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 list_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 hd_struct
*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 (rq
->part
== mi
->part
&& blk_mq_rq_state(rq
) == MQ_RQ_IN_FLIGHT
)
109 mi
->inflight
[rq_data_dir(rq
)]++;
114 unsigned int blk_mq_in_flight(struct request_queue
*q
, struct hd_struct
*part
)
116 struct mq_inflight mi
= { .part
= part
};
118 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight
, &mi
);
120 return mi
.inflight
[0] + mi
.inflight
[1];
123 void blk_mq_in_flight_rw(struct request_queue
*q
, struct hd_struct
*part
,
124 unsigned int inflight
[2])
126 struct mq_inflight mi
= { .part
= part
};
128 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight
, &mi
);
129 inflight
[0] = mi
.inflight
[0];
130 inflight
[1] = mi
.inflight
[1];
133 void blk_freeze_queue_start(struct request_queue
*q
)
135 mutex_lock(&q
->mq_freeze_lock
);
136 if (++q
->mq_freeze_depth
== 1) {
137 percpu_ref_kill(&q
->q_usage_counter
);
138 mutex_unlock(&q
->mq_freeze_lock
);
140 blk_mq_run_hw_queues(q
, false);
142 mutex_unlock(&q
->mq_freeze_lock
);
145 EXPORT_SYMBOL_GPL(blk_freeze_queue_start
);
147 void blk_mq_freeze_queue_wait(struct request_queue
*q
)
149 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
151 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait
);
153 int blk_mq_freeze_queue_wait_timeout(struct request_queue
*q
,
154 unsigned long timeout
)
156 return wait_event_timeout(q
->mq_freeze_wq
,
157 percpu_ref_is_zero(&q
->q_usage_counter
),
160 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout
);
163 * Guarantee no request is in use, so we can change any data structure of
164 * the queue afterward.
166 void blk_freeze_queue(struct request_queue
*q
)
169 * In the !blk_mq case we are only calling this to kill the
170 * q_usage_counter, otherwise this increases the freeze depth
171 * and waits for it to return to zero. For this reason there is
172 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
173 * exported to drivers as the only user for unfreeze is blk_mq.
175 blk_freeze_queue_start(q
);
176 blk_mq_freeze_queue_wait(q
);
179 void blk_mq_freeze_queue(struct request_queue
*q
)
182 * ...just an alias to keep freeze and unfreeze actions balanced
183 * in the blk_mq_* namespace
187 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
189 void blk_mq_unfreeze_queue(struct request_queue
*q
)
191 mutex_lock(&q
->mq_freeze_lock
);
192 q
->mq_freeze_depth
--;
193 WARN_ON_ONCE(q
->mq_freeze_depth
< 0);
194 if (!q
->mq_freeze_depth
) {
195 percpu_ref_resurrect(&q
->q_usage_counter
);
196 wake_up_all(&q
->mq_freeze_wq
);
198 mutex_unlock(&q
->mq_freeze_lock
);
200 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
203 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
204 * mpt3sas driver such that this function can be removed.
206 void blk_mq_quiesce_queue_nowait(struct request_queue
*q
)
208 blk_queue_flag_set(QUEUE_FLAG_QUIESCED
, q
);
210 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait
);
213 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
216 * Note: this function does not prevent that the struct request end_io()
217 * callback function is invoked. Once this function is returned, we make
218 * sure no dispatch can happen until the queue is unquiesced via
219 * blk_mq_unquiesce_queue().
221 void blk_mq_quiesce_queue(struct request_queue
*q
)
223 struct blk_mq_hw_ctx
*hctx
;
227 blk_mq_quiesce_queue_nowait(q
);
229 queue_for_each_hw_ctx(q
, hctx
, i
) {
230 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
231 synchronize_srcu(hctx
->srcu
);
238 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
241 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
244 * This function recovers queue into the state before quiescing
245 * which is done by blk_mq_quiesce_queue.
247 void blk_mq_unquiesce_queue(struct request_queue
*q
)
249 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED
, q
);
251 /* dispatch requests which are inserted during quiescing */
252 blk_mq_run_hw_queues(q
, true);
254 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue
);
256 void blk_mq_wake_waiters(struct request_queue
*q
)
258 struct blk_mq_hw_ctx
*hctx
;
261 queue_for_each_hw_ctx(q
, hctx
, i
)
262 if (blk_mq_hw_queue_mapped(hctx
))
263 blk_mq_tag_wakeup_all(hctx
->tags
, true);
267 * Only need start/end time stamping if we have iostat or
268 * blk stats enabled, or using an IO scheduler.
270 static inline bool blk_mq_need_time_stamp(struct request
*rq
)
272 return (rq
->rq_flags
& (RQF_IO_STAT
| RQF_STATS
)) || rq
->q
->elevator
;
275 static struct request
*blk_mq_rq_ctx_init(struct blk_mq_alloc_data
*data
,
276 unsigned int tag
, u64 alloc_time_ns
)
278 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
279 struct request
*rq
= tags
->static_rqs
[tag
];
281 if (data
->q
->elevator
) {
282 rq
->tag
= BLK_MQ_NO_TAG
;
283 rq
->internal_tag
= tag
;
286 rq
->internal_tag
= BLK_MQ_NO_TAG
;
289 /* csd/requeue_work/fifo_time is initialized before use */
291 rq
->mq_ctx
= data
->ctx
;
292 rq
->mq_hctx
= data
->hctx
;
294 rq
->cmd_flags
= data
->cmd_flags
;
295 if (data
->flags
& BLK_MQ_REQ_PREEMPT
)
296 rq
->rq_flags
|= RQF_PREEMPT
;
297 if (blk_queue_io_stat(data
->q
))
298 rq
->rq_flags
|= RQF_IO_STAT
;
299 INIT_LIST_HEAD(&rq
->queuelist
);
300 INIT_HLIST_NODE(&rq
->hash
);
301 RB_CLEAR_NODE(&rq
->rb_node
);
304 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
305 rq
->alloc_time_ns
= alloc_time_ns
;
307 if (blk_mq_need_time_stamp(rq
))
308 rq
->start_time_ns
= ktime_get_ns();
310 rq
->start_time_ns
= 0;
311 rq
->io_start_time_ns
= 0;
312 rq
->stats_sectors
= 0;
313 rq
->nr_phys_segments
= 0;
314 #if defined(CONFIG_BLK_DEV_INTEGRITY)
315 rq
->nr_integrity_segments
= 0;
317 blk_crypto_rq_set_defaults(rq
);
318 /* tag was already set */
319 WRITE_ONCE(rq
->deadline
, 0);
324 rq
->end_io_data
= NULL
;
326 data
->ctx
->rq_dispatched
[op_is_sync(data
->cmd_flags
)]++;
327 refcount_set(&rq
->ref
, 1);
329 if (!op_is_flush(data
->cmd_flags
)) {
330 struct elevator_queue
*e
= data
->q
->elevator
;
333 if (e
&& e
->type
->ops
.prepare_request
) {
334 if (e
->type
->icq_cache
)
335 blk_mq_sched_assign_ioc(rq
);
337 e
->type
->ops
.prepare_request(rq
);
338 rq
->rq_flags
|= RQF_ELVPRIV
;
342 data
->hctx
->queued
++;
346 static struct request
*__blk_mq_alloc_request(struct blk_mq_alloc_data
*data
)
348 struct request_queue
*q
= data
->q
;
349 struct elevator_queue
*e
= q
->elevator
;
350 u64 alloc_time_ns
= 0;
353 /* alloc_time includes depth and tag waits */
354 if (blk_queue_rq_alloc_time(q
))
355 alloc_time_ns
= ktime_get_ns();
357 if (data
->cmd_flags
& REQ_NOWAIT
)
358 data
->flags
|= BLK_MQ_REQ_NOWAIT
;
362 * Flush requests are special and go directly to the
363 * dispatch list. Don't include reserved tags in the
364 * limiting, as it isn't useful.
