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>
31 #include <linux/sched/isolation.h>
33 #include <trace/events/block.h>
35 #include <linux/t10-pi.h>
38 #include "blk-mq-debugfs.h"
41 #include "blk-mq-sched.h"
42 #include "blk-rq-qos.h"
44 static DEFINE_PER_CPU(struct llist_head
, blk_cpu_done
);
45 static DEFINE_PER_CPU(call_single_data_t
, blk_cpu_csd
);
47 static void blk_mq_insert_request(struct request
*rq
, blk_insert_t flags
);
48 static void blk_mq_request_bypass_insert(struct request
*rq
,
50 static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx
*hctx
,
51 struct list_head
*list
);
52 static int blk_hctx_poll(struct request_queue
*q
, struct blk_mq_hw_ctx
*hctx
,
53 struct io_comp_batch
*iob
, unsigned int flags
);
56 * Check if any of the ctx, dispatch list or elevator
57 * have pending work in this hardware queue.
59 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
61 return !list_empty_careful(&hctx
->dispatch
) ||
62 sbitmap_any_bit_set(&hctx
->ctx_map
) ||
63 blk_mq_sched_has_work(hctx
);
67 * Mark this ctx as having pending work in this hardware queue
69 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
70 struct blk_mq_ctx
*ctx
)
72 const int bit
= ctx
->index_hw
[hctx
->type
];
74 if (!sbitmap_test_bit(&hctx
->ctx_map
, bit
))
75 sbitmap_set_bit(&hctx
->ctx_map
, bit
);
78 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
79 struct blk_mq_ctx
*ctx
)
81 const int bit
= ctx
->index_hw
[hctx
->type
];
83 sbitmap_clear_bit(&hctx
->ctx_map
, bit
);
87 struct block_device
*part
;
88 unsigned int inflight
[2];
91 static bool blk_mq_check_inflight(struct request
*rq
, void *priv
)
93 struct mq_inflight
*mi
= priv
;
95 if (rq
->part
&& blk_do_io_stat(rq
) &&
96 (!bdev_is_partition(mi
->part
) || rq
->part
== mi
->part
) &&
97 blk_mq_rq_state(rq
) == MQ_RQ_IN_FLIGHT
)
98 mi
->inflight
[rq_data_dir(rq
)]++;
103 unsigned int blk_mq_in_flight(struct request_queue
*q
,
104 struct block_device
*part
)
106 struct mq_inflight mi
= { .part
= part
};
108 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight
, &mi
);
110 return mi
.inflight
[0] + mi
.inflight
[1];
113 void blk_mq_in_flight_rw(struct request_queue
*q
, struct block_device
*part
,
114 unsigned int inflight
[2])
116 struct mq_inflight mi
= { .part
= part
};
118 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight
, &mi
);
119 inflight
[0] = mi
.inflight
[0];
120 inflight
[1] = mi
.inflight
[1];
123 void blk_freeze_queue_start(struct request_queue
*q
)
125 mutex_lock(&q
->mq_freeze_lock
);
126 if (++q
->mq_freeze_depth
== 1) {
127 percpu_ref_kill(&q
->q_usage_counter
);
128 mutex_unlock(&q
->mq_freeze_lock
);
130 blk_mq_run_hw_queues(q
, false);
132 mutex_unlock(&q
->mq_freeze_lock
);
135 EXPORT_SYMBOL_GPL(blk_freeze_queue_start
);
137 void blk_mq_freeze_queue_wait(struct request_queue
*q
)
139 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
141 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait
);
143 int blk_mq_freeze_queue_wait_timeout(struct request_queue
*q
,
144 unsigned long timeout
)
146 return wait_event_timeout(q
->mq_freeze_wq
,
147 percpu_ref_is_zero(&q
->q_usage_counter
),
150 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout
);
153 * Guarantee no request is in use, so we can change any data structure of
154 * the queue afterward.
156 void blk_freeze_queue(struct request_queue
*q
)
159 * In the !blk_mq case we are only calling this to kill the
160 * q_usage_counter, otherwise this increases the freeze depth
161 * and waits for it to return to zero. For this reason there is
162 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
163 * exported to drivers as the only user for unfreeze is blk_mq.
165 blk_freeze_queue_start(q
);
166 blk_mq_freeze_queue_wait(q
);
169 void blk_mq_freeze_queue(struct request_queue
*q
)
172 * ...just an alias to keep freeze and unfreeze actions balanced
173 * in the blk_mq_* namespace
177 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
179 void __blk_mq_unfreeze_queue(struct request_queue
*q
, bool force_atomic
)
181 mutex_lock(&q
->mq_freeze_lock
);
183 q
->q_usage_counter
.data
->force_atomic
= true;
184 q
->mq_freeze_depth
--;
185 WARN_ON_ONCE(q
->mq_freeze_depth
< 0);
186 if (!q
->mq_freeze_depth
) {
187 percpu_ref_resurrect(&q
->q_usage_counter
);
188 wake_up_all(&q
->mq_freeze_wq
);
190 mutex_unlock(&q
->mq_freeze_lock
);
193 void blk_mq_unfreeze_queue(struct request_queue
*q
)
195 __blk_mq_unfreeze_queue(q
, false);
197 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
200 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
201 * mpt3sas driver such that this function can be removed.
203 void blk_mq_quiesce_queue_nowait(struct request_queue
*q
)
207 spin_lock_irqsave(&q
->queue_lock
, flags
);
208 if (!q
->quiesce_depth
++)
209 blk_queue_flag_set(QUEUE_FLAG_QUIESCED
, q
);
210 spin_unlock_irqrestore(&q
->queue_lock
, flags
);
212 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait
);
215 * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done
216 * @set: tag_set to wait on
218 * Note: it is driver's responsibility for making sure that quiesce has
219 * been started on or more of the request_queues of the tag_set. This
220 * function only waits for the quiesce on those request_queues that had
221 * the quiesce flag set using blk_mq_quiesce_queue_nowait.
223 void blk_mq_wait_quiesce_done(struct blk_mq_tag_set
*set
)
225 if (set
->flags
& BLK_MQ_F_BLOCKING
)
226 synchronize_srcu(set
->srcu
);
230 EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done
);
233 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
236 * Note: this function does not prevent that the struct request end_io()
237 * callback function is invoked. Once this function is returned, we make
238 * sure no dispatch can happen until the queue is unquiesced via
239 * blk_mq_unquiesce_queue().
241 void blk_mq_quiesce_queue(struct request_queue
*q
)
243 blk_mq_quiesce_queue_nowait(q
);
244 /* nothing to wait for non-mq queues */
246 blk_mq_wait_quiesce_done(q
->tag_set
);
248 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
251 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
254 * This function recovers queue into the state before quiescing
255 * which is done by blk_mq_quiesce_queue.
257 void blk_mq_unquiesce_queue(struct request_queue
*q
)
260 bool run_queue
= false;
262 spin_lock_irqsave(&q
->queue_lock
, flags
);
263 if (WARN_ON_ONCE(q
->quiesce_depth
<= 0)) {
265 } else if (!--q
->quiesce_depth
) {
266 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED
, q
);
269 spin_unlock_irqrestore(&q
->queue_lock
, flags
);
271 /* dispatch requests which are inserted during quiescing */
273 blk_mq_run_hw_queues(q
, true);
275 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue
);
277 void blk_mq_quiesce_tagset(struct blk_mq_tag_set
*set
)
279 struct request_queue
*q
;
281 mutex_lock(&set
->tag_list_lock
);
282 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
283 if (!blk_queue_skip_tagset_quiesce(q
))
284 blk_mq_quiesce_queue_nowait(q
);
286 blk_mq_wait_quiesce_done(set
);
287 mutex_unlock(&set
->tag_list_lock
);
289 EXPORT_SYMBOL_GPL(blk_mq_quiesce_tagset
);
291 void blk_mq_unquiesce_tagset(struct blk_mq_tag_set
*set
)
293 struct request_queue
*q
;
295 mutex_lock(&set
->tag_list_lock
);
296 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
297 if (!blk_queue_skip_tagset_quiesce(q
))
298 blk_mq_unquiesce_queue(q
);
300 mutex_unlock(&set
->tag_list_lock
);
302 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_tagset
);
304 void blk_mq_wake_waiters(struct request_queue
*q
)
306 struct blk_mq_hw_ctx
*hctx
;
309 queue_for_each_hw_ctx(q
, hctx
, i
)
310 if (blk_mq_hw_queue_mapped(hctx
))
311 blk_mq_tag_wakeup_all(hctx
->tags
, true);
314 void blk_rq_init(struct request_queue
*q
, struct request
*rq
)
316 memset(rq
, 0, sizeof(*rq
));
318 INIT_LIST_HEAD(&rq
->queuelist
);
320 rq
->__sector
= (sector_t
) -1;
321 INIT_HLIST_NODE(&rq
->hash
);
322 RB_CLEAR_NODE(&rq
->rb_node
);
323 rq
->tag
= BLK_MQ_NO_TAG
;
324 rq
->internal_tag
= BLK_MQ_NO_TAG
;
325 rq
->start_time_ns
= blk_time_get_ns();
327 blk_crypto_rq_set_defaults(rq
);
329 EXPORT_SYMBOL(blk_rq_init
);
331 /* Set start and alloc time when the allocated request is actually used */
332 static inline void blk_mq_rq_time_init(struct request
*rq
, u64 alloc_time_ns
)
334 if (blk_mq_need_time_stamp(rq
))
335 rq
->start_time_ns
= blk_time_get_ns();
337 rq
->start_time_ns
= 0;
339 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
340 if (blk_queue_rq_alloc_time(rq
->q
))
341 rq
->alloc_time_ns
= alloc_time_ns
?: rq
->start_time_ns
;
343 rq
->alloc_time_ns
= 0;
347 static struct request
*blk_mq_rq_ctx_init(struct blk_mq_alloc_data
*data
,
348 struct blk_mq_tags
*tags
, unsigned int tag
)
350 struct blk_mq_ctx
*ctx
= data
->ctx
;
351 struct blk_mq_hw_ctx
*hctx
= data
->hctx
;
352 struct request_queue
*q
= data
->q
;
353 struct request
*rq
= tags
->static_rqs
[tag
];
358 rq
->cmd_flags
= data
->cmd_flags
;
360 if (data
->flags
& BLK_MQ_REQ_PM
)
361 data
->rq_flags
|= RQF_PM
;
362 if (blk_queue_io_stat(q
))
363 data
->rq_flags
|= RQF_IO_STAT
;
364 rq
->rq_flags
= data
->rq_flags
;
366 if (data
->rq_flags
& RQF_SCHED_TAGS
) {
367 rq
->tag
= BLK_MQ_NO_TAG
;
368 rq
->internal_tag
= tag
;
371 rq
->internal_tag
= BLK_MQ_NO_TAG
;
376 rq
->io_start_time_ns
= 0;
377 rq
->stats_sectors
= 0;
378 rq
->nr_phys_segments
= 0;
379 #if defined(CONFIG_BLK_DEV_INTEGRITY)
380 rq
->nr_integrity_segments
= 0;
383 rq
->end_io_data
= NULL
;
385 blk_crypto_rq_set_defaults(rq
);
386 INIT_LIST_HEAD(&rq
->queuelist
);
387 /* tag was already set */
388 WRITE_ONCE(rq
->deadline
, 0);
391 if (rq
->rq_flags
& RQF_USE_SCHED
) {
392 struct elevator_queue
*e
= data
->q
->elevator
;
394 INIT_HLIST_NODE(&rq
->hash
);
395 RB_CLEAR_NODE(&rq
->rb_node
);
397 if (e
->type
->ops
.prepare_request
)
398 e
->type
->ops
.prepare_request(rq
);
404 static inline struct request
*
405 __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data
*data
)
407 unsigned int tag
, tag_offset
;
408 struct blk_mq_tags
*tags
;
410 unsigned long tag_mask
;
413 tag_mask
= blk_mq_get_tags(data
, data
->nr_tags
, &tag_offset
);
414 if (unlikely(!tag_mask
))
417 tags
= blk_mq_tags_from_data(data
);
418 for (i
= 0; tag_mask
; i
++) {
419 if (!(tag_mask
& (1UL << i
)))
421 tag
= tag_offset
+ i
;
422 prefetch(tags
->static_rqs
[tag
]);
423 tag_mask
&= ~(1UL << i
);
424 rq
= blk_mq_rq_ctx_init(data
, tags
, tag
);
425 rq_list_add(data
->cached_rq
, rq
);
428 if (!(data
->rq_flags
& RQF_SCHED_TAGS
))
429 blk_mq_add_active_requests(data
->hctx
, nr
);
430 /* caller already holds a reference, add for remainder */
431 percpu_ref_get_many(&data
->q
->q_usage_counter
, nr
- 1);
434 return rq_list_pop(data
->cached_rq
);
437 static struct request
*__blk_mq_alloc_requests(struct blk_mq_alloc_data
*data
)
439 struct request_queue
*q
= data
->q
;
440 u64 alloc_time_ns
= 0;
444 /* alloc_time includes depth and tag waits */
445 if (blk_queue_rq_alloc_time(q
))
446 alloc_time_ns
= blk_time_get_ns();
448 if (data
->cmd_flags
& REQ_NOWAIT
)
449 data
->flags
|= BLK_MQ_REQ_NOWAIT
;
452 data
->ctx
= blk_mq_get_ctx(q
);
453 data
->hctx
= blk_mq_map_queue(q
, data
->cmd_flags
, data
->ctx
);
457 * All requests use scheduler tags when an I/O scheduler is
458 * enabled for the queue.
460 data
->rq_flags
|= RQF_SCHED_TAGS
;
463 * Flush/passthrough requests are special and go directly to the
466 if ((data
->cmd_flags
& REQ_OP_MASK
) != REQ_OP_FLUSH
&&
467 !blk_op_is_passthrough(data
->cmd_flags
)) {
468 struct elevator_mq_ops
*ops
= &q
->elevator
->type
->ops
;
470 WARN_ON_ONCE(data
->flags
& BLK_MQ_REQ_RESERVED
);
472 data
->rq_flags
|= RQF_USE_SCHED
;
473 if (ops
->limit_depth
)
474 ops
->limit_depth(data
->cmd_flags
, data
);
477 blk_mq_tag_busy(data
->hctx
);
480 if (data
->flags
& BLK_MQ_REQ_RESERVED
)
481 data
->rq_flags
|= RQF_RESV
;
484 * Try batched alloc if we want more than 1 tag.
486 if (data
->nr_tags
> 1) {
487 rq
= __blk_mq_alloc_requests_batch(data
);
489 blk_mq_rq_time_init(rq
, alloc_time_ns
);
496 * Waiting allocations only fail because of an inactive hctx. In that
497 * case just retry the hctx assignment and tag allocation as CPU hotplug
498 * should have migrated us to an online CPU by now.
500 tag
= blk_mq_get_tag(data
);
501 if (tag
== BLK_MQ_NO_TAG
) {
502 if (data
->flags
& BLK_MQ_REQ_NOWAIT
)
505 * Give up the CPU and sleep for a random short time to
506 * ensure that thread using a realtime scheduling class
507 * are migrated off the CPU, and thus off the hctx that
514 if (!(data
->rq_flags
& RQF_SCHED_TAGS
))
515 blk_mq_inc_active_requests(data
->hctx
);
516 rq
= blk_mq_rq_ctx_init(data
, blk_mq_tags_from_data(data
), tag
);
517 blk_mq_rq_time_init(rq
, alloc_time_ns
);
521 static struct request
*blk_mq_rq_cache_fill(struct request_queue
*q
,
522 struct blk_plug
*plug
,
524 blk_mq_req_flags_t flags
)
526 struct blk_mq_alloc_data data
= {
530 .nr_tags
= plug
->nr_ios
,
531 .cached_rq
= &plug
->cached_rq
,
535 if (blk_queue_enter(q
, flags
))
540 rq
= __blk_mq_alloc_requests(&data
);
546 static struct request
*blk_mq_alloc_cached_request(struct request_queue
*q
,
548 blk_mq_req_flags_t flags
)
550 struct blk_plug
*plug
= current
->plug
;
556 if (rq_list_empty(plug
->cached_rq
)) {
557 if (plug
->nr_ios
== 1)
559 rq
= blk_mq_rq_cache_fill(q
, plug
, opf
, flags
);
563 rq
= rq_list_peek(&plug
->cached_rq
);
564 if (!rq
|| rq
->q
!= q
)
567 if (blk_mq_get_hctx_type(opf
) != rq
->mq_hctx
->type
)
569 if (op_is_flush(rq
->cmd_flags
) != op_is_flush(opf
))
572 plug
->cached_rq
= rq_list_next(rq
);
573 blk_mq_rq_time_init(rq
, 0);
577 INIT_LIST_HEAD(&rq
->queuelist
);
581 struct request
*blk_mq_alloc_request(struct request_queue
*q
, blk_opf_t opf
,
582 blk_mq_req_flags_t flags
)
586 rq
= blk_mq_alloc_cached_request(q
, opf
, flags
);
588 struct blk_mq_alloc_data data
= {
596 ret
= blk_queue_enter(q
, flags
);
600 rq
= __blk_mq_alloc_requests(&data
);
605 rq
->__sector
= (sector_t
) -1;
606 rq
->bio
= rq
->biotail
= NULL
;
610 return ERR_PTR(-EWOULDBLOCK
);
612 EXPORT_SYMBOL(blk_mq_alloc_request
);
614 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
,
615 blk_opf_t opf
, blk_mq_req_flags_t flags
, unsigned int hctx_idx
)
617 struct blk_mq_alloc_data data
= {
623 u64 alloc_time_ns
= 0;
629 /* alloc_time includes depth and tag waits */
630 if (blk_queue_rq_alloc_time(q
))
631 alloc_time_ns
= blk_time_get_ns();
634 * If the tag allocator sleeps we could get an allocation for a
635 * different hardware context. No need to complicate the low level
636 * allocator for this for the rare use case of a command tied to
639 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)) ||
640 WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_RESERVED
)))
641 return ERR_PTR(-EINVAL
);
643 if (hctx_idx
>= q
->nr_hw_queues
)
644 return ERR_PTR(-EIO
);
646 ret
= blk_queue_enter(q
, flags
);
651 * Check if the hardware context is actually mapped to anything.
