1 // SPDX-License-Identifier: GPL-2.0
3 * Block multiqueue core code
5 * Copyright (C) 2013-2014 Jens Axboe
6 * Copyright (C) 2013-2014 Christoph Hellwig
8 #include <linux/kernel.h>
9 #include <linux/module.h>
10 #include <linux/backing-dev.h>
11 #include <linux/bio.h>
12 #include <linux/blkdev.h>
13 #include <linux/blk-integrity.h>
14 #include <linux/kmemleak.h>
16 #include <linux/init.h>
17 #include <linux/slab.h>
18 #include <linux/workqueue.h>
19 #include <linux/smp.h>
20 #include <linux/interrupt.h>
21 #include <linux/llist.h>
22 #include <linux/cpu.h>
23 #include <linux/cache.h>
24 #include <linux/sched/sysctl.h>
25 #include <linux/sched/topology.h>
26 #include <linux/sched/signal.h>
27 #include <linux/delay.h>
28 #include <linux/crash_dump.h>
29 #include <linux/prefetch.h>
30 #include <linux/blk-crypto.h>
31 #include <linux/part_stat.h>
33 #include <trace/events/block.h>
35 #include <linux/blk-mq.h>
36 #include <linux/t10-pi.h>
39 #include "blk-mq-debugfs.h"
40 #include "blk-mq-tag.h"
43 #include "blk-mq-sched.h"
44 #include "blk-rq-qos.h"
46 static DEFINE_PER_CPU(struct llist_head
, blk_cpu_done
);
48 static void blk_mq_poll_stats_start(struct request_queue
*q
);
49 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
);
51 static int blk_mq_poll_stats_bkt(const struct request
*rq
)
53 int ddir
, sectors
, bucket
;
55 ddir
= rq_data_dir(rq
);
56 sectors
= blk_rq_stats_sectors(rq
);
58 bucket
= ddir
+ 2 * ilog2(sectors
);
62 else if (bucket
>= BLK_MQ_POLL_STATS_BKTS
)
63 return ddir
+ BLK_MQ_POLL_STATS_BKTS
- 2;
68 #define BLK_QC_T_SHIFT 16
69 #define BLK_QC_T_INTERNAL (1U << 31)
71 static inline struct blk_mq_hw_ctx
*blk_qc_to_hctx(struct request_queue
*q
,
74 return q
->queue_hw_ctx
[(qc
& ~BLK_QC_T_INTERNAL
) >> BLK_QC_T_SHIFT
];
77 static inline struct request
*blk_qc_to_rq(struct blk_mq_hw_ctx
*hctx
,
80 unsigned int tag
= qc
& ((1U << BLK_QC_T_SHIFT
) - 1);
82 if (qc
& BLK_QC_T_INTERNAL
)
83 return blk_mq_tag_to_rq(hctx
->sched_tags
, tag
);
84 return blk_mq_tag_to_rq(hctx
->tags
, tag
);
87 static inline blk_qc_t
blk_rq_to_qc(struct request
*rq
)
89 return (rq
->mq_hctx
->queue_num
<< BLK_QC_T_SHIFT
) |
91 rq
->tag
: (rq
->internal_tag
| BLK_QC_T_INTERNAL
));
95 * Check if any of the ctx, dispatch list or elevator
96 * have pending work in this hardware queue.
98 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
100 return !list_empty_careful(&hctx
->dispatch
) ||
101 sbitmap_any_bit_set(&hctx
->ctx_map
) ||
102 blk_mq_sched_has_work(hctx
);
106 * Mark this ctx as having pending work in this hardware queue
108 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
109 struct blk_mq_ctx
*ctx
)
111 const int bit
= ctx
->index_hw
[hctx
->type
];
113 if (!sbitmap_test_bit(&hctx
->ctx_map
, bit
))
114 sbitmap_set_bit(&hctx
->ctx_map
, bit
);
117 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
118 struct blk_mq_ctx
*ctx
)
120 const int bit
= ctx
->index_hw
[hctx
->type
];
122 sbitmap_clear_bit(&hctx
->ctx_map
, bit
);
126 struct block_device
*part
;
127 unsigned int inflight
[2];
130 static bool blk_mq_check_inflight(struct request
*rq
, void *priv
,
133 struct mq_inflight
*mi
= priv
;
135 if ((!mi
->part
->bd_partno
|| rq
->part
== mi
->part
) &&
136 blk_mq_rq_state(rq
) == MQ_RQ_IN_FLIGHT
)
137 mi
->inflight
[rq_data_dir(rq
)]++;
142 unsigned int blk_mq_in_flight(struct request_queue
*q
,
143 struct block_device
*part
)
145 struct mq_inflight mi
= { .part
= part
};
147 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight
, &mi
);
149 return mi
.inflight
[0] + mi
.inflight
[1];
152 void blk_mq_in_flight_rw(struct request_queue
*q
, struct block_device
*part
,
153 unsigned int inflight
[2])
155 struct mq_inflight mi
= { .part
= part
};
157 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight
, &mi
);
158 inflight
[0] = mi
.inflight
[0];
159 inflight
[1] = mi
.inflight
[1];
162 void blk_freeze_queue_start(struct request_queue
*q
)
164 mutex_lock(&q
->mq_freeze_lock
);
165 if (++q
->mq_freeze_depth
== 1) {
166 percpu_ref_kill(&q
->q_usage_counter
);
167 mutex_unlock(&q
->mq_freeze_lock
);
169 blk_mq_run_hw_queues(q
, false);
171 mutex_unlock(&q
->mq_freeze_lock
);
174 EXPORT_SYMBOL_GPL(blk_freeze_queue_start
);
176 void blk_mq_freeze_queue_wait(struct request_queue
*q
)
178 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
180 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait
);
182 int blk_mq_freeze_queue_wait_timeout(struct request_queue
*q
,
183 unsigned long timeout
)
185 return wait_event_timeout(q
->mq_freeze_wq
,
186 percpu_ref_is_zero(&q
->q_usage_counter
),
189 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout
);
192 * Guarantee no request is in use, so we can change any data structure of
193 * the queue afterward.
195 void blk_freeze_queue(struct request_queue
*q
)
198 * In the !blk_mq case we are only calling this to kill the
199 * q_usage_counter, otherwise this increases the freeze depth
200 * and waits for it to return to zero. For this reason there is
201 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
202 * exported to drivers as the only user for unfreeze is blk_mq.
204 blk_freeze_queue_start(q
);
205 blk_mq_freeze_queue_wait(q
);
208 void blk_mq_freeze_queue(struct request_queue
*q
)
211 * ...just an alias to keep freeze and unfreeze actions balanced
212 * in the blk_mq_* namespace
216 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
218 void __blk_mq_unfreeze_queue(struct request_queue
*q
, bool force_atomic
)
220 mutex_lock(&q
->mq_freeze_lock
);
222 q
->q_usage_counter
.data
->force_atomic
= true;
223 q
->mq_freeze_depth
--;
224 WARN_ON_ONCE(q
->mq_freeze_depth
< 0);
225 if (!q
->mq_freeze_depth
) {
226 percpu_ref_resurrect(&q
->q_usage_counter
);
227 wake_up_all(&q
->mq_freeze_wq
);
229 mutex_unlock(&q
->mq_freeze_lock
);
232 void blk_mq_unfreeze_queue(struct request_queue
*q
)
234 __blk_mq_unfreeze_queue(q
, false);
236 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
239 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
240 * mpt3sas driver such that this function can be removed.
242 void blk_mq_quiesce_queue_nowait(struct request_queue
*q
)
246 spin_lock_irqsave(&q
->queue_lock
, flags
);
247 if (!q
->quiesce_depth
++)
248 blk_queue_flag_set(QUEUE_FLAG_QUIESCED
, q
);
249 spin_unlock_irqrestore(&q
->queue_lock
, flags
);
251 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait
);
254 * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done
257 * Note: it is driver's responsibility for making sure that quiesce has
260 void blk_mq_wait_quiesce_done(struct request_queue
*q
)
262 if (blk_queue_has_srcu(q
))
263 synchronize_srcu(q
->srcu
);
267 EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done
);
270 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
273 * Note: this function does not prevent that the struct request end_io()
274 * callback function is invoked. Once this function is returned, we make
275 * sure no dispatch can happen until the queue is unquiesced via
276 * blk_mq_unquiesce_queue().
278 void blk_mq_quiesce_queue(struct request_queue
*q
)
280 blk_mq_quiesce_queue_nowait(q
);
281 blk_mq_wait_quiesce_done(q
);
283 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
286 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
289 * This function recovers queue into the state before quiescing
290 * which is done by blk_mq_quiesce_queue.
292 void blk_mq_unquiesce_queue(struct request_queue
*q
)
295 bool run_queue
= false;
297 spin_lock_irqsave(&q
->queue_lock
, flags
);
298 if (WARN_ON_ONCE(q
->quiesce_depth
<= 0)) {
300 } else if (!--q
->quiesce_depth
) {
301 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED
, q
);
304 spin_unlock_irqrestore(&q
->queue_lock
, flags
);
306 /* dispatch requests which are inserted during quiescing */
308 blk_mq_run_hw_queues(q
, true);
310 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue
);
312 void blk_mq_wake_waiters(struct request_queue
*q
)
314 struct blk_mq_hw_ctx
*hctx
;
317 queue_for_each_hw_ctx(q
, hctx
, i
)
318 if (blk_mq_hw_queue_mapped(hctx
))
319 blk_mq_tag_wakeup_all(hctx
->tags
, true);
322 void blk_rq_init(struct request_queue
*q
, struct request
*rq
)
324 memset(rq
, 0, sizeof(*rq
));
326 INIT_LIST_HEAD(&rq
->queuelist
);
328 rq
->__sector
= (sector_t
) -1;
329 INIT_HLIST_NODE(&rq
->hash
);
330 RB_CLEAR_NODE(&rq
->rb_node
);
331 rq
->tag
= BLK_MQ_NO_TAG
;
332 rq
->internal_tag
= BLK_MQ_NO_TAG
;
333 rq
->start_time_ns
= ktime_get_ns();
335 blk_crypto_rq_set_defaults(rq
);
337 EXPORT_SYMBOL(blk_rq_init
);
339 static struct request
*blk_mq_rq_ctx_init(struct blk_mq_alloc_data
*data
,
340 struct blk_mq_tags
*tags
, unsigned int tag
, u64 alloc_time_ns
)
342 struct blk_mq_ctx
*ctx
= data
->ctx
;
343 struct blk_mq_hw_ctx
*hctx
= data
->hctx
;
344 struct request_queue
*q
= data
->q
;
345 struct request
*rq
= tags
->static_rqs
[tag
];
350 rq
->cmd_flags
= data
->cmd_flags
;
352 if (data
->flags
& BLK_MQ_REQ_PM
)
353 data
->rq_flags
|= RQF_PM
;
354 if (blk_queue_io_stat(q
))
355 data
->rq_flags
|= RQF_IO_STAT
;
356 rq
->rq_flags
= data
->rq_flags
;
358 if (!(data
->rq_flags
& RQF_ELV
)) {
360 rq
->internal_tag
= BLK_MQ_NO_TAG
;
362 rq
->tag
= BLK_MQ_NO_TAG
;
363 rq
->internal_tag
= tag
;
367 if (blk_mq_need_time_stamp(rq
))
368 rq
->start_time_ns
= ktime_get_ns();
370 rq
->start_time_ns
= 0;
372 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
373 rq
->alloc_time_ns
= alloc_time_ns
;
375 rq
->io_start_time_ns
= 0;
376 rq
->stats_sectors
= 0;
377 rq
->nr_phys_segments
= 0;
378 #if defined(CONFIG_BLK_DEV_INTEGRITY)
379 rq
->nr_integrity_segments
= 0;
382 rq
->end_io_data
= NULL
;
384 blk_crypto_rq_set_defaults(rq
);
385 INIT_LIST_HEAD(&rq
->queuelist
);
386 /* tag was already set */
387 WRITE_ONCE(rq
->deadline
, 0);
390 if (rq
->rq_flags
& RQF_ELV
) {
391 struct elevator_queue
*e
= data
->q
->elevator
;
393 INIT_HLIST_NODE(&rq
->hash
);
394 RB_CLEAR_NODE(&rq
->rb_node
);
396 if (!op_is_flush(data
->cmd_flags
) &&
397 e
->type
->ops
.prepare_request
) {
398 e
->type
->ops
.prepare_request(rq
);
399 rq
->rq_flags
|= RQF_ELVPRIV
;
406 static inline struct request
*
407 __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data
*data
,
410 unsigned int tag
, tag_offset
;
411 struct blk_mq_tags
*tags
;
413 unsigned long tag_mask
;
416 tag_mask
= blk_mq_get_tags(data
, data
->nr_tags
, &tag_offset
);
417 if (unlikely(!tag_mask
))
420 tags
= blk_mq_tags_from_data(data
);
421 for (i
= 0; tag_mask
; i
++) {
422 if (!(tag_mask
& (1UL << i
)))
424 tag
= tag_offset
+ i
;
425 prefetch(tags
->static_rqs
[tag
]);
426 tag_mask
&= ~(1UL << i
);
427 rq
= blk_mq_rq_ctx_init(data
, tags
, tag
, alloc_time_ns
);
428 rq_list_add(data
->cached_rq
, rq
);
431 /* caller already holds a reference, add for remainder */
432 percpu_ref_get_many(&data
->q
->q_usage_counter
, nr
- 1);
435 return rq_list_pop(data
->cached_rq
);
438 static struct request
*__blk_mq_alloc_requests(struct blk_mq_alloc_data
*data
)
440 struct request_queue
*q
= data
->q
;
441 u64 alloc_time_ns
= 0;
445 /* alloc_time includes depth and tag waits */
446 if (blk_queue_rq_alloc_time(q
))
447 alloc_time_ns
= ktime_get_ns();
449 if (data
->cmd_flags
& REQ_NOWAIT
)
450 data
->flags
|= BLK_MQ_REQ_NOWAIT
;
453 struct elevator_queue
*e
= q
->elevator
;
455 data
->rq_flags
|= RQF_ELV
;
458 * Flush/passthrough requests are special and go directly to the
459 * dispatch list. Don't include reserved tags in the
460 * limiting, as it isn't useful.
462 if (!op_is_flush(data
->cmd_flags
) &&
463 !blk_op_is_passthrough(data
->cmd_flags
) &&
464 e
->type
->ops
.limit_depth
&&
465 !(data
->flags
& BLK_MQ_REQ_RESERVED
))
466 e
->type
->ops
.limit_depth(data
->cmd_flags
, data
);
470 data
->ctx
= blk_mq_get_ctx(q
);
471 data
->hctx
= blk_mq_map_queue(q
, data
->cmd_flags
, data
->ctx
);
472 if (!(data
->rq_flags
& RQF_ELV
))
473 blk_mq_tag_busy(data
->hctx
);
476 * Try batched alloc if we want more than 1 tag.
478 if (data
->nr_tags
> 1) {
479 rq
= __blk_mq_alloc_requests_batch(data
, alloc_time_ns
);
486 * Waiting allocations only fail because of an inactive hctx. In that
487 * case just retry the hctx assignment and tag allocation as CPU hotplug
488 * should have migrated us to an online CPU by now.
490 tag
= blk_mq_get_tag(data
);
491 if (tag
== BLK_MQ_NO_TAG
) {
492 if (data
->flags
& BLK_MQ_REQ_NOWAIT
)
495 * Give up the CPU and sleep for a random short time to
496 * ensure that thread using a realtime scheduling class
497 * are migrated off the CPU, and thus off the hctx that
504 return blk_mq_rq_ctx_init(data
, blk_mq_tags_from_data(data
), tag
,
508 struct request
*blk_mq_alloc_request(struct request_queue
*q
, unsigned int op
,
509 blk_mq_req_flags_t flags
)
511 struct blk_mq_alloc_data data
= {
520 ret
= blk_queue_enter(q
, flags
);
524 rq
= __blk_mq_alloc_requests(&data
);
528 rq
->__sector
= (sector_t
) -1;
529 rq
->bio
= rq
->biotail
= NULL
;
533 return ERR_PTR(-EWOULDBLOCK
);
535 EXPORT_SYMBOL(blk_mq_alloc_request
);
537 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
,
538 unsigned int op
, blk_mq_req_flags_t flags
, unsigned int hctx_idx
)
540 struct blk_mq_alloc_data data
= {
546 u64 alloc_time_ns
= 0;
551 /* alloc_time includes depth and tag waits */
552 if (blk_queue_rq_alloc_time(q
))
553 alloc_time_ns
= ktime_get_ns();
556 * If the tag allocator sleeps we could get an allocation for a
557 * different hardware context. No need to complicate the low level
558 * allocator for this for the rare use case of a command tied to
561 if (WARN_ON_ONCE(!(flags
& (BLK_MQ_REQ_NOWAIT
| BLK_MQ_REQ_RESERVED
))))
562 return ERR_PTR(-EINVAL
);
564 if (hctx_idx
>= q
->nr_hw_queues
)
565 return ERR_PTR(-EIO
);
567 ret
= blk_queue_enter(q
, flags
);
572 * Check if the hardware context is actually mapped to anything.
