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/kmemleak.h>
15 #include <linux/init.h>
16 #include <linux/slab.h>
17 #include <linux/workqueue.h>
18 #include <linux/smp.h>
19 #include <linux/llist.h>
20 #include <linux/list_sort.h>
21 #include <linux/cpu.h>
22 #include <linux/cache.h>
23 #include <linux/sched/sysctl.h>
24 #include <linux/sched/topology.h>
25 #include <linux/sched/signal.h>
26 #include <linux/delay.h>
27 #include <linux/crash_dump.h>
28 #include <linux/prefetch.h>
29 #include <linux/blk-crypto.h>
31 #include <trace/events/block.h>
33 #include <linux/blk-mq.h>
34 #include <linux/t10-pi.h>
37 #include "blk-mq-debugfs.h"
38 #include "blk-mq-tag.h"
41 #include "blk-mq-sched.h"
42 #include "blk-rq-qos.h"
44 static DEFINE_PER_CPU(struct llist_head
, blk_cpu_done
);
46 static void blk_mq_poll_stats_start(struct request_queue
*q
);
47 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
);
49 static int blk_mq_poll_stats_bkt(const struct request
*rq
)
51 int ddir
, sectors
, bucket
;
53 ddir
= rq_data_dir(rq
);
54 sectors
= blk_rq_stats_sectors(rq
);
56 bucket
= ddir
+ 2 * ilog2(sectors
);
60 else if (bucket
>= BLK_MQ_POLL_STATS_BKTS
)
61 return ddir
+ BLK_MQ_POLL_STATS_BKTS
- 2;
67 * Check if any of the ctx, dispatch list or elevator
68 * have pending work in this hardware queue.
70 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
72 return !list_empty_careful(&hctx
->dispatch
) ||
73 sbitmap_any_bit_set(&hctx
->ctx_map
) ||
74 blk_mq_sched_has_work(hctx
);
78 * Mark this ctx as having pending work in this hardware queue
80 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
81 struct blk_mq_ctx
*ctx
)
83 const int bit
= ctx
->index_hw
[hctx
->type
];
85 if (!sbitmap_test_bit(&hctx
->ctx_map
, bit
))
86 sbitmap_set_bit(&hctx
->ctx_map
, bit
);
89 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
90 struct blk_mq_ctx
*ctx
)
92 const int bit
= ctx
->index_hw
[hctx
->type
];
94 sbitmap_clear_bit(&hctx
->ctx_map
, bit
);
98 struct block_device
*part
;
99 unsigned int inflight
[2];
102 static bool blk_mq_check_inflight(struct blk_mq_hw_ctx
*hctx
,
103 struct request
*rq
, void *priv
,
106 struct mq_inflight
*mi
= priv
;
108 if ((!mi
->part
->bd_partno
|| rq
->part
== mi
->part
) &&
109 blk_mq_rq_state(rq
) == MQ_RQ_IN_FLIGHT
)
110 mi
->inflight
[rq_data_dir(rq
)]++;
115 unsigned int blk_mq_in_flight(struct request_queue
*q
,
116 struct block_device
*part
)
118 struct mq_inflight mi
= { .part
= part
};
120 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight
, &mi
);
122 return mi
.inflight
[0] + mi
.inflight
[1];
125 void blk_mq_in_flight_rw(struct request_queue
*q
, struct block_device
*part
,
126 unsigned int inflight
[2])
128 struct mq_inflight mi
= { .part
= part
};
130 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight
, &mi
);
131 inflight
[0] = mi
.inflight
[0];
132 inflight
[1] = mi
.inflight
[1];
135 void blk_freeze_queue_start(struct request_queue
*q
)
137 mutex_lock(&q
->mq_freeze_lock
);
138 if (++q
->mq_freeze_depth
== 1) {
139 percpu_ref_kill(&q
->q_usage_counter
);
140 mutex_unlock(&q
->mq_freeze_lock
);
142 blk_mq_run_hw_queues(q
, false);
144 mutex_unlock(&q
->mq_freeze_lock
);
147 EXPORT_SYMBOL_GPL(blk_freeze_queue_start
);
149 void blk_mq_freeze_queue_wait(struct request_queue
*q
)
151 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
153 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait
);
155 int blk_mq_freeze_queue_wait_timeout(struct request_queue
*q
,
156 unsigned long timeout
)
158 return wait_event_timeout(q
->mq_freeze_wq
,
159 percpu_ref_is_zero(&q
->q_usage_counter
),
162 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout
);
165 * Guarantee no request is in use, so we can change any data structure of
166 * the queue afterward.
168 void blk_freeze_queue(struct request_queue
*q
)
171 * In the !blk_mq case we are only calling this to kill the
172 * q_usage_counter, otherwise this increases the freeze depth
173 * and waits for it to return to zero. For this reason there is
174 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
175 * exported to drivers as the only user for unfreeze is blk_mq.
177 blk_freeze_queue_start(q
);
178 blk_mq_freeze_queue_wait(q
);
181 void blk_mq_freeze_queue(struct request_queue
*q
)
184 * ...just an alias to keep freeze and unfreeze actions balanced
185 * in the blk_mq_* namespace
189 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
191 void blk_mq_unfreeze_queue(struct request_queue
*q
)
193 mutex_lock(&q
->mq_freeze_lock
);
194 q
->mq_freeze_depth
--;
195 WARN_ON_ONCE(q
->mq_freeze_depth
< 0);
196 if (!q
->mq_freeze_depth
) {
197 percpu_ref_resurrect(&q
->q_usage_counter
);
198 wake_up_all(&q
->mq_freeze_wq
);
200 mutex_unlock(&q
->mq_freeze_lock
);
202 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
205 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
206 * mpt3sas driver such that this function can be removed.
208 void blk_mq_quiesce_queue_nowait(struct request_queue
*q
)
210 blk_queue_flag_set(QUEUE_FLAG_QUIESCED
, q
);
212 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait
);
215 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
218 * Note: this function does not prevent that the struct request end_io()
219 * callback function is invoked. Once this function is returned, we make
220 * sure no dispatch can happen until the queue is unquiesced via
221 * blk_mq_unquiesce_queue().
223 void blk_mq_quiesce_queue(struct request_queue
*q
)
225 struct blk_mq_hw_ctx
*hctx
;
229 blk_mq_quiesce_queue_nowait(q
);
231 queue_for_each_hw_ctx(q
, hctx
, i
) {
232 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
233 synchronize_srcu(hctx
->srcu
);
240 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
243 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
246 * This function recovers queue into the state before quiescing
247 * which is done by blk_mq_quiesce_queue.
249 void blk_mq_unquiesce_queue(struct request_queue
*q
)
251 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED
, q
);
253 /* dispatch requests which are inserted during quiescing */
254 blk_mq_run_hw_queues(q
, true);
256 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue
);
258 void blk_mq_wake_waiters(struct request_queue
*q
)
260 struct blk_mq_hw_ctx
*hctx
;
263 queue_for_each_hw_ctx(q
, hctx
, i
)
264 if (blk_mq_hw_queue_mapped(hctx
))
265 blk_mq_tag_wakeup_all(hctx
->tags
, true);
269 * Only need start/end time stamping if we have iostat or
270 * blk stats enabled, or using an IO scheduler.
272 static inline bool blk_mq_need_time_stamp(struct request
*rq
)
274 return (rq
->rq_flags
& (RQF_IO_STAT
| RQF_STATS
)) || rq
->q
->elevator
;
277 static struct request
*blk_mq_rq_ctx_init(struct blk_mq_alloc_data
*data
,
278 unsigned int tag
, u64 alloc_time_ns
)
280 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
281 struct request
*rq
= tags
->static_rqs
[tag
];
283 if (data
->q
->elevator
) {
284 rq
->tag
= BLK_MQ_NO_TAG
;
285 rq
->internal_tag
= tag
;
288 rq
->internal_tag
= BLK_MQ_NO_TAG
;
291 /* csd/requeue_work/fifo_time is initialized before use */
293 rq
->mq_ctx
= data
->ctx
;
294 rq
->mq_hctx
= data
->hctx
;
296 rq
->cmd_flags
= data
->cmd_flags
;
297 if (data
->flags
& BLK_MQ_REQ_PM
)
298 rq
->rq_flags
|= RQF_PM
;
299 if (blk_queue_io_stat(data
->q
))
300 rq
->rq_flags
|= RQF_IO_STAT
;
301 INIT_LIST_HEAD(&rq
->queuelist
);
302 INIT_HLIST_NODE(&rq
->hash
);
303 RB_CLEAR_NODE(&rq
->rb_node
);
306 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
307 rq
->alloc_time_ns
= alloc_time_ns
;
309 if (blk_mq_need_time_stamp(rq
))
310 rq
->start_time_ns
= ktime_get_ns();
312 rq
->start_time_ns
= 0;
313 rq
->io_start_time_ns
= 0;
314 rq
->stats_sectors
= 0;
315 rq
->nr_phys_segments
= 0;
316 #if defined(CONFIG_BLK_DEV_INTEGRITY)
317 rq
->nr_integrity_segments
= 0;
319 blk_crypto_rq_set_defaults(rq
);
320 /* tag was already set */
321 WRITE_ONCE(rq
->deadline
, 0);
326 rq
->end_io_data
= NULL
;
328 data
->ctx
->rq_dispatched
[op_is_sync(data
->cmd_flags
)]++;
329 refcount_set(&rq
->ref
, 1);
331 if (!op_is_flush(data
->cmd_flags
)) {
332 struct elevator_queue
*e
= data
->q
->elevator
;
335 if (e
&& e
->type
->ops
.prepare_request
) {
336 if (e
->type
->icq_cache
)
337 blk_mq_sched_assign_ioc(rq
);
339 e
->type
->ops
.prepare_request(rq
);
340 rq
->rq_flags
|= RQF_ELVPRIV
;
344 data
->hctx
->queued
++;
348 static struct request
*__blk_mq_alloc_request(struct blk_mq_alloc_data
*data
)
350 struct request_queue
*q
= data
->q
;
351 struct elevator_queue
*e
= q
->elevator
;
352 u64 alloc_time_ns
= 0;
355 /* alloc_time includes depth and tag waits */
356 if (blk_queue_rq_alloc_time(q
))
357 alloc_time_ns
= ktime_get_ns();
359 if (data
->cmd_flags
& REQ_NOWAIT
)
360 data
->flags
|= BLK_MQ_REQ_NOWAIT
;
364 * Flush requests are special and go directly to the
365 * dispatch list. Don't include reserved tags in the
366 * limiting, as it isn't useful.
368 if (!op_is_flush(data
->cmd_flags
) &&
369 e
->type
->ops
.limit_depth
&&
370 !(data
->flags
& BLK_MQ_REQ_RESERVED
))
371 e
->type
->ops
.limit_depth(data
->cmd_flags
, data
);
375 data
->ctx
= blk_mq_get_ctx(q
);
376 data
->hctx
= blk_mq_map_queue(q
, data
->cmd_flags
, data
->ctx
);
378 blk_mq_tag_busy(data
->hctx
);
381 * Waiting allocations only fail because of an inactive hctx. In that
382 * case just retry the hctx assignment and tag allocation as CPU hotplug
383 * should have migrated us to an online CPU by now.
385 tag
= blk_mq_get_tag(data
);
386 if (tag
== BLK_MQ_NO_TAG
) {
387 if (data
->flags
& BLK_MQ_REQ_NOWAIT
)
391 * Give up the CPU and sleep for a random short time to ensure
392 * that thread using a realtime scheduling class are migrated
393 * off the CPU, and thus off the hctx that is going away.
398 return blk_mq_rq_ctx_init(data
, tag
, alloc_time_ns
);
401 struct request
*blk_mq_alloc_request(struct request_queue
*q
, unsigned int op
,
402 blk_mq_req_flags_t flags
)
404 struct blk_mq_alloc_data data
= {
412 ret
= blk_queue_enter(q
, flags
);
416 rq
= __blk_mq_alloc_request(&data
);
420 rq
->__sector
= (sector_t
) -1;
421 rq
->bio
= rq
->biotail
= NULL
;
425 return ERR_PTR(-EWOULDBLOCK
);
427 EXPORT_SYMBOL(blk_mq_alloc_request
);
429 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
,
430 unsigned int op
, blk_mq_req_flags_t flags
, unsigned int hctx_idx
)
432 struct blk_mq_alloc_data data
= {
437 u64 alloc_time_ns
= 0;
442 /* alloc_time includes depth and tag waits */
443 if (blk_queue_rq_alloc_time(q
))
444 alloc_time_ns
= ktime_get_ns();
447 * If the tag allocator sleeps we could get an allocation for a
448 * different hardware context. No need to complicate the low level
449 * allocator for this for the rare use case of a command tied to
452 if (WARN_ON_ONCE(!(flags
& (BLK_MQ_REQ_NOWAIT
| BLK_MQ_REQ_RESERVED
))))
453 return ERR_PTR(-EINVAL
);
455 if (hctx_idx
>= q
->nr_hw_queues
)
456 return ERR_PTR(-EIO
);
458 ret
= blk_queue_enter(q
, flags
);
463 * Check if the hardware context is actually mapped to anything.
464 * If not tell the caller that it should skip this queue.