652 * If not tell the caller that it should skip this queue.
655 data
.hctx
= xa_load(&q
->hctx_table
, hctx_idx
);
656 if (!blk_mq_hw_queue_mapped(data
.hctx
))
658 cpu
= cpumask_first_and(data
.hctx
->cpumask
, cpu_online_mask
);
659 if (cpu
>= nr_cpu_ids
)
661 data
.ctx
= __blk_mq_get_ctx(q
, cpu
);
664 data
.rq_flags
|= RQF_SCHED_TAGS
;
666 blk_mq_tag_busy(data
.hctx
);
668 if (flags
& BLK_MQ_REQ_RESERVED
)
669 data
.rq_flags
|= RQF_RESV
;
672 tag
= blk_mq_get_tag(&data
);
673 if (tag
== BLK_MQ_NO_TAG
)
675 if (!(data
.rq_flags
& RQF_SCHED_TAGS
))
676 blk_mq_inc_active_requests(data
.hctx
);
677 rq
= blk_mq_rq_ctx_init(&data
, blk_mq_tags_from_data(&data
), tag
);
678 blk_mq_rq_time_init(rq
, alloc_time_ns
);
680 rq
->__sector
= (sector_t
) -1;
681 rq
->bio
= rq
->biotail
= NULL
;
688 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
690 static void blk_mq_finish_request(struct request
*rq
)
692 struct request_queue
*q
= rq
->q
;
694 blk_zone_finish_request(rq
);
696 if (rq
->rq_flags
& RQF_USE_SCHED
) {
697 q
->elevator
->type
->ops
.finish_request(rq
);
699 * For postflush request that may need to be
700 * completed twice, we should clear this flag
701 * to avoid double finish_request() on the rq.
703 rq
->rq_flags
&= ~RQF_USE_SCHED
;
707 static void __blk_mq_free_request(struct request
*rq
)
709 struct request_queue
*q
= rq
->q
;
710 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
711 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
712 const int sched_tag
= rq
->internal_tag
;
714 blk_crypto_free_request(rq
);
715 blk_pm_mark_last_busy(rq
);
718 if (rq
->tag
!= BLK_MQ_NO_TAG
) {
719 blk_mq_dec_active_requests(hctx
);
720 blk_mq_put_tag(hctx
->tags
, ctx
, rq
->tag
);
722 if (sched_tag
!= BLK_MQ_NO_TAG
)
723 blk_mq_put_tag(hctx
->sched_tags
, ctx
, sched_tag
);
724 blk_mq_sched_restart(hctx
);
728 void blk_mq_free_request(struct request
*rq
)
730 struct request_queue
*q
= rq
->q
;
732 blk_mq_finish_request(rq
);
734 if (unlikely(laptop_mode
&& !blk_rq_is_passthrough(rq
)))
735 laptop_io_completion(q
->disk
->bdi
);
739 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
740 if (req_ref_put_and_test(rq
))
741 __blk_mq_free_request(rq
);
743 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
745 void blk_mq_free_plug_rqs(struct blk_plug
*plug
)
749 while ((rq
= rq_list_pop(&plug
->cached_rq
)) != NULL
)
750 blk_mq_free_request(rq
);
753 void blk_dump_rq_flags(struct request
*rq
, char *msg
)
755 printk(KERN_INFO
"%s: dev %s: flags=%llx\n", msg
,
756 rq
->q
->disk
? rq
->q
->disk
->disk_name
: "?",
757 (__force
unsigned long long) rq
->cmd_flags
);
759 printk(KERN_INFO
" sector %llu, nr/cnr %u/%u\n",
760 (unsigned long long)blk_rq_pos(rq
),
761 blk_rq_sectors(rq
), blk_rq_cur_sectors(rq
));
762 printk(KERN_INFO
" bio %p, biotail %p, len %u\n",
763 rq
->bio
, rq
->biotail
, blk_rq_bytes(rq
));
765 EXPORT_SYMBOL(blk_dump_rq_flags
);
767 static void blk_account_io_completion(struct request
*req
, unsigned int bytes
)
769 if (req
->part
&& blk_do_io_stat(req
)) {
770 const int sgrp
= op_stat_group(req_op(req
));
773 part_stat_add(req
->part
, sectors
[sgrp
], bytes
>> 9);
778 static void blk_print_req_error(struct request
*req
, blk_status_t status
)
780 printk_ratelimited(KERN_ERR
781 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
782 "phys_seg %u prio class %u\n",
783 blk_status_to_str(status
),
784 req
->q
->disk
? req
->q
->disk
->disk_name
: "?",
785 blk_rq_pos(req
), (__force u32
)req_op(req
),
786 blk_op_str(req_op(req
)),
787 (__force u32
)(req
->cmd_flags
& ~REQ_OP_MASK
),
788 req
->nr_phys_segments
,
789 IOPRIO_PRIO_CLASS(req
->ioprio
));
793 * Fully end IO on a request. Does not support partial completions, or
796 static void blk_complete_request(struct request
*req
)
798 const bool is_flush
= (req
->rq_flags
& RQF_FLUSH_SEQ
) != 0;
799 int total_bytes
= blk_rq_bytes(req
);
800 struct bio
*bio
= req
->bio
;
802 trace_block_rq_complete(req
, BLK_STS_OK
, total_bytes
);
807 if (blk_integrity_rq(req
) && req_op(req
) == REQ_OP_READ
)
808 blk_integrity_complete(req
, total_bytes
);
811 * Upper layers may call blk_crypto_evict_key() anytime after the last
812 * bio_endio(). Therefore, the keyslot must be released before that.
814 blk_crypto_rq_put_keyslot(req
);
816 blk_account_io_completion(req
, total_bytes
);
819 struct bio
*next
= bio
->bi_next
;
821 /* Completion has already been traced */
822 bio_clear_flag(bio
, BIO_TRACE_COMPLETION
);
824 blk_zone_update_request_bio(req
, bio
);
832 * Reset counters so that the request stacking driver
833 * can find how many bytes remain in the request
843 * blk_update_request - Complete multiple bytes without completing the request
844 * @req: the request being processed
845 * @error: block status code
846 * @nr_bytes: number of bytes to complete for @req
849 * Ends I/O on a number of bytes attached to @req, but doesn't complete
850 * the request structure even if @req doesn't have leftover.
851 * If @req has leftover, sets it up for the next range of segments.
853 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
854 * %false return from this function.
857 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
858 * except in the consistency check at the end of this function.
861 * %false - this request doesn't have any more data
862 * %true - this request has more data
864 bool blk_update_request(struct request
*req
, blk_status_t error
,
865 unsigned int nr_bytes
)
867 bool is_flush
= req
->rq_flags
& RQF_FLUSH_SEQ
;
868 bool quiet
= req
->rq_flags
& RQF_QUIET
;
871 trace_block_rq_complete(req
, error
, nr_bytes
);
876 if (blk_integrity_rq(req
) && req_op(req
) == REQ_OP_READ
&&
878 blk_integrity_complete(req
, nr_bytes
);
881 * Upper layers may call blk_crypto_evict_key() anytime after the last
882 * bio_endio(). Therefore, the keyslot must be released before that.
884 if (blk_crypto_rq_has_keyslot(req
) && nr_bytes
>= blk_rq_bytes(req
))
885 __blk_crypto_rq_put_keyslot(req
);
887 if (unlikely(error
&& !blk_rq_is_passthrough(req
) && !quiet
) &&
888 !test_bit(GD_DEAD
, &req
->q
->disk
->state
)) {
889 blk_print_req_error(req
, error
);
890 trace_block_rq_error(req
, error
, nr_bytes
);
893 blk_account_io_completion(req
, nr_bytes
);
897 struct bio
*bio
= req
->bio
;
898 unsigned bio_bytes
= min(bio
->bi_iter
.bi_size
, nr_bytes
);
901 bio
->bi_status
= error
;
903 if (bio_bytes
== bio
->bi_iter
.bi_size
) {
904 req
->bio
= bio
->bi_next
;
905 } else if (bio_is_zone_append(bio
) && error
== BLK_STS_OK
) {
907 * Partial zone append completions cannot be supported
908 * as the BIO fragments may end up not being written
911 bio
->bi_status
= BLK_STS_IOERR
;
914 /* Completion has already been traced */
915 bio_clear_flag(bio
, BIO_TRACE_COMPLETION
);
917 bio_set_flag(bio
, BIO_QUIET
);
919 bio_advance(bio
, bio_bytes
);
921 /* Don't actually finish bio if it's part of flush sequence */
922 if (!bio
->bi_iter
.bi_size
) {
923 blk_zone_update_request_bio(req
, bio
);
928 total_bytes
+= bio_bytes
;
929 nr_bytes
-= bio_bytes
;
940 * Reset counters so that the request stacking driver
941 * can find how many bytes remain in the request
948 req
->__data_len
-= total_bytes
;
950 /* update sector only for requests with clear definition of sector */
951 if (!blk_rq_is_passthrough(req
))
952 req
->__sector
+= total_bytes
>> 9;
954 /* mixed attributes always follow the first bio */
955 if (req
->rq_flags
& RQF_MIXED_MERGE
) {
956 req
->cmd_flags
&= ~REQ_FAILFAST_MASK
;
957 req
->cmd_flags
|= req
->bio
->bi_opf
& REQ_FAILFAST_MASK
;
960 if (!(req
->rq_flags
& RQF_SPECIAL_PAYLOAD
)) {
962 * If total number of sectors is less than the first segment
963 * size, something has gone terribly wrong.
965 if (blk_rq_bytes(req
) < blk_rq_cur_bytes(req
)) {
966 blk_dump_rq_flags(req
, "request botched");
967 req
->__data_len
= blk_rq_cur_bytes(req
);
970 /* recalculate the number of segments */
971 req
->nr_phys_segments
= blk_recalc_rq_segments(req
);
976 EXPORT_SYMBOL_GPL(blk_update_request
);
978 static inline void blk_account_io_done(struct request
*req
, u64 now
)
980 trace_block_io_done(req
);
983 * Account IO completion. flush_rq isn't accounted as a
984 * normal IO on queueing nor completion. Accounting the
985 * containing request is enough.
987 if (blk_do_io_stat(req
) && req
->part
&&
988 !(req
->rq_flags
& RQF_FLUSH_SEQ
)) {
989 const int sgrp
= op_stat_group(req_op(req
));
992 update_io_ticks(req
->part
, jiffies
, true);
993 part_stat_inc(req
->part
, ios
[sgrp
]);
994 part_stat_add(req
->part
, nsecs
[sgrp
], now
- req
->start_time_ns
);
995 part_stat_local_dec(req
->part
,
996 in_flight
[op_is_write(req_op(req
))]);
1001 static inline void blk_account_io_start(struct request
*req
)
1003 trace_block_io_start(req
);
1005 if (blk_do_io_stat(req
)) {
1007 * All non-passthrough requests are created from a bio with one
1008 * exception: when a flush command that is part of a flush sequence
1009 * generated by the state machine in blk-flush.c is cloned onto the
1010 * lower device by dm-multipath we can get here without a bio.
1013 req
->part
= req
->bio
->bi_bdev
;
1015 req
->part
= req
->q
->disk
->part0
;
1018 update_io_ticks(req
->part
, jiffies
, false);
1019 part_stat_local_inc(req
->part
,
1020 in_flight
[op_is_write(req_op(req
))]);
1025 static inline void __blk_mq_end_request_acct(struct request
*rq
, u64 now
)
1027 if (rq
->rq_flags
& RQF_STATS
)
1028 blk_stat_add(rq
, now
);
1030 blk_mq_sched_completed_request(rq
, now
);
1031 blk_account_io_done(rq
, now
);
1034 inline void __blk_mq_end_request(struct request
*rq
, blk_status_t error
)
1036 if (blk_mq_need_time_stamp(rq
))
1037 __blk_mq_end_request_acct(rq
, blk_time_get_ns());
1039 blk_mq_finish_request(rq
);
1042 rq_qos_done(rq
->q
, rq
);
1043 if (rq
->end_io(rq
, error
) == RQ_END_IO_FREE
)
1044 blk_mq_free_request(rq
);
1046 blk_mq_free_request(rq
);
1049 EXPORT_SYMBOL(__blk_mq_end_request
);
1051 void blk_mq_end_request(struct request
*rq
, blk_status_t error
)
1053 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
1055 __blk_mq_end_request(rq
, error
);
1057 EXPORT_SYMBOL(blk_mq_end_request
);
1059 #define TAG_COMP_BATCH 32
1061 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx
*hctx
,
1062 int *tag_array
, int nr_tags
)
1064 struct request_queue
*q
= hctx
->queue
;
1066 blk_mq_sub_active_requests(hctx
, nr_tags
);
1068 blk_mq_put_tags(hctx
->tags
, tag_array
, nr_tags
);
1069 percpu_ref_put_many(&q
->q_usage_counter
, nr_tags
);
1072 void blk_mq_end_request_batch(struct io_comp_batch
*iob
)
1074 int tags
[TAG_COMP_BATCH
], nr_tags
= 0;
1075 struct blk_mq_hw_ctx
*cur_hctx
= NULL
;
1080 now
= blk_time_get_ns();
1082 while ((rq
= rq_list_pop(&iob
->req_list
)) != NULL
) {
1084 prefetch(rq
->rq_next
);
1086 blk_complete_request(rq
);
1088 __blk_mq_end_request_acct(rq
, now
);
1090 blk_mq_finish_request(rq
);
1092 rq_qos_done(rq
->q
, rq
);
1095 * If end_io handler returns NONE, then it still has
1096 * ownership of the request.
1098 if (rq
->end_io
&& rq
->end_io(rq
, 0) == RQ_END_IO_NONE
)
1101 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
1102 if (!req_ref_put_and_test(rq
))
1105 blk_crypto_free_request(rq
);
1106 blk_pm_mark_last_busy(rq
);
1108 if (nr_tags
== TAG_COMP_BATCH
|| cur_hctx
!= rq
->mq_hctx
) {
1110 blk_mq_flush_tag_batch(cur_hctx
, tags
, nr_tags
);
1112 cur_hctx
= rq
->mq_hctx
;
1114 tags
[nr_tags
++] = rq
->tag
;
1118 blk_mq_flush_tag_batch(cur_hctx
, tags
, nr_tags
);
1120 EXPORT_SYMBOL_GPL(blk_mq_end_request_batch
);
1122 static void blk_complete_reqs(struct llist_head
*list
)
1124 struct llist_node
*entry
= llist_reverse_order(llist_del_all(list
));
1125 struct request
*rq
, *next
;
1127 llist_for_each_entry_safe(rq
, next
, entry
, ipi_list
)
1128 rq
->q
->mq_ops
->complete(rq
);
1131 static __latent_entropy
void blk_done_softirq(struct softirq_action
*h
)
1133 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done
));
1136 static int blk_softirq_cpu_dead(unsigned int cpu
)
1138 blk_complete_reqs(&per_cpu(blk_cpu_done
, cpu
));
1142 static void __blk_mq_complete_request_remote(void *data
)
1144 __raise_softirq_irqoff(BLOCK_SOFTIRQ
);
1147 static inline bool blk_mq_complete_need_ipi(struct request
*rq
)
1149 int cpu
= raw_smp_processor_id();
1151 if (!IS_ENABLED(CONFIG_SMP
) ||
1152 !test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
))
1155 * With force threaded interrupts enabled, raising softirq from an SMP
1156 * function call will always result in waking the ksoftirqd thread.
1157 * This is probably worse than completing the request on a different
1160 if (force_irqthreads())
1163 /* same CPU or cache domain and capacity? Complete locally */
1164 if (cpu
== rq
->mq_ctx
->cpu
||
1165 (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
) &&
1166 cpus_share_cache(cpu
, rq
->mq_ctx
->cpu
) &&
1167 cpus_equal_capacity(cpu
, rq
->mq_ctx
->cpu
)))
1170 /* don't try to IPI to an offline CPU */
1171 return cpu_online(rq
->mq_ctx
->cpu
);
1174 static void blk_mq_complete_send_ipi(struct request
*rq
)
1178 cpu
= rq
->mq_ctx
->cpu
;
1179 if (llist_add(&rq
->ipi_list
, &per_cpu(blk_cpu_done
, cpu
)))
1180 smp_call_function_single_async(cpu
, &per_cpu(blk_cpu_csd
, cpu
));
1183 static void blk_mq_raise_softirq(struct request
*rq
)
1185 struct llist_head
*list
;
1188 list
= this_cpu_ptr(&blk_cpu_done
);
1189 if (llist_add(&rq
->ipi_list
, list
))
1190 raise_softirq(BLOCK_SOFTIRQ
);
1194 bool blk_mq_complete_request_remote(struct request
*rq
)
1196 WRITE_ONCE(rq
->state
, MQ_RQ_COMPLETE
);
1199 * For request which hctx has only one ctx mapping,
1200 * or a polled request, always complete locally,
1201 * it's pointless to redirect the completion.
1203 if ((rq
->mq_hctx
->nr_ctx
== 1 &&
1204 rq
->mq_ctx
->cpu
== raw_smp_processor_id()) ||
1205 rq
->cmd_flags
& REQ_POLLED
)
1208 if (blk_mq_complete_need_ipi(rq
)) {
1209 blk_mq_complete_send_ipi(rq
);
1213 if (rq
->q
->nr_hw_queues
== 1) {
1214 blk_mq_raise_softirq(rq
);
1219 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote
);
1222 * blk_mq_complete_request - end I/O on a request
1223 * @rq: the request being processed
1226 * Complete a request by scheduling the ->complete_rq operation.
1228 void blk_mq_complete_request(struct request
*rq
)
1230 if (!blk_mq_complete_request_remote(rq
))
1231 rq
->q
->mq_ops
->complete(rq
);
1233 EXPORT_SYMBOL(blk_mq_complete_request
);
1236 * blk_mq_start_request - Start processing a request
1237 * @rq: Pointer to request to be started
1239 * Function used by device drivers to notify the block layer that a request
1240 * is going to be processed now, so blk layer can do proper initializations
1241 * such as starting the timeout timer.
