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/topology.h>
25 #include <linux/sched/signal.h>
26 #include <linux/delay.h>
27 #include <linux/crash_dump.h>
28 #include <linux/prefetch.h>
29 #include <linux/blk-crypto.h>
30 #include <linux/part_stat.h>
32 #include <trace/events/block.h>
34 #include <linux/t10-pi.h>
37 #include "blk-mq-debugfs.h"
40 #include "blk-mq-sched.h"
41 #include "blk-rq-qos.h"
43 static DEFINE_PER_CPU(struct llist_head
, blk_cpu_done
);
44 static DEFINE_PER_CPU(call_single_data_t
, blk_cpu_csd
);
46 static void blk_mq_insert_request(struct request
*rq
, blk_insert_t flags
);
47 static void blk_mq_request_bypass_insert(struct request
*rq
,
49 static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx
*hctx
,
50 struct list_head
*list
);
51 static int blk_hctx_poll(struct request_queue
*q
, struct blk_mq_hw_ctx
*hctx
,
52 struct io_comp_batch
*iob
, unsigned int flags
);
55 * Check if any of the ctx, dispatch list or elevator
56 * have pending work in this hardware queue.
58 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
60 return !list_empty_careful(&hctx
->dispatch
) ||
61 sbitmap_any_bit_set(&hctx
->ctx_map
) ||
62 blk_mq_sched_has_work(hctx
);
66 * Mark this ctx as having pending work in this hardware queue
68 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
69 struct blk_mq_ctx
*ctx
)
71 const int bit
= ctx
->index_hw
[hctx
->type
];
73 if (!sbitmap_test_bit(&hctx
->ctx_map
, bit
))
74 sbitmap_set_bit(&hctx
->ctx_map
, bit
);
77 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
78 struct blk_mq_ctx
*ctx
)
80 const int bit
= ctx
->index_hw
[hctx
->type
];
82 sbitmap_clear_bit(&hctx
->ctx_map
, bit
);
86 struct block_device
*part
;
87 unsigned int inflight
[2];
90 static bool blk_mq_check_inflight(struct request
*rq
, void *priv
)
92 struct mq_inflight
*mi
= priv
;
94 if (rq
->part
&& blk_do_io_stat(rq
) &&
95 (!mi
->part
->bd_partno
|| rq
->part
== mi
->part
) &&
96 blk_mq_rq_state(rq
) == MQ_RQ_IN_FLIGHT
)
97 mi
->inflight
[rq_data_dir(rq
)]++;
102 unsigned int blk_mq_in_flight(struct request_queue
*q
,
103 struct block_device
*part
)
105 struct mq_inflight mi
= { .part
= part
};
107 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight
, &mi
);
109 return mi
.inflight
[0] + mi
.inflight
[1];
112 void blk_mq_in_flight_rw(struct request_queue
*q
, struct block_device
*part
,
113 unsigned int inflight
[2])
115 struct mq_inflight mi
= { .part
= part
};
117 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight
, &mi
);
118 inflight
[0] = mi
.inflight
[0];
119 inflight
[1] = mi
.inflight
[1];
122 void blk_freeze_queue_start(struct request_queue
*q
)
124 mutex_lock(&q
->mq_freeze_lock
);
125 if (++q
->mq_freeze_depth
== 1) {
126 percpu_ref_kill(&q
->q_usage_counter
);
127 mutex_unlock(&q
->mq_freeze_lock
);
129 blk_mq_run_hw_queues(q
, false);
131 mutex_unlock(&q
->mq_freeze_lock
);
134 EXPORT_SYMBOL_GPL(blk_freeze_queue_start
);
136 void blk_mq_freeze_queue_wait(struct request_queue
*q
)
138 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
140 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait
);
142 int blk_mq_freeze_queue_wait_timeout(struct request_queue
*q
,
143 unsigned long timeout
)
145 return wait_event_timeout(q
->mq_freeze_wq
,
146 percpu_ref_is_zero(&q
->q_usage_counter
),
149 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout
);
152 * Guarantee no request is in use, so we can change any data structure of
153 * the queue afterward.
155 void blk_freeze_queue(struct request_queue
*q
)
158 * In the !blk_mq case we are only calling this to kill the
159 * q_usage_counter, otherwise this increases the freeze depth
160 * and waits for it to return to zero. For this reason there is
161 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
162 * exported to drivers as the only user for unfreeze is blk_mq.
164 blk_freeze_queue_start(q
);
165 blk_mq_freeze_queue_wait(q
);
168 void blk_mq_freeze_queue(struct request_queue
*q
)
171 * ...just an alias to keep freeze and unfreeze actions balanced
172 * in the blk_mq_* namespace
176 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
178 void __blk_mq_unfreeze_queue(struct request_queue
*q
, bool force_atomic
)
180 mutex_lock(&q
->mq_freeze_lock
);
182 q
->q_usage_counter
.data
->force_atomic
= true;
183 q
->mq_freeze_depth
--;
184 WARN_ON_ONCE(q
->mq_freeze_depth
< 0);
185 if (!q
->mq_freeze_depth
) {
186 percpu_ref_resurrect(&q
->q_usage_counter
);
187 wake_up_all(&q
->mq_freeze_wq
);
189 mutex_unlock(&q
->mq_freeze_lock
);
192 void blk_mq_unfreeze_queue(struct request_queue
*q
)
194 __blk_mq_unfreeze_queue(q
, false);
196 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
199 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
200 * mpt3sas driver such that this function can be removed.
202 void blk_mq_quiesce_queue_nowait(struct request_queue
*q
)
206 spin_lock_irqsave(&q
->queue_lock
, flags
);
207 if (!q
->quiesce_depth
++)
208 blk_queue_flag_set(QUEUE_FLAG_QUIESCED
, q
);
209 spin_unlock_irqrestore(&q
->queue_lock
, flags
);
211 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait
);
214 * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done
215 * @set: tag_set to wait on
217 * Note: it is driver's responsibility for making sure that quiesce has
218 * been started on or more of the request_queues of the tag_set. This
219 * function only waits for the quiesce on those request_queues that had
220 * the quiesce flag set using blk_mq_quiesce_queue_nowait.
222 void blk_mq_wait_quiesce_done(struct blk_mq_tag_set
*set
)
224 if (set
->flags
& BLK_MQ_F_BLOCKING
)
225 synchronize_srcu(set
->srcu
);
229 EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done
);
232 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
235 * Note: this function does not prevent that the struct request end_io()
236 * callback function is invoked. Once this function is returned, we make
237 * sure no dispatch can happen until the queue is unquiesced via
238 * blk_mq_unquiesce_queue().
240 void blk_mq_quiesce_queue(struct request_queue
*q
)
242 blk_mq_quiesce_queue_nowait(q
);
243 /* nothing to wait for non-mq queues */
245 blk_mq_wait_quiesce_done(q
->tag_set
);
247 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
250 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
253 * This function recovers queue into the state before quiescing
254 * which is done by blk_mq_quiesce_queue.
256 void blk_mq_unquiesce_queue(struct request_queue
*q
)
259 bool run_queue
= false;
261 spin_lock_irqsave(&q
->queue_lock
, flags
);
262 if (WARN_ON_ONCE(q
->quiesce_depth
<= 0)) {
264 } else if (!--q
->quiesce_depth
) {
265 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED
, q
);
268 spin_unlock_irqrestore(&q
->queue_lock
, flags
);
270 /* dispatch requests which are inserted during quiescing */
272 blk_mq_run_hw_queues(q
, true);
274 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue
);
276 void blk_mq_quiesce_tagset(struct blk_mq_tag_set
*set
)
278 struct request_queue
*q
;
280 mutex_lock(&set
->tag_list_lock
);
281 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
282 if (!blk_queue_skip_tagset_quiesce(q
))
283 blk_mq_quiesce_queue_nowait(q
);
285 blk_mq_wait_quiesce_done(set
);
286 mutex_unlock(&set
->tag_list_lock
);
288 EXPORT_SYMBOL_GPL(blk_mq_quiesce_tagset
);
290 void blk_mq_unquiesce_tagset(struct blk_mq_tag_set
*set
)
292 struct request_queue
*q
;
294 mutex_lock(&set
->tag_list_lock
);
295 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
296 if (!blk_queue_skip_tagset_quiesce(q
))
297 blk_mq_unquiesce_queue(q
);
299 mutex_unlock(&set
->tag_list_lock
);
301 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_tagset
);
303 void blk_mq_wake_waiters(struct request_queue
*q
)
305 struct blk_mq_hw_ctx
*hctx
;
308 queue_for_each_hw_ctx(q
, hctx
, i
)
309 if (blk_mq_hw_queue_mapped(hctx
))
310 blk_mq_tag_wakeup_all(hctx
->tags
, true);
313 void blk_rq_init(struct request_queue
*q
, struct request
*rq
)
315 memset(rq
, 0, sizeof(*rq
));
317 INIT_LIST_HEAD(&rq
->queuelist
);
319 rq
->__sector
= (sector_t
) -1;
320 INIT_HLIST_NODE(&rq
->hash
);
321 RB_CLEAR_NODE(&rq
->rb_node
);
322 rq
->tag
= BLK_MQ_NO_TAG
;
323 rq
->internal_tag
= BLK_MQ_NO_TAG
;
324 rq
->start_time_ns
= blk_time_get_ns();
326 blk_crypto_rq_set_defaults(rq
);
328 EXPORT_SYMBOL(blk_rq_init
);
330 /* Set start and alloc time when the allocated request is actually used */
331 static inline void blk_mq_rq_time_init(struct request
*rq
, u64 alloc_time_ns
)
333 if (blk_mq_need_time_stamp(rq
))
334 rq
->start_time_ns
= blk_time_get_ns();
336 rq
->start_time_ns
= 0;
338 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
339 if (blk_queue_rq_alloc_time(rq
->q
))
340 rq
->alloc_time_ns
= alloc_time_ns
?: rq
->start_time_ns
;
342 rq
->alloc_time_ns
= 0;
346 static struct request
*blk_mq_rq_ctx_init(struct blk_mq_alloc_data
*data
,
347 struct blk_mq_tags
*tags
, unsigned int tag
)
349 struct blk_mq_ctx
*ctx
= data
->ctx
;
350 struct blk_mq_hw_ctx
*hctx
= data
->hctx
;
351 struct request_queue
*q
= data
->q
;
352 struct request
*rq
= tags
->static_rqs
[tag
];
357 rq
->cmd_flags
= data
->cmd_flags
;
359 if (data
->flags
& BLK_MQ_REQ_PM
)
360 data
->rq_flags
|= RQF_PM
;
361 if (blk_queue_io_stat(q
))
362 data
->rq_flags
|= RQF_IO_STAT
;
363 rq
->rq_flags
= data
->rq_flags
;
365 if (data
->rq_flags
& RQF_SCHED_TAGS
) {
366 rq
->tag
= BLK_MQ_NO_TAG
;
367 rq
->internal_tag
= tag
;
370 rq
->internal_tag
= BLK_MQ_NO_TAG
;
375 rq
->io_start_time_ns
= 0;
376 rq
->stats_sectors
= 0;
377 rq
->nr_phys_segments
= 0;
378 #if defined(CONFIG_BLK_DEV_INTEGRITY)
379 rq
->nr_integrity_segments
= 0;
382 rq
->end_io_data
= NULL
;
384 blk_crypto_rq_set_defaults(rq
);
385 INIT_LIST_HEAD(&rq
->queuelist
);
386 /* tag was already set */
387 WRITE_ONCE(rq
->deadline
, 0);
390 if (rq
->rq_flags
& RQF_USE_SCHED
) {
391 struct elevator_queue
*e
= data
->q
->elevator
;
393 INIT_HLIST_NODE(&rq
->hash
);
394 RB_CLEAR_NODE(&rq
->rb_node
);
396 if (e
->type
->ops
.prepare_request
)
397 e
->type
->ops
.prepare_request(rq
);
403 static inline struct request
*
404 __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data
*data
)
406 unsigned int tag
, tag_offset
;
407 struct blk_mq_tags
*tags
;
409 unsigned long tag_mask
;
412 tag_mask
= blk_mq_get_tags(data
, data
->nr_tags
, &tag_offset
);
413 if (unlikely(!tag_mask
))
416 tags
= blk_mq_tags_from_data(data
);
417 for (i
= 0; tag_mask
; i
++) {
418 if (!(tag_mask
& (1UL << i
)))
420 tag
= tag_offset
+ i
;
421 prefetch(tags
->static_rqs
[tag
]);
422 tag_mask
&= ~(1UL << i
);
423 rq
= blk_mq_rq_ctx_init(data
, tags
, tag
);
424 rq_list_add(data
->cached_rq
, rq
);
427 if (!(data
->rq_flags
& RQF_SCHED_TAGS
))
428 blk_mq_add_active_requests(data
->hctx
, nr
);
429 /* caller already holds a reference, add for remainder */
430 percpu_ref_get_many(&data
->q
->q_usage_counter
, nr
- 1);
433 return rq_list_pop(data
->cached_rq
);
436 static struct request
*__blk_mq_alloc_requests(struct blk_mq_alloc_data
*data
)
438 struct request_queue
*q
= data
->q
;
439 u64 alloc_time_ns
= 0;
443 /* alloc_time includes depth and tag waits */
444 if (blk_queue_rq_alloc_time(q
))
445 alloc_time_ns
= blk_time_get_ns();
447 if (data
->cmd_flags
& REQ_NOWAIT
)
448 data
->flags
|= BLK_MQ_REQ_NOWAIT
;
452 * All requests use scheduler tags when an I/O scheduler is
453 * enabled for the queue.
455 data
->rq_flags
|= RQF_SCHED_TAGS
;
458 * Flush/passthrough requests are special and go directly to the
461 if ((data
->cmd_flags
& REQ_OP_MASK
) != REQ_OP_FLUSH
&&
462 !blk_op_is_passthrough(data
->cmd_flags
)) {
463 struct elevator_mq_ops
*ops
= &q
->elevator
->type
->ops
;
465 WARN_ON_ONCE(data
->flags
& BLK_MQ_REQ_RESERVED
);
467 data
->rq_flags
|= RQF_USE_SCHED
;
468 if (ops
->limit_depth
)
469 ops
->limit_depth(data
->cmd_flags
, data
);
474 data
->ctx
= blk_mq_get_ctx(q
);
475 data
->hctx
= blk_mq_map_queue(q
, data
->cmd_flags
, data
->ctx
);
476 if (!(data
->rq_flags
& RQF_SCHED_TAGS
))
477 blk_mq_tag_busy(data
->hctx
);
479 if (data
->flags
& BLK_MQ_REQ_RESERVED
)
480 data
->rq_flags
|= RQF_RESV
;
483 * Try batched alloc if we want more than 1 tag.
485 if (data
->nr_tags
> 1) {
486 rq
= __blk_mq_alloc_requests_batch(data
);
488 blk_mq_rq_time_init(rq
, alloc_time_ns
);
495 * Waiting allocations only fail because of an inactive hctx. In that
496 * case just retry the hctx assignment and tag allocation as CPU hotplug
497 * should have migrated us to an online CPU by now.
499 tag
= blk_mq_get_tag(data
);
500 if (tag
== BLK_MQ_NO_TAG
) {
501 if (data
->flags
& BLK_MQ_REQ_NOWAIT
)
504 * Give up the CPU and sleep for a random short time to
505 * ensure that thread using a realtime scheduling class
506 * are migrated off the CPU, and thus off the hctx that
513 if (!(data
->rq_flags
& RQF_SCHED_TAGS
))
514 blk_mq_inc_active_requests(data
->hctx
);
515 rq
= blk_mq_rq_ctx_init(data
, blk_mq_tags_from_data(data
), tag
);
516 blk_mq_rq_time_init(rq
, alloc_time_ns
);
520 static struct request
*blk_mq_rq_cache_fill(struct request_queue
*q
,
521 struct blk_plug
*plug
,
523 blk_mq_req_flags_t flags
)
525 struct blk_mq_alloc_data data
= {
529 .nr_tags
= plug
->nr_ios
,
530 .cached_rq
= &plug
->cached_rq
,
534 if (blk_queue_enter(q
, flags
))
539 rq
= __blk_mq_alloc_requests(&data
);
545 static struct request
*blk_mq_alloc_cached_request(struct request_queue
*q
,
547 blk_mq_req_flags_t flags
)
549 struct blk_plug
*plug
= current
->plug
;
555 if (rq_list_empty(plug
->cached_rq
)) {
556 if (plug
->nr_ios
== 1)
558 rq
= blk_mq_rq_cache_fill(q
, plug
, opf
, flags
);
562 rq
= rq_list_peek(&plug
->cached_rq
);
563 if (!rq
|| rq
->q
!= q
)
566 if (blk_mq_get_hctx_type(opf
) != rq
->mq_hctx
->type
)
568 if (op_is_flush(rq
->cmd_flags
) != op_is_flush(opf
))
571 plug
->cached_rq
= rq_list_next(rq
);
572 blk_mq_rq_time_init(rq
, 0);
576 INIT_LIST_HEAD(&rq
->queuelist
);
580 struct request
*blk_mq_alloc_request(struct request_queue
*q
, blk_opf_t opf
,
581 blk_mq_req_flags_t flags
)
585 rq
= blk_mq_alloc_cached_request(q
, opf
, flags
);
587 struct blk_mq_alloc_data data
= {
595 ret
= blk_queue_enter(q
, flags
);
599 rq
= __blk_mq_alloc_requests(&data
);
604 rq
->__sector
= (sector_t
) -1;
605 rq
->bio
= rq
->biotail
= NULL
;
609 return ERR_PTR(-EWOULDBLOCK
);
611 EXPORT_SYMBOL(blk_mq_alloc_request
);
613 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
,
614 blk_opf_t opf
, blk_mq_req_flags_t flags
, unsigned int hctx_idx
)
616 struct blk_mq_alloc_data data
= {
622 u64 alloc_time_ns
= 0;
628 /* alloc_time includes depth and tag waits */
629 if (blk_queue_rq_alloc_time(q
))
630 alloc_time_ns
= blk_time_get_ns();
633 * If the tag allocator sleeps we could get an allocation for a
634 * different hardware context. No need to complicate the low level
635 * allocator for this for the rare use case of a command tied to
638 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)) ||
639 WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_RESERVED
)))
640 return ERR_PTR(-EINVAL
);
642 if (hctx_idx
>= q
->nr_hw_queues
)
643 return ERR_PTR(-EIO
);
645 ret
= blk_queue_enter(q
, flags
);
650 * Check if the hardware context is actually mapped to anything.