467 data
.hctx
= q
->queue_hw_ctx
[hctx_idx
];
468 if (!blk_mq_hw_queue_mapped(data
.hctx
))
470 cpu
= cpumask_first_and(data
.hctx
->cpumask
, cpu_online_mask
);
471 data
.ctx
= __blk_mq_get_ctx(q
, cpu
);
474 blk_mq_tag_busy(data
.hctx
);
477 tag
= blk_mq_get_tag(&data
);
478 if (tag
== BLK_MQ_NO_TAG
)
480 return blk_mq_rq_ctx_init(&data
, tag
, alloc_time_ns
);
486 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
488 static void __blk_mq_free_request(struct request
*rq
)
490 struct request_queue
*q
= rq
->q
;
491 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
492 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
493 const int sched_tag
= rq
->internal_tag
;
495 blk_crypto_free_request(rq
);
496 blk_pm_mark_last_busy(rq
);
498 if (rq
->tag
!= BLK_MQ_NO_TAG
)
499 blk_mq_put_tag(hctx
->tags
, ctx
, rq
->tag
);
500 if (sched_tag
!= BLK_MQ_NO_TAG
)
501 blk_mq_put_tag(hctx
->sched_tags
, ctx
, sched_tag
);
502 blk_mq_sched_restart(hctx
);
506 void blk_mq_free_request(struct request
*rq
)
508 struct request_queue
*q
= rq
->q
;
509 struct elevator_queue
*e
= q
->elevator
;
510 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
511 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
513 if (rq
->rq_flags
& RQF_ELVPRIV
) {
514 if (e
&& e
->type
->ops
.finish_request
)
515 e
->type
->ops
.finish_request(rq
);
517 put_io_context(rq
->elv
.icq
->ioc
);
522 ctx
->rq_completed
[rq_is_sync(rq
)]++;
523 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
524 __blk_mq_dec_active_requests(hctx
);
526 if (unlikely(laptop_mode
&& !blk_rq_is_passthrough(rq
)))
527 laptop_io_completion(q
->backing_dev_info
);
531 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
532 if (refcount_dec_and_test(&rq
->ref
))
533 __blk_mq_free_request(rq
);
535 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
537 inline void __blk_mq_end_request(struct request
*rq
, blk_status_t error
)
541 if (blk_mq_need_time_stamp(rq
))
542 now
= ktime_get_ns();
544 if (rq
->rq_flags
& RQF_STATS
) {
545 blk_mq_poll_stats_start(rq
->q
);
546 blk_stat_add(rq
, now
);
549 blk_mq_sched_completed_request(rq
, now
);
551 blk_account_io_done(rq
, now
);
554 rq_qos_done(rq
->q
, rq
);
555 rq
->end_io(rq
, error
);
557 blk_mq_free_request(rq
);
560 EXPORT_SYMBOL(__blk_mq_end_request
);
562 void blk_mq_end_request(struct request
*rq
, blk_status_t error
)
564 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
566 __blk_mq_end_request(rq
, error
);
568 EXPORT_SYMBOL(blk_mq_end_request
);
570 static void blk_complete_reqs(struct llist_head
*list
)
572 struct llist_node
*entry
= llist_reverse_order(llist_del_all(list
));
573 struct request
*rq
, *next
;
575 llist_for_each_entry_safe(rq
, next
, entry
, ipi_list
)
576 rq
->q
->mq_ops
->complete(rq
);
579 static __latent_entropy
void blk_done_softirq(struct softirq_action
*h
)
581 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done
));
584 static int blk_softirq_cpu_dead(unsigned int cpu
)
586 blk_complete_reqs(&per_cpu(blk_cpu_done
, cpu
));
590 static void __blk_mq_complete_request_remote(void *data
)
592 __raise_softirq_irqoff(BLOCK_SOFTIRQ
);
595 static inline bool blk_mq_complete_need_ipi(struct request
*rq
)
597 int cpu
= raw_smp_processor_id();
599 if (!IS_ENABLED(CONFIG_SMP
) ||
600 !test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
))
603 * With force threaded interrupts enabled, raising softirq from an SMP
604 * function call will always result in waking the ksoftirqd thread.
605 * This is probably worse than completing the request on a different
608 if (force_irqthreads
)
611 /* same CPU or cache domain? Complete locally */
612 if (cpu
== rq
->mq_ctx
->cpu
||
613 (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
) &&
614 cpus_share_cache(cpu
, rq
->mq_ctx
->cpu
)))
617 /* don't try to IPI to an offline CPU */
618 return cpu_online(rq
->mq_ctx
->cpu
);
621 static void blk_mq_complete_send_ipi(struct request
*rq
)
623 struct llist_head
*list
;
626 cpu
= rq
->mq_ctx
->cpu
;
627 list
= &per_cpu(blk_cpu_done
, cpu
);
628 if (llist_add(&rq
->ipi_list
, list
)) {
629 INIT_CSD(&rq
->csd
, __blk_mq_complete_request_remote
, rq
);
630 smp_call_function_single_async(cpu
, &rq
->csd
);
634 static void blk_mq_raise_softirq(struct request
*rq
)
636 struct llist_head
*list
;
639 list
= this_cpu_ptr(&blk_cpu_done
);
640 if (llist_add(&rq
->ipi_list
, list
))
641 raise_softirq(BLOCK_SOFTIRQ
);
645 bool blk_mq_complete_request_remote(struct request
*rq
)
647 WRITE_ONCE(rq
->state
, MQ_RQ_COMPLETE
);
650 * For a polled request, always complete locallly, it's pointless
651 * to redirect the completion.
653 if (rq
->cmd_flags
& REQ_HIPRI
)
656 if (blk_mq_complete_need_ipi(rq
)) {
657 blk_mq_complete_send_ipi(rq
);
661 if (rq
->q
->nr_hw_queues
== 1) {
662 blk_mq_raise_softirq(rq
);
667 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote
);
670 * blk_mq_complete_request - end I/O on a request
671 * @rq: the request being processed
674 * Complete a request by scheduling the ->complete_rq operation.
676 void blk_mq_complete_request(struct request
*rq
)
678 if (!blk_mq_complete_request_remote(rq
))
679 rq
->q
->mq_ops
->complete(rq
);
681 EXPORT_SYMBOL(blk_mq_complete_request
);
683 static void hctx_unlock(struct blk_mq_hw_ctx
*hctx
, int srcu_idx
)
684 __releases(hctx
->srcu
)
686 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
))
689 srcu_read_unlock(hctx
->srcu
, srcu_idx
);
692 static void hctx_lock(struct blk_mq_hw_ctx
*hctx
, int *srcu_idx
)
693 __acquires(hctx
->srcu
)
695 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
696 /* shut up gcc false positive */
700 *srcu_idx
= srcu_read_lock(hctx
->srcu
);
704 * blk_mq_start_request - Start processing a request
705 * @rq: Pointer to request to be started
707 * Function used by device drivers to notify the block layer that a request
708 * is going to be processed now, so blk layer can do proper initializations
709 * such as starting the timeout timer.
711 void blk_mq_start_request(struct request
*rq
)
713 struct request_queue
*q
= rq
->q
;
715 trace_block_rq_issue(rq
);
717 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
718 rq
->io_start_time_ns
= ktime_get_ns();
719 rq
->stats_sectors
= blk_rq_sectors(rq
);
720 rq
->rq_flags
|= RQF_STATS
;
724 WARN_ON_ONCE(blk_mq_rq_state(rq
) != MQ_RQ_IDLE
);
727 WRITE_ONCE(rq
->state
, MQ_RQ_IN_FLIGHT
);
729 #ifdef CONFIG_BLK_DEV_INTEGRITY
730 if (blk_integrity_rq(rq
) && req_op(rq
) == REQ_OP_WRITE
)
731 q
->integrity
.profile
->prepare_fn(rq
);
734 EXPORT_SYMBOL(blk_mq_start_request
);
736 static void __blk_mq_requeue_request(struct request
*rq
)
738 struct request_queue
*q
= rq
->q
;
740 blk_mq_put_driver_tag(rq
);
742 trace_block_rq_requeue(rq
);
743 rq_qos_requeue(q
, rq
);
745 if (blk_mq_request_started(rq
)) {
746 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
747 rq
->rq_flags
&= ~RQF_TIMED_OUT
;
751 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
753 __blk_mq_requeue_request(rq
);
755 /* this request will be re-inserted to io scheduler queue */
756 blk_mq_sched_requeue_request(rq
);
758 BUG_ON(!list_empty(&rq
->queuelist
));
759 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
761 EXPORT_SYMBOL(blk_mq_requeue_request
);
763 static void blk_mq_requeue_work(struct work_struct
*work
)
765 struct request_queue
*q
=
766 container_of(work
, struct request_queue
, requeue_work
.work
);
768 struct request
*rq
, *next
;
770 spin_lock_irq(&q
->requeue_lock
);
771 list_splice_init(&q
->requeue_list
, &rq_list
);
772 spin_unlock_irq(&q
->requeue_lock
);
774 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
775 if (!(rq
->rq_flags
& (RQF_SOFTBARRIER
| RQF_DONTPREP
)))
778 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
779 list_del_init(&rq
->queuelist
);
781 * If RQF_DONTPREP, rq has contained some driver specific
782 * data, so insert it to hctx dispatch list to avoid any
785 if (rq
->rq_flags
& RQF_DONTPREP
)
786 blk_mq_request_bypass_insert(rq
, false, false);
788 blk_mq_sched_insert_request(rq
, true, false, false);
791 while (!list_empty(&rq_list
)) {
792 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
793 list_del_init(&rq
->queuelist
);
794 blk_mq_sched_insert_request(rq
, false, false, false);
797 blk_mq_run_hw_queues(q
, false);
800 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
801 bool kick_requeue_list
)
803 struct request_queue
*q
= rq
->q
;
807 * We abuse this flag that is otherwise used by the I/O scheduler to
808 * request head insertion from the workqueue.
810 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
812 spin_lock_irqsave(&q
->requeue_lock
, flags
);
814 rq
->rq_flags
|= RQF_SOFTBARRIER
;
815 list_add(&rq
->queuelist
, &q
->requeue_list
);
817 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
819 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
821 if (kick_requeue_list
)
822 blk_mq_kick_requeue_list(q
);
825 void blk_mq_kick_requeue_list(struct request_queue
*q
)
827 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
, 0);
829 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
831 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
834 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
,
835 msecs_to_jiffies(msecs
));
837 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
839 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
841 if (tag
< tags
->nr_tags
) {
842 prefetch(tags
->rqs
[tag
]);
843 return tags
->rqs
[tag
];
848 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
850 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
851 void *priv
, bool reserved
)
854 * If we find a request that isn't idle and the queue matches,
855 * we know the queue is busy. Return false to stop the iteration.
857 if (blk_mq_request_started(rq
) && rq
->q
== hctx
->queue
) {
867 bool blk_mq_queue_inflight(struct request_queue
*q
)
871 blk_mq_queue_tag_busy_iter(q
, blk_mq_rq_inflight
, &busy
);
874 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight
);
876 static void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
878 req
->rq_flags
|= RQF_TIMED_OUT
;
879 if (req
->q
->mq_ops
->timeout
) {
880 enum blk_eh_timer_return ret
;
882 ret
= req
->q
->mq_ops
->timeout(req
, reserved
);
883 if (ret
== BLK_EH_DONE
)
885 WARN_ON_ONCE(ret
!= BLK_EH_RESET_TIMER
);
891 static bool blk_mq_req_expired(struct request
*rq
, unsigned long *next
)
893 unsigned long deadline
;
895 if (blk_mq_rq_state(rq
) != MQ_RQ_IN_FLIGHT
)
897 if (rq
->rq_flags
& RQF_TIMED_OUT
)
900 deadline
= READ_ONCE(rq
->deadline
);
901 if (time_after_eq(jiffies
, deadline
))
906 else if (time_after(*next
, deadline
))
911 static bool blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
912 struct request
*rq
, void *priv
, bool reserved
)
914 unsigned long *next
= priv
;
917 * Just do a quick check if it is expired before locking the request in
918 * so we're not unnecessarilly synchronizing across CPUs.
920 if (!blk_mq_req_expired(rq
, next
))
924 * We have reason to believe the request may be expired. Take a
925 * reference on the request to lock this request lifetime into its
926 * currently allocated context to prevent it from being reallocated in
927 * the event the completion by-passes this timeout handler.
929 * If the reference was already released, then the driver beat the
930 * timeout handler to posting a natural completion.
932 if (!refcount_inc_not_zero(&rq
->ref
))
936 * The request is now locked and cannot be reallocated underneath the
937 * timeout handler's processing. Re-verify this exact request is truly
938 * expired; if it is not expired, then the request was completed and
939 * reallocated as a new request.
941 if (blk_mq_req_expired(rq
, next
))
942 blk_mq_rq_timed_out(rq
, reserved
);
944 if (is_flush_rq(rq
, hctx
))
946 else if (refcount_dec_and_test(&rq
->ref
))
947 __blk_mq_free_request(rq
);
952 static void blk_mq_timeout_work(struct work_struct
*work
)
954 struct request_queue
*q
=
955 container_of(work
, struct request_queue
, timeout_work
);
956 unsigned long next
= 0;
957 struct blk_mq_hw_ctx
*hctx
;
960 /* A deadlock might occur if a request is stuck requiring a
961 * timeout at the same time a queue freeze is waiting
962 * completion, since the timeout code would not be able to
963 * acquire the queue reference here.
