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.
775 * For such case, force the completed nbytes to be equal to
776 * the BIO size so that bio_advance() sets the BIO remaining
777 * size to 0 and we end up calling bio_endio() before returning.
779 if (bio
->bi_iter
.bi_size
!= nbytes
) {
780 bio
->bi_status
= BLK_STS_IOERR
;
781 nbytes
= bio
->bi_iter
.bi_size
;
783 bio
->bi_iter
.bi_sector
= rq
->__sector
;
787 bio_advance(bio
, nbytes
);
789 if (unlikely(rq
->rq_flags
& RQF_QUIET
))
790 bio_set_flag(bio
, BIO_QUIET
);
791 /* don't actually finish bio if it's part of flush sequence */
792 if (bio
->bi_iter
.bi_size
== 0 && !(rq
->rq_flags
& RQF_FLUSH_SEQ
))
796 static void blk_account_io_completion(struct request
*req
, unsigned int bytes
)
798 if (req
->part
&& blk_do_io_stat(req
)) {
799 const int sgrp
= op_stat_group(req_op(req
));
802 part_stat_add(req
->part
, sectors
[sgrp
], bytes
>> 9);
807 static void blk_print_req_error(struct request
*req
, blk_status_t status
)
809 printk_ratelimited(KERN_ERR
810 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
811 "phys_seg %u prio class %u\n",
812 blk_status_to_str(status
),
813 req
->q
->disk
? req
->q
->disk
->disk_name
: "?",
814 blk_rq_pos(req
), (__force u32
)req_op(req
),
815 blk_op_str(req_op(req
)),
816 (__force u32
)(req
->cmd_flags
& ~REQ_OP_MASK
),
817 req
->nr_phys_segments
,
818 IOPRIO_PRIO_CLASS(req
->ioprio
));
822 * Fully end IO on a request. Does not support partial completions, or
825 static void blk_complete_request(struct request
*req
)
827 const bool is_flush
= (req
->rq_flags
& RQF_FLUSH_SEQ
) != 0;
828 int total_bytes
= blk_rq_bytes(req
);
829 struct bio
*bio
= req
->bio
;
831 trace_block_rq_complete(req
, BLK_STS_OK
, total_bytes
);
836 #ifdef CONFIG_BLK_DEV_INTEGRITY
837 if (blk_integrity_rq(req
) && req_op(req
) == REQ_OP_READ
)
838 req
->q
->integrity
.profile
->complete_fn(req
, total_bytes
);
842 * Upper layers may call blk_crypto_evict_key() anytime after the last
843 * bio_endio(). Therefore, the keyslot must be released before that.
845 blk_crypto_rq_put_keyslot(req
);
847 blk_account_io_completion(req
, total_bytes
);
850 struct bio
*next
= bio
->bi_next
;
852 /* Completion has already been traced */
853 bio_clear_flag(bio
, BIO_TRACE_COMPLETION
);
855 if (req_op(req
) == REQ_OP_ZONE_APPEND
)
856 bio
->bi_iter
.bi_sector
= req
->__sector
;
864 * Reset counters so that the request stacking driver
865 * can find how many bytes remain in the request
875 * blk_update_request - Complete multiple bytes without completing the request
876 * @req: the request being processed
877 * @error: block status code
878 * @nr_bytes: number of bytes to complete for @req
881 * Ends I/O on a number of bytes attached to @req, but doesn't complete
882 * the request structure even if @req doesn't have leftover.
883 * If @req has leftover, sets it up for the next range of segments.
885 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
886 * %false return from this function.
889 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
890 * except in the consistency check at the end of this function.
893 * %false - this request doesn't have any more data
894 * %true - this request has more data
896 bool blk_update_request(struct request
*req
, blk_status_t error
,
897 unsigned int nr_bytes
)
901 trace_block_rq_complete(req
, error
, nr_bytes
);
906 #ifdef CONFIG_BLK_DEV_INTEGRITY
907 if (blk_integrity_rq(req
) && req_op(req
) == REQ_OP_READ
&&
909 req
->q
->integrity
.profile
->complete_fn(req
, nr_bytes
);
913 * Upper layers may call blk_crypto_evict_key() anytime after the last
914 * bio_endio(). Therefore, the keyslot must be released before that.
916 if (blk_crypto_rq_has_keyslot(req
) && nr_bytes
>= blk_rq_bytes(req
))
917 __blk_crypto_rq_put_keyslot(req
);
919 if (unlikely(error
&& !blk_rq_is_passthrough(req
) &&
920 !(req
->rq_flags
& RQF_QUIET
)) &&
921 !test_bit(GD_DEAD
, &req
->q
->disk
->state
)) {
922 blk_print_req_error(req
, error
);
923 trace_block_rq_error(req
, error
, nr_bytes
);
926 blk_account_io_completion(req
, nr_bytes
);
930 struct bio
*bio
= req
->bio
;
931 unsigned bio_bytes
= min(bio
->bi_iter
.bi_size
, nr_bytes
);
933 if (bio_bytes
== bio
->bi_iter
.bi_size
)
934 req
->bio
= bio
->bi_next
;
936 /* Completion has already been traced */
937 bio_clear_flag(bio
, BIO_TRACE_COMPLETION
);
938 req_bio_endio(req
, bio
, bio_bytes
, error
);
940 total_bytes
+= bio_bytes
;
941 nr_bytes
-= bio_bytes
;
952 * Reset counters so that the request stacking driver
953 * can find how many bytes remain in the request
960 req
->__data_len
-= total_bytes
;
962 /* update sector only for requests with clear definition of sector */
963 if (!blk_rq_is_passthrough(req
))
964 req
->__sector
+= total_bytes
>> 9;
966 /* mixed attributes always follow the first bio */
967 if (req
->rq_flags
& RQF_MIXED_MERGE
) {
968 req
->cmd_flags
&= ~REQ_FAILFAST_MASK
;
969 req
->cmd_flags
|= req
->bio
->bi_opf
& REQ_FAILFAST_MASK
;
972 if (!(req
->rq_flags
& RQF_SPECIAL_PAYLOAD
)) {
974 * If total number of sectors is less than the first segment
975 * size, something has gone terribly wrong.
977 if (blk_rq_bytes(req
) < blk_rq_cur_bytes(req
)) {
978 blk_dump_rq_flags(req
, "request botched");
979 req
->__data_len
= blk_rq_cur_bytes(req
);
982 /* recalculate the number of segments */
983 req
->nr_phys_segments
= blk_recalc_rq_segments(req
);
988 EXPORT_SYMBOL_GPL(blk_update_request
);
990 static inline void blk_account_io_done(struct request
*req
, u64 now
)
992 trace_block_io_done(req
);
995 * Account IO completion. flush_rq isn't accounted as a
996 * normal IO on queueing nor completion. Accounting the
997 * containing request is enough.
999 if (blk_do_io_stat(req
) && req
->part
&&
1000 !(req
->rq_flags
& RQF_FLUSH_SEQ
)) {
1001 const int sgrp
= op_stat_group(req_op(req
));
1004 update_io_ticks(req
->part
, jiffies
, true);
1005 part_stat_inc(req
->part
, ios
[sgrp
]);
1006 part_stat_add(req
->part
, nsecs
[sgrp
], now
- req
->start_time_ns
);
1011 static inline void blk_account_io_start(struct request
*req
)
1013 trace_block_io_start(req
);
1015 if (blk_do_io_stat(req
)) {
1017 * All non-passthrough requests are created from a bio with one
1018 * exception: when a flush command that is part of a flush sequence
1019 * generated by the state machine in blk-flush.c is cloned onto the
1020 * lower device by dm-multipath we can get here without a bio.
1023 req
->part
= req
->bio
->bi_bdev
;
1025 req
->part
= req
->q
->disk
->part0
;
1028 update_io_ticks(req
->part
, jiffies
, false);
1033 static inline void __blk_mq_end_request_acct(struct request
*rq
, u64 now
)
1035 if (rq
->rq_flags
& RQF_STATS
)
1036 blk_stat_add(rq
, now
);
1038 blk_mq_sched_completed_request(rq
, now
);
1039 blk_account_io_done(rq
, now
);
1042 inline void __blk_mq_end_request(struct request
*rq
, blk_status_t error
)
1044 if (blk_mq_need_time_stamp(rq
))
1045 __blk_mq_end_request_acct(rq
, ktime_get_ns());
1047 blk_mq_finish_request(rq
);
1050 rq_qos_done(rq
->q
, rq
);
1051 if (rq
->end_io(rq
, error
) == RQ_END_IO_FREE
)
1052 blk_mq_free_request(rq
);
1054 blk_mq_free_request(rq
);
1057 EXPORT_SYMBOL(__blk_mq_end_request
);
1059 void blk_mq_end_request(struct request
*rq
, blk_status_t error
)
1061 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
1063 __blk_mq_end_request(rq
, error
);
1065 EXPORT_SYMBOL(blk_mq_end_request
);
1067 #define TAG_COMP_BATCH 32
1069 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx
*hctx
,
1070 int *tag_array
, int nr_tags
)
1072 struct request_queue
*q
= hctx
->queue
;
1074 blk_mq_sub_active_requests(hctx
, nr_tags
);
1076 blk_mq_put_tags(hctx
->tags
, tag_array
, nr_tags
);
1077 percpu_ref_put_many(&q
->q_usage_counter
, nr_tags
);
1080 void blk_mq_end_request_batch(struct io_comp_batch
*iob
)
1082 int tags
[TAG_COMP_BATCH
], nr_tags
= 0;
1083 struct blk_mq_hw_ctx
*cur_hctx
= NULL
;
1088 now
= ktime_get_ns();
1090 while ((rq
= rq_list_pop(&iob
->req_list
)) != NULL
) {
1092 prefetch(rq
->rq_next
);
1094 blk_complete_request(rq
);
1096 __blk_mq_end_request_acct(rq
, now
);
1098 blk_mq_finish_request(rq
);
1100 rq_qos_done(rq
->q
, rq
);
1103 * If end_io handler returns NONE, then it still has
1104 * ownership of the request.
1106 if (rq
->end_io
&& rq
->end_io(rq
, 0) == RQ_END_IO_NONE
)
1109 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
1110 if (!req_ref_put_and_test(rq
))
1113 blk_crypto_free_request(rq
);
1114 blk_pm_mark_last_busy(rq
);
1116 if (nr_tags
== TAG_COMP_BATCH
|| cur_hctx
!= rq
->mq_hctx
) {
1118 blk_mq_flush_tag_batch(cur_hctx
, tags
, nr_tags
);
1120 cur_hctx
= rq
->mq_hctx
;
1122 tags
[nr_tags
++] = rq
->tag
;
1126 blk_mq_flush_tag_batch(cur_hctx
, tags
, nr_tags
);
1128 EXPORT_SYMBOL_GPL(blk_mq_end_request_batch
);
1130 static void blk_complete_reqs(struct llist_head
*list
)
1132 struct llist_node
*entry
= llist_reverse_order(llist_del_all(list
));
1133 struct request
*rq
, *next
;
1135 llist_for_each_entry_safe(rq
, next
, entry
, ipi_list
)
1136 rq
->q
->mq_ops
->complete(rq
);
1139 static __latent_entropy
void blk_done_softirq(struct softirq_action
*h
)
1141 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done
));
1144 static int blk_softirq_cpu_dead(unsigned int cpu
)
1146 blk_complete_reqs(&per_cpu(blk_cpu_done
, cpu
));
1150 static void __blk_mq_complete_request_remote(void *data
)
1152 __raise_softirq_irqoff(BLOCK_SOFTIRQ
);
1155 static inline bool blk_mq_complete_need_ipi(struct request
*rq
)
1157 int cpu
= raw_smp_processor_id();
1159 if (!IS_ENABLED(CONFIG_SMP
) ||
1160 !test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
))
1163 * With force threaded interrupts enabled, raising softirq from an SMP
1164 * function call will always result in waking the ksoftirqd thread.
1165 * This is probably worse than completing the request on a different
1168 if (force_irqthreads())
1171 /* same CPU or cache domain? Complete locally */
1172 if (cpu
== rq
->mq_ctx
->cpu
||
1173 (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
) &&
1174 cpus_share_cache(cpu
, rq
->mq_ctx
->cpu
)))
1177 /* don't try to IPI to an offline CPU */
1178 return cpu_online(rq
->mq_ctx
->cpu
);
1181 static void blk_mq_complete_send_ipi(struct request
*rq
)
1185 cpu
= rq
->mq_ctx
->cpu
;
1186 if (llist_add(&rq
->ipi_list
, &per_cpu(blk_cpu_done
, cpu
)))
1187 smp_call_function_single_async(cpu
, &per_cpu(blk_cpu_csd
, cpu
));
1190 static void blk_mq_raise_softirq(struct request
*rq
)
1192 struct llist_head
*list
;
1195 list
= this_cpu_ptr(&blk_cpu_done
);
1196 if (llist_add(&rq
->ipi_list
, list
))
1197 raise_softirq(BLOCK_SOFTIRQ
);
1201 bool blk_mq_complete_request_remote(struct request
*rq
)
1203 WRITE_ONCE(rq
->state
, MQ_RQ_COMPLETE
);
1206 * For request which hctx has only one ctx mapping,
1207 * or a polled request, always complete locally,
1208 * it's pointless to redirect the completion.
1210 if ((rq
->mq_hctx
->nr_ctx
== 1 &&
1211 rq
->mq_ctx
->cpu
== raw_smp_processor_id()) ||
1212 rq
->cmd_flags
& REQ_POLLED
)
1215 if (blk_mq_complete_need_ipi(rq
)) {
1216 blk_mq_complete_send_ipi(rq
);
1220 if (rq
->q
->nr_hw_queues
== 1) {
1221 blk_mq_raise_softirq(rq
);
1226 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote
);
1229 * blk_mq_complete_request - end I/O on a request
1230 * @rq: the request being processed
1233 * Complete a request by scheduling the ->complete_rq operation.
1235 void blk_mq_complete_request(struct request
*rq
)
1237 if (!blk_mq_complete_request_remote(rq
))
1238 rq
->q
->mq_ops
->complete(rq
);
1240 EXPORT_SYMBOL(blk_mq_complete_request
);
1243 * blk_mq_start_request - Start processing a request
1244 * @rq: Pointer to request to be started
1246 * Function used by device drivers to notify the block layer that a request
1247 * is going to be processed now, so blk layer can do proper initializations
1248 * such as starting the timeout timer.