573 * If not tell the caller that it should skip this queue.
576 data
.hctx
= q
->queue_hw_ctx
[hctx_idx
];
577 if (!blk_mq_hw_queue_mapped(data
.hctx
))
579 cpu
= cpumask_first_and(data
.hctx
->cpumask
, cpu_online_mask
);
580 data
.ctx
= __blk_mq_get_ctx(q
, cpu
);
583 blk_mq_tag_busy(data
.hctx
);
585 data
.rq_flags
|= RQF_ELV
;
588 tag
= blk_mq_get_tag(&data
);
589 if (tag
== BLK_MQ_NO_TAG
)
591 return blk_mq_rq_ctx_init(&data
, blk_mq_tags_from_data(&data
), tag
,
598 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
600 static void __blk_mq_free_request(struct request
*rq
)
602 struct request_queue
*q
= rq
->q
;
603 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
604 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
605 const int sched_tag
= rq
->internal_tag
;
607 blk_crypto_free_request(rq
);
608 blk_pm_mark_last_busy(rq
);
610 if (rq
->tag
!= BLK_MQ_NO_TAG
)
611 blk_mq_put_tag(hctx
->tags
, ctx
, rq
->tag
);
612 if (sched_tag
!= BLK_MQ_NO_TAG
)
613 blk_mq_put_tag(hctx
->sched_tags
, ctx
, sched_tag
);
614 blk_mq_sched_restart(hctx
);
618 void blk_mq_free_request(struct request
*rq
)
620 struct request_queue
*q
= rq
->q
;
621 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
623 if ((rq
->rq_flags
& RQF_ELVPRIV
) &&
624 q
->elevator
->type
->ops
.finish_request
)
625 q
->elevator
->type
->ops
.finish_request(rq
);
627 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
628 __blk_mq_dec_active_requests(hctx
);
630 if (unlikely(laptop_mode
&& !blk_rq_is_passthrough(rq
)))
631 laptop_io_completion(q
->disk
->bdi
);
635 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
636 if (req_ref_put_and_test(rq
))
637 __blk_mq_free_request(rq
);
639 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
641 void blk_mq_free_plug_rqs(struct blk_plug
*plug
)
645 while ((rq
= rq_list_pop(&plug
->cached_rq
)) != NULL
)
646 blk_mq_free_request(rq
);
649 void blk_dump_rq_flags(struct request
*rq
, char *msg
)
651 printk(KERN_INFO
"%s: dev %s: flags=%llx\n", msg
,
652 rq
->q
->disk
? rq
->q
->disk
->disk_name
: "?",
653 (unsigned long long) rq
->cmd_flags
);
655 printk(KERN_INFO
" sector %llu, nr/cnr %u/%u\n",
656 (unsigned long long)blk_rq_pos(rq
),
657 blk_rq_sectors(rq
), blk_rq_cur_sectors(rq
));
658 printk(KERN_INFO
" bio %p, biotail %p, len %u\n",
659 rq
->bio
, rq
->biotail
, blk_rq_bytes(rq
));
661 EXPORT_SYMBOL(blk_dump_rq_flags
);
663 static void req_bio_endio(struct request
*rq
, struct bio
*bio
,
664 unsigned int nbytes
, blk_status_t error
)
666 if (unlikely(error
)) {
667 bio
->bi_status
= error
;
668 } else if (req_op(rq
) == REQ_OP_ZONE_APPEND
) {
670 * Partial zone append completions cannot be supported as the
671 * BIO fragments may end up not being written sequentially.
673 if (bio
->bi_iter
.bi_size
!= nbytes
)
674 bio
->bi_status
= BLK_STS_IOERR
;
676 bio
->bi_iter
.bi_sector
= rq
->__sector
;
679 bio_advance(bio
, nbytes
);
681 if (unlikely(rq
->rq_flags
& RQF_QUIET
))
682 bio_set_flag(bio
, BIO_QUIET
);
683 /* don't actually finish bio if it's part of flush sequence */
684 if (bio
->bi_iter
.bi_size
== 0 && !(rq
->rq_flags
& RQF_FLUSH_SEQ
))
688 static void blk_account_io_completion(struct request
*req
, unsigned int bytes
)
690 if (req
->part
&& blk_do_io_stat(req
)) {
691 const int sgrp
= op_stat_group(req_op(req
));
694 part_stat_add(req
->part
, sectors
[sgrp
], bytes
>> 9);
699 static void blk_print_req_error(struct request
*req
, blk_status_t status
)
701 printk_ratelimited(KERN_ERR
702 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
703 "phys_seg %u prio class %u\n",
704 blk_status_to_str(status
),
705 req
->q
->disk
? req
->q
->disk
->disk_name
: "?",
706 blk_rq_pos(req
), req_op(req
), blk_op_str(req_op(req
)),
707 req
->cmd_flags
& ~REQ_OP_MASK
,
708 req
->nr_phys_segments
,
709 IOPRIO_PRIO_CLASS(req
->ioprio
));
713 * Fully end IO on a request. Does not support partial completions, or
716 static void blk_complete_request(struct request
*req
)
718 const bool is_flush
= (req
->rq_flags
& RQF_FLUSH_SEQ
) != 0;
719 int total_bytes
= blk_rq_bytes(req
);
720 struct bio
*bio
= req
->bio
;
722 trace_block_rq_complete(req
, BLK_STS_OK
, total_bytes
);
727 #ifdef CONFIG_BLK_DEV_INTEGRITY
728 if (blk_integrity_rq(req
) && req_op(req
) == REQ_OP_READ
)
729 req
->q
->integrity
.profile
->complete_fn(req
, total_bytes
);
732 blk_account_io_completion(req
, total_bytes
);
735 struct bio
*next
= bio
->bi_next
;
737 /* Completion has already been traced */
738 bio_clear_flag(bio
, BIO_TRACE_COMPLETION
);
745 * Reset counters so that the request stacking driver
746 * can find how many bytes remain in the request
754 * blk_update_request - Complete multiple bytes without completing the request
755 * @req: the request being processed
756 * @error: block status code
757 * @nr_bytes: number of bytes to complete for @req
760 * Ends I/O on a number of bytes attached to @req, but doesn't complete
761 * the request structure even if @req doesn't have leftover.
762 * If @req has leftover, sets it up for the next range of segments.
764 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
765 * %false return from this function.
768 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
769 * except in the consistency check at the end of this function.
772 * %false - this request doesn't have any more data
773 * %true - this request has more data
775 bool blk_update_request(struct request
*req
, blk_status_t error
,
776 unsigned int nr_bytes
)
780 trace_block_rq_complete(req
, error
, nr_bytes
);
785 #ifdef CONFIG_BLK_DEV_INTEGRITY
786 if (blk_integrity_rq(req
) && req_op(req
) == REQ_OP_READ
&&
788 req
->q
->integrity
.profile
->complete_fn(req
, nr_bytes
);
791 if (unlikely(error
&& !blk_rq_is_passthrough(req
) &&
792 !(req
->rq_flags
& RQF_QUIET
)))
793 blk_print_req_error(req
, error
);
795 blk_account_io_completion(req
, nr_bytes
);
799 struct bio
*bio
= req
->bio
;
800 unsigned bio_bytes
= min(bio
->bi_iter
.bi_size
, nr_bytes
);
802 if (bio_bytes
== bio
->bi_iter
.bi_size
)
803 req
->bio
= bio
->bi_next
;
805 /* Completion has already been traced */
806 bio_clear_flag(bio
, BIO_TRACE_COMPLETION
);
807 req_bio_endio(req
, bio
, bio_bytes
, error
);
809 total_bytes
+= bio_bytes
;
810 nr_bytes
-= bio_bytes
;
821 * Reset counters so that the request stacking driver
822 * can find how many bytes remain in the request
829 req
->__data_len
-= total_bytes
;
831 /* update sector only for requests with clear definition of sector */
832 if (!blk_rq_is_passthrough(req
))
833 req
->__sector
+= total_bytes
>> 9;
835 /* mixed attributes always follow the first bio */
836 if (req
->rq_flags
& RQF_MIXED_MERGE
) {
837 req
->cmd_flags
&= ~REQ_FAILFAST_MASK
;
838 req
->cmd_flags
|= req
->bio
->bi_opf
& REQ_FAILFAST_MASK
;
841 if (!(req
->rq_flags
& RQF_SPECIAL_PAYLOAD
)) {
843 * If total number of sectors is less than the first segment
844 * size, something has gone terribly wrong.
846 if (blk_rq_bytes(req
) < blk_rq_cur_bytes(req
)) {
847 blk_dump_rq_flags(req
, "request botched");
848 req
->__data_len
= blk_rq_cur_bytes(req
);
851 /* recalculate the number of segments */
852 req
->nr_phys_segments
= blk_recalc_rq_segments(req
);
857 EXPORT_SYMBOL_GPL(blk_update_request
);
859 static void __blk_account_io_done(struct request
*req
, u64 now
)
861 const int sgrp
= op_stat_group(req_op(req
));
864 update_io_ticks(req
->part
, jiffies
, true);
865 part_stat_inc(req
->part
, ios
[sgrp
]);
866 part_stat_add(req
->part
, nsecs
[sgrp
], now
- req
->start_time_ns
);
870 static inline void blk_account_io_done(struct request
*req
, u64 now
)
873 * Account IO completion. flush_rq isn't accounted as a
874 * normal IO on queueing nor completion. Accounting the
875 * containing request is enough.
877 if (blk_do_io_stat(req
) && req
->part
&&
878 !(req
->rq_flags
& RQF_FLUSH_SEQ
))
879 __blk_account_io_done(req
, now
);
882 static void __blk_account_io_start(struct request
*rq
)
884 /* passthrough requests can hold bios that do not have ->bi_bdev set */
885 if (rq
->bio
&& rq
->bio
->bi_bdev
)
886 rq
->part
= rq
->bio
->bi_bdev
;
887 else if (rq
->q
->disk
)
888 rq
->part
= rq
->q
->disk
->part0
;
891 update_io_ticks(rq
->part
, jiffies
, false);
895 static inline void blk_account_io_start(struct request
*req
)
897 if (blk_do_io_stat(req
))
898 __blk_account_io_start(req
);
901 static inline void __blk_mq_end_request_acct(struct request
*rq
, u64 now
)
903 if (rq
->rq_flags
& RQF_STATS
) {
904 blk_mq_poll_stats_start(rq
->q
);
905 blk_stat_add(rq
, now
);
908 blk_mq_sched_completed_request(rq
, now
);
909 blk_account_io_done(rq
, now
);
912 inline void __blk_mq_end_request(struct request
*rq
, blk_status_t error
)
914 if (blk_mq_need_time_stamp(rq
))
915 __blk_mq_end_request_acct(rq
, ktime_get_ns());
918 rq_qos_done(rq
->q
, rq
);
919 rq
->end_io(rq
, error
);
921 blk_mq_free_request(rq
);
924 EXPORT_SYMBOL(__blk_mq_end_request
);
926 void blk_mq_end_request(struct request
*rq
, blk_status_t error
)
928 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
930 __blk_mq_end_request(rq
, error
);
932 EXPORT_SYMBOL(blk_mq_end_request
);
934 #define TAG_COMP_BATCH 32
936 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx
*hctx
,
937 int *tag_array
, int nr_tags
)
939 struct request_queue
*q
= hctx
->queue
;
942 * All requests should have been marked as RQF_MQ_INFLIGHT, so
943 * update hctx->nr_active in batch
945 if (hctx
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
)
946 __blk_mq_sub_active_requests(hctx
, nr_tags
);
948 blk_mq_put_tags(hctx
->tags
, tag_array
, nr_tags
);
949 percpu_ref_put_many(&q
->q_usage_counter
, nr_tags
);
952 void blk_mq_end_request_batch(struct io_comp_batch
*iob
)
954 int tags
[TAG_COMP_BATCH
], nr_tags
= 0;
955 struct blk_mq_hw_ctx
*cur_hctx
= NULL
;
960 now
= ktime_get_ns();
962 while ((rq
= rq_list_pop(&iob
->req_list
)) != NULL
) {
964 prefetch(rq
->rq_next
);
966 blk_complete_request(rq
);
968 __blk_mq_end_request_acct(rq
, now
);
970 rq_qos_done(rq
->q
, rq
);
972 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
973 if (!req_ref_put_and_test(rq
))
976 blk_crypto_free_request(rq
);
977 blk_pm_mark_last_busy(rq
);
979 if (nr_tags
== TAG_COMP_BATCH
|| cur_hctx
!= rq
->mq_hctx
) {
981 blk_mq_flush_tag_batch(cur_hctx
, tags
, nr_tags
);
983 cur_hctx
= rq
->mq_hctx
;
985 tags
[nr_tags
++] = rq
->tag
;
989 blk_mq_flush_tag_batch(cur_hctx
, tags
, nr_tags
);
991 EXPORT_SYMBOL_GPL(blk_mq_end_request_batch
);
993 static void blk_complete_reqs(struct llist_head
*list
)
995 struct llist_node
*entry
= llist_reverse_order(llist_del_all(list
));
996 struct request
*rq
, *next
;
998 llist_for_each_entry_safe(rq
, next
, entry
, ipi_list
)
999 rq
->q
->mq_ops
->complete(rq
);
1002 static __latent_entropy
void blk_done_softirq(struct softirq_action
*h
)
1004 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done
));
1007 static int blk_softirq_cpu_dead(unsigned int cpu
)
1009 blk_complete_reqs(&per_cpu(blk_cpu_done
, cpu
));
1013 static void __blk_mq_complete_request_remote(void *data
)
1015 __raise_softirq_irqoff(BLOCK_SOFTIRQ
);
1018 static inline bool blk_mq_complete_need_ipi(struct request
*rq
)
1020 int cpu
= raw_smp_processor_id();
1022 if (!IS_ENABLED(CONFIG_SMP
) ||
1023 !test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
))
1026 * With force threaded interrupts enabled, raising softirq from an SMP
1027 * function call will always result in waking the ksoftirqd thread.
1028 * This is probably worse than completing the request on a different
1031 if (force_irqthreads())
1034 /* same CPU or cache domain? Complete locally */
1035 if (cpu
== rq
->mq_ctx
->cpu
||
1036 (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
) &&
1037 cpus_share_cache(cpu
, rq
->mq_ctx
->cpu
)))
1040 /* don't try to IPI to an offline CPU */
1041 return cpu_online(rq
->mq_ctx
->cpu
);
1044 static void blk_mq_complete_send_ipi(struct request
*rq
)
1046 struct llist_head
*list
;
1049 cpu
= rq
->mq_ctx
->cpu
;
1050 list
= &per_cpu(blk_cpu_done
, cpu
);
1051 if (llist_add(&rq
->ipi_list
, list
)) {
1052 INIT_CSD(&rq
->csd
, __blk_mq_complete_request_remote
, rq
);
1053 smp_call_function_single_async(cpu
, &rq
->csd
);
1057 static void blk_mq_raise_softirq(struct request
*rq
)
1059 struct llist_head
*list
;
1062 list
= this_cpu_ptr(&blk_cpu_done
);
1063 if (llist_add(&rq
->ipi_list
, list
))
1064 raise_softirq(BLOCK_SOFTIRQ
);
1068 bool blk_mq_complete_request_remote(struct request
*rq
)
1070 WRITE_ONCE(rq
->state
, MQ_RQ_COMPLETE
);
1073 * For a polled request, always complete locallly, it's pointless
1074 * to redirect the completion.
1076 if (rq
->cmd_flags
& REQ_POLLED
)
1079 if (blk_mq_complete_need_ipi(rq
)) {
1080 blk_mq_complete_send_ipi(rq
);
1084 if (rq
->q
->nr_hw_queues
== 1) {
1085 blk_mq_raise_softirq(rq
);
1090 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote
);
1093 * blk_mq_complete_request - end I/O on a request
1094 * @rq: the request being processed
1097 * Complete a request by scheduling the ->complete_rq operation.
1099 void blk_mq_complete_request(struct request
*rq
)
1101 if (!blk_mq_complete_request_remote(rq
))
1102 rq
->q
->mq_ops
->complete(rq
);
1104 EXPORT_SYMBOL(blk_mq_complete_request
);
1107 * blk_mq_start_request - Start processing a request
1108 * @rq: Pointer to request to be started
1110 * Function used by device drivers to notify the block layer that a request
1111 * is going to be processed now, so blk layer can do proper initializations
1112 * such as starting the timeout timer.