651 * If not tell the caller that it should skip this queue.
654 data
.hctx
= xa_load(&q
->hctx_table
, hctx_idx
);
655 if (!blk_mq_hw_queue_mapped(data
.hctx
))
657 cpu
= cpumask_first_and(data
.hctx
->cpumask
, cpu_online_mask
);
658 if (cpu
>= nr_cpu_ids
)
660 data
.ctx
= __blk_mq_get_ctx(q
, cpu
);
663 data
.rq_flags
|= RQF_SCHED_TAGS
;
665 blk_mq_tag_busy(data
.hctx
);
667 if (flags
& BLK_MQ_REQ_RESERVED
)
668 data
.rq_flags
|= RQF_RESV
;
671 tag
= blk_mq_get_tag(&data
);
672 if (tag
== BLK_MQ_NO_TAG
)
674 if (!(data
.rq_flags
& RQF_SCHED_TAGS
))
675 blk_mq_inc_active_requests(data
.hctx
);
676 rq
= blk_mq_rq_ctx_init(&data
, blk_mq_tags_from_data(&data
), tag
);
677 blk_mq_rq_time_init(rq
, alloc_time_ns
);
679 rq
->__sector
= (sector_t
) -1;
680 rq
->bio
= rq
->biotail
= NULL
;
687 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
689 static void blk_mq_finish_request(struct request
*rq
)
691 struct request_queue
*q
= rq
->q
;
693 if (rq
->rq_flags
& RQF_USE_SCHED
) {
694 q
->elevator
->type
->ops
.finish_request(rq
);
696 * For postflush request that may need to be
697 * completed twice, we should clear this flag
698 * to avoid double finish_request() on the rq.
700 rq
->rq_flags
&= ~RQF_USE_SCHED
;
704 static void __blk_mq_free_request(struct request
*rq
)
706 struct request_queue
*q
= rq
->q
;
707 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
708 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
709 const int sched_tag
= rq
->internal_tag
;
711 blk_crypto_free_request(rq
);
712 blk_pm_mark_last_busy(rq
);
715 if (rq
->tag
!= BLK_MQ_NO_TAG
) {
716 blk_mq_dec_active_requests(hctx
);
717 blk_mq_put_tag(hctx
->tags
, ctx
, rq
->tag
);
719 if (sched_tag
!= BLK_MQ_NO_TAG
)
720 blk_mq_put_tag(hctx
->sched_tags
, ctx
, sched_tag
);
721 blk_mq_sched_restart(hctx
);
725 void blk_mq_free_request(struct request
*rq
)
727 struct request_queue
*q
= rq
->q
;
729 blk_mq_finish_request(rq
);
731 if (unlikely(laptop_mode
&& !blk_rq_is_passthrough(rq
)))
732 laptop_io_completion(q
->disk
->bdi
);
736 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
737 if (req_ref_put_and_test(rq
))
738 __blk_mq_free_request(rq
);
740 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
742 void blk_mq_free_plug_rqs(struct blk_plug
*plug
)
746 while ((rq
= rq_list_pop(&plug
->cached_rq
)) != NULL
)
747 blk_mq_free_request(rq
);
750 void blk_dump_rq_flags(struct request
*rq
, char *msg
)
752 printk(KERN_INFO
"%s: dev %s: flags=%llx\n", msg
,
753 rq
->q
->disk
? rq
->q
->disk
->disk_name
: "?",
754 (__force
unsigned long long) rq
->cmd_flags
);
756 printk(KERN_INFO
" sector %llu, nr/cnr %u/%u\n",
757 (unsigned long long)blk_rq_pos(rq
),
758 blk_rq_sectors(rq
), blk_rq_cur_sectors(rq
));
759 printk(KERN_INFO
" bio %p, biotail %p, len %u\n",
760 rq
->bio
, rq
->biotail
, blk_rq_bytes(rq
));
762 EXPORT_SYMBOL(blk_dump_rq_flags
);
764 static void req_bio_endio(struct request
*rq
, struct bio
*bio
,
765 unsigned int nbytes
, blk_status_t error
)
767 if (unlikely(error
)) {
768 bio
->bi_status
= error
;
769 } else if (req_op(rq
) == REQ_OP_ZONE_APPEND
) {
771 * Partial zone append completions cannot be supported as the
772 * BIO fragments may end up not being written sequentially.
774 if (bio
->bi_iter
.bi_size
!= nbytes
)
775 bio
->bi_status
= BLK_STS_IOERR
;
777 bio
->bi_iter
.bi_sector
= rq
->__sector
;
780 bio_advance(bio
, nbytes
);
782 if (unlikely(rq
->rq_flags
& RQF_QUIET
))
783 bio_set_flag(bio
, BIO_QUIET
);
784 /* don't actually finish bio if it's part of flush sequence */
785 if (bio
->bi_iter
.bi_size
== 0 && !(rq
->rq_flags
& RQF_FLUSH_SEQ
))
789 static void blk_account_io_completion(struct request
*req
, unsigned int bytes
)
791 if (req
->part
&& blk_do_io_stat(req
)) {
792 const int sgrp
= op_stat_group(req_op(req
));
795 part_stat_add(req
->part
, sectors
[sgrp
], bytes
>> 9);
800 static void blk_print_req_error(struct request
*req
, blk_status_t status
)
802 printk_ratelimited(KERN_ERR
803 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
804 "phys_seg %u prio class %u\n",
805 blk_status_to_str(status
),
806 req
->q
->disk
? req
->q
->disk
->disk_name
: "?",
807 blk_rq_pos(req
), (__force u32
)req_op(req
),
808 blk_op_str(req_op(req
)),
809 (__force u32
)(req
->cmd_flags
& ~REQ_OP_MASK
),
810 req
->nr_phys_segments
,
811 IOPRIO_PRIO_CLASS(req
->ioprio
));
815 * Fully end IO on a request. Does not support partial completions, or
818 static void blk_complete_request(struct request
*req
)
820 const bool is_flush
= (req
->rq_flags
& RQF_FLUSH_SEQ
) != 0;
821 int total_bytes
= blk_rq_bytes(req
);
822 struct bio
*bio
= req
->bio
;
824 trace_block_rq_complete(req
, BLK_STS_OK
, total_bytes
);
829 #ifdef CONFIG_BLK_DEV_INTEGRITY
830 if (blk_integrity_rq(req
) && req_op(req
) == REQ_OP_READ
)
831 req
->q
->integrity
.profile
->complete_fn(req
, total_bytes
);
835 * Upper layers may call blk_crypto_evict_key() anytime after the last
836 * bio_endio(). Therefore, the keyslot must be released before that.
838 blk_crypto_rq_put_keyslot(req
);
840 blk_account_io_completion(req
, total_bytes
);
843 struct bio
*next
= bio
->bi_next
;
845 /* Completion has already been traced */
846 bio_clear_flag(bio
, BIO_TRACE_COMPLETION
);
848 if (req_op(req
) == REQ_OP_ZONE_APPEND
)
849 bio
->bi_iter
.bi_sector
= req
->__sector
;
857 * Reset counters so that the request stacking driver
858 * can find how many bytes remain in the request
868 * blk_update_request - Complete multiple bytes without completing the request
869 * @req: the request being processed
870 * @error: block status code
871 * @nr_bytes: number of bytes to complete for @req
874 * Ends I/O on a number of bytes attached to @req, but doesn't complete
875 * the request structure even if @req doesn't have leftover.
876 * If @req has leftover, sets it up for the next range of segments.
878 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
879 * %false return from this function.
882 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
883 * except in the consistency check at the end of this function.
886 * %false - this request doesn't have any more data
887 * %true - this request has more data
889 bool blk_update_request(struct request
*req
, blk_status_t error
,
890 unsigned int nr_bytes
)
894 trace_block_rq_complete(req
, error
, nr_bytes
);
899 #ifdef CONFIG_BLK_DEV_INTEGRITY
900 if (blk_integrity_rq(req
) && req_op(req
) == REQ_OP_READ
&&
902 req
->q
->integrity
.profile
->complete_fn(req
, nr_bytes
);
906 * Upper layers may call blk_crypto_evict_key() anytime after the last
907 * bio_endio(). Therefore, the keyslot must be released before that.
909 if (blk_crypto_rq_has_keyslot(req
) && nr_bytes
>= blk_rq_bytes(req
))
910 __blk_crypto_rq_put_keyslot(req
);
912 if (unlikely(error
&& !blk_rq_is_passthrough(req
) &&
913 !(req
->rq_flags
& RQF_QUIET
)) &&
914 !test_bit(GD_DEAD
, &req
->q
->disk
->state
)) {
915 blk_print_req_error(req
, error
);
916 trace_block_rq_error(req
, error
, nr_bytes
);
919 blk_account_io_completion(req
, nr_bytes
);
923 struct bio
*bio
= req
->bio
;
924 unsigned bio_bytes
= min(bio
->bi_iter
.bi_size
, nr_bytes
);
926 if (bio_bytes
== bio
->bi_iter
.bi_size
)
927 req
->bio
= bio
->bi_next
;
929 /* Completion has already been traced */
930 bio_clear_flag(bio
, BIO_TRACE_COMPLETION
);
931 req_bio_endio(req
, bio
, bio_bytes
, error
);
933 total_bytes
+= bio_bytes
;
934 nr_bytes
-= bio_bytes
;
945 * Reset counters so that the request stacking driver
946 * can find how many bytes remain in the request
953 req
->__data_len
-= total_bytes
;
955 /* update sector only for requests with clear definition of sector */
956 if (!blk_rq_is_passthrough(req
))
957 req
->__sector
+= total_bytes
>> 9;
959 /* mixed attributes always follow the first bio */
960 if (req
->rq_flags
& RQF_MIXED_MERGE
) {
961 req
->cmd_flags
&= ~REQ_FAILFAST_MASK
;
962 req
->cmd_flags
|= req
->bio
->bi_opf
& REQ_FAILFAST_MASK
;
965 if (!(req
->rq_flags
& RQF_SPECIAL_PAYLOAD
)) {
967 * If total number of sectors is less than the first segment
968 * size, something has gone terribly wrong.
970 if (blk_rq_bytes(req
) < blk_rq_cur_bytes(req
)) {
971 blk_dump_rq_flags(req
, "request botched");
972 req
->__data_len
= blk_rq_cur_bytes(req
);
975 /* recalculate the number of segments */
976 req
->nr_phys_segments
= blk_recalc_rq_segments(req
);
981 EXPORT_SYMBOL_GPL(blk_update_request
);
983 static inline void blk_account_io_done(struct request
*req
, u64 now
)
985 trace_block_io_done(req
);
988 * Account IO completion. flush_rq isn't accounted as a
989 * normal IO on queueing nor completion. Accounting the
990 * containing request is enough.
992 if (blk_do_io_stat(req
) && req
->part
&&
993 !(req
->rq_flags
& RQF_FLUSH_SEQ
)) {
994 const int sgrp
= op_stat_group(req_op(req
));
997 update_io_ticks(req
->part
, jiffies
, true);
998 part_stat_inc(req
->part
, ios
[sgrp
]);
999 part_stat_add(req
->part
, nsecs
[sgrp
], now
- req
->start_time_ns
);
1004 static inline void blk_account_io_start(struct request
*req
)
1006 trace_block_io_start(req
);
1008 if (blk_do_io_stat(req
)) {
1010 * All non-passthrough requests are created from a bio with one
1011 * exception: when a flush command that is part of a flush sequence
1012 * generated by the state machine in blk-flush.c is cloned onto the
1013 * lower device by dm-multipath we can get here without a bio.
1016 req
->part
= req
->bio
->bi_bdev
;
1018 req
->part
= req
->q
->disk
->part0
;
1021 update_io_ticks(req
->part
, jiffies
, false);
1026 static inline void __blk_mq_end_request_acct(struct request
*rq
, u64 now
)
1028 if (rq
->rq_flags
& RQF_STATS
)
1029 blk_stat_add(rq
, now
);
1031 blk_mq_sched_completed_request(rq
, now
);
1032 blk_account_io_done(rq
, now
);
1035 inline void __blk_mq_end_request(struct request
*rq
, blk_status_t error
)
1037 if (blk_mq_need_time_stamp(rq
))
1038 __blk_mq_end_request_acct(rq
, blk_time_get_ns());
1040 blk_mq_finish_request(rq
);
1043 rq_qos_done(rq
->q
, rq
);
1044 if (rq
->end_io(rq
, error
) == RQ_END_IO_FREE
)
1045 blk_mq_free_request(rq
);
1047 blk_mq_free_request(rq
);
1050 EXPORT_SYMBOL(__blk_mq_end_request
);
1052 void blk_mq_end_request(struct request
*rq
, blk_status_t error
)
1054 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
1056 __blk_mq_end_request(rq
, error
);
1058 EXPORT_SYMBOL(blk_mq_end_request
);
1060 #define TAG_COMP_BATCH 32
1062 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx
*hctx
,
1063 int *tag_array
, int nr_tags
)
1065 struct request_queue
*q
= hctx
->queue
;
1067 blk_mq_sub_active_requests(hctx
, nr_tags
);
1069 blk_mq_put_tags(hctx
->tags
, tag_array
, nr_tags
);
1070 percpu_ref_put_many(&q
->q_usage_counter
, nr_tags
);
1073 void blk_mq_end_request_batch(struct io_comp_batch
*iob
)
1075 int tags
[TAG_COMP_BATCH
], nr_tags
= 0;
1076 struct blk_mq_hw_ctx
*cur_hctx
= NULL
;
1081 now
= blk_time_get_ns();
1083 while ((rq
= rq_list_pop(&iob
->req_list
)) != NULL
) {
1085 prefetch(rq
->rq_next
);
1087 blk_complete_request(rq
);
1089 __blk_mq_end_request_acct(rq
, now
);
1091 blk_mq_finish_request(rq
);
1093 rq_qos_done(rq
->q
, rq
);
1096 * If end_io handler returns NONE, then it still has
1097 * ownership of the request.
1099 if (rq
->end_io
&& rq
->end_io(rq
, 0) == RQ_END_IO_NONE
)
1102 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
1103 if (!req_ref_put_and_test(rq
))
1106 blk_crypto_free_request(rq
);
1107 blk_pm_mark_last_busy(rq
);
1109 if (nr_tags
== TAG_COMP_BATCH
|| cur_hctx
!= rq
->mq_hctx
) {
1111 blk_mq_flush_tag_batch(cur_hctx
, tags
, nr_tags
);
1113 cur_hctx
= rq
->mq_hctx
;
1115 tags
[nr_tags
++] = rq
->tag
;
1119 blk_mq_flush_tag_batch(cur_hctx
, tags
, nr_tags
);
1121 EXPORT_SYMBOL_GPL(blk_mq_end_request_batch
);
1123 static void blk_complete_reqs(struct llist_head
*list
)
1125 struct llist_node
*entry
= llist_reverse_order(llist_del_all(list
));
1126 struct request
*rq
, *next
;
1128 llist_for_each_entry_safe(rq
, next
, entry
, ipi_list
)
1129 rq
->q
->mq_ops
->complete(rq
);
1132 static __latent_entropy
void blk_done_softirq(struct softirq_action
*h
)
1134 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done
));
1137 static int blk_softirq_cpu_dead(unsigned int cpu
)
1139 blk_complete_reqs(&per_cpu(blk_cpu_done
, cpu
));
1143 static void __blk_mq_complete_request_remote(void *data
)
1145 __raise_softirq_irqoff(BLOCK_SOFTIRQ
);
1148 static inline bool blk_mq_complete_need_ipi(struct request
*rq
)
1150 int cpu
= raw_smp_processor_id();
1152 if (!IS_ENABLED(CONFIG_SMP
) ||
1153 !test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
))
1156 * With force threaded interrupts enabled, raising softirq from an SMP
1157 * function call will always result in waking the ksoftirqd thread.
1158 * This is probably worse than completing the request on a different
1161 if (force_irqthreads())
1164 /* same CPU or cache domain and capacity? Complete locally */
1165 if (cpu
== rq
->mq_ctx
->cpu
||
1166 (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
) &&
1167 cpus_share_cache(cpu
, rq
->mq_ctx
->cpu
) &&
1168 cpus_equal_capacity(cpu
, rq
->mq_ctx
->cpu
)))
1171 /* don't try to IPI to an offline CPU */
1172 return cpu_online(rq
->mq_ctx
->cpu
);
1175 static void blk_mq_complete_send_ipi(struct request
*rq
)
1179 cpu
= rq
->mq_ctx
->cpu
;
1180 if (llist_add(&rq
->ipi_list
, &per_cpu(blk_cpu_done
, cpu
)))
1181 smp_call_function_single_async(cpu
, &per_cpu(blk_cpu_csd
, cpu
));
1184 static void blk_mq_raise_softirq(struct request
*rq
)
1186 struct llist_head
*list
;
1189 list
= this_cpu_ptr(&blk_cpu_done
);
1190 if (llist_add(&rq
->ipi_list
, list
))
1191 raise_softirq(BLOCK_SOFTIRQ
);
1195 bool blk_mq_complete_request_remote(struct request
*rq
)
1197 WRITE_ONCE(rq
->state
, MQ_RQ_COMPLETE
);
1200 * For request which hctx has only one ctx mapping,
1201 * or a polled request, always complete locally,
1202 * it's pointless to redirect the completion.
