sysfs_register_hctxs(q
);
4755 blk_mq_debugfs_register_hctxs(q
);
4759 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
4760 blk_mq_elv_switch_back(&head
, q
);
4762 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
4763 blk_mq_unfreeze_queue(q
);
4765 /* Free the excess tags when nr_hw_queues shrink. */
4766 for (i
= set
->nr_hw_queues
; i
< prev_nr_hw_queues
; i
++)
4767 __blk_mq_free_map_and_rqs(set
, i
);
4770 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
4772 mutex_lock(&set
->tag_list_lock
);
4773 __blk_mq_update_nr_hw_queues(set
, nr_hw_queues
);
4774 mutex_unlock(&set
->tag_list_lock
);
4776 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
4778 static int blk_hctx_poll(struct request_queue
*q
, struct blk_mq_hw_ctx
*hctx
,
4779 struct io_comp_batch
*iob
, unsigned int flags
)
4781 long state
= get_current_state();
4785 ret
= q
->mq_ops
->poll(hctx
, iob
);
4787 __set_current_state(TASK_RUNNING
);
4791 if (signal_pending_state(state
, current
))
4792 __set_current_state(TASK_RUNNING
);
4793 if (task_is_running(current
))
4796 if (ret
< 0 || (flags
& BLK_POLL_ONESHOT
))
4799 } while (!need_resched());
4801 __set_current_state(TASK_RUNNING
);
4805 int blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
,
4806 struct io_comp_batch
*iob
, unsigned int flags
)
4808 struct blk_mq_hw_ctx
*hctx
= xa_load(&q
->hctx_table
, cookie
);
4810 return blk_hctx_poll(q
, hctx
, iob
, flags
);
4813 int blk_rq_poll(struct request
*rq
, struct io_comp_batch
*iob
,
4814 unsigned int poll_flags
)
4816 struct request_queue
*q
= rq
->q
;
4819 if (!blk_rq_is_poll(rq
))
4821 if (!percpu_ref_tryget(&q
->q_usage_counter
))
4824 ret
= blk_hctx_poll(q
, rq
->mq_hctx
, iob
, poll_flags
);
4829 EXPORT_SYMBOL_GPL(blk_rq_poll
);
4831 unsigned int blk_mq_rq_cpu(struct request
*rq
)
4833 return rq
->mq_ctx
->cpu
;
4835 EXPORT_SYMBOL(blk_mq_rq_cpu
);
4837 void blk_mq_cancel_work_sync(struct request_queue
*q
)
4839 struct blk_mq_hw_ctx
*hctx
;
4842 cancel_delayed_work_sync(&q
->requeue_work
);
4844 queue_for_each_hw_ctx(q
, hctx
, i
)
4845 cancel_delayed_work_sync(&hctx
->run_work
);
4848 static int __init
blk_mq_init(void)
4852 for_each_possible_cpu(i
)
4853 init_llist_head(&per_cpu(blk_cpu_done
, i
));
4854 for_each_possible_cpu(i
)
4855 INIT_CSD(&per_cpu(blk_cpu_csd
, i
),
4856 __blk_mq_complete_request_remote
, NULL
);
4857 open_softirq(BLOCK_SOFTIRQ
, blk_done_softirq
);
4859 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD
,
4860 "block/softirq:dead", NULL
,
4861 blk_softirq_cpu_dead
);
4862 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
4863 blk_mq_hctx_notify_dead
);
4864 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE
, "block/mq:online",
4865 blk_mq_hctx_notify_online
,
4866 blk_mq_hctx_notify_offline
);
4869 subsys_initcall(blk_mq_init
);