1250 void blk_mq_start_request(struct request
*rq
)
1252 struct request_queue
*q
= rq
->q
;
1254 trace_block_rq_issue(rq
);
1256 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
) &&
1257 !blk_rq_is_passthrough(rq
)) {
1258 rq
->io_start_time_ns
= ktime_get_ns();
1259 rq
->stats_sectors
= blk_rq_sectors(rq
);
1260 rq
->rq_flags
|= RQF_STATS
;
1261 rq_qos_issue(q
, rq
);
1264 WARN_ON_ONCE(blk_mq_rq_state(rq
) != MQ_RQ_IDLE
);
1267 WRITE_ONCE(rq
->state
, MQ_RQ_IN_FLIGHT
);
1268 rq
->mq_hctx
->tags
->rqs
[rq
->tag
] = rq
;
1270 #ifdef CONFIG_BLK_DEV_INTEGRITY
1271 if (blk_integrity_rq(rq
) && req_op(rq
) == REQ_OP_WRITE
)
1272 q
->integrity
.profile
->prepare_fn(rq
);
1274 if (rq
->bio
&& rq
->bio
->bi_opf
& REQ_POLLED
)
1275 WRITE_ONCE(rq
->bio
->bi_cookie
, rq
->mq_hctx
->queue_num
);
1277 EXPORT_SYMBOL(blk_mq_start_request
);
1280 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1281 * queues. This is important for md arrays to benefit from merging
1284 static inline unsigned short blk_plug_max_rq_count(struct blk_plug
*plug
)
1286 if (plug
->multiple_queues
)
1287 return BLK_MAX_REQUEST_COUNT
* 2;
1288 return BLK_MAX_REQUEST_COUNT
;
1291 static void blk_add_rq_to_plug(struct blk_plug
*plug
, struct request
*rq
)
1293 struct request
*last
= rq_list_peek(&plug
->mq_list
);
1295 if (!plug
->rq_count
) {
1296 trace_block_plug(rq
->q
);
1297 } else if (plug
->rq_count
>= blk_plug_max_rq_count(plug
) ||
1298 (!blk_queue_nomerges(rq
->q
) &&
1299 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1300 blk_mq_flush_plug_list(plug
, false);
1302 trace_block_plug(rq
->q
);
1305 if (!plug
->multiple_queues
&& last
&& last
->q
!= rq
->q
)
1306 plug
->multiple_queues
= true;
1308 * Any request allocated from sched tags can't be issued to
1309 * ->queue_rqs() directly
1311 if (!plug
->has_elevator
&& (rq
->rq_flags
& RQF_SCHED_TAGS
))
1312 plug
->has_elevator
= true;
1314 rq_list_add(&plug
->mq_list
, rq
);
1319 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1320 * @rq: request to insert
1321 * @at_head: insert request at head or tail of queue
1324 * Insert a fully prepared request at the back of the I/O scheduler queue
1325 * for execution. Don't wait for completion.
1328 * This function will invoke @done directly if the queue is dead.
1330 void blk_execute_rq_nowait(struct request
*rq
, bool at_head
)
1332 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1334 WARN_ON(irqs_disabled());
1335 WARN_ON(!blk_rq_is_passthrough(rq
));
1337 blk_account_io_start(rq
);
1340 * As plugging can be enabled for passthrough requests on a zoned
1341 * device, directly accessing the plug instead of using blk_mq_plug()
1342 * should not have any consequences.
1344 if (current
->plug
&& !at_head
) {
1345 blk_add_rq_to_plug(current
->plug
, rq
);
1349 blk_mq_insert_request(rq
, at_head
? BLK_MQ_INSERT_AT_HEAD
: 0);
1350 blk_mq_run_hw_queue(hctx
, hctx
->flags
& BLK_MQ_F_BLOCKING
);
1352 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait
);
1354 struct blk_rq_wait
{
1355 struct completion done
;
1359 static enum rq_end_io_ret
blk_end_sync_rq(struct request
*rq
, blk_status_t ret
)
1361 struct blk_rq_wait
*wait
= rq
->end_io_data
;
1364 complete(&wait
->done
);
1365 return RQ_END_IO_NONE
;
1368 bool blk_rq_is_poll(struct request
*rq
)
1372 if (rq
->mq_hctx
->type
!= HCTX_TYPE_POLL
)
1376 EXPORT_SYMBOL_GPL(blk_rq_is_poll
);
1378 static void blk_rq_poll_completion(struct request
*rq
, struct completion
*wait
)
1381 blk_hctx_poll(rq
->q
, rq
->mq_hctx
, NULL
, 0);
1383 } while (!completion_done(wait
));
1387 * blk_execute_rq - insert a request into queue for execution
1388 * @rq: request to insert
1389 * @at_head: insert request at head or tail of queue
1392 * Insert a fully prepared request at the back of the I/O scheduler queue
1393 * for execution and wait for completion.
1394 * Return: The blk_status_t result provided to blk_mq_end_request().
1396 blk_status_t
blk_execute_rq(struct request
*rq
, bool at_head
)
1398 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1399 struct blk_rq_wait wait
= {
1400 .done
= COMPLETION_INITIALIZER_ONSTACK(wait
.done
),
1403 WARN_ON(irqs_disabled());
1404 WARN_ON(!blk_rq_is_passthrough(rq
));
1406 rq
->end_io_data
= &wait
;
1407 rq
->end_io
= blk_end_sync_rq
;
1409 blk_account_io_start(rq
);
1410 blk_mq_insert_request(rq
, at_head
? BLK_MQ_INSERT_AT_HEAD
: 0);
1411 blk_mq_run_hw_queue(hctx
, false);
1413 if (blk_rq_is_poll(rq
)) {
1414 blk_rq_poll_completion(rq
, &wait
.done
);
1417 * Prevent hang_check timer from firing at us during very long
1420 unsigned long hang_check
= sysctl_hung_task_timeout_secs
;
1423 while (!wait_for_completion_io_timeout(&wait
.done
,
1424 hang_check
* (HZ
/2)))
1427 wait_for_completion_io(&wait
.done
);
1432 EXPORT_SYMBOL(blk_execute_rq
);
1434 static void __blk_mq_requeue_request(struct request
*rq
)
1436 struct request_queue
*q
= rq
->q
;
1438 blk_mq_put_driver_tag(rq
);
1440 trace_block_rq_requeue(rq
);
1441 rq_qos_requeue(q
, rq
);
1443 if (blk_mq_request_started(rq
)) {
1444 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
1445 rq
->rq_flags
&= ~RQF_TIMED_OUT
;
1449 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
1451 struct request_queue
*q
= rq
->q
;
1452 unsigned long flags
;
1454 __blk_mq_requeue_request(rq
);
1456 /* this request will be re-inserted to io scheduler queue */
1457 blk_mq_sched_requeue_request(rq
);
1459 spin_lock_irqsave(&q
->requeue_lock
, flags
);
1460 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
1461 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
1463 if (kick_requeue_list
)
1464 blk_mq_kick_requeue_list(q
);
1466 EXPORT_SYMBOL(blk_mq_requeue_request
);
1468 static void blk_mq_requeue_work(struct work_struct
*work
)
1470 struct request_queue
*q
=
1471 container_of(work
, struct request_queue
, requeue_work
.work
);
1473 LIST_HEAD(flush_list
);
1476 spin_lock_irq(&q
->requeue_lock
);
1477 list_splice_init(&q
->requeue_list
, &rq_list
);
1478 list_splice_init(&q
->flush_list
, &flush_list
);
1479 spin_unlock_irq(&q
->requeue_lock
);
1481 while (!list_empty(&rq_list
)) {
1482 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
1484 * If RQF_DONTPREP ist set, the request has been started by the
1485 * driver already and might have driver-specific data allocated
1486 * already. Insert it into the hctx dispatch list to avoid
1487 * block layer merges for the request.
1489 if (rq
->rq_flags
& RQF_DONTPREP
) {
1490 list_del_init(&rq
->queuelist
);
1491 blk_mq_request_bypass_insert(rq
, 0);
1493 list_del_init(&rq
->queuelist
);
1494 blk_mq_insert_request(rq
, BLK_MQ_INSERT_AT_HEAD
);
1498 while (!list_empty(&flush_list
)) {
1499 rq
= list_entry(flush_list
.next
, struct request
, queuelist
);
1500 list_del_init(&rq
->queuelist
);
1501 blk_mq_insert_request(rq
, 0);
1504 blk_mq_run_hw_queues(q
, false);
1507 void blk_mq_kick_requeue_list(struct request_queue
*q
)
1509 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
, 0);
1511 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
1513 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
1514 unsigned long msecs
)
1516 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
,
1517 msecs_to_jiffies(msecs
));
1519 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
1521 static bool blk_mq_rq_inflight(struct request
*rq
, void *priv
)
1524 * If we find a request that isn't idle we know the queue is busy
1525 * as it's checked in the iter.
1526 * Return false to stop the iteration.
1528 if (blk_mq_request_started(rq
)) {
1538 bool blk_mq_queue_inflight(struct request_queue
*q
)
1542 blk_mq_queue_tag_busy_iter(q
, blk_mq_rq_inflight
, &busy
);
1545 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight
);
1547 static void blk_mq_rq_timed_out(struct request
*req
)
1549 req
->rq_flags
|= RQF_TIMED_OUT
;
1550 if (req
->q
->mq_ops
->timeout
) {
1551 enum blk_eh_timer_return ret
;
1553 ret
= req
->q
->mq_ops
->timeout(req
);
1554 if (ret
== BLK_EH_DONE
)
1556 WARN_ON_ONCE(ret
!= BLK_EH_RESET_TIMER
);
1562 struct blk_expired_data
{
1563 bool has_timedout_rq
;
1565 unsigned long timeout_start
;
1568 static bool blk_mq_req_expired(struct request
*rq
, struct blk_expired_data
*expired
)
1570 unsigned long deadline
;
1572 if (blk_mq_rq_state(rq
) != MQ_RQ_IN_FLIGHT
)
1574 if (rq
->rq_flags
& RQF_TIMED_OUT
)
1577 deadline
= READ_ONCE(rq
->deadline
);
1578 if (time_after_eq(expired
->timeout_start
, deadline
))
1581 if (expired
->next
== 0)
1582 expired
->next
= deadline
;
1583 else if (time_after(expired
->next
, deadline
))
1584 expired
->next
= deadline
;
1588 void blk_mq_put_rq_ref(struct request
*rq
)
1590 if (is_flush_rq(rq
)) {
1591 if (rq
->end_io(rq
, 0) == RQ_END_IO_FREE
)
1592 blk_mq_free_request(rq
);
1593 } else if (req_ref_put_and_test(rq
)) {
1594 __blk_mq_free_request(rq
);
1598 static bool blk_mq_check_expired(struct request
*rq
, void *priv
)
1600 struct blk_expired_data
*expired
= priv
;
1603 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1604 * be reallocated underneath the timeout handler's processing, then
1605 * the expire check is reliable. If the request is not expired, then
1606 * it was completed and reallocated as a new request after returning
1607 * from blk_mq_check_expired().
1609 if (blk_mq_req_expired(rq
, expired
)) {
1610 expired
->has_timedout_rq
= true;
1616 static bool blk_mq_handle_expired(struct request
*rq
, void *priv
)
1618 struct blk_expired_data
*expired
= priv
;
1620 if (blk_mq_req_expired(rq
, expired
))
1621 blk_mq_rq_timed_out(rq
);
1625 static void blk_mq_timeout_work(struct work_struct
*work
)
1627 struct request_queue
*q
=
1628 container_of(work
, struct request_queue
, timeout_work
);
1629 struct blk_expired_data expired
= {
1630 .timeout_start
= jiffies
,
1632 struct blk_mq_hw_ctx
*hctx
;
1635 /* A deadlock might occur if a request is stuck requiring a
1636 * timeout at the same time a queue freeze is waiting
1637 * completion, since the timeout code would not be able to
1638 * acquire the queue reference here.
1640 * That's why we don't use blk_queue_enter here; instead, we use
1641 * percpu_ref_tryget directly, because we need to be able to
1642 * obtain a reference even in the short window between the queue
1643 * starting to freeze, by dropping the first reference in
1644 * blk_freeze_queue_start, and the moment the last request is
1645 * consumed, marked by the instant q_usage_counter reaches
1648 if (!percpu_ref_tryget(&q
->q_usage_counter
))
1651 /* check if there is any timed-out request */
1652 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &expired
);
1653 if (expired
.has_timedout_rq
) {
1655 * Before walking tags, we must ensure any submit started
1656 * before the current time has finished. Since the submit
1657 * uses srcu or rcu, wait for a synchronization point to
1658 * ensure all running submits have finished
1660 blk_mq_wait_quiesce_done(q
->tag_set
);
1663 blk_mq_queue_tag_busy_iter(q
, blk_mq_handle_expired
, &expired
);
1666 if (expired
.next
!= 0) {
1667 mod_timer(&q
->timeout
, expired
.next
);
1670 * Request timeouts are handled as a forward rolling timer. If
1671 * we end up here it means that no requests are pending and
1672 * also that no request has been pending for a while. Mark
1673 * each hctx as idle.
1675 queue_for_each_hw_ctx(q
, hctx
, i
) {
1676 /* the hctx may be unmapped, so check it here */
1677 if (blk_mq_hw_queue_mapped(hctx
))
1678 blk_mq_tag_idle(hctx
);
1684 struct flush_busy_ctx_data
{
1685 struct blk_mq_hw_ctx
*hctx
;
1686 struct list_head
*list
;
1689 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
1691 struct flush_busy_ctx_data
*flush_data
= data
;
1692 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
1693 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
1694 enum hctx_type type
= hctx
->type
;
1696 spin_lock(&ctx
->lock
);
1697 list_splice_tail_init(&ctx
->rq_lists
[type
], flush_data
->list
);
1698 sbitmap_clear_bit(sb
, bitnr
);
1699 spin_unlock(&ctx
->lock
);
1704 * Process software queues that have been marked busy, splicing them
1705 * to the for-dispatch
1707 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
1709 struct flush_busy_ctx_data data
= {
1714 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
1716 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
1718 struct dispatch_rq_data
{
1719 struct blk_mq_hw_ctx
*hctx
;
1723 static bool dispatch_rq_from_ctx(struct sbitmap
*sb
, unsigned int bitnr
,
1726 struct dispatch_rq_data
*dispatch_data
= data
;
1727 struct blk_mq_hw_ctx
*hctx
= dispatch_data
->hctx
;
1728 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
1729 enum hctx_type type
= hctx
->type
;
1731 spin_lock(&ctx
->lock
);
1732 if (!list_empty(&ctx
->rq_lists
[type
])) {
1733 dispatch_data
->rq
= list_entry_rq(ctx
->rq_lists
[type
].next
);
1734 list_del_init(&dispatch_data
->rq
->queuelist
);
1735 if (list_empty(&ctx
->rq_lists
[type
]))
1736 sbitmap_clear_bit(sb
, bitnr
);
1738 spin_unlock(&ctx
->lock
);
1740 return !dispatch_data
->rq
;
1743 struct request
*blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx
*hctx
,
1744 struct blk_mq_ctx
*start
)
1746 unsigned off
= start
? start
->index_hw
[hctx
->type
] : 0;
1747 struct dispatch_rq_data data
= {
1752 __sbitmap_for_each_set(&hctx
->ctx_map
, off
,
1753 dispatch_rq_from_ctx
, &data
);
1758 bool __blk_mq_alloc_driver_tag(struct request
*rq
)
1760 struct sbitmap_queue
*bt
= &rq
->mq_hctx
->tags
->bitmap_tags
;
1761 unsigned int tag_offset
= rq
->mq_hctx
->tags
->nr_reserved_tags
;
1764 blk_mq_tag_busy(rq
->mq_hctx
);
1766 if (blk_mq_tag_is_reserved(rq
->mq_hctx
->sched_tags
, rq
->internal_tag
)) {
1767 bt
= &rq
->mq_hctx
->tags
->breserved_tags
;
1770 if (!hctx_may_queue(rq
->mq_hctx
, bt
))
1774 tag
= __sbitmap_queue_get(bt
);
1775 if (tag
== BLK_MQ_NO_TAG
)
1778 rq
->tag
= tag
+ tag_offset
;
1779 blk_mq_inc_active_requests(rq
->mq_hctx
);
1783 static int blk_mq_dispatch_wake(wait_queue_entry_t
*wait
, unsigned mode
,
1784 int flags
, void *key
)
1786 struct blk_mq_hw_ctx
*hctx
;
1788 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
1790 spin_lock(&hctx
->dispatch_wait_lock
);
1791 if (!list_empty(&wait
->entry
)) {
1792 struct sbitmap_queue
*sbq
;
1794 list_del_init(&wait
->entry
);
1795 sbq
= &hctx
->tags
->bitmap_tags
;
1796 atomic_dec(&sbq
->ws_active
);
1798 spin_unlock(&hctx
->dispatch_wait_lock
);
1800 blk_mq_run_hw_queue(hctx
, true);
1805 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1806 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1807 * restart. For both cases, take care to check the condition again after
1808 * marking us as waiting.