1204 if ((rq
->mq_hctx
->nr_ctx
== 1 &&
1205 rq
->mq_ctx
->cpu
== raw_smp_processor_id()) ||
1206 rq
->cmd_flags
& REQ_POLLED
)
1209 if (blk_mq_complete_need_ipi(rq
)) {
1210 blk_mq_complete_send_ipi(rq
);
1214 if (rq
->q
->nr_hw_queues
== 1) {
1215 blk_mq_raise_softirq(rq
);
1220 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote
);
1223 * blk_mq_complete_request - end I/O on a request
1224 * @rq: the request being processed
1227 * Complete a request by scheduling the ->complete_rq operation.
1229 void blk_mq_complete_request(struct request
*rq
)
1231 if (!blk_mq_complete_request_remote(rq
))
1232 rq
->q
->mq_ops
->complete(rq
);
1234 EXPORT_SYMBOL(blk_mq_complete_request
);
1237 * blk_mq_start_request - Start processing a request
1238 * @rq: Pointer to request to be started
1240 * Function used by device drivers to notify the block layer that a request
1241 * is going to be processed now, so blk layer can do proper initializations
1242 * such as starting the timeout timer.
1244 void blk_mq_start_request(struct request
*rq
)
1246 struct request_queue
*q
= rq
->q
;
1248 trace_block_rq_issue(rq
);
1250 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
) &&
1251 !blk_rq_is_passthrough(rq
)) {
1252 rq
->io_start_time_ns
= blk_time_get_ns();
1253 rq
->stats_sectors
= blk_rq_sectors(rq
);
1254 rq
->rq_flags
|= RQF_STATS
;
1255 rq_qos_issue(q
, rq
);
1258 WARN_ON_ONCE(blk_mq_rq_state(rq
) != MQ_RQ_IDLE
);
1261 WRITE_ONCE(rq
->state
, MQ_RQ_IN_FLIGHT
);
1262 rq
->mq_hctx
->tags
->rqs
[rq
->tag
] = rq
;
1264 #ifdef CONFIG_BLK_DEV_INTEGRITY
1265 if (blk_integrity_rq(rq
) && req_op(rq
) == REQ_OP_WRITE
)
1266 q
->integrity
.profile
->prepare_fn(rq
);
1268 if (rq
->bio
&& rq
->bio
->bi_opf
& REQ_POLLED
)
1269 WRITE_ONCE(rq
->bio
->bi_cookie
, rq
->mq_hctx
->queue_num
);
1271 EXPORT_SYMBOL(blk_mq_start_request
);
1274 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1275 * queues. This is important for md arrays to benefit from merging
1278 static inline unsigned short blk_plug_max_rq_count(struct blk_plug
*plug
)
1280 if (plug
->multiple_queues
)
1281 return BLK_MAX_REQUEST_COUNT
* 2;
1282 return BLK_MAX_REQUEST_COUNT
;
1285 static void blk_add_rq_to_plug(struct blk_plug
*plug
, struct request
*rq
)
1287 struct request
*last
= rq_list_peek(&plug
->mq_list
);
1289 if (!plug
->rq_count
) {
1290 trace_block_plug(rq
->q
);
1291 } else if (plug
->rq_count
>= blk_plug_max_rq_count(plug
) ||
1292 (!blk_queue_nomerges(rq
->q
) &&
1293 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1294 blk_mq_flush_plug_list(plug
, false);
1296 trace_block_plug(rq
->q
);
1299 if (!plug
->multiple_queues
&& last
&& last
->q
!= rq
->q
)
1300 plug
->multiple_queues
= true;
1302 * Any request allocated from sched tags can't be issued to
1303 * ->queue_rqs() directly
1305 if (!plug
->has_elevator
&& (rq
->rq_flags
& RQF_SCHED_TAGS
))
1306 plug
->has_elevator
= true;
1308 rq_list_add(&plug
->mq_list
, rq
);
1313 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1314 * @rq: request to insert
1315 * @at_head: insert request at head or tail of queue
1318 * Insert a fully prepared request at the back of the I/O scheduler queue
1319 * for execution. Don't wait for completion.
1322 * This function will invoke @done directly if the queue is dead.
1324 void blk_execute_rq_nowait(struct request
*rq
, bool at_head
)
1326 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1328 WARN_ON(irqs_disabled());
1329 WARN_ON(!blk_rq_is_passthrough(rq
));
1331 blk_account_io_start(rq
);
1334 * As plugging can be enabled for passthrough requests on a zoned
1335 * device, directly accessing the plug instead of using blk_mq_plug()
1336 * should not have any consequences.
1338 if (current
->plug
&& !at_head
) {
1339 blk_add_rq_to_plug(current
->plug
, rq
);
1343 blk_mq_insert_request(rq
, at_head
? BLK_MQ_INSERT_AT_HEAD
: 0);
1344 blk_mq_run_hw_queue(hctx
, hctx
->flags
& BLK_MQ_F_BLOCKING
);
1346 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait
);
1348 struct blk_rq_wait
{
1349 struct completion done
;
1353 static enum rq_end_io_ret
blk_end_sync_rq(struct request
*rq
, blk_status_t ret
)
1355 struct blk_rq_wait
*wait
= rq
->end_io_data
;
1358 complete(&wait
->done
);
1359 return RQ_END_IO_NONE
;
1362 bool blk_rq_is_poll(struct request
*rq
)
1366 if (rq
->mq_hctx
->type
!= HCTX_TYPE_POLL
)
1370 EXPORT_SYMBOL_GPL(blk_rq_is_poll
);
1372 static void blk_rq_poll_completion(struct request
*rq
, struct completion
*wait
)
1375 blk_hctx_poll(rq
->q
, rq
->mq_hctx
, NULL
, 0);
1377 } while (!completion_done(wait
));
1381 * blk_execute_rq - insert a request into queue for execution
1382 * @rq: request to insert
1383 * @at_head: insert request at head or tail of queue
1386 * Insert a fully prepared request at the back of the I/O scheduler queue
1387 * for execution and wait for completion.
1388 * Return: The blk_status_t result provided to blk_mq_end_request().
1390 blk_status_t
blk_execute_rq(struct request
*rq
, bool at_head
)
1392 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1393 struct blk_rq_wait wait
= {
1394 .done
= COMPLETION_INITIALIZER_ONSTACK(wait
.done
),
1397 WARN_ON(irqs_disabled());
1398 WARN_ON(!blk_rq_is_passthrough(rq
));
1400 rq
->end_io_data
= &wait
;
1401 rq
->end_io
= blk_end_sync_rq
;
1403 blk_account_io_start(rq
);
1404 blk_mq_insert_request(rq
, at_head
? BLK_MQ_INSERT_AT_HEAD
: 0);
1405 blk_mq_run_hw_queue(hctx
, false);
1407 if (blk_rq_is_poll(rq
))
1408 blk_rq_poll_completion(rq
, &wait
.done
);
1410 blk_wait_io(&wait
.done
);
1414 EXPORT_SYMBOL(blk_execute_rq
);
1416 static void __blk_mq_requeue_request(struct request
*rq
)
1418 struct request_queue
*q
= rq
->q
;
1420 blk_mq_put_driver_tag(rq
);
1422 trace_block_rq_requeue(rq
);
1423 rq_qos_requeue(q
, rq
);
1425 if (blk_mq_request_started(rq
)) {
1426 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
1427 rq
->rq_flags
&= ~RQF_TIMED_OUT
;
1431 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
1433 struct request_queue
*q
= rq
->q
;
1434 unsigned long flags
;
1436 __blk_mq_requeue_request(rq
);
1438 /* this request will be re-inserted to io scheduler queue */
1439 blk_mq_sched_requeue_request(rq
);
1441 spin_lock_irqsave(&q
->requeue_lock
, flags
);
1442 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
1443 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
1445 if (kick_requeue_list
)
1446 blk_mq_kick_requeue_list(q
);
1448 EXPORT_SYMBOL(blk_mq_requeue_request
);
1450 static void blk_mq_requeue_work(struct work_struct
*work
)
1452 struct request_queue
*q
=
1453 container_of(work
, struct request_queue
, requeue_work
.work
);
1455 LIST_HEAD(flush_list
);
1458 spin_lock_irq(&q
->requeue_lock
);
1459 list_splice_init(&q
->requeue_list
, &rq_list
);
1460 list_splice_init(&q
->flush_list
, &flush_list
);
1461 spin_unlock_irq(&q
->requeue_lock
);
1463 while (!list_empty(&rq_list
)) {
1464 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
1466 * If RQF_DONTPREP ist set, the request has been started by the
1467 * driver already and might have driver-specific data allocated
1468 * already. Insert it into the hctx dispatch list to avoid
1469 * block layer merges for the request.
1471 if (rq
->rq_flags
& RQF_DONTPREP
) {
1472 list_del_init(&rq
->queuelist
);
1473 blk_mq_request_bypass_insert(rq
, 0);
1475 list_del_init(&rq
->queuelist
);
1476 blk_mq_insert_request(rq
, BLK_MQ_INSERT_AT_HEAD
);
1480 while (!list_empty(&flush_list
)) {
1481 rq
= list_entry(flush_list
.next
, struct request
, queuelist
);
1482 list_del_init(&rq
->queuelist
);
1483 blk_mq_insert_request(rq
, 0);
1486 blk_mq_run_hw_queues(q
, false);
1489 void blk_mq_kick_requeue_list(struct request_queue
*q
)
1491 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
, 0);
1493 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
1495 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
1496 unsigned long msecs
)
1498 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
,
1499 msecs_to_jiffies(msecs
));
1501 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
1503 static bool blk_is_flush_data_rq(struct request
*rq
)
1505 return (rq
->rq_flags
& RQF_FLUSH_SEQ
) && !is_flush_rq(rq
);
1508 static bool blk_mq_rq_inflight(struct request
*rq
, void *priv
)
1511 * If we find a request that isn't idle we know the queue is busy
1512 * as it's checked in the iter.
1513 * Return false to stop the iteration.
1515 * In case of queue quiesce, if one flush data request is completed,
1516 * don't count it as inflight given the flush sequence is suspended,
1517 * and the original flush data request is invisible to driver, just
1518 * like other pending requests because of quiesce
1520 if (blk_mq_request_started(rq
) && !(blk_queue_quiesced(rq
->q
) &&
1521 blk_is_flush_data_rq(rq
) &&
1522 blk_mq_request_completed(rq
))) {
1532 bool blk_mq_queue_inflight(struct request_queue
*q
)
1536 blk_mq_queue_tag_busy_iter(q
, blk_mq_rq_inflight
, &busy
);
1539 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight
);
1541 static void blk_mq_rq_timed_out(struct request
*req
)
1543 req
->rq_flags
|= RQF_TIMED_OUT
;
1544 if (req
->q
->mq_ops
->timeout
) {
1545 enum blk_eh_timer_return ret
;
1547 ret
= req
->q
->mq_ops
->timeout(req
);
1548 if (ret
== BLK_EH_DONE
)
1550 WARN_ON_ONCE(ret
!= BLK_EH_RESET_TIMER
);
1556 struct blk_expired_data
{
1557 bool has_timedout_rq
;
1559 unsigned long timeout_start
;
1562 static bool blk_mq_req_expired(struct request
*rq
, struct blk_expired_data
*expired
)
1564 unsigned long deadline
;
1566 if (blk_mq_rq_state(rq
) != MQ_RQ_IN_FLIGHT
)
1568 if (rq
->rq_flags
& RQF_TIMED_OUT
)
1571 deadline
= READ_ONCE(rq
->deadline
);
1572 if (time_after_eq(expired
->timeout_start
, deadline
))
1575 if (expired
->next
== 0)
1576 expired
->next
= deadline
;
1577 else if (time_after(expired
->next
, deadline
))
1578 expired
->next
= deadline
;
1582 void blk_mq_put_rq_ref(struct request
*rq
)
1584 if (is_flush_rq(rq
)) {
1585 if (rq
->end_io(rq
, 0) == RQ_END_IO_FREE
)
1586 blk_mq_free_request(rq
);
1587 } else if (req_ref_put_and_test(rq
)) {
1588 __blk_mq_free_request(rq
);
1592 static bool blk_mq_check_expired(struct request
*rq
, void *priv
)
1594 struct blk_expired_data
*expired
= priv
;
1597 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1598 * be reallocated underneath the timeout handler's processing, then
1599 * the expire check is reliable. If the request is not expired, then
1600 * it was completed and reallocated as a new request after returning
1601 * from blk_mq_check_expired().
1603 if (blk_mq_req_expired(rq
, expired
)) {
1604 expired
->has_timedout_rq
= true;
1610 static bool blk_mq_handle_expired(struct request
*rq
, void *priv
)
1612 struct blk_expired_data
*expired
= priv
;
1614 if (blk_mq_req_expired(rq
, expired
))
1615 blk_mq_rq_timed_out(rq
);
1619 static void blk_mq_timeout_work(struct work_struct
*work
)
1621 struct request_queue
*q
=
1622 container_of(work
, struct request_queue
, timeout_work
);
1623 struct blk_expired_data expired
= {
1624 .timeout_start
= jiffies
,
1626 struct blk_mq_hw_ctx
*hctx
;
1629 /* A deadlock might occur if a request is stuck requiring a
1630 * timeout at the same time a queue freeze is waiting
1631 * completion, since the timeout code would not be able to
1632 * acquire the queue reference here.
1634 * That's why we don't use blk_queue_enter here; instead, we use
1635 * percpu_ref_tryget directly, because we need to be able to
1636 * obtain a reference even in the short window between the queue
1637 * starting to freeze, by dropping the first reference in
1638 * blk_freeze_queue_start, and the moment the last request is
1639 * consumed, marked by the instant q_usage_counter reaches
1642 if (!percpu_ref_tryget(&q
->q_usage_counter
))
1645 /* check if there is any timed-out request */
1646 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &expired
);
1647 if (expired
.has_timedout_rq
) {
1649 * Before walking tags, we must ensure any submit started
1650 * before the current time has finished. Since the submit
1651 * uses srcu or rcu, wait for a synchronization point to
1652 * ensure all running submits have finished
1654 blk_mq_wait_quiesce_done(q
->tag_set
);
1657 blk_mq_queue_tag_busy_iter(q
, blk_mq_handle_expired
, &expired
);
1660 if (expired
.next
!= 0) {
1661 mod_timer(&q
->timeout
, expired
.next
);
1664 * Request timeouts are handled as a forward rolling timer. If
1665 * we end up here it means that no requests are pending and
1666 * also that no request has been pending for a while. Mark
1667 * each hctx as idle.
1669 queue_for_each_hw_ctx(q
, hctx
, i
) {
1670 /* the hctx may be unmapped, so check it here */
1671 if (blk_mq_hw_queue_mapped(hctx
))
1672 blk_mq_tag_idle(hctx
);
1678 struct flush_busy_ctx_data
{
1679 struct blk_mq_hw_ctx
*hctx
;
1680 struct list_head
*list
;
1683 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
1685 struct flush_busy_ctx_data
*flush_data
= data
;
1686 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
1687 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
1688 enum hctx_type type
= hctx
->type
;
1690 spin_lock(&ctx
->lock
);
1691 list_splice_tail_init(&ctx
->rq_lists
[type
], flush_data
->list
);
1692 sbitmap_clear_bit(sb
, bitnr
);
1693 spin_unlock(&ctx
->lock
);
1698 * Process software queues that have been marked busy, splicing them
1699 * to the for-dispatch
1701 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
1703 struct flush_busy_ctx_data data
= {
1708 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
1710 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
1712 struct dispatch_rq_data
{
1713 struct blk_mq_hw_ctx
*hctx
;
1717 static bool dispatch_rq_from_ctx(struct sbitmap
*sb
, unsigned int bitnr
,
1720 struct dispatch_rq_data
*dispatch_data
= data
;
1721 struct blk_mq_hw_ctx
*hctx
= dispatch_data
->hctx
;
1722 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
1723 enum hctx_type type
= hctx
->type
;
1725 spin_lock(&ctx
->lock
);
1726 if (!list_empty(&ctx
->rq_lists
[type
])) {
1727 dispatch_data
->rq
= list_entry_rq(ctx
->rq_lists
[type
].next
);
1728 list_del_init(&dispatch_data
->rq
->queuelist
);
1729 if (list_empty(&ctx
->rq_lists
[type
]))
1730 sbitmap_clear_bit(sb
, bitnr
);
1732 spin_unlock(&ctx
->lock
);
1734 return !dispatch_data
->rq
;
1737 struct request
*blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx
*hctx
,
1738 struct blk_mq_ctx
*start
)
1740 unsigned off
= start
? start
->index_hw
[hctx
->type
] : 0;
1741 struct dispatch_rq_data data
= {
1746 __sbitmap_for_each_set(&hctx
->ctx_map
, off
,
1747 dispatch_rq_from_ctx
, &data
);
1752 bool __blk_mq_alloc_driver_tag(struct request
*rq
)
1754 struct sbitmap_queue
*bt
= &rq
->mq_hctx
->tags
->bitmap_tags
;
1755 unsigned int tag_offset
= rq
->mq_hctx
->tags
->nr_reserved_tags
;
1758 blk_mq_tag_busy(rq
->mq_hctx
);
1760 if (blk_mq_tag_is_reserved(rq
->mq_hctx
->sched_tags
, rq
->internal_tag
)) {
1761 bt
= &rq
->mq_hctx
->tags
->breserved_tags
;
1764 if (!hctx_may_queue(rq
->mq_hctx
, bt
))
1768 tag
= __sbitmap_queue_get(bt
);
1769 if (tag
== BLK_MQ_NO_TAG
)
1772 rq
->tag
= tag
+ tag_offset
;
1773 blk_mq_inc_active_requests(rq
->mq_hctx
);
1777 static int blk_mq_dispatch_wake(wait_queue_entry_t
*wait
, unsigned mode
,
1778 int flags
, void *key
)
1780 struct blk_mq_hw_ctx
*hctx
;
1782 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
1784 spin_lock(&hctx
->dispatch_wait_lock
);
1785 if (!list_empty(&wait
->entry
)) {
1786 struct sbitmap_queue
*sbq
;
1788 list_del_init(&wait
->entry
);
1789 sbq
= &hctx
->tags
->bitmap_tags
;
1790 atomic_dec(&sbq
->ws_active
);
1792 spin_unlock(&hctx
->dispatch_wait_lock
);
1794 blk_mq_run_hw_queue(hctx
, true);
1799 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1800 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1801 * restart. For both cases, take care to check the condition again after
1802 * marking us as waiting.