1114 void blk_mq_start_request(struct request
*rq
)
1116 struct request_queue
*q
= rq
->q
;
1118 trace_block_rq_issue(rq
);
1120 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
1122 #ifdef CONFIG_BLK_CGROUP
1124 start_time
= bio_issue_time(&rq
->bio
->bi_issue
);
1127 start_time
= ktime_get_ns();
1128 rq
->io_start_time_ns
= start_time
;
1129 rq
->stats_sectors
= blk_rq_sectors(rq
);
1130 rq
->rq_flags
|= RQF_STATS
;
1131 rq_qos_issue(q
, rq
);
1134 WARN_ON_ONCE(blk_mq_rq_state(rq
) != MQ_RQ_IDLE
);
1137 WRITE_ONCE(rq
->state
, MQ_RQ_IN_FLIGHT
);
1139 #ifdef CONFIG_BLK_DEV_INTEGRITY
1140 if (blk_integrity_rq(rq
) && req_op(rq
) == REQ_OP_WRITE
)
1141 q
->integrity
.profile
->prepare_fn(rq
);
1143 if (rq
->bio
&& rq
->bio
->bi_opf
& REQ_POLLED
)
1144 WRITE_ONCE(rq
->bio
->bi_cookie
, blk_rq_to_qc(rq
));
1146 EXPORT_SYMBOL(blk_mq_start_request
);
1149 * blk_end_sync_rq - executes a completion event on a request
1150 * @rq: request to complete
1151 * @error: end I/O status of the request
1153 static void blk_end_sync_rq(struct request
*rq
, blk_status_t error
)
1155 struct completion
*waiting
= rq
->end_io_data
;
1157 rq
->end_io_data
= (void *)(uintptr_t)error
;
1160 * complete last, if this is a stack request the process (and thus
1161 * the rq pointer) could be invalid right after this complete()
1167 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1168 * @rq: request to insert
1169 * @at_head: insert request at head or tail of queue
1170 * @done: I/O completion handler
1173 * Insert a fully prepared request at the back of the I/O scheduler queue
1174 * for execution. Don't wait for completion.
1177 * This function will invoke @done directly if the queue is dead.
1179 void blk_execute_rq_nowait(struct request
*rq
, bool at_head
, rq_end_io_fn
*done
)
1181 WARN_ON(irqs_disabled());
1182 WARN_ON(!blk_rq_is_passthrough(rq
));
1186 blk_account_io_start(rq
);
1189 * don't check dying flag for MQ because the request won't
1190 * be reused after dying flag is set
1192 blk_mq_sched_insert_request(rq
, at_head
, true, false);
1194 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait
);
1196 static bool blk_rq_is_poll(struct request
*rq
)
1200 if (rq
->mq_hctx
->type
!= HCTX_TYPE_POLL
)
1202 if (WARN_ON_ONCE(!rq
->bio
))
1207 static void blk_rq_poll_completion(struct request
*rq
, struct completion
*wait
)
1210 bio_poll(rq
->bio
, NULL
, 0);
1212 } while (!completion_done(wait
));
1216 * blk_execute_rq - insert a request into queue for execution
1217 * @rq: request to insert
1218 * @at_head: insert request at head or tail of queue
1221 * Insert a fully prepared request at the back of the I/O scheduler queue
1222 * for execution and wait for completion.
1223 * Return: The blk_status_t result provided to blk_mq_end_request().
1225 blk_status_t
blk_execute_rq(struct request
*rq
, bool at_head
)
1227 DECLARE_COMPLETION_ONSTACK(wait
);
1228 unsigned long hang_check
;
1230 rq
->end_io_data
= &wait
;
1231 blk_execute_rq_nowait(rq
, at_head
, blk_end_sync_rq
);
1233 /* Prevent hang_check timer from firing at us during very long I/O */
1234 hang_check
= sysctl_hung_task_timeout_secs
;
1236 if (blk_rq_is_poll(rq
))
1237 blk_rq_poll_completion(rq
, &wait
);
1238 else if (hang_check
)
1239 while (!wait_for_completion_io_timeout(&wait
,
1240 hang_check
* (HZ
/2)))
1243 wait_for_completion_io(&wait
);
1245 return (blk_status_t
)(uintptr_t)rq
->end_io_data
;
1247 EXPORT_SYMBOL(blk_execute_rq
);
1249 static void __blk_mq_requeue_request(struct request
*rq
)
1251 struct request_queue
*q
= rq
->q
;
1253 blk_mq_put_driver_tag(rq
);
1255 trace_block_rq_requeue(rq
);
1256 rq_qos_requeue(q
, rq
);
1258 if (blk_mq_request_started(rq
)) {
1259 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
1260 rq
->rq_flags
&= ~RQF_TIMED_OUT
;
1264 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
1266 __blk_mq_requeue_request(rq
);
1268 /* this request will be re-inserted to io scheduler queue */
1269 blk_mq_sched_requeue_request(rq
);
1271 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
1273 EXPORT_SYMBOL(blk_mq_requeue_request
);
1275 static void blk_mq_requeue_work(struct work_struct
*work
)
1277 struct request_queue
*q
=
1278 container_of(work
, struct request_queue
, requeue_work
.work
);
1280 struct request
*rq
, *next
;
1282 spin_lock_irq(&q
->requeue_lock
);
1283 list_splice_init(&q
->requeue_list
, &rq_list
);
1284 spin_unlock_irq(&q
->requeue_lock
);
1286 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
1287 if (!(rq
->rq_flags
& (RQF_SOFTBARRIER
| RQF_DONTPREP
)))
1290 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
1291 list_del_init(&rq
->queuelist
);
1293 * If RQF_DONTPREP, rq has contained some driver specific
1294 * data, so insert it to hctx dispatch list to avoid any
1297 if (rq
->rq_flags
& RQF_DONTPREP
)
1298 blk_mq_request_bypass_insert(rq
, false, false);
1300 blk_mq_sched_insert_request(rq
, true, false, false);
1303 while (!list_empty(&rq_list
)) {
1304 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
1305 list_del_init(&rq
->queuelist
);
1306 blk_mq_sched_insert_request(rq
, false, false, false);
1309 blk_mq_run_hw_queues(q
, false);
1312 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
1313 bool kick_requeue_list
)
1315 struct request_queue
*q
= rq
->q
;
1316 unsigned long flags
;
1319 * We abuse this flag that is otherwise used by the I/O scheduler to
1320 * request head insertion from the workqueue.
1322 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
1324 spin_lock_irqsave(&q
->requeue_lock
, flags
);
1326 rq
->rq_flags
|= RQF_SOFTBARRIER
;
1327 list_add(&rq
->queuelist
, &q
->requeue_list
);
1329 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
1331 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
1333 if (kick_requeue_list
)
1334 blk_mq_kick_requeue_list(q
);
1337 void blk_mq_kick_requeue_list(struct request_queue
*q
)
1339 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
, 0);
1341 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
1343 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
1344 unsigned long msecs
)
1346 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
,
1347 msecs_to_jiffies(msecs
));
1349 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
1351 static bool blk_mq_rq_inflight(struct request
*rq
, void *priv
,
1355 * If we find a request that isn't idle we know the queue is busy
1356 * as it's checked in the iter.
1357 * Return false to stop the iteration.
1359 if (blk_mq_request_started(rq
)) {
1369 bool blk_mq_queue_inflight(struct request_queue
*q
)
1373 blk_mq_queue_tag_busy_iter(q
, blk_mq_rq_inflight
, &busy
);
1376 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight
);
1378 static void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
1380 req
->rq_flags
|= RQF_TIMED_OUT
;
1381 if (req
->q
->mq_ops
->timeout
) {
1382 enum blk_eh_timer_return ret
;
1384 ret
= req
->q
->mq_ops
->timeout(req
, reserved
);
1385 if (ret
== BLK_EH_DONE
)
1387 WARN_ON_ONCE(ret
!= BLK_EH_RESET_TIMER
);
1393 static bool blk_mq_req_expired(struct request
*rq
, unsigned long *next
)
1395 unsigned long deadline
;
1397 if (blk_mq_rq_state(rq
) != MQ_RQ_IN_FLIGHT
)
1399 if (rq
->rq_flags
& RQF_TIMED_OUT
)
1402 deadline
= READ_ONCE(rq
->deadline
);
1403 if (time_after_eq(jiffies
, deadline
))
1408 else if (time_after(*next
, deadline
))
1413 void blk_mq_put_rq_ref(struct request
*rq
)
1415 if (is_flush_rq(rq
))
1417 else if (req_ref_put_and_test(rq
))
1418 __blk_mq_free_request(rq
);
1421 static bool blk_mq_check_expired(struct request
*rq
, void *priv
, bool reserved
)
1423 unsigned long *next
= priv
;
1426 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1427 * be reallocated underneath the timeout handler's processing, then
1428 * the expire check is reliable. If the request is not expired, then
1429 * it was completed and reallocated as a new request after returning
1430 * from blk_mq_check_expired().
1432 if (blk_mq_req_expired(rq
, next
))
1433 blk_mq_rq_timed_out(rq
, reserved
);
1437 static void blk_mq_timeout_work(struct work_struct
*work
)
1439 struct request_queue
*q
=
1440 container_of(work
, struct request_queue
, timeout_work
);
1441 unsigned long next
= 0;
1442 struct blk_mq_hw_ctx
*hctx
;
1445 /* A deadlock might occur if a request is stuck requiring a
1446 * timeout at the same time a queue freeze is waiting
1447 * completion, since the timeout code would not be able to
1448 * acquire the queue reference here.
1450 * That's why we don't use blk_queue_enter here; instead, we use
1451 * percpu_ref_tryget directly, because we need to be able to
1452 * obtain a reference even in the short window between the queue
1453 * starting to freeze, by dropping the first reference in
1454 * blk_freeze_queue_start, and the moment the last request is
1455 * consumed, marked by the instant q_usage_counter reaches
1458 if (!percpu_ref_tryget(&q
->q_usage_counter
))
1461 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &next
);
1464 mod_timer(&q
->timeout
, next
);
1467 * Request timeouts are handled as a forward rolling timer. If
1468 * we end up here it means that no requests are pending and
1469 * also that no request has been pending for a while. Mark
1470 * each hctx as idle.
1472 queue_for_each_hw_ctx(q
, hctx
, i
) {
1473 /* the hctx may be unmapped, so check it here */
1474 if (blk_mq_hw_queue_mapped(hctx
))
1475 blk_mq_tag_idle(hctx
);
1481 struct flush_busy_ctx_data
{
1482 struct blk_mq_hw_ctx
*hctx
;
1483 struct list_head
*list
;
1486 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
1488 struct flush_busy_ctx_data
*flush_data
= data
;
1489 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
1490 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
1491 enum hctx_type type
= hctx
->type
;
1493 spin_lock(&ctx
->lock
);
1494 list_splice_tail_init(&ctx
->rq_lists
[type
], flush_data
->list
);
1495 sbitmap_clear_bit(sb
, bitnr
);
1496 spin_unlock(&ctx
->lock
);
1501 * Process software queues that have been marked busy, splicing them
1502 * to the for-dispatch
1504 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
1506 struct flush_busy_ctx_data data
= {
1511 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
1513 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
1515 struct dispatch_rq_data
{
1516 struct blk_mq_hw_ctx
*hctx
;
1520 static bool dispatch_rq_from_ctx(struct sbitmap
*sb
, unsigned int bitnr
,
1523 struct dispatch_rq_data
*dispatch_data
= data
;
1524 struct blk_mq_hw_ctx
*hctx
= dispatch_data
->hctx
;
1525 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
1526 enum hctx_type type
= hctx
->type
;
1528 spin_lock(&ctx
->lock
);
1529 if (!list_empty(&ctx
->rq_lists
[type
])) {
1530 dispatch_data
->rq
= list_entry_rq(ctx
->rq_lists
[type
].next
);
1531 list_del_init(&dispatch_data
->rq
->queuelist
);
1532 if (list_empty(&ctx
->rq_lists
[type
]))
1533 sbitmap_clear_bit(sb
, bitnr
);
1535 spin_unlock(&ctx
->lock
);
1537 return !dispatch_data
->rq
;
1540 struct request
*blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx
*hctx
,
1541 struct blk_mq_ctx
*start
)
1543 unsigned off
= start
? start
->index_hw
[hctx
->type
] : 0;
1544 struct dispatch_rq_data data
= {
1549 __sbitmap_for_each_set(&hctx
->ctx_map
, off
,
1550 dispatch_rq_from_ctx
, &data
);
1555 static bool __blk_mq_alloc_driver_tag(struct request
*rq
)
1557 struct sbitmap_queue
*bt
= &rq
->mq_hctx
->tags
->bitmap_tags
;
1558 unsigned int tag_offset
= rq
->mq_hctx
->tags
->nr_reserved_tags
;
1561 blk_mq_tag_busy(rq
->mq_hctx
);
1563 if (blk_mq_tag_is_reserved(rq
->mq_hctx
->sched_tags
, rq
->internal_tag
)) {
1564 bt
= &rq
->mq_hctx
->tags
->breserved_tags
;
1567 if (!hctx_may_queue(rq
->mq_hctx
, bt
))
1571 tag
= __sbitmap_queue_get(bt
);
1572 if (tag
== BLK_MQ_NO_TAG
)
1575 rq
->tag
= tag
+ tag_offset
;
1579 bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
1581 if (rq
->tag
== BLK_MQ_NO_TAG
&& !__blk_mq_alloc_driver_tag(rq
))
1584 if ((hctx
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
) &&
1585 !(rq
->rq_flags
& RQF_MQ_INFLIGHT
)) {
1586 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
1587 __blk_mq_inc_active_requests(hctx
);
1589 hctx
->tags
->rqs
[rq
->tag
] = rq
;
1593 static int blk_mq_dispatch_wake(wait_queue_entry_t
*wait
, unsigned mode
,
1594 int flags
, void *key
)
1596 struct blk_mq_hw_ctx
*hctx
;
1598 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
1600 spin_lock(&hctx
->dispatch_wait_lock
);
1601 if (!list_empty(&wait
->entry
)) {
1602 struct sbitmap_queue
*sbq
;
1604 list_del_init(&wait
->entry
);
1605 sbq
= &hctx
->tags
->bitmap_tags
;
1606 atomic_dec(&sbq
->ws_active
);
1608 spin_unlock(&hctx
->dispatch_wait_lock
);
1610 blk_mq_run_hw_queue(hctx
, true);
1615 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1616 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1617 * restart. For both cases, take care to check the condition again after
1618 * marking us as waiting.
1620 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx
*hctx
,
1623 struct sbitmap_queue
*sbq
= &hctx
->tags
->bitmap_tags
;
1624 struct wait_queue_head
*wq
;
1625 wait_queue_entry_t
*wait
;
1628 if (!(hctx
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
)) {
1629 blk_mq_sched_mark_restart_hctx(hctx
);
1632 * It's possible that a tag was freed in the window between the
1633 * allocation failure and adding the hardware queue to the wait
1636 * Don't clear RESTART here, someone else could have set it.
1637 * At most this will cost an extra queue run.
1639 return blk_mq_get_driver_tag(rq
);
1642 wait
= &hctx
->dispatch_wait
;
1643 if (!list_empty_careful(&wait
->entry
))
1646 wq
= &bt_wait_ptr(sbq
, hctx
)->wait
;
1648 spin_lock_irq(&wq
->lock
);
1649 spin_lock(&hctx
->dispatch_wait_lock
);
1650 if (!list_empty(&wait
->entry
)) {
1651 spin_unlock(&hctx
->dispatch_wait_lock
);
1652 spin_unlock_irq(&wq
->lock
);
1656 atomic_inc(&sbq
->ws_active
);
1657 wait
->flags
&= ~WQ_FLAG_EXCLUSIVE
;
1658 __add_wait_queue(wq
, wait
);
1661 * It's possible that a tag was freed in the window between the
1662 * allocation failure and adding the hardware queue to the wait
1665 ret
= blk_mq_get_driver_tag(rq
);
1667 spin_unlock(&hctx
->dispatch_wait_lock
);
1668 spin_unlock_irq(&wq
->lock
);
1673 * We got a tag, remove ourselves from the wait queue to ensure
1674 * someone else gets the wakeup.
1676 list_del_init(&wait
->entry
);
1677 atomic_dec(&sbq
->ws_active
);
1678 spin_unlock(&hctx
->dispatch_wait_lock
);
1679 spin_unlock_irq(&wq
->lock
);
1684 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1685 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1687 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1688 * - EWMA is one simple way to compute running average value
1689 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1690 * - take 4 as factor for avoiding to get too small(0) result, and this
1691 * factor doesn't matter because EWMA decreases exponentially
1693 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx
*hctx
, bool busy
)
1697 ewma
= hctx
->dispatch_busy
;
1702 ewma
*= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
- 1;
1704 ewma
+= 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR
;
1705 ewma
/= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
;
1707 hctx
->dispatch_busy
= ewma
;
1710 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1712 static void blk_mq_handle_dev_resource(struct request
*rq
,
1713 struct list_head
*list
)
1715 struct request
*next
=
1716 list_first_entry_or_null(list
, struct request
, queuelist
);
1719 * If an I/O scheduler has been configured and we got a driver tag for
1720 * the next request already, free it.