1810 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx
*hctx
,
1813 struct sbitmap_queue
*sbq
;
1814 struct wait_queue_head
*wq
;
1815 wait_queue_entry_t
*wait
;
1818 if (!(hctx
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
) &&
1819 !(blk_mq_is_shared_tags(hctx
->flags
))) {
1820 blk_mq_sched_mark_restart_hctx(hctx
);
1823 * It's possible that a tag was freed in the window between the
1824 * allocation failure and adding the hardware queue to the wait
1827 * Don't clear RESTART here, someone else could have set it.
1828 * At most this will cost an extra queue run.
1830 return blk_mq_get_driver_tag(rq
);
1833 wait
= &hctx
->dispatch_wait
;
1834 if (!list_empty_careful(&wait
->entry
))
1837 if (blk_mq_tag_is_reserved(rq
->mq_hctx
->sched_tags
, rq
->internal_tag
))
1838 sbq
= &hctx
->tags
->breserved_tags
;
1840 sbq
= &hctx
->tags
->bitmap_tags
;
1841 wq
= &bt_wait_ptr(sbq
, hctx
)->wait
;
1843 spin_lock_irq(&wq
->lock
);
1844 spin_lock(&hctx
->dispatch_wait_lock
);
1845 if (!list_empty(&wait
->entry
)) {
1846 spin_unlock(&hctx
->dispatch_wait_lock
);
1847 spin_unlock_irq(&wq
->lock
);
1851 atomic_inc(&sbq
->ws_active
);
1852 wait
->flags
&= ~WQ_FLAG_EXCLUSIVE
;
1853 __add_wait_queue(wq
, wait
);
1856 * It's possible that a tag was freed in the window between the
1857 * allocation failure and adding the hardware queue to the wait
1860 ret
= blk_mq_get_driver_tag(rq
);
1862 spin_unlock(&hctx
->dispatch_wait_lock
);
1863 spin_unlock_irq(&wq
->lock
);
1868 * We got a tag, remove ourselves from the wait queue to ensure
1869 * someone else gets the wakeup.
1871 list_del_init(&wait
->entry
);
1872 atomic_dec(&sbq
->ws_active
);
1873 spin_unlock(&hctx
->dispatch_wait_lock
);
1874 spin_unlock_irq(&wq
->lock
);
1879 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1880 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1882 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1883 * - EWMA is one simple way to compute running average value
1884 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1885 * - take 4 as factor for avoiding to get too small(0) result, and this
1886 * factor doesn't matter because EWMA decreases exponentially
1888 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx
*hctx
, bool busy
)
1892 ewma
= hctx
->dispatch_busy
;
1897 ewma
*= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
- 1;
1899 ewma
+= 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR
;
1900 ewma
/= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
;
1902 hctx
->dispatch_busy
= ewma
;
1905 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1907 static void blk_mq_handle_dev_resource(struct request
*rq
,
1908 struct list_head
*list
)
1910 list_add(&rq
->queuelist
, list
);
1911 __blk_mq_requeue_request(rq
);
1914 static void blk_mq_handle_zone_resource(struct request
*rq
,
1915 struct list_head
*zone_list
)
1918 * If we end up here it is because we cannot dispatch a request to a
1919 * specific zone due to LLD level zone-write locking or other zone
1920 * related resource not being available. In this case, set the request
1921 * aside in zone_list for retrying it later.
1923 list_add(&rq
->queuelist
, zone_list
);
1924 __blk_mq_requeue_request(rq
);
1927 enum prep_dispatch
{
1929 PREP_DISPATCH_NO_TAG
,
1930 PREP_DISPATCH_NO_BUDGET
,
1933 static enum prep_dispatch
blk_mq_prep_dispatch_rq(struct request
*rq
,
1936 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1937 int budget_token
= -1;
1940 budget_token
= blk_mq_get_dispatch_budget(rq
->q
);
1941 if (budget_token
< 0) {
1942 blk_mq_put_driver_tag(rq
);
1943 return PREP_DISPATCH_NO_BUDGET
;
1945 blk_mq_set_rq_budget_token(rq
, budget_token
);
1948 if (!blk_mq_get_driver_tag(rq
)) {
1950 * The initial allocation attempt failed, so we need to
1951 * rerun the hardware queue when a tag is freed. The
1952 * waitqueue takes care of that. If the queue is run
1953 * before we add this entry back on the dispatch list,
1954 * we'll re-run it below.
1956 if (!blk_mq_mark_tag_wait(hctx
, rq
)) {
1958 * All budgets not got from this function will be put
1959 * together during handling partial dispatch
1962 blk_mq_put_dispatch_budget(rq
->q
, budget_token
);
1963 return PREP_DISPATCH_NO_TAG
;
1967 return PREP_DISPATCH_OK
;
1970 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1971 static void blk_mq_release_budgets(struct request_queue
*q
,
1972 struct list_head
*list
)
1976 list_for_each_entry(rq
, list
, queuelist
) {
1977 int budget_token
= blk_mq_get_rq_budget_token(rq
);
1979 if (budget_token
>= 0)
1980 blk_mq_put_dispatch_budget(q
, budget_token
);
1985 * blk_mq_commit_rqs will notify driver using bd->last that there is no
1986 * more requests. (See comment in struct blk_mq_ops for commit_rqs for
1988 * Attention, we should explicitly call this in unusual cases:
1989 * 1) did not queue everything initially scheduled to queue
1990 * 2) the last attempt to queue a request failed
1992 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx
*hctx
, int queued
,
1995 if (hctx
->queue
->mq_ops
->commit_rqs
&& queued
) {
1996 trace_block_unplug(hctx
->queue
, queued
, !from_schedule
);
1997 hctx
->queue
->mq_ops
->commit_rqs(hctx
);
2002 * Returns true if we did some work AND can potentially do more.
2004 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
,
2005 unsigned int nr_budgets
)
2007 enum prep_dispatch prep
;
2008 struct request_queue
*q
= hctx
->queue
;
2011 blk_status_t ret
= BLK_STS_OK
;
2012 LIST_HEAD(zone_list
);
2013 bool needs_resource
= false;
2015 if (list_empty(list
))
2019 * Now process all the entries, sending them to the driver.
2023 struct blk_mq_queue_data bd
;
2025 rq
= list_first_entry(list
, struct request
, queuelist
);
2027 WARN_ON_ONCE(hctx
!= rq
->mq_hctx
);
2028 prep
= blk_mq_prep_dispatch_rq(rq
, !nr_budgets
);
2029 if (prep
!= PREP_DISPATCH_OK
)
2032 list_del_init(&rq
->queuelist
);
2035 bd
.last
= list_empty(list
);
2038 * once the request is queued to lld, no need to cover the
2043 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
2048 case BLK_STS_RESOURCE
:
2049 needs_resource
= true;
2051 case BLK_STS_DEV_RESOURCE
:
2052 blk_mq_handle_dev_resource(rq
, list
);
2054 case BLK_STS_ZONE_RESOURCE
:
2056 * Move the request to zone_list and keep going through
2057 * the dispatch list to find more requests the drive can
2060 blk_mq_handle_zone_resource(rq
, &zone_list
);
2061 needs_resource
= true;
2064 blk_mq_end_request(rq
, ret
);
2066 } while (!list_empty(list
));
2068 if (!list_empty(&zone_list
))
2069 list_splice_tail_init(&zone_list
, list
);
2071 /* If we didn't flush the entire list, we could have told the driver
2072 * there was more coming, but that turned out to be a lie.
2074 if (!list_empty(list
) || ret
!= BLK_STS_OK
)
2075 blk_mq_commit_rqs(hctx
, queued
, false);
2078 * Any items that need requeuing? Stuff them into hctx->dispatch,
2079 * that is where we will continue on next queue run.
2081 if (!list_empty(list
)) {
2083 /* For non-shared tags, the RESTART check will suffice */
2084 bool no_tag
= prep
== PREP_DISPATCH_NO_TAG
&&
2085 ((hctx
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
) ||
2086 blk_mq_is_shared_tags(hctx
->flags
));
2089 blk_mq_release_budgets(q
, list
);
2091 spin_lock(&hctx
->lock
);
2092 list_splice_tail_init(list
, &hctx
->dispatch
);
2093 spin_unlock(&hctx
->lock
);
2096 * Order adding requests to hctx->dispatch and checking
2097 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2098 * in blk_mq_sched_restart(). Avoid restart code path to
2099 * miss the new added requests to hctx->dispatch, meantime
2100 * SCHED_RESTART is observed here.
2105 * If SCHED_RESTART was set by the caller of this function and
2106 * it is no longer set that means that it was cleared by another
2107 * thread and hence that a queue rerun is needed.
2109 * If 'no_tag' is set, that means that we failed getting
2110 * a driver tag with an I/O scheduler attached. If our dispatch
2111 * waitqueue is no longer active, ensure that we run the queue
2112 * AFTER adding our entries back to the list.
2114 * If no I/O scheduler has been configured it is possible that
2115 * the hardware queue got stopped and restarted before requests
2116 * were pushed back onto the dispatch list. Rerun the queue to
2117 * avoid starvation. Notes:
2118 * - blk_mq_run_hw_queue() checks whether or not a queue has
2119 * been stopped before rerunning a queue.
2120 * - Some but not all block drivers stop a queue before
2121 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2124 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2125 * bit is set, run queue after a delay to avoid IO stalls
2126 * that could otherwise occur if the queue is idle. We'll do
2127 * similar if we couldn't get budget or couldn't lock a zone
2128 * and SCHED_RESTART is set.
2130 needs_restart
= blk_mq_sched_needs_restart(hctx
);
2131 if (prep
== PREP_DISPATCH_NO_BUDGET
)
2132 needs_resource
= true;
2133 if (!needs_restart
||
2134 (no_tag
&& list_empty_careful(&hctx
->dispatch_wait
.entry
)))
2135 blk_mq_run_hw_queue(hctx
, true);
2136 else if (needs_resource
)
2137 blk_mq_delay_run_hw_queue(hctx
, BLK_MQ_RESOURCE_DELAY
);
2139 blk_mq_update_dispatch_busy(hctx
, true);
2143 blk_mq_update_dispatch_busy(hctx
, false);
2147 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx
*hctx
)
2149 int cpu
= cpumask_first_and(hctx
->cpumask
, cpu_online_mask
);
2151 if (cpu
>= nr_cpu_ids
)
2152 cpu
= cpumask_first(hctx
->cpumask
);
2157 * It'd be great if the workqueue API had a way to pass
2158 * in a mask and had some smarts for more clever placement.
2159 * For now we just round-robin here, switching for every
2160 * BLK_MQ_CPU_WORK_BATCH queued items.
2162 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
2165 int next_cpu
= hctx
->next_cpu
;
2167 if (hctx
->queue
->nr_hw_queues
== 1)
2168 return WORK_CPU_UNBOUND
;
2170 if (--hctx
->next_cpu_batch
<= 0) {
2172 next_cpu
= cpumask_next_and(next_cpu
, hctx
->cpumask
,
2174 if (next_cpu
>= nr_cpu_ids
)
2175 next_cpu
= blk_mq_first_mapped_cpu(hctx
);
2176 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2180 * Do unbound schedule if we can't find a online CPU for this hctx,
2181 * and it should only happen in the path of handling CPU DEAD.
2183 if (!cpu_online(next_cpu
)) {
2190 * Make sure to re-select CPU next time once after CPUs
2191 * in hctx->cpumask become online again.
2193 hctx
->next_cpu
= next_cpu
;
2194 hctx
->next_cpu_batch
= 1;
2195 return WORK_CPU_UNBOUND
;
2198 hctx
->next_cpu
= next_cpu
;
2203 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2204 * @hctx: Pointer to the hardware queue to run.
2205 * @msecs: Milliseconds of delay to wait before running the queue.
2207 * Run a hardware queue asynchronously with a delay of @msecs.
2209 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
2211 if (unlikely(blk_mq_hctx_stopped(hctx
)))
2213 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
,
2214 msecs_to_jiffies(msecs
));
2216 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue
);
2219 * blk_mq_run_hw_queue - Start to run a hardware queue.
2220 * @hctx: Pointer to the hardware queue to run.
2221 * @async: If we want to run the queue asynchronously.
2223 * Check if the request queue is not in a quiesced state and if there are
2224 * pending requests to be sent. If this is true, run the queue to send requests
2227 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
2232 * We can't run the queue inline with interrupts disabled.
2234 WARN_ON_ONCE(!async
&& in_interrupt());
2236 might_sleep_if(!async
&& hctx
->flags
& BLK_MQ_F_BLOCKING
);
2239 * When queue is quiesced, we may be switching io scheduler, or
2240 * updating nr_hw_queues, or other things, and we can't run queue
2241 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2243 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2246 __blk_mq_run_dispatch_ops(hctx
->queue
, false,
2247 need_run
= !blk_queue_quiesced(hctx
->queue
) &&
2248 blk_mq_hctx_has_pending(hctx
));
2253 if (async
|| !cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
)) {
2254 blk_mq_delay_run_hw_queue(hctx
, 0);
2258 blk_mq_run_dispatch_ops(hctx
->queue
,
2259 blk_mq_sched_dispatch_requests(hctx
));
2261 EXPORT_SYMBOL(blk_mq_run_hw_queue
);
2264 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2267 static struct blk_mq_hw_ctx
*blk_mq_get_sq_hctx(struct request_queue
*q
)
2269 struct blk_mq_ctx
*ctx
= blk_mq_get_ctx(q
);
2271 * If the IO scheduler does not respect hardware queues when
2272 * dispatching, we just don't bother with multiple HW queues and
2273 * dispatch from hctx for the current CPU since running multiple queues
2274 * just causes lock contention inside the scheduler and pointless cache
2277 struct blk_mq_hw_ctx
*hctx
= ctx
->hctxs
[HCTX_TYPE_DEFAULT
];
2279 if (!blk_mq_hctx_stopped(hctx
))
2285 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2286 * @q: Pointer to the request queue to run.