1804 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx
*hctx
,
1807 struct sbitmap_queue
*sbq
;
1808 struct wait_queue_head
*wq
;
1809 wait_queue_entry_t
*wait
;
1812 if (!(hctx
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
) &&
1813 !(blk_mq_is_shared_tags(hctx
->flags
))) {
1814 blk_mq_sched_mark_restart_hctx(hctx
);
1817 * It's possible that a tag was freed in the window between the
1818 * allocation failure and adding the hardware queue to the wait
1821 * Don't clear RESTART here, someone else could have set it.
1822 * At most this will cost an extra queue run.
1824 return blk_mq_get_driver_tag(rq
);
1827 wait
= &hctx
->dispatch_wait
;
1828 if (!list_empty_careful(&wait
->entry
))
1831 if (blk_mq_tag_is_reserved(rq
->mq_hctx
->sched_tags
, rq
->internal_tag
))
1832 sbq
= &hctx
->tags
->breserved_tags
;
1834 sbq
= &hctx
->tags
->bitmap_tags
;
1835 wq
= &bt_wait_ptr(sbq
, hctx
)->wait
;
1837 spin_lock_irq(&wq
->lock
);
1838 spin_lock(&hctx
->dispatch_wait_lock
);
1839 if (!list_empty(&wait
->entry
)) {
1840 spin_unlock(&hctx
->dispatch_wait_lock
);
1841 spin_unlock_irq(&wq
->lock
);
1845 atomic_inc(&sbq
->ws_active
);
1846 wait
->flags
&= ~WQ_FLAG_EXCLUSIVE
;
1847 __add_wait_queue(wq
, wait
);
1850 * Add one explicit barrier since blk_mq_get_driver_tag() may
1851 * not imply barrier in case of failure.
1853 * Order adding us to wait queue and allocating driver tag.
1855 * The pair is the one implied in sbitmap_queue_wake_up() which
1856 * orders clearing sbitmap tag bits and waitqueue_active() in
1857 * __sbitmap_queue_wake_up(), since waitqueue_active() is lockless
1859 * Otherwise, re-order of adding wait queue and getting driver tag
1860 * may cause __sbitmap_queue_wake_up() to wake up nothing because
1861 * the waitqueue_active() may not observe us in wait queue.
1866 * It's possible that a tag was freed in the window between the
1867 * allocation failure and adding the hardware queue to the wait
1870 ret
= blk_mq_get_driver_tag(rq
);
1872 spin_unlock(&hctx
->dispatch_wait_lock
);
1873 spin_unlock_irq(&wq
->lock
);
1878 * We got a tag, remove ourselves from the wait queue to ensure
1879 * someone else gets the wakeup.
1881 list_del_init(&wait
->entry
);
1882 atomic_dec(&sbq
->ws_active
);
1883 spin_unlock(&hctx
->dispatch_wait_lock
);
1884 spin_unlock_irq(&wq
->lock
);
1889 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1890 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1892 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1893 * - EWMA is one simple way to compute running average value
1894 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1895 * - take 4 as factor for avoiding to get too small(0) result, and this
1896 * factor doesn't matter because EWMA decreases exponentially
1898 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx
*hctx
, bool busy
)
1902 ewma
= hctx
->dispatch_busy
;
1907 ewma
*= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
- 1;
1909 ewma
+= 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR
;
1910 ewma
/= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
;
1912 hctx
->dispatch_busy
= ewma
;
1915 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1917 static void blk_mq_handle_dev_resource(struct request
*rq
,
1918 struct list_head
*list
)
1920 list_add(&rq
->queuelist
, list
);
1921 __blk_mq_requeue_request(rq
);
1924 static void blk_mq_handle_zone_resource(struct request
*rq
,
1925 struct list_head
*zone_list
)
1928 * If we end up here it is because we cannot dispatch a request to a
1929 * specific zone due to LLD level zone-write locking or other zone
1930 * related resource not being available. In this case, set the request
1931 * aside in zone_list for retrying it later.
1933 list_add(&rq
->queuelist
, zone_list
);
1934 __blk_mq_requeue_request(rq
);
1937 enum prep_dispatch
{
1939 PREP_DISPATCH_NO_TAG
,
1940 PREP_DISPATCH_NO_BUDGET
,
1943 static enum prep_dispatch
blk_mq_prep_dispatch_rq(struct request
*rq
,
1946 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1947 int budget_token
= -1;
1950 budget_token
= blk_mq_get_dispatch_budget(rq
->q
);
1951 if (budget_token
< 0) {
1952 blk_mq_put_driver_tag(rq
);
1953 return PREP_DISPATCH_NO_BUDGET
;
1955 blk_mq_set_rq_budget_token(rq
, budget_token
);
1958 if (!blk_mq_get_driver_tag(rq
)) {
1960 * The initial allocation attempt failed, so we need to
1961 * rerun the hardware queue when a tag is freed. The
1962 * waitqueue takes care of that. If the queue is run
1963 * before we add this entry back on the dispatch list,
1964 * we'll re-run it below.
1966 if (!blk_mq_mark_tag_wait(hctx
, rq
)) {
1968 * All budgets not got from this function will be put
1969 * together during handling partial dispatch
1972 blk_mq_put_dispatch_budget(rq
->q
, budget_token
);
1973 return PREP_DISPATCH_NO_TAG
;
1977 return PREP_DISPATCH_OK
;
1980 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1981 static void blk_mq_release_budgets(struct request_queue
*q
,
1982 struct list_head
*list
)
1986 list_for_each_entry(rq
, list
, queuelist
) {
1987 int budget_token
= blk_mq_get_rq_budget_token(rq
);
1989 if (budget_token
>= 0)
1990 blk_mq_put_dispatch_budget(q
, budget_token
);
1995 * blk_mq_commit_rqs will notify driver using bd->last that there is no
1996 * more requests. (See comment in struct blk_mq_ops for commit_rqs for
1998 * Attention, we should explicitly call this in unusual cases:
1999 * 1) did not queue everything initially scheduled to queue
2000 * 2) the last attempt to queue a request failed
2002 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx
*hctx
, int queued
,
2005 if (hctx
->queue
->mq_ops
->commit_rqs
&& queued
) {
2006 trace_block_unplug(hctx
->queue
, queued
, !from_schedule
);
2007 hctx
->queue
->mq_ops
->commit_rqs(hctx
);
2012 * Returns true if we did some work AND can potentially do more.
2014 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
,
2015 unsigned int nr_budgets
)
2017 enum prep_dispatch prep
;
2018 struct request_queue
*q
= hctx
->queue
;
2021 blk_status_t ret
= BLK_STS_OK
;
2022 LIST_HEAD(zone_list
);
2023 bool needs_resource
= false;
2025 if (list_empty(list
))
2029 * Now process all the entries, sending them to the driver.
2033 struct blk_mq_queue_data bd
;
2035 rq
= list_first_entry(list
, struct request
, queuelist
);
2037 WARN_ON_ONCE(hctx
!= rq
->mq_hctx
);
2038 prep
= blk_mq_prep_dispatch_rq(rq
, !nr_budgets
);
2039 if (prep
!= PREP_DISPATCH_OK
)
2042 list_del_init(&rq
->queuelist
);
2045 bd
.last
= list_empty(list
);
2048 * once the request is queued to lld, no need to cover the
2053 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
2058 case BLK_STS_RESOURCE
:
2059 needs_resource
= true;
2061 case BLK_STS_DEV_RESOURCE
:
2062 blk_mq_handle_dev_resource(rq
, list
);
2064 case BLK_STS_ZONE_RESOURCE
:
2066 * Move the request to zone_list and keep going through
2067 * the dispatch list to find more requests the drive can
2070 blk_mq_handle_zone_resource(rq
, &zone_list
);
2071 needs_resource
= true;
2074 blk_mq_end_request(rq
, ret
);
2076 } while (!list_empty(list
));
2078 if (!list_empty(&zone_list
))
2079 list_splice_tail_init(&zone_list
, list
);
2081 /* If we didn't flush the entire list, we could have told the driver
2082 * there was more coming, but that turned out to be a lie.
2084 if (!list_empty(list
) || ret
!= BLK_STS_OK
)
2085 blk_mq_commit_rqs(hctx
, queued
, false);
2088 * Any items that need requeuing? Stuff them into hctx->dispatch,
2089 * that is where we will continue on next queue run.
2091 if (!list_empty(list
)) {
2093 /* For non-shared tags, the RESTART check will suffice */
2094 bool no_tag
= prep
== PREP_DISPATCH_NO_TAG
&&
2095 ((hctx
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
) ||
2096 blk_mq_is_shared_tags(hctx
->flags
));
2099 blk_mq_release_budgets(q
, list
);
2101 spin_lock(&hctx
->lock
);
2102 list_splice_tail_init(list
, &hctx
->dispatch
);
2103 spin_unlock(&hctx
->lock
);
2106 * Order adding requests to hctx->dispatch and checking
2107 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2108 * in blk_mq_sched_restart(). Avoid restart code path to
2109 * miss the new added requests to hctx->dispatch, meantime
2110 * SCHED_RESTART is observed here.
2115 * If SCHED_RESTART was set by the caller of this function and
2116 * it is no longer set that means that it was cleared by another
2117 * thread and hence that a queue rerun is needed.
2119 * If 'no_tag' is set, that means that we failed getting
2120 * a driver tag with an I/O scheduler attached. If our dispatch
2121 * waitqueue is no longer active, ensure that we run the queue
2122 * AFTER adding our entries back to the list.
2124 * If no I/O scheduler has been configured it is possible that
2125 * the hardware queue got stopped and restarted before requests
2126 * were pushed back onto the dispatch list. Rerun the queue to
2127 * avoid starvation. Notes:
2128 * - blk_mq_run_hw_queue() checks whether or not a queue has
2129 * been stopped before rerunning a queue.
2130 * - Some but not all block drivers stop a queue before
2131 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2134 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2135 * bit is set, run queue after a delay to avoid IO stalls
2136 * that could otherwise occur if the queue is idle. We'll do
2137 * similar if we couldn't get budget or couldn't lock a zone
2138 * and SCHED_RESTART is set.
2140 needs_restart
= blk_mq_sched_needs_restart(hctx
);
2141 if (prep
== PREP_DISPATCH_NO_BUDGET
)
2142 needs_resource
= true;
2143 if (!needs_restart
||
2144 (no_tag
&& list_empty_careful(&hctx
->dispatch_wait
.entry
)))
2145 blk_mq_run_hw_queue(hctx
, true);
2146 else if (needs_resource
)
2147 blk_mq_delay_run_hw_queue(hctx
, BLK_MQ_RESOURCE_DELAY
);
2149 blk_mq_update_dispatch_busy(hctx
, true);
2153 blk_mq_update_dispatch_busy(hctx
, false);
2157 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx
*hctx
)
2159 int cpu
= cpumask_first_and(hctx
->cpumask
, cpu_online_mask
);
2161 if (cpu
>= nr_cpu_ids
)
2162 cpu
= cpumask_first(hctx
->cpumask
);
2167 * It'd be great if the workqueue API had a way to pass
2168 * in a mask and had some smarts for more clever placement.
2169 * For now we just round-robin here, switching for every
2170 * BLK_MQ_CPU_WORK_BATCH queued items.
2172 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
2175 int next_cpu
= hctx
->next_cpu
;
2177 if (hctx
->queue
->nr_hw_queues
== 1)
2178 return WORK_CPU_UNBOUND
;
2180 if (--hctx
->next_cpu_batch
<= 0) {
2182 next_cpu
= cpumask_next_and(next_cpu
, hctx
->cpumask
,
2184 if (next_cpu
>= nr_cpu_ids
)
2185 next_cpu
= blk_mq_first_mapped_cpu(hctx
);
2186 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2190 * Do unbound schedule if we can't find a online CPU for this hctx,
2191 * and it should only happen in the path of handling CPU DEAD.
2193 if (!cpu_online(next_cpu
)) {
2200 * Make sure to re-select CPU next time once after CPUs
2201 * in hctx->cpumask become online again.
2203 hctx
->next_cpu
= next_cpu
;
2204 hctx
->next_cpu_batch
= 1;
2205 return WORK_CPU_UNBOUND
;
2208 hctx
->next_cpu
= next_cpu
;
2213 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2214 * @hctx: Pointer to the hardware queue to run.
2215 * @msecs: Milliseconds of delay to wait before running the queue.
2217 * Run a hardware queue asynchronously with a delay of @msecs.
2219 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
2221 if (unlikely(blk_mq_hctx_stopped(hctx
)))
2223 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
,
2224 msecs_to_jiffies(msecs
));
2226 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue
);
2229 * blk_mq_run_hw_queue - Start to run a hardware queue.
2230 * @hctx: Pointer to the hardware queue to run.
2231 * @async: If we want to run the queue asynchronously.
2233 * Check if the request queue is not in a quiesced state and if there are
2234 * pending requests to be sent. If this is true, run the queue to send requests
2237 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
2242 * We can't run the queue inline with interrupts disabled.
2244 WARN_ON_ONCE(!async
&& in_interrupt());
2246 might_sleep_if(!async
&& hctx
->flags
& BLK_MQ_F_BLOCKING
);
2249 * When queue is quiesced, we may be switching io scheduler, or
2250 * updating nr_hw_queues, or other things, and we can't run queue
2251 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2253 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2256 __blk_mq_run_dispatch_ops(hctx
->queue
, false,
2257 need_run
= !blk_queue_quiesced(hctx
->queue
) &&
2258 blk_mq_hctx_has_pending(hctx
));
2263 if (async
|| !cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
)) {
2264 blk_mq_delay_run_hw_queue(hctx
, 0);
2268 blk_mq_run_dispatch_ops(hctx
->queue
,
2269 blk_mq_sched_dispatch_requests(hctx
));
2271 EXPORT_SYMBOL(blk_mq_run_hw_queue
);
2274 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2277 static struct blk_mq_hw_ctx
*blk_mq_get_sq_hctx(struct request_queue
*q
)
2279 struct blk_mq_ctx
*ctx
= blk_mq_get_ctx(q
);
2281 * If the IO scheduler does not respect hardware queues when
2282 * dispatching, we just don't bother with multiple HW queues and
2283 * dispatch from hctx for the current CPU since running multiple queues
2284 * just causes lock contention inside the scheduler and pointless cache
2287 struct blk_mq_hw_ctx
*hctx
= ctx
->hctxs
[HCTX_TYPE_DEFAULT
];
2289 if (!blk_mq_hctx_stopped(hctx
))
2295 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2296 * @q: Pointer to the request queue to run.
2297 * @async: If we want to run the queue asynchronously.
2299 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
2301 struct blk_mq_hw_ctx
*hctx
, *sq_hctx
;
2305 if (blk_queue_sq_sched(q
))
2306 sq_hctx
= blk_mq_get_sq_hctx(q
);
2307 queue_for_each_hw_ctx(q
, hctx
, i
) {
2308 if (blk_mq_hctx_stopped(hctx
))
2311 * Dispatch from this hctx either if there's no hctx preferred
2312 * by IO scheduler or if it has requests that bypass the
2315 if (!sq_hctx
|| sq_hctx
== hctx
||
2316 !list_empty_careful(&hctx
->dispatch
))
2317 blk_mq_run_hw_queue(hctx
, async
);
2320 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
2323 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2324 * @q: Pointer to the request queue to run.
2325 * @msecs: Milliseconds of delay to wait before running the queues.
2327 void blk_mq_delay_run_hw_queues(struct request_queue
*q
, unsigned long msecs
)
2329 struct blk_mq_hw_ctx
*hctx
, *sq_hctx
;
2333 if (blk_queue_sq_sched(q
))
2334 sq_hctx
= blk_mq_get_sq_hctx(q
);
2335 queue_for_each_hw_ctx(q
, hctx
, i
) {
2336 if (blk_mq_hctx_stopped(hctx
))
2339 * If there is already a run_work pending, leave the
2340 * pending delay untouched. Otherwise, a hctx can stall
2341 * if another hctx is re-delaying the other's work
2342 * before the work executes.
2344 if (delayed_work_pending(&hctx
->run_work
))
2347 * Dispatch from this hctx either if there's no hctx preferred
2348 * by IO scheduler or if it has requests that bypass the
2351 if (!sq_hctx
|| sq_hctx
== hctx
||
2352 !list_empty_careful(&hctx
->dispatch
))
2353 blk_mq_delay_run_hw_queue(hctx
, msecs
);
2356 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues
);
2359 * This function is often used for pausing .queue_rq() by driver when
2360 * there isn't enough resource or some conditions aren't satisfied, and
2361 * BLK_STS_RESOURCE is usually returned.
2363 * We do not guarantee that dispatch can be drained or blocked
2364 * after blk_mq_stop_hw_queue() returns. Please use
2365 * blk_mq_quiesce_queue() for that requirement.
2367 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
2369 cancel_delayed_work(&hctx
->run_work
);
2371 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
2373 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
2376 * This function is often used for pausing .queue_rq() by driver when
2377 * there isn't enough resource or some conditions aren't satisfied, and
2378 * BLK_STS_RESOURCE is usually returned.
2380 * We do not guarantee that dispatch can be drained or blocked
2381 * after blk_mq_stop_hw_queues() returns. Please use
2382 * blk_mq_quiesce_queue() for that requirement.