1243 void blk_mq_start_request(struct request
*rq
)
1245 struct request_queue
*q
= rq
->q
;
1247 trace_block_rq_issue(rq
);
1249 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
) &&
1250 !blk_rq_is_passthrough(rq
)) {
1251 rq
->io_start_time_ns
= blk_time_get_ns();
1252 rq
->stats_sectors
= blk_rq_sectors(rq
);
1253 rq
->rq_flags
|= RQF_STATS
;
1254 rq_qos_issue(q
, rq
);
1257 WARN_ON_ONCE(blk_mq_rq_state(rq
) != MQ_RQ_IDLE
);
1260 WRITE_ONCE(rq
->state
, MQ_RQ_IN_FLIGHT
);
1261 rq
->mq_hctx
->tags
->rqs
[rq
->tag
] = rq
;
1263 if (blk_integrity_rq(rq
) && req_op(rq
) == REQ_OP_WRITE
)
1264 blk_integrity_prepare(rq
);
1266 if (rq
->bio
&& rq
->bio
->bi_opf
& REQ_POLLED
)
1267 WRITE_ONCE(rq
->bio
->bi_cookie
, rq
->mq_hctx
->queue_num
);
1269 EXPORT_SYMBOL(blk_mq_start_request
);
1272 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1273 * queues. This is important for md arrays to benefit from merging
1276 static inline unsigned short blk_plug_max_rq_count(struct blk_plug
*plug
)
1278 if (plug
->multiple_queues
)
1279 return BLK_MAX_REQUEST_COUNT
* 2;
1280 return BLK_MAX_REQUEST_COUNT
;
1283 static void blk_add_rq_to_plug(struct blk_plug
*plug
, struct request
*rq
)
1285 struct request
*last
= rq_list_peek(&plug
->mq_list
);
1287 if (!plug
->rq_count
) {
1288 trace_block_plug(rq
->q
);
1289 } else if (plug
->rq_count
>= blk_plug_max_rq_count(plug
) ||
1290 (!blk_queue_nomerges(rq
->q
) &&
1291 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1292 blk_mq_flush_plug_list(plug
, false);
1294 trace_block_plug(rq
->q
);
1297 if (!plug
->multiple_queues
&& last
&& last
->q
!= rq
->q
)
1298 plug
->multiple_queues
= true;
1300 * Any request allocated from sched tags can't be issued to
1301 * ->queue_rqs() directly
1303 if (!plug
->has_elevator
&& (rq
->rq_flags
& RQF_SCHED_TAGS
))
1304 plug
->has_elevator
= true;
1306 rq_list_add(&plug
->mq_list
, rq
);
1311 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1312 * @rq: request to insert
1313 * @at_head: insert request at head or tail of queue
1316 * Insert a fully prepared request at the back of the I/O scheduler queue
1317 * for execution. Don't wait for completion.
1320 * This function will invoke @done directly if the queue is dead.
1322 void blk_execute_rq_nowait(struct request
*rq
, bool at_head
)
1324 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1326 WARN_ON(irqs_disabled());
1327 WARN_ON(!blk_rq_is_passthrough(rq
));
1329 blk_account_io_start(rq
);
1331 if (current
->plug
&& !at_head
) {
1332 blk_add_rq_to_plug(current
->plug
, rq
);
1336 blk_mq_insert_request(rq
, at_head
? BLK_MQ_INSERT_AT_HEAD
: 0);
1337 blk_mq_run_hw_queue(hctx
, hctx
->flags
& BLK_MQ_F_BLOCKING
);
1339 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait
);
1341 struct blk_rq_wait
{
1342 struct completion done
;
1346 static enum rq_end_io_ret
blk_end_sync_rq(struct request
*rq
, blk_status_t ret
)
1348 struct blk_rq_wait
*wait
= rq
->end_io_data
;
1351 complete(&wait
->done
);
1352 return RQ_END_IO_NONE
;
1355 bool blk_rq_is_poll(struct request
*rq
)
1359 if (rq
->mq_hctx
->type
!= HCTX_TYPE_POLL
)
1363 EXPORT_SYMBOL_GPL(blk_rq_is_poll
);
1365 static void blk_rq_poll_completion(struct request
*rq
, struct completion
*wait
)
1368 blk_hctx_poll(rq
->q
, rq
->mq_hctx
, NULL
, 0);
1370 } while (!completion_done(wait
));
1374 * blk_execute_rq - insert a request into queue for execution
1375 * @rq: request to insert
1376 * @at_head: insert request at head or tail of queue
1379 * Insert a fully prepared request at the back of the I/O scheduler queue
1380 * for execution and wait for completion.
1381 * Return: The blk_status_t result provided to blk_mq_end_request().
1383 blk_status_t
blk_execute_rq(struct request
*rq
, bool at_head
)
1385 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1386 struct blk_rq_wait wait
= {
1387 .done
= COMPLETION_INITIALIZER_ONSTACK(wait
.done
),
1390 WARN_ON(irqs_disabled());
1391 WARN_ON(!blk_rq_is_passthrough(rq
));
1393 rq
->end_io_data
= &wait
;
1394 rq
->end_io
= blk_end_sync_rq
;
1396 blk_account_io_start(rq
);
1397 blk_mq_insert_request(rq
, at_head
? BLK_MQ_INSERT_AT_HEAD
: 0);
1398 blk_mq_run_hw_queue(hctx
, false);
1400 if (blk_rq_is_poll(rq
))
1401 blk_rq_poll_completion(rq
, &wait
.done
);
1403 blk_wait_io(&wait
.done
);
1407 EXPORT_SYMBOL(blk_execute_rq
);
1409 static void __blk_mq_requeue_request(struct request
*rq
)
1411 struct request_queue
*q
= rq
->q
;
1413 blk_mq_put_driver_tag(rq
);
1415 trace_block_rq_requeue(rq
);
1416 rq_qos_requeue(q
, rq
);
1418 if (blk_mq_request_started(rq
)) {
1419 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
1420 rq
->rq_flags
&= ~RQF_TIMED_OUT
;
1424 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
1426 struct request_queue
*q
= rq
->q
;
1427 unsigned long flags
;
1429 __blk_mq_requeue_request(rq
);
1431 /* this request will be re-inserted to io scheduler queue */
1432 blk_mq_sched_requeue_request(rq
);
1434 spin_lock_irqsave(&q
->requeue_lock
, flags
);
1435 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
1436 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
1438 if (kick_requeue_list
)
1439 blk_mq_kick_requeue_list(q
);
1441 EXPORT_SYMBOL(blk_mq_requeue_request
);
1443 static void blk_mq_requeue_work(struct work_struct
*work
)
1445 struct request_queue
*q
=
1446 container_of(work
, struct request_queue
, requeue_work
.work
);
1448 LIST_HEAD(flush_list
);
1451 spin_lock_irq(&q
->requeue_lock
);
1452 list_splice_init(&q
->requeue_list
, &rq_list
);
1453 list_splice_init(&q
->flush_list
, &flush_list
);
1454 spin_unlock_irq(&q
->requeue_lock
);
1456 while (!list_empty(&rq_list
)) {
1457 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
1459 * If RQF_DONTPREP ist set, the request has been started by the
1460 * driver already and might have driver-specific data allocated
1461 * already. Insert it into the hctx dispatch list to avoid
1462 * block layer merges for the request.
1464 if (rq
->rq_flags
& RQF_DONTPREP
) {
1465 list_del_init(&rq
->queuelist
);
1466 blk_mq_request_bypass_insert(rq
, 0);
1468 list_del_init(&rq
->queuelist
);
1469 blk_mq_insert_request(rq
, BLK_MQ_INSERT_AT_HEAD
);
1473 while (!list_empty(&flush_list
)) {
1474 rq
= list_entry(flush_list
.next
, struct request
, queuelist
);
1475 list_del_init(&rq
->queuelist
);
1476 blk_mq_insert_request(rq
, 0);
1479 blk_mq_run_hw_queues(q
, false);
1482 void blk_mq_kick_requeue_list(struct request_queue
*q
)
1484 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
, 0);
1486 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
1488 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
1489 unsigned long msecs
)
1491 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
,
1492 msecs_to_jiffies(msecs
));
1494 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
1496 static bool blk_is_flush_data_rq(struct request
*rq
)
1498 return (rq
->rq_flags
& RQF_FLUSH_SEQ
) && !is_flush_rq(rq
);
1501 static bool blk_mq_rq_inflight(struct request
*rq
, void *priv
)
1504 * If we find a request that isn't idle we know the queue is busy
1505 * as it's checked in the iter.
1506 * Return false to stop the iteration.
1508 * In case of queue quiesce, if one flush data request is completed,
1509 * don't count it as inflight given the flush sequence is suspended,
1510 * and the original flush data request is invisible to driver, just
1511 * like other pending requests because of quiesce
1513 if (blk_mq_request_started(rq
) && !(blk_queue_quiesced(rq
->q
) &&
1514 blk_is_flush_data_rq(rq
) &&
1515 blk_mq_request_completed(rq
))) {
1525 bool blk_mq_queue_inflight(struct request_queue
*q
)
1529 blk_mq_queue_tag_busy_iter(q
, blk_mq_rq_inflight
, &busy
);
1532 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight
);
1534 static void blk_mq_rq_timed_out(struct request
*req
)
1536 req
->rq_flags
|= RQF_TIMED_OUT
;
1537 if (req
->q
->mq_ops
->timeout
) {
1538 enum blk_eh_timer_return ret
;
1540 ret
= req
->q
->mq_ops
->timeout(req
);
1541 if (ret
== BLK_EH_DONE
)
1543 WARN_ON_ONCE(ret
!= BLK_EH_RESET_TIMER
);
1549 struct blk_expired_data
{
1550 bool has_timedout_rq
;
1552 unsigned long timeout_start
;
1555 static bool blk_mq_req_expired(struct request
*rq
, struct blk_expired_data
*expired
)
1557 unsigned long deadline
;
1559 if (blk_mq_rq_state(rq
) != MQ_RQ_IN_FLIGHT
)
1561 if (rq
->rq_flags
& RQF_TIMED_OUT
)
1564 deadline
= READ_ONCE(rq
->deadline
);
1565 if (time_after_eq(expired
->timeout_start
, deadline
))
1568 if (expired
->next
== 0)
1569 expired
->next
= deadline
;
1570 else if (time_after(expired
->next
, deadline
))
1571 expired
->next
= deadline
;
1575 void blk_mq_put_rq_ref(struct request
*rq
)
1577 if (is_flush_rq(rq
)) {
1578 if (rq
->end_io(rq
, 0) == RQ_END_IO_FREE
)
1579 blk_mq_free_request(rq
);
1580 } else if (req_ref_put_and_test(rq
)) {
1581 __blk_mq_free_request(rq
);
1585 static bool blk_mq_check_expired(struct request
*rq
, void *priv
)
1587 struct blk_expired_data
*expired
= priv
;
1590 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1591 * be reallocated underneath the timeout handler's processing, then
1592 * the expire check is reliable. If the request is not expired, then
1593 * it was completed and reallocated as a new request after returning
1594 * from blk_mq_check_expired().
1596 if (blk_mq_req_expired(rq
, expired
)) {
1597 expired
->has_timedout_rq
= true;
1603 static bool blk_mq_handle_expired(struct request
*rq
, void *priv
)
1605 struct blk_expired_data
*expired
= priv
;
1607 if (blk_mq_req_expired(rq
, expired
))
1608 blk_mq_rq_timed_out(rq
);
1612 static void blk_mq_timeout_work(struct work_struct
*work
)
1614 struct request_queue
*q
=
1615 container_of(work
, struct request_queue
, timeout_work
);
1616 struct blk_expired_data expired
= {
1617 .timeout_start
= jiffies
,
1619 struct blk_mq_hw_ctx
*hctx
;
1622 /* A deadlock might occur if a request is stuck requiring a
1623 * timeout at the same time a queue freeze is waiting
1624 * completion, since the timeout code would not be able to
1625 * acquire the queue reference here.
1627 * That's why we don't use blk_queue_enter here; instead, we use
1628 * percpu_ref_tryget directly, because we need to be able to
1629 * obtain a reference even in the short window between the queue
1630 * starting to freeze, by dropping the first reference in
1631 * blk_freeze_queue_start, and the moment the last request is
1632 * consumed, marked by the instant q_usage_counter reaches
1635 if (!percpu_ref_tryget(&q
->q_usage_counter
))
1638 /* check if there is any timed-out request */
1639 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &expired
);
1640 if (expired
.has_timedout_rq
) {
1642 * Before walking tags, we must ensure any submit started
1643 * before the current time has finished. Since the submit
1644 * uses srcu or rcu, wait for a synchronization point to
1645 * ensure all running submits have finished
1647 blk_mq_wait_quiesce_done(q
->tag_set
);
1650 blk_mq_queue_tag_busy_iter(q
, blk_mq_handle_expired
, &expired
);
1653 if (expired
.next
!= 0) {
1654 mod_timer(&q
->timeout
, expired
.next
);
1657 * Request timeouts are handled as a forward rolling timer. If
1658 * we end up here it means that no requests are pending and
1659 * also that no request has been pending for a while. Mark
1660 * each hctx as idle.
1662 queue_for_each_hw_ctx(q
, hctx
, i
) {
1663 /* the hctx may be unmapped, so check it here */
1664 if (blk_mq_hw_queue_mapped(hctx
))
1665 blk_mq_tag_idle(hctx
);
1671 struct flush_busy_ctx_data
{
1672 struct blk_mq_hw_ctx
*hctx
;
1673 struct list_head
*list
;
1676 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
1678 struct flush_busy_ctx_data
*flush_data
= data
;
1679 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
1680 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
1681 enum hctx_type type
= hctx
->type
;
1683 spin_lock(&ctx
->lock
);
1684 list_splice_tail_init(&ctx
->rq_lists
[type
], flush_data
->list
);
1685 sbitmap_clear_bit(sb
, bitnr
);
1686 spin_unlock(&ctx
->lock
);
1691 * Process software queues that have been marked busy, splicing them
1692 * to the for-dispatch
1694 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
1696 struct flush_busy_ctx_data data
= {
1701 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
1703 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
1705 struct dispatch_rq_data
{
1706 struct blk_mq_hw_ctx
*hctx
;
1710 static bool dispatch_rq_from_ctx(struct sbitmap
*sb
, unsigned int bitnr
,
1713 struct dispatch_rq_data
*dispatch_data
= data
;
1714 struct blk_mq_hw_ctx
*hctx
= dispatch_data
->hctx
;
1715 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
1716 enum hctx_type type
= hctx
->type
;
1718 spin_lock(&ctx
->lock
);
1719 if (!list_empty(&ctx
->rq_lists
[type
])) {
1720 dispatch_data
->rq
= list_entry_rq(ctx
->rq_lists
[type
].next
);
1721 list_del_init(&dispatch_data
->rq
->queuelist
);
1722 if (list_empty(&ctx
->rq_lists
[type
]))
1723 sbitmap_clear_bit(sb
, bitnr
);
1725 spin_unlock(&ctx
->lock
);
1727 return !dispatch_data
->rq
;
1730 struct request
*blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx
*hctx
,
1731 struct blk_mq_ctx
*start
)
1733 unsigned off
= start
? start
->index_hw
[hctx
->type
] : 0;
1734 struct dispatch_rq_data data
= {
1739 __sbitmap_for_each_set(&hctx
->ctx_map
, off
,
1740 dispatch_rq_from_ctx
, &data
);
1745 bool __blk_mq_alloc_driver_tag(struct request
*rq
)
1747 struct sbitmap_queue
*bt
= &rq
->mq_hctx
->tags
->bitmap_tags
;
1748 unsigned int tag_offset
= rq
->mq_hctx
->tags
->nr_reserved_tags
;
1751 blk_mq_tag_busy(rq
->mq_hctx
);
1753 if (blk_mq_tag_is_reserved(rq
->mq_hctx
->sched_tags
, rq
->internal_tag
)) {
1754 bt
= &rq
->mq_hctx
->tags
->breserved_tags
;
1757 if (!hctx_may_queue(rq
->mq_hctx
, bt
))
1761 tag
= __sbitmap_queue_get(bt
);
1762 if (tag
== BLK_MQ_NO_TAG
)
1765 rq
->tag
= tag
+ tag_offset
;
1766 blk_mq_inc_active_requests(rq
->mq_hctx
);
1770 static int blk_mq_dispatch_wake(wait_queue_entry_t
*wait
, unsigned mode
,
1771 int flags
, void *key
)
1773 struct blk_mq_hw_ctx
*hctx
;
1775 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
1777 spin_lock(&hctx
->dispatch_wait_lock
);
1778 if (!list_empty(&wait
->entry
)) {
1779 struct sbitmap_queue
*sbq
;
1781 list_del_init(&wait
->entry
);
1782 sbq
= &hctx
->tags
->bitmap_tags
;
1783 atomic_dec(&sbq
->ws_active
);
1785 spin_unlock(&hctx
->dispatch_wait_lock
);
1787 blk_mq_run_hw_queue(hctx
, true);
1792 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1793 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1794 * restart. For both cases, take care to check the condition again after
1795 * marking us as waiting.
1797 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx
*hctx
,
1800 struct sbitmap_queue
*sbq
;
1801 struct wait_queue_head
*wq
;
1802 wait_queue_entry_t
*wait
;
1805 if (!(hctx
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
) &&
1806 !(blk_mq_is_shared_tags(hctx
->flags
))) {
1807 blk_mq_sched_mark_restart_hctx(hctx
);
1810 * It's possible that a tag was freed in the window between the
1811 * allocation failure and adding the hardware queue to the wait
1814 * Don't clear RESTART here, someone else could have set it.
1815 * At most this will cost an extra queue run.