2287 * @async: If we want to run the queue asynchronously.
2289 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
2291 struct blk_mq_hw_ctx
*hctx
, *sq_hctx
;
2295 if (blk_queue_sq_sched(q
))
2296 sq_hctx
= blk_mq_get_sq_hctx(q
);
2297 queue_for_each_hw_ctx(q
, hctx
, i
) {
2298 if (blk_mq_hctx_stopped(hctx
))
2301 * Dispatch from this hctx either if there's no hctx preferred
2302 * by IO scheduler or if it has requests that bypass the
2305 if (!sq_hctx
|| sq_hctx
== hctx
||
2306 !list_empty_careful(&hctx
->dispatch
))
2307 blk_mq_run_hw_queue(hctx
, async
);
2310 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
2313 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2314 * @q: Pointer to the request queue to run.
2315 * @msecs: Milliseconds of delay to wait before running the queues.
2317 void blk_mq_delay_run_hw_queues(struct request_queue
*q
, unsigned long msecs
)
2319 struct blk_mq_hw_ctx
*hctx
, *sq_hctx
;
2323 if (blk_queue_sq_sched(q
))
2324 sq_hctx
= blk_mq_get_sq_hctx(q
);
2325 queue_for_each_hw_ctx(q
, hctx
, i
) {
2326 if (blk_mq_hctx_stopped(hctx
))
2329 * If there is already a run_work pending, leave the
2330 * pending delay untouched. Otherwise, a hctx can stall
2331 * if another hctx is re-delaying the other's work
2332 * before the work executes.
2334 if (delayed_work_pending(&hctx
->run_work
))
2337 * Dispatch from this hctx either if there's no hctx preferred
2338 * by IO scheduler or if it has requests that bypass the
2341 if (!sq_hctx
|| sq_hctx
== hctx
||
2342 !list_empty_careful(&hctx
->dispatch
))
2343 blk_mq_delay_run_hw_queue(hctx
, msecs
);
2346 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues
);
2349 * This function is often used for pausing .queue_rq() by driver when
2350 * there isn't enough resource or some conditions aren't satisfied, and
2351 * BLK_STS_RESOURCE is usually returned.
2353 * We do not guarantee that dispatch can be drained or blocked
2354 * after blk_mq_stop_hw_queue() returns. Please use
2355 * blk_mq_quiesce_queue() for that requirement.
2357 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
2359 cancel_delayed_work(&hctx
->run_work
);
2361 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
2363 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
2366 * This function is often used for pausing .queue_rq() by driver when
2367 * there isn't enough resource or some conditions aren't satisfied, and
2368 * BLK_STS_RESOURCE is usually returned.
2370 * We do not guarantee that dispatch can be drained or blocked
2371 * after blk_mq_stop_hw_queues() returns. Please use
2372 * blk_mq_quiesce_queue() for that requirement.
2374 void blk_mq_stop_hw_queues(struct request_queue
*q
)
2376 struct blk_mq_hw_ctx
*hctx
;
2379 queue_for_each_hw_ctx(q
, hctx
, i
)
2380 blk_mq_stop_hw_queue(hctx
);
2382 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
2384 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
2386 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
2388 blk_mq_run_hw_queue(hctx
, hctx
->flags
& BLK_MQ_F_BLOCKING
);
2390 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
2392 void blk_mq_start_hw_queues(struct request_queue
*q
)
2394 struct blk_mq_hw_ctx
*hctx
;
2397 queue_for_each_hw_ctx(q
, hctx
, i
)
2398 blk_mq_start_hw_queue(hctx
);
2400 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
2402 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
2404 if (!blk_mq_hctx_stopped(hctx
))
2407 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
2408 blk_mq_run_hw_queue(hctx
, async
);
2410 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
2412 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
2414 struct blk_mq_hw_ctx
*hctx
;
2417 queue_for_each_hw_ctx(q
, hctx
, i
)
2418 blk_mq_start_stopped_hw_queue(hctx
, async
||
2419 (hctx
->flags
& BLK_MQ_F_BLOCKING
));
2421 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
2423 static void blk_mq_run_work_fn(struct work_struct
*work
)
2425 struct blk_mq_hw_ctx
*hctx
=
2426 container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
2428 blk_mq_run_dispatch_ops(hctx
->queue
,
2429 blk_mq_sched_dispatch_requests(hctx
));
2433 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2434 * @rq: Pointer to request to be inserted.
2435 * @flags: BLK_MQ_INSERT_*
2437 * Should only be used carefully, when the caller knows we want to
2438 * bypass a potential IO scheduler on the target device.
2440 static void blk_mq_request_bypass_insert(struct request
*rq
, blk_insert_t flags
)
2442 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
2444 spin_lock(&hctx
->lock
);
2445 if (flags
& BLK_MQ_INSERT_AT_HEAD
)
2446 list_add(&rq
->queuelist
, &hctx
->dispatch
);
2448 list_add_tail(&rq
->queuelist
, &hctx
->dispatch
);
2449 spin_unlock(&hctx
->lock
);
2452 static void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
,
2453 struct blk_mq_ctx
*ctx
, struct list_head
*list
,
2454 bool run_queue_async
)
2457 enum hctx_type type
= hctx
->type
;
2460 * Try to issue requests directly if the hw queue isn't busy to save an
2461 * extra enqueue & dequeue to the sw queue.
2463 if (!hctx
->dispatch_busy
&& !run_queue_async
) {
2464 blk_mq_run_dispatch_ops(hctx
->queue
,
2465 blk_mq_try_issue_list_directly(hctx
, list
));
2466 if (list_empty(list
))
2471 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2474 list_for_each_entry(rq
, list
, queuelist
) {
2475 BUG_ON(rq
->mq_ctx
!= ctx
);
2476 trace_block_rq_insert(rq
);
2477 if (rq
->cmd_flags
& REQ_NOWAIT
)
2478 run_queue_async
= true;
2481 spin_lock(&ctx
->lock
);
2482 list_splice_tail_init(list
, &ctx
->rq_lists
[type
]);
2483 blk_mq_hctx_mark_pending(hctx
, ctx
);
2484 spin_unlock(&ctx
->lock
);
2486 blk_mq_run_hw_queue(hctx
, run_queue_async
);
2489 static void blk_mq_insert_request(struct request
*rq
, blk_insert_t flags
)
2491 struct request_queue
*q
= rq
->q
;
2492 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
2493 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
2495 if (blk_rq_is_passthrough(rq
)) {
2497 * Passthrough request have to be added to hctx->dispatch
2498 * directly. The device may be in a situation where it can't
2499 * handle FS request, and always returns BLK_STS_RESOURCE for
2500 * them, which gets them added to hctx->dispatch.
2502 * If a passthrough request is required to unblock the queues,
2503 * and it is added to the scheduler queue, there is no chance to
2504 * dispatch it given we prioritize requests in hctx->dispatch.
2506 blk_mq_request_bypass_insert(rq
, flags
);
2507 } else if (req_op(rq
) == REQ_OP_FLUSH
) {
2509 * Firstly normal IO request is inserted to scheduler queue or
2510 * sw queue, meantime we add flush request to dispatch queue(
2511 * hctx->dispatch) directly and there is at most one in-flight
2512 * flush request for each hw queue, so it doesn't matter to add
2513 * flush request to tail or front of the dispatch queue.
2515 * Secondly in case of NCQ, flush request belongs to non-NCQ
2516 * command, and queueing it will fail when there is any
2517 * in-flight normal IO request(NCQ command). When adding flush
2518 * rq to the front of hctx->dispatch, it is easier to introduce
2519 * extra time to flush rq's latency because of S_SCHED_RESTART
2520 * compared with adding to the tail of dispatch queue, then
2521 * chance of flush merge is increased, and less flush requests
2522 * will be issued to controller. It is observed that ~10% time
2523 * is saved in blktests block/004 on disk attached to AHCI/NCQ
2524 * drive when adding flush rq to the front of hctx->dispatch.
2526 * Simply queue flush rq to the front of hctx->dispatch so that
2527 * intensive flush workloads can benefit in case of NCQ HW.
2529 blk_mq_request_bypass_insert(rq
, BLK_MQ_INSERT_AT_HEAD
);
2530 } else if (q
->elevator
) {
2533 WARN_ON_ONCE(rq
->tag
!= BLK_MQ_NO_TAG
);
2535 list_add(&rq
->queuelist
, &list
);
2536 q
->elevator
->type
->ops
.insert_requests(hctx
, &list
, flags
);
2538 trace_block_rq_insert(rq
);
2540 spin_lock(&ctx
->lock
);
2541 if (flags
& BLK_MQ_INSERT_AT_HEAD
)
2542 list_add(&rq
->queuelist
, &ctx
->rq_lists
[hctx
->type
]);
2544 list_add_tail(&rq
->queuelist
,
2545 &ctx
->rq_lists
[hctx
->type
]);
2546 blk_mq_hctx_mark_pending(hctx
, ctx
);
2547 spin_unlock(&ctx
->lock
);
2551 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
,
2552 unsigned int nr_segs
)
2556 if (bio
->bi_opf
& REQ_RAHEAD
)
2557 rq
->cmd_flags
|= REQ_FAILFAST_MASK
;
2559 rq
->__sector
= bio
->bi_iter
.bi_sector
;
2560 blk_rq_bio_prep(rq
, bio
, nr_segs
);
2562 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2563 err
= blk_crypto_rq_bio_prep(rq
, bio
, GFP_NOIO
);
2566 blk_account_io_start(rq
);
2569 static blk_status_t
__blk_mq_issue_directly(struct blk_mq_hw_ctx
*hctx
,
2570 struct request
*rq
, bool last
)
2572 struct request_queue
*q
= rq
->q
;
2573 struct blk_mq_queue_data bd
= {
2580 * For OK queue, we are done. For error, caller may kill it.
2581 * Any other error (busy), just add it to our list as we
2582 * previously would have done.
2584 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
2587 blk_mq_update_dispatch_busy(hctx
, false);
2589 case BLK_STS_RESOURCE
:
2590 case BLK_STS_DEV_RESOURCE
:
2591 blk_mq_update_dispatch_busy(hctx
, true);
2592 __blk_mq_requeue_request(rq
);
2595 blk_mq_update_dispatch_busy(hctx
, false);
2602 static bool blk_mq_get_budget_and_tag(struct request
*rq
)
2606 budget_token
= blk_mq_get_dispatch_budget(rq
->q
);
2607 if (budget_token
< 0)
2609 blk_mq_set_rq_budget_token(rq
, budget_token
);
2610 if (!blk_mq_get_driver_tag(rq
)) {
2611 blk_mq_put_dispatch_budget(rq
->q
, budget_token
);
2618 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2619 * @hctx: Pointer of the associated hardware queue.
2620 * @rq: Pointer to request to be sent.
2622 * If the device has enough resources to accept a new request now, send the
2623 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2624 * we can try send it another time in the future. Requests inserted at this
2625 * queue have higher priority.
2627 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
2632 if (blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(rq
->q
)) {
2633 blk_mq_insert_request(rq
, 0);
2637 if ((rq
->rq_flags
& RQF_USE_SCHED
) || !blk_mq_get_budget_and_tag(rq
)) {
2638 blk_mq_insert_request(rq
, 0);
2639 blk_mq_run_hw_queue(hctx
, rq
->cmd_flags
& REQ_NOWAIT
);
2643 ret
= __blk_mq_issue_directly(hctx
, rq
, true);
2647 case BLK_STS_RESOURCE
:
2648 case BLK_STS_DEV_RESOURCE
:
2649 blk_mq_request_bypass_insert(rq
, 0);
2650 blk_mq_run_hw_queue(hctx
, false);
2653 blk_mq_end_request(rq
, ret
);
2658 static blk_status_t
blk_mq_request_issue_directly(struct request
*rq
, bool last
)
2660 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
2662 if (blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(rq
->q
)) {
2663 blk_mq_insert_request(rq
, 0);
2667 if (!blk_mq_get_budget_and_tag(rq
))
2668 return BLK_STS_RESOURCE
;
2669 return __blk_mq_issue_directly(hctx
, rq
, last
);
2672 static void blk_mq_plug_issue_direct(struct blk_plug
*plug
)
2674 struct blk_mq_hw_ctx
*hctx
= NULL
;
2677 blk_status_t ret
= BLK_STS_OK
;
2679 while ((rq
= rq_list_pop(&plug
->mq_list
))) {
2680 bool last
= rq_list_empty(plug
->mq_list
);
2682 if (hctx
!= rq
->mq_hctx
) {
2684 blk_mq_commit_rqs(hctx
, queued
, false);
2690 ret
= blk_mq_request_issue_directly(rq
, last
);
2695 case BLK_STS_RESOURCE
:
2696 case BLK_STS_DEV_RESOURCE
:
2697 blk_mq_request_bypass_insert(rq
, 0);
2698 blk_mq_run_hw_queue(hctx
, false);
2701 blk_mq_end_request(rq
, ret
);
2707 if (ret
!= BLK_STS_OK
)
2708 blk_mq_commit_rqs(hctx
, queued
, false);
2711 static void __blk_mq_flush_plug_list(struct request_queue
*q
,
2712 struct blk_plug
*plug
)
2714 if (blk_queue_quiesced(q
))
2716 q
->mq_ops
->queue_rqs(&plug
->mq_list
);
2719 static void blk_mq_dispatch_plug_list(struct blk_plug
*plug
, bool from_sched
)
2721 struct blk_mq_hw_ctx
*this_hctx
= NULL
;
2722 struct blk_mq_ctx
*this_ctx
= NULL
;
2723 struct request
*requeue_list
= NULL
;
2724 struct request
**requeue_lastp
= &requeue_list
;
2725 unsigned int depth
= 0;
2726 bool is_passthrough
= false;
2730 struct request
*rq
= rq_list_pop(&plug
->mq_list
);
2733 this_hctx
= rq
->mq_hctx
;
2734 this_ctx
= rq
->mq_ctx
;
2735 is_passthrough
= blk_rq_is_passthrough(rq
);
2736 } else if (this_hctx
!= rq
->mq_hctx
|| this_ctx
!= rq
->mq_ctx
||
2737 is_passthrough
!= blk_rq_is_passthrough(rq
)) {
2738 rq_list_add_tail(&requeue_lastp
, rq
);
2741 list_add(&rq
->queuelist
, &list
);
2743 } while (!rq_list_empty(plug
->mq_list
));
2745 plug
->mq_list
= requeue_list
;
2746 trace_block_unplug(this_hctx
->queue
, depth
, !from_sched
);
2748 percpu_ref_get(&this_hctx
->queue
->q_usage_counter
);
2749 /* passthrough requests should never be issued to the I/O scheduler */
2750 if (is_passthrough
) {
2751 spin_lock(&this_hctx
->lock
);
2752 list_splice_tail_init(&list
, &this_hctx
->dispatch
);
2753 spin_unlock(&this_hctx
->lock
);
2754 blk_mq_run_hw_queue(this_hctx
, from_sched
);
2755 } else if (this_hctx
->queue
->elevator
) {
2756 this_hctx
->queue
->elevator
->type
->ops
.insert_requests(this_hctx
,
2758 blk_mq_run_hw_queue(this_hctx
, from_sched
);
2760 blk_mq_insert_requests(this_hctx
, this_ctx
, &list
, from_sched
);
2762 percpu_ref_put(&this_hctx
->queue
->q_usage_counter
);
2765 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
2770 * We may have been called recursively midway through handling
2771 * plug->mq_list via a schedule() in the driver's queue_rq() callback.