1723 blk_mq_put_driver_tag(next
);
1725 list_add(&rq
->queuelist
, list
);
1726 __blk_mq_requeue_request(rq
);
1729 static void blk_mq_handle_zone_resource(struct request
*rq
,
1730 struct list_head
*zone_list
)
1733 * If we end up here it is because we cannot dispatch a request to a
1734 * specific zone due to LLD level zone-write locking or other zone
1735 * related resource not being available. In this case, set the request
1736 * aside in zone_list for retrying it later.
1738 list_add(&rq
->queuelist
, zone_list
);
1739 __blk_mq_requeue_request(rq
);
1742 enum prep_dispatch
{
1744 PREP_DISPATCH_NO_TAG
,
1745 PREP_DISPATCH_NO_BUDGET
,
1748 static enum prep_dispatch
blk_mq_prep_dispatch_rq(struct request
*rq
,
1751 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1752 int budget_token
= -1;
1755 budget_token
= blk_mq_get_dispatch_budget(rq
->q
);
1756 if (budget_token
< 0) {
1757 blk_mq_put_driver_tag(rq
);
1758 return PREP_DISPATCH_NO_BUDGET
;
1760 blk_mq_set_rq_budget_token(rq
, budget_token
);
1763 if (!blk_mq_get_driver_tag(rq
)) {
1765 * The initial allocation attempt failed, so we need to
1766 * rerun the hardware queue when a tag is freed. The
1767 * waitqueue takes care of that. If the queue is run
1768 * before we add this entry back on the dispatch list,
1769 * we'll re-run it below.
1771 if (!blk_mq_mark_tag_wait(hctx
, rq
)) {
1773 * All budgets not got from this function will be put
1774 * together during handling partial dispatch
1777 blk_mq_put_dispatch_budget(rq
->q
, budget_token
);
1778 return PREP_DISPATCH_NO_TAG
;
1782 return PREP_DISPATCH_OK
;
1785 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1786 static void blk_mq_release_budgets(struct request_queue
*q
,
1787 struct list_head
*list
)
1791 list_for_each_entry(rq
, list
, queuelist
) {
1792 int budget_token
= blk_mq_get_rq_budget_token(rq
);
1794 if (budget_token
>= 0)
1795 blk_mq_put_dispatch_budget(q
, budget_token
);
1800 * Returns true if we did some work AND can potentially do more.
1802 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
,
1803 unsigned int nr_budgets
)
1805 enum prep_dispatch prep
;
1806 struct request_queue
*q
= hctx
->queue
;
1807 struct request
*rq
, *nxt
;
1809 blk_status_t ret
= BLK_STS_OK
;
1810 LIST_HEAD(zone_list
);
1811 bool needs_resource
= false;
1813 if (list_empty(list
))
1817 * Now process all the entries, sending them to the driver.
1819 errors
= queued
= 0;
1821 struct blk_mq_queue_data bd
;
1823 rq
= list_first_entry(list
, struct request
, queuelist
);
1825 WARN_ON_ONCE(hctx
!= rq
->mq_hctx
);
1826 prep
= blk_mq_prep_dispatch_rq(rq
, !nr_budgets
);
1827 if (prep
!= PREP_DISPATCH_OK
)
1830 list_del_init(&rq
->queuelist
);
1835 * Flag last if we have no more requests, or if we have more
1836 * but can't assign a driver tag to it.
1838 if (list_empty(list
))
1841 nxt
= list_first_entry(list
, struct request
, queuelist
);
1842 bd
.last
= !blk_mq_get_driver_tag(nxt
);
1846 * once the request is queued to lld, no need to cover the
1851 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1856 case BLK_STS_RESOURCE
:
1857 needs_resource
= true;
1859 case BLK_STS_DEV_RESOURCE
:
1860 blk_mq_handle_dev_resource(rq
, list
);
1862 case BLK_STS_ZONE_RESOURCE
:
1864 * Move the request to zone_list and keep going through
1865 * the dispatch list to find more requests the drive can
1868 blk_mq_handle_zone_resource(rq
, &zone_list
);
1869 needs_resource
= true;
1873 blk_mq_end_request(rq
, ret
);
1875 } while (!list_empty(list
));
1877 if (!list_empty(&zone_list
))
1878 list_splice_tail_init(&zone_list
, list
);
1880 /* If we didn't flush the entire list, we could have told the driver
1881 * there was more coming, but that turned out to be a lie.
1883 if ((!list_empty(list
) || errors
) && q
->mq_ops
->commit_rqs
&& queued
)
1884 q
->mq_ops
->commit_rqs(hctx
);
1886 * Any items that need requeuing? Stuff them into hctx->dispatch,
1887 * that is where we will continue on next queue run.
1889 if (!list_empty(list
)) {
1891 /* For non-shared tags, the RESTART check will suffice */
1892 bool no_tag
= prep
== PREP_DISPATCH_NO_TAG
&&
1893 (hctx
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
);
1896 blk_mq_release_budgets(q
, list
);
1898 spin_lock(&hctx
->lock
);
1899 list_splice_tail_init(list
, &hctx
->dispatch
);
1900 spin_unlock(&hctx
->lock
);
1903 * Order adding requests to hctx->dispatch and checking
1904 * SCHED_RESTART flag. The pair of this smp_mb() is the one
1905 * in blk_mq_sched_restart(). Avoid restart code path to
1906 * miss the new added requests to hctx->dispatch, meantime
1907 * SCHED_RESTART is observed here.
1912 * If SCHED_RESTART was set by the caller of this function and
1913 * it is no longer set that means that it was cleared by another
1914 * thread and hence that a queue rerun is needed.
1916 * If 'no_tag' is set, that means that we failed getting
1917 * a driver tag with an I/O scheduler attached. If our dispatch
1918 * waitqueue is no longer active, ensure that we run the queue
1919 * AFTER adding our entries back to the list.
1921 * If no I/O scheduler has been configured it is possible that
1922 * the hardware queue got stopped and restarted before requests
1923 * were pushed back onto the dispatch list. Rerun the queue to
1924 * avoid starvation. Notes:
1925 * - blk_mq_run_hw_queue() checks whether or not a queue has
1926 * been stopped before rerunning a queue.
1927 * - Some but not all block drivers stop a queue before
1928 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1931 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1932 * bit is set, run queue after a delay to avoid IO stalls
1933 * that could otherwise occur if the queue is idle. We'll do
1934 * similar if we couldn't get budget or couldn't lock a zone
1935 * and SCHED_RESTART is set.
1937 needs_restart
= blk_mq_sched_needs_restart(hctx
);
1938 if (prep
== PREP_DISPATCH_NO_BUDGET
)
1939 needs_resource
= true;
1940 if (!needs_restart
||
1941 (no_tag
&& list_empty_careful(&hctx
->dispatch_wait
.entry
)))
1942 blk_mq_run_hw_queue(hctx
, true);
1943 else if (needs_restart
&& needs_resource
)
1944 blk_mq_delay_run_hw_queue(hctx
, BLK_MQ_RESOURCE_DELAY
);
1946 blk_mq_update_dispatch_busy(hctx
, true);
1949 blk_mq_update_dispatch_busy(hctx
, false);
1951 return (queued
+ errors
) != 0;
1955 * __blk_mq_run_hw_queue - Run a hardware queue.
1956 * @hctx: Pointer to the hardware queue to run.
1958 * Send pending requests to the hardware.
1960 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1963 * We can't run the queue inline with ints disabled. Ensure that
1964 * we catch bad users of this early.
1966 WARN_ON_ONCE(in_interrupt());
1968 blk_mq_run_dispatch_ops(hctx
->queue
,
1969 blk_mq_sched_dispatch_requests(hctx
));
1972 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx
*hctx
)
1974 int cpu
= cpumask_first_and(hctx
->cpumask
, cpu_online_mask
);
1976 if (cpu
>= nr_cpu_ids
)
1977 cpu
= cpumask_first(hctx
->cpumask
);
1982 * It'd be great if the workqueue API had a way to pass
1983 * in a mask and had some smarts for more clever placement.
1984 * For now we just round-robin here, switching for every
1985 * BLK_MQ_CPU_WORK_BATCH queued items.
1987 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1990 int next_cpu
= hctx
->next_cpu
;
1992 if (hctx
->queue
->nr_hw_queues
== 1)
1993 return WORK_CPU_UNBOUND
;
1995 if (--hctx
->next_cpu_batch
<= 0) {
1997 next_cpu
= cpumask_next_and(next_cpu
, hctx
->cpumask
,
1999 if (next_cpu
>= nr_cpu_ids
)
2000 next_cpu
= blk_mq_first_mapped_cpu(hctx
);
2001 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2005 * Do unbound schedule if we can't find a online CPU for this hctx,
2006 * and it should only happen in the path of handling CPU DEAD.
2008 if (!cpu_online(next_cpu
)) {
2015 * Make sure to re-select CPU next time once after CPUs
2016 * in hctx->cpumask become online again.
2018 hctx
->next_cpu
= next_cpu
;
2019 hctx
->next_cpu_batch
= 1;
2020 return WORK_CPU_UNBOUND
;
2023 hctx
->next_cpu
= next_cpu
;
2028 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
2029 * @hctx: Pointer to the hardware queue to run.
2030 * @async: If we want to run the queue asynchronously.
2031 * @msecs: Milliseconds of delay to wait before running the queue.
2033 * If !@async, try to run the queue now. Else, run the queue asynchronously and
2034 * with a delay of @msecs.
2036 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
,
2037 unsigned long msecs
)
2039 if (unlikely(blk_mq_hctx_stopped(hctx
)))
2042 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
2043 int cpu
= get_cpu();
2044 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
2045 __blk_mq_run_hw_queue(hctx
);
2053 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
,
2054 msecs_to_jiffies(msecs
));
2058 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2059 * @hctx: Pointer to the hardware queue to run.
2060 * @msecs: Milliseconds of delay to wait before running the queue.
2062 * Run a hardware queue asynchronously with a delay of @msecs.
2064 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
2066 __blk_mq_delay_run_hw_queue(hctx
, true, msecs
);
2068 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue
);
2071 * blk_mq_run_hw_queue - Start to run a hardware queue.
2072 * @hctx: Pointer to the hardware queue to run.
2073 * @async: If we want to run the queue asynchronously.
2075 * Check if the request queue is not in a quiesced state and if there are
2076 * pending requests to be sent. If this is true, run the queue to send requests
2079 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
2084 * When queue is quiesced, we may be switching io scheduler, or
2085 * updating nr_hw_queues, or other things, and we can't run queue
2086 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2088 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2091 __blk_mq_run_dispatch_ops(hctx
->queue
, false,
2092 need_run
= !blk_queue_quiesced(hctx
->queue
) &&
2093 blk_mq_hctx_has_pending(hctx
));
2096 __blk_mq_delay_run_hw_queue(hctx
, async
, 0);
2098 EXPORT_SYMBOL(blk_mq_run_hw_queue
);
2101 * Is the request queue handled by an IO scheduler that does not respect
2102 * hardware queues when dispatching?
2104 static bool blk_mq_has_sqsched(struct request_queue
*q
)
2106 struct elevator_queue
*e
= q
->elevator
;
2108 if (e
&& e
->type
->ops
.dispatch_request
&&
2109 !(e
->type
->elevator_features
& ELEVATOR_F_MQ_AWARE
))
2115 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2118 static struct blk_mq_hw_ctx
*blk_mq_get_sq_hctx(struct request_queue
*q
)
2120 struct blk_mq_hw_ctx
*hctx
;
2123 * If the IO scheduler does not respect hardware queues when
2124 * dispatching, we just don't bother with multiple HW queues and
2125 * dispatch from hctx for the current CPU since running multiple queues
2126 * just causes lock contention inside the scheduler and pointless cache
2129 hctx
= blk_mq_map_queue_type(q
, HCTX_TYPE_DEFAULT
,
2130 raw_smp_processor_id());
2131 if (!blk_mq_hctx_stopped(hctx
))
2137 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2138 * @q: Pointer to the request queue to run.
2139 * @async: If we want to run the queue asynchronously.
2141 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
2143 struct blk_mq_hw_ctx
*hctx
, *sq_hctx
;
2147 if (blk_mq_has_sqsched(q
))
2148 sq_hctx
= blk_mq_get_sq_hctx(q
);
2149 queue_for_each_hw_ctx(q
, hctx
, i
) {
2150 if (blk_mq_hctx_stopped(hctx
))
2153 * Dispatch from this hctx either if there's no hctx preferred
2154 * by IO scheduler or if it has requests that bypass the
2157 if (!sq_hctx
|| sq_hctx
== hctx
||
2158 !list_empty_careful(&hctx
->dispatch
))
2159 blk_mq_run_hw_queue(hctx
, async
);
2162 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
2165 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2166 * @q: Pointer to the request queue to run.
2167 * @msecs: Milliseconds of delay to wait before running the queues.
2169 void blk_mq_delay_run_hw_queues(struct request_queue
*q
, unsigned long msecs
)
2171 struct blk_mq_hw_ctx
*hctx
, *sq_hctx
;
2175 if (blk_mq_has_sqsched(q
))
2176 sq_hctx
= blk_mq_get_sq_hctx(q
);
2177 queue_for_each_hw_ctx(q
, hctx
, i
) {
2178 if (blk_mq_hctx_stopped(hctx
))
2181 * Dispatch from this hctx either if there's no hctx preferred
2182 * by IO scheduler or if it has requests that bypass the
2185 if (!sq_hctx
|| sq_hctx
== hctx
||
2186 !list_empty_careful(&hctx
->dispatch
))
2187 blk_mq_delay_run_hw_queue(hctx
, msecs
);
2190 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues
);
2193 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
2194 * @q: request queue.
2196 * The caller is responsible for serializing this function against
2197 * blk_mq_{start,stop}_hw_queue().
2199 bool blk_mq_queue_stopped(struct request_queue
*q
)
2201 struct blk_mq_hw_ctx
*hctx
;
2204 queue_for_each_hw_ctx(q
, hctx
, i
)
2205 if (blk_mq_hctx_stopped(hctx
))
2210 EXPORT_SYMBOL(blk_mq_queue_stopped
);
2213 * This function is often used for pausing .queue_rq() by driver when
2214 * there isn't enough resource or some conditions aren't satisfied, and
2215 * BLK_STS_RESOURCE is usually returned.
2217 * We do not guarantee that dispatch can be drained or blocked
2218 * after blk_mq_stop_hw_queue() returns. Please use
2219 * blk_mq_quiesce_queue() for that requirement.
2221 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
2223 cancel_delayed_work(&hctx
->run_work
);
2225 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
2227 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
2230 * This function is often used for pausing .queue_rq() by driver when
2231 * there isn't enough resource or some conditions aren't satisfied, and
2232 * BLK_STS_RESOURCE is usually returned.
2234 * We do not guarantee that dispatch can be drained or blocked
2235 * after blk_mq_stop_hw_queues() returns. Please use
2236 * blk_mq_quiesce_queue() for that requirement.
2238 void blk_mq_stop_hw_queues(struct request_queue
*q
)
2240 struct blk_mq_hw_ctx
*hctx
;
2243 queue_for_each_hw_ctx(q
, hctx
, i
)
2244 blk_mq_stop_hw_queue(hctx
);
2246 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
2248 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
2250 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
2252 blk_mq_run_hw_queue(hctx
, false);
2254 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
2256 void blk_mq_start_hw_queues(struct request_queue
*q
)
2258 struct blk_mq_hw_ctx
*hctx
;
2261 queue_for_each_hw_ctx(q
, hctx
, i
)
2262 blk_mq_start_hw_queue(hctx
);
2264 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
2266 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
2268 if (!blk_mq_hctx_stopped(hctx
))
2271 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
2272 blk_mq_run_hw_queue(hctx
, async
);
2274 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
2276 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
2278 struct blk_mq_hw_ctx
*hctx
;
2281 queue_for_each_hw_ctx(q
, hctx
, i
)
2282 blk_mq_start_stopped_hw_queue(hctx
, async
);
2284 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
2286 static void blk_mq_run_work_fn(struct work_struct
*work
)
2288 struct blk_mq_hw_ctx
*hctx
;
2290 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
2293 * If we are stopped, don't run the queue.
2295 if (blk_mq_hctx_stopped(hctx
))
2298 __blk_mq_run_hw_queue(hctx
);
2301 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
2305 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
2306 enum hctx_type type
= hctx
->type
;
2308 lockdep_assert_held(&ctx
->lock
);
2310 trace_block_rq_insert(rq
);
2313 list_add(&rq
->queuelist
, &ctx
->rq_lists
[type
]);
2315 list_add_tail(&rq
->queuelist
, &ctx
->rq_lists
[type
]);
2318 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
2321 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
2323 lockdep_assert_held(&ctx
->lock
);
2325 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
2326 blk_mq_hctx_mark_pending(hctx
, ctx
);
2330 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2331 * @rq: Pointer to request to be inserted.