1817 return blk_mq_get_driver_tag(rq
);
1820 wait
= &hctx
->dispatch_wait
;
1821 if (!list_empty_careful(&wait
->entry
))
1824 if (blk_mq_tag_is_reserved(rq
->mq_hctx
->sched_tags
, rq
->internal_tag
))
1825 sbq
= &hctx
->tags
->breserved_tags
;
1827 sbq
= &hctx
->tags
->bitmap_tags
;
1828 wq
= &bt_wait_ptr(sbq
, hctx
)->wait
;
1830 spin_lock_irq(&wq
->lock
);
1831 spin_lock(&hctx
->dispatch_wait_lock
);
1832 if (!list_empty(&wait
->entry
)) {
1833 spin_unlock(&hctx
->dispatch_wait_lock
);
1834 spin_unlock_irq(&wq
->lock
);
1838 atomic_inc(&sbq
->ws_active
);
1839 wait
->flags
&= ~WQ_FLAG_EXCLUSIVE
;
1840 __add_wait_queue(wq
, wait
);
1843 * Add one explicit barrier since blk_mq_get_driver_tag() may
1844 * not imply barrier in case of failure.
1846 * Order adding us to wait queue and allocating driver tag.
1848 * The pair is the one implied in sbitmap_queue_wake_up() which
1849 * orders clearing sbitmap tag bits and waitqueue_active() in
1850 * __sbitmap_queue_wake_up(), since waitqueue_active() is lockless
1852 * Otherwise, re-order of adding wait queue and getting driver tag
1853 * may cause __sbitmap_queue_wake_up() to wake up nothing because
1854 * the waitqueue_active() may not observe us in wait queue.
1859 * It's possible that a tag was freed in the window between the
1860 * allocation failure and adding the hardware queue to the wait
1863 ret
= blk_mq_get_driver_tag(rq
);
1865 spin_unlock(&hctx
->dispatch_wait_lock
);
1866 spin_unlock_irq(&wq
->lock
);
1871 * We got a tag, remove ourselves from the wait queue to ensure
1872 * someone else gets the wakeup.
1874 list_del_init(&wait
->entry
);
1875 atomic_dec(&sbq
->ws_active
);
1876 spin_unlock(&hctx
->dispatch_wait_lock
);
1877 spin_unlock_irq(&wq
->lock
);
1882 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1883 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1885 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1886 * - EWMA is one simple way to compute running average value
1887 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1888 * - take 4 as factor for avoiding to get too small(0) result, and this
1889 * factor doesn't matter because EWMA decreases exponentially
1891 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx
*hctx
, bool busy
)
1895 ewma
= hctx
->dispatch_busy
;
1900 ewma
*= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
- 1;
1902 ewma
+= 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR
;
1903 ewma
/= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
;
1905 hctx
->dispatch_busy
= ewma
;
1908 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1910 static void blk_mq_handle_dev_resource(struct request
*rq
,
1911 struct list_head
*list
)
1913 list_add(&rq
->queuelist
, list
);
1914 __blk_mq_requeue_request(rq
);
1917 enum prep_dispatch
{
1919 PREP_DISPATCH_NO_TAG
,
1920 PREP_DISPATCH_NO_BUDGET
,
1923 static enum prep_dispatch
blk_mq_prep_dispatch_rq(struct request
*rq
,
1926 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1927 int budget_token
= -1;
1930 budget_token
= blk_mq_get_dispatch_budget(rq
->q
);
1931 if (budget_token
< 0) {
1932 blk_mq_put_driver_tag(rq
);
1933 return PREP_DISPATCH_NO_BUDGET
;
1935 blk_mq_set_rq_budget_token(rq
, budget_token
);
1938 if (!blk_mq_get_driver_tag(rq
)) {
1940 * The initial allocation attempt failed, so we need to
1941 * rerun the hardware queue when a tag is freed. The
1942 * waitqueue takes care of that. If the queue is run
1943 * before we add this entry back on the dispatch list,
1944 * we'll re-run it below.
1946 if (!blk_mq_mark_tag_wait(hctx
, rq
)) {
1948 * All budgets not got from this function will be put
1949 * together during handling partial dispatch
1952 blk_mq_put_dispatch_budget(rq
->q
, budget_token
);
1953 return PREP_DISPATCH_NO_TAG
;
1957 return PREP_DISPATCH_OK
;
1960 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1961 static void blk_mq_release_budgets(struct request_queue
*q
,
1962 struct list_head
*list
)
1966 list_for_each_entry(rq
, list
, queuelist
) {
1967 int budget_token
= blk_mq_get_rq_budget_token(rq
);
1969 if (budget_token
>= 0)
1970 blk_mq_put_dispatch_budget(q
, budget_token
);
1975 * blk_mq_commit_rqs will notify driver using bd->last that there is no
1976 * more requests. (See comment in struct blk_mq_ops for commit_rqs for
1978 * Attention, we should explicitly call this in unusual cases:
1979 * 1) did not queue everything initially scheduled to queue
1980 * 2) the last attempt to queue a request failed
1982 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx
*hctx
, int queued
,
1985 if (hctx
->queue
->mq_ops
->commit_rqs
&& queued
) {
1986 trace_block_unplug(hctx
->queue
, queued
, !from_schedule
);
1987 hctx
->queue
->mq_ops
->commit_rqs(hctx
);
1992 * Returns true if we did some work AND can potentially do more.
1994 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
,
1995 unsigned int nr_budgets
)
1997 enum prep_dispatch prep
;
1998 struct request_queue
*q
= hctx
->queue
;
2001 blk_status_t ret
= BLK_STS_OK
;
2002 bool needs_resource
= false;
2004 if (list_empty(list
))
2008 * Now process all the entries, sending them to the driver.
2012 struct blk_mq_queue_data bd
;
2014 rq
= list_first_entry(list
, struct request
, queuelist
);
2016 WARN_ON_ONCE(hctx
!= rq
->mq_hctx
);
2017 prep
= blk_mq_prep_dispatch_rq(rq
, !nr_budgets
);
2018 if (prep
!= PREP_DISPATCH_OK
)
2021 list_del_init(&rq
->queuelist
);
2024 bd
.last
= list_empty(list
);
2027 * once the request is queued to lld, no need to cover the
2032 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
2037 case BLK_STS_RESOURCE
:
2038 needs_resource
= true;
2040 case BLK_STS_DEV_RESOURCE
:
2041 blk_mq_handle_dev_resource(rq
, list
);
2044 blk_mq_end_request(rq
, ret
);
2046 } while (!list_empty(list
));
2048 /* If we didn't flush the entire list, we could have told the driver
2049 * there was more coming, but that turned out to be a lie.
2051 if (!list_empty(list
) || ret
!= BLK_STS_OK
)
2052 blk_mq_commit_rqs(hctx
, queued
, false);
2055 * Any items that need requeuing? Stuff them into hctx->dispatch,
2056 * that is where we will continue on next queue run.
2058 if (!list_empty(list
)) {
2060 /* For non-shared tags, the RESTART check will suffice */
2061 bool no_tag
= prep
== PREP_DISPATCH_NO_TAG
&&
2062 ((hctx
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
) ||
2063 blk_mq_is_shared_tags(hctx
->flags
));
2066 blk_mq_release_budgets(q
, list
);
2068 spin_lock(&hctx
->lock
);
2069 list_splice_tail_init(list
, &hctx
->dispatch
);
2070 spin_unlock(&hctx
->lock
);
2073 * Order adding requests to hctx->dispatch and checking
2074 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2075 * in blk_mq_sched_restart(). Avoid restart code path to
2076 * miss the new added requests to hctx->dispatch, meantime
2077 * SCHED_RESTART is observed here.
2082 * If SCHED_RESTART was set by the caller of this function and
2083 * it is no longer set that means that it was cleared by another
2084 * thread and hence that a queue rerun is needed.
2086 * If 'no_tag' is set, that means that we failed getting
2087 * a driver tag with an I/O scheduler attached. If our dispatch
2088 * waitqueue is no longer active, ensure that we run the queue
2089 * AFTER adding our entries back to the list.
2091 * If no I/O scheduler has been configured it is possible that
2092 * the hardware queue got stopped and restarted before requests
2093 * were pushed back onto the dispatch list. Rerun the queue to
2094 * avoid starvation. Notes:
2095 * - blk_mq_run_hw_queue() checks whether or not a queue has
2096 * been stopped before rerunning a queue.
2097 * - Some but not all block drivers stop a queue before
2098 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2101 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2102 * bit is set, run queue after a delay to avoid IO stalls
2103 * that could otherwise occur if the queue is idle. We'll do
2104 * similar if we couldn't get budget or couldn't lock a zone
2105 * and SCHED_RESTART is set.
2107 needs_restart
= blk_mq_sched_needs_restart(hctx
);
2108 if (prep
== PREP_DISPATCH_NO_BUDGET
)
2109 needs_resource
= true;
2110 if (!needs_restart
||
2111 (no_tag
&& list_empty_careful(&hctx
->dispatch_wait
.entry
)))
2112 blk_mq_run_hw_queue(hctx
, true);
2113 else if (needs_resource
)
2114 blk_mq_delay_run_hw_queue(hctx
, BLK_MQ_RESOURCE_DELAY
);
2116 blk_mq_update_dispatch_busy(hctx
, true);
2120 blk_mq_update_dispatch_busy(hctx
, false);
2124 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx
*hctx
)
2126 int cpu
= cpumask_first_and(hctx
->cpumask
, cpu_online_mask
);
2128 if (cpu
>= nr_cpu_ids
)
2129 cpu
= cpumask_first(hctx
->cpumask
);
2134 * ->next_cpu is always calculated from hctx->cpumask, so simply use
2135 * it for speeding up the check
2137 static bool blk_mq_hctx_empty_cpumask(struct blk_mq_hw_ctx
*hctx
)
2139 return hctx
->next_cpu
>= nr_cpu_ids
;
2143 * It'd be great if the workqueue API had a way to pass
2144 * in a mask and had some smarts for more clever placement.
2145 * For now we just round-robin here, switching for every
2146 * BLK_MQ_CPU_WORK_BATCH queued items.
2148 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
2151 int next_cpu
= hctx
->next_cpu
;
2153 /* Switch to unbound if no allowable CPUs in this hctx */
2154 if (hctx
->queue
->nr_hw_queues
== 1 || blk_mq_hctx_empty_cpumask(hctx
))
2155 return WORK_CPU_UNBOUND
;
2157 if (--hctx
->next_cpu_batch
<= 0) {
2159 next_cpu
= cpumask_next_and(next_cpu
, hctx
->cpumask
,
2161 if (next_cpu
>= nr_cpu_ids
)
2162 next_cpu
= blk_mq_first_mapped_cpu(hctx
);
2163 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2167 * Do unbound schedule if we can't find a online CPU for this hctx,
2168 * and it should only happen in the path of handling CPU DEAD.
2170 if (!cpu_online(next_cpu
)) {
2177 * Make sure to re-select CPU next time once after CPUs
2178 * in hctx->cpumask become online again.
2180 hctx
->next_cpu
= next_cpu
;
2181 hctx
->next_cpu_batch
= 1;
2182 return WORK_CPU_UNBOUND
;
2185 hctx
->next_cpu
= next_cpu
;
2190 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2191 * @hctx: Pointer to the hardware queue to run.
2192 * @msecs: Milliseconds of delay to wait before running the queue.
2194 * Run a hardware queue asynchronously with a delay of @msecs.
2196 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
2198 if (unlikely(blk_mq_hctx_stopped(hctx
)))
2200 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
,
2201 msecs_to_jiffies(msecs
));
2203 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue
);
2206 * blk_mq_run_hw_queue - Start to run a hardware queue.
2207 * @hctx: Pointer to the hardware queue to run.
2208 * @async: If we want to run the queue asynchronously.
2210 * Check if the request queue is not in a quiesced state and if there are
2211 * pending requests to be sent. If this is true, run the queue to send requests
2214 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
2219 * We can't run the queue inline with interrupts disabled.
2221 WARN_ON_ONCE(!async
&& in_interrupt());
2223 might_sleep_if(!async
&& hctx
->flags
& BLK_MQ_F_BLOCKING
);
2226 * When queue is quiesced, we may be switching io scheduler, or
2227 * updating nr_hw_queues, or other things, and we can't run queue
2228 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2230 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2233 __blk_mq_run_dispatch_ops(hctx
->queue
, false,
2234 need_run
= !blk_queue_quiesced(hctx
->queue
) &&
2235 blk_mq_hctx_has_pending(hctx
));
2240 if (async
|| !cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
)) {
2241 blk_mq_delay_run_hw_queue(hctx
, 0);
2245 blk_mq_run_dispatch_ops(hctx
->queue
,
2246 blk_mq_sched_dispatch_requests(hctx
));
2248 EXPORT_SYMBOL(blk_mq_run_hw_queue
);
2251 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2254 static struct blk_mq_hw_ctx
*blk_mq_get_sq_hctx(struct request_queue
*q
)
2256 struct blk_mq_ctx
*ctx
= blk_mq_get_ctx(q
);
2258 * If the IO scheduler does not respect hardware queues when
2259 * dispatching, we just don't bother with multiple HW queues and
2260 * dispatch from hctx for the current CPU since running multiple queues
2261 * just causes lock contention inside the scheduler and pointless cache
2264 struct blk_mq_hw_ctx
*hctx
= ctx
->hctxs
[HCTX_TYPE_DEFAULT
];
2266 if (!blk_mq_hctx_stopped(hctx
))
2272 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2273 * @q: Pointer to the request queue to run.
2274 * @async: If we want to run the queue asynchronously.
2276 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
2278 struct blk_mq_hw_ctx
*hctx
, *sq_hctx
;
2282 if (blk_queue_sq_sched(q
))
2283 sq_hctx
= blk_mq_get_sq_hctx(q
);
2284 queue_for_each_hw_ctx(q
, hctx
, i
) {
2285 if (blk_mq_hctx_stopped(hctx
))
2288 * Dispatch from this hctx either if there's no hctx preferred
2289 * by IO scheduler or if it has requests that bypass the
2292 if (!sq_hctx
|| sq_hctx
== hctx
||
2293 !list_empty_careful(&hctx
->dispatch
))
2294 blk_mq_run_hw_queue(hctx
, async
);
2297 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
2300 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2301 * @q: Pointer to the request queue to run.
2302 * @msecs: Milliseconds of delay to wait before running the queues.
2304 void blk_mq_delay_run_hw_queues(struct request_queue
*q
, unsigned long msecs
)
2306 struct blk_mq_hw_ctx
*hctx
, *sq_hctx
;
2310 if (blk_queue_sq_sched(q
))
2311 sq_hctx
= blk_mq_get_sq_hctx(q
);
2312 queue_for_each_hw_ctx(q
, hctx
, i
) {
2313 if (blk_mq_hctx_stopped(hctx
))
2316 * If there is already a run_work pending, leave the
2317 * pending delay untouched. Otherwise, a hctx can stall
2318 * if another hctx is re-delaying the other's work
2319 * before the work executes.
2321 if (delayed_work_pending(&hctx
->run_work
))
2324 * Dispatch from this hctx either if there's no hctx preferred
2325 * by IO scheduler or if it has requests that bypass the
2328 if (!sq_hctx
|| sq_hctx
== hctx
||
2329 !list_empty_careful(&hctx
->dispatch
))
2330 blk_mq_delay_run_hw_queue(hctx
, msecs
);
2333 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues
);
2336 * This function is often used for pausing .queue_rq() by driver when
2337 * there isn't enough resource or some conditions aren't satisfied, and
2338 * BLK_STS_RESOURCE is usually returned.
2340 * We do not guarantee that dispatch can be drained or blocked
2341 * after blk_mq_stop_hw_queue() returns. Please use
2342 * blk_mq_quiesce_queue() for that requirement.
2344 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
2346 cancel_delayed_work(&hctx
->run_work
);
2348 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
2350 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
2353 * This function is often used for pausing .queue_rq() by driver when
2354 * there isn't enough resource or some conditions aren't satisfied, and
2355 * BLK_STS_RESOURCE is usually returned.
2357 * We do not guarantee that dispatch can be drained or blocked
2358 * after blk_mq_stop_hw_queues() returns. Please use
2359 * blk_mq_quiesce_queue() for that requirement.
2361 void blk_mq_stop_hw_queues(struct request_queue
*q
)
2363 struct blk_mq_hw_ctx
*hctx
;
2366 queue_for_each_hw_ctx(q
, hctx
, i
)
2367 blk_mq_stop_hw_queue(hctx
);
2369 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
2371 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
2373 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
2375 blk_mq_run_hw_queue(hctx
, hctx
->flags
& BLK_MQ_F_BLOCKING
);
2377 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
2379 void blk_mq_start_hw_queues(struct request_queue
*q
)
2381 struct blk_mq_hw_ctx
*hctx
;
2384 queue_for_each_hw_ctx(q
, hctx
, i
)
2385 blk_mq_start_hw_queue(hctx
);
2387 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
2389 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
2391 if (!blk_mq_hctx_stopped(hctx
))
2394 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
2395 blk_mq_run_hw_queue(hctx
, async
);
2397 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
2399 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
2401 struct blk_mq_hw_ctx
*hctx
;
2404 queue_for_each_hw_ctx(q
, hctx
, i
)
2405 blk_mq_start_stopped_hw_queue(hctx
, async
||
2406 (hctx
->flags
& BLK_MQ_F_BLOCKING
));
2408 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
2410 static void blk_mq_run_work_fn(struct work_struct
*work
)
2412 struct blk_mq_hw_ctx
*hctx
=
2413 container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
2415 blk_mq_run_dispatch_ops(hctx
->queue
,
2416 blk_mq_sched_dispatch_requests(hctx
));
2420 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2421 * @rq: Pointer to request to be inserted.
2422 * @flags: BLK_MQ_INSERT_*
2424 * Should only be used carefully, when the caller knows we want to
2425 * bypass a potential IO scheduler on the target device.