366 if (!op_is_flush(data
->cmd_flags
) &&
367 e
->type
->ops
.limit_depth
&&
368 !(data
->flags
& BLK_MQ_REQ_RESERVED
))
369 e
->type
->ops
.limit_depth(data
->cmd_flags
, data
);
373 data
->ctx
= blk_mq_get_ctx(q
);
374 data
->hctx
= blk_mq_map_queue(q
, data
->cmd_flags
, data
->ctx
);
376 blk_mq_tag_busy(data
->hctx
);
379 * Waiting allocations only fail because of an inactive hctx. In that
380 * case just retry the hctx assignment and tag allocation as CPU hotplug
381 * should have migrated us to an online CPU by now.
383 tag
= blk_mq_get_tag(data
);
384 if (tag
== BLK_MQ_NO_TAG
) {
385 if (data
->flags
& BLK_MQ_REQ_NOWAIT
)
389 * Give up the CPU and sleep for a random short time to ensure
390 * that thread using a realtime scheduling class are migrated
391 * off the CPU, and thus off the hctx that is going away.
396 return blk_mq_rq_ctx_init(data
, tag
, alloc_time_ns
);
399 struct request
*blk_mq_alloc_request(struct request_queue
*q
, unsigned int op
,
400 blk_mq_req_flags_t flags
)
402 struct blk_mq_alloc_data data
= {
410 ret
= blk_queue_enter(q
, flags
);
414 rq
= __blk_mq_alloc_request(&data
);
418 rq
->__sector
= (sector_t
) -1;
419 rq
->bio
= rq
->biotail
= NULL
;
423 return ERR_PTR(-EWOULDBLOCK
);
425 EXPORT_SYMBOL(blk_mq_alloc_request
);
427 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
,
428 unsigned int op
, blk_mq_req_flags_t flags
, unsigned int hctx_idx
)
430 struct blk_mq_alloc_data data
= {
435 u64 alloc_time_ns
= 0;
440 /* alloc_time includes depth and tag waits */
441 if (blk_queue_rq_alloc_time(q
))
442 alloc_time_ns
= ktime_get_ns();
445 * If the tag allocator sleeps we could get an allocation for a
446 * different hardware context. No need to complicate the low level
447 * allocator for this for the rare use case of a command tied to
450 if (WARN_ON_ONCE(!(flags
& (BLK_MQ_REQ_NOWAIT
| BLK_MQ_REQ_RESERVED
))))
451 return ERR_PTR(-EINVAL
);
453 if (hctx_idx
>= q
->nr_hw_queues
)
454 return ERR_PTR(-EIO
);
456 ret
= blk_queue_enter(q
, flags
);
461 * Check if the hardware context is actually mapped to anything.
462 * If not tell the caller that it should skip this queue.
465 data
.hctx
= q
->queue_hw_ctx
[hctx_idx
];
466 if (!blk_mq_hw_queue_mapped(data
.hctx
))
468 cpu
= cpumask_first_and(data
.hctx
->cpumask
, cpu_online_mask
);
469 data
.ctx
= __blk_mq_get_ctx(q
, cpu
);
472 blk_mq_tag_busy(data
.hctx
);
475 tag
= blk_mq_get_tag(&data
);
476 if (tag
== BLK_MQ_NO_TAG
)
478 return blk_mq_rq_ctx_init(&data
, tag
, alloc_time_ns
);
484 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
486 static void __blk_mq_free_request(struct request
*rq
)
488 struct request_queue
*q
= rq
->q
;
489 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
490 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
491 const int sched_tag
= rq
->internal_tag
;
493 blk_crypto_free_request(rq
);
494 blk_pm_mark_last_busy(rq
);
496 if (rq
->tag
!= BLK_MQ_NO_TAG
)
497 blk_mq_put_tag(hctx
->tags
, ctx
, rq
->tag
);
498 if (sched_tag
!= BLK_MQ_NO_TAG
)
499 blk_mq_put_tag(hctx
->sched_tags
, ctx
, sched_tag
);
500 blk_mq_sched_restart(hctx
);
504 void blk_mq_free_request(struct request
*rq
)
506 struct request_queue
*q
= rq
->q
;
507 struct elevator_queue
*e
= q
->elevator
;
508 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
509 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
511 if (rq
->rq_flags
& RQF_ELVPRIV
) {
512 if (e
&& e
->type
->ops
.finish_request
)
513 e
->type
->ops
.finish_request(rq
);
515 put_io_context(rq
->elv
.icq
->ioc
);
520 ctx
->rq_completed
[rq_is_sync(rq
)]++;
521 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
522 __blk_mq_dec_active_requests(hctx
);
524 if (unlikely(laptop_mode
&& !blk_rq_is_passthrough(rq
)))
525 laptop_io_completion(q
->backing_dev_info
);
529 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
530 if (refcount_dec_and_test(&rq
->ref
))
531 __blk_mq_free_request(rq
);
533 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
535 inline void __blk_mq_end_request(struct request
*rq
, blk_status_t error
)
539 if (blk_mq_need_time_stamp(rq
))
540 now
= ktime_get_ns();
542 if (rq
->rq_flags
& RQF_STATS
) {
543 blk_mq_poll_stats_start(rq
->q
);
544 blk_stat_add(rq
, now
);
547 blk_mq_sched_completed_request(rq
, now
);
549 blk_account_io_done(rq
, now
);
552 rq_qos_done(rq
->q
, rq
);
553 rq
->end_io(rq
, error
);
555 blk_mq_free_request(rq
);
558 EXPORT_SYMBOL(__blk_mq_end_request
);
560 void blk_mq_end_request(struct request
*rq
, blk_status_t error
)
562 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
564 __blk_mq_end_request(rq
, error
);
566 EXPORT_SYMBOL(blk_mq_end_request
);
569 * Softirq action handler - move entries to local list and loop over them
570 * while passing them to the queue registered handler.
572 static __latent_entropy
void blk_done_softirq(struct softirq_action
*h
)
574 struct list_head
*cpu_list
, local_list
;
577 cpu_list
= this_cpu_ptr(&blk_cpu_done
);
578 list_replace_init(cpu_list
, &local_list
);
581 while (!list_empty(&local_list
)) {
584 rq
= list_entry(local_list
.next
, struct request
, ipi_list
);
585 list_del_init(&rq
->ipi_list
);
586 rq
->q
->mq_ops
->complete(rq
);
590 static void blk_mq_trigger_softirq(struct request
*rq
)
592 struct list_head
*list
;
595 local_irq_save(flags
);
596 list
= this_cpu_ptr(&blk_cpu_done
);
597 list_add_tail(&rq
->ipi_list
, list
);
600 * If the list only contains our just added request, signal a raise of
601 * the softirq. If there are already entries there, someone already
602 * raised the irq but it hasn't run yet.