2772 * To avoid mq_list changing under our feet, clear rq_count early and
2773 * bail out specifically if rq_count is 0 rather than checking
2774 * whether the mq_list is empty.
2776 if (plug
->rq_count
== 0)
2780 if (!plug
->multiple_queues
&& !plug
->has_elevator
&& !from_schedule
) {
2781 struct request_queue
*q
;
2783 rq
= rq_list_peek(&plug
->mq_list
);
2787 * Peek first request and see if we have a ->queue_rqs() hook.
2788 * If we do, we can dispatch the whole plug list in one go. We
2789 * already know at this point that all requests belong to the
2790 * same queue, caller must ensure that's the case.
2792 if (q
->mq_ops
->queue_rqs
) {
2793 blk_mq_run_dispatch_ops(q
,
2794 __blk_mq_flush_plug_list(q
, plug
));
2795 if (rq_list_empty(plug
->mq_list
))
2799 blk_mq_run_dispatch_ops(q
,
2800 blk_mq_plug_issue_direct(plug
));
2801 if (rq_list_empty(plug
->mq_list
))
2806 blk_mq_dispatch_plug_list(plug
, from_schedule
);
2807 } while (!rq_list_empty(plug
->mq_list
));
2810 static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx
*hctx
,
2811 struct list_head
*list
)
2814 blk_status_t ret
= BLK_STS_OK
;
2816 while (!list_empty(list
)) {
2817 struct request
*rq
= list_first_entry(list
, struct request
,
2820 list_del_init(&rq
->queuelist
);
2821 ret
= blk_mq_request_issue_directly(rq
, list_empty(list
));
2826 case BLK_STS_RESOURCE
:
2827 case BLK_STS_DEV_RESOURCE
:
2828 blk_mq_request_bypass_insert(rq
, 0);
2829 if (list_empty(list
))
2830 blk_mq_run_hw_queue(hctx
, false);
2833 blk_mq_end_request(rq
, ret
);
2839 if (ret
!= BLK_STS_OK
)
2840 blk_mq_commit_rqs(hctx
, queued
, false);
2843 static bool blk_mq_attempt_bio_merge(struct request_queue
*q
,
2844 struct bio
*bio
, unsigned int nr_segs
)
2846 if (!blk_queue_nomerges(q
) && bio_mergeable(bio
)) {
2847 if (blk_attempt_plug_merge(q
, bio
, nr_segs
))
2849 if (blk_mq_sched_bio_merge(q
, bio
, nr_segs
))
2855 static struct request
*blk_mq_get_new_requests(struct request_queue
*q
,
2856 struct blk_plug
*plug
,
2860 struct blk_mq_alloc_data data
= {
2863 .cmd_flags
= bio
->bi_opf
,
2867 if (blk_mq_attempt_bio_merge(q
, bio
, nsegs
))
2870 rq_qos_throttle(q
, bio
);
2873 data
.nr_tags
= plug
->nr_ios
;
2875 data
.cached_rq
= &plug
->cached_rq
;
2878 rq
= __blk_mq_alloc_requests(&data
);
2881 rq_qos_cleanup(q
, bio
);
2882 if (bio
->bi_opf
& REQ_NOWAIT
)
2883 bio_wouldblock_error(bio
);
2887 /* return true if this @rq can be used for @bio */
2888 static bool blk_mq_can_use_cached_rq(struct request
*rq
, struct blk_plug
*plug
,
2891 enum hctx_type type
= blk_mq_get_hctx_type(bio
->bi_opf
);
2892 enum hctx_type hctx_type
= rq
->mq_hctx
->type
;
2894 WARN_ON_ONCE(rq_list_peek(&plug
->cached_rq
) != rq
);
2896 if (type
!= hctx_type
&&
2897 !(type
== HCTX_TYPE_READ
&& hctx_type
== HCTX_TYPE_DEFAULT
))
2899 if (op_is_flush(rq
->cmd_flags
) != op_is_flush(bio
->bi_opf
))
2903 * If any qos ->throttle() end up blocking, we will have flushed the
2904 * plug and hence killed the cached_rq list as well. Pop this entry
2905 * before we throttle.
2907 plug
->cached_rq
= rq_list_next(rq
);
2908 rq_qos_throttle(rq
->q
, bio
);
2910 blk_mq_rq_time_init(rq
, 0);
2911 rq
->cmd_flags
= bio
->bi_opf
;
2912 INIT_LIST_HEAD(&rq
->queuelist
);
2916 static void bio_set_ioprio(struct bio
*bio
)
2918 /* Nobody set ioprio so far? Initialize it based on task's nice value */
2919 if (IOPRIO_PRIO_CLASS(bio
->bi_ioprio
) == IOPRIO_CLASS_NONE
)
2920 bio
->bi_ioprio
= get_current_ioprio();
2921 blkcg_set_ioprio(bio
);
2925 * blk_mq_submit_bio - Create and send a request to block device.
2926 * @bio: Bio pointer.
2928 * Builds up a request structure from @q and @bio and send to the device. The
2929 * request may not be queued directly to hardware if:
2930 * * This request can be merged with another one
2931 * * We want to place request at plug queue for possible future merging
2932 * * There is an IO scheduler active at this queue
2934 * It will not queue the request if there is an error with the bio, or at the
2937 void blk_mq_submit_bio(struct bio
*bio
)
2939 struct request_queue
*q
= bdev_get_queue(bio
->bi_bdev
);
2940 struct blk_plug
*plug
= blk_mq_plug(bio
);
2941 const int is_sync
= op_is_sync(bio
->bi_opf
);
2942 struct blk_mq_hw_ctx
*hctx
;
2943 struct request
*rq
= NULL
;
2944 unsigned int nr_segs
= 1;
2947 bio
= blk_queue_bounce(bio
, q
);
2948 if (bio_may_exceed_limits(bio
, &q
->limits
)) {
2949 bio
= __bio_split_to_limits(bio
, &q
->limits
, &nr_segs
);
2954 bio_set_ioprio(bio
);
2957 rq
= rq_list_peek(&plug
->cached_rq
);
2958 if (rq
&& rq
->q
!= q
)
2962 if (!bio_integrity_prep(bio
))
2964 if (blk_mq_attempt_bio_merge(q
, bio
, nr_segs
))
2966 if (blk_mq_can_use_cached_rq(rq
, plug
, bio
))
2968 percpu_ref_get(&q
->q_usage_counter
);
2970 if (unlikely(bio_queue_enter(bio
)))
2972 if (!bio_integrity_prep(bio
))
2976 rq
= blk_mq_get_new_requests(q
, plug
, bio
, nr_segs
);
2977 if (unlikely(!rq
)) {
2984 trace_block_getrq(bio
);
2986 rq_qos_track(q
, rq
, bio
);
2988 blk_mq_bio_to_request(rq
, bio
, nr_segs
);
2990 ret
= blk_crypto_rq_get_keyslot(rq
);
2991 if (ret
!= BLK_STS_OK
) {
2992 bio
->bi_status
= ret
;
2994 blk_mq_free_request(rq
);
2998 if (op_is_flush(bio
->bi_opf
) && blk_insert_flush(rq
))
3002 blk_add_rq_to_plug(plug
, rq
);
3007 if ((rq
->rq_flags
& RQF_USE_SCHED
) ||
3008 (hctx
->dispatch_busy
&& (q
->nr_hw_queues
== 1 || !is_sync
))) {
3009 blk_mq_insert_request(rq
, 0);
3010 blk_mq_run_hw_queue(hctx
, true);
3012 blk_mq_run_dispatch_ops(q
, blk_mq_try_issue_directly(hctx
, rq
));
3016 #ifdef CONFIG_BLK_MQ_STACKING
3018 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
3019 * @rq: the request being queued
3021 blk_status_t
blk_insert_cloned_request(struct request
*rq
)
3023 struct request_queue
*q
= rq
->q
;
3024 unsigned int max_sectors
= blk_queue_get_max_sectors(q
, req_op(rq
));
3025 unsigned int max_segments
= blk_rq_get_max_segments(rq
);
3028 if (blk_rq_sectors(rq
) > max_sectors
) {
3030 * SCSI device does not have a good way to return if
3031 * Write Same/Zero is actually supported. If a device rejects
3032 * a non-read/write command (discard, write same,etc.) the
3033 * low-level device driver will set the relevant queue limit to
3034 * 0 to prevent blk-lib from issuing more of the offending
3035 * operations. Commands queued prior to the queue limit being
3036 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
3037 * errors being propagated to upper layers.
3039 if (max_sectors
== 0)
3040 return BLK_STS_NOTSUPP
;
3042 printk(KERN_ERR
"%s: over max size limit. (%u > %u)\n",
3043 __func__
, blk_rq_sectors(rq
), max_sectors
);
3044 return BLK_STS_IOERR
;
3048 * The queue settings related to segment counting may differ from the
3051 rq
->nr_phys_segments
= blk_recalc_rq_segments(rq
);
3052 if (rq
->nr_phys_segments
> max_segments
) {
3053 printk(KERN_ERR
"%s: over max segments limit. (%u > %u)\n",
3054 __func__
, rq
->nr_phys_segments
, max_segments
);
3055 return BLK_STS_IOERR
;
3058 if (q
->disk
&& should_fail_request(q
->disk
->part0
, blk_rq_bytes(rq
)))
3059 return BLK_STS_IOERR
;
3061 ret
= blk_crypto_rq_get_keyslot(rq
);
3062 if (ret
!= BLK_STS_OK
)
3065 blk_account_io_start(rq
);
3068 * Since we have a scheduler attached on the top device,
3069 * bypass a potential scheduler on the bottom device for
3072 blk_mq_run_dispatch_ops(q
,
3073 ret
= blk_mq_request_issue_directly(rq
, true));
3075 blk_account_io_done(rq
, ktime_get_ns());
3078 EXPORT_SYMBOL_GPL(blk_insert_cloned_request
);
3081 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3082 * @rq: the clone request to be cleaned up
3085 * Free all bios in @rq for a cloned request.
3087 void blk_rq_unprep_clone(struct request
*rq
)
3091 while ((bio
= rq
->bio
) != NULL
) {
3092 rq
->bio
= bio
->bi_next
;
3097 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone
);
3100 * blk_rq_prep_clone - Helper function to setup clone request
3101 * @rq: the request to be setup
3102 * @rq_src: original request to be cloned
3103 * @bs: bio_set that bios for clone are allocated from
3104 * @gfp_mask: memory allocation mask for bio
3105 * @bio_ctr: setup function to be called for each clone bio.
3106 * Returns %0 for success, non %0 for failure.
3107 * @data: private data to be passed to @bio_ctr
3110 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3111 * Also, pages which the original bios are pointing to are not copied
3112 * and the cloned bios just point same pages.
3113 * So cloned bios must be completed before original bios, which means
3114 * the caller must complete @rq before @rq_src.
3116 int blk_rq_prep_clone(struct request
*rq
, struct request
*rq_src
,
3117 struct bio_set
*bs
, gfp_t gfp_mask
,
3118 int (*bio_ctr
)(struct bio
*, struct bio
*, void *),
3121 struct bio
*bio
, *bio_src
;
3126 __rq_for_each_bio(bio_src
, rq_src
) {
3127 bio
= bio_alloc_clone(rq
->q
->disk
->part0
, bio_src
, gfp_mask
,
3132 if (bio_ctr
&& bio_ctr(bio
, bio_src
, data
))
3136 rq
->biotail
->bi_next
= bio
;
3139 rq
->bio
= rq
->biotail
= bio
;
3144 /* Copy attributes of the original request to the clone request. */
3145 rq
->__sector
= blk_rq_pos(rq_src
);
3146 rq
->__data_len
= blk_rq_bytes(rq_src
);
3147 if (rq_src
->rq_flags
& RQF_SPECIAL_PAYLOAD
) {
3148 rq
->rq_flags
|= RQF_SPECIAL_PAYLOAD
;
3149 rq
->special_vec
= rq_src
->special_vec
;
3151 rq
->nr_phys_segments
= rq_src
->nr_phys_segments
;
3152 rq
->ioprio
= rq_src
->ioprio
;
3154 if (rq
->bio
&& blk_crypto_rq_bio_prep(rq
, rq
->bio
, gfp_mask
) < 0)
3162 blk_rq_unprep_clone(rq
);
3166 EXPORT_SYMBOL_GPL(blk_rq_prep_clone
);
3167 #endif /* CONFIG_BLK_MQ_STACKING */
3170 * Steal bios from a request and add them to a bio list.
3171 * The request must not have been partially completed before.
3173 void blk_steal_bios(struct bio_list
*list
, struct request
*rq
)
3177 list
->tail
->bi_next
= rq
->bio
;
3179 list
->head
= rq
->bio
;
3180 list
->tail
= rq
->biotail
;
3188 EXPORT_SYMBOL_GPL(blk_steal_bios
);
3190 static size_t order_to_size(unsigned int order
)
3192 return (size_t)PAGE_SIZE
<< order
;
3195 /* called before freeing request pool in @tags */
3196 static void blk_mq_clear_rq_mapping(struct blk_mq_tags
*drv_tags
,
3197 struct blk_mq_tags
*tags
)
3200 unsigned long flags
;
3203 * There is no need to clear mapping if driver tags is not initialized
3204 * or the mapping belongs to the driver tags.
3206 if (!drv_tags
|| drv_tags
== tags
)
3209 list_for_each_entry(page
, &tags
->page_list
, lru
) {
3210 unsigned long start
= (unsigned long)page_address(page
);
3211 unsigned long end
= start
+ order_to_size(page
->private);
3214 for (i
= 0; i
< drv_tags
->nr_tags
; i
++) {
3215 struct request
*rq
= drv_tags
->rqs
[i
];
3216 unsigned long rq_addr
= (unsigned long)rq
;
3218 if (rq_addr
>= start
&& rq_addr
< end
) {
3219 WARN_ON_ONCE(req_ref_read(rq
) != 0);
3220 cmpxchg(&drv_tags
->rqs
[i
], rq
, NULL
);
3226 * Wait until all pending iteration is done.
3228 * Request reference is cleared and it is guaranteed to be observed
3229 * after the ->lock is released.