2427 static void blk_mq_request_bypass_insert(struct request
*rq
, blk_insert_t flags
)
2429 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
2431 spin_lock(&hctx
->lock
);
2432 if (flags
& BLK_MQ_INSERT_AT_HEAD
)
2433 list_add(&rq
->queuelist
, &hctx
->dispatch
);
2435 list_add_tail(&rq
->queuelist
, &hctx
->dispatch
);
2436 spin_unlock(&hctx
->lock
);
2439 static void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
,
2440 struct blk_mq_ctx
*ctx
, struct list_head
*list
,
2441 bool run_queue_async
)
2444 enum hctx_type type
= hctx
->type
;
2447 * Try to issue requests directly if the hw queue isn't busy to save an
2448 * extra enqueue & dequeue to the sw queue.
2450 if (!hctx
->dispatch_busy
&& !run_queue_async
) {
2451 blk_mq_run_dispatch_ops(hctx
->queue
,
2452 blk_mq_try_issue_list_directly(hctx
, list
));
2453 if (list_empty(list
))
2458 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2461 list_for_each_entry(rq
, list
, queuelist
) {
2462 BUG_ON(rq
->mq_ctx
!= ctx
);
2463 trace_block_rq_insert(rq
);
2464 if (rq
->cmd_flags
& REQ_NOWAIT
)
2465 run_queue_async
= true;
2468 spin_lock(&ctx
->lock
);
2469 list_splice_tail_init(list
, &ctx
->rq_lists
[type
]);
2470 blk_mq_hctx_mark_pending(hctx
, ctx
);
2471 spin_unlock(&ctx
->lock
);
2473 blk_mq_run_hw_queue(hctx
, run_queue_async
);
2476 static void blk_mq_insert_request(struct request
*rq
, blk_insert_t flags
)
2478 struct request_queue
*q
= rq
->q
;
2479 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
2480 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
2482 if (blk_rq_is_passthrough(rq
)) {
2484 * Passthrough request have to be added to hctx->dispatch
2485 * directly. The device may be in a situation where it can't
2486 * handle FS request, and always returns BLK_STS_RESOURCE for
2487 * them, which gets them added to hctx->dispatch.
2489 * If a passthrough request is required to unblock the queues,
2490 * and it is added to the scheduler queue, there is no chance to
2491 * dispatch it given we prioritize requests in hctx->dispatch.
2493 blk_mq_request_bypass_insert(rq
, flags
);
2494 } else if (req_op(rq
) == REQ_OP_FLUSH
) {
2496 * Firstly normal IO request is inserted to scheduler queue or
2497 * sw queue, meantime we add flush request to dispatch queue(
2498 * hctx->dispatch) directly and there is at most one in-flight
2499 * flush request for each hw queue, so it doesn't matter to add
2500 * flush request to tail or front of the dispatch queue.
2502 * Secondly in case of NCQ, flush request belongs to non-NCQ
2503 * command, and queueing it will fail when there is any
2504 * in-flight normal IO request(NCQ command). When adding flush
2505 * rq to the front of hctx->dispatch, it is easier to introduce
2506 * extra time to flush rq's latency because of S_SCHED_RESTART
2507 * compared with adding to the tail of dispatch queue, then
2508 * chance of flush merge is increased, and less flush requests
2509 * will be issued to controller. It is observed that ~10% time
2510 * is saved in blktests block/004 on disk attached to AHCI/NCQ
2511 * drive when adding flush rq to the front of hctx->dispatch.
2513 * Simply queue flush rq to the front of hctx->dispatch so that
2514 * intensive flush workloads can benefit in case of NCQ HW.
2516 blk_mq_request_bypass_insert(rq
, BLK_MQ_INSERT_AT_HEAD
);
2517 } else if (q
->elevator
) {
2520 WARN_ON_ONCE(rq
->tag
!= BLK_MQ_NO_TAG
);
2522 list_add(&rq
->queuelist
, &list
);
2523 q
->elevator
->type
->ops
.insert_requests(hctx
, &list
, flags
);
2525 trace_block_rq_insert(rq
);
2527 spin_lock(&ctx
->lock
);
2528 if (flags
& BLK_MQ_INSERT_AT_HEAD
)
2529 list_add(&rq
->queuelist
, &ctx
->rq_lists
[hctx
->type
]);
2531 list_add_tail(&rq
->queuelist
,
2532 &ctx
->rq_lists
[hctx
->type
]);
2533 blk_mq_hctx_mark_pending(hctx
, ctx
);
2534 spin_unlock(&ctx
->lock
);
2538 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
,
2539 unsigned int nr_segs
)
2543 if (bio
->bi_opf
& REQ_RAHEAD
)
2544 rq
->cmd_flags
|= REQ_FAILFAST_MASK
;
2546 rq
->__sector
= bio
->bi_iter
.bi_sector
;
2547 rq
->write_hint
= bio
->bi_write_hint
;
2548 blk_rq_bio_prep(rq
, bio
, nr_segs
);
2550 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2551 err
= blk_crypto_rq_bio_prep(rq
, bio
, GFP_NOIO
);
2554 blk_account_io_start(rq
);
2557 static blk_status_t
__blk_mq_issue_directly(struct blk_mq_hw_ctx
*hctx
,
2558 struct request
*rq
, bool last
)
2560 struct request_queue
*q
= rq
->q
;
2561 struct blk_mq_queue_data bd
= {
2568 * For OK queue, we are done. For error, caller may kill it.
2569 * Any other error (busy), just add it to our list as we
2570 * previously would have done.
2572 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
2575 blk_mq_update_dispatch_busy(hctx
, false);
2577 case BLK_STS_RESOURCE
:
2578 case BLK_STS_DEV_RESOURCE
:
2579 blk_mq_update_dispatch_busy(hctx
, true);
2580 __blk_mq_requeue_request(rq
);
2583 blk_mq_update_dispatch_busy(hctx
, false);
2590 static bool blk_mq_get_budget_and_tag(struct request
*rq
)
2594 budget_token
= blk_mq_get_dispatch_budget(rq
->q
);
2595 if (budget_token
< 0)
2597 blk_mq_set_rq_budget_token(rq
, budget_token
);
2598 if (!blk_mq_get_driver_tag(rq
)) {
2599 blk_mq_put_dispatch_budget(rq
->q
, budget_token
);
2606 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2607 * @hctx: Pointer of the associated hardware queue.
2608 * @rq: Pointer to request to be sent.
2610 * If the device has enough resources to accept a new request now, send the
2611 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2612 * we can try send it another time in the future. Requests inserted at this
2613 * queue have higher priority.
2615 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
2620 if (blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(rq
->q
)) {
2621 blk_mq_insert_request(rq
, 0);
2625 if ((rq
->rq_flags
& RQF_USE_SCHED
) || !blk_mq_get_budget_and_tag(rq
)) {
2626 blk_mq_insert_request(rq
, 0);
2627 blk_mq_run_hw_queue(hctx
, rq
->cmd_flags
& REQ_NOWAIT
);
2631 ret
= __blk_mq_issue_directly(hctx
, rq
, true);
2635 case BLK_STS_RESOURCE
:
2636 case BLK_STS_DEV_RESOURCE
:
2637 blk_mq_request_bypass_insert(rq
, 0);
2638 blk_mq_run_hw_queue(hctx
, false);
2641 blk_mq_end_request(rq
, ret
);
2646 static blk_status_t
blk_mq_request_issue_directly(struct request
*rq
, bool last
)
2648 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
2650 if (blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(rq
->q
)) {
2651 blk_mq_insert_request(rq
, 0);
2655 if (!blk_mq_get_budget_and_tag(rq
))
2656 return BLK_STS_RESOURCE
;
2657 return __blk_mq_issue_directly(hctx
, rq
, last
);
2660 static void blk_mq_plug_issue_direct(struct blk_plug
*plug
)
2662 struct blk_mq_hw_ctx
*hctx
= NULL
;
2665 blk_status_t ret
= BLK_STS_OK
;
2667 while ((rq
= rq_list_pop(&plug
->mq_list
))) {
2668 bool last
= rq_list_empty(plug
->mq_list
);
2670 if (hctx
!= rq
->mq_hctx
) {
2672 blk_mq_commit_rqs(hctx
, queued
, false);
2678 ret
= blk_mq_request_issue_directly(rq
, last
);
2683 case BLK_STS_RESOURCE
:
2684 case BLK_STS_DEV_RESOURCE
:
2685 blk_mq_request_bypass_insert(rq
, 0);
2686 blk_mq_run_hw_queue(hctx
, false);
2689 blk_mq_end_request(rq
, ret
);
2695 if (ret
!= BLK_STS_OK
)
2696 blk_mq_commit_rqs(hctx
, queued
, false);
2699 static void __blk_mq_flush_plug_list(struct request_queue
*q
,
2700 struct blk_plug
*plug
)
2702 if (blk_queue_quiesced(q
))
2704 q
->mq_ops
->queue_rqs(&plug
->mq_list
);
2707 static void blk_mq_dispatch_plug_list(struct blk_plug
*plug
, bool from_sched
)
2709 struct blk_mq_hw_ctx
*this_hctx
= NULL
;
2710 struct blk_mq_ctx
*this_ctx
= NULL
;
2711 struct request
*requeue_list
= NULL
;
2712 struct request
**requeue_lastp
= &requeue_list
;
2713 unsigned int depth
= 0;
2714 bool is_passthrough
= false;
2718 struct request
*rq
= rq_list_pop(&plug
->mq_list
);
2721 this_hctx
= rq
->mq_hctx
;
2722 this_ctx
= rq
->mq_ctx
;
2723 is_passthrough
= blk_rq_is_passthrough(rq
);
2724 } else if (this_hctx
!= rq
->mq_hctx
|| this_ctx
!= rq
->mq_ctx
||
2725 is_passthrough
!= blk_rq_is_passthrough(rq
)) {
2726 rq_list_add_tail(&requeue_lastp
, rq
);
2729 list_add(&rq
->queuelist
, &list
);
2731 } while (!rq_list_empty(plug
->mq_list
));
2733 plug
->mq_list
= requeue_list
;
2734 trace_block_unplug(this_hctx
->queue
, depth
, !from_sched
);
2736 percpu_ref_get(&this_hctx
->queue
->q_usage_counter
);
2737 /* passthrough requests should never be issued to the I/O scheduler */
2738 if (is_passthrough
) {
2739 spin_lock(&this_hctx
->lock
);
2740 list_splice_tail_init(&list
, &this_hctx
->dispatch
);
2741 spin_unlock(&this_hctx
->lock
);
2742 blk_mq_run_hw_queue(this_hctx
, from_sched
);
2743 } else if (this_hctx
->queue
->elevator
) {
2744 this_hctx
->queue
->elevator
->type
->ops
.insert_requests(this_hctx
,
2746 blk_mq_run_hw_queue(this_hctx
, from_sched
);
2748 blk_mq_insert_requests(this_hctx
, this_ctx
, &list
, from_sched
);
2750 percpu_ref_put(&this_hctx
->queue
->q_usage_counter
);
2753 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
2758 * We may have been called recursively midway through handling
2759 * plug->mq_list via a schedule() in the driver's queue_rq() callback.
2760 * To avoid mq_list changing under our feet, clear rq_count early and
2761 * bail out specifically if rq_count is 0 rather than checking
2762 * whether the mq_list is empty.
2764 if (plug
->rq_count
== 0)
2768 if (!plug
->multiple_queues
&& !plug
->has_elevator
&& !from_schedule
) {
2769 struct request_queue
*q
;
2771 rq
= rq_list_peek(&plug
->mq_list
);
2775 * Peek first request and see if we have a ->queue_rqs() hook.
2776 * If we do, we can dispatch the whole plug list in one go. We
2777 * already know at this point that all requests belong to the
2778 * same queue, caller must ensure that's the case.
2780 if (q
->mq_ops
->queue_rqs
) {
2781 blk_mq_run_dispatch_ops(q
,
2782 __blk_mq_flush_plug_list(q
, plug
));
2783 if (rq_list_empty(plug
->mq_list
))
2787 blk_mq_run_dispatch_ops(q
,
2788 blk_mq_plug_issue_direct(plug
));
2789 if (rq_list_empty(plug
->mq_list
))
2794 blk_mq_dispatch_plug_list(plug
, from_schedule
);
2795 } while (!rq_list_empty(plug
->mq_list
));
2798 static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx
*hctx
,
2799 struct list_head
*list
)
2802 blk_status_t ret
= BLK_STS_OK
;
2804 while (!list_empty(list
)) {
2805 struct request
*rq
= list_first_entry(list
, struct request
,
2808 list_del_init(&rq
->queuelist
);
2809 ret
= blk_mq_request_issue_directly(rq
, list_empty(list
));
2814 case BLK_STS_RESOURCE
:
2815 case BLK_STS_DEV_RESOURCE
:
2816 blk_mq_request_bypass_insert(rq
, 0);
2817 if (list_empty(list
))
2818 blk_mq_run_hw_queue(hctx
, false);
2821 blk_mq_end_request(rq
, ret
);
2827 if (ret
!= BLK_STS_OK
)
2828 blk_mq_commit_rqs(hctx
, queued
, false);
2831 static bool blk_mq_attempt_bio_merge(struct request_queue
*q
,
2832 struct bio
*bio
, unsigned int nr_segs
)
2834 if (!blk_queue_nomerges(q
) && bio_mergeable(bio
)) {
2835 if (blk_attempt_plug_merge(q
, bio
, nr_segs
))
2837 if (blk_mq_sched_bio_merge(q
, bio
, nr_segs
))
2843 static struct request
*blk_mq_get_new_requests(struct request_queue
*q
,
2844 struct blk_plug
*plug
,
2848 struct blk_mq_alloc_data data
= {
2851 .cmd_flags
= bio
->bi_opf
,
2855 rq_qos_throttle(q
, bio
);
2858 data
.nr_tags
= plug
->nr_ios
;
2860 data
.cached_rq
= &plug
->cached_rq
;
2863 rq
= __blk_mq_alloc_requests(&data
);
2866 rq_qos_cleanup(q
, bio
);
2867 if (bio
->bi_opf
& REQ_NOWAIT
)
2868 bio_wouldblock_error(bio
);
2873 * Check if there is a suitable cached request and return it.
2875 static struct request
*blk_mq_peek_cached_request(struct blk_plug
*plug
,
2876 struct request_queue
*q
, blk_opf_t opf
)
2878 enum hctx_type type
= blk_mq_get_hctx_type(opf
);
2883 rq
= rq_list_peek(&plug
->cached_rq
);
2884 if (!rq
|| rq
->q
!= q
)
2886 if (type
!= rq
->mq_hctx
->type
&&
2887 (type
!= HCTX_TYPE_READ
|| rq
->mq_hctx
->type
!= HCTX_TYPE_DEFAULT
))
2889 if (op_is_flush(rq
->cmd_flags
) != op_is_flush(opf
))
2894 static void blk_mq_use_cached_rq(struct request
*rq
, struct blk_plug
*plug
,
2897 WARN_ON_ONCE(rq_list_peek(&plug
->cached_rq
) != rq
);
2900 * If any qos ->throttle() end up blocking, we will have flushed the
2901 * plug and hence killed the cached_rq list as well. Pop this entry
2902 * before we throttle.
2904 plug
->cached_rq
= rq_list_next(rq
);
2905 rq_qos_throttle(rq
->q
, bio
);
2907 blk_mq_rq_time_init(rq
, 0);
2908 rq
->cmd_flags
= bio
->bi_opf
;
2909 INIT_LIST_HEAD(&rq
->queuelist
);
2912 static bool bio_unaligned(const struct bio
*bio
, struct request_queue
*q
)
2914 unsigned int bs_mask
= queue_logical_block_size(q
) - 1;
2916 /* .bi_sector of any zero sized bio need to be initialized */
2917 if ((bio
->bi_iter
.bi_size
& bs_mask
) ||
2918 ((bio
->bi_iter
.bi_sector
<< SECTOR_SHIFT
) & bs_mask
))
2924 * blk_mq_submit_bio - Create and send a request to block device.
2925 * @bio: Bio pointer.
2927 * Builds up a request structure from @q and @bio and send to the device. The
2928 * request may not be queued directly to hardware if:
2929 * * This request can be merged with another one
2930 * * We want to place request at plug queue for possible future merging
2931 * * There is an IO scheduler active at this queue
2933 * It will not queue the request if there is an error with the bio, or at the
2936 void blk_mq_submit_bio(struct bio
*bio
)
2938 struct request_queue
*q
= bdev_get_queue(bio
->bi_bdev
);
2939 struct blk_plug
*plug
= current
->plug
;
2940 const int is_sync
= op_is_sync(bio
->bi_opf
);
2941 struct blk_mq_hw_ctx
*hctx
;
2942 unsigned int nr_segs
= 1;
2947 * If the plug has a cached request for this queue, try to use it.
2949 rq
= blk_mq_peek_cached_request(plug
, q
, bio
->bi_opf
);
2952 * A BIO that was released from a zone write plug has already been
2953 * through the preparation in this function, already holds a reference
2954 * on the queue usage counter, and is the only write BIO in-flight for
2955 * the target zone. Go straight to preparing a request for it.
2957 if (bio_zone_write_plugging(bio
)) {
2958 nr_segs
= bio
->__bi_nr_segments
;
2964 bio
= blk_queue_bounce(bio
, q
);
2967 * The cached request already holds a q_usage_counter reference and we
2968 * don't have to acquire a new one if we use it.