604 if (list
->next
== &rq
->ipi_list
)
605 raise_softirq_irqoff(BLOCK_SOFTIRQ
);
606 local_irq_restore(flags
);
609 static int blk_softirq_cpu_dead(unsigned int cpu
)
612 * If a CPU goes away, splice its entries to the current CPU
613 * and trigger a run of the softirq
616 list_splice_init(&per_cpu(blk_cpu_done
, cpu
),
617 this_cpu_ptr(&blk_cpu_done
));
618 raise_softirq_irqoff(BLOCK_SOFTIRQ
);
625 static void __blk_mq_complete_request_remote(void *data
)
627 struct request
*rq
= data
;
630 * For most of single queue controllers, there is only one irq vector
631 * for handling I/O completion, and the only irq's affinity is set
632 * to all possible CPUs. On most of ARCHs, this affinity means the irq
633 * is handled on one specific CPU.
635 * So complete I/O requests in softirq context in case of single queue
636 * devices to avoid degrading I/O performance due to irqsoff latency.
638 if (rq
->q
->nr_hw_queues
== 1)
639 blk_mq_trigger_softirq(rq
);
641 rq
->q
->mq_ops
->complete(rq
);
644 static inline bool blk_mq_complete_need_ipi(struct request
*rq
)
646 int cpu
= raw_smp_processor_id();
648 if (!IS_ENABLED(CONFIG_SMP
) ||
649 !test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
))
652 /* same CPU or cache domain? Complete locally */
653 if (cpu
== rq
->mq_ctx
->cpu
||
654 (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
) &&
655 cpus_share_cache(cpu
, rq
->mq_ctx
->cpu
)))
658 /* don't try to IPI to an offline CPU */
659 return cpu_online(rq
->mq_ctx
->cpu
);
662 bool blk_mq_complete_request_remote(struct request
*rq
)
664 WRITE_ONCE(rq
->state
, MQ_RQ_COMPLETE
);
667 * For a polled request, always complete locallly, it's pointless
668 * to redirect the completion.
670 if (rq
->cmd_flags
& REQ_HIPRI
)
673 if (blk_mq_complete_need_ipi(rq
)) {
674 INIT_CSD(&rq
->csd
, __blk_mq_complete_request_remote
, rq
);
675 smp_call_function_single_async(rq
->mq_ctx
->cpu
, &rq
->csd
);
677 if (rq
->q
->nr_hw_queues
> 1)
679 blk_mq_trigger_softirq(rq
);
684 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote
);
687 * blk_mq_complete_request - end I/O on a request
688 * @rq: the request being processed
691 * Complete a request by scheduling the ->complete_rq operation.
693 void blk_mq_complete_request(struct request
*rq
)
695 if (!blk_mq_complete_request_remote(rq
))
696 rq
->q
->mq_ops
->complete(rq
);
698 EXPORT_SYMBOL(blk_mq_complete_request
);
700 static void hctx_unlock(struct blk_mq_hw_ctx
*hctx
, int srcu_idx
)
701 __releases(hctx
->srcu
)
703 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
))
706 srcu_read_unlock(hctx
->srcu
, srcu_idx
);
709 static void hctx_lock(struct blk_mq_hw_ctx
*hctx
, int *srcu_idx
)
710 __acquires(hctx
->srcu
)
712 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
713 /* shut up gcc false positive */
717 *srcu_idx
= srcu_read_lock(hctx
->srcu
);
721 * blk_mq_start_request - Start processing a request
722 * @rq: Pointer to request to be started
724 * Function used by device drivers to notify the block layer that a request
725 * is going to be processed now, so blk layer can do proper initializations
726 * such as starting the timeout timer.
728 void blk_mq_start_request(struct request
*rq
)
730 struct request_queue
*q
= rq
->q
;
732 trace_block_rq_issue(q
, rq
);
734 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
735 rq
->io_start_time_ns
= ktime_get_ns();
736 rq
->stats_sectors
= blk_rq_sectors(rq
);
737 rq
->rq_flags
|= RQF_STATS
;
741 WARN_ON_ONCE(blk_mq_rq_state(rq
) != MQ_RQ_IDLE
);
744 WRITE_ONCE(rq
->state
, MQ_RQ_IN_FLIGHT
);
746 #ifdef CONFIG_BLK_DEV_INTEGRITY
747 if (blk_integrity_rq(rq
) && req_op(rq
) == REQ_OP_WRITE
)
748 q
->integrity
.profile
->prepare_fn(rq
);
751 EXPORT_SYMBOL(blk_mq_start_request
);
753 static void __blk_mq_requeue_request(struct request
*rq
)
755 struct request_queue
*q
= rq
->q
;
757 blk_mq_put_driver_tag(rq
);
759 trace_block_rq_requeue(q
, rq
);
760 rq_qos_requeue(q
, rq
);
762 if (blk_mq_request_started(rq
)) {
763 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
764 rq
->rq_flags
&= ~RQF_TIMED_OUT
;
768 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
770 __blk_mq_requeue_request(rq
);
772 /* this request will be re-inserted to io scheduler queue */
773 blk_mq_sched_requeue_request(rq
);
775 BUG_ON(!list_empty(&rq
->queuelist
));
776 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
778 EXPORT_SYMBOL(blk_mq_requeue_request
);
780 static void blk_mq_requeue_work(struct work_struct
*work
)
782 struct request_queue
*q
=
783 container_of(work
, struct request_queue
, requeue_work
.work
);
785 struct request
*rq
, *next
;
787 spin_lock_irq(&q
->requeue_lock
);
788 list_splice_init(&q
->requeue_list
, &rq_list
);
789 spin_unlock_irq(&q
->requeue_lock
);
791 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
792 if (!(rq
->rq_flags
& (RQF_SOFTBARRIER
| RQF_DONTPREP
)))
795 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
796 list_del_init(&rq
->queuelist
);
798 * If RQF_DONTPREP, rq has contained some driver specific
799 * data, so insert it to hctx dispatch list to avoid any
802 if (rq
->rq_flags
& RQF_DONTPREP
)
803 blk_mq_request_bypass_insert(rq
, false, false);
805 blk_mq_sched_insert_request(rq
, true, false, false);
808 while (!list_empty(&rq_list
)) {
809 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
810 list_del_init(&rq
->queuelist
);
811 blk_mq_sched_insert_request(rq
, false, false, false);
814 blk_mq_run_hw_queues(q
, false);
817 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
818 bool kick_requeue_list
)
820 struct request_queue
*q
= rq
->q
;
824 * We abuse this flag that is otherwise used by the I/O scheduler to
825 * request head insertion from the workqueue.
827 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
829 spin_lock_irqsave(&q
->requeue_lock
, flags
);
831 rq
->rq_flags
|= RQF_SOFTBARRIER
;
832 list_add(&rq
->queuelist
, &q
->requeue_list
);
834 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
836 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
838 if (kick_requeue_list
)
839 blk_mq_kick_requeue_list(q
);
842 void blk_mq_kick_requeue_list(struct request_queue
*q
)
844 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
, 0);
846 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
848 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
851 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
,
852 msecs_to_jiffies(msecs
));
854 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
856 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
858 if (tag
< tags
->nr_tags
) {
859 prefetch(tags
->rqs
[tag
]);
860 return tags
->rqs
[tag
];
865 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
867 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
868 void *priv
, bool reserved
)
871 * If we find a request that isn't idle and the queue matches,
872 * we know the queue is busy. Return false to stop the iteration.