2332 * @at_head: true if the request should be inserted at the head of the list.
2333 * @run_queue: If we should run the hardware queue after inserting the request.
2335 * Should only be used carefully, when the caller knows we want to
2336 * bypass a potential IO scheduler on the target device.
2338 void blk_mq_request_bypass_insert(struct request
*rq
, bool at_head
,
2341 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
2343 spin_lock(&hctx
->lock
);
2345 list_add(&rq
->queuelist
, &hctx
->dispatch
);
2347 list_add_tail(&rq
->queuelist
, &hctx
->dispatch
);
2348 spin_unlock(&hctx
->lock
);
2351 blk_mq_run_hw_queue(hctx
, false);
2354 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
2355 struct list_head
*list
)
2359 enum hctx_type type
= hctx
->type
;
2362 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2365 list_for_each_entry(rq
, list
, queuelist
) {
2366 BUG_ON(rq
->mq_ctx
!= ctx
);
2367 trace_block_rq_insert(rq
);
2370 spin_lock(&ctx
->lock
);
2371 list_splice_tail_init(list
, &ctx
->rq_lists
[type
]);
2372 blk_mq_hctx_mark_pending(hctx
, ctx
);
2373 spin_unlock(&ctx
->lock
);
2376 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx
*hctx
, int *queued
,
2379 if (hctx
->queue
->mq_ops
->commit_rqs
) {
2380 trace_block_unplug(hctx
->queue
, *queued
, !from_schedule
);
2381 hctx
->queue
->mq_ops
->commit_rqs(hctx
);
2386 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
,
2387 unsigned int nr_segs
)
2391 if (bio
->bi_opf
& REQ_RAHEAD
)
2392 rq
->cmd_flags
|= REQ_FAILFAST_MASK
;
2394 rq
->__sector
= bio
->bi_iter
.bi_sector
;
2395 rq
->write_hint
= bio
->bi_write_hint
;
2396 blk_rq_bio_prep(rq
, bio
, nr_segs
);
2398 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2399 err
= blk_crypto_rq_bio_prep(rq
, bio
, GFP_NOIO
);
2402 blk_account_io_start(rq
);
2405 static blk_status_t
__blk_mq_issue_directly(struct blk_mq_hw_ctx
*hctx
,
2406 struct request
*rq
, bool last
)
2408 struct request_queue
*q
= rq
->q
;
2409 struct blk_mq_queue_data bd
= {
2416 * For OK queue, we are done. For error, caller may kill it.
2417 * Any other error (busy), just add it to our list as we
2418 * previously would have done.
2420 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
2423 blk_mq_update_dispatch_busy(hctx
, false);
2425 case BLK_STS_RESOURCE
:
2426 case BLK_STS_DEV_RESOURCE
:
2427 blk_mq_update_dispatch_busy(hctx
, true);
2428 __blk_mq_requeue_request(rq
);
2431 blk_mq_update_dispatch_busy(hctx
, false);
2438 static blk_status_t
__blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
2440 bool bypass_insert
, bool last
)
2442 struct request_queue
*q
= rq
->q
;
2443 bool run_queue
= true;
2447 * RCU or SRCU read lock is needed before checking quiesced flag.
2449 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2450 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2451 * and avoid driver to try to dispatch again.
2453 if (blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(q
)) {
2455 bypass_insert
= false;
2459 if ((rq
->rq_flags
& RQF_ELV
) && !bypass_insert
)
2462 budget_token
= blk_mq_get_dispatch_budget(q
);
2463 if (budget_token
< 0)
2466 blk_mq_set_rq_budget_token(rq
, budget_token
);
2468 if (!blk_mq_get_driver_tag(rq
)) {
2469 blk_mq_put_dispatch_budget(q
, budget_token
);
2473 return __blk_mq_issue_directly(hctx
, rq
, last
);
2476 return BLK_STS_RESOURCE
;
2478 blk_mq_sched_insert_request(rq
, false, run_queue
, false);
2484 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2485 * @hctx: Pointer of the associated hardware queue.
2486 * @rq: Pointer to request to be sent.
2488 * If the device has enough resources to accept a new request now, send the
2489 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2490 * we can try send it another time in the future. Requests inserted at this
2491 * queue have higher priority.
2493 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
2497 __blk_mq_try_issue_directly(hctx
, rq
, false, true);
2499 if (ret
== BLK_STS_RESOURCE
|| ret
== BLK_STS_DEV_RESOURCE
)
2500 blk_mq_request_bypass_insert(rq
, false, true);
2501 else if (ret
!= BLK_STS_OK
)
2502 blk_mq_end_request(rq
, ret
);
2505 static blk_status_t
blk_mq_request_issue_directly(struct request
*rq
, bool last
)
2507 return __blk_mq_try_issue_directly(rq
->mq_hctx
, rq
, true, last
);
2510 static void blk_mq_plug_issue_direct(struct blk_plug
*plug
, bool from_schedule
)
2512 struct blk_mq_hw_ctx
*hctx
= NULL
;
2517 while ((rq
= rq_list_pop(&plug
->mq_list
))) {
2518 bool last
= rq_list_empty(plug
->mq_list
);
2521 if (hctx
!= rq
->mq_hctx
) {
2523 blk_mq_commit_rqs(hctx
, &queued
, from_schedule
);
2527 ret
= blk_mq_request_issue_directly(rq
, last
);
2532 case BLK_STS_RESOURCE
:
2533 case BLK_STS_DEV_RESOURCE
:
2534 blk_mq_request_bypass_insert(rq
, false, last
);
2535 blk_mq_commit_rqs(hctx
, &queued
, from_schedule
);
2538 blk_mq_end_request(rq
, ret
);
2545 * If we didn't flush the entire list, we could have told the driver
2546 * there was more coming, but that turned out to be a lie.
2549 blk_mq_commit_rqs(hctx
, &queued
, from_schedule
);
2552 static void __blk_mq_flush_plug_list(struct request_queue
*q
,
2553 struct blk_plug
*plug
)
2555 if (blk_queue_quiesced(q
))
2557 q
->mq_ops
->queue_rqs(&plug
->mq_list
);
2560 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
2562 struct blk_mq_hw_ctx
*this_hctx
;
2563 struct blk_mq_ctx
*this_ctx
;
2568 if (rq_list_empty(plug
->mq_list
))
2572 if (!plug
->multiple_queues
&& !plug
->has_elevator
&& !from_schedule
) {
2573 struct request_queue
*q
;
2575 rq
= rq_list_peek(&plug
->mq_list
);
2579 * Peek first request and see if we have a ->queue_rqs() hook.
2580 * If we do, we can dispatch the whole plug list in one go. We
2581 * already know at this point that all requests belong to the
2582 * same queue, caller must ensure that's the case.
2584 * Since we pass off the full list to the driver at this point,
2585 * we do not increment the active request count for the queue.
2586 * Bypass shared tags for now because of that.
2588 if (q
->mq_ops
->queue_rqs
&&
2589 !(rq
->mq_hctx
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
)) {
2590 blk_mq_run_dispatch_ops(q
,
2591 __blk_mq_flush_plug_list(q
, plug
));
2592 if (rq_list_empty(plug
->mq_list
))
2596 blk_mq_run_dispatch_ops(q
,
2597 blk_mq_plug_issue_direct(plug
, false));
2598 if (rq_list_empty(plug
->mq_list
))
2606 rq
= rq_list_pop(&plug
->mq_list
);
2609 this_hctx
= rq
->mq_hctx
;
2610 this_ctx
= rq
->mq_ctx
;
2611 } else if (this_hctx
!= rq
->mq_hctx
|| this_ctx
!= rq
->mq_ctx
) {
2612 trace_block_unplug(this_hctx
->queue
, depth
,
2614 blk_mq_sched_insert_requests(this_hctx
, this_ctx
,
2615 &list
, from_schedule
);
2617 this_hctx
= rq
->mq_hctx
;
2618 this_ctx
= rq
->mq_ctx
;
2622 list_add(&rq
->queuelist
, &list
);
2624 } while (!rq_list_empty(plug
->mq_list
));
2626 if (!list_empty(&list
)) {
2627 trace_block_unplug(this_hctx
->queue
, depth
, !from_schedule
);
2628 blk_mq_sched_insert_requests(this_hctx
, this_ctx
, &list
,
2633 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx
*hctx
,
2634 struct list_head
*list
)
2639 while (!list_empty(list
)) {
2641 struct request
*rq
= list_first_entry(list
, struct request
,
2644 list_del_init(&rq
->queuelist
);
2645 ret
= blk_mq_request_issue_directly(rq
, list_empty(list
));
2646 if (ret
!= BLK_STS_OK
) {
2647 if (ret
== BLK_STS_RESOURCE
||
2648 ret
== BLK_STS_DEV_RESOURCE
) {
2649 blk_mq_request_bypass_insert(rq
, false,
2653 blk_mq_end_request(rq
, ret
);
2660 * If we didn't flush the entire list, we could have told
2661 * the driver there was more coming, but that turned out to
2664 if ((!list_empty(list
) || errors
) &&
2665 hctx
->queue
->mq_ops
->commit_rqs
&& queued
)
2666 hctx
->queue
->mq_ops
->commit_rqs(hctx
);
2670 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
2671 * queues. This is important for md arrays to benefit from merging
2674 static inline unsigned short blk_plug_max_rq_count(struct blk_plug
*plug
)
2676 if (plug
->multiple_queues
)
2677 return BLK_MAX_REQUEST_COUNT
* 2;
2678 return BLK_MAX_REQUEST_COUNT
;
2681 static void blk_add_rq_to_plug(struct blk_plug
*plug
, struct request
*rq
)
2683 struct request
*last
= rq_list_peek(&plug
->mq_list
);
2685 if (!plug
->rq_count
) {
2686 trace_block_plug(rq
->q
);
2687 } else if (plug
->rq_count
>= blk_plug_max_rq_count(plug
) ||
2688 (!blk_queue_nomerges(rq
->q
) &&
2689 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
2690 blk_mq_flush_plug_list(plug
, false);
2691 trace_block_plug(rq
->q
);
2694 if (!plug
->multiple_queues
&& last
&& last
->q
!= rq
->q
)
2695 plug
->multiple_queues
= true;
2696 if (!plug
->has_elevator
&& (rq
->rq_flags
& RQF_ELV
))
2697 plug
->has_elevator
= true;
2699 rq_list_add(&plug
->mq_list
, rq
);
2703 static bool blk_mq_attempt_bio_merge(struct request_queue
*q
,
2704 struct bio
*bio
, unsigned int nr_segs
)
2706 if (!blk_queue_nomerges(q
) && bio_mergeable(bio
)) {
2707 if (blk_attempt_plug_merge(q
, bio
, nr_segs
))
2709 if (blk_mq_sched_bio_merge(q
, bio
, nr_segs
))
2715 static struct request
*blk_mq_get_new_requests(struct request_queue
*q
,
2716 struct blk_plug
*plug
,
2719 struct blk_mq_alloc_data data
= {
2722 .cmd_flags
= bio
->bi_opf
,
2726 if (unlikely(bio_queue_enter(bio
)))
2730 data
.nr_tags
= plug
->nr_ios
;
2732 data
.cached_rq
= &plug
->cached_rq
;
2735 rq
= __blk_mq_alloc_requests(&data
);
2738 rq_qos_cleanup(q
, bio
);
2739 if (bio
->bi_opf
& REQ_NOWAIT
)
2740 bio_wouldblock_error(bio
);
2745 static inline struct request
*blk_mq_get_cached_request(struct request_queue
*q
,
2746 struct blk_plug
*plug
, struct bio
*bio
)
2752 rq
= rq_list_peek(&plug
->cached_rq
);
2753 if (!rq
|| rq
->q
!= q
)
2756 if (blk_mq_get_hctx_type(bio
->bi_opf
) != rq
->mq_hctx
->type
)
2758 if (op_is_flush(rq
->cmd_flags
) != op_is_flush(bio
->bi_opf
))
2761 rq
->cmd_flags
= bio
->bi_opf
;
2762 plug
->cached_rq
= rq_list_next(rq
);
2763 INIT_LIST_HEAD(&rq
->queuelist
);
2768 * blk_mq_submit_bio - Create and send a request to block device.
2769 * @bio: Bio pointer.
2771 * Builds up a request structure from @q and @bio and send to the device. The
2772 * request may not be queued directly to hardware if:
2773 * * This request can be merged with another one
2774 * * We want to place request at plug queue for possible future merging
2775 * * There is an IO scheduler active at this queue
2777 * It will not queue the request if there is an error with the bio, or at the
2780 void blk_mq_submit_bio(struct bio
*bio
)
2782 struct request_queue
*q
= bdev_get_queue(bio
->bi_bdev
);
2783 struct blk_plug
*plug
= blk_mq_plug(q
, bio
);
2784 const int is_sync
= op_is_sync(bio
->bi_opf
);
2786 unsigned int nr_segs
= 1;
2789 if (unlikely(!blk_crypto_bio_prep(&bio
)))
2792 blk_queue_bounce(q
, &bio
);
2793 if (blk_may_split(q
, bio
))
2794 __blk_queue_split(q
, &bio
, &nr_segs
);
2796 if (!bio_integrity_prep(bio
))
2799 if (blk_mq_attempt_bio_merge(q
, bio
, nr_segs
))
2802 rq_qos_throttle(q
, bio
);
2804 rq
= blk_mq_get_cached_request(q
, plug
, bio
);
2806 rq
= blk_mq_get_new_requests(q
, plug
, bio
);
2811 trace_block_getrq(bio
);
2813 rq_qos_track(q
, rq
, bio
);
2815 blk_mq_bio_to_request(rq
, bio
, nr_segs
);
2817 ret
= blk_crypto_init_request(rq
);
2818 if (ret
!= BLK_STS_OK
) {
2819 bio
->bi_status
= ret
;
2821 blk_mq_free_request(rq
);
2825 if (op_is_flush(bio
->bi_opf
)) {
2826 blk_insert_flush(rq
);
2831 blk_add_rq_to_plug(plug
, rq
);
2832 else if ((rq
->rq_flags
& RQF_ELV
) ||
2833 (rq
->mq_hctx
->dispatch_busy
&&
2834 (q
->nr_hw_queues
== 1 || !is_sync
)))
2835 blk_mq_sched_insert_request(rq
, false, true, true);
2837 blk_mq_run_dispatch_ops(rq
->q
,
2838 blk_mq_try_issue_directly(rq
->mq_hctx
, rq
));
2842 * blk_cloned_rq_check_limits - Helper function to check a cloned request
2843 * for the new queue limits
2845 * @rq: the request being checked
2848 * @rq may have been made based on weaker limitations of upper-level queues
2849 * in request stacking drivers, and it may violate the limitation of @q.
2850 * Since the block layer and the underlying device driver trust @rq
2851 * after it is inserted to @q, it should be checked against @q before
2852 * the insertion using this generic function.
2854 * Request stacking drivers like request-based dm may change the queue
2855 * limits when retrying requests on other queues. Those requests need
2856 * to be checked against the new queue limits again during dispatch.
2858 static blk_status_t
blk_cloned_rq_check_limits(struct request_queue
*q
,
2861 unsigned int max_sectors
= blk_queue_get_max_sectors(q
, req_op(rq
));
2863 if (blk_rq_sectors(rq
) > max_sectors
) {
2865 * SCSI device does not have a good way to return if
2866 * Write Same/Zero is actually supported. If a device rejects
2867 * a non-read/write command (discard, write same,etc.) the
2868 * low-level device driver will set the relevant queue limit to
2869 * 0 to prevent blk-lib from issuing more of the offending
2870 * operations. Commands queued prior to the queue limit being
2871 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
2872 * errors being propagated to upper layers.
2874 if (max_sectors
== 0)
2875 return BLK_STS_NOTSUPP
;
2877 printk(KERN_ERR
"%s: over max size limit. (%u > %u)\n",
2878 __func__
, blk_rq_sectors(rq
), max_sectors
);
2879 return BLK_STS_IOERR
;
2883 * The queue settings related to segment counting may differ from the
2886 rq
->nr_phys_segments
= blk_recalc_rq_segments(rq
);
2887 if (rq
->nr_phys_segments
> queue_max_segments(q
)) {
2888 printk(KERN_ERR
"%s: over max segments limit. (%hu > %hu)\n",
2889 __func__
, rq
->nr_phys_segments
, queue_max_segments(q
));
2890 return BLK_STS_IOERR
;
2897 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
2898 * @q: the queue to submit the request
2899 * @rq: the request being queued
2901 blk_status_t
blk_insert_cloned_request(struct request_queue
*q
, struct request
*rq
)
2905 ret
= blk_cloned_rq_check_limits(q
, rq
);
2906 if (ret
!= BLK_STS_OK
)
2910 should_fail_request(rq
->q
->disk
->part0
, blk_rq_bytes(rq
)))
2911 return BLK_STS_IOERR
;
2913 if (blk_crypto_insert_cloned_request(rq
))
2914 return BLK_STS_IOERR
;
2916 blk_account_io_start(rq
);
2919 * Since we have a scheduler attached on the top device,
2920 * bypass a potential scheduler on the bottom device for
2923 blk_mq_run_dispatch_ops(rq
->q
,
2924 ret
= blk_mq_request_issue_directly(rq
, true));
2926 blk_account_io_done(rq
, ktime_get_ns());
2929 EXPORT_SYMBOL_GPL(blk_insert_cloned_request
);
2932 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
2933 * @rq: the clone request to be cleaned up
2936 * Free all bios in @rq for a cloned request.