2971 if (unlikely(bio_queue_enter(bio
)))
2976 * Device reconfiguration may change logical block size, so alignment
2977 * check has to be done with queue usage counter held
2979 if (unlikely(bio_unaligned(bio
, q
))) {
2984 if (unlikely(bio_may_exceed_limits(bio
, &q
->limits
))) {
2985 bio
= __bio_split_to_limits(bio
, &q
->limits
, &nr_segs
);
2989 if (!bio_integrity_prep(bio
))
2992 if (blk_mq_attempt_bio_merge(q
, bio
, nr_segs
))
2995 if (blk_queue_is_zoned(q
) && blk_zone_plug_bio(bio
, nr_segs
))
3000 rq
= blk_mq_get_new_requests(q
, plug
, bio
, nr_segs
);
3004 blk_mq_use_cached_rq(rq
, plug
, bio
);
3007 trace_block_getrq(bio
);
3009 rq_qos_track(q
, rq
, bio
);
3011 blk_mq_bio_to_request(rq
, bio
, nr_segs
);
3013 ret
= blk_crypto_rq_get_keyslot(rq
);
3014 if (ret
!= BLK_STS_OK
) {
3015 bio
->bi_status
= ret
;
3017 blk_mq_free_request(rq
);
3021 if (bio_zone_write_plugging(bio
))
3022 blk_zone_write_plug_init_request(rq
);
3024 if (op_is_flush(bio
->bi_opf
) && blk_insert_flush(rq
))
3028 blk_add_rq_to_plug(plug
, rq
);
3033 if ((rq
->rq_flags
& RQF_USE_SCHED
) ||
3034 (hctx
->dispatch_busy
&& (q
->nr_hw_queues
== 1 || !is_sync
))) {
3035 blk_mq_insert_request(rq
, 0);
3036 blk_mq_run_hw_queue(hctx
, true);
3038 blk_mq_run_dispatch_ops(q
, blk_mq_try_issue_directly(hctx
, rq
));
3044 * Don't drop the queue reference if we were trying to use a cached
3045 * request and thus didn't acquire one.
3051 #ifdef CONFIG_BLK_MQ_STACKING
3053 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
3054 * @rq: the request being queued
3056 blk_status_t
blk_insert_cloned_request(struct request
*rq
)
3058 struct request_queue
*q
= rq
->q
;
3059 unsigned int max_sectors
= blk_queue_get_max_sectors(rq
);
3060 unsigned int max_segments
= blk_rq_get_max_segments(rq
);
3063 if (blk_rq_sectors(rq
) > max_sectors
) {
3065 * SCSI device does not have a good way to return if
3066 * Write Same/Zero is actually supported. If a device rejects
3067 * a non-read/write command (discard, write same,etc.) the
3068 * low-level device driver will set the relevant queue limit to
3069 * 0 to prevent blk-lib from issuing more of the offending
3070 * operations. Commands queued prior to the queue limit being
3071 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
3072 * errors being propagated to upper layers.
3074 if (max_sectors
== 0)
3075 return BLK_STS_NOTSUPP
;
3077 printk(KERN_ERR
"%s: over max size limit. (%u > %u)\n",
3078 __func__
, blk_rq_sectors(rq
), max_sectors
);
3079 return BLK_STS_IOERR
;
3083 * The queue settings related to segment counting may differ from the
3086 rq
->nr_phys_segments
= blk_recalc_rq_segments(rq
);
3087 if (rq
->nr_phys_segments
> max_segments
) {
3088 printk(KERN_ERR
"%s: over max segments limit. (%u > %u)\n",
3089 __func__
, rq
->nr_phys_segments
, max_segments
);
3090 return BLK_STS_IOERR
;
3093 if (q
->disk
&& should_fail_request(q
->disk
->part0
, blk_rq_bytes(rq
)))
3094 return BLK_STS_IOERR
;
3096 ret
= blk_crypto_rq_get_keyslot(rq
);
3097 if (ret
!= BLK_STS_OK
)
3100 blk_account_io_start(rq
);
3103 * Since we have a scheduler attached on the top device,
3104 * bypass a potential scheduler on the bottom device for
3107 blk_mq_run_dispatch_ops(q
,
3108 ret
= blk_mq_request_issue_directly(rq
, true));
3110 blk_account_io_done(rq
, blk_time_get_ns());
3113 EXPORT_SYMBOL_GPL(blk_insert_cloned_request
);
3116 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3117 * @rq: the clone request to be cleaned up
3120 * Free all bios in @rq for a cloned request.
3122 void blk_rq_unprep_clone(struct request
*rq
)
3126 while ((bio
= rq
->bio
) != NULL
) {
3127 rq
->bio
= bio
->bi_next
;
3132 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone
);
3135 * blk_rq_prep_clone - Helper function to setup clone request
3136 * @rq: the request to be setup
3137 * @rq_src: original request to be cloned
3138 * @bs: bio_set that bios for clone are allocated from
3139 * @gfp_mask: memory allocation mask for bio
3140 * @bio_ctr: setup function to be called for each clone bio.
3141 * Returns %0 for success, non %0 for failure.
3142 * @data: private data to be passed to @bio_ctr
3145 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3146 * Also, pages which the original bios are pointing to are not copied
3147 * and the cloned bios just point same pages.
3148 * So cloned bios must be completed before original bios, which means
3149 * the caller must complete @rq before @rq_src.
3151 int blk_rq_prep_clone(struct request
*rq
, struct request
*rq_src
,
3152 struct bio_set
*bs
, gfp_t gfp_mask
,
3153 int (*bio_ctr
)(struct bio
*, struct bio
*, void *),
3156 struct bio
*bio
, *bio_src
;
3161 __rq_for_each_bio(bio_src
, rq_src
) {
3162 bio
= bio_alloc_clone(rq
->q
->disk
->part0
, bio_src
, gfp_mask
,
3167 if (bio_ctr
&& bio_ctr(bio
, bio_src
, data
))
3171 rq
->biotail
->bi_next
= bio
;
3174 rq
->bio
= rq
->biotail
= bio
;
3179 /* Copy attributes of the original request to the clone request. */
3180 rq
->__sector
= blk_rq_pos(rq_src
);
3181 rq
->__data_len
= blk_rq_bytes(rq_src
);
3182 if (rq_src
->rq_flags
& RQF_SPECIAL_PAYLOAD
) {
3183 rq
->rq_flags
|= RQF_SPECIAL_PAYLOAD
;
3184 rq
->special_vec
= rq_src
->special_vec
;
3186 rq
->nr_phys_segments
= rq_src
->nr_phys_segments
;
3187 rq
->ioprio
= rq_src
->ioprio
;
3188 rq
->write_hint
= rq_src
->write_hint
;
3190 if (rq
->bio
&& blk_crypto_rq_bio_prep(rq
, rq
->bio
, gfp_mask
) < 0)
3198 blk_rq_unprep_clone(rq
);
3202 EXPORT_SYMBOL_GPL(blk_rq_prep_clone
);
3203 #endif /* CONFIG_BLK_MQ_STACKING */
3206 * Steal bios from a request and add them to a bio list.
3207 * The request must not have been partially completed before.
3209 void blk_steal_bios(struct bio_list
*list
, struct request
*rq
)
3213 list
->tail
->bi_next
= rq
->bio
;
3215 list
->head
= rq
->bio
;
3216 list
->tail
= rq
->biotail
;
3224 EXPORT_SYMBOL_GPL(blk_steal_bios
);
3226 static size_t order_to_size(unsigned int order
)
3228 return (size_t)PAGE_SIZE
<< order
;
3231 /* called before freeing request pool in @tags */
3232 static void blk_mq_clear_rq_mapping(struct blk_mq_tags
*drv_tags
,
3233 struct blk_mq_tags
*tags
)
3236 unsigned long flags
;
3239 * There is no need to clear mapping if driver tags is not initialized
3240 * or the mapping belongs to the driver tags.
3242 if (!drv_tags
|| drv_tags
== tags
)
3245 list_for_each_entry(page
, &tags
->page_list
, lru
) {
3246 unsigned long start
= (unsigned long)page_address(page
);
3247 unsigned long end
= start
+ order_to_size(page
->private);
3250 for (i
= 0; i
< drv_tags
->nr_tags
; i
++) {
3251 struct request
*rq
= drv_tags
->rqs
[i
];
3252 unsigned long rq_addr
= (unsigned long)rq
;
3254 if (rq_addr
>= start
&& rq_addr
< end
) {
3255 WARN_ON_ONCE(req_ref_read(rq
) != 0);
3256 cmpxchg(&drv_tags
->rqs
[i
], rq
, NULL
);
3262 * Wait until all pending iteration is done.
3264 * Request reference is cleared and it is guaranteed to be observed
3265 * after the ->lock is released.
3267 spin_lock_irqsave(&drv_tags
->lock
, flags
);
3268 spin_unlock_irqrestore(&drv_tags
->lock
, flags
);
3271 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
3272 unsigned int hctx_idx
)
3274 struct blk_mq_tags
*drv_tags
;
3277 if (list_empty(&tags
->page_list
))
3280 if (blk_mq_is_shared_tags(set
->flags
))
3281 drv_tags
= set
->shared_tags
;
3283 drv_tags
= set
->tags
[hctx_idx
];
3285 if (tags
->static_rqs
&& set
->ops
->exit_request
) {
3288 for (i
= 0; i
< tags
->nr_tags
; i
++) {
3289 struct request
*rq
= tags
->static_rqs
[i
];
3293 set
->ops
->exit_request(set
, rq
, hctx_idx
);
3294 tags
->static_rqs
[i
] = NULL
;
3298 blk_mq_clear_rq_mapping(drv_tags
, tags
);
3300 while (!list_empty(&tags
->page_list
)) {
3301 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
3302 list_del_init(&page
->lru
);
3304 * Remove kmemleak object previously allocated in
3305 * blk_mq_alloc_rqs().
3307 kmemleak_free(page_address(page
));
3308 __free_pages(page
, page
->private);
3312 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
3316 kfree(tags
->static_rqs
);
3317 tags
->static_rqs
= NULL
;
3319 blk_mq_free_tags(tags
);
3322 static enum hctx_type
hctx_idx_to_type(struct blk_mq_tag_set
*set
,
3323 unsigned int hctx_idx
)
3327 for (i
= 0; i
< set
->nr_maps
; i
++) {
3328 unsigned int start
= set
->map
[i
].queue_offset
;
3329 unsigned int end
= start
+ set
->map
[i
].nr_queues
;
3331 if (hctx_idx
>= start
&& hctx_idx
< end
)
3335 if (i
>= set
->nr_maps
)
3336 i
= HCTX_TYPE_DEFAULT
;
3341 static int blk_mq_get_hctx_node(struct blk_mq_tag_set
*set
,
3342 unsigned int hctx_idx
)
3344 enum hctx_type type
= hctx_idx_to_type(set
, hctx_idx
);
3346 return blk_mq_hw_queue_to_node(&set
->map
[type
], hctx_idx
);
3349 static struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
3350 unsigned int hctx_idx
,
3351 unsigned int nr_tags
,
3352 unsigned int reserved_tags
)
3354 int node
= blk_mq_get_hctx_node(set
, hctx_idx
);
3355 struct blk_mq_tags
*tags
;
3357 if (node
== NUMA_NO_NODE
)
3358 node
= set
->numa_node
;
3360 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
,
3361 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
3365 tags
->rqs
= kcalloc_node(nr_tags
, sizeof(struct request
*),
3366 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
3371 tags
->static_rqs
= kcalloc_node(nr_tags
, sizeof(struct request
*),
3372 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
3374 if (!tags
->static_rqs
)
3382 blk_mq_free_tags(tags
);
3386 static int blk_mq_init_request(struct blk_mq_tag_set
*set
, struct request
*rq
,
3387 unsigned int hctx_idx
, int node
)
3391 if (set
->ops
->init_request
) {
3392 ret
= set
->ops
->init_request(set
, rq
, hctx_idx
, node
);
3397 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
3401 static int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
,
3402 struct blk_mq_tags
*tags
,
3403 unsigned int hctx_idx
, unsigned int depth
)
3405 unsigned int i
, j
, entries_per_page
, max_order
= 4;
3406 int node
= blk_mq_get_hctx_node(set
, hctx_idx
);
3407 size_t rq_size
, left
;
3409 if (node
== NUMA_NO_NODE
)
3410 node
= set
->numa_node
;
3412 INIT_LIST_HEAD(&tags
->page_list
);
3415 * rq_size is the size of the request plus driver payload, rounded
3416 * to the cacheline size
3418 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
3420 left
= rq_size
* depth
;
3422 for (i
= 0; i
< depth
; ) {
3423 int this_order
= max_order
;
3428 while (this_order
&& left
< order_to_size(this_order
- 1))
3432 page
= alloc_pages_node(node
,
3433 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
3439 if (order_to_size(this_order
) < rq_size
)
3446 page
->private = this_order
;
3447 list_add_tail(&page
->lru
, &tags
->page_list
);
3449 p
= page_address(page
);
3451 * Allow kmemleak to scan these pages as they contain pointers
3452 * to additional allocations like via ops->init_request().
3454 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
3455 entries_per_page
= order_to_size(this_order
) / rq_size
;
3456 to_do
= min(entries_per_page
, depth
- i
);
3457 left
-= to_do
* rq_size
;
3458 for (j
= 0; j
< to_do
; j
++) {
3459 struct request
*rq
= p
;
3461 tags
->static_rqs
[i
] = rq
;
3462 if (blk_mq_init_request(set
, rq
, hctx_idx
, node
)) {
3463 tags
->static_rqs
[i
] = NULL
;
3474 blk_mq_free_rqs(set
, tags
, hctx_idx
);
3478 struct rq_iter_data
{
3479 struct blk_mq_hw_ctx
*hctx
;
3483 static bool blk_mq_has_request(struct request
*rq
, void *data
)
3485 struct rq_iter_data
*iter_data
= data
;
3487 if (rq
->mq_hctx
!= iter_data
->hctx
)
3489 iter_data
->has_rq
= true;
3493 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx
*hctx
)
3495 struct blk_mq_tags
*tags
= hctx
->sched_tags
?
3496 hctx
->sched_tags
: hctx
->tags
;
3497 struct rq_iter_data data
= {
3501 blk_mq_all_tag_iter(tags
, blk_mq_has_request
, &data
);
3505 static bool blk_mq_hctx_has_online_cpu(struct blk_mq_hw_ctx
*hctx
,
3506 unsigned int this_cpu
)
3508 enum hctx_type type
= hctx
->type
;
3512 * hctx->cpumask has to rule out isolated CPUs, but userspace still
3513 * might submit IOs on these isolated CPUs, so use the queue map to
3514 * check if all CPUs mapped to this hctx are offline
3516 for_each_online_cpu(cpu
) {
3517 struct blk_mq_hw_ctx
*h
= blk_mq_map_queue_type(hctx
->queue
,
3523 /* this hctx has at least one online CPU */
3524 if (this_cpu
!= cpu
)
3531 static int blk_mq_hctx_notify_offline(unsigned int cpu
, struct hlist_node
*node
)
3533 struct blk_mq_hw_ctx
*hctx
= hlist_entry_safe(node
,
3534 struct blk_mq_hw_ctx
, cpuhp_online
);
3536 if (blk_mq_hctx_has_online_cpu(hctx
, cpu
))
3540 * Prevent new request from being allocated on the current hctx.
3542 * The smp_mb__after_atomic() Pairs with the implied barrier in
3543 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3544 * seen once we return from the tag allocator.
3546 set_bit(BLK_MQ_S_INACTIVE
, &hctx
->state
);
3547 smp_mb__after_atomic();
3550 * Try to grab a reference to the queue and wait for any outstanding
3551 * requests. If we could not grab a reference the queue has been
3552 * frozen and there are no requests.
3554 if (percpu_ref_tryget(&hctx
->queue
->q_usage_counter
)) {
3555 while (blk_mq_hctx_has_requests(hctx
))
3557 percpu_ref_put(&hctx
->queue
->q_usage_counter
);
3564 * Check if one CPU is mapped to the specified hctx
3566 * Isolated CPUs have been ruled out from hctx->cpumask, which is supposed
3567 * to be used for scheduling kworker only. For other usage, please call this
3568 * helper for checking if one CPU belongs to the specified hctx
3570 static bool blk_mq_cpu_mapped_to_hctx(unsigned int cpu
,
3571 const struct blk_mq_hw_ctx
*hctx
)
3573 struct blk_mq_hw_ctx
*mapped_hctx
= blk_mq_map_queue_type(hctx
->queue
,
3576 return mapped_hctx
== hctx
;
3579 static int blk_mq_hctx_notify_online(unsigned int cpu
, struct hlist_node
*node
)
3581 struct blk_mq_hw_ctx
*hctx
= hlist_entry_safe(node
,
3582 struct blk_mq_hw_ctx
, cpuhp_online
);
3584 if (blk_mq_cpu_mapped_to_hctx(cpu
, hctx
))
3585 clear_bit(BLK_MQ_S_INACTIVE
, &hctx
->state
);
3590 * 'cpu' is going away. splice any existing rq_list entries from this
3591 * software queue to the hw queue dispatch list, and ensure that it
3594 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
3596 struct blk_mq_hw_ctx
*hctx
;
3597 struct blk_mq_ctx
*ctx
;
3599 enum hctx_type type
;
3601 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
3602 if (!blk_mq_cpu_mapped_to_hctx(cpu
, hctx
))
3605 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
3608 spin_lock(&ctx
->lock
);
3609 if (!list_empty(&ctx
->rq_lists
[type
])) {
3610 list_splice_init(&ctx
->rq_lists
[type
], &tmp
);
3611 blk_mq_hctx_clear_pending(hctx
, ctx
);
3613 spin_unlock(&ctx
->lock
);
3615 if (list_empty(&tmp
))
3618 spin_lock(&hctx
->lock
);
3619 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
3620 spin_unlock(&hctx
->lock
);
3622 blk_mq_run_hw_queue(hctx
, true);
3626 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
3628 if (!(hctx
->flags
& BLK_MQ_F_STACKING
))
3629 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE
,
3630 &hctx
->cpuhp_online
);
3631 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
3636 * Before freeing hw queue, clearing the flush request reference in
3637 * tags->rqs[] for avoiding potential UAF.
3639 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags
*tags
,
3640 unsigned int queue_depth
, struct request
*flush_rq
)
3643 unsigned long flags
;
3645 /* The hw queue may not be mapped yet */
3649 WARN_ON_ONCE(req_ref_read(flush_rq
) != 0);
3651 for (i
= 0; i
< queue_depth
; i
++)
3652 cmpxchg(&tags
->rqs
[i
], flush_rq
, NULL
);
3655 * Wait until all pending iteration is done.
3657 * Request reference is cleared and it is guaranteed to be observed
3658 * after the ->lock is released.