2384 void blk_mq_stop_hw_queues(struct request_queue
*q
)
2386 struct blk_mq_hw_ctx
*hctx
;
2389 queue_for_each_hw_ctx(q
, hctx
, i
)
2390 blk_mq_stop_hw_queue(hctx
);
2392 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
2394 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
2396 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
2398 blk_mq_run_hw_queue(hctx
, hctx
->flags
& BLK_MQ_F_BLOCKING
);
2400 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
2402 void blk_mq_start_hw_queues(struct request_queue
*q
)
2404 struct blk_mq_hw_ctx
*hctx
;
2407 queue_for_each_hw_ctx(q
, hctx
, i
)
2408 blk_mq_start_hw_queue(hctx
);
2410 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
2412 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
2414 if (!blk_mq_hctx_stopped(hctx
))
2417 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
2418 blk_mq_run_hw_queue(hctx
, async
);
2420 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
2422 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
2424 struct blk_mq_hw_ctx
*hctx
;
2427 queue_for_each_hw_ctx(q
, hctx
, i
)
2428 blk_mq_start_stopped_hw_queue(hctx
, async
||
2429 (hctx
->flags
& BLK_MQ_F_BLOCKING
));
2431 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
2433 static void blk_mq_run_work_fn(struct work_struct
*work
)
2435 struct blk_mq_hw_ctx
*hctx
=
2436 container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
2438 blk_mq_run_dispatch_ops(hctx
->queue
,
2439 blk_mq_sched_dispatch_requests(hctx
));
2443 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2444 * @rq: Pointer to request to be inserted.
2445 * @flags: BLK_MQ_INSERT_*
2447 * Should only be used carefully, when the caller knows we want to
2448 * bypass a potential IO scheduler on the target device.
2450 static void blk_mq_request_bypass_insert(struct request
*rq
, blk_insert_t flags
)
2452 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
2454 spin_lock(&hctx
->lock
);
2455 if (flags
& BLK_MQ_INSERT_AT_HEAD
)
2456 list_add(&rq
->queuelist
, &hctx
->dispatch
);
2458 list_add_tail(&rq
->queuelist
, &hctx
->dispatch
);
2459 spin_unlock(&hctx
->lock
);
2462 static void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
,
2463 struct blk_mq_ctx
*ctx
, struct list_head
*list
,
2464 bool run_queue_async
)
2467 enum hctx_type type
= hctx
->type
;
2470 * Try to issue requests directly if the hw queue isn't busy to save an
2471 * extra enqueue & dequeue to the sw queue.
2473 if (!hctx
->dispatch_busy
&& !run_queue_async
) {
2474 blk_mq_run_dispatch_ops(hctx
->queue
,
2475 blk_mq_try_issue_list_directly(hctx
, list
));
2476 if (list_empty(list
))
2481 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2484 list_for_each_entry(rq
, list
, queuelist
) {
2485 BUG_ON(rq
->mq_ctx
!= ctx
);
2486 trace_block_rq_insert(rq
);
2487 if (rq
->cmd_flags
& REQ_NOWAIT
)
2488 run_queue_async
= true;
2491 spin_lock(&ctx
->lock
);
2492 list_splice_tail_init(list
, &ctx
->rq_lists
[type
]);
2493 blk_mq_hctx_mark_pending(hctx
, ctx
);
2494 spin_unlock(&ctx
->lock
);
2496 blk_mq_run_hw_queue(hctx
, run_queue_async
);
2499 static void blk_mq_insert_request(struct request
*rq
, blk_insert_t flags
)
2501 struct request_queue
*q
= rq
->q
;
2502 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
2503 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
2505 if (blk_rq_is_passthrough(rq
)) {
2507 * Passthrough request have to be added to hctx->dispatch
2508 * directly. The device may be in a situation where it can't
2509 * handle FS request, and always returns BLK_STS_RESOURCE for
2510 * them, which gets them added to hctx->dispatch.
2512 * If a passthrough request is required to unblock the queues,
2513 * and it is added to the scheduler queue, there is no chance to
2514 * dispatch it given we prioritize requests in hctx->dispatch.
2516 blk_mq_request_bypass_insert(rq
, flags
);
2517 } else if (req_op(rq
) == REQ_OP_FLUSH
) {
2519 * Firstly normal IO request is inserted to scheduler queue or
2520 * sw queue, meantime we add flush request to dispatch queue(
2521 * hctx->dispatch) directly and there is at most one in-flight
2522 * flush request for each hw queue, so it doesn't matter to add
2523 * flush request to tail or front of the dispatch queue.
2525 * Secondly in case of NCQ, flush request belongs to non-NCQ
2526 * command, and queueing it will fail when there is any
2527 * in-flight normal IO request(NCQ command). When adding flush
2528 * rq to the front of hctx->dispatch, it is easier to introduce
2529 * extra time to flush rq's latency because of S_SCHED_RESTART
2530 * compared with adding to the tail of dispatch queue, then
2531 * chance of flush merge is increased, and less flush requests
2532 * will be issued to controller. It is observed that ~10% time
2533 * is saved in blktests block/004 on disk attached to AHCI/NCQ
2534 * drive when adding flush rq to the front of hctx->dispatch.
2536 * Simply queue flush rq to the front of hctx->dispatch so that
2537 * intensive flush workloads can benefit in case of NCQ HW.
2539 blk_mq_request_bypass_insert(rq
, BLK_MQ_INSERT_AT_HEAD
);
2540 } else if (q
->elevator
) {
2543 WARN_ON_ONCE(rq
->tag
!= BLK_MQ_NO_TAG
);
2545 list_add(&rq
->queuelist
, &list
);
2546 q
->elevator
->type
->ops
.insert_requests(hctx
, &list
, flags
);
2548 trace_block_rq_insert(rq
);
2550 spin_lock(&ctx
->lock
);
2551 if (flags
& BLK_MQ_INSERT_AT_HEAD
)
2552 list_add(&rq
->queuelist
, &ctx
->rq_lists
[hctx
->type
]);
2554 list_add_tail(&rq
->queuelist
,
2555 &ctx
->rq_lists
[hctx
->type
]);
2556 blk_mq_hctx_mark_pending(hctx
, ctx
);
2557 spin_unlock(&ctx
->lock
);
2561 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
,
2562 unsigned int nr_segs
)
2566 if (bio
->bi_opf
& REQ_RAHEAD
)
2567 rq
->cmd_flags
|= REQ_FAILFAST_MASK
;
2569 rq
->__sector
= bio
->bi_iter
.bi_sector
;
2570 rq
->write_hint
= bio
->bi_write_hint
;
2571 blk_rq_bio_prep(rq
, bio
, nr_segs
);
2573 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2574 err
= blk_crypto_rq_bio_prep(rq
, bio
, GFP_NOIO
);
2577 blk_account_io_start(rq
);
2580 static blk_status_t
__blk_mq_issue_directly(struct blk_mq_hw_ctx
*hctx
,
2581 struct request
*rq
, bool last
)
2583 struct request_queue
*q
= rq
->q
;
2584 struct blk_mq_queue_data bd
= {
2591 * For OK queue, we are done. For error, caller may kill it.
2592 * Any other error (busy), just add it to our list as we
2593 * previously would have done.
2595 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
2598 blk_mq_update_dispatch_busy(hctx
, false);
2600 case BLK_STS_RESOURCE
:
2601 case BLK_STS_DEV_RESOURCE
:
2602 blk_mq_update_dispatch_busy(hctx
, true);
2603 __blk_mq_requeue_request(rq
);
2606 blk_mq_update_dispatch_busy(hctx
, false);
2613 static bool blk_mq_get_budget_and_tag(struct request
*rq
)
2617 budget_token
= blk_mq_get_dispatch_budget(rq
->q
);
2618 if (budget_token
< 0)
2620 blk_mq_set_rq_budget_token(rq
, budget_token
);
2621 if (!blk_mq_get_driver_tag(rq
)) {
2622 blk_mq_put_dispatch_budget(rq
->q
, budget_token
);
2629 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2630 * @hctx: Pointer of the associated hardware queue.
2631 * @rq: Pointer to request to be sent.
2633 * If the device has enough resources to accept a new request now, send the
2634 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2635 * we can try send it another time in the future. Requests inserted at this
2636 * queue have higher priority.
2638 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
2643 if (blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(rq
->q
)) {
2644 blk_mq_insert_request(rq
, 0);
2648 if ((rq
->rq_flags
& RQF_USE_SCHED
) || !blk_mq_get_budget_and_tag(rq
)) {
2649 blk_mq_insert_request(rq
, 0);
2650 blk_mq_run_hw_queue(hctx
, rq
->cmd_flags
& REQ_NOWAIT
);
2654 ret
= __blk_mq_issue_directly(hctx
, rq
, true);
2658 case BLK_STS_RESOURCE
:
2659 case BLK_STS_DEV_RESOURCE
:
2660 blk_mq_request_bypass_insert(rq
, 0);
2661 blk_mq_run_hw_queue(hctx
, false);
2664 blk_mq_end_request(rq
, ret
);
2669 static blk_status_t
blk_mq_request_issue_directly(struct request
*rq
, bool last
)
2671 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
2673 if (blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(rq
->q
)) {
2674 blk_mq_insert_request(rq
, 0);
2678 if (!blk_mq_get_budget_and_tag(rq
))
2679 return BLK_STS_RESOURCE
;
2680 return __blk_mq_issue_directly(hctx
, rq
, last
);
2683 static void blk_mq_plug_issue_direct(struct blk_plug
*plug
)
2685 struct blk_mq_hw_ctx
*hctx
= NULL
;
2688 blk_status_t ret
= BLK_STS_OK
;
2690 while ((rq
= rq_list_pop(&plug
->mq_list
))) {
2691 bool last
= rq_list_empty(plug
->mq_list
);
2693 if (hctx
!= rq
->mq_hctx
) {
2695 blk_mq_commit_rqs(hctx
, queued
, false);
2701 ret
= blk_mq_request_issue_directly(rq
, last
);
2706 case BLK_STS_RESOURCE
:
2707 case BLK_STS_DEV_RESOURCE
:
2708 blk_mq_request_bypass_insert(rq
, 0);
2709 blk_mq_run_hw_queue(hctx
, false);
2712 blk_mq_end_request(rq
, ret
);
2718 if (ret
!= BLK_STS_OK
)
2719 blk_mq_commit_rqs(hctx
, queued
, false);
2722 static void __blk_mq_flush_plug_list(struct request_queue
*q
,
2723 struct blk_plug
*plug
)
2725 if (blk_queue_quiesced(q
))
2727 q
->mq_ops
->queue_rqs(&plug
->mq_list
);
2730 static void blk_mq_dispatch_plug_list(struct blk_plug
*plug
, bool from_sched
)
2732 struct blk_mq_hw_ctx
*this_hctx
= NULL
;
2733 struct blk_mq_ctx
*this_ctx
= NULL
;
2734 struct request
*requeue_list
= NULL
;
2735 struct request
**requeue_lastp
= &requeue_list
;
2736 unsigned int depth
= 0;
2737 bool is_passthrough
= false;
2741 struct request
*rq
= rq_list_pop(&plug
->mq_list
);
2744 this_hctx
= rq
->mq_hctx
;
2745 this_ctx
= rq
->mq_ctx
;
2746 is_passthrough
= blk_rq_is_passthrough(rq
);
2747 } else if (this_hctx
!= rq
->mq_hctx
|| this_ctx
!= rq
->mq_ctx
||
2748 is_passthrough
!= blk_rq_is_passthrough(rq
)) {
2749 rq_list_add_tail(&requeue_lastp
, rq
);
2752 list_add(&rq
->queuelist
, &list
);
2754 } while (!rq_list_empty(plug
->mq_list
));
2756 plug
->mq_list
= requeue_list
;
2757 trace_block_unplug(this_hctx
->queue
, depth
, !from_sched
);
2759 percpu_ref_get(&this_hctx
->queue
->q_usage_counter
);
2760 /* passthrough requests should never be issued to the I/O scheduler */
2761 if (is_passthrough
) {
2762 spin_lock(&this_hctx
->lock
);
2763 list_splice_tail_init(&list
, &this_hctx
->dispatch
);
2764 spin_unlock(&this_hctx
->lock
);
2765 blk_mq_run_hw_queue(this_hctx
, from_sched
);
2766 } else if (this_hctx
->queue
->elevator
) {
2767 this_hctx
->queue
->elevator
->type
->ops
.insert_requests(this_hctx
,
2769 blk_mq_run_hw_queue(this_hctx
, from_sched
);
2771 blk_mq_insert_requests(this_hctx
, this_ctx
, &list
, from_sched
);
2773 percpu_ref_put(&this_hctx
->queue
->q_usage_counter
);
2776 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
2781 * We may have been called recursively midway through handling
2782 * plug->mq_list via a schedule() in the driver's queue_rq() callback.
2783 * To avoid mq_list changing under our feet, clear rq_count early and
2784 * bail out specifically if rq_count is 0 rather than checking
2785 * whether the mq_list is empty.
2787 if (plug
->rq_count
== 0)
2791 if (!plug
->multiple_queues
&& !plug
->has_elevator
&& !from_schedule
) {
2792 struct request_queue
*q
;
2794 rq
= rq_list_peek(&plug
->mq_list
);
2798 * Peek first request and see if we have a ->queue_rqs() hook.
2799 * If we do, we can dispatch the whole plug list in one go. We
2800 * already know at this point that all requests belong to the
2801 * same queue, caller must ensure that's the case.
2803 if (q
->mq_ops
->queue_rqs
) {
2804 blk_mq_run_dispatch_ops(q
,
2805 __blk_mq_flush_plug_list(q
, plug
));
2806 if (rq_list_empty(plug
->mq_list
))
2810 blk_mq_run_dispatch_ops(q
,
2811 blk_mq_plug_issue_direct(plug
));
2812 if (rq_list_empty(plug
->mq_list
))
2817 blk_mq_dispatch_plug_list(plug
, from_schedule
);
2818 } while (!rq_list_empty(plug
->mq_list
));
2821 static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx
*hctx
,
2822 struct list_head
*list
)
2825 blk_status_t ret
= BLK_STS_OK
;
2827 while (!list_empty(list
)) {
2828 struct request
*rq
= list_first_entry(list
, struct request
,
2831 list_del_init(&rq
->queuelist
);
2832 ret
= blk_mq_request_issue_directly(rq
, list_empty(list
));
2837 case BLK_STS_RESOURCE
:
2838 case BLK_STS_DEV_RESOURCE
:
2839 blk_mq_request_bypass_insert(rq
, 0);
2840 if (list_empty(list
))
2841 blk_mq_run_hw_queue(hctx
, false);
2844 blk_mq_end_request(rq
, ret
);
2850 if (ret
!= BLK_STS_OK
)
2851 blk_mq_commit_rqs(hctx
, queued
, false);
2854 static bool blk_mq_attempt_bio_merge(struct request_queue
*q
,
2855 struct bio
*bio
, unsigned int nr_segs
)
2857 if (!blk_queue_nomerges(q
) && bio_mergeable(bio
)) {
2858 if (blk_attempt_plug_merge(q
, bio
, nr_segs
))
2860 if (blk_mq_sched_bio_merge(q
, bio
, nr_segs
))
2866 static struct request
*blk_mq_get_new_requests(struct request_queue
*q
,
2867 struct blk_plug
*plug
,
2871 struct blk_mq_alloc_data data
= {
2874 .cmd_flags
= bio
->bi_opf
,
2878 rq_qos_throttle(q
, bio
);
2881 data
.nr_tags
= plug
->nr_ios
;
2883 data
.cached_rq
= &plug
->cached_rq
;
2886 rq
= __blk_mq_alloc_requests(&data
);
2889 rq_qos_cleanup(q
, bio
);
2890 if (bio
->bi_opf
& REQ_NOWAIT
)
2891 bio_wouldblock_error(bio
);
2896 * Check if there is a suitable cached request and return it.
2898 static struct request
*blk_mq_peek_cached_request(struct blk_plug
*plug
,
2899 struct request_queue
*q
, blk_opf_t opf
)
2901 enum hctx_type type
= blk_mq_get_hctx_type(opf
);
2906 rq
= rq_list_peek(&plug
->cached_rq
);
2907 if (!rq
|| rq
->q
!= q
)
2909 if (type
!= rq
->mq_hctx
->type
&&
2910 (type
!= HCTX_TYPE_READ
|| rq
->mq_hctx
->type
!= HCTX_TYPE_DEFAULT
))
2912 if (op_is_flush(rq
->cmd_flags
) != op_is_flush(opf
))
2917 static void blk_mq_use_cached_rq(struct request
*rq
, struct blk_plug
*plug
,
2920 WARN_ON_ONCE(rq_list_peek(&plug
->cached_rq
) != rq
);
2923 * If any qos ->throttle() end up blocking, we will have flushed the
2924 * plug and hence killed the cached_rq list as well. Pop this entry
2925 * before we throttle.
2927 plug
->cached_rq
= rq_list_next(rq
);
2928 rq_qos_throttle(rq
->q
, bio
);
2930 blk_mq_rq_time_init(rq
, 0);
2931 rq
->cmd_flags
= bio
->bi_opf
;
2932 INIT_LIST_HEAD(&rq
->queuelist
);
2936 * blk_mq_submit_bio - Create and send a request to block device.
2937 * @bio: Bio pointer.
2939 * Builds up a request structure from @q and @bio and send to the device. The
2940 * request may not be queued directly to hardware if:
2941 * * This request can be merged with another one
2942 * * We want to place request at plug queue for possible future merging
2943 * * There is an IO scheduler active at this queue
2945 * It will not queue the request if there is an error with the bio, or at the
2948 void blk_mq_submit_bio(struct bio
*bio
)
2950 struct request_queue
*q
= bdev_get_queue(bio
->bi_bdev
);
2951 struct blk_plug
*plug
= blk_mq_plug(bio
);
2952 const int is_sync
= op_is_sync(bio
->bi_opf
);
2953 struct blk_mq_hw_ctx
*hctx
;
2954 unsigned int nr_segs
= 1;
2958 bio
= blk_queue_bounce(bio
, q
);
2961 * If the plug has a cached request for this queue, try use it.