3231 spin_lock_irqsave(&drv_tags
->lock
, flags
);
3232 spin_unlock_irqrestore(&drv_tags
->lock
, flags
);
3235 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
3236 unsigned int hctx_idx
)
3238 struct blk_mq_tags
*drv_tags
;
3241 if (list_empty(&tags
->page_list
))
3244 if (blk_mq_is_shared_tags(set
->flags
))
3245 drv_tags
= set
->shared_tags
;
3247 drv_tags
= set
->tags
[hctx_idx
];
3249 if (tags
->static_rqs
&& set
->ops
->exit_request
) {
3252 for (i
= 0; i
< tags
->nr_tags
; i
++) {
3253 struct request
*rq
= tags
->static_rqs
[i
];
3257 set
->ops
->exit_request(set
, rq
, hctx_idx
);
3258 tags
->static_rqs
[i
] = NULL
;
3262 blk_mq_clear_rq_mapping(drv_tags
, tags
);
3264 while (!list_empty(&tags
->page_list
)) {
3265 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
3266 list_del_init(&page
->lru
);
3268 * Remove kmemleak object previously allocated in
3269 * blk_mq_alloc_rqs().
3271 kmemleak_free(page_address(page
));
3272 __free_pages(page
, page
->private);
3276 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
3280 kfree(tags
->static_rqs
);
3281 tags
->static_rqs
= NULL
;
3283 blk_mq_free_tags(tags
);
3286 static enum hctx_type
hctx_idx_to_type(struct blk_mq_tag_set
*set
,
3287 unsigned int hctx_idx
)
3291 for (i
= 0; i
< set
->nr_maps
; i
++) {
3292 unsigned int start
= set
->map
[i
].queue_offset
;
3293 unsigned int end
= start
+ set
->map
[i
].nr_queues
;
3295 if (hctx_idx
>= start
&& hctx_idx
< end
)
3299 if (i
>= set
->nr_maps
)
3300 i
= HCTX_TYPE_DEFAULT
;
3305 static int blk_mq_get_hctx_node(struct blk_mq_tag_set
*set
,
3306 unsigned int hctx_idx
)
3308 enum hctx_type type
= hctx_idx_to_type(set
, hctx_idx
);
3310 return blk_mq_hw_queue_to_node(&set
->map
[type
], hctx_idx
);
3313 static struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
3314 unsigned int hctx_idx
,
3315 unsigned int nr_tags
,
3316 unsigned int reserved_tags
)
3318 int node
= blk_mq_get_hctx_node(set
, hctx_idx
);
3319 struct blk_mq_tags
*tags
;
3321 if (node
== NUMA_NO_NODE
)
3322 node
= set
->numa_node
;
3324 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
,
3325 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
3329 tags
->rqs
= kcalloc_node(nr_tags
, sizeof(struct request
*),
3330 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
3335 tags
->static_rqs
= kcalloc_node(nr_tags
, sizeof(struct request
*),
3336 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
3338 if (!tags
->static_rqs
)
3346 blk_mq_free_tags(tags
);
3350 static int blk_mq_init_request(struct blk_mq_tag_set
*set
, struct request
*rq
,
3351 unsigned int hctx_idx
, int node
)
3355 if (set
->ops
->init_request
) {
3356 ret
= set
->ops
->init_request(set
, rq
, hctx_idx
, node
);
3361 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
3365 static int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
,
3366 struct blk_mq_tags
*tags
,
3367 unsigned int hctx_idx
, unsigned int depth
)
3369 unsigned int i
, j
, entries_per_page
, max_order
= 4;
3370 int node
= blk_mq_get_hctx_node(set
, hctx_idx
);
3371 size_t rq_size
, left
;
3373 if (node
== NUMA_NO_NODE
)
3374 node
= set
->numa_node
;
3376 INIT_LIST_HEAD(&tags
->page_list
);
3379 * rq_size is the size of the request plus driver payload, rounded
3380 * to the cacheline size
3382 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
3384 left
= rq_size
* depth
;
3386 for (i
= 0; i
< depth
; ) {
3387 int this_order
= max_order
;
3392 while (this_order
&& left
< order_to_size(this_order
- 1))
3396 page
= alloc_pages_node(node
,
3397 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
3403 if (order_to_size(this_order
) < rq_size
)
3410 page
->private = this_order
;
3411 list_add_tail(&page
->lru
, &tags
->page_list
);
3413 p
= page_address(page
);
3415 * Allow kmemleak to scan these pages as they contain pointers
3416 * to additional allocations like via ops->init_request().
3418 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
3419 entries_per_page
= order_to_size(this_order
) / rq_size
;
3420 to_do
= min(entries_per_page
, depth
- i
);
3421 left
-= to_do
* rq_size
;
3422 for (j
= 0; j
< to_do
; j
++) {
3423 struct request
*rq
= p
;
3425 tags
->static_rqs
[i
] = rq
;
3426 if (blk_mq_init_request(set
, rq
, hctx_idx
, node
)) {
3427 tags
->static_rqs
[i
] = NULL
;
3438 blk_mq_free_rqs(set
, tags
, hctx_idx
);
3442 struct rq_iter_data
{
3443 struct blk_mq_hw_ctx
*hctx
;
3447 static bool blk_mq_has_request(struct request
*rq
, void *data
)
3449 struct rq_iter_data
*iter_data
= data
;
3451 if (rq
->mq_hctx
!= iter_data
->hctx
)
3453 iter_data
->has_rq
= true;
3457 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx
*hctx
)
3459 struct blk_mq_tags
*tags
= hctx
->sched_tags
?
3460 hctx
->sched_tags
: hctx
->tags
;
3461 struct rq_iter_data data
= {
3465 blk_mq_all_tag_iter(tags
, blk_mq_has_request
, &data
);
3469 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu
,
3470 struct blk_mq_hw_ctx
*hctx
)
3472 if (cpumask_first_and(hctx
->cpumask
, cpu_online_mask
) != cpu
)
3474 if (cpumask_next_and(cpu
, hctx
->cpumask
, cpu_online_mask
) < nr_cpu_ids
)
3479 static int blk_mq_hctx_notify_offline(unsigned int cpu
, struct hlist_node
*node
)
3481 struct blk_mq_hw_ctx
*hctx
= hlist_entry_safe(node
,
3482 struct blk_mq_hw_ctx
, cpuhp_online
);
3484 if (!cpumask_test_cpu(cpu
, hctx
->cpumask
) ||
3485 !blk_mq_last_cpu_in_hctx(cpu
, hctx
))
3489 * Prevent new request from being allocated on the current hctx.
3491 * The smp_mb__after_atomic() Pairs with the implied barrier in
3492 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3493 * seen once we return from the tag allocator.
3495 set_bit(BLK_MQ_S_INACTIVE
, &hctx
->state
);
3496 smp_mb__after_atomic();
3499 * Try to grab a reference to the queue and wait for any outstanding
3500 * requests. If we could not grab a reference the queue has been
3501 * frozen and there are no requests.
3503 if (percpu_ref_tryget(&hctx
->queue
->q_usage_counter
)) {
3504 while (blk_mq_hctx_has_requests(hctx
))
3506 percpu_ref_put(&hctx
->queue
->q_usage_counter
);
3512 static int blk_mq_hctx_notify_online(unsigned int cpu
, struct hlist_node
*node
)
3514 struct blk_mq_hw_ctx
*hctx
= hlist_entry_safe(node
,
3515 struct blk_mq_hw_ctx
, cpuhp_online
);
3517 if (cpumask_test_cpu(cpu
, hctx
->cpumask
))
3518 clear_bit(BLK_MQ_S_INACTIVE
, &hctx
->state
);
3523 * 'cpu' is going away. splice any existing rq_list entries from this
3524 * software queue to the hw queue dispatch list, and ensure that it
3527 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
3529 struct blk_mq_hw_ctx
*hctx
;
3530 struct blk_mq_ctx
*ctx
;
3532 enum hctx_type type
;
3534 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
3535 if (!cpumask_test_cpu(cpu
, hctx
->cpumask
))
3538 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
3541 spin_lock(&ctx
->lock
);
3542 if (!list_empty(&ctx
->rq_lists
[type
])) {
3543 list_splice_init(&ctx
->rq_lists
[type
], &tmp
);
3544 blk_mq_hctx_clear_pending(hctx
, ctx
);
3546 spin_unlock(&ctx
->lock
);
3548 if (list_empty(&tmp
))
3551 spin_lock(&hctx
->lock
);
3552 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
3553 spin_unlock(&hctx
->lock
);
3555 blk_mq_run_hw_queue(hctx
, true);
3559 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
3561 if (!(hctx
->flags
& BLK_MQ_F_STACKING
))
3562 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE
,
3563 &hctx
->cpuhp_online
);
3564 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
3569 * Before freeing hw queue, clearing the flush request reference in
3570 * tags->rqs[] for avoiding potential UAF.
3572 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags
*tags
,
3573 unsigned int queue_depth
, struct request
*flush_rq
)
3576 unsigned long flags
;
3578 /* The hw queue may not be mapped yet */
3582 WARN_ON_ONCE(req_ref_read(flush_rq
) != 0);
3584 for (i
= 0; i
< queue_depth
; i
++)
3585 cmpxchg(&tags
->rqs
[i
], flush_rq
, NULL
);
3588 * Wait until all pending iteration is done.
3590 * Request reference is cleared and it is guaranteed to be observed
3591 * after the ->lock is released.
3593 spin_lock_irqsave(&tags
->lock
, flags
);
3594 spin_unlock_irqrestore(&tags
->lock
, flags
);
3597 /* hctx->ctxs will be freed in queue's release handler */
3598 static void blk_mq_exit_hctx(struct request_queue
*q
,
3599 struct blk_mq_tag_set
*set
,
3600 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
3602 struct request
*flush_rq
= hctx
->fq
->flush_rq
;
3604 if (blk_mq_hw_queue_mapped(hctx
))
3605 blk_mq_tag_idle(hctx
);
3607 if (blk_queue_init_done(q
))
3608 blk_mq_clear_flush_rq_mapping(set
->tags
[hctx_idx
],
3609 set
->queue_depth
, flush_rq
);
3610 if (set
->ops
->exit_request
)
3611 set
->ops
->exit_request(set
, flush_rq
, hctx_idx
);
3613 if (set
->ops
->exit_hctx
)
3614 set
->ops
->exit_hctx(hctx
, hctx_idx
);
3616 blk_mq_remove_cpuhp(hctx
);
3618 xa_erase(&q
->hctx_table
, hctx_idx
);
3620 spin_lock(&q
->unused_hctx_lock
);
3621 list_add(&hctx
->hctx_list
, &q
->unused_hctx_list
);
3622 spin_unlock(&q
->unused_hctx_lock
);
3625 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
3626 struct blk_mq_tag_set
*set
, int nr_queue
)
3628 struct blk_mq_hw_ctx
*hctx
;
3631 queue_for_each_hw_ctx(q
, hctx
, i
) {
3634 blk_mq_exit_hctx(q
, set
, hctx
, i
);
3638 static int blk_mq_init_hctx(struct request_queue
*q
,
3639 struct blk_mq_tag_set
*set
,
3640 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
3642 hctx
->queue_num
= hctx_idx
;
3644 if (!(hctx
->flags
& BLK_MQ_F_STACKING
))
3645 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE
,
3646 &hctx
->cpuhp_online
);
3647 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
3649 hctx
->tags
= set
->tags
[hctx_idx
];
3651 if (set
->ops
->init_hctx
&&
3652 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
3653 goto unregister_cpu_notifier
;
3655 if (blk_mq_init_request(set
, hctx
->fq
->flush_rq
, hctx_idx
,
3659 if (xa_insert(&q
->hctx_table
, hctx_idx
, hctx
, GFP_KERNEL
))
3665 if (set
->ops
->exit_request
)
3666 set
->ops
->exit_request(set
, hctx
->fq
->flush_rq
, hctx_idx
);
3668 if (set
->ops
->exit_hctx
)
3669 set
->ops
->exit_hctx(hctx
, hctx_idx
);
3670 unregister_cpu_notifier
:
3671 blk_mq_remove_cpuhp(hctx
);
3675 static struct blk_mq_hw_ctx
*
3676 blk_mq_alloc_hctx(struct request_queue
*q
, struct blk_mq_tag_set
*set
,
3679 struct blk_mq_hw_ctx
*hctx
;
3680 gfp_t gfp
= GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
;
3682 hctx
= kzalloc_node(sizeof(struct blk_mq_hw_ctx
), gfp
, node
);
3684 goto fail_alloc_hctx
;
3686 if (!zalloc_cpumask_var_node(&hctx
->cpumask
, gfp
, node
))
3689 atomic_set(&hctx
->nr_active
, 0);
3690 if (node
== NUMA_NO_NODE
)
3691 node
= set
->numa_node
;
3692 hctx
->numa_node
= node
;
3694 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
3695 spin_lock_init(&hctx
->lock
);
3696 INIT_LIST_HEAD(&hctx
->dispatch
);
3698 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_QUEUE_SHARED
;
3700 INIT_LIST_HEAD(&hctx
->hctx_list
);
3703 * Allocate space for all possible cpus to avoid allocation at
3706 hctx
->ctxs
= kmalloc_array_node(nr_cpu_ids
, sizeof(void *),
3711 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8),
3712 gfp
, node
, false, false))
3716 spin_lock_init(&hctx
->dispatch_wait_lock
);
3717 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
3718 INIT_LIST_HEAD(&hctx
->dispatch_wait
.entry
);
3720 hctx
->fq
= blk_alloc_flush_queue(hctx
->numa_node
, set
->cmd_size
, gfp
);
3724 blk_mq_hctx_kobj_init(hctx
);
3729 sbitmap_free(&hctx
->ctx_map
);
3733 free_cpumask_var(hctx
->cpumask
);
3740 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
3741 unsigned int nr_hw_queues
)
3743 struct blk_mq_tag_set
*set
= q
->tag_set
;
3746 for_each_possible_cpu(i
) {
3747 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
3748 struct blk_mq_hw_ctx
*hctx
;
3752 spin_lock_init(&__ctx
->lock
);
3753 for (k
= HCTX_TYPE_DEFAULT
; k
< HCTX_MAX_TYPES
; k
++)
3754 INIT_LIST_HEAD(&__ctx
->rq_lists
[k
]);
3759 * Set local node, IFF we have more than one hw queue. If
3760 * not, we remain on the home node of the device
3762 for (j
= 0; j
< set
->nr_maps
; j
++) {
3763 hctx
= blk_mq_map_queue_type(q
, j
, i
);
3764 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
3765 hctx
->numa_node
= cpu_to_node(i
);
3770 struct blk_mq_tags
*blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set
*set
,
3771 unsigned int hctx_idx
,
3774 struct blk_mq_tags
*tags
;
3777 tags
= blk_mq_alloc_rq_map(set
, hctx_idx
, depth
, set
->reserved_tags
);
3781 ret
= blk_mq_alloc_rqs(set
, tags
, hctx_idx
, depth
);
3783 blk_mq_free_rq_map(tags
);
3790 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set
*set
,
3793 if (blk_mq_is_shared_tags(set
->flags
)) {
3794 set
->tags
[hctx_idx
] = set
->shared_tags
;
3799 set
->tags
[hctx_idx
] = blk_mq_alloc_map_and_rqs(set
, hctx_idx
,
3802 return set
->tags
[hctx_idx
];
3805 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set
*set
,
3806 struct blk_mq_tags
*tags
,
3807 unsigned int hctx_idx
)
3810 blk_mq_free_rqs(set
, tags
, hctx_idx
);
3811 blk_mq_free_rq_map(tags
);
3815 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set
*set
,
3816 unsigned int hctx_idx
)
3818 if (!blk_mq_is_shared_tags(set
->flags
))
3819 blk_mq_free_map_and_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
3821 set
->tags
[hctx_idx
] = NULL
;
3824 static void blk_mq_map_swqueue(struct request_queue
*q
)
3826 unsigned int j
, hctx_idx
;
3828 struct blk_mq_hw_ctx
*hctx
;
3829 struct blk_mq_ctx
*ctx
;
3830 struct blk_mq_tag_set
*set
= q
->tag_set
;
3832 queue_for_each_hw_ctx(q
, hctx
, i
) {
3833 cpumask_clear(hctx
->cpumask
);
3835 hctx
->dispatch_from
= NULL
;
3839 * Map software to hardware queues.