965 * That's why we don't use blk_queue_enter here; instead, we use
966 * percpu_ref_tryget directly, because we need to be able to
967 * obtain a reference even in the short window between the queue
968 * starting to freeze, by dropping the first reference in
969 * blk_freeze_queue_start, and the moment the last request is
970 * consumed, marked by the instant q_usage_counter reaches
973 if (!percpu_ref_tryget(&q
->q_usage_counter
))
976 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &next
);
979 mod_timer(&q
->timeout
, next
);
982 * Request timeouts are handled as a forward rolling timer. If
983 * we end up here it means that no requests are pending and
984 * also that no request has been pending for a while. Mark
987 queue_for_each_hw_ctx(q
, hctx
, i
) {
988 /* the hctx may be unmapped, so check it here */
989 if (blk_mq_hw_queue_mapped(hctx
))
990 blk_mq_tag_idle(hctx
);
996 struct flush_busy_ctx_data
{
997 struct blk_mq_hw_ctx
*hctx
;
998 struct list_head
*list
;
1001 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
1003 struct flush_busy_ctx_data
*flush_data
= data
;
1004 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
1005 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
1006 enum hctx_type type
= hctx
->type
;
1008 spin_lock(&ctx
->lock
);
1009 list_splice_tail_init(&ctx
->rq_lists
[type
], flush_data
->list
);
1010 sbitmap_clear_bit(sb
, bitnr
);
1011 spin_unlock(&ctx
->lock
);
1016 * Process software queues that have been marked busy, splicing them
1017 * to the for-dispatch
1019 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
1021 struct flush_busy_ctx_data data
= {
1026 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
1028 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
1030 struct dispatch_rq_data
{
1031 struct blk_mq_hw_ctx
*hctx
;
1035 static bool dispatch_rq_from_ctx(struct sbitmap
*sb
, unsigned int bitnr
,
1038 struct dispatch_rq_data
*dispatch_data
= data
;
1039 struct blk_mq_hw_ctx
*hctx
= dispatch_data
->hctx
;
1040 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
1041 enum hctx_type type
= hctx
->type
;
1043 spin_lock(&ctx
->lock
);
1044 if (!list_empty(&ctx
->rq_lists
[type
])) {
1045 dispatch_data
->rq
= list_entry_rq(ctx
->rq_lists
[type
].next
);
1046 list_del_init(&dispatch_data
->rq
->queuelist
);
1047 if (list_empty(&ctx
->rq_lists
[type
]))
1048 sbitmap_clear_bit(sb
, bitnr
);
1050 spin_unlock(&ctx
->lock
);
1052 return !dispatch_data
->rq
;
1055 struct request
*blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx
*hctx
,
1056 struct blk_mq_ctx
*start
)
1058 unsigned off
= start
? start
->index_hw
[hctx
->type
] : 0;
1059 struct dispatch_rq_data data
= {
1064 __sbitmap_for_each_set(&hctx
->ctx_map
, off
,
1065 dispatch_rq_from_ctx
, &data
);
1070 static inline unsigned int queued_to_index(unsigned int queued
)
1075 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
1078 static bool __blk_mq_get_driver_tag(struct request
*rq
)
1080 struct sbitmap_queue
*bt
= rq
->mq_hctx
->tags
->bitmap_tags
;
1081 unsigned int tag_offset
= rq
->mq_hctx
->tags
->nr_reserved_tags
;
1084 blk_mq_tag_busy(rq
->mq_hctx
);
1086 if (blk_mq_tag_is_reserved(rq
->mq_hctx
->sched_tags
, rq
->internal_tag
)) {
1087 bt
= rq
->mq_hctx
->tags
->breserved_tags
;
1090 if (!hctx_may_queue(rq
->mq_hctx
, bt
))
1094 tag
= __sbitmap_queue_get(bt
);
1095 if (tag
== BLK_MQ_NO_TAG
)
1098 rq
->tag
= tag
+ tag_offset
;
1102 static bool blk_mq_get_driver_tag(struct request
*rq
)
1104 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1106 if (rq
->tag
== BLK_MQ_NO_TAG
&& !__blk_mq_get_driver_tag(rq
))
1109 if ((hctx
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
) &&
1110 !(rq
->rq_flags
& RQF_MQ_INFLIGHT
)) {
1111 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
1112 __blk_mq_inc_active_requests(hctx
);
1114 hctx
->tags
->rqs
[rq
->tag
] = rq
;
1118 static int blk_mq_dispatch_wake(wait_queue_entry_t
*wait
, unsigned mode
,
1119 int flags
, void *key
)
1121 struct blk_mq_hw_ctx
*hctx
;
1123 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
1125 spin_lock(&hctx
->dispatch_wait_lock
);
1126 if (!list_empty(&wait
->entry
)) {
1127 struct sbitmap_queue
*sbq
;
1129 list_del_init(&wait
->entry
);
1130 sbq
= hctx
->tags
->bitmap_tags
;
1131 atomic_dec(&sbq
->ws_active
);
1133 spin_unlock(&hctx
->dispatch_wait_lock
);
1135 blk_mq_run_hw_queue(hctx
, true);
1140 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1141 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1142 * restart. For both cases, take care to check the condition again after
1143 * marking us as waiting.
1145 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx
*hctx
,
1148 struct sbitmap_queue
*sbq
= hctx
->tags
->bitmap_tags
;
1149 struct wait_queue_head
*wq
;
1150 wait_queue_entry_t
*wait
;
1153 if (!(hctx
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
)) {
1154 blk_mq_sched_mark_restart_hctx(hctx
);
1157 * It's possible that a tag was freed in the window between the
1158 * allocation failure and adding the hardware queue to the wait
1161 * Don't clear RESTART here, someone else could have set it.
1162 * At most this will cost an extra queue run.
1164 return blk_mq_get_driver_tag(rq
);
1167 wait
= &hctx
->dispatch_wait
;
1168 if (!list_empty_careful(&wait
->entry
))
1171 wq
= &bt_wait_ptr(sbq
, hctx
)->wait
;
1173 spin_lock_irq(&wq
->lock
);
1174 spin_lock(&hctx
->dispatch_wait_lock
);
1175 if (!list_empty(&wait
->entry
)) {
1176 spin_unlock(&hctx
->dispatch_wait_lock
);
1177 spin_unlock_irq(&wq
->lock
);
1181 atomic_inc(&sbq
->ws_active
);
1182 wait
->flags
&= ~WQ_FLAG_EXCLUSIVE
;
1183 __add_wait_queue(wq
, wait
);
1186 * It's possible that a tag was freed in the window between the
1187 * allocation failure and adding the hardware queue to the wait
1190 ret
= blk_mq_get_driver_tag(rq
);
1192 spin_unlock(&hctx
->dispatch_wait_lock
);
1193 spin_unlock_irq(&wq
->lock
);
1198 * We got a tag, remove ourselves from the wait queue to ensure
1199 * someone else gets the wakeup.
1201 list_del_init(&wait
->entry
);
1202 atomic_dec(&sbq
->ws_active
);
1203 spin_unlock(&hctx
->dispatch_wait_lock
);
1204 spin_unlock_irq(&wq
->lock
);
1209 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1210 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1212 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1213 * - EWMA is one simple way to compute running average value
1214 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1215 * - take 4 as factor for avoiding to get too small(0) result, and this
1216 * factor doesn't matter because EWMA decreases exponentially
1218 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx
*hctx
, bool busy
)
1222 if (hctx
->queue
->elevator
)
1225 ewma
= hctx
->dispatch_busy
;
1230 ewma
*= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
- 1;
1232 ewma
+= 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR
;
1233 ewma
/= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
;
1235 hctx
->dispatch_busy
= ewma
;
1238 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1240 static void blk_mq_handle_dev_resource(struct request
*rq
,
1241 struct list_head
*list
)
1243 struct request
*next
=
1244 list_first_entry_or_null(list
, struct request
, queuelist
);
1247 * If an I/O scheduler has been configured and we got a driver tag for
1248 * the next request already, free it.
1251 blk_mq_put_driver_tag(next
);
1253 list_add(&rq
->queuelist
, list
);
1254 __blk_mq_requeue_request(rq
);
1257 static void blk_mq_handle_zone_resource(struct request
*rq
,
1258 struct list_head
*zone_list
)
1261 * If we end up here it is because we cannot dispatch a request to a
1262 * specific zone due to LLD level zone-write locking or other zone
1263 * related resource not being available. In this case, set the request
1264 * aside in zone_list for retrying it later.
1266 list_add(&rq
->queuelist
, zone_list
);
1267 __blk_mq_requeue_request(rq
);
1270 enum prep_dispatch
{
1272 PREP_DISPATCH_NO_TAG
,
1273 PREP_DISPATCH_NO_BUDGET
,
1276 static enum prep_dispatch
blk_mq_prep_dispatch_rq(struct request
*rq
,
1279 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1281 if (need_budget
&& !blk_mq_get_dispatch_budget(rq
->q
)) {
1282 blk_mq_put_driver_tag(rq
);
1283 return PREP_DISPATCH_NO_BUDGET
;
1286 if (!blk_mq_get_driver_tag(rq
)) {
1288 * The initial allocation attempt failed, so we need to
1289 * rerun the hardware queue when a tag is freed. The
1290 * waitqueue takes care of that. If the queue is run
1291 * before we add this entry back on the dispatch list,
1292 * we'll re-run it below.
1294 if (!blk_mq_mark_tag_wait(hctx
, rq
)) {
1296 * All budgets not got from this function will be put
1297 * together during handling partial dispatch
1300 blk_mq_put_dispatch_budget(rq
->q
);
1301 return PREP_DISPATCH_NO_TAG
;
1305 return PREP_DISPATCH_OK
;
1308 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1309 static void blk_mq_release_budgets(struct request_queue
*q
,
1310 unsigned int nr_budgets
)
1314 for (i
= 0; i
< nr_budgets
; i
++)
1315 blk_mq_put_dispatch_budget(q
);
1319 * Returns true if we did some work AND can potentially do more.
1321 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
,
1322 unsigned int nr_budgets
)
1324 enum prep_dispatch prep
;
1325 struct request_queue
*q
= hctx
->queue
;
1326 struct request
*rq
, *nxt
;
1328 blk_status_t ret
= BLK_STS_OK
;
1329 LIST_HEAD(zone_list
);
1331 if (list_empty(list
))
1335 * Now process all the entries, sending them to the driver.
1337 errors
= queued
= 0;
1339 struct blk_mq_queue_data bd
;
1341 rq
= list_first_entry(list
, struct request
, queuelist
);
1343 WARN_ON_ONCE(hctx
!= rq
->mq_hctx
);
1344 prep
= blk_mq_prep_dispatch_rq(rq
, !nr_budgets
);
1345 if (prep
!= PREP_DISPATCH_OK
)
1348 list_del_init(&rq
->queuelist
);
1353 * Flag last if we have no more requests, or if we have more
1354 * but can't assign a driver tag to it.
1356 if (list_empty(list
))
1359 nxt
= list_first_entry(list
, struct request
, queuelist
);
1360 bd
.last
= !blk_mq_get_driver_tag(nxt
);
1364 * once the request is queued to lld, no need to cover the
1369 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1374 case BLK_STS_RESOURCE
:
1375 case BLK_STS_DEV_RESOURCE
:
1376 blk_mq_handle_dev_resource(rq
, list
);
1378 case BLK_STS_ZONE_RESOURCE
:
1380 * Move the request to zone_list and keep going through
1381 * the dispatch list to find more requests the drive can
1384 blk_mq_handle_zone_resource(rq
, &zone_list
);
1388 blk_mq_end_request(rq
, ret
);
1390 } while (!list_empty(list
));
1392 if (!list_empty(&zone_list
))
1393 list_splice_tail_init(&zone_list
, list
);
1395 hctx
->dispatched
[queued_to_index(queued
)]++;
1397 /* If we didn't flush the entire list, we could have told the driver
1398 * there was more coming, but that turned out to be a lie.
1400 if ((!list_empty(list
) || errors
) && q
->mq_ops
->commit_rqs
&& queued
)
1401 q
->mq_ops
->commit_rqs(hctx
);
1403 * Any items that need requeuing? Stuff them into hctx->dispatch,
1404 * that is where we will continue on next queue run.
1406 if (!list_empty(list
)) {
1408 /* For non-shared tags, the RESTART check will suffice */
1409 bool no_tag
= prep
== PREP_DISPATCH_NO_TAG
&&
1410 (hctx
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
);
1411 bool no_budget_avail
= prep
== PREP_DISPATCH_NO_BUDGET
;
1413 blk_mq_release_budgets(q
, nr_budgets
);
1415 spin_lock(&hctx
->lock
);
1416 list_splice_tail_init(list
, &hctx
->dispatch
);
1417 spin_unlock(&hctx
->lock
);
1420 * Order adding requests to hctx->dispatch and checking
1421 * SCHED_RESTART flag. The pair of this smp_mb() is the one
1422 * in blk_mq_sched_restart(). Avoid restart code path to
1423 * miss the new added requests to hctx->dispatch, meantime
1424 * SCHED_RESTART is observed here.
1429 * If SCHED_RESTART was set by the caller of this function and
1430 * it is no longer set that means that it was cleared by another
1431 * thread and hence that a queue rerun is needed.
1433 * If 'no_tag' is set, that means that we failed getting
1434 * a driver tag with an I/O scheduler attached. If our dispatch
1435 * waitqueue is no longer active, ensure that we run the queue
1436 * AFTER adding our entries back to the list.
1438 * If no I/O scheduler has been configured it is possible that
1439 * the hardware queue got stopped and restarted before requests
1440 * were pushed back onto the dispatch list. Rerun the queue to
1441 * avoid starvation. Notes:
1442 * - blk_mq_run_hw_queue() checks whether or not a queue has
1443 * been stopped before rerunning a queue.
1444 * - Some but not all block drivers stop a queue before
1445 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1448 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1449 * bit is set, run queue after a delay to avoid IO stalls
1450 * that could otherwise occur if the queue is idle. We'll do
1451 * similar if we couldn't get budget and SCHED_RESTART is set.