2963 * The cached request already holds a q_usage_counter reference and we
2964 * don't have to acquire a new one if we use it.
2966 rq
= blk_mq_peek_cached_request(plug
, q
, bio
->bi_opf
);
2968 if (unlikely(bio_queue_enter(bio
)))
2972 if (unlikely(bio_may_exceed_limits(bio
, &q
->limits
))) {
2973 bio
= __bio_split_to_limits(bio
, &q
->limits
, &nr_segs
);
2977 if (!bio_integrity_prep(bio
))
2980 if (blk_mq_attempt_bio_merge(q
, bio
, nr_segs
))
2984 rq
= blk_mq_get_new_requests(q
, plug
, bio
, nr_segs
);
2988 blk_mq_use_cached_rq(rq
, plug
, bio
);
2991 trace_block_getrq(bio
);
2993 rq_qos_track(q
, rq
, bio
);
2995 blk_mq_bio_to_request(rq
, bio
, nr_segs
);
2997 ret
= blk_crypto_rq_get_keyslot(rq
);
2998 if (ret
!= BLK_STS_OK
) {
2999 bio
->bi_status
= ret
;
3001 blk_mq_free_request(rq
);
3005 if (op_is_flush(bio
->bi_opf
) && blk_insert_flush(rq
))
3009 blk_add_rq_to_plug(plug
, rq
);
3014 if ((rq
->rq_flags
& RQF_USE_SCHED
) ||
3015 (hctx
->dispatch_busy
&& (q
->nr_hw_queues
== 1 || !is_sync
))) {
3016 blk_mq_insert_request(rq
, 0);
3017 blk_mq_run_hw_queue(hctx
, true);
3019 blk_mq_run_dispatch_ops(q
, blk_mq_try_issue_directly(hctx
, rq
));
3025 * Don't drop the queue reference if we were trying to use a cached
3026 * request and thus didn't acquire one.
3032 #ifdef CONFIG_BLK_MQ_STACKING
3034 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
3035 * @rq: the request being queued
3037 blk_status_t
blk_insert_cloned_request(struct request
*rq
)
3039 struct request_queue
*q
= rq
->q
;
3040 unsigned int max_sectors
= blk_queue_get_max_sectors(q
, req_op(rq
));
3041 unsigned int max_segments
= blk_rq_get_max_segments(rq
);
3044 if (blk_rq_sectors(rq
) > max_sectors
) {
3046 * SCSI device does not have a good way to return if
3047 * Write Same/Zero is actually supported. If a device rejects
3048 * a non-read/write command (discard, write same,etc.) the
3049 * low-level device driver will set the relevant queue limit to
3050 * 0 to prevent blk-lib from issuing more of the offending
3051 * operations. Commands queued prior to the queue limit being
3052 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
3053 * errors being propagated to upper layers.
3055 if (max_sectors
== 0)
3056 return BLK_STS_NOTSUPP
;
3058 printk(KERN_ERR
"%s: over max size limit. (%u > %u)\n",
3059 __func__
, blk_rq_sectors(rq
), max_sectors
);
3060 return BLK_STS_IOERR
;
3064 * The queue settings related to segment counting may differ from the
3067 rq
->nr_phys_segments
= blk_recalc_rq_segments(rq
);
3068 if (rq
->nr_phys_segments
> max_segments
) {
3069 printk(KERN_ERR
"%s: over max segments limit. (%u > %u)\n",
3070 __func__
, rq
->nr_phys_segments
, max_segments
);
3071 return BLK_STS_IOERR
;
3074 if (q
->disk
&& should_fail_request(q
->disk
->part0
, blk_rq_bytes(rq
)))
3075 return BLK_STS_IOERR
;
3077 ret
= blk_crypto_rq_get_keyslot(rq
);
3078 if (ret
!= BLK_STS_OK
)
3081 blk_account_io_start(rq
);
3084 * Since we have a scheduler attached on the top device,
3085 * bypass a potential scheduler on the bottom device for
3088 blk_mq_run_dispatch_ops(q
,
3089 ret
= blk_mq_request_issue_directly(rq
, true));
3091 blk_account_io_done(rq
, blk_time_get_ns());
3094 EXPORT_SYMBOL_GPL(blk_insert_cloned_request
);
3097 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3098 * @rq: the clone request to be cleaned up
3101 * Free all bios in @rq for a cloned request.
3103 void blk_rq_unprep_clone(struct request
*rq
)
3107 while ((bio
= rq
->bio
) != NULL
) {
3108 rq
->bio
= bio
->bi_next
;
3113 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone
);
3116 * blk_rq_prep_clone - Helper function to setup clone request
3117 * @rq: the request to be setup
3118 * @rq_src: original request to be cloned
3119 * @bs: bio_set that bios for clone are allocated from
3120 * @gfp_mask: memory allocation mask for bio
3121 * @bio_ctr: setup function to be called for each clone bio.
3122 * Returns %0 for success, non %0 for failure.
3123 * @data: private data to be passed to @bio_ctr
3126 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3127 * Also, pages which the original bios are pointing to are not copied
3128 * and the cloned bios just point same pages.
3129 * So cloned bios must be completed before original bios, which means
3130 * the caller must complete @rq before @rq_src.
3132 int blk_rq_prep_clone(struct request
*rq
, struct request
*rq_src
,
3133 struct bio_set
*bs
, gfp_t gfp_mask
,
3134 int (*bio_ctr
)(struct bio
*, struct bio
*, void *),
3137 struct bio
*bio
, *bio_src
;
3142 __rq_for_each_bio(bio_src
, rq_src
) {
3143 bio
= bio_alloc_clone(rq
->q
->disk
->part0
, bio_src
, gfp_mask
,
3148 if (bio_ctr
&& bio_ctr(bio
, bio_src
, data
))
3152 rq
->biotail
->bi_next
= bio
;
3155 rq
->bio
= rq
->biotail
= bio
;
3160 /* Copy attributes of the original request to the clone request. */
3161 rq
->__sector
= blk_rq_pos(rq_src
);
3162 rq
->__data_len
= blk_rq_bytes(rq_src
);
3163 if (rq_src
->rq_flags
& RQF_SPECIAL_PAYLOAD
) {
3164 rq
->rq_flags
|= RQF_SPECIAL_PAYLOAD
;
3165 rq
->special_vec
= rq_src
->special_vec
;
3167 rq
->nr_phys_segments
= rq_src
->nr_phys_segments
;
3168 rq
->ioprio
= rq_src
->ioprio
;
3169 rq
->write_hint
= rq_src
->write_hint
;
3171 if (rq
->bio
&& blk_crypto_rq_bio_prep(rq
, rq
->bio
, gfp_mask
) < 0)
3179 blk_rq_unprep_clone(rq
);
3183 EXPORT_SYMBOL_GPL(blk_rq_prep_clone
);
3184 #endif /* CONFIG_BLK_MQ_STACKING */
3187 * Steal bios from a request and add them to a bio list.
3188 * The request must not have been partially completed before.
3190 void blk_steal_bios(struct bio_list
*list
, struct request
*rq
)
3194 list
->tail
->bi_next
= rq
->bio
;
3196 list
->head
= rq
->bio
;
3197 list
->tail
= rq
->biotail
;
3205 EXPORT_SYMBOL_GPL(blk_steal_bios
);
3207 static size_t order_to_size(unsigned int order
)
3209 return (size_t)PAGE_SIZE
<< order
;
3212 /* called before freeing request pool in @tags */
3213 static void blk_mq_clear_rq_mapping(struct blk_mq_tags
*drv_tags
,
3214 struct blk_mq_tags
*tags
)
3217 unsigned long flags
;
3220 * There is no need to clear mapping if driver tags is not initialized
3221 * or the mapping belongs to the driver tags.
3223 if (!drv_tags
|| drv_tags
== tags
)
3226 list_for_each_entry(page
, &tags
->page_list
, lru
) {
3227 unsigned long start
= (unsigned long)page_address(page
);
3228 unsigned long end
= start
+ order_to_size(page
->private);
3231 for (i
= 0; i
< drv_tags
->nr_tags
; i
++) {
3232 struct request
*rq
= drv_tags
->rqs
[i
];
3233 unsigned long rq_addr
= (unsigned long)rq
;
3235 if (rq_addr
>= start
&& rq_addr
< end
) {
3236 WARN_ON_ONCE(req_ref_read(rq
) != 0);
3237 cmpxchg(&drv_tags
->rqs
[i
], rq
, NULL
);
3243 * Wait until all pending iteration is done.
3245 * Request reference is cleared and it is guaranteed to be observed
3246 * after the ->lock is released.
3248 spin_lock_irqsave(&drv_tags
->lock
, flags
);
3249 spin_unlock_irqrestore(&drv_tags
->lock
, flags
);
3252 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
3253 unsigned int hctx_idx
)
3255 struct blk_mq_tags
*drv_tags
;
3258 if (list_empty(&tags
->page_list
))
3261 if (blk_mq_is_shared_tags(set
->flags
))
3262 drv_tags
= set
->shared_tags
;
3264 drv_tags
= set
->tags
[hctx_idx
];
3266 if (tags
->static_rqs
&& set
->ops
->exit_request
) {
3269 for (i
= 0; i
< tags
->nr_tags
; i
++) {
3270 struct request
*rq
= tags
->static_rqs
[i
];
3274 set
->ops
->exit_request(set
, rq
, hctx_idx
);
3275 tags
->static_rqs
[i
] = NULL
;
3279 blk_mq_clear_rq_mapping(drv_tags
, tags
);
3281 while (!list_empty(&tags
->page_list
)) {
3282 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
3283 list_del_init(&page
->lru
);
3285 * Remove kmemleak object previously allocated in
3286 * blk_mq_alloc_rqs().
3288 kmemleak_free(page_address(page
));
3289 __free_pages(page
, page
->private);
3293 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
3297 kfree(tags
->static_rqs
);
3298 tags
->static_rqs
= NULL
;
3300 blk_mq_free_tags(tags
);
3303 static enum hctx_type
hctx_idx_to_type(struct blk_mq_tag_set
*set
,
3304 unsigned int hctx_idx
)
3308 for (i
= 0; i
< set
->nr_maps
; i
++) {
3309 unsigned int start
= set
->map
[i
].queue_offset
;
3310 unsigned int end
= start
+ set
->map
[i
].nr_queues
;
3312 if (hctx_idx
>= start
&& hctx_idx
< end
)
3316 if (i
>= set
->nr_maps
)
3317 i
= HCTX_TYPE_DEFAULT
;
3322 static int blk_mq_get_hctx_node(struct blk_mq_tag_set
*set
,
3323 unsigned int hctx_idx
)
3325 enum hctx_type type
= hctx_idx_to_type(set
, hctx_idx
);
3327 return blk_mq_hw_queue_to_node(&set
->map
[type
], hctx_idx
);
3330 static struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
3331 unsigned int hctx_idx
,
3332 unsigned int nr_tags
,
3333 unsigned int reserved_tags
)
3335 int node
= blk_mq_get_hctx_node(set
, hctx_idx
);
3336 struct blk_mq_tags
*tags
;
3338 if (node
== NUMA_NO_NODE
)
3339 node
= set
->numa_node
;
3341 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
,
3342 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
3346 tags
->rqs
= kcalloc_node(nr_tags
, sizeof(struct request
*),
3347 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
3352 tags
->static_rqs
= kcalloc_node(nr_tags
, sizeof(struct request
*),
3353 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
3355 if (!tags
->static_rqs
)
3363 blk_mq_free_tags(tags
);
3367 static int blk_mq_init_request(struct blk_mq_tag_set
*set
, struct request
*rq
,
3368 unsigned int hctx_idx
, int node
)
3372 if (set
->ops
->init_request
) {
3373 ret
= set
->ops
->init_request(set
, rq
, hctx_idx
, node
);
3378 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
3382 static int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
,
3383 struct blk_mq_tags
*tags
,
3384 unsigned int hctx_idx
, unsigned int depth
)
3386 unsigned int i
, j
, entries_per_page
, max_order
= 4;
3387 int node
= blk_mq_get_hctx_node(set
, hctx_idx
);
3388 size_t rq_size
, left
;
3390 if (node
== NUMA_NO_NODE
)
3391 node
= set
->numa_node
;
3393 INIT_LIST_HEAD(&tags
->page_list
);
3396 * rq_size is the size of the request plus driver payload, rounded
3397 * to the cacheline size
3399 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
3401 left
= rq_size
* depth
;
3403 for (i
= 0; i
< depth
; ) {
3404 int this_order
= max_order
;
3409 while (this_order
&& left
< order_to_size(this_order
- 1))
3413 page
= alloc_pages_node(node
,
3414 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
3420 if (order_to_size(this_order
) < rq_size
)
3427 page
->private = this_order
;
3428 list_add_tail(&page
->lru
, &tags
->page_list
);
3430 p
= page_address(page
);
3432 * Allow kmemleak to scan these pages as they contain pointers
3433 * to additional allocations like via ops->init_request().
3435 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
3436 entries_per_page
= order_to_size(this_order
) / rq_size
;
3437 to_do
= min(entries_per_page
, depth
- i
);
3438 left
-= to_do
* rq_size
;
3439 for (j
= 0; j
< to_do
; j
++) {
3440 struct request
*rq
= p
;
3442 tags
->static_rqs
[i
] = rq
;
3443 if (blk_mq_init_request(set
, rq
, hctx_idx
, node
)) {
3444 tags
->static_rqs
[i
] = NULL
;
3455 blk_mq_free_rqs(set
, tags
, hctx_idx
);
3459 struct rq_iter_data
{
3460 struct blk_mq_hw_ctx
*hctx
;
3464 static bool blk_mq_has_request(struct request
*rq
, void *data
)
3466 struct rq_iter_data
*iter_data
= data
;
3468 if (rq
->mq_hctx
!= iter_data
->hctx
)
3470 iter_data
->has_rq
= true;
3474 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx
*hctx
)
3476 struct blk_mq_tags
*tags
= hctx
->sched_tags
?
3477 hctx
->sched_tags
: hctx
->tags
;
3478 struct rq_iter_data data
= {
3482 blk_mq_all_tag_iter(tags
, blk_mq_has_request
, &data
);
3486 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu
,
3487 struct blk_mq_hw_ctx
*hctx
)
3489 if (cpumask_first_and(hctx
->cpumask
, cpu_online_mask
) != cpu
)
3491 if (cpumask_next_and(cpu
, hctx
->cpumask
, cpu_online_mask
) < nr_cpu_ids
)
3496 static int blk_mq_hctx_notify_offline(unsigned int cpu
, struct hlist_node
*node
)
3498 struct blk_mq_hw_ctx
*hctx
= hlist_entry_safe(node
,
3499 struct blk_mq_hw_ctx
, cpuhp_online
);
3501 if (!cpumask_test_cpu(cpu
, hctx
->cpumask
) ||
3502 !blk_mq_last_cpu_in_hctx(cpu
, hctx
))
3506 * Prevent new request from being allocated on the current hctx.
3508 * The smp_mb__after_atomic() Pairs with the implied barrier in
3509 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3510 * seen once we return from the tag allocator.
3512 set_bit(BLK_MQ_S_INACTIVE
, &hctx
->state
);
3513 smp_mb__after_atomic();
3516 * Try to grab a reference to the queue and wait for any outstanding
3517 * requests. If we could not grab a reference the queue has been
3518 * frozen and there are no requests.
3520 if (percpu_ref_tryget(&hctx
->queue
->q_usage_counter
)) {
3521 while (blk_mq_hctx_has_requests(hctx
))
3523 percpu_ref_put(&hctx
->queue
->q_usage_counter
);
3529 static int blk_mq_hctx_notify_online(unsigned int cpu
, struct hlist_node
*node
)
3531 struct blk_mq_hw_ctx
*hctx
= hlist_entry_safe(node
,
3532 struct blk_mq_hw_ctx
, cpuhp_online
);
3534 if (cpumask_test_cpu(cpu
, hctx
->cpumask
))
3535 clear_bit(BLK_MQ_S_INACTIVE
, &hctx
->state
);
3540 * 'cpu' is going away. splice any existing rq_list entries from this
3541 * software queue to the hw queue dispatch list, and ensure that it
3544 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
3546 struct blk_mq_hw_ctx
*hctx
;
3547 struct blk_mq_ctx
*ctx
;
3549 enum hctx_type type
;
3551 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
3552 if (!cpumask_test_cpu(cpu
, hctx
->cpumask
))
3555 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
3558 spin_lock(&ctx
->lock
);
3559 if (!list_empty(&ctx
->rq_lists
[type
])) {
3560 list_splice_init(&ctx
->rq_lists
[type
], &tmp
);
3561 blk_mq_hctx_clear_pending(hctx
, ctx
);
3563 spin_unlock(&ctx
->lock
);
3565 if (list_empty(&tmp
))
3568 spin_lock(&hctx
->lock
);
3569 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
3570 spin_unlock(&hctx
->lock
);
3572 blk_mq_run_hw_queue(hctx
, true);
3576 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
3578 if (!(hctx
->flags
& BLK_MQ_F_STACKING
))
3579 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE
,
3580 &hctx
->cpuhp_online
);
3581 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
3586 * Before freeing hw queue, clearing the flush request reference in
3587 * tags->rqs[] for avoiding potential UAF.
3589 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags
*tags
,
3590 unsigned int queue_depth
, struct request
*flush_rq
)
3593 unsigned long flags
;
3595 /* The hw queue may not be mapped yet */
3599 WARN_ON_ONCE(req_ref_read(flush_rq
) != 0);
3601 for (i
= 0; i
< queue_depth
; i
++)
3602 cmpxchg(&tags
->rqs
[i
], flush_rq
, NULL
);
3605 * Wait until all pending iteration is done.