3841 * If the cpu isn't present, the cpu is mapped to first hctx.
3843 for_each_possible_cpu(i
) {
3845 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
3846 for (j
= 0; j
< set
->nr_maps
; j
++) {
3847 if (!set
->map
[j
].nr_queues
) {
3848 ctx
->hctxs
[j
] = blk_mq_map_queue_type(q
,
3849 HCTX_TYPE_DEFAULT
, i
);
3852 hctx_idx
= set
->map
[j
].mq_map
[i
];
3853 /* unmapped hw queue can be remapped after CPU topo changed */
3854 if (!set
->tags
[hctx_idx
] &&
3855 !__blk_mq_alloc_map_and_rqs(set
, hctx_idx
)) {
3857 * If tags initialization fail for some hctx,
3858 * that hctx won't be brought online. In this
3859 * case, remap the current ctx to hctx[0] which
3860 * is guaranteed to always have tags allocated
3862 set
->map
[j
].mq_map
[i
] = 0;
3865 hctx
= blk_mq_map_queue_type(q
, j
, i
);
3866 ctx
->hctxs
[j
] = hctx
;
3868 * If the CPU is already set in the mask, then we've
3869 * mapped this one already. This can happen if
3870 * devices share queues across queue maps.
3872 if (cpumask_test_cpu(i
, hctx
->cpumask
))
3875 cpumask_set_cpu(i
, hctx
->cpumask
);
3877 ctx
->index_hw
[hctx
->type
] = hctx
->nr_ctx
;
3878 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
3881 * If the nr_ctx type overflows, we have exceeded the
3882 * amount of sw queues we can support.
3884 BUG_ON(!hctx
->nr_ctx
);
3887 for (; j
< HCTX_MAX_TYPES
; j
++)
3888 ctx
->hctxs
[j
] = blk_mq_map_queue_type(q
,
3889 HCTX_TYPE_DEFAULT
, i
);
3892 queue_for_each_hw_ctx(q
, hctx
, i
) {
3894 * If no software queues are mapped to this hardware queue,
3895 * disable it and free the request entries.
3897 if (!hctx
->nr_ctx
) {
3898 /* Never unmap queue 0. We need it as a
3899 * fallback in case of a new remap fails
3903 __blk_mq_free_map_and_rqs(set
, i
);
3909 hctx
->tags
= set
->tags
[i
];
3910 WARN_ON(!hctx
->tags
);
3913 * Set the map size to the number of mapped software queues.
3914 * This is more accurate and more efficient than looping
3915 * over all possibly mapped software queues.
3917 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
3920 * Initialize batch roundrobin counts
3922 hctx
->next_cpu
= blk_mq_first_mapped_cpu(hctx
);
3923 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
3928 * Caller needs to ensure that we're either frozen/quiesced, or that
3929 * the queue isn't live yet.
3931 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
3933 struct blk_mq_hw_ctx
*hctx
;
3936 queue_for_each_hw_ctx(q
, hctx
, i
) {
3938 hctx
->flags
|= BLK_MQ_F_TAG_QUEUE_SHARED
;
3940 blk_mq_tag_idle(hctx
);
3941 hctx
->flags
&= ~BLK_MQ_F_TAG_QUEUE_SHARED
;
3946 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set
*set
,
3949 struct request_queue
*q
;
3951 lockdep_assert_held(&set
->tag_list_lock
);
3953 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3954 blk_mq_freeze_queue(q
);
3955 queue_set_hctx_shared(q
, shared
);
3956 blk_mq_unfreeze_queue(q
);
3960 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
3962 struct blk_mq_tag_set
*set
= q
->tag_set
;
3964 mutex_lock(&set
->tag_list_lock
);
3965 list_del(&q
->tag_set_list
);
3966 if (list_is_singular(&set
->tag_list
)) {
3967 /* just transitioned to unshared */
3968 set
->flags
&= ~BLK_MQ_F_TAG_QUEUE_SHARED
;
3969 /* update existing queue */
3970 blk_mq_update_tag_set_shared(set
, false);
3972 mutex_unlock(&set
->tag_list_lock
);
3973 INIT_LIST_HEAD(&q
->tag_set_list
);
3976 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
3977 struct request_queue
*q
)
3979 mutex_lock(&set
->tag_list_lock
);
3982 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3984 if (!list_empty(&set
->tag_list
) &&
3985 !(set
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
)) {
3986 set
->flags
|= BLK_MQ_F_TAG_QUEUE_SHARED
;
3987 /* update existing queue */
3988 blk_mq_update_tag_set_shared(set
, true);
3990 if (set
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
)
3991 queue_set_hctx_shared(q
, true);
3992 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
3994 mutex_unlock(&set
->tag_list_lock
);
3997 /* All allocations will be freed in release handler of q->mq_kobj */
3998 static int blk_mq_alloc_ctxs(struct request_queue
*q
)
4000 struct blk_mq_ctxs
*ctxs
;
4003 ctxs
= kzalloc(sizeof(*ctxs
), GFP_KERNEL
);
4007 ctxs
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
4008 if (!ctxs
->queue_ctx
)
4011 for_each_possible_cpu(cpu
) {
4012 struct blk_mq_ctx
*ctx
= per_cpu_ptr(ctxs
->queue_ctx
, cpu
);
4016 q
->mq_kobj
= &ctxs
->kobj
;
4017 q
->queue_ctx
= ctxs
->queue_ctx
;
4026 * It is the actual release handler for mq, but we do it from
4027 * request queue's release handler for avoiding use-after-free
4028 * and headache because q->mq_kobj shouldn't have been introduced,
4029 * but we can't group ctx/kctx kobj without it.
4031 void blk_mq_release(struct request_queue
*q
)
4033 struct blk_mq_hw_ctx
*hctx
, *next
;
4036 queue_for_each_hw_ctx(q
, hctx
, i
)
4037 WARN_ON_ONCE(hctx
&& list_empty(&hctx
->hctx_list
));
4039 /* all hctx are in .unused_hctx_list now */
4040 list_for_each_entry_safe(hctx
, next
, &q
->unused_hctx_list
, hctx_list
) {
4041 list_del_init(&hctx
->hctx_list
);
4042 kobject_put(&hctx
->kobj
);
4045 xa_destroy(&q
->hctx_table
);
4048 * release .mq_kobj and sw queue's kobject now because
4049 * both share lifetime with request queue.
4051 blk_mq_sysfs_deinit(q
);
4054 static struct request_queue
*blk_mq_init_queue_data(struct blk_mq_tag_set
*set
,
4057 struct request_queue
*q
;
4060 q
= blk_alloc_queue(set
->numa_node
);
4062 return ERR_PTR(-ENOMEM
);
4063 q
->queuedata
= queuedata
;
4064 ret
= blk_mq_init_allocated_queue(set
, q
);
4067 return ERR_PTR(ret
);
4072 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
4074 return blk_mq_init_queue_data(set
, NULL
);
4076 EXPORT_SYMBOL(blk_mq_init_queue
);
4079 * blk_mq_destroy_queue - shutdown a request queue
4080 * @q: request queue to shutdown
4082 * This shuts down a request queue allocated by blk_mq_init_queue(). All future
4083 * requests will be failed with -ENODEV. The caller is responsible for dropping
4084 * the reference from blk_mq_init_queue() by calling blk_put_queue().
4086 * Context: can sleep
4088 void blk_mq_destroy_queue(struct request_queue
*q
)
4090 WARN_ON_ONCE(!queue_is_mq(q
));
4091 WARN_ON_ONCE(blk_queue_registered(q
));
4095 blk_queue_flag_set(QUEUE_FLAG_DYING
, q
);
4096 blk_queue_start_drain(q
);
4097 blk_mq_freeze_queue_wait(q
);
4100 blk_mq_cancel_work_sync(q
);
4101 blk_mq_exit_queue(q
);
4103 EXPORT_SYMBOL(blk_mq_destroy_queue
);
4105 struct gendisk
*__blk_mq_alloc_disk(struct blk_mq_tag_set
*set
, void *queuedata
,
4106 struct lock_class_key
*lkclass
)
4108 struct request_queue
*q
;
4109 struct gendisk
*disk
;
4111 q
= blk_mq_init_queue_data(set
, queuedata
);
4115 disk
= __alloc_disk_node(q
, set
->numa_node
, lkclass
);
4117 blk_mq_destroy_queue(q
);
4119 return ERR_PTR(-ENOMEM
);
4121 set_bit(GD_OWNS_QUEUE
, &disk
->state
);
4124 EXPORT_SYMBOL(__blk_mq_alloc_disk
);
4126 struct gendisk
*blk_mq_alloc_disk_for_queue(struct request_queue
*q
,
4127 struct lock_class_key
*lkclass
)
4129 struct gendisk
*disk
;
4131 if (!blk_get_queue(q
))
4133 disk
= __alloc_disk_node(q
, NUMA_NO_NODE
, lkclass
);
4138 EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue
);
4140 static struct blk_mq_hw_ctx
*blk_mq_alloc_and_init_hctx(
4141 struct blk_mq_tag_set
*set
, struct request_queue
*q
,
4142 int hctx_idx
, int node
)
4144 struct blk_mq_hw_ctx
*hctx
= NULL
, *tmp
;
4146 /* reuse dead hctx first */
4147 spin_lock(&q
->unused_hctx_lock
);
4148 list_for_each_entry(tmp
, &q
->unused_hctx_list
, hctx_list
) {
4149 if (tmp
->numa_node
== node
) {
4155 list_del_init(&hctx
->hctx_list
);
4156 spin_unlock(&q
->unused_hctx_lock
);
4159 hctx
= blk_mq_alloc_hctx(q
, set
, node
);
4163 if (blk_mq_init_hctx(q
, set
, hctx
, hctx_idx
))
4169 kobject_put(&hctx
->kobj
);
4174 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
4175 struct request_queue
*q
)
4177 struct blk_mq_hw_ctx
*hctx
;
4180 /* protect against switching io scheduler */
4181 mutex_lock(&q
->sysfs_lock
);
4182 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
4184 int node
= blk_mq_get_hctx_node(set
, i
);
4185 struct blk_mq_hw_ctx
*old_hctx
= xa_load(&q
->hctx_table
, i
);
4188 old_node
= old_hctx
->numa_node
;
4189 blk_mq_exit_hctx(q
, set
, old_hctx
, i
);
4192 if (!blk_mq_alloc_and_init_hctx(set
, q
, i
, node
)) {
4195 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4197 hctx
= blk_mq_alloc_and_init_hctx(set
, q
, i
, old_node
);
4198 WARN_ON_ONCE(!hctx
);
4202 * Increasing nr_hw_queues fails. Free the newly allocated
4203 * hctxs and keep the previous q->nr_hw_queues.
4205 if (i
!= set
->nr_hw_queues
) {
4206 j
= q
->nr_hw_queues
;
4209 q
->nr_hw_queues
= set
->nr_hw_queues
;
4212 xa_for_each_start(&q
->hctx_table
, j
, hctx
, j
)
4213 blk_mq_exit_hctx(q
, set
, hctx
, j
);
4214 mutex_unlock(&q
->sysfs_lock
);
4217 static void blk_mq_update_poll_flag(struct request_queue
*q
)
4219 struct blk_mq_tag_set
*set
= q
->tag_set
;
4221 if (set
->nr_maps
> HCTX_TYPE_POLL
&&
4222 set
->map
[HCTX_TYPE_POLL
].nr_queues
)
4223 blk_queue_flag_set(QUEUE_FLAG_POLL
, q
);
4225 blk_queue_flag_clear(QUEUE_FLAG_POLL
, q
);
4228 int blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
4229 struct request_queue
*q
)
4231 /* mark the queue as mq asap */
4232 q
->mq_ops
= set
->ops
;
4234 if (blk_mq_alloc_ctxs(q
))
4237 /* init q->mq_kobj and sw queues' kobjects */
4238 blk_mq_sysfs_init(q
);
4240 INIT_LIST_HEAD(&q
->unused_hctx_list
);
4241 spin_lock_init(&q
->unused_hctx_lock
);
4243 xa_init(&q
->hctx_table
);
4245 blk_mq_realloc_hw_ctxs(set
, q
);
4246 if (!q
->nr_hw_queues
)
4249 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
4250 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
4254 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
4255 blk_mq_update_poll_flag(q
);
4257 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
4258 INIT_LIST_HEAD(&q
->flush_list
);
4259 INIT_LIST_HEAD(&q
->requeue_list
);
4260 spin_lock_init(&q
->requeue_lock
);
4262 q
->nr_requests
= set
->queue_depth
;
4264 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
4265 blk_mq_add_queue_tag_set(set
, q
);
4266 blk_mq_map_swqueue(q
);
4275 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
4277 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4278 void blk_mq_exit_queue(struct request_queue
*q
)
4280 struct blk_mq_tag_set
*set
= q
->tag_set
;
4282 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4283 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
4284 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4285 blk_mq_del_queue_tag_set(q
);
4288 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
4292 if (blk_mq_is_shared_tags(set
->flags
)) {
4293 set
->shared_tags
= blk_mq_alloc_map_and_rqs(set
,
4296 if (!set
->shared_tags
)
4300 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
4301 if (!__blk_mq_alloc_map_and_rqs(set
, i
))
4310 __blk_mq_free_map_and_rqs(set
, i
);
4312 if (blk_mq_is_shared_tags(set
->flags
)) {
4313 blk_mq_free_map_and_rqs(set
, set
->shared_tags
,
4314 BLK_MQ_NO_HCTX_IDX
);
4321 * Allocate the request maps associated with this tag_set. Note that this
4322 * may reduce the depth asked for, if memory is tight. set->queue_depth
4323 * will be updated to reflect the allocated depth.