2938 void blk_rq_unprep_clone(struct request
*rq
)
2942 while ((bio
= rq
->bio
) != NULL
) {
2943 rq
->bio
= bio
->bi_next
;
2948 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone
);
2951 * blk_rq_prep_clone - Helper function to setup clone request
2952 * @rq: the request to be setup
2953 * @rq_src: original request to be cloned
2954 * @bs: bio_set that bios for clone are allocated from
2955 * @gfp_mask: memory allocation mask for bio
2956 * @bio_ctr: setup function to be called for each clone bio.
2957 * Returns %0 for success, non %0 for failure.
2958 * @data: private data to be passed to @bio_ctr
2961 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
2962 * Also, pages which the original bios are pointing to are not copied
2963 * and the cloned bios just point same pages.
2964 * So cloned bios must be completed before original bios, which means
2965 * the caller must complete @rq before @rq_src.
2967 int blk_rq_prep_clone(struct request
*rq
, struct request
*rq_src
,
2968 struct bio_set
*bs
, gfp_t gfp_mask
,
2969 int (*bio_ctr
)(struct bio
*, struct bio
*, void *),
2972 struct bio
*bio
, *bio_src
;
2977 __rq_for_each_bio(bio_src
, rq_src
) {
2978 bio
= bio_alloc_clone(rq
->q
->disk
->part0
, bio_src
, gfp_mask
,
2983 if (bio_ctr
&& bio_ctr(bio
, bio_src
, data
))
2987 rq
->biotail
->bi_next
= bio
;
2990 rq
->bio
= rq
->biotail
= bio
;
2995 /* Copy attributes of the original request to the clone request. */
2996 rq
->__sector
= blk_rq_pos(rq_src
);
2997 rq
->__data_len
= blk_rq_bytes(rq_src
);
2998 if (rq_src
->rq_flags
& RQF_SPECIAL_PAYLOAD
) {
2999 rq
->rq_flags
|= RQF_SPECIAL_PAYLOAD
;
3000 rq
->special_vec
= rq_src
->special_vec
;
3002 rq
->nr_phys_segments
= rq_src
->nr_phys_segments
;
3003 rq
->ioprio
= rq_src
->ioprio
;
3005 if (rq
->bio
&& blk_crypto_rq_bio_prep(rq
, rq
->bio
, gfp_mask
) < 0)
3013 blk_rq_unprep_clone(rq
);
3017 EXPORT_SYMBOL_GPL(blk_rq_prep_clone
);
3020 * Steal bios from a request and add them to a bio list.
3021 * The request must not have been partially completed before.
3023 void blk_steal_bios(struct bio_list
*list
, struct request
*rq
)
3027 list
->tail
->bi_next
= rq
->bio
;
3029 list
->head
= rq
->bio
;
3030 list
->tail
= rq
->biotail
;
3038 EXPORT_SYMBOL_GPL(blk_steal_bios
);
3040 static size_t order_to_size(unsigned int order
)
3042 return (size_t)PAGE_SIZE
<< order
;
3045 /* called before freeing request pool in @tags */
3046 static void blk_mq_clear_rq_mapping(struct blk_mq_tags
*drv_tags
,
3047 struct blk_mq_tags
*tags
)
3050 unsigned long flags
;
3052 /* There is no need to clear a driver tags own mapping */
3053 if (drv_tags
== tags
)
3056 list_for_each_entry(page
, &tags
->page_list
, lru
) {
3057 unsigned long start
= (unsigned long)page_address(page
);
3058 unsigned long end
= start
+ order_to_size(page
->private);
3061 for (i
= 0; i
< drv_tags
->nr_tags
; i
++) {
3062 struct request
*rq
= drv_tags
->rqs
[i
];
3063 unsigned long rq_addr
= (unsigned long)rq
;
3065 if (rq_addr
>= start
&& rq_addr
< end
) {
3066 WARN_ON_ONCE(req_ref_read(rq
) != 0);
3067 cmpxchg(&drv_tags
->rqs
[i
], rq
, NULL
);
3073 * Wait until all pending iteration is done.
3075 * Request reference is cleared and it is guaranteed to be observed
3076 * after the ->lock is released.
3078 spin_lock_irqsave(&drv_tags
->lock
, flags
);
3079 spin_unlock_irqrestore(&drv_tags
->lock
, flags
);
3082 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
3083 unsigned int hctx_idx
)
3085 struct blk_mq_tags
*drv_tags
;
3088 if (blk_mq_is_shared_tags(set
->flags
))
3089 drv_tags
= set
->shared_tags
;
3091 drv_tags
= set
->tags
[hctx_idx
];
3093 if (tags
->static_rqs
&& set
->ops
->exit_request
) {
3096 for (i
= 0; i
< tags
->nr_tags
; i
++) {
3097 struct request
*rq
= tags
->static_rqs
[i
];
3101 set
->ops
->exit_request(set
, rq
, hctx_idx
);
3102 tags
->static_rqs
[i
] = NULL
;
3106 blk_mq_clear_rq_mapping(drv_tags
, tags
);
3108 while (!list_empty(&tags
->page_list
)) {
3109 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
3110 list_del_init(&page
->lru
);
3112 * Remove kmemleak object previously allocated in
3113 * blk_mq_alloc_rqs().
3115 kmemleak_free(page_address(page
));
3116 __free_pages(page
, page
->private);
3120 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
3124 kfree(tags
->static_rqs
);
3125 tags
->static_rqs
= NULL
;
3127 blk_mq_free_tags(tags
);
3130 static struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
3131 unsigned int hctx_idx
,
3132 unsigned int nr_tags
,
3133 unsigned int reserved_tags
)
3135 struct blk_mq_tags
*tags
;
3138 node
= blk_mq_hw_queue_to_node(&set
->map
[HCTX_TYPE_DEFAULT
], hctx_idx
);
3139 if (node
== NUMA_NO_NODE
)
3140 node
= set
->numa_node
;
3142 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
,
3143 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
3147 tags
->rqs
= kcalloc_node(nr_tags
, sizeof(struct request
*),
3148 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
3151 blk_mq_free_tags(tags
);
3155 tags
->static_rqs
= kcalloc_node(nr_tags
, sizeof(struct request
*),
3156 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
3158 if (!tags
->static_rqs
) {
3160 blk_mq_free_tags(tags
);
3167 static int blk_mq_init_request(struct blk_mq_tag_set
*set
, struct request
*rq
,
3168 unsigned int hctx_idx
, int node
)
3172 if (set
->ops
->init_request
) {
3173 ret
= set
->ops
->init_request(set
, rq
, hctx_idx
, node
);
3178 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
3182 static int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
,
3183 struct blk_mq_tags
*tags
,
3184 unsigned int hctx_idx
, unsigned int depth
)
3186 unsigned int i
, j
, entries_per_page
, max_order
= 4;
3187 size_t rq_size
, left
;
3190 node
= blk_mq_hw_queue_to_node(&set
->map
[HCTX_TYPE_DEFAULT
], hctx_idx
);
3191 if (node
== NUMA_NO_NODE
)
3192 node
= set
->numa_node
;
3194 INIT_LIST_HEAD(&tags
->page_list
);
3197 * rq_size is the size of the request plus driver payload, rounded
3198 * to the cacheline size
3200 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
3202 left
= rq_size
* depth
;
3204 for (i
= 0; i
< depth
; ) {
3205 int this_order
= max_order
;
3210 while (this_order
&& left
< order_to_size(this_order
- 1))
3214 page
= alloc_pages_node(node
,
3215 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
3221 if (order_to_size(this_order
) < rq_size
)
3228 page
->private = this_order
;
3229 list_add_tail(&page
->lru
, &tags
->page_list
);
3231 p
= page_address(page
);
3233 * Allow kmemleak to scan these pages as they contain pointers
3234 * to additional allocations like via ops->init_request().
3236 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
3237 entries_per_page
= order_to_size(this_order
) / rq_size
;
3238 to_do
= min(entries_per_page
, depth
- i
);
3239 left
-= to_do
* rq_size
;
3240 for (j
= 0; j
< to_do
; j
++) {
3241 struct request
*rq
= p
;
3243 tags
->static_rqs
[i
] = rq
;
3244 if (blk_mq_init_request(set
, rq
, hctx_idx
, node
)) {
3245 tags
->static_rqs
[i
] = NULL
;
3256 blk_mq_free_rqs(set
, tags
, hctx_idx
);
3260 struct rq_iter_data
{
3261 struct blk_mq_hw_ctx
*hctx
;
3265 static bool blk_mq_has_request(struct request
*rq
, void *data
, bool reserved
)
3267 struct rq_iter_data
*iter_data
= data
;
3269 if (rq
->mq_hctx
!= iter_data
->hctx
)
3271 iter_data
->has_rq
= true;
3275 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx
*hctx
)
3277 struct blk_mq_tags
*tags
= hctx
->sched_tags
?
3278 hctx
->sched_tags
: hctx
->tags
;
3279 struct rq_iter_data data
= {
3283 blk_mq_all_tag_iter(tags
, blk_mq_has_request
, &data
);
3287 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu
,
3288 struct blk_mq_hw_ctx
*hctx
)
3290 if (cpumask_first_and(hctx
->cpumask
, cpu_online_mask
) != cpu
)
3292 if (cpumask_next_and(cpu
, hctx
->cpumask
, cpu_online_mask
) < nr_cpu_ids
)
3297 static int blk_mq_hctx_notify_offline(unsigned int cpu
, struct hlist_node
*node
)
3299 struct blk_mq_hw_ctx
*hctx
= hlist_entry_safe(node
,
3300 struct blk_mq_hw_ctx
, cpuhp_online
);
3302 if (!cpumask_test_cpu(cpu
, hctx
->cpumask
) ||
3303 !blk_mq_last_cpu_in_hctx(cpu
, hctx
))
3307 * Prevent new request from being allocated on the current hctx.
3309 * The smp_mb__after_atomic() Pairs with the implied barrier in
3310 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3311 * seen once we return from the tag allocator.
3313 set_bit(BLK_MQ_S_INACTIVE
, &hctx
->state
);
3314 smp_mb__after_atomic();
3317 * Try to grab a reference to the queue and wait for any outstanding
3318 * requests. If we could not grab a reference the queue has been
3319 * frozen and there are no requests.
3321 if (percpu_ref_tryget(&hctx
->queue
->q_usage_counter
)) {
3322 while (blk_mq_hctx_has_requests(hctx
))
3324 percpu_ref_put(&hctx
->queue
->q_usage_counter
);
3330 static int blk_mq_hctx_notify_online(unsigned int cpu
, struct hlist_node
*node
)
3332 struct blk_mq_hw_ctx
*hctx
= hlist_entry_safe(node
,
3333 struct blk_mq_hw_ctx
, cpuhp_online
);
3335 if (cpumask_test_cpu(cpu
, hctx
->cpumask
))
3336 clear_bit(BLK_MQ_S_INACTIVE
, &hctx
->state
);
3341 * 'cpu' is going away. splice any existing rq_list entries from this
3342 * software queue to the hw queue dispatch list, and ensure that it
3345 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
3347 struct blk_mq_hw_ctx
*hctx
;
3348 struct blk_mq_ctx
*ctx
;
3350 enum hctx_type type
;
3352 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
3353 if (!cpumask_test_cpu(cpu
, hctx
->cpumask
))
3356 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
3359 spin_lock(&ctx
->lock
);
3360 if (!list_empty(&ctx
->rq_lists
[type
])) {
3361 list_splice_init(&ctx
->rq_lists
[type
], &tmp
);
3362 blk_mq_hctx_clear_pending(hctx
, ctx
);
3364 spin_unlock(&ctx
->lock
);
3366 if (list_empty(&tmp
))
3369 spin_lock(&hctx
->lock
);
3370 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
3371 spin_unlock(&hctx
->lock
);
3373 blk_mq_run_hw_queue(hctx
, true);
3377 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
3379 if (!(hctx
->flags
& BLK_MQ_F_STACKING
))
3380 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE
,
3381 &hctx
->cpuhp_online
);
3382 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
3387 * Before freeing hw queue, clearing the flush request reference in
3388 * tags->rqs[] for avoiding potential UAF.
3390 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags
*tags
,
3391 unsigned int queue_depth
, struct request
*flush_rq
)
3394 unsigned long flags
;
3396 /* The hw queue may not be mapped yet */
3400 WARN_ON_ONCE(req_ref_read(flush_rq
) != 0);
3402 for (i
= 0; i
< queue_depth
; i
++)
3403 cmpxchg(&tags
->rqs
[i
], flush_rq
, NULL
);
3406 * Wait until all pending iteration is done.
3408 * Request reference is cleared and it is guaranteed to be observed
3409 * after the ->lock is released.
3411 spin_lock_irqsave(&tags
->lock
, flags
);
3412 spin_unlock_irqrestore(&tags
->lock
, flags
);
3415 /* hctx->ctxs will be freed in queue's release handler */
3416 static void blk_mq_exit_hctx(struct request_queue
*q
,
3417 struct blk_mq_tag_set
*set
,
3418 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
3420 struct request
*flush_rq
= hctx
->fq
->flush_rq
;
3422 if (blk_mq_hw_queue_mapped(hctx
))
3423 blk_mq_tag_idle(hctx
);
3425 blk_mq_clear_flush_rq_mapping(set
->tags
[hctx_idx
],
3426 set
->queue_depth
, flush_rq
);
3427 if (set
->ops
->exit_request
)
3428 set
->ops
->exit_request(set
, flush_rq
, hctx_idx
);
3430 if (set
->ops
->exit_hctx
)
3431 set
->ops
->exit_hctx(hctx
, hctx_idx
);
3433 blk_mq_remove_cpuhp(hctx
);
3435 spin_lock(&q
->unused_hctx_lock
);
3436 list_add(&hctx
->hctx_list
, &q
->unused_hctx_list
);
3437 spin_unlock(&q
->unused_hctx_lock
);
3440 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
3441 struct blk_mq_tag_set
*set
, int nr_queue
)
3443 struct blk_mq_hw_ctx
*hctx
;
3446 queue_for_each_hw_ctx(q
, hctx
, i
) {
3449 blk_mq_debugfs_unregister_hctx(hctx
);
3450 blk_mq_exit_hctx(q
, set
, hctx
, i
);
3454 static int blk_mq_init_hctx(struct request_queue
*q
,
3455 struct blk_mq_tag_set
*set
,
3456 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
3458 hctx
->queue_num
= hctx_idx
;
3460 if (!(hctx
->flags
& BLK_MQ_F_STACKING
))
3461 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE
,
3462 &hctx
->cpuhp_online
);
3463 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
3465 hctx
->tags
= set
->tags
[hctx_idx
];
3467 if (set
->ops
->init_hctx
&&
3468 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
3469 goto unregister_cpu_notifier
;
3471 if (blk_mq_init_request(set
, hctx
->fq
->flush_rq
, hctx_idx
,
3477 if (set
->ops
->exit_hctx
)
3478 set
->ops
->exit_hctx(hctx
, hctx_idx
);
3479 unregister_cpu_notifier
:
3480 blk_mq_remove_cpuhp(hctx
);
3484 static struct blk_mq_hw_ctx
*
3485 blk_mq_alloc_hctx(struct request_queue
*q
, struct blk_mq_tag_set
*set
,
3488 struct blk_mq_hw_ctx
*hctx
;
3489 gfp_t gfp
= GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
;
3491 hctx
= kzalloc_node(sizeof(struct blk_mq_hw_ctx
), gfp
, node
);
3493 goto fail_alloc_hctx
;
3495 if (!zalloc_cpumask_var_node(&hctx
->cpumask
, gfp
, node
))
3498 atomic_set(&hctx
->nr_active
, 0);
3499 if (node
== NUMA_NO_NODE
)
3500 node
= set
->numa_node
;
3501 hctx
->numa_node
= node
;
3503 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
3504 spin_lock_init(&hctx
->lock
);
3505 INIT_LIST_HEAD(&hctx
->dispatch
);
3507 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_QUEUE_SHARED
;
3509 INIT_LIST_HEAD(&hctx
->hctx_list
);
3512 * Allocate space for all possible cpus to avoid allocation at
3515 hctx
->ctxs
= kmalloc_array_node(nr_cpu_ids
, sizeof(void *),
3520 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8),
3521 gfp
, node
, false, false))
3525 spin_lock_init(&hctx
->dispatch_wait_lock
);
3526 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
3527 INIT_LIST_HEAD(&hctx
->dispatch_wait
.entry
);
3529 hctx
->fq
= blk_alloc_flush_queue(hctx
->numa_node
, set
->cmd_size
, gfp
);
3533 blk_mq_hctx_kobj_init(hctx
);
3538 sbitmap_free(&hctx
->ctx_map
);
3542 free_cpumask_var(hctx
->cpumask
);
3549 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
3550 unsigned int nr_hw_queues
)
3552 struct blk_mq_tag_set
*set
= q
->tag_set
;
3555 for_each_possible_cpu(i
) {
3556 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
3557 struct blk_mq_hw_ctx
*hctx
;
3561 spin_lock_init(&__ctx
->lock
);
3562 for (k
= HCTX_TYPE_DEFAULT
; k
< HCTX_MAX_TYPES
; k
++)
3563 INIT_LIST_HEAD(&__ctx
->rq_lists
[k
]);
3568 * Set local node, IFF we have more than one hw queue. If
3569 * not, we remain on the home node of the device
3571 for (j
= 0; j
< set
->nr_maps
; j
++) {
3572 hctx
= blk_mq_map_queue_type(q
, j
, i
);
3573 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
3574 hctx
->numa_node
= cpu_to_node(i
);
3579 struct blk_mq_tags
*blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set
*set
,
3580 unsigned int hctx_idx
,
3583 struct blk_mq_tags
*tags
;
3586 tags
= blk_mq_alloc_rq_map(set
, hctx_idx
, depth
, set
->reserved_tags
);
3590 ret
= blk_mq_alloc_rqs(set
, tags
, hctx_idx
, depth
);
3592 blk_mq_free_rq_map(tags
);
3599 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set
*set
,
3602 if (blk_mq_is_shared_tags(set
->flags
)) {
3603 set
->tags
[hctx_idx
] = set
->shared_tags
;
3608 set
->tags
[hctx_idx
] = blk_mq_alloc_map_and_rqs(set
, hctx_idx
,
3611 return set
->tags
[hctx_idx
];
3614 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set
*set
,
3615 struct blk_mq_tags
*tags
,
3616 unsigned int hctx_idx
)
3619 blk_mq_free_rqs(set
, tags
, hctx_idx
);
3620 blk_mq_free_rq_map(tags
);
3624 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set
*set
,
3625 unsigned int hctx_idx
)
3627 if (!blk_mq_is_shared_tags(set
->flags
))
3628 blk_mq_free_map_and_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
3630 set
->tags
[hctx_idx
] = NULL
;
3633 static void blk_mq_map_swqueue(struct request_queue
*q
)
3635 unsigned int i
, j
, hctx_idx
;
3636 struct blk_mq_hw_ctx
*hctx
;
3637 struct blk_mq_ctx
*ctx
;
3638 struct blk_mq_tag_set
*set
= q
->tag_set
;
3640 queue_for_each_hw_ctx(q
, hctx
, i
) {
3641 cpumask_clear(hctx
->cpumask
);
3643 hctx
->dispatch_from
= NULL
;
3647 * Map software to hardware queues.