1453 needs_restart
= blk_mq_sched_needs_restart(hctx
);
1454 if (!needs_restart
||
1455 (no_tag
&& list_empty_careful(&hctx
->dispatch_wait
.entry
)))
1456 blk_mq_run_hw_queue(hctx
, true);
1457 else if (needs_restart
&& (ret
== BLK_STS_RESOURCE
||
1459 blk_mq_delay_run_hw_queue(hctx
, BLK_MQ_RESOURCE_DELAY
);
1461 blk_mq_update_dispatch_busy(hctx
, true);
1464 blk_mq_update_dispatch_busy(hctx
, false);
1466 return (queued
+ errors
) != 0;
1470 * __blk_mq_run_hw_queue - Run a hardware queue.
1471 * @hctx: Pointer to the hardware queue to run.
1473 * Send pending requests to the hardware.
1475 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1480 * We can't run the queue inline with ints disabled. Ensure that
1481 * we catch bad users of this early.
1483 WARN_ON_ONCE(in_interrupt());
1485 might_sleep_if(hctx
->flags
& BLK_MQ_F_BLOCKING
);
1487 hctx_lock(hctx
, &srcu_idx
);
1488 blk_mq_sched_dispatch_requests(hctx
);
1489 hctx_unlock(hctx
, srcu_idx
);
1492 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx
*hctx
)
1494 int cpu
= cpumask_first_and(hctx
->cpumask
, cpu_online_mask
);
1496 if (cpu
>= nr_cpu_ids
)
1497 cpu
= cpumask_first(hctx
->cpumask
);
1502 * It'd be great if the workqueue API had a way to pass
1503 * in a mask and had some smarts for more clever placement.
1504 * For now we just round-robin here, switching for every
1505 * BLK_MQ_CPU_WORK_BATCH queued items.
1507 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1510 int next_cpu
= hctx
->next_cpu
;
1512 if (hctx
->queue
->nr_hw_queues
== 1)
1513 return WORK_CPU_UNBOUND
;
1515 if (--hctx
->next_cpu_batch
<= 0) {
1517 next_cpu
= cpumask_next_and(next_cpu
, hctx
->cpumask
,
1519 if (next_cpu
>= nr_cpu_ids
)
1520 next_cpu
= blk_mq_first_mapped_cpu(hctx
);
1521 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1525 * Do unbound schedule if we can't find a online CPU for this hctx,
1526 * and it should only happen in the path of handling CPU DEAD.
1528 if (!cpu_online(next_cpu
)) {
1535 * Make sure to re-select CPU next time once after CPUs
1536 * in hctx->cpumask become online again.
1538 hctx
->next_cpu
= next_cpu
;
1539 hctx
->next_cpu_batch
= 1;
1540 return WORK_CPU_UNBOUND
;
1543 hctx
->next_cpu
= next_cpu
;
1548 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
1549 * @hctx: Pointer to the hardware queue to run.
1550 * @async: If we want to run the queue asynchronously.
1551 * @msecs: Milliseconds of delay to wait before running the queue.
1553 * If !@async, try to run the queue now. Else, run the queue asynchronously and
1554 * with a delay of @msecs.
1556 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
,
1557 unsigned long msecs
)
1559 if (unlikely(blk_mq_hctx_stopped(hctx
)))
1562 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1563 int cpu
= get_cpu();
1564 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1565 __blk_mq_run_hw_queue(hctx
);
1573 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
,
1574 msecs_to_jiffies(msecs
));
1578 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
1579 * @hctx: Pointer to the hardware queue to run.
1580 * @msecs: Milliseconds of delay to wait before running the queue.
1582 * Run a hardware queue asynchronously with a delay of @msecs.
1584 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1586 __blk_mq_delay_run_hw_queue(hctx
, true, msecs
);
1588 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue
);
1591 * blk_mq_run_hw_queue - Start to run a hardware queue.
1592 * @hctx: Pointer to the hardware queue to run.
1593 * @async: If we want to run the queue asynchronously.
1595 * Check if the request queue is not in a quiesced state and if there are
1596 * pending requests to be sent. If this is true, run the queue to send requests
1599 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1605 * When queue is quiesced, we may be switching io scheduler, or
1606 * updating nr_hw_queues, or other things, and we can't run queue
1607 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1609 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1612 hctx_lock(hctx
, &srcu_idx
);
1613 need_run
= !blk_queue_quiesced(hctx
->queue
) &&
1614 blk_mq_hctx_has_pending(hctx
);
1615 hctx_unlock(hctx
, srcu_idx
);
1618 __blk_mq_delay_run_hw_queue(hctx
, async
, 0);
1620 EXPORT_SYMBOL(blk_mq_run_hw_queue
);
1623 * Is the request queue handled by an IO scheduler that does not respect
1624 * hardware queues when dispatching?
1626 static bool blk_mq_has_sqsched(struct request_queue
*q
)
1628 struct elevator_queue
*e
= q
->elevator
;
1630 if (e
&& e
->type
->ops
.dispatch_request
&&
1631 !(e
->type
->elevator_features
& ELEVATOR_F_MQ_AWARE
))
1637 * Return prefered queue to dispatch from (if any) for non-mq aware IO
1640 static struct blk_mq_hw_ctx
*blk_mq_get_sq_hctx(struct request_queue
*q
)
1642 struct blk_mq_hw_ctx
*hctx
;
1645 * If the IO scheduler does not respect hardware queues when
1646 * dispatching, we just don't bother with multiple HW queues and
1647 * dispatch from hctx for the current CPU since running multiple queues
1648 * just causes lock contention inside the scheduler and pointless cache
1651 hctx
= blk_mq_map_queue_type(q
, HCTX_TYPE_DEFAULT
,
1652 raw_smp_processor_id());
1653 if (!blk_mq_hctx_stopped(hctx
))
1659 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
1660 * @q: Pointer to the request queue to run.
1661 * @async: If we want to run the queue asynchronously.
1663 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1665 struct blk_mq_hw_ctx
*hctx
, *sq_hctx
;
1669 if (blk_mq_has_sqsched(q
))
1670 sq_hctx
= blk_mq_get_sq_hctx(q
);
1671 queue_for_each_hw_ctx(q
, hctx
, i
) {
1672 if (blk_mq_hctx_stopped(hctx
))
1675 * Dispatch from this hctx either if there's no hctx preferred
1676 * by IO scheduler or if it has requests that bypass the
1679 if (!sq_hctx
|| sq_hctx
== hctx
||
1680 !list_empty_careful(&hctx
->dispatch
))
1681 blk_mq_run_hw_queue(hctx
, async
);
1684 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1687 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
1688 * @q: Pointer to the request queue to run.
1689 * @msecs: Milliseconds of delay to wait before running the queues.
1691 void blk_mq_delay_run_hw_queues(struct request_queue
*q
, unsigned long msecs
)
1693 struct blk_mq_hw_ctx
*hctx
, *sq_hctx
;
1697 if (blk_mq_has_sqsched(q
))
1698 sq_hctx
= blk_mq_get_sq_hctx(q
);
1699 queue_for_each_hw_ctx(q
, hctx
, i
) {
1700 if (blk_mq_hctx_stopped(hctx
))
1703 * Dispatch from this hctx either if there's no hctx preferred
1704 * by IO scheduler or if it has requests that bypass the
1707 if (!sq_hctx
|| sq_hctx
== hctx
||
1708 !list_empty_careful(&hctx
->dispatch
))
1709 blk_mq_delay_run_hw_queue(hctx
, msecs
);
1712 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues
);
1715 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1716 * @q: request queue.
1718 * The caller is responsible for serializing this function against
1719 * blk_mq_{start,stop}_hw_queue().
1721 bool blk_mq_queue_stopped(struct request_queue
*q
)
1723 struct blk_mq_hw_ctx
*hctx
;
1726 queue_for_each_hw_ctx(q
, hctx
, i
)
1727 if (blk_mq_hctx_stopped(hctx
))
1732 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1735 * This function is often used for pausing .queue_rq() by driver when
1736 * there isn't enough resource or some conditions aren't satisfied, and
1737 * BLK_STS_RESOURCE is usually returned.
1739 * We do not guarantee that dispatch can be drained or blocked
1740 * after blk_mq_stop_hw_queue() returns. Please use
1741 * blk_mq_quiesce_queue() for that requirement.
1743 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1745 cancel_delayed_work(&hctx
->run_work
);
1747 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1749 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1752 * This function is often used for pausing .queue_rq() by driver when
1753 * there isn't enough resource or some conditions aren't satisfied, and
1754 * BLK_STS_RESOURCE is usually returned.
1756 * We do not guarantee that dispatch can be drained or blocked
1757 * after blk_mq_stop_hw_queues() returns. Please use
1758 * blk_mq_quiesce_queue() for that requirement.
1760 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1762 struct blk_mq_hw_ctx
*hctx
;
1765 queue_for_each_hw_ctx(q
, hctx
, i
)
1766 blk_mq_stop_hw_queue(hctx
);
1768 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1770 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1772 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1774 blk_mq_run_hw_queue(hctx
, false);
1776 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1778 void blk_mq_start_hw_queues(struct request_queue
*q
)
1780 struct blk_mq_hw_ctx
*hctx
;
1783 queue_for_each_hw_ctx(q
, hctx
, i
)
1784 blk_mq_start_hw_queue(hctx
);
1786 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1788 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1790 if (!blk_mq_hctx_stopped(hctx
))
1793 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1794 blk_mq_run_hw_queue(hctx
, async
);
1796 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1798 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1800 struct blk_mq_hw_ctx
*hctx
;
1803 queue_for_each_hw_ctx(q
, hctx
, i
)
1804 blk_mq_start_stopped_hw_queue(hctx
, async
);
1806 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1808 static void blk_mq_run_work_fn(struct work_struct
*work
)
1810 struct blk_mq_hw_ctx
*hctx
;
1812 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
1815 * If we are stopped, don't run the queue.
1817 if (blk_mq_hctx_stopped(hctx
))
1820 __blk_mq_run_hw_queue(hctx
);
1823 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1827 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1828 enum hctx_type type
= hctx
->type
;
1830 lockdep_assert_held(&ctx
->lock
);
1832 trace_block_rq_insert(rq
);
1835 list_add(&rq
->queuelist
, &ctx
->rq_lists
[type
]);
1837 list_add_tail(&rq
->queuelist
, &ctx
->rq_lists
[type
]);
1840 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1843 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1845 lockdep_assert_held(&ctx
->lock
);
1847 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1848 blk_mq_hctx_mark_pending(hctx
, ctx
);
1852 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
1853 * @rq: Pointer to request to be inserted.
1854 * @at_head: true if the request should be inserted at the head of the list.
1855 * @run_queue: If we should run the hardware queue after inserting the request.
1857 * Should only be used carefully, when the caller knows we want to
1858 * bypass a potential IO scheduler on the target device.
1860 void blk_mq_request_bypass_insert(struct request
*rq
, bool at_head
,
1863 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1865 spin_lock(&hctx
->lock
);
1867 list_add(&rq
->queuelist
, &hctx
->dispatch
);
1869 list_add_tail(&rq
->queuelist
, &hctx
->dispatch
);
1870 spin_unlock(&hctx
->lock
);
1873 blk_mq_run_hw_queue(hctx
, false);
1876 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1877 struct list_head
*list
)
1881 enum hctx_type type
= hctx
->type
;
1884 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1887 list_for_each_entry(rq
, list
, queuelist
) {
1888 BUG_ON(rq
->mq_ctx
!= ctx
);
1889 trace_block_rq_insert(rq
);
1892 spin_lock(&ctx
->lock
);
1893 list_splice_tail_init(list
, &ctx
->rq_lists
[type
]);
1894 blk_mq_hctx_mark_pending(hctx
, ctx
);
1895 spin_unlock(&ctx
->lock
);
1898 static int plug_rq_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1900 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1901 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1903 if (rqa
->mq_ctx
!= rqb
->mq_ctx
)
1904 return rqa
->mq_ctx
> rqb
->mq_ctx
;
1905 if (rqa
->mq_hctx
!= rqb
->mq_hctx
)
1906 return rqa
->mq_hctx
> rqb
->mq_hctx
;
1908 return blk_rq_pos(rqa
) > blk_rq_pos(rqb
);
1911 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1915 if (list_empty(&plug
->mq_list
))
1917 list_splice_init(&plug
->mq_list
, &list
);
1919 if (plug
->rq_count
> 2 && plug
->multiple_queues
)
1920 list_sort(NULL
, &list
, plug_rq_cmp
);
1925 struct list_head rq_list
;
1926 struct request
*rq
, *head_rq
= list_entry_rq(list
.next
);
1927 struct list_head
*pos
= &head_rq
->queuelist
; /* skip first */
1928 struct blk_mq_hw_ctx
*this_hctx
= head_rq
->mq_hctx
;
1929 struct blk_mq_ctx
*this_ctx
= head_rq
->mq_ctx
;
1930 unsigned int depth
= 1;
1932 list_for_each_continue(pos
, &list
) {
1933 rq
= list_entry_rq(pos
);
1935 if (rq
->mq_hctx
!= this_hctx
|| rq
->mq_ctx
!= this_ctx
)
1940 list_cut_before(&rq_list
, &list
, pos
);
1941 trace_block_unplug(head_rq
->q
, depth
, !from_schedule
);
1942 blk_mq_sched_insert_requests(this_hctx
, this_ctx
, &rq_list
,
1944 } while(!list_empty(&list
));
1947 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
,
1948 unsigned int nr_segs
)
1952 if (bio
->bi_opf
& REQ_RAHEAD
)
1953 rq
->cmd_flags
|= REQ_FAILFAST_MASK
;
1955 rq
->__sector
= bio
->bi_iter
.bi_sector
;
1956 rq
->write_hint
= bio
->bi_write_hint
;
1957 blk_rq_bio_prep(rq
, bio
, nr_segs
);
1959 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
1960 err
= blk_crypto_rq_bio_prep(rq
, bio
, GFP_NOIO
);
1963 blk_account_io_start(rq
);
1966 static blk_status_t
__blk_mq_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1968 blk_qc_t
*cookie
, bool last
)
1970 struct request_queue
*q
= rq
->q
;
1971 struct blk_mq_queue_data bd
= {
1975 blk_qc_t new_cookie
;
1978 new_cookie
= request_to_qc_t(hctx
, rq
);
1981 * For OK queue, we are done. For error, caller may kill it.