874 if (blk_mq_request_started(rq
) && rq
->q
== hctx
->queue
) {
884 bool blk_mq_queue_inflight(struct request_queue
*q
)
888 blk_mq_queue_tag_busy_iter(q
, blk_mq_rq_inflight
, &busy
);
891 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight
);
893 static void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
895 req
->rq_flags
|= RQF_TIMED_OUT
;
896 if (req
->q
->mq_ops
->timeout
) {
897 enum blk_eh_timer_return ret
;
899 ret
= req
->q
->mq_ops
->timeout(req
, reserved
);
900 if (ret
== BLK_EH_DONE
)
902 WARN_ON_ONCE(ret
!= BLK_EH_RESET_TIMER
);
908 static bool blk_mq_req_expired(struct request
*rq
, unsigned long *next
)
910 unsigned long deadline
;
912 if (blk_mq_rq_state(rq
) != MQ_RQ_IN_FLIGHT
)
914 if (rq
->rq_flags
& RQF_TIMED_OUT
)
917 deadline
= READ_ONCE(rq
->deadline
);
918 if (time_after_eq(jiffies
, deadline
))
923 else if (time_after(*next
, deadline
))
928 static bool blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
929 struct request
*rq
, void *priv
, bool reserved
)
931 unsigned long *next
= priv
;
934 * Just do a quick check if it is expired before locking the request in
935 * so we're not unnecessarilly synchronizing across CPUs.
937 if (!blk_mq_req_expired(rq
, next
))
941 * We have reason to believe the request may be expired. Take a
942 * reference on the request to lock this request lifetime into its
943 * currently allocated context to prevent it from being reallocated in
944 * the event the completion by-passes this timeout handler.
946 * If the reference was already released, then the driver beat the
947 * timeout handler to posting a natural completion.
949 if (!refcount_inc_not_zero(&rq
->ref
))
953 * The request is now locked and cannot be reallocated underneath the
954 * timeout handler's processing. Re-verify this exact request is truly
955 * expired; if it is not expired, then the request was completed and
956 * reallocated as a new request.
958 if (blk_mq_req_expired(rq
, next
))
959 blk_mq_rq_timed_out(rq
, reserved
);
961 if (is_flush_rq(rq
, hctx
))
963 else if (refcount_dec_and_test(&rq
->ref
))
964 __blk_mq_free_request(rq
);
969 static void blk_mq_timeout_work(struct work_struct
*work
)
971 struct request_queue
*q
=
972 container_of(work
, struct request_queue
, timeout_work
);
973 unsigned long next
= 0;
974 struct blk_mq_hw_ctx
*hctx
;
977 /* A deadlock might occur if a request is stuck requiring a
978 * timeout at the same time a queue freeze is waiting
979 * completion, since the timeout code would not be able to
980 * acquire the queue reference here.
982 * That's why we don't use blk_queue_enter here; instead, we use
983 * percpu_ref_tryget directly, because we need to be able to
984 * obtain a reference even in the short window between the queue
985 * starting to freeze, by dropping the first reference in
986 * blk_freeze_queue_start, and the moment the last request is
987 * consumed, marked by the instant q_usage_counter reaches
990 if (!percpu_ref_tryget(&q
->q_usage_counter
))
993 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &next
);
996 mod_timer(&q
->timeout
, next
);
999 * Request timeouts are handled as a forward rolling timer. If
1000 * we end up here it means that no requests are pending and
1001 * also that no request has been pending for a while. Mark
1002 * each hctx as idle.
1004 queue_for_each_hw_ctx(q
, hctx
, i
) {
1005 /* the hctx may be unmapped, so check it here */
1006 if (blk_mq_hw_queue_mapped(hctx
))
1007 blk_mq_tag_idle(hctx
);
1013 struct flush_busy_ctx_data
{
1014 struct blk_mq_hw_ctx
*hctx
;
1015 struct list_head
*list
;
1018 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
1020 struct flush_busy_ctx_data
*flush_data
= data
;
1021 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
1022 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
1023 enum hctx_type type
= hctx
->type
;
1025 spin_lock(&ctx
->lock
);
1026 list_splice_tail_init(&ctx
->rq_lists
[type
], flush_data
->list
);
1027 sbitmap_clear_bit(sb
, bitnr
);
1028 spin_unlock(&ctx
->lock
);
1033 * Process software queues that have been marked busy, splicing them
1034 * to the for-dispatch
1036 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
1038 struct flush_busy_ctx_data data
= {
1043 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
1045 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
1047 struct dispatch_rq_data
{
1048 struct blk_mq_hw_ctx
*hctx
;
1052 static bool dispatch_rq_from_ctx(struct sbitmap
*sb
, unsigned int bitnr
,
1055 struct dispatch_rq_data
*dispatch_data
= data
;
1056 struct blk_mq_hw_ctx
*hctx
= dispatch_data
->hctx
;
1057 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
1058 enum hctx_type type
= hctx
->type
;
1060 spin_lock(&ctx
->lock
);
1061 if (!list_empty(&ctx
->rq_lists
[type
])) {
1062 dispatch_data
->rq
= list_entry_rq(ctx
->rq_lists
[type
].next
);
1063 list_del_init(&dispatch_data
->rq
->queuelist
);
1064 if (list_empty(&ctx
->rq_lists
[type
]))
1065 sbitmap_clear_bit(sb
, bitnr
);
1067 spin_unlock(&ctx
->lock
);
1069 return !dispatch_data
->rq
;
1072 struct request
*blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx
*hctx
,
1073 struct blk_mq_ctx
*start
)
1075 unsigned off
= start
? start
->index_hw
[hctx
->type
] : 0;
1076 struct dispatch_rq_data data
= {
1081 __sbitmap_for_each_set(&hctx
->ctx_map
, off
,
1082 dispatch_rq_from_ctx
, &data
);
1087 static inline unsigned int queued_to_index(unsigned int queued
)
1092 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
1095 static bool __blk_mq_get_driver_tag(struct request
*rq
)
1097 struct sbitmap_queue
*bt
= rq
->mq_hctx
->tags
->bitmap_tags
;
1098 unsigned int tag_offset
= rq
->mq_hctx
->tags
->nr_reserved_tags
;
1101 blk_mq_tag_busy(rq
->mq_hctx
);
1103 if (blk_mq_tag_is_reserved(rq
->mq_hctx
->sched_tags
, rq
->internal_tag
)) {
1104 bt
= rq
->mq_hctx
->tags
->breserved_tags
;
1107 if (!hctx_may_queue(rq
->mq_hctx
, bt
))
1111 tag
= __sbitmap_queue_get(bt
);
1112 if (tag
== BLK_MQ_NO_TAG
)
1115 rq
->tag
= tag
+ tag_offset
;
1119 static bool blk_mq_get_driver_tag(struct request
*rq
)
1121 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1123 if (rq
->tag
== BLK_MQ_NO_TAG
&& !__blk_mq_get_driver_tag(rq
))
1126 if ((hctx
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
) &&
1127 !(rq
->rq_flags
& RQF_MQ_INFLIGHT
)) {
1128 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
1129 __blk_mq_inc_active_requests(hctx
);
1131 hctx
->tags
->rqs
[rq
->tag
] = rq
;
1135 static int blk_mq_dispatch_wake(wait_queue_entry_t
*wait
, unsigned mode
,
1136 int flags
, void *key
)
1138 struct blk_mq_hw_ctx
*hctx
;
1140 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
1142 spin_lock(&hctx
->dispatch_wait_lock
);
1143 if (!list_empty(&wait
->entry
)) {
1144 struct sbitmap_queue
*sbq
;
1146 list_del_init(&wait
->entry
);
1147 sbq
= hctx
->tags
->bitmap_tags
;
1148 atomic_dec(&sbq
->ws_active
);
1150 spin_unlock(&hctx
->dispatch_wait_lock
);
1152 blk_mq_run_hw_queue(hctx
, true);
1157 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1158 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1159 * restart. For both cases, take care to check the condition again after
1160 * marking us as waiting.