3660 spin_lock_irqsave(&tags
->lock
, flags
);
3661 spin_unlock_irqrestore(&tags
->lock
, flags
);
3664 /* hctx->ctxs will be freed in queue's release handler */
3665 static void blk_mq_exit_hctx(struct request_queue
*q
,
3666 struct blk_mq_tag_set
*set
,
3667 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
3669 struct request
*flush_rq
= hctx
->fq
->flush_rq
;
3671 if (blk_mq_hw_queue_mapped(hctx
))
3672 blk_mq_tag_idle(hctx
);
3674 if (blk_queue_init_done(q
))
3675 blk_mq_clear_flush_rq_mapping(set
->tags
[hctx_idx
],
3676 set
->queue_depth
, flush_rq
);
3677 if (set
->ops
->exit_request
)
3678 set
->ops
->exit_request(set
, flush_rq
, hctx_idx
);
3680 if (set
->ops
->exit_hctx
)
3681 set
->ops
->exit_hctx(hctx
, hctx_idx
);
3683 blk_mq_remove_cpuhp(hctx
);
3685 xa_erase(&q
->hctx_table
, hctx_idx
);
3687 spin_lock(&q
->unused_hctx_lock
);
3688 list_add(&hctx
->hctx_list
, &q
->unused_hctx_list
);
3689 spin_unlock(&q
->unused_hctx_lock
);
3692 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
3693 struct blk_mq_tag_set
*set
, int nr_queue
)
3695 struct blk_mq_hw_ctx
*hctx
;
3698 queue_for_each_hw_ctx(q
, hctx
, i
) {
3701 blk_mq_exit_hctx(q
, set
, hctx
, i
);
3705 static int blk_mq_init_hctx(struct request_queue
*q
,
3706 struct blk_mq_tag_set
*set
,
3707 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
3709 hctx
->queue_num
= hctx_idx
;
3711 if (!(hctx
->flags
& BLK_MQ_F_STACKING
))
3712 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE
,
3713 &hctx
->cpuhp_online
);
3714 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
3716 hctx
->tags
= set
->tags
[hctx_idx
];
3718 if (set
->ops
->init_hctx
&&
3719 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
3720 goto unregister_cpu_notifier
;
3722 if (blk_mq_init_request(set
, hctx
->fq
->flush_rq
, hctx_idx
,
3726 if (xa_insert(&q
->hctx_table
, hctx_idx
, hctx
, GFP_KERNEL
))
3732 if (set
->ops
->exit_request
)
3733 set
->ops
->exit_request(set
, hctx
->fq
->flush_rq
, hctx_idx
);
3735 if (set
->ops
->exit_hctx
)
3736 set
->ops
->exit_hctx(hctx
, hctx_idx
);
3737 unregister_cpu_notifier
:
3738 blk_mq_remove_cpuhp(hctx
);
3742 static struct blk_mq_hw_ctx
*
3743 blk_mq_alloc_hctx(struct request_queue
*q
, struct blk_mq_tag_set
*set
,
3746 struct blk_mq_hw_ctx
*hctx
;
3747 gfp_t gfp
= GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
;
3749 hctx
= kzalloc_node(sizeof(struct blk_mq_hw_ctx
), gfp
, node
);
3751 goto fail_alloc_hctx
;
3753 if (!zalloc_cpumask_var_node(&hctx
->cpumask
, gfp
, node
))
3756 atomic_set(&hctx
->nr_active
, 0);
3757 if (node
== NUMA_NO_NODE
)
3758 node
= set
->numa_node
;
3759 hctx
->numa_node
= node
;
3761 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
3762 spin_lock_init(&hctx
->lock
);
3763 INIT_LIST_HEAD(&hctx
->dispatch
);
3765 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_QUEUE_SHARED
;
3767 INIT_LIST_HEAD(&hctx
->hctx_list
);
3770 * Allocate space for all possible cpus to avoid allocation at
3773 hctx
->ctxs
= kmalloc_array_node(nr_cpu_ids
, sizeof(void *),
3778 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8),
3779 gfp
, node
, false, false))
3783 spin_lock_init(&hctx
->dispatch_wait_lock
);
3784 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
3785 INIT_LIST_HEAD(&hctx
->dispatch_wait
.entry
);
3787 hctx
->fq
= blk_alloc_flush_queue(hctx
->numa_node
, set
->cmd_size
, gfp
);
3791 blk_mq_hctx_kobj_init(hctx
);
3796 sbitmap_free(&hctx
->ctx_map
);
3800 free_cpumask_var(hctx
->cpumask
);
3807 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
3808 unsigned int nr_hw_queues
)
3810 struct blk_mq_tag_set
*set
= q
->tag_set
;
3813 for_each_possible_cpu(i
) {
3814 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
3815 struct blk_mq_hw_ctx
*hctx
;
3819 spin_lock_init(&__ctx
->lock
);
3820 for (k
= HCTX_TYPE_DEFAULT
; k
< HCTX_MAX_TYPES
; k
++)
3821 INIT_LIST_HEAD(&__ctx
->rq_lists
[k
]);
3826 * Set local node, IFF we have more than one hw queue. If
3827 * not, we remain on the home node of the device
3829 for (j
= 0; j
< set
->nr_maps
; j
++) {
3830 hctx
= blk_mq_map_queue_type(q
, j
, i
);
3831 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
3832 hctx
->numa_node
= cpu_to_node(i
);
3837 struct blk_mq_tags
*blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set
*set
,
3838 unsigned int hctx_idx
,
3841 struct blk_mq_tags
*tags
;
3844 tags
= blk_mq_alloc_rq_map(set
, hctx_idx
, depth
, set
->reserved_tags
);
3848 ret
= blk_mq_alloc_rqs(set
, tags
, hctx_idx
, depth
);
3850 blk_mq_free_rq_map(tags
);
3857 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set
*set
,
3860 if (blk_mq_is_shared_tags(set
->flags
)) {
3861 set
->tags
[hctx_idx
] = set
->shared_tags
;
3866 set
->tags
[hctx_idx
] = blk_mq_alloc_map_and_rqs(set
, hctx_idx
,
3869 return set
->tags
[hctx_idx
];
3872 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set
*set
,
3873 struct blk_mq_tags
*tags
,
3874 unsigned int hctx_idx
)
3877 blk_mq_free_rqs(set
, tags
, hctx_idx
);
3878 blk_mq_free_rq_map(tags
);
3882 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set
*set
,
3883 unsigned int hctx_idx
)
3885 if (!blk_mq_is_shared_tags(set
->flags
))
3886 blk_mq_free_map_and_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
3888 set
->tags
[hctx_idx
] = NULL
;
3891 static void blk_mq_map_swqueue(struct request_queue
*q
)
3893 unsigned int j
, hctx_idx
;
3895 struct blk_mq_hw_ctx
*hctx
;
3896 struct blk_mq_ctx
*ctx
;
3897 struct blk_mq_tag_set
*set
= q
->tag_set
;
3899 queue_for_each_hw_ctx(q
, hctx
, i
) {
3900 cpumask_clear(hctx
->cpumask
);
3902 hctx
->dispatch_from
= NULL
;
3906 * Map software to hardware queues.
3908 * If the cpu isn't present, the cpu is mapped to first hctx.
3910 for_each_possible_cpu(i
) {
3912 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
3913 for (j
= 0; j
< set
->nr_maps
; j
++) {
3914 if (!set
->map
[j
].nr_queues
) {
3915 ctx
->hctxs
[j
] = blk_mq_map_queue_type(q
,
3916 HCTX_TYPE_DEFAULT
, i
);
3919 hctx_idx
= set
->map
[j
].mq_map
[i
];
3920 /* unmapped hw queue can be remapped after CPU topo changed */
3921 if (!set
->tags
[hctx_idx
] &&
3922 !__blk_mq_alloc_map_and_rqs(set
, hctx_idx
)) {
3924 * If tags initialization fail for some hctx,
3925 * that hctx won't be brought online. In this
3926 * case, remap the current ctx to hctx[0] which
3927 * is guaranteed to always have tags allocated
3929 set
->map
[j
].mq_map
[i
] = 0;
3932 hctx
= blk_mq_map_queue_type(q
, j
, i
);
3933 ctx
->hctxs
[j
] = hctx
;
3935 * If the CPU is already set in the mask, then we've
3936 * mapped this one already. This can happen if
3937 * devices share queues across queue maps.
3939 if (cpumask_test_cpu(i
, hctx
->cpumask
))
3942 cpumask_set_cpu(i
, hctx
->cpumask
);
3944 ctx
->index_hw
[hctx
->type
] = hctx
->nr_ctx
;
3945 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
3948 * If the nr_ctx type overflows, we have exceeded the
3949 * amount of sw queues we can support.
3951 BUG_ON(!hctx
->nr_ctx
);
3954 for (; j
< HCTX_MAX_TYPES
; j
++)
3955 ctx
->hctxs
[j
] = blk_mq_map_queue_type(q
,
3956 HCTX_TYPE_DEFAULT
, i
);
3959 queue_for_each_hw_ctx(q
, hctx
, i
) {
3963 * If no software queues are mapped to this hardware queue,
3964 * disable it and free the request entries.
3966 if (!hctx
->nr_ctx
) {
3967 /* Never unmap queue 0. We need it as a
3968 * fallback in case of a new remap fails
3972 __blk_mq_free_map_and_rqs(set
, i
);
3978 hctx
->tags
= set
->tags
[i
];
3979 WARN_ON(!hctx
->tags
);
3982 * Set the map size to the number of mapped software queues.
3983 * This is more accurate and more efficient than looping
3984 * over all possibly mapped software queues.
3986 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
3989 * Rule out isolated CPUs from hctx->cpumask to avoid
3990 * running block kworker on isolated CPUs
3992 for_each_cpu(cpu
, hctx
->cpumask
) {
3993 if (cpu_is_isolated(cpu
))
3994 cpumask_clear_cpu(cpu
, hctx
->cpumask
);
3998 * Initialize batch roundrobin counts
4000 hctx
->next_cpu
= blk_mq_first_mapped_cpu(hctx
);
4001 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
4006 * Caller needs to ensure that we're either frozen/quiesced, or that
4007 * the queue isn't live yet.
4009 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
4011 struct blk_mq_hw_ctx
*hctx
;
4014 queue_for_each_hw_ctx(q
, hctx
, i
) {
4016 hctx
->flags
|= BLK_MQ_F_TAG_QUEUE_SHARED
;
4018 blk_mq_tag_idle(hctx
);
4019 hctx
->flags
&= ~BLK_MQ_F_TAG_QUEUE_SHARED
;
4024 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set
*set
,
4027 struct request_queue
*q
;
4029 lockdep_assert_held(&set
->tag_list_lock
);
4031 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
4032 blk_mq_freeze_queue(q
);
4033 queue_set_hctx_shared(q
, shared
);
4034 blk_mq_unfreeze_queue(q
);
4038 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
4040 struct blk_mq_tag_set
*set
= q
->tag_set
;
4042 mutex_lock(&set
->tag_list_lock
);
4043 list_del(&q
->tag_set_list
);
4044 if (list_is_singular(&set
->tag_list
)) {
4045 /* just transitioned to unshared */
4046 set
->flags
&= ~BLK_MQ_F_TAG_QUEUE_SHARED
;
4047 /* update existing queue */
4048 blk_mq_update_tag_set_shared(set
, false);
4050 mutex_unlock(&set
->tag_list_lock
);
4051 INIT_LIST_HEAD(&q
->tag_set_list
);
4054 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
4055 struct request_queue
*q
)
4057 mutex_lock(&set
->tag_list_lock
);
4060 * Check to see if we're transitioning to shared (from 1 to 2 queues).
4062 if (!list_empty(&set
->tag_list
) &&
4063 !(set
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
)) {
4064 set
->flags
|= BLK_MQ_F_TAG_QUEUE_SHARED
;
4065 /* update existing queue */
4066 blk_mq_update_tag_set_shared(set
, true);
4068 if (set
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
)
4069 queue_set_hctx_shared(q
, true);
4070 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
4072 mutex_unlock(&set
->tag_list_lock
);
4075 /* All allocations will be freed in release handler of q->mq_kobj */
4076 static int blk_mq_alloc_ctxs(struct request_queue
*q
)
4078 struct blk_mq_ctxs
*ctxs
;
4081 ctxs
= kzalloc(sizeof(*ctxs
), GFP_KERNEL
);
4085 ctxs
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
4086 if (!ctxs
->queue_ctx
)
4089 for_each_possible_cpu(cpu
) {
4090 struct blk_mq_ctx
*ctx
= per_cpu_ptr(ctxs
->queue_ctx
, cpu
);
4094 q
->mq_kobj
= &ctxs
->kobj
;
4095 q
->queue_ctx
= ctxs
->queue_ctx
;
4104 * It is the actual release handler for mq, but we do it from
4105 * request queue's release handler for avoiding use-after-free
4106 * and headache because q->mq_kobj shouldn't have been introduced,
4107 * but we can't group ctx/kctx kobj without it.
4109 void blk_mq_release(struct request_queue
*q
)
4111 struct blk_mq_hw_ctx
*hctx
, *next
;
4114 queue_for_each_hw_ctx(q
, hctx
, i
)
4115 WARN_ON_ONCE(hctx
&& list_empty(&hctx
->hctx_list
));
4117 /* all hctx are in .unused_hctx_list now */
4118 list_for_each_entry_safe(hctx
, next
, &q
->unused_hctx_list
, hctx_list
) {
4119 list_del_init(&hctx
->hctx_list
);
4120 kobject_put(&hctx
->kobj
);
4123 xa_destroy(&q
->hctx_table
);
4126 * release .mq_kobj and sw queue's kobject now because
4127 * both share lifetime with request queue.
4129 blk_mq_sysfs_deinit(q
);
4132 static bool blk_mq_can_poll(struct blk_mq_tag_set
*set
)
4134 return set
->nr_maps
> HCTX_TYPE_POLL
&&
4135 set
->map
[HCTX_TYPE_POLL
].nr_queues
;
4138 struct request_queue
*blk_mq_alloc_queue(struct blk_mq_tag_set
*set
,
4139 struct queue_limits
*lim
, void *queuedata
)
4141 struct queue_limits default_lim
= { };
4142 struct request_queue
*q
;
4147 lim
->features
|= BLK_FEAT_IO_STAT
| BLK_FEAT_NOWAIT
;
4148 if (blk_mq_can_poll(set
))
4149 lim
->features
|= BLK_FEAT_POLL
;
4151 q
= blk_alloc_queue(lim
, set
->numa_node
);
4154 q
->queuedata
= queuedata
;
4155 ret
= blk_mq_init_allocated_queue(set
, q
);
4158 return ERR_PTR(ret
);
4162 EXPORT_SYMBOL(blk_mq_alloc_queue
);
4165 * blk_mq_destroy_queue - shutdown a request queue
4166 * @q: request queue to shutdown
4168 * This shuts down a request queue allocated by blk_mq_alloc_queue(). All future
4169 * requests will be failed with -ENODEV. The caller is responsible for dropping
4170 * the reference from blk_mq_alloc_queue() by calling blk_put_queue().
4172 * Context: can sleep
4174 void blk_mq_destroy_queue(struct request_queue
*q
)
4176 WARN_ON_ONCE(!queue_is_mq(q
));
4177 WARN_ON_ONCE(blk_queue_registered(q
));
4181 blk_queue_flag_set(QUEUE_FLAG_DYING
, q
);
4182 blk_queue_start_drain(q
);
4183 blk_mq_freeze_queue_wait(q
);
4186 blk_mq_cancel_work_sync(q
);
4187 blk_mq_exit_queue(q
);
4189 EXPORT_SYMBOL(blk_mq_destroy_queue
);
4191 struct gendisk
*__blk_mq_alloc_disk(struct blk_mq_tag_set
*set
,
4192 struct queue_limits
*lim
, void *queuedata
,
4193 struct lock_class_key
*lkclass
)
4195 struct request_queue
*q
;
4196 struct gendisk
*disk
;
4198 q
= blk_mq_alloc_queue(set
, lim
, queuedata
);
4202 disk
= __alloc_disk_node(q
, set
->numa_node
, lkclass
);
4204 blk_mq_destroy_queue(q
);
4206 return ERR_PTR(-ENOMEM
);
4208 set_bit(GD_OWNS_QUEUE
, &disk
->state
);
4211 EXPORT_SYMBOL(__blk_mq_alloc_disk
);
4213 struct gendisk
*blk_mq_alloc_disk_for_queue(struct request_queue
*q
,
4214 struct lock_class_key
*lkclass
)
4216 struct gendisk
*disk
;
4218 if (!blk_get_queue(q
))
4220 disk
= __alloc_disk_node(q
, NUMA_NO_NODE
, lkclass
);
4225 EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue
);
4227 static struct blk_mq_hw_ctx
*blk_mq_alloc_and_init_hctx(
4228 struct blk_mq_tag_set
*set
, struct request_queue
*q
,
4229 int hctx_idx
, int node
)
4231 struct blk_mq_hw_ctx
*hctx
= NULL
, *tmp
;
4233 /* reuse dead hctx first */
4234 spin_lock(&q
->unused_hctx_lock
);
4235 list_for_each_entry(tmp
, &q
->unused_hctx_list
, hctx_list
) {
4236 if (tmp
->numa_node
== node
) {
4242 list_del_init(&hctx
->hctx_list
);
4243 spin_unlock(&q
->unused_hctx_lock
);
4246 hctx
= blk_mq_alloc_hctx(q
, set
, node
);
4250 if (blk_mq_init_hctx(q
, set
, hctx
, hctx_idx
))
4256 kobject_put(&hctx
->kobj
);
4261 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
4262 struct request_queue
*q
)
4264 struct blk_mq_hw_ctx
*hctx
;
4267 /* protect against switching io scheduler */
4268 mutex_lock(&q
->sysfs_lock
);
4269 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
4271 int node
= blk_mq_get_hctx_node(set
, i
);
4272 struct blk_mq_hw_ctx
*old_hctx
= xa_load(&q
->hctx_table
, i
);
4275 old_node
= old_hctx
->numa_node
;
4276 blk_mq_exit_hctx(q
, set
, old_hctx
, i
);
4279 if (!blk_mq_alloc_and_init_hctx(set
, q
, i
, node
)) {
4282 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4284 hctx
= blk_mq_alloc_and_init_hctx(set
, q
, i
, old_node
);
4285 WARN_ON_ONCE(!hctx
);
4289 * Increasing nr_hw_queues fails. Free the newly allocated
4290 * hctxs and keep the previous q->nr_hw_queues.