1982 * Any other error (busy), just add it to our list as we
1983 * previously would have done.
1985 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1988 blk_mq_update_dispatch_busy(hctx
, false);
1989 *cookie
= new_cookie
;
1991 case BLK_STS_RESOURCE
:
1992 case BLK_STS_DEV_RESOURCE
:
1993 blk_mq_update_dispatch_busy(hctx
, true);
1994 __blk_mq_requeue_request(rq
);
1997 blk_mq_update_dispatch_busy(hctx
, false);
1998 *cookie
= BLK_QC_T_NONE
;
2005 static blk_status_t
__blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
2008 bool bypass_insert
, bool last
)
2010 struct request_queue
*q
= rq
->q
;
2011 bool run_queue
= true;
2014 * RCU or SRCU read lock is needed before checking quiesced flag.
2016 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2017 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2018 * and avoid driver to try to dispatch again.
2020 if (blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(q
)) {
2022 bypass_insert
= false;
2026 if (q
->elevator
&& !bypass_insert
)
2029 if (!blk_mq_get_dispatch_budget(q
))
2032 if (!blk_mq_get_driver_tag(rq
)) {
2033 blk_mq_put_dispatch_budget(q
);
2037 return __blk_mq_issue_directly(hctx
, rq
, cookie
, last
);
2040 return BLK_STS_RESOURCE
;
2042 blk_mq_sched_insert_request(rq
, false, run_queue
, false);
2048 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2049 * @hctx: Pointer of the associated hardware queue.
2050 * @rq: Pointer to request to be sent.
2051 * @cookie: Request queue cookie.
2053 * If the device has enough resources to accept a new request now, send the
2054 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2055 * we can try send it another time in the future. Requests inserted at this
2056 * queue have higher priority.
2058 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
2059 struct request
*rq
, blk_qc_t
*cookie
)
2064 might_sleep_if(hctx
->flags
& BLK_MQ_F_BLOCKING
);
2066 hctx_lock(hctx
, &srcu_idx
);
2068 ret
= __blk_mq_try_issue_directly(hctx
, rq
, cookie
, false, true);
2069 if (ret
== BLK_STS_RESOURCE
|| ret
== BLK_STS_DEV_RESOURCE
)
2070 blk_mq_request_bypass_insert(rq
, false, true);
2071 else if (ret
!= BLK_STS_OK
)
2072 blk_mq_end_request(rq
, ret
);
2074 hctx_unlock(hctx
, srcu_idx
);
2077 blk_status_t
blk_mq_request_issue_directly(struct request
*rq
, bool last
)
2081 blk_qc_t unused_cookie
;
2082 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
2084 hctx_lock(hctx
, &srcu_idx
);
2085 ret
= __blk_mq_try_issue_directly(hctx
, rq
, &unused_cookie
, true, last
);
2086 hctx_unlock(hctx
, srcu_idx
);
2091 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx
*hctx
,
2092 struct list_head
*list
)
2097 while (!list_empty(list
)) {
2099 struct request
*rq
= list_first_entry(list
, struct request
,
2102 list_del_init(&rq
->queuelist
);
2103 ret
= blk_mq_request_issue_directly(rq
, list_empty(list
));
2104 if (ret
!= BLK_STS_OK
) {
2105 if (ret
== BLK_STS_RESOURCE
||
2106 ret
== BLK_STS_DEV_RESOURCE
) {
2107 blk_mq_request_bypass_insert(rq
, false,
2111 blk_mq_end_request(rq
, ret
);
2118 * If we didn't flush the entire list, we could have told
2119 * the driver there was more coming, but that turned out to
2122 if ((!list_empty(list
) || errors
) &&
2123 hctx
->queue
->mq_ops
->commit_rqs
&& queued
)
2124 hctx
->queue
->mq_ops
->commit_rqs(hctx
);
2127 static void blk_add_rq_to_plug(struct blk_plug
*plug
, struct request
*rq
)
2129 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
2131 if (!plug
->multiple_queues
&& !list_is_singular(&plug
->mq_list
)) {
2132 struct request
*tmp
;
2134 tmp
= list_first_entry(&plug
->mq_list
, struct request
,
2136 if (tmp
->q
!= rq
->q
)
2137 plug
->multiple_queues
= true;
2142 * blk_mq_submit_bio - Create and send a request to block device.
2143 * @bio: Bio pointer.
2145 * Builds up a request structure from @q and @bio and send to the device. The
2146 * request may not be queued directly to hardware if:
2147 * * This request can be merged with another one
2148 * * We want to place request at plug queue for possible future merging
2149 * * There is an IO scheduler active at this queue
2151 * It will not queue the request if there is an error with the bio, or at the
2154 * Returns: Request queue cookie.
2156 blk_qc_t
blk_mq_submit_bio(struct bio
*bio
)
2158 struct request_queue
*q
= bio
->bi_bdev
->bd_disk
->queue
;
2159 const int is_sync
= op_is_sync(bio
->bi_opf
);
2160 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
2161 struct blk_mq_alloc_data data
= {
2165 struct blk_plug
*plug
;
2166 struct request
*same_queue_rq
= NULL
;
2167 unsigned int nr_segs
;
2172 blk_queue_bounce(q
, &bio
);
2173 __blk_queue_split(&bio
, &nr_segs
);
2175 if (!bio_integrity_prep(bio
))
2178 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
2179 blk_attempt_plug_merge(q
, bio
, nr_segs
, &same_queue_rq
))
2182 if (blk_mq_sched_bio_merge(q
, bio
, nr_segs
))
2185 rq_qos_throttle(q
, bio
);
2187 hipri
= bio
->bi_opf
& REQ_HIPRI
;
2189 data
.cmd_flags
= bio
->bi_opf
;
2190 rq
= __blk_mq_alloc_request(&data
);
2191 if (unlikely(!rq
)) {
2192 rq_qos_cleanup(q
, bio
);
2193 if (bio
->bi_opf
& REQ_NOWAIT
)
2194 bio_wouldblock_error(bio
);
2198 trace_block_getrq(bio
);
2200 rq_qos_track(q
, rq
, bio
);
2202 cookie
= request_to_qc_t(data
.hctx
, rq
);
2204 blk_mq_bio_to_request(rq
, bio
, nr_segs
);
2206 ret
= blk_crypto_init_request(rq
);
2207 if (ret
!= BLK_STS_OK
) {
2208 bio
->bi_status
= ret
;
2210 blk_mq_free_request(rq
);
2211 return BLK_QC_T_NONE
;
2214 plug
= blk_mq_plug(q
, bio
);
2215 if (unlikely(is_flush_fua
)) {
2216 /* Bypass scheduler for flush requests */
2217 blk_insert_flush(rq
);
2218 blk_mq_run_hw_queue(data
.hctx
, true);
2219 } else if (plug
&& (q
->nr_hw_queues
== 1 || q
->mq_ops
->commit_rqs
||
2220 !blk_queue_nonrot(q
))) {
2222 * Use plugging if we have a ->commit_rqs() hook as well, as
2223 * we know the driver uses bd->last in a smart fashion.
2225 * Use normal plugging if this disk is slow HDD, as sequential
2226 * IO may benefit a lot from plug merging.
2228 unsigned int request_count
= plug
->rq_count
;
2229 struct request
*last
= NULL
;
2232 trace_block_plug(q
);
2234 last
= list_entry_rq(plug
->mq_list
.prev
);
2236 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
2237 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
2238 blk_flush_plug_list(plug
, false);
2239 trace_block_plug(q
);
2242 blk_add_rq_to_plug(plug
, rq
);
2243 } else if (q
->elevator
) {
2244 /* Insert the request at the IO scheduler queue */
2245 blk_mq_sched_insert_request(rq
, false, true, true);
2246 } else if (plug
&& !blk_queue_nomerges(q
)) {
2248 * We do limited plugging. If the bio can be merged, do that.
2249 * Otherwise the existing request in the plug list will be
2250 * issued. So the plug list will have one request at most
2251 * The plug list might get flushed before this. If that happens,
2252 * the plug list is empty, and same_queue_rq is invalid.
2254 if (list_empty(&plug
->mq_list
))
2255 same_queue_rq
= NULL
;
2256 if (same_queue_rq
) {
2257 list_del_init(&same_queue_rq
->queuelist
);
2260 blk_add_rq_to_plug(plug
, rq
);
2261 trace_block_plug(q
);
2263 if (same_queue_rq
) {
2264 data
.hctx
= same_queue_rq
->mq_hctx
;
2265 trace_block_unplug(q
, 1, true);
2266 blk_mq_try_issue_directly(data
.hctx
, same_queue_rq
,
2269 } else if ((q
->nr_hw_queues
> 1 && is_sync
) ||
2270 !data
.hctx
->dispatch_busy
) {
2272 * There is no scheduler and we can try to send directly
2275 blk_mq_try_issue_directly(data
.hctx
, rq
, &cookie
);
2278 blk_mq_sched_insert_request(rq
, false, true, true);
2282 return BLK_QC_T_NONE
;
2286 return BLK_QC_T_NONE
;
2289 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
2290 unsigned int hctx_idx
)
2294 if (tags
->rqs
&& set
->ops
->exit_request
) {
2297 for (i
= 0; i
< tags
->nr_tags
; i
++) {
2298 struct request
*rq
= tags
->static_rqs
[i
];
2302 set
->ops
->exit_request(set
, rq
, hctx_idx
);
2303 tags
->static_rqs
[i
] = NULL
;
2307 while (!list_empty(&tags
->page_list
)) {
2308 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
2309 list_del_init(&page
->lru
);
2311 * Remove kmemleak object previously allocated in
2312 * blk_mq_alloc_rqs().
2314 kmemleak_free(page_address(page
));
2315 __free_pages(page
, page
->private);
2319 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
, unsigned int flags
)
2323 kfree(tags
->static_rqs
);
2324 tags
->static_rqs
= NULL
;
2326 blk_mq_free_tags(tags
, flags
);
2329 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
2330 unsigned int hctx_idx
,
2331 unsigned int nr_tags
,
2332 unsigned int reserved_tags
,
2335 struct blk_mq_tags
*tags
;
2338 node
= blk_mq_hw_queue_to_node(&set
->map
[HCTX_TYPE_DEFAULT
], hctx_idx
);
2339 if (node
== NUMA_NO_NODE
)
2340 node
= set
->numa_node
;
2342 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
, flags
);
2346 tags
->rqs
= kcalloc_node(nr_tags
, sizeof(struct request
*),
2347 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
2350 blk_mq_free_tags(tags
, flags
);
2354 tags
->static_rqs
= kcalloc_node(nr_tags
, sizeof(struct request
*),
2355 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
2357 if (!tags
->static_rqs
) {
2359 blk_mq_free_tags(tags
, flags
);
2366 static size_t order_to_size(unsigned int order
)
2368 return (size_t)PAGE_SIZE
<< order
;
2371 static int blk_mq_init_request(struct blk_mq_tag_set
*set
, struct request
*rq
,
2372 unsigned int hctx_idx
, int node
)
2376 if (set
->ops
->init_request
) {
2377 ret
= set
->ops
->init_request(set
, rq
, hctx_idx
, node
);
2382 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
2386 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
2387 unsigned int hctx_idx
, unsigned int depth
)
2389 unsigned int i
, j
, entries_per_page
, max_order
= 4;
2390 size_t rq_size
, left
;
2393 node
= blk_mq_hw_queue_to_node(&set
->map
[HCTX_TYPE_DEFAULT
], hctx_idx
);
2394 if (node
== NUMA_NO_NODE
)
2395 node
= set
->numa_node
;
2397 INIT_LIST_HEAD(&tags
->page_list
);
2400 * rq_size is the size of the request plus driver payload, rounded
2401 * to the cacheline size
2403 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
2405 left
= rq_size
* depth
;
2407 for (i
= 0; i
< depth
; ) {
2408 int this_order
= max_order
;
2413 while (this_order
&& left
< order_to_size(this_order
- 1))
2417 page
= alloc_pages_node(node
,
2418 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
2424 if (order_to_size(this_order
) < rq_size
)
2431 page
->private = this_order
;
2432 list_add_tail(&page
->lru
, &tags
->page_list
);
2434 p
= page_address(page
);
2436 * Allow kmemleak to scan these pages as they contain pointers
2437 * to additional allocations like via ops->init_request().