1162 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx
*hctx
,
1165 struct sbitmap_queue
*sbq
= hctx
->tags
->bitmap_tags
;
1166 struct wait_queue_head
*wq
;
1167 wait_queue_entry_t
*wait
;
1170 if (!(hctx
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
)) {
1171 blk_mq_sched_mark_restart_hctx(hctx
);
1174 * It's possible that a tag was freed in the window between the
1175 * allocation failure and adding the hardware queue to the wait
1178 * Don't clear RESTART here, someone else could have set it.
1179 * At most this will cost an extra queue run.
1181 return blk_mq_get_driver_tag(rq
);
1184 wait
= &hctx
->dispatch_wait
;
1185 if (!list_empty_careful(&wait
->entry
))
1188 wq
= &bt_wait_ptr(sbq
, hctx
)->wait
;
1190 spin_lock_irq(&wq
->lock
);
1191 spin_lock(&hctx
->dispatch_wait_lock
);
1192 if (!list_empty(&wait
->entry
)) {
1193 spin_unlock(&hctx
->dispatch_wait_lock
);
1194 spin_unlock_irq(&wq
->lock
);
1198 atomic_inc(&sbq
->ws_active
);
1199 wait
->flags
&= ~WQ_FLAG_EXCLUSIVE
;
1200 __add_wait_queue(wq
, wait
);
1203 * It's possible that a tag was freed in the window between the
1204 * allocation failure and adding the hardware queue to the wait
1207 ret
= blk_mq_get_driver_tag(rq
);
1209 spin_unlock(&hctx
->dispatch_wait_lock
);
1210 spin_unlock_irq(&wq
->lock
);
1215 * We got a tag, remove ourselves from the wait queue to ensure
1216 * someone else gets the wakeup.
1218 list_del_init(&wait
->entry
);
1219 atomic_dec(&sbq
->ws_active
);
1220 spin_unlock(&hctx
->dispatch_wait_lock
);
1221 spin_unlock_irq(&wq
->lock
);
1226 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1227 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1229 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1230 * - EWMA is one simple way to compute running average value
1231 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1232 * - take 4 as factor for avoiding to get too small(0) result, and this
1233 * factor doesn't matter because EWMA decreases exponentially
1235 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx
*hctx
, bool busy
)
1239 if (hctx
->queue
->elevator
)
1242 ewma
= hctx
->dispatch_busy
;
1247 ewma
*= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
- 1;
1249 ewma
+= 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR
;
1250 ewma
/= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
;
1252 hctx
->dispatch_busy
= ewma
;
1255 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1257 static void blk_mq_handle_dev_resource(struct request
*rq
,
1258 struct list_head
*list
)
1260 struct request
*next
=
1261 list_first_entry_or_null(list
, struct request
, queuelist
);
1264 * If an I/O scheduler has been configured and we got a driver tag for
1265 * the next request already, free it.
1268 blk_mq_put_driver_tag(next
);
1270 list_add(&rq
->queuelist
, list
);
1271 __blk_mq_requeue_request(rq
);
1274 static void blk_mq_handle_zone_resource(struct request
*rq
,
1275 struct list_head
*zone_list
)
1278 * If we end up here it is because we cannot dispatch a request to a
1279 * specific zone due to LLD level zone-write locking or other zone
1280 * related resource not being available. In this case, set the request
1281 * aside in zone_list for retrying it later.
1283 list_add(&rq
->queuelist
, zone_list
);
1284 __blk_mq_requeue_request(rq
);
1287 enum prep_dispatch
{
1289 PREP_DISPATCH_NO_TAG
,
1290 PREP_DISPATCH_NO_BUDGET
,
1293 static enum prep_dispatch
blk_mq_prep_dispatch_rq(struct request
*rq
,
1296 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1298 if (need_budget
&& !blk_mq_get_dispatch_budget(rq
->q
)) {
1299 blk_mq_put_driver_tag(rq
);
1300 return PREP_DISPATCH_NO_BUDGET
;
1303 if (!blk_mq_get_driver_tag(rq
)) {
1305 * The initial allocation attempt failed, so we need to
1306 * rerun the hardware queue when a tag is freed. The
1307 * waitqueue takes care of that. If the queue is run
1308 * before we add this entry back on the dispatch list,
1309 * we'll re-run it below.
1311 if (!blk_mq_mark_tag_wait(hctx
, rq
)) {
1313 * All budgets not got from this function will be put
1314 * together during handling partial dispatch
1317 blk_mq_put_dispatch_budget(rq
->q
);
1318 return PREP_DISPATCH_NO_TAG
;
1322 return PREP_DISPATCH_OK
;
1325 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1326 static void blk_mq_release_budgets(struct request_queue
*q
,
1327 unsigned int nr_budgets
)
1331 for (i
= 0; i
< nr_budgets
; i
++)
1332 blk_mq_put_dispatch_budget(q
);
1336 * Returns true if we did some work AND can potentially do more.
1338 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
,
1339 unsigned int nr_budgets
)
1341 enum prep_dispatch prep
;
1342 struct request_queue
*q
= hctx
->queue
;
1343 struct request
*rq
, *nxt
;
1345 blk_status_t ret
= BLK_STS_OK
;
1346 LIST_HEAD(zone_list
);
1348 if (list_empty(list
))
1352 * Now process all the entries, sending them to the driver.
1354 errors
= queued
= 0;
1356 struct blk_mq_queue_data bd
;
1358 rq
= list_first_entry(list
, struct request
, queuelist
);
1360 WARN_ON_ONCE(hctx
!= rq
->mq_hctx
);
1361 prep
= blk_mq_prep_dispatch_rq(rq
, !nr_budgets
);
1362 if (prep
!= PREP_DISPATCH_OK
)
1365 list_del_init(&rq
->queuelist
);
1370 * Flag last if we have no more requests, or if we have more
1371 * but can't assign a driver tag to it.
1373 if (list_empty(list
))
1376 nxt
= list_first_entry(list
, struct request
, queuelist
);
1377 bd
.last
= !blk_mq_get_driver_tag(nxt
);
1381 * once the request is queued to lld, no need to cover the
1386 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1391 case BLK_STS_RESOURCE
:
1392 case BLK_STS_DEV_RESOURCE
:
1393 blk_mq_handle_dev_resource(rq
, list
);
1395 case BLK_STS_ZONE_RESOURCE
:
1397 * Move the request to zone_list and keep going through
1398 * the dispatch list to find more requests the drive can
1401 blk_mq_handle_zone_resource(rq
, &zone_list
);
1405 blk_mq_end_request(rq
, BLK_STS_IOERR
);
1407 } while (!list_empty(list
));
1409 if (!list_empty(&zone_list
))
1410 list_splice_tail_init(&zone_list
, list
);
1412 hctx
->dispatched
[queued_to_index(queued
)]++;
1414 /* If we didn't flush the entire list, we could have told the driver
1415 * there was more coming, but that turned out to be a lie.
1417 if ((!list_empty(list
) || errors
) && q
->mq_ops
->commit_rqs
&& queued
)
1418 q
->mq_ops
->commit_rqs(hctx
);
1420 * Any items that need requeuing? Stuff them into hctx->dispatch,
1421 * that is where we will continue on next queue run.