3607 * Request reference is cleared and it is guaranteed to be observed
3608 * after the ->lock is released.
3610 spin_lock_irqsave(&tags
->lock
, flags
);
3611 spin_unlock_irqrestore(&tags
->lock
, flags
);
3614 /* hctx->ctxs will be freed in queue's release handler */
3615 static void blk_mq_exit_hctx(struct request_queue
*q
,
3616 struct blk_mq_tag_set
*set
,
3617 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
3619 struct request
*flush_rq
= hctx
->fq
->flush_rq
;
3621 if (blk_mq_hw_queue_mapped(hctx
))
3622 blk_mq_tag_idle(hctx
);
3624 if (blk_queue_init_done(q
))
3625 blk_mq_clear_flush_rq_mapping(set
->tags
[hctx_idx
],
3626 set
->queue_depth
, flush_rq
);
3627 if (set
->ops
->exit_request
)
3628 set
->ops
->exit_request(set
, flush_rq
, hctx_idx
);
3630 if (set
->ops
->exit_hctx
)
3631 set
->ops
->exit_hctx(hctx
, hctx_idx
);
3633 blk_mq_remove_cpuhp(hctx
);
3635 xa_erase(&q
->hctx_table
, hctx_idx
);
3637 spin_lock(&q
->unused_hctx_lock
);
3638 list_add(&hctx
->hctx_list
, &q
->unused_hctx_list
);
3639 spin_unlock(&q
->unused_hctx_lock
);
3642 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
3643 struct blk_mq_tag_set
*set
, int nr_queue
)
3645 struct blk_mq_hw_ctx
*hctx
;
3648 queue_for_each_hw_ctx(q
, hctx
, i
) {
3651 blk_mq_exit_hctx(q
, set
, hctx
, i
);
3655 static int blk_mq_init_hctx(struct request_queue
*q
,
3656 struct blk_mq_tag_set
*set
,
3657 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
3659 hctx
->queue_num
= hctx_idx
;
3661 if (!(hctx
->flags
& BLK_MQ_F_STACKING
))
3662 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE
,
3663 &hctx
->cpuhp_online
);
3664 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
3666 hctx
->tags
= set
->tags
[hctx_idx
];
3668 if (set
->ops
->init_hctx
&&
3669 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
3670 goto unregister_cpu_notifier
;
3672 if (blk_mq_init_request(set
, hctx
->fq
->flush_rq
, hctx_idx
,
3676 if (xa_insert(&q
->hctx_table
, hctx_idx
, hctx
, GFP_KERNEL
))
3682 if (set
->ops
->exit_request
)
3683 set
->ops
->exit_request(set
, hctx
->fq
->flush_rq
, hctx_idx
);
3685 if (set
->ops
->exit_hctx
)
3686 set
->ops
->exit_hctx(hctx
, hctx_idx
);
3687 unregister_cpu_notifier
:
3688 blk_mq_remove_cpuhp(hctx
);
3692 static struct blk_mq_hw_ctx
*
3693 blk_mq_alloc_hctx(struct request_queue
*q
, struct blk_mq_tag_set
*set
,
3696 struct blk_mq_hw_ctx
*hctx
;
3697 gfp_t gfp
= GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
;
3699 hctx
= kzalloc_node(sizeof(struct blk_mq_hw_ctx
), gfp
, node
);
3701 goto fail_alloc_hctx
;
3703 if (!zalloc_cpumask_var_node(&hctx
->cpumask
, gfp
, node
))
3706 atomic_set(&hctx
->nr_active
, 0);
3707 if (node
== NUMA_NO_NODE
)
3708 node
= set
->numa_node
;
3709 hctx
->numa_node
= node
;
3711 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
3712 spin_lock_init(&hctx
->lock
);
3713 INIT_LIST_HEAD(&hctx
->dispatch
);
3715 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_QUEUE_SHARED
;
3717 INIT_LIST_HEAD(&hctx
->hctx_list
);
3720 * Allocate space for all possible cpus to avoid allocation at
3723 hctx
->ctxs
= kmalloc_array_node(nr_cpu_ids
, sizeof(void *),
3728 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8),
3729 gfp
, node
, false, false))
3733 spin_lock_init(&hctx
->dispatch_wait_lock
);
3734 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
3735 INIT_LIST_HEAD(&hctx
->dispatch_wait
.entry
);
3737 hctx
->fq
= blk_alloc_flush_queue(hctx
->numa_node
, set
->cmd_size
, gfp
);
3741 blk_mq_hctx_kobj_init(hctx
);
3746 sbitmap_free(&hctx
->ctx_map
);
3750 free_cpumask_var(hctx
->cpumask
);
3757 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
3758 unsigned int nr_hw_queues
)
3760 struct blk_mq_tag_set
*set
= q
->tag_set
;
3763 for_each_possible_cpu(i
) {
3764 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
3765 struct blk_mq_hw_ctx
*hctx
;
3769 spin_lock_init(&__ctx
->lock
);
3770 for (k
= HCTX_TYPE_DEFAULT
; k
< HCTX_MAX_TYPES
; k
++)
3771 INIT_LIST_HEAD(&__ctx
->rq_lists
[k
]);
3776 * Set local node, IFF we have more than one hw queue. If
3777 * not, we remain on the home node of the device
3779 for (j
= 0; j
< set
->nr_maps
; j
++) {
3780 hctx
= blk_mq_map_queue_type(q
, j
, i
);
3781 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
3782 hctx
->numa_node
= cpu_to_node(i
);
3787 struct blk_mq_tags
*blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set
*set
,
3788 unsigned int hctx_idx
,
3791 struct blk_mq_tags
*tags
;
3794 tags
= blk_mq_alloc_rq_map(set
, hctx_idx
, depth
, set
->reserved_tags
);
3798 ret
= blk_mq_alloc_rqs(set
, tags
, hctx_idx
, depth
);
3800 blk_mq_free_rq_map(tags
);
3807 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set
*set
,
3810 if (blk_mq_is_shared_tags(set
->flags
)) {
3811 set
->tags
[hctx_idx
] = set
->shared_tags
;
3816 set
->tags
[hctx_idx
] = blk_mq_alloc_map_and_rqs(set
, hctx_idx
,
3819 return set
->tags
[hctx_idx
];
3822 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set
*set
,
3823 struct blk_mq_tags
*tags
,
3824 unsigned int hctx_idx
)
3827 blk_mq_free_rqs(set
, tags
, hctx_idx
);
3828 blk_mq_free_rq_map(tags
);
3832 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set
*set
,
3833 unsigned int hctx_idx
)
3835 if (!blk_mq_is_shared_tags(set
->flags
))
3836 blk_mq_free_map_and_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
3838 set
->tags
[hctx_idx
] = NULL
;
3841 static void blk_mq_map_swqueue(struct request_queue
*q
)
3843 unsigned int j
, hctx_idx
;
3845 struct blk_mq_hw_ctx
*hctx
;
3846 struct blk_mq_ctx
*ctx
;
3847 struct blk_mq_tag_set
*set
= q
->tag_set
;
3849 queue_for_each_hw_ctx(q
, hctx
, i
) {
3850 cpumask_clear(hctx
->cpumask
);
3852 hctx
->dispatch_from
= NULL
;
3856 * Map software to hardware queues.
3858 * If the cpu isn't present, the cpu is mapped to first hctx.
3860 for_each_possible_cpu(i
) {
3862 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
3863 for (j
= 0; j
< set
->nr_maps
; j
++) {
3864 if (!set
->map
[j
].nr_queues
) {
3865 ctx
->hctxs
[j
] = blk_mq_map_queue_type(q
,
3866 HCTX_TYPE_DEFAULT
, i
);
3869 hctx_idx
= set
->map
[j
].mq_map
[i
];
3870 /* unmapped hw queue can be remapped after CPU topo changed */
3871 if (!set
->tags
[hctx_idx
] &&
3872 !__blk_mq_alloc_map_and_rqs(set
, hctx_idx
)) {
3874 * If tags initialization fail for some hctx,
3875 * that hctx won't be brought online. In this
3876 * case, remap the current ctx to hctx[0] which
3877 * is guaranteed to always have tags allocated
3879 set
->map
[j
].mq_map
[i
] = 0;
3882 hctx
= blk_mq_map_queue_type(q
, j
, i
);
3883 ctx
->hctxs
[j
] = hctx
;
3885 * If the CPU is already set in the mask, then we've
3886 * mapped this one already. This can happen if
3887 * devices share queues across queue maps.
3889 if (cpumask_test_cpu(i
, hctx
->cpumask
))
3892 cpumask_set_cpu(i
, hctx
->cpumask
);
3894 ctx
->index_hw
[hctx
->type
] = hctx
->nr_ctx
;
3895 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
3898 * If the nr_ctx type overflows, we have exceeded the
3899 * amount of sw queues we can support.
3901 BUG_ON(!hctx
->nr_ctx
);
3904 for (; j
< HCTX_MAX_TYPES
; j
++)
3905 ctx
->hctxs
[j
] = blk_mq_map_queue_type(q
,
3906 HCTX_TYPE_DEFAULT
, i
);
3909 queue_for_each_hw_ctx(q
, hctx
, i
) {
3911 * If no software queues are mapped to this hardware queue,
3912 * disable it and free the request entries.
3914 if (!hctx
->nr_ctx
) {
3915 /* Never unmap queue 0. We need it as a
3916 * fallback in case of a new remap fails
3920 __blk_mq_free_map_and_rqs(set
, i
);
3926 hctx
->tags
= set
->tags
[i
];
3927 WARN_ON(!hctx
->tags
);
3930 * Set the map size to the number of mapped software queues.
3931 * This is more accurate and more efficient than looping
3932 * over all possibly mapped software queues.
3934 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
3937 * Initialize batch roundrobin counts
3939 hctx
->next_cpu
= blk_mq_first_mapped_cpu(hctx
);
3940 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
3945 * Caller needs to ensure that we're either frozen/quiesced, or that
3946 * the queue isn't live yet.
3948 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
3950 struct blk_mq_hw_ctx
*hctx
;
3953 queue_for_each_hw_ctx(q
, hctx
, i
) {
3955 hctx
->flags
|= BLK_MQ_F_TAG_QUEUE_SHARED
;
3957 blk_mq_tag_idle(hctx
);
3958 hctx
->flags
&= ~BLK_MQ_F_TAG_QUEUE_SHARED
;
3963 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set
*set
,
3966 struct request_queue
*q
;
3968 lockdep_assert_held(&set
->tag_list_lock
);
3970 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3971 blk_mq_freeze_queue(q
);
3972 queue_set_hctx_shared(q
, shared
);
3973 blk_mq_unfreeze_queue(q
);
3977 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
3979 struct blk_mq_tag_set
*set
= q
->tag_set
;
3981 mutex_lock(&set
->tag_list_lock
);
3982 list_del(&q
->tag_set_list
);
3983 if (list_is_singular(&set
->tag_list
)) {
3984 /* just transitioned to unshared */
3985 set
->flags
&= ~BLK_MQ_F_TAG_QUEUE_SHARED
;
3986 /* update existing queue */
3987 blk_mq_update_tag_set_shared(set
, false);
3989 mutex_unlock(&set
->tag_list_lock
);
3990 INIT_LIST_HEAD(&q
->tag_set_list
);
3993 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
3994 struct request_queue
*q
)
3996 mutex_lock(&set
->tag_list_lock
);
3999 * Check to see if we're transitioning to shared (from 1 to 2 queues).
4001 if (!list_empty(&set
->tag_list
) &&
4002 !(set
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
)) {
4003 set
->flags
|= BLK_MQ_F_TAG_QUEUE_SHARED
;
4004 /* update existing queue */
4005 blk_mq_update_tag_set_shared(set
, true);
4007 if (set
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
)
4008 queue_set_hctx_shared(q
, true);
4009 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
4011 mutex_unlock(&set
->tag_list_lock
);
4014 /* All allocations will be freed in release handler of q->mq_kobj */
4015 static int blk_mq_alloc_ctxs(struct request_queue
*q
)
4017 struct blk_mq_ctxs
*ctxs
;
4020 ctxs
= kzalloc(sizeof(*ctxs
), GFP_KERNEL
);
4024 ctxs
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
4025 if (!ctxs
->queue_ctx
)
4028 for_each_possible_cpu(cpu
) {
4029 struct blk_mq_ctx
*ctx
= per_cpu_ptr(ctxs
->queue_ctx
, cpu
);
4033 q
->mq_kobj
= &ctxs
->kobj
;
4034 q
->queue_ctx
= ctxs
->queue_ctx
;
4043 * It is the actual release handler for mq, but we do it from
4044 * request queue's release handler for avoiding use-after-free
4045 * and headache because q->mq_kobj shouldn't have been introduced,
4046 * but we can't group ctx/kctx kobj without it.
4048 void blk_mq_release(struct request_queue
*q
)
4050 struct blk_mq_hw_ctx
*hctx
, *next
;
4053 queue_for_each_hw_ctx(q
, hctx
, i
)
4054 WARN_ON_ONCE(hctx
&& list_empty(&hctx
->hctx_list
));
4056 /* all hctx are in .unused_hctx_list now */
4057 list_for_each_entry_safe(hctx
, next
, &q
->unused_hctx_list
, hctx_list
) {
4058 list_del_init(&hctx
->hctx_list
);
4059 kobject_put(&hctx
->kobj
);
4062 xa_destroy(&q
->hctx_table
);
4065 * release .mq_kobj and sw queue's kobject now because
4066 * both share lifetime with request queue.
4068 blk_mq_sysfs_deinit(q
);
4071 struct request_queue
*blk_mq_alloc_queue(struct blk_mq_tag_set
*set
,
4072 struct queue_limits
*lim
, void *queuedata
)
4074 struct queue_limits default_lim
= { };
4075 struct request_queue
*q
;
4078 q
= blk_alloc_queue(lim
? lim
: &default_lim
, set
->numa_node
);
4081 q
->queuedata
= queuedata
;
4082 ret
= blk_mq_init_allocated_queue(set
, q
);
4085 return ERR_PTR(ret
);
4089 EXPORT_SYMBOL(blk_mq_alloc_queue
);
4092 * blk_mq_destroy_queue - shutdown a request queue
4093 * @q: request queue to shutdown
4095 * This shuts down a request queue allocated by blk_mq_alloc_queue(). All future
4096 * requests will be failed with -ENODEV. The caller is responsible for dropping
4097 * the reference from blk_mq_alloc_queue() by calling blk_put_queue().
4099 * Context: can sleep
4101 void blk_mq_destroy_queue(struct request_queue
*q
)
4103 WARN_ON_ONCE(!queue_is_mq(q
));
4104 WARN_ON_ONCE(blk_queue_registered(q
));
4108 blk_queue_flag_set(QUEUE_FLAG_DYING
, q
);
4109 blk_queue_start_drain(q
);
4110 blk_mq_freeze_queue_wait(q
);
4113 blk_mq_cancel_work_sync(q
);
4114 blk_mq_exit_queue(q
);
4116 EXPORT_SYMBOL(blk_mq_destroy_queue
);
4118 struct gendisk
*__blk_mq_alloc_disk(struct blk_mq_tag_set
*set
,
4119 struct queue_limits
*lim
, void *queuedata
,
4120 struct lock_class_key
*lkclass
)
4122 struct request_queue
*q
;
4123 struct gendisk
*disk
;
4125 q
= blk_mq_alloc_queue(set
, lim
, queuedata
);
4129 disk
= __alloc_disk_node(q
, set
->numa_node
, lkclass
);
4131 blk_mq_destroy_queue(q
);
4133 return ERR_PTR(-ENOMEM
);
4135 set_bit(GD_OWNS_QUEUE
, &disk
->state
);
4138 EXPORT_SYMBOL(__blk_mq_alloc_disk
);
4140 struct gendisk
*blk_mq_alloc_disk_for_queue(struct request_queue
*q
,
4141 struct lock_class_key
*lkclass
)
4143 struct gendisk
*disk
;
4145 if (!blk_get_queue(q
))
4147 disk
= __alloc_disk_node(q
, NUMA_NO_NODE
, lkclass
);
4152 EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue
);
4154 static struct blk_mq_hw_ctx
*blk_mq_alloc_and_init_hctx(
4155 struct blk_mq_tag_set
*set
, struct request_queue
*q
,
4156 int hctx_idx
, int node
)
4158 struct blk_mq_hw_ctx
*hctx
= NULL
, *tmp
;
4160 /* reuse dead hctx first */
4161 spin_lock(&q
->unused_hctx_lock
);
4162 list_for_each_entry(tmp
, &q
->unused_hctx_list
, hctx_list
) {
4163 if (tmp
->numa_node
== node
) {
4169 list_del_init(&hctx
->hctx_list
);
4170 spin_unlock(&q
->unused_hctx_lock
);
4173 hctx
= blk_mq_alloc_hctx(q
, set
, node
);
4177 if (blk_mq_init_hctx(q
, set
, hctx
, hctx_idx
))
4183 kobject_put(&hctx
->kobj
);
4188 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
4189 struct request_queue
*q
)
4191 struct blk_mq_hw_ctx
*hctx
;
4194 /* protect against switching io scheduler */
4195 mutex_lock(&q
->sysfs_lock
);
4196 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
4198 int node
= blk_mq_get_hctx_node(set
, i
);
4199 struct blk_mq_hw_ctx
*old_hctx
= xa_load(&q
->hctx_table
, i
);
4202 old_node
= old_hctx
->numa_node
;
4203 blk_mq_exit_hctx(q
, set
, old_hctx
, i
);
4206 if (!blk_mq_alloc_and_init_hctx(set
, q
, i
, node
)) {
4209 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4211 hctx
= blk_mq_alloc_and_init_hctx(set
, q
, i
, old_node
);
4212 WARN_ON_ONCE(!hctx
);
4216 * Increasing nr_hw_queues fails. Free the newly allocated
4217 * hctxs and keep the previous q->nr_hw_queues.