2439 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
2440 entries_per_page
= order_to_size(this_order
) / rq_size
;
2441 to_do
= min(entries_per_page
, depth
- i
);
2442 left
-= to_do
* rq_size
;
2443 for (j
= 0; j
< to_do
; j
++) {
2444 struct request
*rq
= p
;
2446 tags
->static_rqs
[i
] = rq
;
2447 if (blk_mq_init_request(set
, rq
, hctx_idx
, node
)) {
2448 tags
->static_rqs
[i
] = NULL
;
2459 blk_mq_free_rqs(set
, tags
, hctx_idx
);
2463 struct rq_iter_data
{
2464 struct blk_mq_hw_ctx
*hctx
;
2468 static bool blk_mq_has_request(struct request
*rq
, void *data
, bool reserved
)
2470 struct rq_iter_data
*iter_data
= data
;
2472 if (rq
->mq_hctx
!= iter_data
->hctx
)
2474 iter_data
->has_rq
= true;
2478 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx
*hctx
)
2480 struct blk_mq_tags
*tags
= hctx
->sched_tags
?
2481 hctx
->sched_tags
: hctx
->tags
;
2482 struct rq_iter_data data
= {
2486 blk_mq_all_tag_iter(tags
, blk_mq_has_request
, &data
);
2490 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu
,
2491 struct blk_mq_hw_ctx
*hctx
)
2493 if (cpumask_next_and(-1, hctx
->cpumask
, cpu_online_mask
) != cpu
)
2495 if (cpumask_next_and(cpu
, hctx
->cpumask
, cpu_online_mask
) < nr_cpu_ids
)
2500 static int blk_mq_hctx_notify_offline(unsigned int cpu
, struct hlist_node
*node
)
2502 struct blk_mq_hw_ctx
*hctx
= hlist_entry_safe(node
,
2503 struct blk_mq_hw_ctx
, cpuhp_online
);
2505 if (!cpumask_test_cpu(cpu
, hctx
->cpumask
) ||
2506 !blk_mq_last_cpu_in_hctx(cpu
, hctx
))
2510 * Prevent new request from being allocated on the current hctx.
2512 * The smp_mb__after_atomic() Pairs with the implied barrier in
2513 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
2514 * seen once we return from the tag allocator.
2516 set_bit(BLK_MQ_S_INACTIVE
, &hctx
->state
);
2517 smp_mb__after_atomic();
2520 * Try to grab a reference to the queue and wait for any outstanding
2521 * requests. If we could not grab a reference the queue has been
2522 * frozen and there are no requests.
2524 if (percpu_ref_tryget(&hctx
->queue
->q_usage_counter
)) {
2525 while (blk_mq_hctx_has_requests(hctx
))
2527 percpu_ref_put(&hctx
->queue
->q_usage_counter
);
2533 static int blk_mq_hctx_notify_online(unsigned int cpu
, struct hlist_node
*node
)
2535 struct blk_mq_hw_ctx
*hctx
= hlist_entry_safe(node
,
2536 struct blk_mq_hw_ctx
, cpuhp_online
);
2538 if (cpumask_test_cpu(cpu
, hctx
->cpumask
))
2539 clear_bit(BLK_MQ_S_INACTIVE
, &hctx
->state
);
2544 * 'cpu' is going away. splice any existing rq_list entries from this
2545 * software queue to the hw queue dispatch list, and ensure that it
2548 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
2550 struct blk_mq_hw_ctx
*hctx
;
2551 struct blk_mq_ctx
*ctx
;
2553 enum hctx_type type
;
2555 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
2556 if (!cpumask_test_cpu(cpu
, hctx
->cpumask
))
2559 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
2562 spin_lock(&ctx
->lock
);
2563 if (!list_empty(&ctx
->rq_lists
[type
])) {
2564 list_splice_init(&ctx
->rq_lists
[type
], &tmp
);
2565 blk_mq_hctx_clear_pending(hctx
, ctx
);
2567 spin_unlock(&ctx
->lock
);
2569 if (list_empty(&tmp
))
2572 spin_lock(&hctx
->lock
);
2573 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
2574 spin_unlock(&hctx
->lock
);
2576 blk_mq_run_hw_queue(hctx
, true);
2580 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
2582 if (!(hctx
->flags
& BLK_MQ_F_STACKING
))
2583 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE
,
2584 &hctx
->cpuhp_online
);
2585 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
2589 /* hctx->ctxs will be freed in queue's release handler */
2590 static void blk_mq_exit_hctx(struct request_queue
*q
,
2591 struct blk_mq_tag_set
*set
,
2592 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
2594 if (blk_mq_hw_queue_mapped(hctx
))
2595 blk_mq_tag_idle(hctx
);
2597 if (set
->ops
->exit_request
)
2598 set
->ops
->exit_request(set
, hctx
->fq
->flush_rq
, hctx_idx
);
2600 if (set
->ops
->exit_hctx
)
2601 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2603 blk_mq_remove_cpuhp(hctx
);
2605 spin_lock(&q
->unused_hctx_lock
);
2606 list_add(&hctx
->hctx_list
, &q
->unused_hctx_list
);
2607 spin_unlock(&q
->unused_hctx_lock
);
2610 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
2611 struct blk_mq_tag_set
*set
, int nr_queue
)
2613 struct blk_mq_hw_ctx
*hctx
;
2616 queue_for_each_hw_ctx(q
, hctx
, i
) {
2619 blk_mq_debugfs_unregister_hctx(hctx
);
2620 blk_mq_exit_hctx(q
, set
, hctx
, i
);
2624 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set
*tag_set
)
2626 int hw_ctx_size
= sizeof(struct blk_mq_hw_ctx
);
2628 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx
, srcu
),
2629 __alignof__(struct blk_mq_hw_ctx
)) !=
2630 sizeof(struct blk_mq_hw_ctx
));
2632 if (tag_set
->flags
& BLK_MQ_F_BLOCKING
)
2633 hw_ctx_size
+= sizeof(struct srcu_struct
);
2638 static int blk_mq_init_hctx(struct request_queue
*q
,
2639 struct blk_mq_tag_set
*set
,
2640 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
2642 hctx
->queue_num
= hctx_idx
;
2644 if (!(hctx
->flags
& BLK_MQ_F_STACKING
))
2645 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE
,
2646 &hctx
->cpuhp_online
);
2647 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
2649 hctx
->tags
= set
->tags
[hctx_idx
];
2651 if (set
->ops
->init_hctx
&&
2652 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
2653 goto unregister_cpu_notifier
;
2655 if (blk_mq_init_request(set
, hctx
->fq
->flush_rq
, hctx_idx
,
2661 if (set
->ops
->exit_hctx
)
2662 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2663 unregister_cpu_notifier
:
2664 blk_mq_remove_cpuhp(hctx
);
2668 static struct blk_mq_hw_ctx
*
2669 blk_mq_alloc_hctx(struct request_queue
*q
, struct blk_mq_tag_set
*set
,
2672 struct blk_mq_hw_ctx
*hctx
;
2673 gfp_t gfp
= GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
;
2675 hctx
= kzalloc_node(blk_mq_hw_ctx_size(set
), gfp
, node
);
2677 goto fail_alloc_hctx
;
2679 if (!zalloc_cpumask_var_node(&hctx
->cpumask
, gfp
, node
))
2682 atomic_set(&hctx
->nr_active
, 0);
2683 if (node
== NUMA_NO_NODE
)
2684 node
= set
->numa_node
;
2685 hctx
->numa_node
= node
;
2687 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
2688 spin_lock_init(&hctx
->lock
);
2689 INIT_LIST_HEAD(&hctx
->dispatch
);
2691 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_QUEUE_SHARED
;
2693 INIT_LIST_HEAD(&hctx
->hctx_list
);
2696 * Allocate space for all possible cpus to avoid allocation at
2699 hctx
->ctxs
= kmalloc_array_node(nr_cpu_ids
, sizeof(void *),
2704 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8),
2709 spin_lock_init(&hctx
->dispatch_wait_lock
);
2710 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
2711 INIT_LIST_HEAD(&hctx
->dispatch_wait
.entry
);
2713 hctx
->fq
= blk_alloc_flush_queue(hctx
->numa_node
, set
->cmd_size
, gfp
);
2717 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
2718 init_srcu_struct(hctx
->srcu
);
2719 blk_mq_hctx_kobj_init(hctx
);
2724 sbitmap_free(&hctx
->ctx_map
);
2728 free_cpumask_var(hctx
->cpumask
);
2735 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
2736 unsigned int nr_hw_queues
)
2738 struct blk_mq_tag_set
*set
= q
->tag_set
;
2741 for_each_possible_cpu(i
) {
2742 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2743 struct blk_mq_hw_ctx
*hctx
;
2747 spin_lock_init(&__ctx
->lock
);
2748 for (k
= HCTX_TYPE_DEFAULT
; k
< HCTX_MAX_TYPES
; k
++)
2749 INIT_LIST_HEAD(&__ctx
->rq_lists
[k
]);
2754 * Set local node, IFF we have more than one hw queue. If
2755 * not, we remain on the home node of the device
2757 for (j
= 0; j
< set
->nr_maps
; j
++) {
2758 hctx
= blk_mq_map_queue_type(q
, j
, i
);
2759 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
2760 hctx
->numa_node
= cpu_to_node(i
);
2765 static bool __blk_mq_alloc_map_and_request(struct blk_mq_tag_set
*set
,
2768 unsigned int flags
= set
->flags
;
2771 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
2772 set
->queue_depth
, set
->reserved_tags
, flags
);
2773 if (!set
->tags
[hctx_idx
])
2776 ret
= blk_mq_alloc_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
,
2781 blk_mq_free_rq_map(set
->tags
[hctx_idx
], flags
);
2782 set
->tags
[hctx_idx
] = NULL
;
2786 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set
*set
,
2787 unsigned int hctx_idx
)
2789 unsigned int flags
= set
->flags
;
2791 if (set
->tags
&& set
->tags
[hctx_idx
]) {
2792 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
2793 blk_mq_free_rq_map(set
->tags
[hctx_idx
], flags
);
2794 set
->tags
[hctx_idx
] = NULL
;
2798 static void blk_mq_map_swqueue(struct request_queue
*q
)
2800 unsigned int i
, j
, hctx_idx
;
2801 struct blk_mq_hw_ctx
*hctx
;
2802 struct blk_mq_ctx
*ctx
;
2803 struct blk_mq_tag_set
*set
= q
->tag_set
;
2805 queue_for_each_hw_ctx(q
, hctx
, i
) {
2806 cpumask_clear(hctx
->cpumask
);
2808 hctx
->dispatch_from
= NULL
;
2812 * Map software to hardware queues.
2814 * If the cpu isn't present, the cpu is mapped to first hctx.
2816 for_each_possible_cpu(i
) {
2818 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2819 for (j
= 0; j
< set
->nr_maps
; j
++) {
2820 if (!set
->map
[j
].nr_queues
) {
2821 ctx
->hctxs
[j
] = blk_mq_map_queue_type(q
,
2822 HCTX_TYPE_DEFAULT
, i
);
2825 hctx_idx
= set
->map
[j
].mq_map
[i
];
2826 /* unmapped hw queue can be remapped after CPU topo changed */
2827 if (!set
->tags
[hctx_idx
] &&
2828 !__blk_mq_alloc_map_and_request(set
, hctx_idx
)) {
2830 * If tags initialization fail for some hctx,
2831 * that hctx won't be brought online. In this
2832 * case, remap the current ctx to hctx[0] which
2833 * is guaranteed to always have tags allocated
2835 set
->map
[j
].mq_map
[i
] = 0;
2838 hctx
= blk_mq_map_queue_type(q
, j
, i
);
2839 ctx
->hctxs
[j
] = hctx
;
2841 * If the CPU is already set in the mask, then we've
2842 * mapped this one already. This can happen if
2843 * devices share queues across queue maps.
2845 if (cpumask_test_cpu(i
, hctx
->cpumask
))
2848 cpumask_set_cpu(i
, hctx
->cpumask
);
2850 ctx
->index_hw
[hctx
->type
] = hctx
->nr_ctx
;
2851 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2854 * If the nr_ctx type overflows, we have exceeded the
2855 * amount of sw queues we can support.
2857 BUG_ON(!hctx
->nr_ctx
);
2860 for (; j
< HCTX_MAX_TYPES
; j
++)
2861 ctx
->hctxs
[j
] = blk_mq_map_queue_type(q
,
2862 HCTX_TYPE_DEFAULT
, i
);
2865 queue_for_each_hw_ctx(q
, hctx
, i
) {
2867 * If no software queues are mapped to this hardware queue,
2868 * disable it and free the request entries.
2870 if (!hctx
->nr_ctx
) {
2871 /* Never unmap queue 0. We need it as a
2872 * fallback in case of a new remap fails
2875 if (i
&& set
->tags
[i
])
2876 blk_mq_free_map_and_requests(set
, i
);
2882 hctx
->tags
= set
->tags
[i
];
2883 WARN_ON(!hctx
->tags
);
2886 * Set the map size to the number of mapped software queues.
2887 * This is more accurate and more efficient than looping
2888 * over all possibly mapped software queues.
2890 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2893 * Initialize batch roundrobin counts
2895 hctx
->next_cpu
= blk_mq_first_mapped_cpu(hctx
);
2896 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2901 * Caller needs to ensure that we're either frozen/quiesced, or that
2902 * the queue isn't live yet.