1423 if (!list_empty(list
)) {
1425 /* For non-shared tags, the RESTART check will suffice */
1426 bool no_tag
= prep
== PREP_DISPATCH_NO_TAG
&&
1427 (hctx
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
);
1428 bool no_budget_avail
= prep
== PREP_DISPATCH_NO_BUDGET
;
1430 blk_mq_release_budgets(q
, nr_budgets
);
1432 spin_lock(&hctx
->lock
);
1433 list_splice_tail_init(list
, &hctx
->dispatch
);
1434 spin_unlock(&hctx
->lock
);
1437 * Order adding requests to hctx->dispatch and checking
1438 * SCHED_RESTART flag. The pair of this smp_mb() is the one
1439 * in blk_mq_sched_restart(). Avoid restart code path to
1440 * miss the new added requests to hctx->dispatch, meantime
1441 * SCHED_RESTART is observed here.
1446 * If SCHED_RESTART was set by the caller of this function and
1447 * it is no longer set that means that it was cleared by another
1448 * thread and hence that a queue rerun is needed.
1450 * If 'no_tag' is set, that means that we failed getting
1451 * a driver tag with an I/O scheduler attached. If our dispatch
1452 * waitqueue is no longer active, ensure that we run the queue
1453 * AFTER adding our entries back to the list.
1455 * If no I/O scheduler has been configured it is possible that
1456 * the hardware queue got stopped and restarted before requests
1457 * were pushed back onto the dispatch list. Rerun the queue to
1458 * avoid starvation. Notes:
1459 * - blk_mq_run_hw_queue() checks whether or not a queue has
1460 * been stopped before rerunning a queue.
1461 * - Some but not all block drivers stop a queue before
1462 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1465 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1466 * bit is set, run queue after a delay to avoid IO stalls
1467 * that could otherwise occur if the queue is idle. We'll do
1468 * similar if we couldn't get budget and SCHED_RESTART is set.
1470 needs_restart
= blk_mq_sched_needs_restart(hctx
);
1471 if (!needs_restart
||
1472 (no_tag
&& list_empty_careful(&hctx
->dispatch_wait
.entry
)))
1473 blk_mq_run_hw_queue(hctx
, true);
1474 else if (needs_restart
&& (ret
== BLK_STS_RESOURCE
||
1476 blk_mq_delay_run_hw_queue(hctx
, BLK_MQ_RESOURCE_DELAY
);
1478 blk_mq_update_dispatch_busy(hctx
, true);
1481 blk_mq_update_dispatch_busy(hctx
, false);
1483 return (queued
+ errors
) != 0;
1487 * __blk_mq_run_hw_queue - Run a hardware queue.
1488 * @hctx: Pointer to the hardware queue to run.
1490 * Send pending requests to the hardware.
1492 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1497 * We should be running this queue from one of the CPUs that
1500 * There are at least two related races now between setting
1501 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1502 * __blk_mq_run_hw_queue():
1504 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1505 * but later it becomes online, then this warning is harmless
1508 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1509 * but later it becomes offline, then the warning can't be
1510 * triggered, and we depend on blk-mq timeout handler to
1511 * handle dispatched requests to this hctx
1513 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
1514 cpu_online(hctx
->next_cpu
)) {
1515 printk(KERN_WARNING
"run queue from wrong CPU %d, hctx %s\n",
1516 raw_smp_processor_id(),
1517 cpumask_empty(hctx
->cpumask
) ? "inactive": "active");
1522 * We can't run the queue inline with ints disabled. Ensure that
1523 * we catch bad users of this early.
1525 WARN_ON_ONCE(in_interrupt());
1527 might_sleep_if(hctx
->flags
& BLK_MQ_F_BLOCKING
);
1529 hctx_lock(hctx
, &srcu_idx
);
1530 blk_mq_sched_dispatch_requests(hctx
);
1531 hctx_unlock(hctx
, srcu_idx
);
1534 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx
*hctx
)
1536 int cpu
= cpumask_first_and(hctx
->cpumask
, cpu_online_mask
);
1538 if (cpu
>= nr_cpu_ids
)
1539 cpu
= cpumask_first(hctx
->cpumask
);
1544 * It'd be great if the workqueue API had a way to pass
1545 * in a mask and had some smarts for more clever placement.
1546 * For now we just round-robin here, switching for every
1547 * BLK_MQ_CPU_WORK_BATCH queued items.
1549 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1552 int next_cpu
= hctx
->next_cpu
;
1554 if (hctx
->queue
->nr_hw_queues
== 1)
1555 return WORK_CPU_UNBOUND
;
1557 if (--hctx
->next_cpu_batch
<= 0) {
1559 next_cpu
= cpumask_next_and(next_cpu
, hctx
->cpumask
,
1561 if (next_cpu
>= nr_cpu_ids
)
1562 next_cpu
= blk_mq_first_mapped_cpu(hctx
);
1563 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1567 * Do unbound schedule if we can't find a online CPU for this hctx,
1568 * and it should only happen in the path of handling CPU DEAD.
1570 if (!cpu_online(next_cpu
)) {
1577 * Make sure to re-select CPU next time once after CPUs
1578 * in hctx->cpumask become online again.
1580 hctx
->next_cpu
= next_cpu
;
1581 hctx
->next_cpu_batch
= 1;
1582 return WORK_CPU_UNBOUND
;
1585 hctx
->next_cpu
= next_cpu
;
1590 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
1591 * @hctx: Pointer to the hardware queue to run.
1592 * @async: If we want to run the queue asynchronously.
1593 * @msecs: Microseconds of delay to wait before running the queue.
1595 * If !@async, try to run the queue now. Else, run the queue asynchronously and
1596 * with a delay of @msecs.
1598 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
,
1599 unsigned long msecs
)
1601 if (unlikely(blk_mq_hctx_stopped(hctx
)))
1604 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1605 int cpu
= get_cpu();
1606 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1607 __blk_mq_run_hw_queue(hctx
);
1615 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
,
1616 msecs_to_jiffies(msecs
));
1620 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
1621 * @hctx: Pointer to the hardware queue to run.
1622 * @msecs: Microseconds of delay to wait before running the queue.
1624 * Run a hardware queue asynchronously with a delay of @msecs.
1626 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1628 __blk_mq_delay_run_hw_queue(hctx
, true, msecs
);
1630 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue
);
1633 * blk_mq_run_hw_queue - Start to run a hardware queue.
1634 * @hctx: Pointer to the hardware queue to run.
1635 * @async: If we want to run the queue asynchronously.
1637 * Check if the request queue is not in a quiesced state and if there are
1638 * pending requests to be sent. If this is true, run the queue to send requests
1641 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1647 * When queue is quiesced, we may be switching io scheduler, or
1648 * updating nr_hw_queues, or other things, and we can't run queue
1649 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1651 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1654 hctx_lock(hctx
, &srcu_idx
);
1655 need_run
= !blk_queue_quiesced(hctx
->queue
) &&
1656 blk_mq_hctx_has_pending(hctx
);
1657 hctx_unlock(hctx
, srcu_idx
);
1660 __blk_mq_delay_run_hw_queue(hctx
, async
, 0);
1662 EXPORT_SYMBOL(blk_mq_run_hw_queue
);
1665 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
1666 * @q: Pointer to the request queue to run.
1667 * @async: If we want to run the queue asynchronously.
1669 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1671 struct blk_mq_hw_ctx
*hctx
;
1674 queue_for_each_hw_ctx(q
, hctx
, i
) {
1675 if (blk_mq_hctx_stopped(hctx
))
1678 blk_mq_run_hw_queue(hctx
, async
);
1681 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1684 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
1685 * @q: Pointer to the request queue to run.