3649 * If the cpu isn't present, the cpu is mapped to first hctx.
3651 for_each_possible_cpu(i
) {
3653 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
3654 for (j
= 0; j
< set
->nr_maps
; j
++) {
3655 if (!set
->map
[j
].nr_queues
) {
3656 ctx
->hctxs
[j
] = blk_mq_map_queue_type(q
,
3657 HCTX_TYPE_DEFAULT
, i
);
3660 hctx_idx
= set
->map
[j
].mq_map
[i
];
3661 /* unmapped hw queue can be remapped after CPU topo changed */
3662 if (!set
->tags
[hctx_idx
] &&
3663 !__blk_mq_alloc_map_and_rqs(set
, hctx_idx
)) {
3665 * If tags initialization fail for some hctx,
3666 * that hctx won't be brought online. In this
3667 * case, remap the current ctx to hctx[0] which
3668 * is guaranteed to always have tags allocated
3670 set
->map
[j
].mq_map
[i
] = 0;
3673 hctx
= blk_mq_map_queue_type(q
, j
, i
);
3674 ctx
->hctxs
[j
] = hctx
;
3676 * If the CPU is already set in the mask, then we've
3677 * mapped this one already. This can happen if
3678 * devices share queues across queue maps.
3680 if (cpumask_test_cpu(i
, hctx
->cpumask
))
3683 cpumask_set_cpu(i
, hctx
->cpumask
);
3685 ctx
->index_hw
[hctx
->type
] = hctx
->nr_ctx
;
3686 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
3689 * If the nr_ctx type overflows, we have exceeded the
3690 * amount of sw queues we can support.
3692 BUG_ON(!hctx
->nr_ctx
);
3695 for (; j
< HCTX_MAX_TYPES
; j
++)
3696 ctx
->hctxs
[j
] = blk_mq_map_queue_type(q
,
3697 HCTX_TYPE_DEFAULT
, i
);
3700 queue_for_each_hw_ctx(q
, hctx
, i
) {
3702 * If no software queues are mapped to this hardware queue,
3703 * disable it and free the request entries.
3705 if (!hctx
->nr_ctx
) {
3706 /* Never unmap queue 0. We need it as a
3707 * fallback in case of a new remap fails
3711 __blk_mq_free_map_and_rqs(set
, i
);
3717 hctx
->tags
= set
->tags
[i
];
3718 WARN_ON(!hctx
->tags
);
3721 * Set the map size to the number of mapped software queues.
3722 * This is more accurate and more efficient than looping
3723 * over all possibly mapped software queues.
3725 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
3728 * Initialize batch roundrobin counts
3730 hctx
->next_cpu
= blk_mq_first_mapped_cpu(hctx
);
3731 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
3736 * Caller needs to ensure that we're either frozen/quiesced, or that
3737 * the queue isn't live yet.
3739 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
3741 struct blk_mq_hw_ctx
*hctx
;
3744 queue_for_each_hw_ctx(q
, hctx
, i
) {
3746 hctx
->flags
|= BLK_MQ_F_TAG_QUEUE_SHARED
;
3748 blk_mq_tag_idle(hctx
);
3749 hctx
->flags
&= ~BLK_MQ_F_TAG_QUEUE_SHARED
;
3754 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set
*set
,
3757 struct request_queue
*q
;
3759 lockdep_assert_held(&set
->tag_list_lock
);
3761 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3762 blk_mq_freeze_queue(q
);
3763 queue_set_hctx_shared(q
, shared
);
3764 blk_mq_unfreeze_queue(q
);
3768 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
3770 struct blk_mq_tag_set
*set
= q
->tag_set
;
3772 mutex_lock(&set
->tag_list_lock
);
3773 list_del(&q
->tag_set_list
);
3774 if (list_is_singular(&set
->tag_list
)) {
3775 /* just transitioned to unshared */
3776 set
->flags
&= ~BLK_MQ_F_TAG_QUEUE_SHARED
;
3777 /* update existing queue */
3778 blk_mq_update_tag_set_shared(set
, false);
3780 mutex_unlock(&set
->tag_list_lock
);
3781 INIT_LIST_HEAD(&q
->tag_set_list
);
3784 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
3785 struct request_queue
*q
)
3787 mutex_lock(&set
->tag_list_lock
);
3790 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3792 if (!list_empty(&set
->tag_list
) &&
3793 !(set
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
)) {
3794 set
->flags
|= BLK_MQ_F_TAG_QUEUE_SHARED
;
3795 /* update existing queue */
3796 blk_mq_update_tag_set_shared(set
, true);
3798 if (set
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
)
3799 queue_set_hctx_shared(q
, true);
3800 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
3802 mutex_unlock(&set
->tag_list_lock
);
3805 /* All allocations will be freed in release handler of q->mq_kobj */
3806 static int blk_mq_alloc_ctxs(struct request_queue
*q
)
3808 struct blk_mq_ctxs
*ctxs
;
3811 ctxs
= kzalloc(sizeof(*ctxs
), GFP_KERNEL
);
3815 ctxs
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
3816 if (!ctxs
->queue_ctx
)
3819 for_each_possible_cpu(cpu
) {
3820 struct blk_mq_ctx
*ctx
= per_cpu_ptr(ctxs
->queue_ctx
, cpu
);
3824 q
->mq_kobj
= &ctxs
->kobj
;
3825 q
->queue_ctx
= ctxs
->queue_ctx
;
3834 * It is the actual release handler for mq, but we do it from
3835 * request queue's release handler for avoiding use-after-free
3836 * and headache because q->mq_kobj shouldn't have been introduced,
3837 * but we can't group ctx/kctx kobj without it.
3839 void blk_mq_release(struct request_queue
*q
)
3841 struct blk_mq_hw_ctx
*hctx
, *next
;
3844 queue_for_each_hw_ctx(q
, hctx
, i
)
3845 WARN_ON_ONCE(hctx
&& list_empty(&hctx
->hctx_list
));
3847 /* all hctx are in .unused_hctx_list now */
3848 list_for_each_entry_safe(hctx
, next
, &q
->unused_hctx_list
, hctx_list
) {
3849 list_del_init(&hctx
->hctx_list
);
3850 kobject_put(&hctx
->kobj
);
3853 kfree(q
->queue_hw_ctx
);
3856 * release .mq_kobj and sw queue's kobject now because
3857 * both share lifetime with request queue.
3859 blk_mq_sysfs_deinit(q
);
3862 static struct request_queue
*blk_mq_init_queue_data(struct blk_mq_tag_set
*set
,
3865 struct request_queue
*q
;
3868 q
= blk_alloc_queue(set
->numa_node
, set
->flags
& BLK_MQ_F_BLOCKING
);
3870 return ERR_PTR(-ENOMEM
);
3871 q
->queuedata
= queuedata
;
3872 ret
= blk_mq_init_allocated_queue(set
, q
);
3874 blk_cleanup_queue(q
);
3875 return ERR_PTR(ret
);
3880 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
3882 return blk_mq_init_queue_data(set
, NULL
);
3884 EXPORT_SYMBOL(blk_mq_init_queue
);
3886 struct gendisk
*__blk_mq_alloc_disk(struct blk_mq_tag_set
*set
, void *queuedata
,
3887 struct lock_class_key
*lkclass
)
3889 struct request_queue
*q
;
3890 struct gendisk
*disk
;
3892 q
= blk_mq_init_queue_data(set
, queuedata
);
3896 disk
= __alloc_disk_node(q
, set
->numa_node
, lkclass
);
3898 blk_cleanup_queue(q
);
3899 return ERR_PTR(-ENOMEM
);
3903 EXPORT_SYMBOL(__blk_mq_alloc_disk
);
3905 static struct blk_mq_hw_ctx
*blk_mq_alloc_and_init_hctx(
3906 struct blk_mq_tag_set
*set
, struct request_queue
*q
,
3907 int hctx_idx
, int node
)
3909 struct blk_mq_hw_ctx
*hctx
= NULL
, *tmp
;
3911 /* reuse dead hctx first */
3912 spin_lock(&q
->unused_hctx_lock
);
3913 list_for_each_entry(tmp
, &q
->unused_hctx_list
, hctx_list
) {
3914 if (tmp
->numa_node
== node
) {
3920 list_del_init(&hctx
->hctx_list
);
3921 spin_unlock(&q
->unused_hctx_lock
);
3924 hctx
= blk_mq_alloc_hctx(q
, set
, node
);
3928 if (blk_mq_init_hctx(q
, set
, hctx
, hctx_idx
))
3934 kobject_put(&hctx
->kobj
);
3939 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
3940 struct request_queue
*q
)
3943 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
3945 if (q
->nr_hw_queues
< set
->nr_hw_queues
) {
3946 struct blk_mq_hw_ctx
**new_hctxs
;
3948 new_hctxs
= kcalloc_node(set
->nr_hw_queues
,
3949 sizeof(*new_hctxs
), GFP_KERNEL
,
3954 memcpy(new_hctxs
, hctxs
, q
->nr_hw_queues
*
3956 q
->queue_hw_ctx
= new_hctxs
;
3961 /* protect against switching io scheduler */
3962 mutex_lock(&q
->sysfs_lock
);
3963 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
3965 struct blk_mq_hw_ctx
*hctx
;
3967 node
= blk_mq_hw_queue_to_node(&set
->map
[HCTX_TYPE_DEFAULT
], i
);
3969 * If the hw queue has been mapped to another numa node,
3970 * we need to realloc the hctx. If allocation fails, fallback
3971 * to use the previous one.
3973 if (hctxs
[i
] && (hctxs
[i
]->numa_node
== node
))
3976 hctx
= blk_mq_alloc_and_init_hctx(set
, q
, i
, node
);
3979 blk_mq_exit_hctx(q
, set
, hctxs
[i
], i
);
3983 pr_warn("Allocate new hctx on node %d fails,\
3984 fallback to previous one on node %d\n",
3985 node
, hctxs
[i
]->numa_node
);
3991 * Increasing nr_hw_queues fails. Free the newly allocated
3992 * hctxs and keep the previous q->nr_hw_queues.
3994 if (i
!= set
->nr_hw_queues
) {
3995 j
= q
->nr_hw_queues
;
3999 end
= q
->nr_hw_queues
;
4000 q
->nr_hw_queues
= set
->nr_hw_queues
;
4003 for (; j
< end
; j
++) {
4004 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
4007 blk_mq_exit_hctx(q
, set
, hctx
, j
);
4011 mutex_unlock(&q
->sysfs_lock
);
4014 int blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
4015 struct request_queue
*q
)
4017 WARN_ON_ONCE(blk_queue_has_srcu(q
) !=
4018 !!(set
->flags
& BLK_MQ_F_BLOCKING
));
4020 /* mark the queue as mq asap */
4021 q
->mq_ops
= set
->ops
;
4023 q
->poll_cb
= blk_stat_alloc_callback(blk_mq_poll_stats_fn
,
4024 blk_mq_poll_stats_bkt
,
4025 BLK_MQ_POLL_STATS_BKTS
, q
);
4029 if (blk_mq_alloc_ctxs(q
))
4032 /* init q->mq_kobj and sw queues' kobjects */
4033 blk_mq_sysfs_init(q
);
4035 INIT_LIST_HEAD(&q
->unused_hctx_list
);
4036 spin_lock_init(&q
->unused_hctx_lock
);
4038 blk_mq_realloc_hw_ctxs(set
, q
);
4039 if (!q
->nr_hw_queues
)
4042 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
4043 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
4047 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
4048 if (set
->nr_maps
> HCTX_TYPE_POLL
&&
4049 set
->map
[HCTX_TYPE_POLL
].nr_queues
)
4050 blk_queue_flag_set(QUEUE_FLAG_POLL
, q
);
4052 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
4053 INIT_LIST_HEAD(&q
->requeue_list
);
4054 spin_lock_init(&q
->requeue_lock
);
4056 q
->nr_requests
= set
->queue_depth
;
4059 * Default to classic polling
4061 q
->poll_nsec
= BLK_MQ_POLL_CLASSIC
;
4063 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
4064 blk_mq_add_queue_tag_set(set
, q
);
4065 blk_mq_map_swqueue(q
);
4069 kfree(q
->queue_hw_ctx
);
4070 q
->nr_hw_queues
= 0;
4071 blk_mq_sysfs_deinit(q
);
4073 blk_stat_free_callback(q
->poll_cb
);
4079 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
4081 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4082 void blk_mq_exit_queue(struct request_queue
*q
)
4084 struct blk_mq_tag_set
*set
= q
->tag_set
;
4086 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4087 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
4088 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4089 blk_mq_del_queue_tag_set(q
);
4092 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
4096 if (blk_mq_is_shared_tags(set
->flags
)) {
4097 set
->shared_tags
= blk_mq_alloc_map_and_rqs(set
,
4100 if (!set
->shared_tags
)
4104 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
4105 if (!__blk_mq_alloc_map_and_rqs(set
, i
))
4114 __blk_mq_free_map_and_rqs(set
, i
);
4116 if (blk_mq_is_shared_tags(set
->flags
)) {
4117 blk_mq_free_map_and_rqs(set
, set
->shared_tags
,
4118 BLK_MQ_NO_HCTX_IDX
);
4125 * Allocate the request maps associated with this tag_set. Note that this
4126 * may reduce the depth asked for, if memory is tight. set->queue_depth
4127 * will be updated to reflect the allocated depth.