2904 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2906 struct blk_mq_hw_ctx
*hctx
;
2909 queue_for_each_hw_ctx(q
, hctx
, i
) {
2911 hctx
->flags
|= BLK_MQ_F_TAG_QUEUE_SHARED
;
2913 hctx
->flags
&= ~BLK_MQ_F_TAG_QUEUE_SHARED
;
2917 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set
*set
,
2920 struct request_queue
*q
;
2922 lockdep_assert_held(&set
->tag_list_lock
);
2924 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2925 blk_mq_freeze_queue(q
);
2926 queue_set_hctx_shared(q
, shared
);
2927 blk_mq_unfreeze_queue(q
);
2931 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
2933 struct blk_mq_tag_set
*set
= q
->tag_set
;
2935 mutex_lock(&set
->tag_list_lock
);
2936 list_del(&q
->tag_set_list
);
2937 if (list_is_singular(&set
->tag_list
)) {
2938 /* just transitioned to unshared */
2939 set
->flags
&= ~BLK_MQ_F_TAG_QUEUE_SHARED
;
2940 /* update existing queue */
2941 blk_mq_update_tag_set_shared(set
, false);
2943 mutex_unlock(&set
->tag_list_lock
);
2944 INIT_LIST_HEAD(&q
->tag_set_list
);
2947 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
2948 struct request_queue
*q
)
2950 mutex_lock(&set
->tag_list_lock
);
2953 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2955 if (!list_empty(&set
->tag_list
) &&
2956 !(set
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
)) {
2957 set
->flags
|= BLK_MQ_F_TAG_QUEUE_SHARED
;
2958 /* update existing queue */
2959 blk_mq_update_tag_set_shared(set
, true);
2961 if (set
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
)
2962 queue_set_hctx_shared(q
, true);
2963 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
2965 mutex_unlock(&set
->tag_list_lock
);
2968 /* All allocations will be freed in release handler of q->mq_kobj */
2969 static int blk_mq_alloc_ctxs(struct request_queue
*q
)
2971 struct blk_mq_ctxs
*ctxs
;
2974 ctxs
= kzalloc(sizeof(*ctxs
), GFP_KERNEL
);
2978 ctxs
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2979 if (!ctxs
->queue_ctx
)
2982 for_each_possible_cpu(cpu
) {
2983 struct blk_mq_ctx
*ctx
= per_cpu_ptr(ctxs
->queue_ctx
, cpu
);
2987 q
->mq_kobj
= &ctxs
->kobj
;
2988 q
->queue_ctx
= ctxs
->queue_ctx
;
2997 * It is the actual release handler for mq, but we do it from
2998 * request queue's release handler for avoiding use-after-free
2999 * and headache because q->mq_kobj shouldn't have been introduced,
3000 * but we can't group ctx/kctx kobj without it.
3002 void blk_mq_release(struct request_queue
*q
)
3004 struct blk_mq_hw_ctx
*hctx
, *next
;
3007 queue_for_each_hw_ctx(q
, hctx
, i
)
3008 WARN_ON_ONCE(hctx
&& list_empty(&hctx
->hctx_list
));
3010 /* all hctx are in .unused_hctx_list now */
3011 list_for_each_entry_safe(hctx
, next
, &q
->unused_hctx_list
, hctx_list
) {
3012 list_del_init(&hctx
->hctx_list
);
3013 kobject_put(&hctx
->kobj
);
3016 kfree(q
->queue_hw_ctx
);
3019 * release .mq_kobj and sw queue's kobject now because
3020 * both share lifetime with request queue.
3022 blk_mq_sysfs_deinit(q
);
3025 struct request_queue
*blk_mq_init_queue_data(struct blk_mq_tag_set
*set
,
3028 struct request_queue
*uninit_q
, *q
;
3030 uninit_q
= blk_alloc_queue(set
->numa_node
);
3032 return ERR_PTR(-ENOMEM
);
3033 uninit_q
->queuedata
= queuedata
;
3036 * Initialize the queue without an elevator. device_add_disk() will do
3037 * the initialization.
3039 q
= blk_mq_init_allocated_queue(set
, uninit_q
, false);
3041 blk_cleanup_queue(uninit_q
);
3045 EXPORT_SYMBOL_GPL(blk_mq_init_queue_data
);
3047 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
3049 return blk_mq_init_queue_data(set
, NULL
);
3051 EXPORT_SYMBOL(blk_mq_init_queue
);
3054 * Helper for setting up a queue with mq ops, given queue depth, and
3055 * the passed in mq ops flags.
3057 struct request_queue
*blk_mq_init_sq_queue(struct blk_mq_tag_set
*set
,
3058 const struct blk_mq_ops
*ops
,
3059 unsigned int queue_depth
,
3060 unsigned int set_flags
)
3062 struct request_queue
*q
;
3065 memset(set
, 0, sizeof(*set
));
3067 set
->nr_hw_queues
= 1;
3069 set
->queue_depth
= queue_depth
;
3070 set
->numa_node
= NUMA_NO_NODE
;
3071 set
->flags
= set_flags
;
3073 ret
= blk_mq_alloc_tag_set(set
);
3075 return ERR_PTR(ret
);
3077 q
= blk_mq_init_queue(set
);
3079 blk_mq_free_tag_set(set
);
3085 EXPORT_SYMBOL(blk_mq_init_sq_queue
);
3087 static struct blk_mq_hw_ctx
*blk_mq_alloc_and_init_hctx(
3088 struct blk_mq_tag_set
*set
, struct request_queue
*q
,
3089 int hctx_idx
, int node
)
3091 struct blk_mq_hw_ctx
*hctx
= NULL
, *tmp
;
3093 /* reuse dead hctx first */
3094 spin_lock(&q
->unused_hctx_lock
);
3095 list_for_each_entry(tmp
, &q
->unused_hctx_list
, hctx_list
) {
3096 if (tmp
->numa_node
== node
) {
3102 list_del_init(&hctx
->hctx_list
);
3103 spin_unlock(&q
->unused_hctx_lock
);
3106 hctx
= blk_mq_alloc_hctx(q
, set
, node
);
3110 if (blk_mq_init_hctx(q
, set
, hctx
, hctx_idx
))
3116 kobject_put(&hctx
->kobj
);
3121 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
3122 struct request_queue
*q
)
3125 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
3127 if (q
->nr_hw_queues
< set
->nr_hw_queues
) {
3128 struct blk_mq_hw_ctx
**new_hctxs
;
3130 new_hctxs
= kcalloc_node(set
->nr_hw_queues
,
3131 sizeof(*new_hctxs
), GFP_KERNEL
,
3136 memcpy(new_hctxs
, hctxs
, q
->nr_hw_queues
*
3138 q
->queue_hw_ctx
= new_hctxs
;
3143 /* protect against switching io scheduler */
3144 mutex_lock(&q
->sysfs_lock
);
3145 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
3147 struct blk_mq_hw_ctx
*hctx
;
3149 node
= blk_mq_hw_queue_to_node(&set
->map
[HCTX_TYPE_DEFAULT
], i
);
3151 * If the hw queue has been mapped to another numa node,
3152 * we need to realloc the hctx. If allocation fails, fallback
3153 * to use the previous one.
3155 if (hctxs
[i
] && (hctxs
[i
]->numa_node
== node
))
3158 hctx
= blk_mq_alloc_and_init_hctx(set
, q
, i
, node
);
3161 blk_mq_exit_hctx(q
, set
, hctxs
[i
], i
);
3165 pr_warn("Allocate new hctx on node %d fails,\
3166 fallback to previous one on node %d\n",
3167 node
, hctxs
[i
]->numa_node
);
3173 * Increasing nr_hw_queues fails. Free the newly allocated
3174 * hctxs and keep the previous q->nr_hw_queues.
3176 if (i
!= set
->nr_hw_queues
) {
3177 j
= q
->nr_hw_queues
;
3181 end
= q
->nr_hw_queues
;
3182 q
->nr_hw_queues
= set
->nr_hw_queues
;
3185 for (; j
< end
; j
++) {
3186 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
3190 blk_mq_free_map_and_requests(set
, j
);
3191 blk_mq_exit_hctx(q
, set
, hctx
, j
);
3195 mutex_unlock(&q
->sysfs_lock
);
3198 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
3199 struct request_queue
*q
,
3202 /* mark the queue as mq asap */
3203 q
->mq_ops
= set
->ops
;
3205 q
->poll_cb
= blk_stat_alloc_callback(blk_mq_poll_stats_fn
,
3206 blk_mq_poll_stats_bkt
,
3207 BLK_MQ_POLL_STATS_BKTS
, q
);
3211 if (blk_mq_alloc_ctxs(q
))
3214 /* init q->mq_kobj and sw queues' kobjects */
3215 blk_mq_sysfs_init(q
);
3217 INIT_LIST_HEAD(&q
->unused_hctx_list
);
3218 spin_lock_init(&q
->unused_hctx_lock
);
3220 blk_mq_realloc_hw_ctxs(set
, q
);
3221 if (!q
->nr_hw_queues
)
3224 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
3225 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
3229 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
3230 if (set
->nr_maps
> HCTX_TYPE_POLL
&&
3231 set
->map
[HCTX_TYPE_POLL
].nr_queues
)
3232 blk_queue_flag_set(QUEUE_FLAG_POLL
, q
);
3234 q
->sg_reserved_size
= INT_MAX
;
3236 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
3237 INIT_LIST_HEAD(&q
->requeue_list
);
3238 spin_lock_init(&q
->requeue_lock
);
3240 q
->nr_requests
= set
->queue_depth
;
3243 * Default to classic polling
3245 q
->poll_nsec
= BLK_MQ_POLL_CLASSIC
;
3247 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
3248 blk_mq_add_queue_tag_set(set
, q
);
3249 blk_mq_map_swqueue(q
);
3252 elevator_init_mq(q
);
3257 kfree(q
->queue_hw_ctx
);
3258 q
->nr_hw_queues
= 0;
3259 blk_mq_sysfs_deinit(q
);
3261 blk_stat_free_callback(q
->poll_cb
);
3265 return ERR_PTR(-ENOMEM
);
3267 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
3269 /* tags can _not_ be used after returning from blk_mq_exit_queue */
3270 void blk_mq_exit_queue(struct request_queue
*q
)
3272 struct blk_mq_tag_set
*set
= q
->tag_set
;
3274 blk_mq_del_queue_tag_set(q
);
3275 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
3278 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
3282 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
3283 if (!__blk_mq_alloc_map_and_request(set
, i
))
3292 blk_mq_free_map_and_requests(set
, i
);
3298 * Allocate the request maps associated with this tag_set. Note that this
3299 * may reduce the depth asked for, if memory is tight. set->queue_depth
3300 * will be updated to reflect the allocated depth.
3302 static int blk_mq_alloc_map_and_requests(struct blk_mq_tag_set
*set
)
3307 depth
= set
->queue_depth
;
3309 err
= __blk_mq_alloc_rq_maps(set
);
3313 set
->queue_depth
>>= 1;
3314 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
3318 } while (set
->queue_depth
);
3320 if (!set
->queue_depth
|| err
) {
3321 pr_err("blk-mq: failed to allocate request map\n");
3325 if (depth
!= set
->queue_depth
)
3326 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3327 depth
, set
->queue_depth
);
3332 static int blk_mq_update_queue_map(struct blk_mq_tag_set
*set
)
3335 * blk_mq_map_queues() and multiple .map_queues() implementations
3336 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
3337 * number of hardware queues.
3339 if (set
->nr_maps
== 1)
3340 set
->map
[HCTX_TYPE_DEFAULT
].nr_queues
= set
->nr_hw_queues
;
3342 if (set
->ops
->map_queues
&& !is_kdump_kernel()) {
3346 * transport .map_queues is usually done in the following
3349 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3350 * mask = get_cpu_mask(queue)
3351 * for_each_cpu(cpu, mask)
3352 * set->map[x].mq_map[cpu] = queue;
3355 * When we need to remap, the table has to be cleared for
3356 * killing stale mapping since one CPU may not be mapped
3359 for (i
= 0; i
< set
->nr_maps
; i
++)
3360 blk_mq_clear_mq_map(&set
->map
[i
]);
3362 return set
->ops
->map_queues(set
);
3364 BUG_ON(set
->nr_maps
> 1);
3365 return blk_mq_map_queues(&set
->map
[HCTX_TYPE_DEFAULT
]);
3369 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set
*set
,
3370 int cur_nr_hw_queues
, int new_nr_hw_queues
)
3372 struct blk_mq_tags
**new_tags
;
3374 if (cur_nr_hw_queues
>= new_nr_hw_queues
)
3377 new_tags
= kcalloc_node(new_nr_hw_queues
, sizeof(struct blk_mq_tags
*),
3378 GFP_KERNEL
, set
->numa_node
);
3383 memcpy(new_tags
, set
->tags
, cur_nr_hw_queues
*
3384 sizeof(*set
->tags
));
3386 set
->tags
= new_tags
;
3387 set
->nr_hw_queues
= new_nr_hw_queues
;
3392 static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set
*set
,
3393 int new_nr_hw_queues
)
3395 return blk_mq_realloc_tag_set_tags(set
, 0, new_nr_hw_queues
);
3399 * Alloc a tag set to be associated with one or more request queues.
3400 * May fail with EINVAL for various error conditions. May adjust the
3401 * requested depth down, if it's too large. In that case, the set
3402 * value will be stored in set->queue_depth.
3404 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
3408 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
3410 if (!set
->nr_hw_queues
)
3412 if (!set
->queue_depth
)
3414 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
3417 if (!set
->ops
->queue_rq
)
3420 if (!set
->ops
->get_budget
^ !set
->ops
->put_budget
)
3423 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
3424 pr_info("blk-mq: reduced tag depth to %u\n",
3426 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
3431 else if (set
->nr_maps
> HCTX_MAX_TYPES
)
3435 * If a crashdump is active, then we are potentially in a very
3436 * memory constrained environment. Limit us to 1 queue and
3437 * 64 tags to prevent using too much memory.