1686 * @msecs: Microseconds of delay to wait before running the queues.
1688 void blk_mq_delay_run_hw_queues(struct request_queue
*q
, unsigned long msecs
)
1690 struct blk_mq_hw_ctx
*hctx
;
1693 queue_for_each_hw_ctx(q
, hctx
, i
) {
1694 if (blk_mq_hctx_stopped(hctx
))
1697 blk_mq_delay_run_hw_queue(hctx
, msecs
);
1700 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues
);
1703 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1704 * @q: request queue.
1706 * The caller is responsible for serializing this function against
1707 * blk_mq_{start,stop}_hw_queue().
1709 bool blk_mq_queue_stopped(struct request_queue
*q
)
1711 struct blk_mq_hw_ctx
*hctx
;
1714 queue_for_each_hw_ctx(q
, hctx
, i
)
1715 if (blk_mq_hctx_stopped(hctx
))
1720 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1723 * This function is often used for pausing .queue_rq() by driver when
1724 * there isn't enough resource or some conditions aren't satisfied, and
1725 * BLK_STS_RESOURCE is usually returned.
1727 * We do not guarantee that dispatch can be drained or blocked
1728 * after blk_mq_stop_hw_queue() returns. Please use
1729 * blk_mq_quiesce_queue() for that requirement.
1731 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1733 cancel_delayed_work(&hctx
->run_work
);
1735 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1737 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1740 * This function is often used for pausing .queue_rq() by driver when
1741 * there isn't enough resource or some conditions aren't satisfied, and
1742 * BLK_STS_RESOURCE is usually returned.
1744 * We do not guarantee that dispatch can be drained or blocked
1745 * after blk_mq_stop_hw_queues() returns. Please use
1746 * blk_mq_quiesce_queue() for that requirement.
1748 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1750 struct blk_mq_hw_ctx
*hctx
;
1753 queue_for_each_hw_ctx(q
, hctx
, i
)
1754 blk_mq_stop_hw_queue(hctx
);
1756 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1758 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1760 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1762 blk_mq_run_hw_queue(hctx
, false);
1764 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1766 void blk_mq_start_hw_queues(struct request_queue
*q
)
1768 struct blk_mq_hw_ctx
*hctx
;
1771 queue_for_each_hw_ctx(q
, hctx
, i
)
1772 blk_mq_start_hw_queue(hctx
);
1774 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1776 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1778 if (!blk_mq_hctx_stopped(hctx
))
1781 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1782 blk_mq_run_hw_queue(hctx
, async
);
1784 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1786 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1788 struct blk_mq_hw_ctx
*hctx
;
1791 queue_for_each_hw_ctx(q
, hctx
, i
)
1792 blk_mq_start_stopped_hw_queue(hctx
, async
);
1794 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1796 static void blk_mq_run_work_fn(struct work_struct
*work
)
1798 struct blk_mq_hw_ctx
*hctx
;
1800 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
1803 * If we are stopped, don't run the queue.
1805 if (blk_mq_hctx_stopped(hctx
))
1808 __blk_mq_run_hw_queue(hctx
);
1811 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1815 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1816 enum hctx_type type
= hctx
->type
;
1818 lockdep_assert_held(&ctx
->lock
);
1820 trace_block_rq_insert(hctx
->queue
, rq
);
1823 list_add(&rq
->queuelist
, &ctx
->rq_lists
[type
]);
1825 list_add_tail(&rq
->queuelist
, &ctx
->rq_lists
[type
]);
1828 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1831 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1833 lockdep_assert_held(&ctx
->lock
);
1835 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1836 blk_mq_hctx_mark_pending(hctx
, ctx
);
1840 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
1841 * @rq: Pointer to request to be inserted.
1842 * @at_head: true if the request should be inserted at the head of the list.
1843 * @run_queue: If we should run the hardware queue after inserting the request.
1845 * Should only be used carefully, when the caller knows we want to
1846 * bypass a potential IO scheduler on the target device.
1848 void blk_mq_request_bypass_insert(struct request
*rq
, bool at_head
,
1851 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1853 spin_lock(&hctx
->lock
);
1855 list_add(&rq
->queuelist
, &hctx
->dispatch
);
1857 list_add_tail(&rq
->queuelist
, &hctx
->dispatch
);
1858 spin_unlock(&hctx
->lock
);
1861 blk_mq_run_hw_queue(hctx
, false);
1864 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1865 struct list_head
*list
)
1869 enum hctx_type type
= hctx
->type
;
1872 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1875 list_for_each_entry(rq
, list
, queuelist
) {
1876 BUG_ON(rq
->mq_ctx
!= ctx
);
1877 trace_block_rq_insert(hctx
->queue
, rq
);
1880 spin_lock(&ctx
->lock
);
1881 list_splice_tail_init(list
, &ctx
->rq_lists
[type
]);
1882 blk_mq_hctx_mark_pending(hctx
, ctx
);
1883 spin_unlock(&ctx
->lock
);
1886 static int plug_rq_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1888 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1889 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1891 if (rqa
->mq_ctx
!= rqb
->mq_ctx
)
1892 return rqa
->mq_ctx
> rqb
->mq_ctx
;
1893 if (rqa
->mq_hctx
!= rqb
->mq_hctx
)
1894 return rqa
->mq_hctx
> rqb
->mq_hctx
;
1896 return blk_rq_pos(rqa
) > blk_rq_pos(rqb
);
1899 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1903 if (list_empty(&plug
->mq_list
))
1905 list_splice_init(&plug
->mq_list
, &list
);
1907 if (plug
->rq_count
> 2 && plug
->multiple_queues
)
1908 list_sort(NULL
, &list
, plug_rq_cmp
);
1913 struct list_head rq_list
;
1914 struct request
*rq
, *head_rq
= list_entry_rq(list
.next
);
1915 struct list_head
*pos
= &head_rq
->queuelist
; /* skip first */
1916 struct blk_mq_hw_ctx
*this_hctx
= head_rq
->mq_hctx
;
1917 struct blk_mq_ctx
*this_ctx
= head_rq
->mq_ctx
;
1918 unsigned int depth
= 1;
1920 list_for_each_continue(pos
, &list
) {
1921 rq
= list_entry_rq(pos
);
1923 if (rq
->mq_hctx
!= this_hctx
|| rq
->mq_ctx
!= this_ctx
)
1928 list_cut_before(&rq_list
, &list
, pos
);
1929 trace_block_unplug(head_rq
->q
, depth
, !from_schedule
);
1930 blk_mq_sched_insert_requests(this_hctx
, this_ctx
, &rq_list
,
1932 } while(!list_empty(&list
));
1935 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
,
1936 unsigned int nr_segs
)
1940 if (bio
->bi_opf
& REQ_RAHEAD
)
1941 rq
->cmd_flags
|= REQ_FAILFAST_MASK
;
1943 rq
->__sector
= bio
->bi_iter
.bi_sector
;
1944 rq
->write_hint
= bio
->bi_write_hint
;
1945 blk_rq_bio_prep(rq
, bio
, nr_segs
);
1947 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
1948 err
= blk_crypto_rq_bio_prep(rq
, bio
, GFP_NOIO
);
1951 blk_account_io_start(rq
);
1954 static blk_status_t
__blk_mq_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1956 blk_qc_t
*cookie
, bool last
)
1958 struct request_queue
*q
= rq
->q
;
1959 struct blk_mq_queue_data bd
= {
1963 blk_qc_t new_cookie
;
1966 new_cookie
= request_to_qc_t(hctx
, rq
);
1969 * For OK queue, we are done. For error, caller may kill it.