4292 if (i
!= set
->nr_hw_queues
) {
4293 j
= q
->nr_hw_queues
;
4296 q
->nr_hw_queues
= set
->nr_hw_queues
;
4299 xa_for_each_start(&q
->hctx_table
, j
, hctx
, j
)
4300 blk_mq_exit_hctx(q
, set
, hctx
, j
);
4301 mutex_unlock(&q
->sysfs_lock
);
4304 int blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
4305 struct request_queue
*q
)
4307 /* mark the queue as mq asap */
4308 q
->mq_ops
= set
->ops
;
4310 if (blk_mq_alloc_ctxs(q
))
4313 /* init q->mq_kobj and sw queues' kobjects */
4314 blk_mq_sysfs_init(q
);
4316 INIT_LIST_HEAD(&q
->unused_hctx_list
);
4317 spin_lock_init(&q
->unused_hctx_lock
);
4319 xa_init(&q
->hctx_table
);
4321 blk_mq_realloc_hw_ctxs(set
, q
);
4322 if (!q
->nr_hw_queues
)
4325 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
4326 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
4330 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
4332 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
4333 INIT_LIST_HEAD(&q
->flush_list
);
4334 INIT_LIST_HEAD(&q
->requeue_list
);
4335 spin_lock_init(&q
->requeue_lock
);
4337 q
->nr_requests
= set
->queue_depth
;
4339 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
4340 blk_mq_add_queue_tag_set(set
, q
);
4341 blk_mq_map_swqueue(q
);
4350 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
4352 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4353 void blk_mq_exit_queue(struct request_queue
*q
)
4355 struct blk_mq_tag_set
*set
= q
->tag_set
;
4357 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4358 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
4359 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4360 blk_mq_del_queue_tag_set(q
);
4363 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
4367 if (blk_mq_is_shared_tags(set
->flags
)) {
4368 set
->shared_tags
= blk_mq_alloc_map_and_rqs(set
,
4371 if (!set
->shared_tags
)
4375 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
4376 if (!__blk_mq_alloc_map_and_rqs(set
, i
))
4385 __blk_mq_free_map_and_rqs(set
, i
);
4387 if (blk_mq_is_shared_tags(set
->flags
)) {
4388 blk_mq_free_map_and_rqs(set
, set
->shared_tags
,
4389 BLK_MQ_NO_HCTX_IDX
);
4396 * Allocate the request maps associated with this tag_set. Note that this
4397 * may reduce the depth asked for, if memory is tight. set->queue_depth
4398 * will be updated to reflect the allocated depth.
4400 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set
*set
)
4405 depth
= set
->queue_depth
;
4407 err
= __blk_mq_alloc_rq_maps(set
);
4411 set
->queue_depth
>>= 1;
4412 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
4416 } while (set
->queue_depth
);
4418 if (!set
->queue_depth
|| err
) {
4419 pr_err("blk-mq: failed to allocate request map\n");
4423 if (depth
!= set
->queue_depth
)
4424 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4425 depth
, set
->queue_depth
);
4430 static void blk_mq_update_queue_map(struct blk_mq_tag_set
*set
)
4433 * blk_mq_map_queues() and multiple .map_queues() implementations
4434 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4435 * number of hardware queues.
4437 if (set
->nr_maps
== 1)
4438 set
->map
[HCTX_TYPE_DEFAULT
].nr_queues
= set
->nr_hw_queues
;
4440 if (set
->ops
->map_queues
) {
4444 * transport .map_queues is usually done in the following
4447 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4448 * mask = get_cpu_mask(queue)
4449 * for_each_cpu(cpu, mask)
4450 * set->map[x].mq_map[cpu] = queue;
4453 * When we need to remap, the table has to be cleared for
4454 * killing stale mapping since one CPU may not be mapped
4457 for (i
= 0; i
< set
->nr_maps
; i
++)
4458 blk_mq_clear_mq_map(&set
->map
[i
]);
4460 set
->ops
->map_queues(set
);
4462 BUG_ON(set
->nr_maps
> 1);
4463 blk_mq_map_queues(&set
->map
[HCTX_TYPE_DEFAULT
]);
4467 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set
*set
,
4468 int new_nr_hw_queues
)
4470 struct blk_mq_tags
**new_tags
;
4473 if (set
->nr_hw_queues
>= new_nr_hw_queues
)
4476 new_tags
= kcalloc_node(new_nr_hw_queues
, sizeof(struct blk_mq_tags
*),
4477 GFP_KERNEL
, set
->numa_node
);
4482 memcpy(new_tags
, set
->tags
, set
->nr_hw_queues
*
4483 sizeof(*set
->tags
));
4485 set
->tags
= new_tags
;
4487 for (i
= set
->nr_hw_queues
; i
< new_nr_hw_queues
; i
++) {
4488 if (!__blk_mq_alloc_map_and_rqs(set
, i
)) {
4489 while (--i
>= set
->nr_hw_queues
)
4490 __blk_mq_free_map_and_rqs(set
, i
);
4497 set
->nr_hw_queues
= new_nr_hw_queues
;
4502 * Alloc a tag set to be associated with one or more request queues.
4503 * May fail with EINVAL for various error conditions. May adjust the
4504 * requested depth down, if it's too large. In that case, the set
4505 * value will be stored in set->queue_depth.
4507 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
4511 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
4513 if (!set
->nr_hw_queues
)
4515 if (!set
->queue_depth
)
4517 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
4520 if (!set
->ops
->queue_rq
)
4523 if (!set
->ops
->get_budget
^ !set
->ops
->put_budget
)
4526 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
4527 pr_info("blk-mq: reduced tag depth to %u\n",
4529 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
4534 else if (set
->nr_maps
> HCTX_MAX_TYPES
)
4538 * If a crashdump is active, then we are potentially in a very
4539 * memory constrained environment. Limit us to 64 tags to prevent
4540 * using too much memory.
4542 if (is_kdump_kernel())
4543 set
->queue_depth
= min(64U, set
->queue_depth
);
4546 * There is no use for more h/w queues than cpus if we just have
4549 if (set
->nr_maps
== 1 && set
->nr_hw_queues
> nr_cpu_ids
)
4550 set
->nr_hw_queues
= nr_cpu_ids
;
4552 if (set
->flags
& BLK_MQ_F_BLOCKING
) {
4553 set
->srcu
= kmalloc(sizeof(*set
->srcu
), GFP_KERNEL
);
4556 ret
= init_srcu_struct(set
->srcu
);
4562 set
->tags
= kcalloc_node(set
->nr_hw_queues
,
4563 sizeof(struct blk_mq_tags
*), GFP_KERNEL
,
4566 goto out_cleanup_srcu
;
4568 for (i
= 0; i
< set
->nr_maps
; i
++) {
4569 set
->map
[i
].mq_map
= kcalloc_node(nr_cpu_ids
,
4570 sizeof(set
->map
[i
].mq_map
[0]),
4571 GFP_KERNEL
, set
->numa_node
);
4572 if (!set
->map
[i
].mq_map
)
4573 goto out_free_mq_map
;
4574 set
->map
[i
].nr_queues
= set
->nr_hw_queues
;
4577 blk_mq_update_queue_map(set
);
4579 ret
= blk_mq_alloc_set_map_and_rqs(set
);
4581 goto out_free_mq_map
;
4583 mutex_init(&set
->tag_list_lock
);
4584 INIT_LIST_HEAD(&set
->tag_list
);
4589 for (i
= 0; i
< set
->nr_maps
; i
++) {
4590 kfree(set
->map
[i
].mq_map
);
4591 set
->map
[i
].mq_map
= NULL
;
4596 if (set
->flags
& BLK_MQ_F_BLOCKING
)
4597 cleanup_srcu_struct(set
->srcu
);
4599 if (set
->flags
& BLK_MQ_F_BLOCKING
)
4603 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
4605 /* allocate and initialize a tagset for a simple single-queue device */
4606 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set
*set
,
4607 const struct blk_mq_ops
*ops
, unsigned int queue_depth
,
4608 unsigned int set_flags
)
4610 memset(set
, 0, sizeof(*set
));
4612 set
->nr_hw_queues
= 1;
4614 set
->queue_depth
= queue_depth
;
4615 set
->numa_node
= NUMA_NO_NODE
;
4616 set
->flags
= set_flags
;
4617 return blk_mq_alloc_tag_set(set
);
4619 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set
);
4621 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
4625 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
4626 __blk_mq_free_map_and_rqs(set
, i
);
4628 if (blk_mq_is_shared_tags(set
->flags
)) {
4629 blk_mq_free_map_and_rqs(set
, set
->shared_tags
,
4630 BLK_MQ_NO_HCTX_IDX
);
4633 for (j
= 0; j
< set
->nr_maps
; j
++) {
4634 kfree(set
->map
[j
].mq_map
);
4635 set
->map
[j
].mq_map
= NULL
;
4640 if (set
->flags
& BLK_MQ_F_BLOCKING
) {
4641 cleanup_srcu_struct(set
->srcu
);
4645 EXPORT_SYMBOL(blk_mq_free_tag_set
);
4647 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
4649 struct blk_mq_tag_set
*set
= q
->tag_set
;
4650 struct blk_mq_hw_ctx
*hctx
;
4654 if (WARN_ON_ONCE(!q
->mq_freeze_depth
))
4660 if (q
->nr_requests
== nr
)
4663 blk_mq_quiesce_queue(q
);
4666 queue_for_each_hw_ctx(q
, hctx
, i
) {
4670 * If we're using an MQ scheduler, just update the scheduler
4671 * queue depth. This is similar to what the old code would do.
4673 if (hctx
->sched_tags
) {
4674 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
4677 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
, nr
,
4682 if (q
->elevator
&& q
->elevator
->type
->ops
.depth_updated
)
4683 q
->elevator
->type
->ops
.depth_updated(hctx
);
4686 q
->nr_requests
= nr
;
4687 if (blk_mq_is_shared_tags(set
->flags
)) {
4689 blk_mq_tag_update_sched_shared_tags(q
);
4691 blk_mq_tag_resize_shared_tags(set
, nr
);
4695 blk_mq_unquiesce_queue(q
);
4701 * request_queue and elevator_type pair.
4702 * It is just used by __blk_mq_update_nr_hw_queues to cache
4703 * the elevator_type associated with a request_queue.
4705 struct blk_mq_qe_pair
{
4706 struct list_head node
;
4707 struct request_queue
*q
;
4708 struct elevator_type
*type
;
4712 * Cache the elevator_type in qe pair list and switch the
4713 * io scheduler to 'none'
4715 static bool blk_mq_elv_switch_none(struct list_head
*head
,
4716 struct request_queue
*q
)
4718 struct blk_mq_qe_pair
*qe
;
4720 qe
= kmalloc(sizeof(*qe
), GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
);
4724 /* q->elevator needs protection from ->sysfs_lock */
4725 mutex_lock(&q
->sysfs_lock
);
4727 /* the check has to be done with holding sysfs_lock */
4733 INIT_LIST_HEAD(&qe
->node
);
4735 qe
->type
= q
->elevator
->type
;
4736 /* keep a reference to the elevator module as we'll switch back */
4737 __elevator_get(qe
->type
);
4738 list_add(&qe
->node
, head
);
4739 elevator_disable(q
);
4741 mutex_unlock(&q
->sysfs_lock
);
4746 static struct blk_mq_qe_pair
*blk_lookup_qe_pair(struct list_head
*head
,
4747 struct request_queue
*q
)
4749 struct blk_mq_qe_pair
*qe
;
4751 list_for_each_entry(qe
, head
, node
)
4758 static void blk_mq_elv_switch_back(struct list_head
*head
,
4759 struct request_queue
*q
)
4761 struct blk_mq_qe_pair
*qe
;
4762 struct elevator_type
*t
;
4764 qe
= blk_lookup_qe_pair(head
, q
);
4768 list_del(&qe
->node
);
4771 mutex_lock(&q
->sysfs_lock
);
4772 elevator_switch(q
, t
);
4773 /* drop the reference acquired in blk_mq_elv_switch_none */
4775 mutex_unlock(&q
->sysfs_lock
);
4778 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
,
4781 struct request_queue
*q
;
4783 int prev_nr_hw_queues
= set
->nr_hw_queues
;
4786 lockdep_assert_held(&set
->tag_list_lock
);
4788 if (set
->nr_maps
== 1 && nr_hw_queues
> nr_cpu_ids
)
4789 nr_hw_queues
= nr_cpu_ids
;
4790 if (nr_hw_queues
< 1)
4792 if (set
->nr_maps
== 1 && nr_hw_queues
== set
->nr_hw_queues
)
4795 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
4796 blk_mq_freeze_queue(q
);
4798 * Switch IO scheduler to 'none', cleaning up the data associated
4799 * with the previous scheduler. We will switch back once we are done
4800 * updating the new sw to hw queue mappings.
4802 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
4803 if (!blk_mq_elv_switch_none(&head
, q
))
4806 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
4807 blk_mq_debugfs_unregister_hctxs(q
);
4808 blk_mq_sysfs_unregister_hctxs(q
);
4811 if (blk_mq_realloc_tag_set_tags(set
, nr_hw_queues
) < 0)
4815 blk_mq_update_queue_map(set
);
4816 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
4817 struct queue_limits lim
;
4819 blk_mq_realloc_hw_ctxs(set
, q
);
4821 if (q
->nr_hw_queues
!= set
->nr_hw_queues
) {
4822 int i
= prev_nr_hw_queues
;
4824 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4825 nr_hw_queues
, prev_nr_hw_queues
);
4826 for (; i
< set
->nr_hw_queues
; i
++)
4827 __blk_mq_free_map_and_rqs(set
, i
);
4829 set
->nr_hw_queues
= prev_nr_hw_queues
;
4832 lim
= queue_limits_start_update(q
);
4833 if (blk_mq_can_poll(set
))
4834 lim
.features
|= BLK_FEAT_POLL
;
4836 lim
.features
&= ~BLK_FEAT_POLL
;
4837 if (queue_limits_commit_update(q
, &lim
) < 0)
4838 pr_warn("updating the poll flag failed\n");
4839 blk_mq_map_swqueue(q
);
4843 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
4844 blk_mq_sysfs_register_hctxs(q
);
4845 blk_mq_debugfs_register_hctxs(q
);
4849 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
4850 blk_mq_elv_switch_back(&head
, q
);
4852 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
4853 blk_mq_unfreeze_queue(q
);
4855 /* Free the excess tags when nr_hw_queues shrink. */
4856 for (i
= set
->nr_hw_queues
; i
< prev_nr_hw_queues
; i
++)
4857 __blk_mq_free_map_and_rqs(set
, i
);
4860 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
4862 mutex_lock(&set
->tag_list_lock
);
4863 __blk_mq_update_nr_hw_queues(set
, nr_hw_queues
);
4864 mutex_unlock(&set
->tag_list_lock
);
4866 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
4868 static int blk_hctx_poll(struct request_queue
*q
, struct blk_mq_hw_ctx
*hctx
,
4869 struct io_comp_batch
*iob
, unsigned int flags
)
4871 long state
= get_current_state();
4875 ret
= q
->mq_ops
->poll(hctx
, iob
);
4877 __set_current_state(TASK_RUNNING
);
4881 if (signal_pending_state(state
, current
))
4882 __set_current_state(TASK_RUNNING
);
4883 if (task_is_running(current
))
4886 if (ret
< 0 || (flags
& BLK_POLL_ONESHOT
))
4889 } while (!need_resched());
4891 __set_current_state(TASK_RUNNING
);
4895 int blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
,
4896 struct io_comp_batch
*iob
, unsigned int flags
)
4898 struct blk_mq_hw_ctx
*hctx
= xa_load(&q
->hctx_table
, cookie
);
4900 return blk_hctx_poll(q
, hctx
, iob
, flags
);
4903 int blk_rq_poll(struct request
*rq
, struct io_comp_batch
*iob
,
4904 unsigned int poll_flags
)
4906 struct request_queue
*q
= rq
->q
;
4909 if (!blk_rq_is_poll(rq
))
4911 if (!percpu_ref_tryget(&q
->q_usage_counter
))
4914 ret
= blk_hctx_poll(q
, rq
->mq_hctx
, iob
, poll_flags
);
4919 EXPORT_SYMBOL_GPL(blk_rq_poll
);
4921 unsigned int blk_mq_rq_cpu(struct request
*rq
)
4923 return rq
->mq_ctx
->cpu
;
4925 EXPORT_SYMBOL(blk_mq_rq_cpu
);
4927 void blk_mq_cancel_work_sync(struct request_queue
*q
)
4929 struct blk_mq_hw_ctx
*hctx
;
4932 cancel_delayed_work_sync(&q
->requeue_work
);
4934 queue_for_each_hw_ctx(q
, hctx
, i
)
4935 cancel_delayed_work_sync(&hctx
->run_work
);
4938 static int __init
blk_mq_init(void)
4942 for_each_possible_cpu(i
)
4943 init_llist_head(&per_cpu(blk_cpu_done
, i
));
4944 for_each_possible_cpu(i
)
4945 INIT_CSD(&per_cpu(blk_cpu_csd
, i
),
4946 __blk_mq_complete_request_remote
, NULL
);
4947 open_softirq(BLOCK_SOFTIRQ
, blk_done_softirq
);
4949 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD
,
4950 "block/softirq:dead", NULL
,
4951 blk_softirq_cpu_dead
);
4952 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
4953 blk_mq_hctx_notify_dead
);
4954 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE
, "block/mq:online",
4955 blk_mq_hctx_notify_online
,
4956 blk_mq_hctx_notify_offline
);
4959 subsys_initcall(blk_mq_init
);