4325 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set
*set
)
4330 depth
= set
->queue_depth
;
4332 err
= __blk_mq_alloc_rq_maps(set
);
4336 set
->queue_depth
>>= 1;
4337 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
4341 } while (set
->queue_depth
);
4343 if (!set
->queue_depth
|| err
) {
4344 pr_err("blk-mq: failed to allocate request map\n");
4348 if (depth
!= set
->queue_depth
)
4349 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4350 depth
, set
->queue_depth
);
4355 static void blk_mq_update_queue_map(struct blk_mq_tag_set
*set
)
4358 * blk_mq_map_queues() and multiple .map_queues() implementations
4359 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4360 * number of hardware queues.
4362 if (set
->nr_maps
== 1)
4363 set
->map
[HCTX_TYPE_DEFAULT
].nr_queues
= set
->nr_hw_queues
;
4365 if (set
->ops
->map_queues
&& !is_kdump_kernel()) {
4369 * transport .map_queues is usually done in the following
4372 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4373 * mask = get_cpu_mask(queue)
4374 * for_each_cpu(cpu, mask)
4375 * set->map[x].mq_map[cpu] = queue;
4378 * When we need to remap, the table has to be cleared for
4379 * killing stale mapping since one CPU may not be mapped
4382 for (i
= 0; i
< set
->nr_maps
; i
++)
4383 blk_mq_clear_mq_map(&set
->map
[i
]);
4385 set
->ops
->map_queues(set
);
4387 BUG_ON(set
->nr_maps
> 1);
4388 blk_mq_map_queues(&set
->map
[HCTX_TYPE_DEFAULT
]);
4392 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set
*set
,
4393 int new_nr_hw_queues
)
4395 struct blk_mq_tags
**new_tags
;
4398 if (set
->nr_hw_queues
>= new_nr_hw_queues
)
4401 new_tags
= kcalloc_node(new_nr_hw_queues
, sizeof(struct blk_mq_tags
*),
4402 GFP_KERNEL
, set
->numa_node
);
4407 memcpy(new_tags
, set
->tags
, set
->nr_hw_queues
*
4408 sizeof(*set
->tags
));
4410 set
->tags
= new_tags
;
4412 for (i
= set
->nr_hw_queues
; i
< new_nr_hw_queues
; i
++) {
4413 if (!__blk_mq_alloc_map_and_rqs(set
, i
)) {
4414 while (--i
>= set
->nr_hw_queues
)
4415 __blk_mq_free_map_and_rqs(set
, i
);
4422 set
->nr_hw_queues
= new_nr_hw_queues
;
4427 * Alloc a tag set to be associated with one or more request queues.
4428 * May fail with EINVAL for various error conditions. May adjust the
4429 * requested depth down, if it's too large. In that case, the set
4430 * value will be stored in set->queue_depth.
4432 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
4436 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
4438 if (!set
->nr_hw_queues
)
4440 if (!set
->queue_depth
)
4442 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
4445 if (!set
->ops
->queue_rq
)
4448 if (!set
->ops
->get_budget
^ !set
->ops
->put_budget
)
4451 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
4452 pr_info("blk-mq: reduced tag depth to %u\n",
4454 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
4459 else if (set
->nr_maps
> HCTX_MAX_TYPES
)
4463 * If a crashdump is active, then we are potentially in a very
4464 * memory constrained environment. Limit us to 1 queue and
4465 * 64 tags to prevent using too much memory.
4467 if (is_kdump_kernel()) {
4468 set
->nr_hw_queues
= 1;
4470 set
->queue_depth
= min(64U, set
->queue_depth
);
4473 * There is no use for more h/w queues than cpus if we just have
4476 if (set
->nr_maps
== 1 && set
->nr_hw_queues
> nr_cpu_ids
)
4477 set
->nr_hw_queues
= nr_cpu_ids
;
4479 if (set
->flags
& BLK_MQ_F_BLOCKING
) {
4480 set
->srcu
= kmalloc(sizeof(*set
->srcu
), GFP_KERNEL
);
4483 ret
= init_srcu_struct(set
->srcu
);
4489 set
->tags
= kcalloc_node(set
->nr_hw_queues
,
4490 sizeof(struct blk_mq_tags
*), GFP_KERNEL
,
4493 goto out_cleanup_srcu
;
4495 for (i
= 0; i
< set
->nr_maps
; i
++) {
4496 set
->map
[i
].mq_map
= kcalloc_node(nr_cpu_ids
,
4497 sizeof(set
->map
[i
].mq_map
[0]),
4498 GFP_KERNEL
, set
->numa_node
);
4499 if (!set
->map
[i
].mq_map
)
4500 goto out_free_mq_map
;
4501 set
->map
[i
].nr_queues
= is_kdump_kernel() ? 1 : set
->nr_hw_queues
;
4504 blk_mq_update_queue_map(set
);
4506 ret
= blk_mq_alloc_set_map_and_rqs(set
);
4508 goto out_free_mq_map
;
4510 mutex_init(&set
->tag_list_lock
);
4511 INIT_LIST_HEAD(&set
->tag_list
);
4516 for (i
= 0; i
< set
->nr_maps
; i
++) {
4517 kfree(set
->map
[i
].mq_map
);
4518 set
->map
[i
].mq_map
= NULL
;
4523 if (set
->flags
& BLK_MQ_F_BLOCKING
)
4524 cleanup_srcu_struct(set
->srcu
);
4526 if (set
->flags
& BLK_MQ_F_BLOCKING
)
4530 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
4532 /* allocate and initialize a tagset for a simple single-queue device */
4533 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set
*set
,
4534 const struct blk_mq_ops
*ops
, unsigned int queue_depth
,
4535 unsigned int set_flags
)
4537 memset(set
, 0, sizeof(*set
));
4539 set
->nr_hw_queues
= 1;
4541 set
->queue_depth
= queue_depth
;
4542 set
->numa_node
= NUMA_NO_NODE
;
4543 set
->flags
= set_flags
;
4544 return blk_mq_alloc_tag_set(set
);
4546 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set
);
4548 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
4552 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
4553 __blk_mq_free_map_and_rqs(set
, i
);
4555 if (blk_mq_is_shared_tags(set
->flags
)) {
4556 blk_mq_free_map_and_rqs(set
, set
->shared_tags
,
4557 BLK_MQ_NO_HCTX_IDX
);
4560 for (j
= 0; j
< set
->nr_maps
; j
++) {
4561 kfree(set
->map
[j
].mq_map
);
4562 set
->map
[j
].mq_map
= NULL
;
4567 if (set
->flags
& BLK_MQ_F_BLOCKING
) {
4568 cleanup_srcu_struct(set
->srcu
);
4572 EXPORT_SYMBOL(blk_mq_free_tag_set
);
4574 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
4576 struct blk_mq_tag_set
*set
= q
->tag_set
;
4577 struct blk_mq_hw_ctx
*hctx
;
4584 if (q
->nr_requests
== nr
)
4587 blk_mq_freeze_queue(q
);
4588 blk_mq_quiesce_queue(q
);
4591 queue_for_each_hw_ctx(q
, hctx
, i
) {
4595 * If we're using an MQ scheduler, just update the scheduler
4596 * queue depth. This is similar to what the old code would do.
4598 if (hctx
->sched_tags
) {
4599 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
4602 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
, nr
,
4607 if (q
->elevator
&& q
->elevator
->type
->ops
.depth_updated
)
4608 q
->elevator
->type
->ops
.depth_updated(hctx
);
4611 q
->nr_requests
= nr
;
4612 if (blk_mq_is_shared_tags(set
->flags
)) {
4614 blk_mq_tag_update_sched_shared_tags(q
);
4616 blk_mq_tag_resize_shared_tags(set
, nr
);
4620 blk_mq_unquiesce_queue(q
);
4621 blk_mq_unfreeze_queue(q
);
4627 * request_queue and elevator_type pair.
4628 * It is just used by __blk_mq_update_nr_hw_queues to cache
4629 * the elevator_type associated with a request_queue.
4631 struct blk_mq_qe_pair
{
4632 struct list_head node
;
4633 struct request_queue
*q
;
4634 struct elevator_type
*type
;
4638 * Cache the elevator_type in qe pair list and switch the
4639 * io scheduler to 'none'
4641 static bool blk_mq_elv_switch_none(struct list_head
*head
,
4642 struct request_queue
*q
)
4644 struct blk_mq_qe_pair
*qe
;
4646 qe
= kmalloc(sizeof(*qe
), GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
);
4650 /* q->elevator needs protection from ->sysfs_lock */
4651 mutex_lock(&q
->sysfs_lock
);
4653 /* the check has to be done with holding sysfs_lock */
4659 INIT_LIST_HEAD(&qe
->node
);
4661 qe
->type
= q
->elevator
->type
;
4662 /* keep a reference to the elevator module as we'll switch back */
4663 __elevator_get(qe
->type
);
4664 list_add(&qe
->node
, head
);
4665 elevator_disable(q
);
4667 mutex_unlock(&q
->sysfs_lock
);
4672 static struct blk_mq_qe_pair
*blk_lookup_qe_pair(struct list_head
*head
,
4673 struct request_queue
*q
)
4675 struct blk_mq_qe_pair
*qe
;
4677 list_for_each_entry(qe
, head
, node
)
4684 static void blk_mq_elv_switch_back(struct list_head
*head
,
4685 struct request_queue
*q
)
4687 struct blk_mq_qe_pair
*qe
;
4688 struct elevator_type
*t
;
4690 qe
= blk_lookup_qe_pair(head
, q
);
4694 list_del(&qe
->node
);
4697 mutex_lock(&q
->sysfs_lock
);
4698 elevator_switch(q
, t
);
4699 /* drop the reference acquired in blk_mq_elv_switch_none */
4701 mutex_unlock(&q
->sysfs_lock
);
4704 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
,
4707 struct request_queue
*q
;
4709 int prev_nr_hw_queues
= set
->nr_hw_queues
;
4712 lockdep_assert_held(&set
->tag_list_lock
);
4714 if (set
->nr_maps
== 1 && nr_hw_queues
> nr_cpu_ids
)
4715 nr_hw_queues
= nr_cpu_ids
;
4716 if (nr_hw_queues
< 1)
4718 if (set
->nr_maps
== 1 && nr_hw_queues
== set
->nr_hw_queues
)
4721 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
4722 blk_mq_freeze_queue(q
);
4724 * Switch IO scheduler to 'none', cleaning up the data associated
4725 * with the previous scheduler. We will switch back once we are done
4726 * updating the new sw to hw queue mappings.
4728 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
4729 if (!blk_mq_elv_switch_none(&head
, q
))
4732 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
4733 blk_mq_debugfs_unregister_hctxs(q
);
4734 blk_mq_sysfs_unregister_hctxs(q
);
4737 if (blk_mq_realloc_tag_set_tags(set
, nr_hw_queues
) < 0)
4741 blk_mq_update_queue_map(set
);
4742 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
4743 blk_mq_realloc_hw_ctxs(set
, q
);
4744 blk_mq_update_poll_flag(q
);
4745 if (q
->nr_hw_queues
!= set
->nr_hw_queues
) {
4746 int i
= prev_nr_hw_queues
;
4748 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4749 nr_hw_queues
, prev_nr_hw_queues
);
4750 for (; i
< set
->nr_hw_queues
; i
++)
4751 __blk_mq_free_map_and_rqs(set
, i
);
4753 set
->nr_hw_queues
= prev_nr_hw_queues
;
4756 blk_mq_map_swqueue(q
);
4760 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
4761 blk_mq_sysfs_register_hctxs(q
);
4762 blk_mq_debugfs_register_hctxs(q
);
4766 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
4767 blk_mq_elv_switch_back(&head
, q
);
4769 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
4770 blk_mq_unfreeze_queue(q
);
4772 /* Free the excess tags when nr_hw_queues shrink. */
4773 for (i
= set
->nr_hw_queues
; i
< prev_nr_hw_queues
; i
++)
4774 __blk_mq_free_map_and_rqs(set
, i
);
4777 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
4779 mutex_lock(&set
->tag_list_lock
);
4780 __blk_mq_update_nr_hw_queues(set
, nr_hw_queues
);
4781 mutex_unlock(&set
->tag_list_lock
);
4783 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
4785 static int blk_hctx_poll(struct request_queue
*q
, struct blk_mq_hw_ctx
*hctx
,
4786 struct io_comp_batch
*iob
, unsigned int flags
)
4788 long state
= get_current_state();
4792 ret
= q
->mq_ops
->poll(hctx
, iob
);
4794 __set_current_state(TASK_RUNNING
);
4798 if (signal_pending_state(state
, current
))
4799 __set_current_state(TASK_RUNNING
);
4800 if (task_is_running(current
))
4803 if (ret
< 0 || (flags
& BLK_POLL_ONESHOT
))
4806 } while (!need_resched());
4808 __set_current_state(TASK_RUNNING
);
4812 int blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
,
4813 struct io_comp_batch
*iob
, unsigned int flags
)
4815 struct blk_mq_hw_ctx
*hctx
= xa_load(&q
->hctx_table
, cookie
);
4817 return blk_hctx_poll(q
, hctx
, iob
, flags
);
4820 int blk_rq_poll(struct request
*rq
, struct io_comp_batch
*iob
,
4821 unsigned int poll_flags
)
4823 struct request_queue
*q
= rq
->q
;
4826 if (!blk_rq_is_poll(rq
))
4828 if (!percpu_ref_tryget(&q
->q_usage_counter
))
4831 ret
= blk_hctx_poll(q
, rq
->mq_hctx
, iob
, poll_flags
);
4836 EXPORT_SYMBOL_GPL(blk_rq_poll
);
4838 unsigned int blk_mq_rq_cpu(struct request
*rq
)
4840 return rq
->mq_ctx
->cpu
;
4842 EXPORT_SYMBOL(blk_mq_rq_cpu
);
4844 void blk_mq_cancel_work_sync(struct request_queue
*q
)
4846 struct blk_mq_hw_ctx
*hctx
;
4849 cancel_delayed_work_sync(&q
->requeue_work
);
4851 queue_for_each_hw_ctx(q
, hctx
, i
)
4852 cancel_delayed_work_sync(&hctx
->run_work
);
4855 static int __init
blk_mq_init(void)
4859 for_each_possible_cpu(i
)
4860 init_llist_head(&per_cpu(blk_cpu_done
, i
));
4861 for_each_possible_cpu(i
)
4862 INIT_CSD(&per_cpu(blk_cpu_csd
, i
),
4863 __blk_mq_complete_request_remote
, NULL
);
4864 open_softirq(BLOCK_SOFTIRQ
, blk_done_softirq
);
4866 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD
,
4867 "block/softirq:dead", NULL
,
4868 blk_softirq_cpu_dead
);
4869 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
4870 blk_mq_hctx_notify_dead
);
4871 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE
, "block/mq:online",
4872 blk_mq_hctx_notify_online
,
4873 blk_mq_hctx_notify_offline
);
4876 subsys_initcall(blk_mq_init
);