3439 if (is_kdump_kernel()) {
3440 set
->nr_hw_queues
= 1;
3442 set
->queue_depth
= min(64U, set
->queue_depth
);
3445 * There is no use for more h/w queues than cpus if we just have
3448 if (set
->nr_maps
== 1 && set
->nr_hw_queues
> nr_cpu_ids
)
3449 set
->nr_hw_queues
= nr_cpu_ids
;
3451 if (blk_mq_alloc_tag_set_tags(set
, set
->nr_hw_queues
) < 0)
3455 for (i
= 0; i
< set
->nr_maps
; i
++) {
3456 set
->map
[i
].mq_map
= kcalloc_node(nr_cpu_ids
,
3457 sizeof(set
->map
[i
].mq_map
[0]),
3458 GFP_KERNEL
, set
->numa_node
);
3459 if (!set
->map
[i
].mq_map
)
3460 goto out_free_mq_map
;
3461 set
->map
[i
].nr_queues
= is_kdump_kernel() ? 1 : set
->nr_hw_queues
;
3464 ret
= blk_mq_update_queue_map(set
);
3466 goto out_free_mq_map
;
3468 ret
= blk_mq_alloc_map_and_requests(set
);
3470 goto out_free_mq_map
;
3472 if (blk_mq_is_sbitmap_shared(set
->flags
)) {
3473 atomic_set(&set
->active_queues_shared_sbitmap
, 0);
3475 if (blk_mq_init_shared_sbitmap(set
, set
->flags
)) {
3477 goto out_free_mq_rq_maps
;
3481 mutex_init(&set
->tag_list_lock
);
3482 INIT_LIST_HEAD(&set
->tag_list
);
3486 out_free_mq_rq_maps
:
3487 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
3488 blk_mq_free_map_and_requests(set
, i
);
3490 for (i
= 0; i
< set
->nr_maps
; i
++) {
3491 kfree(set
->map
[i
].mq_map
);
3492 set
->map
[i
].mq_map
= NULL
;
3498 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
3500 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
3504 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
3505 blk_mq_free_map_and_requests(set
, i
);
3507 if (blk_mq_is_sbitmap_shared(set
->flags
))
3508 blk_mq_exit_shared_sbitmap(set
);
3510 for (j
= 0; j
< set
->nr_maps
; j
++) {
3511 kfree(set
->map
[j
].mq_map
);
3512 set
->map
[j
].mq_map
= NULL
;
3518 EXPORT_SYMBOL(blk_mq_free_tag_set
);
3520 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
3522 struct blk_mq_tag_set
*set
= q
->tag_set
;
3523 struct blk_mq_hw_ctx
*hctx
;
3529 if (q
->nr_requests
== nr
)
3532 blk_mq_freeze_queue(q
);
3533 blk_mq_quiesce_queue(q
);
3536 queue_for_each_hw_ctx(q
, hctx
, i
) {
3540 * If we're using an MQ scheduler, just update the scheduler
3541 * queue depth. This is similar to what the old code would do.
3543 if (!hctx
->sched_tags
) {
3544 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
, nr
,
3546 if (!ret
&& blk_mq_is_sbitmap_shared(set
->flags
))
3547 blk_mq_tag_resize_shared_sbitmap(set
, nr
);
3549 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
3554 if (q
->elevator
&& q
->elevator
->type
->ops
.depth_updated
)
3555 q
->elevator
->type
->ops
.depth_updated(hctx
);
3559 q
->nr_requests
= nr
;
3561 blk_mq_unquiesce_queue(q
);
3562 blk_mq_unfreeze_queue(q
);
3568 * request_queue and elevator_type pair.
3569 * It is just used by __blk_mq_update_nr_hw_queues to cache
3570 * the elevator_type associated with a request_queue.
3572 struct blk_mq_qe_pair
{
3573 struct list_head node
;
3574 struct request_queue
*q
;
3575 struct elevator_type
*type
;
3579 * Cache the elevator_type in qe pair list and switch the
3580 * io scheduler to 'none'
3582 static bool blk_mq_elv_switch_none(struct list_head
*head
,
3583 struct request_queue
*q
)
3585 struct blk_mq_qe_pair
*qe
;
3590 qe
= kmalloc(sizeof(*qe
), GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
);
3594 INIT_LIST_HEAD(&qe
->node
);
3596 qe
->type
= q
->elevator
->type
;
3597 list_add(&qe
->node
, head
);
3599 mutex_lock(&q
->sysfs_lock
);
3601 * After elevator_switch_mq, the previous elevator_queue will be
3602 * released by elevator_release. The reference of the io scheduler
3603 * module get by elevator_get will also be put. So we need to get
3604 * a reference of the io scheduler module here to prevent it to be
3607 __module_get(qe
->type
->elevator_owner
);
3608 elevator_switch_mq(q
, NULL
);
3609 mutex_unlock(&q
->sysfs_lock
);
3614 static void blk_mq_elv_switch_back(struct list_head
*head
,
3615 struct request_queue
*q
)
3617 struct blk_mq_qe_pair
*qe
;
3618 struct elevator_type
*t
= NULL
;
3620 list_for_each_entry(qe
, head
, node
)
3629 list_del(&qe
->node
);
3632 mutex_lock(&q
->sysfs_lock
);
3633 elevator_switch_mq(q
, t
);
3634 mutex_unlock(&q
->sysfs_lock
);
3637 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
,
3640 struct request_queue
*q
;
3642 int prev_nr_hw_queues
;
3644 lockdep_assert_held(&set
->tag_list_lock
);
3646 if (set
->nr_maps
== 1 && nr_hw_queues
> nr_cpu_ids
)
3647 nr_hw_queues
= nr_cpu_ids
;
3648 if (nr_hw_queues
< 1)
3650 if (set
->nr_maps
== 1 && nr_hw_queues
== set
->nr_hw_queues
)
3653 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3654 blk_mq_freeze_queue(q
);
3656 * Switch IO scheduler to 'none', cleaning up the data associated
3657 * with the previous scheduler. We will switch back once we are done
3658 * updating the new sw to hw queue mappings.
3660 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3661 if (!blk_mq_elv_switch_none(&head
, q
))
3664 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3665 blk_mq_debugfs_unregister_hctxs(q
);
3666 blk_mq_sysfs_unregister(q
);
3669 prev_nr_hw_queues
= set
->nr_hw_queues
;
3670 if (blk_mq_realloc_tag_set_tags(set
, set
->nr_hw_queues
, nr_hw_queues
) <
3674 set
->nr_hw_queues
= nr_hw_queues
;
3676 blk_mq_update_queue_map(set
);
3677 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3678 blk_mq_realloc_hw_ctxs(set
, q
);
3679 if (q
->nr_hw_queues
!= set
->nr_hw_queues
) {
3680 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3681 nr_hw_queues
, prev_nr_hw_queues
);
3682 set
->nr_hw_queues
= prev_nr_hw_queues
;
3683 blk_mq_map_queues(&set
->map
[HCTX_TYPE_DEFAULT
]);
3686 blk_mq_map_swqueue(q
);
3690 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3691 blk_mq_sysfs_register(q
);
3692 blk_mq_debugfs_register_hctxs(q
);
3696 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3697 blk_mq_elv_switch_back(&head
, q
);
3699 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3700 blk_mq_unfreeze_queue(q
);
3703 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
3705 mutex_lock(&set
->tag_list_lock
);
3706 __blk_mq_update_nr_hw_queues(set
, nr_hw_queues
);
3707 mutex_unlock(&set
->tag_list_lock
);
3709 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
3711 /* Enable polling stats and return whether they were already enabled. */
3712 static bool blk_poll_stats_enable(struct request_queue
*q
)
3714 if (test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
3715 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS
, q
))
3717 blk_stat_add_callback(q
, q
->poll_cb
);
3721 static void blk_mq_poll_stats_start(struct request_queue
*q
)
3724 * We don't arm the callback if polling stats are not enabled or the
3725 * callback is already active.
3727 if (!test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
3728 blk_stat_is_active(q
->poll_cb
))
3731 blk_stat_activate_msecs(q
->poll_cb
, 100);
3734 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
)
3736 struct request_queue
*q
= cb
->data
;
3739 for (bucket
= 0; bucket
< BLK_MQ_POLL_STATS_BKTS
; bucket
++) {
3740 if (cb
->stat
[bucket
].nr_samples
)
3741 q
->poll_stat
[bucket
] = cb
->stat
[bucket
];
3745 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
3748 unsigned long ret
= 0;
3752 * If stats collection isn't on, don't sleep but turn it on for
3755 if (!blk_poll_stats_enable(q
))
3759 * As an optimistic guess, use half of the mean service time
3760 * for this type of request. We can (and should) make this smarter.
3761 * For instance, if the completion latencies are tight, we can
3762 * get closer than just half the mean. This is especially
3763 * important on devices where the completion latencies are longer
3764 * than ~10 usec. We do use the stats for the relevant IO size
3765 * if available which does lead to better estimates.
3767 bucket
= blk_mq_poll_stats_bkt(rq
);
3771 if (q
->poll_stat
[bucket
].nr_samples
)
3772 ret
= (q
->poll_stat
[bucket
].mean
+ 1) / 2;
3777 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
3780 struct hrtimer_sleeper hs
;
3781 enum hrtimer_mode mode
;
3785 if (rq
->rq_flags
& RQF_MQ_POLL_SLEPT
)
3789 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3791 * 0: use half of prev avg
3792 * >0: use this specific value
3794 if (q
->poll_nsec
> 0)
3795 nsecs
= q
->poll_nsec
;
3797 nsecs
= blk_mq_poll_nsecs(q
, rq
);
3802 rq
->rq_flags
|= RQF_MQ_POLL_SLEPT
;
3805 * This will be replaced with the stats tracking code, using
3806 * 'avg_completion_time / 2' as the pre-sleep target.
3810 mode
= HRTIMER_MODE_REL
;
3811 hrtimer_init_sleeper_on_stack(&hs
, CLOCK_MONOTONIC
, mode
);
3812 hrtimer_set_expires(&hs
.timer
, kt
);
3815 if (blk_mq_rq_state(rq
) == MQ_RQ_COMPLETE
)
3817 set_current_state(TASK_UNINTERRUPTIBLE
);
3818 hrtimer_sleeper_start_expires(&hs
, mode
);
3821 hrtimer_cancel(&hs
.timer
);
3822 mode
= HRTIMER_MODE_ABS
;
3823 } while (hs
.task
&& !signal_pending(current
));
3825 __set_current_state(TASK_RUNNING
);
3826 destroy_hrtimer_on_stack(&hs
.timer
);
3830 static bool blk_mq_poll_hybrid(struct request_queue
*q
,
3831 struct blk_mq_hw_ctx
*hctx
, blk_qc_t cookie
)
3835 if (q
->poll_nsec
== BLK_MQ_POLL_CLASSIC
)
3838 if (!blk_qc_t_is_internal(cookie
))
3839 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
3841 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
3843 * With scheduling, if the request has completed, we'll
3844 * get a NULL return here, as we clear the sched tag when
3845 * that happens. The request still remains valid, like always,
3846 * so we should be safe with just the NULL check.
3852 return blk_mq_poll_hybrid_sleep(q
, rq
);
3856 * blk_poll - poll for IO completions
3858 * @cookie: cookie passed back at IO submission time
3859 * @spin: whether to spin for completions
3862 * Poll for completions on the passed in queue. Returns number of
3863 * completed entries found. If @spin is true, then blk_poll will continue
3864 * looping until at least one completion is found, unless the task is
3865 * otherwise marked running (or we need to reschedule).
3867 int blk_poll(struct request_queue
*q
, blk_qc_t cookie
, bool spin
)
3869 struct blk_mq_hw_ctx
*hctx
;
3872 if (!blk_qc_t_valid(cookie
) ||
3873 !test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
3877 blk_flush_plug_list(current
->plug
, false);
3879 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
3882 * If we sleep, have the caller restart the poll loop to reset
3883 * the state. Like for the other success return cases, the
3884 * caller is responsible for checking if the IO completed. If
3885 * the IO isn't complete, we'll get called again and will go
3886 * straight to the busy poll loop. If specified not to spin,
3887 * we also should not sleep.
3889 if (spin
&& blk_mq_poll_hybrid(q
, hctx
, cookie
))
3892 hctx
->poll_considered
++;
3894 state
= current
->state
;
3898 hctx
->poll_invoked
++;
3900 ret
= q
->mq_ops
->poll(hctx
);
3902 hctx
->poll_success
++;
3903 __set_current_state(TASK_RUNNING
);
3907 if (signal_pending_state(state
, current
))
3908 __set_current_state(TASK_RUNNING
);
3910 if (current
->state
== TASK_RUNNING
)
3912 if (ret
< 0 || !spin
)
3915 } while (!need_resched());
3917 __set_current_state(TASK_RUNNING
);
3920 EXPORT_SYMBOL_GPL(blk_poll
);
3922 unsigned int blk_mq_rq_cpu(struct request
*rq
)
3924 return rq
->mq_ctx
->cpu
;
3926 EXPORT_SYMBOL(blk_mq_rq_cpu
);
3928 static int __init
blk_mq_init(void)
3932 for_each_possible_cpu(i
)
3933 init_llist_head(&per_cpu(blk_cpu_done
, i
));
3934 open_softirq(BLOCK_SOFTIRQ
, blk_done_softirq
);
3936 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD
,
3937 "block/softirq:dead", NULL
,
3938 blk_softirq_cpu_dead
);
3939 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
3940 blk_mq_hctx_notify_dead
);
3941 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE
, "block/mq:online",
3942 blk_mq_hctx_notify_online
,
3943 blk_mq_hctx_notify_offline
);
3946 subsys_initcall(blk_mq_init
);