4219 if (i
!= set
->nr_hw_queues
) {
4220 j
= q
->nr_hw_queues
;
4223 q
->nr_hw_queues
= set
->nr_hw_queues
;
4226 xa_for_each_start(&q
->hctx_table
, j
, hctx
, j
)
4227 blk_mq_exit_hctx(q
, set
, hctx
, j
);
4228 mutex_unlock(&q
->sysfs_lock
);
4231 static void blk_mq_update_poll_flag(struct request_queue
*q
)
4233 struct blk_mq_tag_set
*set
= q
->tag_set
;
4235 if (set
->nr_maps
> HCTX_TYPE_POLL
&&
4236 set
->map
[HCTX_TYPE_POLL
].nr_queues
)
4237 blk_queue_flag_set(QUEUE_FLAG_POLL
, q
);
4239 blk_queue_flag_clear(QUEUE_FLAG_POLL
, q
);
4242 int blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
4243 struct request_queue
*q
)
4245 /* mark the queue as mq asap */
4246 q
->mq_ops
= set
->ops
;
4248 if (blk_mq_alloc_ctxs(q
))
4251 /* init q->mq_kobj and sw queues' kobjects */
4252 blk_mq_sysfs_init(q
);
4254 INIT_LIST_HEAD(&q
->unused_hctx_list
);
4255 spin_lock_init(&q
->unused_hctx_lock
);
4257 xa_init(&q
->hctx_table
);
4259 blk_mq_realloc_hw_ctxs(set
, q
);
4260 if (!q
->nr_hw_queues
)
4263 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
4264 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
4268 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
4269 blk_mq_update_poll_flag(q
);
4271 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
4272 INIT_LIST_HEAD(&q
->flush_list
);
4273 INIT_LIST_HEAD(&q
->requeue_list
);
4274 spin_lock_init(&q
->requeue_lock
);
4276 q
->nr_requests
= set
->queue_depth
;
4278 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
4279 blk_mq_add_queue_tag_set(set
, q
);
4280 blk_mq_map_swqueue(q
);
4289 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
4291 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4292 void blk_mq_exit_queue(struct request_queue
*q
)
4294 struct blk_mq_tag_set
*set
= q
->tag_set
;
4296 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4297 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
4298 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4299 blk_mq_del_queue_tag_set(q
);
4302 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
4306 if (blk_mq_is_shared_tags(set
->flags
)) {
4307 set
->shared_tags
= blk_mq_alloc_map_and_rqs(set
,
4310 if (!set
->shared_tags
)
4314 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
4315 if (!__blk_mq_alloc_map_and_rqs(set
, i
))
4324 __blk_mq_free_map_and_rqs(set
, i
);
4326 if (blk_mq_is_shared_tags(set
->flags
)) {
4327 blk_mq_free_map_and_rqs(set
, set
->shared_tags
,
4328 BLK_MQ_NO_HCTX_IDX
);
4335 * Allocate the request maps associated with this tag_set. Note that this
4336 * may reduce the depth asked for, if memory is tight. set->queue_depth
4337 * will be updated to reflect the allocated depth.
4339 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set
*set
)
4344 depth
= set
->queue_depth
;
4346 err
= __blk_mq_alloc_rq_maps(set
);
4350 set
->queue_depth
>>= 1;
4351 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
4355 } while (set
->queue_depth
);
4357 if (!set
->queue_depth
|| err
) {
4358 pr_err("blk-mq: failed to allocate request map\n");
4362 if (depth
!= set
->queue_depth
)
4363 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4364 depth
, set
->queue_depth
);
4369 static void blk_mq_update_queue_map(struct blk_mq_tag_set
*set
)
4372 * blk_mq_map_queues() and multiple .map_queues() implementations
4373 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4374 * number of hardware queues.
4376 if (set
->nr_maps
== 1)
4377 set
->map
[HCTX_TYPE_DEFAULT
].nr_queues
= set
->nr_hw_queues
;
4379 if (set
->ops
->map_queues
) {
4383 * transport .map_queues is usually done in the following
4386 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4387 * mask = get_cpu_mask(queue)
4388 * for_each_cpu(cpu, mask)
4389 * set->map[x].mq_map[cpu] = queue;
4392 * When we need to remap, the table has to be cleared for
4393 * killing stale mapping since one CPU may not be mapped
4396 for (i
= 0; i
< set
->nr_maps
; i
++)
4397 blk_mq_clear_mq_map(&set
->map
[i
]);
4399 set
->ops
->map_queues(set
);
4401 BUG_ON(set
->nr_maps
> 1);
4402 blk_mq_map_queues(&set
->map
[HCTX_TYPE_DEFAULT
]);
4406 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set
*set
,
4407 int new_nr_hw_queues
)
4409 struct blk_mq_tags
**new_tags
;
4412 if (set
->nr_hw_queues
>= new_nr_hw_queues
)
4415 new_tags
= kcalloc_node(new_nr_hw_queues
, sizeof(struct blk_mq_tags
*),
4416 GFP_KERNEL
, set
->numa_node
);
4421 memcpy(new_tags
, set
->tags
, set
->nr_hw_queues
*
4422 sizeof(*set
->tags
));
4424 set
->tags
= new_tags
;
4426 for (i
= set
->nr_hw_queues
; i
< new_nr_hw_queues
; i
++) {
4427 if (!__blk_mq_alloc_map_and_rqs(set
, i
)) {
4428 while (--i
>= set
->nr_hw_queues
)
4429 __blk_mq_free_map_and_rqs(set
, i
);
4436 set
->nr_hw_queues
= new_nr_hw_queues
;
4441 * Alloc a tag set to be associated with one or more request queues.
4442 * May fail with EINVAL for various error conditions. May adjust the
4443 * requested depth down, if it's too large. In that case, the set
4444 * value will be stored in set->queue_depth.
4446 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
4450 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
4452 if (!set
->nr_hw_queues
)
4454 if (!set
->queue_depth
)
4456 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
4459 if (!set
->ops
->queue_rq
)
4462 if (!set
->ops
->get_budget
^ !set
->ops
->put_budget
)
4465 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
4466 pr_info("blk-mq: reduced tag depth to %u\n",
4468 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
4473 else if (set
->nr_maps
> HCTX_MAX_TYPES
)
4477 * If a crashdump is active, then we are potentially in a very
4478 * memory constrained environment. Limit us to 64 tags to prevent
4479 * using too much memory.
4481 if (is_kdump_kernel())
4482 set
->queue_depth
= min(64U, set
->queue_depth
);
4485 * There is no use for more h/w queues than cpus if we just have
4488 if (set
->nr_maps
== 1 && set
->nr_hw_queues
> nr_cpu_ids
)
4489 set
->nr_hw_queues
= nr_cpu_ids
;
4491 if (set
->flags
& BLK_MQ_F_BLOCKING
) {
4492 set
->srcu
= kmalloc(sizeof(*set
->srcu
), GFP_KERNEL
);
4495 ret
= init_srcu_struct(set
->srcu
);
4501 set
->tags
= kcalloc_node(set
->nr_hw_queues
,
4502 sizeof(struct blk_mq_tags
*), GFP_KERNEL
,
4505 goto out_cleanup_srcu
;
4507 for (i
= 0; i
< set
->nr_maps
; i
++) {
4508 set
->map
[i
].mq_map
= kcalloc_node(nr_cpu_ids
,
4509 sizeof(set
->map
[i
].mq_map
[0]),
4510 GFP_KERNEL
, set
->numa_node
);
4511 if (!set
->map
[i
].mq_map
)
4512 goto out_free_mq_map
;
4513 set
->map
[i
].nr_queues
= set
->nr_hw_queues
;
4516 blk_mq_update_queue_map(set
);
4518 ret
= blk_mq_alloc_set_map_and_rqs(set
);
4520 goto out_free_mq_map
;
4522 mutex_init(&set
->tag_list_lock
);
4523 INIT_LIST_HEAD(&set
->tag_list
);
4528 for (i
= 0; i
< set
->nr_maps
; i
++) {
4529 kfree(set
->map
[i
].mq_map
);
4530 set
->map
[i
].mq_map
= NULL
;
4535 if (set
->flags
& BLK_MQ_F_BLOCKING
)
4536 cleanup_srcu_struct(set
->srcu
);
4538 if (set
->flags
& BLK_MQ_F_BLOCKING
)
4542 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
4544 /* allocate and initialize a tagset for a simple single-queue device */
4545 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set
*set
,
4546 const struct blk_mq_ops
*ops
, unsigned int queue_depth
,
4547 unsigned int set_flags
)
4549 memset(set
, 0, sizeof(*set
));
4551 set
->nr_hw_queues
= 1;
4553 set
->queue_depth
= queue_depth
;
4554 set
->numa_node
= NUMA_NO_NODE
;
4555 set
->flags
= set_flags
;
4556 return blk_mq_alloc_tag_set(set
);
4558 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set
);
4560 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
4564 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
4565 __blk_mq_free_map_and_rqs(set
, i
);
4567 if (blk_mq_is_shared_tags(set
->flags
)) {
4568 blk_mq_free_map_and_rqs(set
, set
->shared_tags
,
4569 BLK_MQ_NO_HCTX_IDX
);
4572 for (j
= 0; j
< set
->nr_maps
; j
++) {
4573 kfree(set
->map
[j
].mq_map
);
4574 set
->map
[j
].mq_map
= NULL
;
4579 if (set
->flags
& BLK_MQ_F_BLOCKING
) {
4580 cleanup_srcu_struct(set
->srcu
);
4584 EXPORT_SYMBOL(blk_mq_free_tag_set
);
4586 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
4588 struct blk_mq_tag_set
*set
= q
->tag_set
;
4589 struct blk_mq_hw_ctx
*hctx
;
4596 if (q
->nr_requests
== nr
)
4599 blk_mq_freeze_queue(q
);
4600 blk_mq_quiesce_queue(q
);
4603 queue_for_each_hw_ctx(q
, hctx
, i
) {
4607 * If we're using an MQ scheduler, just update the scheduler
4608 * queue depth. This is similar to what the old code would do.
4610 if (hctx
->sched_tags
) {
4611 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
4614 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
, nr
,
4619 if (q
->elevator
&& q
->elevator
->type
->ops
.depth_updated
)
4620 q
->elevator
->type
->ops
.depth_updated(hctx
);
4623 q
->nr_requests
= nr
;
4624 if (blk_mq_is_shared_tags(set
->flags
)) {
4626 blk_mq_tag_update_sched_shared_tags(q
);
4628 blk_mq_tag_resize_shared_tags(set
, nr
);
4632 blk_mq_unquiesce_queue(q
);
4633 blk_mq_unfreeze_queue(q
);
4639 * request_queue and elevator_type pair.
4640 * It is just used by __blk_mq_update_nr_hw_queues to cache
4641 * the elevator_type associated with a request_queue.
4643 struct blk_mq_qe_pair
{
4644 struct list_head node
;
4645 struct request_queue
*q
;
4646 struct elevator_type
*type
;
4650 * Cache the elevator_type in qe pair list and switch the
4651 * io scheduler to 'none'
4653 static bool blk_mq_elv_switch_none(struct list_head
*head
,
4654 struct request_queue
*q
)
4656 struct blk_mq_qe_pair
*qe
;
4658 qe
= kmalloc(sizeof(*qe
), GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
);
4662 /* q->elevator needs protection from ->sysfs_lock */
4663 mutex_lock(&q
->sysfs_lock
);
4665 /* the check has to be done with holding sysfs_lock */
4671 INIT_LIST_HEAD(&qe
->node
);
4673 qe
->type
= q
->elevator
->type
;
4674 /* keep a reference to the elevator module as we'll switch back */
4675 __elevator_get(qe
->type
);
4676 list_add(&qe
->node
, head
);
4677 elevator_disable(q
);
4679 mutex_unlock(&q
->sysfs_lock
);
4684 static struct blk_mq_qe_pair
*blk_lookup_qe_pair(struct list_head
*head
,
4685 struct request_queue
*q
)
4687 struct blk_mq_qe_pair
*qe
;
4689 list_for_each_entry(qe
, head
, node
)
4696 static void blk_mq_elv_switch_back(struct list_head
*head
,
4697 struct request_queue
*q
)
4699 struct blk_mq_qe_pair
*qe
;
4700 struct elevator_type
*t
;
4702 qe
= blk_lookup_qe_pair(head
, q
);
4706 list_del(&qe
->node
);
4709 mutex_lock(&q
->sysfs_lock
);
4710 elevator_switch(q
, t
);
4711 /* drop the reference acquired in blk_mq_elv_switch_none */
4713 mutex_unlock(&q
->sysfs_lock
);
4716 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
,
4719 struct request_queue
*q
;
4721 int prev_nr_hw_queues
= set
->nr_hw_queues
;
4724 lockdep_assert_held(&set
->tag_list_lock
);
4726 if (set
->nr_maps
== 1 && nr_hw_queues
> nr_cpu_ids
)
4727 nr_hw_queues
= nr_cpu_ids
;
4728 if (nr_hw_queues
< 1)
4730 if (set
->nr_maps
== 1 && nr_hw_queues
== set
->nr_hw_queues
)
4733 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
4734 blk_mq_freeze_queue(q
);
4736 * Switch IO scheduler to 'none', cleaning up the data associated
4737 * with the previous scheduler. We will switch back once we are done
4738 * updating the new sw to hw queue mappings.
4740 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
4741 if (!blk_mq_elv_switch_none(&head
, q
))
4744 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
4745 blk_mq_debugfs_unregister_hctxs(q
);
4746 blk_mq_sysfs_unregister_hctxs(q
);
4749 if (blk_mq_realloc_tag_set_tags(set
, nr_hw_queues
) < 0)
4753 blk_mq_update_queue_map(set
);
4754 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
4755 blk_mq_realloc_hw_ctxs(set
, q
);
4756 blk_mq_update_poll_flag(q
);
4757 if (q
->nr_hw_queues
!= set
->nr_hw_queues
) {
4758 int i
= prev_nr_hw_queues
;
4760 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4761 nr_hw_queues
, prev_nr_hw_queues
);
4762 for (; i
< set
->nr_hw_queues
; i
++)
4763 __blk_mq_free_map_and_rqs(set
, i
);
4765 set
->nr_hw_queues
= prev_nr_hw_queues
;
4768 blk_mq_map_swqueue(q
);
4772 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
4773 blk_mq_sysfs_register_hctxs(q
);
4774 blk_mq_debugfs_register_hctxs(q
);
4778 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
4779 blk_mq_elv_switch_back(&head
, q
);
4781 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
4782 blk_mq_unfreeze_queue(q
);
4784 /* Free the excess tags when nr_hw_queues shrink. */
4785 for (i
= set
->nr_hw_queues
; i
< prev_nr_hw_queues
; i
++)
4786 __blk_mq_free_map_and_rqs(set
, i
);
4789 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
4791 mutex_lock(&set
->tag_list_lock
);
4792 __blk_mq_update_nr_hw_queues(set
, nr_hw_queues
);
4793 mutex_unlock(&set
->tag_list_lock
);
4795 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
4797 static int blk_hctx_poll(struct request_queue
*q
, struct blk_mq_hw_ctx
*hctx
,
4798 struct io_comp_batch
*iob
, unsigned int flags
)
4800 long state
= get_current_state();
4804 ret
= q
->mq_ops
->poll(hctx
, iob
);
4806 __set_current_state(TASK_RUNNING
);
4810 if (signal_pending_state(state
, current
))
4811 __set_current_state(TASK_RUNNING
);
4812 if (task_is_running(current
))
4815 if (ret
< 0 || (flags
& BLK_POLL_ONESHOT
))
4818 } while (!need_resched());
4820 __set_current_state(TASK_RUNNING
);
4824 int blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
,
4825 struct io_comp_batch
*iob
, unsigned int flags
)
4827 struct blk_mq_hw_ctx
*hctx
= xa_load(&q
->hctx_table
, cookie
);
4829 return blk_hctx_poll(q
, hctx
, iob
, flags
);
4832 int blk_rq_poll(struct request
*rq
, struct io_comp_batch
*iob
,
4833 unsigned int poll_flags
)
4835 struct request_queue
*q
= rq
->q
;
4838 if (!blk_rq_is_poll(rq
))
4840 if (!percpu_ref_tryget(&q
->q_usage_counter
))
4843 ret
= blk_hctx_poll(q
, rq
->mq_hctx
, iob
, poll_flags
);
4848 EXPORT_SYMBOL_GPL(blk_rq_poll
);
4850 unsigned int blk_mq_rq_cpu(struct request
*rq
)
4852 return rq
->mq_ctx
->cpu
;
4854 EXPORT_SYMBOL(blk_mq_rq_cpu
);
4856 void blk_mq_cancel_work_sync(struct request_queue
*q
)
4858 struct blk_mq_hw_ctx
*hctx
;
4861 cancel_delayed_work_sync(&q
->requeue_work
);
4863 queue_for_each_hw_ctx(q
, hctx
, i
)
4864 cancel_delayed_work_sync(&hctx
->run_work
);
4867 static int __init
blk_mq_init(void)
4871 for_each_possible_cpu(i
)
4872 init_llist_head(&per_cpu(blk_cpu_done
, i
));
4873 for_each_possible_cpu(i
)
4874 INIT_CSD(&per_cpu(blk_cpu_csd
, i
),
4875 __blk_mq_complete_request_remote
, NULL
);
4876 open_softirq(BLOCK_SOFTIRQ
, blk_done_softirq
);
4878 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD
,
4879 "block/softirq:dead", NULL
,
4880 blk_softirq_cpu_dead
);
4881 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
4882 blk_mq_hctx_notify_dead
);
4883 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE
, "block/mq:online",
4884 blk_mq_hctx_notify_online
,
4885 blk_mq_hctx_notify_offline
);
4888 subsys_initcall(blk_mq_init
);