4129 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set
*set
)
4134 depth
= set
->queue_depth
;
4136 err
= __blk_mq_alloc_rq_maps(set
);
4140 set
->queue_depth
>>= 1;
4141 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
4145 } while (set
->queue_depth
);
4147 if (!set
->queue_depth
|| err
) {
4148 pr_err("blk-mq: failed to allocate request map\n");
4152 if (depth
!= set
->queue_depth
)
4153 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4154 depth
, set
->queue_depth
);
4159 static int blk_mq_update_queue_map(struct blk_mq_tag_set
*set
)
4162 * blk_mq_map_queues() and multiple .map_queues() implementations
4163 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4164 * number of hardware queues.
4166 if (set
->nr_maps
== 1)
4167 set
->map
[HCTX_TYPE_DEFAULT
].nr_queues
= set
->nr_hw_queues
;
4169 if (set
->ops
->map_queues
&& !is_kdump_kernel()) {
4173 * transport .map_queues is usually done in the following
4176 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4177 * mask = get_cpu_mask(queue)
4178 * for_each_cpu(cpu, mask)
4179 * set->map[x].mq_map[cpu] = queue;
4182 * When we need to remap, the table has to be cleared for
4183 * killing stale mapping since one CPU may not be mapped
4186 for (i
= 0; i
< set
->nr_maps
; i
++)
4187 blk_mq_clear_mq_map(&set
->map
[i
]);
4189 return set
->ops
->map_queues(set
);
4191 BUG_ON(set
->nr_maps
> 1);
4192 return blk_mq_map_queues(&set
->map
[HCTX_TYPE_DEFAULT
]);
4196 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set
*set
,
4197 int cur_nr_hw_queues
, int new_nr_hw_queues
)
4199 struct blk_mq_tags
**new_tags
;
4201 if (cur_nr_hw_queues
>= new_nr_hw_queues
)
4204 new_tags
= kcalloc_node(new_nr_hw_queues
, sizeof(struct blk_mq_tags
*),
4205 GFP_KERNEL
, set
->numa_node
);
4210 memcpy(new_tags
, set
->tags
, cur_nr_hw_queues
*
4211 sizeof(*set
->tags
));
4213 set
->tags
= new_tags
;
4214 set
->nr_hw_queues
= new_nr_hw_queues
;
4219 static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set
*set
,
4220 int new_nr_hw_queues
)
4222 return blk_mq_realloc_tag_set_tags(set
, 0, new_nr_hw_queues
);
4226 * Alloc a tag set to be associated with one or more request queues.
4227 * May fail with EINVAL for various error conditions. May adjust the
4228 * requested depth down, if it's too large. In that case, the set
4229 * value will be stored in set->queue_depth.
4231 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
4235 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
4237 if (!set
->nr_hw_queues
)
4239 if (!set
->queue_depth
)
4241 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
4244 if (!set
->ops
->queue_rq
)
4247 if (!set
->ops
->get_budget
^ !set
->ops
->put_budget
)
4250 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
4251 pr_info("blk-mq: reduced tag depth to %u\n",
4253 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
4258 else if (set
->nr_maps
> HCTX_MAX_TYPES
)
4262 * If a crashdump is active, then we are potentially in a very
4263 * memory constrained environment. Limit us to 1 queue and
4264 * 64 tags to prevent using too much memory.
4266 if (is_kdump_kernel()) {
4267 set
->nr_hw_queues
= 1;
4269 set
->queue_depth
= min(64U, set
->queue_depth
);
4272 * There is no use for more h/w queues than cpus if we just have
4275 if (set
->nr_maps
== 1 && set
->nr_hw_queues
> nr_cpu_ids
)
4276 set
->nr_hw_queues
= nr_cpu_ids
;
4278 if (blk_mq_alloc_tag_set_tags(set
, set
->nr_hw_queues
) < 0)
4282 for (i
= 0; i
< set
->nr_maps
; i
++) {
4283 set
->map
[i
].mq_map
= kcalloc_node(nr_cpu_ids
,
4284 sizeof(set
->map
[i
].mq_map
[0]),
4285 GFP_KERNEL
, set
->numa_node
);
4286 if (!set
->map
[i
].mq_map
)
4287 goto out_free_mq_map
;
4288 set
->map
[i
].nr_queues
= is_kdump_kernel() ? 1 : set
->nr_hw_queues
;
4291 ret
= blk_mq_update_queue_map(set
);
4293 goto out_free_mq_map
;
4295 ret
= blk_mq_alloc_set_map_and_rqs(set
);
4297 goto out_free_mq_map
;
4299 mutex_init(&set
->tag_list_lock
);
4300 INIT_LIST_HEAD(&set
->tag_list
);
4305 for (i
= 0; i
< set
->nr_maps
; i
++) {
4306 kfree(set
->map
[i
].mq_map
);
4307 set
->map
[i
].mq_map
= NULL
;
4313 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
4315 /* allocate and initialize a tagset for a simple single-queue device */
4316 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set
*set
,
4317 const struct blk_mq_ops
*ops
, unsigned int queue_depth
,
4318 unsigned int set_flags
)
4320 memset(set
, 0, sizeof(*set
));
4322 set
->nr_hw_queues
= 1;
4324 set
->queue_depth
= queue_depth
;
4325 set
->numa_node
= NUMA_NO_NODE
;
4326 set
->flags
= set_flags
;
4327 return blk_mq_alloc_tag_set(set
);
4329 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set
);
4331 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
4335 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
4336 __blk_mq_free_map_and_rqs(set
, i
);
4338 if (blk_mq_is_shared_tags(set
->flags
)) {
4339 blk_mq_free_map_and_rqs(set
, set
->shared_tags
,
4340 BLK_MQ_NO_HCTX_IDX
);
4343 for (j
= 0; j
< set
->nr_maps
; j
++) {
4344 kfree(set
->map
[j
].mq_map
);
4345 set
->map
[j
].mq_map
= NULL
;
4351 EXPORT_SYMBOL(blk_mq_free_tag_set
);
4353 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
4355 struct blk_mq_tag_set
*set
= q
->tag_set
;
4356 struct blk_mq_hw_ctx
*hctx
;
4362 if (q
->nr_requests
== nr
)
4365 blk_mq_freeze_queue(q
);
4366 blk_mq_quiesce_queue(q
);
4369 queue_for_each_hw_ctx(q
, hctx
, i
) {
4373 * If we're using an MQ scheduler, just update the scheduler
4374 * queue depth. This is similar to what the old code would do.
4376 if (hctx
->sched_tags
) {
4377 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
4380 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
, nr
,
4385 if (q
->elevator
&& q
->elevator
->type
->ops
.depth_updated
)
4386 q
->elevator
->type
->ops
.depth_updated(hctx
);
4389 q
->nr_requests
= nr
;
4390 if (blk_mq_is_shared_tags(set
->flags
)) {
4392 blk_mq_tag_update_sched_shared_tags(q
);
4394 blk_mq_tag_resize_shared_tags(set
, nr
);
4398 blk_mq_unquiesce_queue(q
);
4399 blk_mq_unfreeze_queue(q
);
4405 * request_queue and elevator_type pair.
4406 * It is just used by __blk_mq_update_nr_hw_queues to cache
4407 * the elevator_type associated with a request_queue.
4409 struct blk_mq_qe_pair
{
4410 struct list_head node
;
4411 struct request_queue
*q
;
4412 struct elevator_type
*type
;
4416 * Cache the elevator_type in qe pair list and switch the
4417 * io scheduler to 'none'
4419 static bool blk_mq_elv_switch_none(struct list_head
*head
,
4420 struct request_queue
*q
)
4422 struct blk_mq_qe_pair
*qe
;
4427 qe
= kmalloc(sizeof(*qe
), GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
);
4431 INIT_LIST_HEAD(&qe
->node
);
4433 qe
->type
= q
->elevator
->type
;
4434 list_add(&qe
->node
, head
);
4436 mutex_lock(&q
->sysfs_lock
);
4438 * After elevator_switch_mq, the previous elevator_queue will be
4439 * released by elevator_release. The reference of the io scheduler
4440 * module get by elevator_get will also be put. So we need to get
4441 * a reference of the io scheduler module here to prevent it to be
4444 __module_get(qe
->type
->elevator_owner
);
4445 elevator_switch_mq(q
, NULL
);
4446 mutex_unlock(&q
->sysfs_lock
);
4451 static void blk_mq_elv_switch_back(struct list_head
*head
,
4452 struct request_queue
*q
)
4454 struct blk_mq_qe_pair
*qe
;
4455 struct elevator_type
*t
= NULL
;
4457 list_for_each_entry(qe
, head
, node
)
4466 list_del(&qe
->node
);
4469 mutex_lock(&q
->sysfs_lock
);
4470 elevator_switch_mq(q
, t
);
4471 mutex_unlock(&q
->sysfs_lock
);
4474 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
,
4477 struct request_queue
*q
;
4479 int prev_nr_hw_queues
;
4481 lockdep_assert_held(&set
->tag_list_lock
);
4483 if (set
->nr_maps
== 1 && nr_hw_queues
> nr_cpu_ids
)
4484 nr_hw_queues
= nr_cpu_ids
;
4485 if (nr_hw_queues
< 1)
4487 if (set
->nr_maps
== 1 && nr_hw_queues
== set
->nr_hw_queues
)
4490 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
4491 blk_mq_freeze_queue(q
);
4493 * Switch IO scheduler to 'none', cleaning up the data associated
4494 * with the previous scheduler. We will switch back once we are done
4495 * updating the new sw to hw queue mappings.
4497 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
4498 if (!blk_mq_elv_switch_none(&head
, q
))
4501 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
4502 blk_mq_debugfs_unregister_hctxs(q
);
4503 blk_mq_sysfs_unregister(q
);
4506 prev_nr_hw_queues
= set
->nr_hw_queues
;
4507 if (blk_mq_realloc_tag_set_tags(set
, set
->nr_hw_queues
, nr_hw_queues
) <
4511 set
->nr_hw_queues
= nr_hw_queues
;
4513 blk_mq_update_queue_map(set
);
4514 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
4515 blk_mq_realloc_hw_ctxs(set
, q
);
4516 if (q
->nr_hw_queues
!= set
->nr_hw_queues
) {
4517 int i
= prev_nr_hw_queues
;
4519 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4520 nr_hw_queues
, prev_nr_hw_queues
);
4521 for (; i
< set
->nr_hw_queues
; i
++)
4522 __blk_mq_free_map_and_rqs(set
, i
);
4524 set
->nr_hw_queues
= prev_nr_hw_queues
;
4525 blk_mq_map_queues(&set
->map
[HCTX_TYPE_DEFAULT
]);
4528 blk_mq_map_swqueue(q
);
4532 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
4533 blk_mq_sysfs_register(q
);
4534 blk_mq_debugfs_register_hctxs(q
);
4538 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
4539 blk_mq_elv_switch_back(&head
, q
);
4541 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
4542 blk_mq_unfreeze_queue(q
);
4545 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
4547 mutex_lock(&set
->tag_list_lock
);
4548 __blk_mq_update_nr_hw_queues(set
, nr_hw_queues
);
4549 mutex_unlock(&set
->tag_list_lock
);
4551 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
4553 /* Enable polling stats and return whether they were already enabled. */
4554 static bool blk_poll_stats_enable(struct request_queue
*q
)
4559 return blk_stats_alloc_enable(q
);
4562 static void blk_mq_poll_stats_start(struct request_queue
*q
)
4565 * We don't arm the callback if polling stats are not enabled or the
4566 * callback is already active.
4568 if (!q
->poll_stat
|| blk_stat_is_active(q
->poll_cb
))
4571 blk_stat_activate_msecs(q
->poll_cb
, 100);
4574 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
)
4576 struct request_queue
*q
= cb
->data
;
4579 for (bucket
= 0; bucket
< BLK_MQ_POLL_STATS_BKTS
; bucket
++) {
4580 if (cb
->stat
[bucket
].nr_samples
)
4581 q
->poll_stat
[bucket
] = cb
->stat
[bucket
];
4585 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
4588 unsigned long ret
= 0;
4592 * If stats collection isn't on, don't sleep but turn it on for
4595 if (!blk_poll_stats_enable(q
))
4599 * As an optimistic guess, use half of the mean service time
4600 * for this type of request. We can (and should) make this smarter.
4601 * For instance, if the completion latencies are tight, we can
4602 * get closer than just half the mean. This is especially
4603 * important on devices where the completion latencies are longer
4604 * than ~10 usec. We do use the stats for the relevant IO size
4605 * if available which does lead to better estimates.
4607 bucket
= blk_mq_poll_stats_bkt(rq
);
4611 if (q
->poll_stat
[bucket
].nr_samples
)
4612 ret
= (q
->poll_stat
[bucket
].mean
+ 1) / 2;
4617 static bool blk_mq_poll_hybrid(struct request_queue
*q
, blk_qc_t qc
)
4619 struct blk_mq_hw_ctx
*hctx
= blk_qc_to_hctx(q
, qc
);
4620 struct request
*rq
= blk_qc_to_rq(hctx
, qc
);
4621 struct hrtimer_sleeper hs
;
4622 enum hrtimer_mode mode
;
4627 * If a request has completed on queue that uses an I/O scheduler, we
4628 * won't get back a request from blk_qc_to_rq.
4630 if (!rq
|| (rq
->rq_flags
& RQF_MQ_POLL_SLEPT
))
4634 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
4636 * 0: use half of prev avg
4637 * >0: use this specific value
4639 if (q
->poll_nsec
> 0)
4640 nsecs
= q
->poll_nsec
;
4642 nsecs
= blk_mq_poll_nsecs(q
, rq
);
4647 rq
->rq_flags
|= RQF_MQ_POLL_SLEPT
;
4650 * This will be replaced with the stats tracking code, using
4651 * 'avg_completion_time / 2' as the pre-sleep target.
4655 mode
= HRTIMER_MODE_REL
;
4656 hrtimer_init_sleeper_on_stack(&hs
, CLOCK_MONOTONIC
, mode
);
4657 hrtimer_set_expires(&hs
.timer
, kt
);
4660 if (blk_mq_rq_state(rq
) == MQ_RQ_COMPLETE
)
4662 set_current_state(TASK_UNINTERRUPTIBLE
);
4663 hrtimer_sleeper_start_expires(&hs
, mode
);
4666 hrtimer_cancel(&hs
.timer
);
4667 mode
= HRTIMER_MODE_ABS
;
4668 } while (hs
.task
&& !signal_pending(current
));
4670 __set_current_state(TASK_RUNNING
);
4671 destroy_hrtimer_on_stack(&hs
.timer
);
4674 * If we sleep, have the caller restart the poll loop to reset the
4675 * state. Like for the other success return cases, the caller is
4676 * responsible for checking if the IO completed. If the IO isn't
4677 * complete, we'll get called again and will go straight to the busy
4683 static int blk_mq_poll_classic(struct request_queue
*q
, blk_qc_t cookie
,
4684 struct io_comp_batch
*iob
, unsigned int flags
)
4686 struct blk_mq_hw_ctx
*hctx
= blk_qc_to_hctx(q
, cookie
);
4687 long state
= get_current_state();
4691 ret
= q
->mq_ops
->poll(hctx
, iob
);
4693 __set_current_state(TASK_RUNNING
);
4697 if (signal_pending_state(state
, current
))
4698 __set_current_state(TASK_RUNNING
);
4699 if (task_is_running(current
))
4702 if (ret
< 0 || (flags
& BLK_POLL_ONESHOT
))
4705 } while (!need_resched());
4707 __set_current_state(TASK_RUNNING
);
4711 int blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
, struct io_comp_batch
*iob
,
4714 if (!(flags
& BLK_POLL_NOSLEEP
) &&
4715 q
->poll_nsec
!= BLK_MQ_POLL_CLASSIC
) {
4716 if (blk_mq_poll_hybrid(q
, cookie
))
4719 return blk_mq_poll_classic(q
, cookie
, iob
, flags
);
4722 unsigned int blk_mq_rq_cpu(struct request
*rq
)
4724 return rq
->mq_ctx
->cpu
;
4726 EXPORT_SYMBOL(blk_mq_rq_cpu
);
4728 void blk_mq_cancel_work_sync(struct request_queue
*q
)
4730 if (queue_is_mq(q
)) {
4731 struct blk_mq_hw_ctx
*hctx
;
4734 cancel_delayed_work_sync(&q
->requeue_work
);
4736 queue_for_each_hw_ctx(q
, hctx
, i
)
4737 cancel_delayed_work_sync(&hctx
->run_work
);
4741 static int __init
blk_mq_init(void)
4745 for_each_possible_cpu(i
)
4746 init_llist_head(&per_cpu(blk_cpu_done
, i
));
4747 open_softirq(BLOCK_SOFTIRQ
, blk_done_softirq
);
4749 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD
,
4750 "block/softirq:dead", NULL
,
4751 blk_softirq_cpu_dead
);
4752 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
4753 blk_mq_hctx_notify_dead
);
4754 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE
, "block/mq:online",
4755 blk_mq_hctx_notify_online
,
4756 blk_mq_hctx_notify_offline
);
4759 subsys_initcall(blk_mq_init
);