1970 * Any other error (busy), just add it to our list as we
1971 * previously would have done.
1973 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1976 blk_mq_update_dispatch_busy(hctx
, false);
1977 *cookie
= new_cookie
;
1979 case BLK_STS_RESOURCE
:
1980 case BLK_STS_DEV_RESOURCE
:
1981 blk_mq_update_dispatch_busy(hctx
, true);
1982 __blk_mq_requeue_request(rq
);
1985 blk_mq_update_dispatch_busy(hctx
, false);
1986 *cookie
= BLK_QC_T_NONE
;
1993 static blk_status_t
__blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1996 bool bypass_insert
, bool last
)
1998 struct request_queue
*q
= rq
->q
;
1999 bool run_queue
= true;
2002 * RCU or SRCU read lock is needed before checking quiesced flag.
2004 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2005 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2006 * and avoid driver to try to dispatch again.
2008 if (blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(q
)) {
2010 bypass_insert
= false;
2014 if (q
->elevator
&& !bypass_insert
)
2017 if (!blk_mq_get_dispatch_budget(q
))
2020 if (!blk_mq_get_driver_tag(rq
)) {
2021 blk_mq_put_dispatch_budget(q
);
2025 return __blk_mq_issue_directly(hctx
, rq
, cookie
, last
);
2028 return BLK_STS_RESOURCE
;
2030 blk_mq_sched_insert_request(rq
, false, run_queue
, false);
2036 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2037 * @hctx: Pointer of the associated hardware queue.
2038 * @rq: Pointer to request to be sent.
2039 * @cookie: Request queue cookie.
2041 * If the device has enough resources to accept a new request now, send the
2042 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2043 * we can try send it another time in the future. Requests inserted at this
2044 * queue have higher priority.
2046 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
2047 struct request
*rq
, blk_qc_t
*cookie
)
2052 might_sleep_if(hctx
->flags
& BLK_MQ_F_BLOCKING
);
2054 hctx_lock(hctx
, &srcu_idx
);
2056 ret
= __blk_mq_try_issue_directly(hctx
, rq
, cookie
, false, true);
2057 if (ret
== BLK_STS_RESOURCE
|| ret
== BLK_STS_DEV_RESOURCE
)
2058 blk_mq_request_bypass_insert(rq
, false, true);
2059 else if (ret
!= BLK_STS_OK
)
2060 blk_mq_end_request(rq
, ret
);
2062 hctx_unlock(hctx
, srcu_idx
);
2065 blk_status_t
blk_mq_request_issue_directly(struct request
*rq
, bool last
)
2069 blk_qc_t unused_cookie
;
2070 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
2072 hctx_lock(hctx
, &srcu_idx
);
2073 ret
= __blk_mq_try_issue_directly(hctx
, rq
, &unused_cookie
, true, last
);
2074 hctx_unlock(hctx
, srcu_idx
);
2079 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx
*hctx
,
2080 struct list_head
*list
)
2085 while (!list_empty(list
)) {
2087 struct request
*rq
= list_first_entry(list
, struct request
,
2090 list_del_init(&rq
->queuelist
);
2091 ret
= blk_mq_request_issue_directly(rq
, list_empty(list
));
2092 if (ret
!= BLK_STS_OK
) {
2093 if (ret
== BLK_STS_RESOURCE
||
2094 ret
== BLK_STS_DEV_RESOURCE
) {
2095 blk_mq_request_bypass_insert(rq
, false,
2099 blk_mq_end_request(rq
, ret
);
2106 * If we didn't flush the entire list, we could have told
2107 * the driver there was more coming, but that turned out to
2110 if ((!list_empty(list
) || errors
) &&
2111 hctx
->queue
->mq_ops
->commit_rqs
&& queued
)
2112 hctx
->queue
->mq_ops
->commit_rqs(hctx
);
2115 static void blk_add_rq_to_plug(struct blk_plug
*plug
, struct request
*rq
)
2117 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
2119 if (!plug
->multiple_queues
&& !list_is_singular(&plug
->mq_list
)) {
2120 struct request
*tmp
;
2122 tmp
= list_first_entry(&plug
->mq_list
, struct request
,
2124 if (tmp
->q
!= rq
->q
)
2125 plug
->multiple_queues
= true;
2130 * blk_mq_submit_bio - Create and send a request to block device.
2131 * @bio: Bio pointer.
2133 * Builds up a request structure from @q and @bio and send to the device. The
2134 * request may not be queued directly to hardware if:
2135 * * This request can be merged with another one
2136 * * We want to place request at plug queue for possible future merging
2137 * * There is an IO scheduler active at this queue
2139 * It will not queue the request if there is an error with the bio, or at the
2142 * Returns: Request queue cookie.
2144 blk_qc_t
blk_mq_submit_bio(struct bio
*bio
)
2146 struct request_queue
*q
= bio
->bi_disk
->queue
;
2147 const int is_sync
= op_is_sync(bio
->bi_opf
);
2148 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
2149 struct blk_mq_alloc_data data
= {
2153 struct blk_plug
*plug
;
2154 struct request
*same_queue_rq
= NULL
;
2155 unsigned int nr_segs
;
2159 blk_queue_bounce(q
, &bio
);
2160 __blk_queue_split(&bio
, &nr_segs
);
2162 if (!bio_integrity_prep(bio
))
2165 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
2166 blk_attempt_plug_merge(q
, bio
, nr_segs
, &same_queue_rq
))
2169 if (blk_mq_sched_bio_merge(q
, bio
, nr_segs
))
2172 rq_qos_throttle(q
, bio
);
2174 data
.cmd_flags
= bio
->bi_opf
;
2175 rq
= __blk_mq_alloc_request(&data
);
2176 if (unlikely(!rq
)) {
2177 rq_qos_cleanup(q
, bio
);
2178 if (bio
->bi_opf
& REQ_NOWAIT
)
2179 bio_wouldblock_error(bio
);
2183 trace_block_getrq(q
, bio
, bio
->bi_opf
);
2185 rq_qos_track(q
, rq
, bio
);
2187 cookie
= request_to_qc_t(data
.hctx
, rq
);
2189 blk_mq_bio_to_request(rq
, bio
, nr_segs
);
2191 ret
= blk_crypto_init_request(rq
);
2192 if (ret
!= BLK_STS_OK
) {
2193 bio
->bi_status
= ret
;
2195 blk_mq_free_request(rq
);
2196 return BLK_QC_T_NONE
;
2199 plug
= blk_mq_plug(q
, bio
);
2200 if (unlikely(is_flush_fua
)) {
2201 /* Bypass scheduler for flush requests */
2202 blk_insert_flush(rq
);
2203 blk_mq_run_hw_queue(data
.hctx
, true);
2204 } else if (plug
&& (q
->nr_hw_queues
== 1 || q
->mq_ops
->commit_rqs
||
2205 !blk_queue_nonrot(q
))) {
2207 * Use plugging if we have a ->commit_rqs() hook as well, as
2208 * we know the driver uses bd->last in a smart fashion.
2210 * Use normal plugging if this disk is slow HDD, as sequential
2211 * IO may benefit a lot from plug merging.
2213 unsigned int request_count
= plug
->rq_count
;
2214 struct request
*last
= NULL
;
2217 trace_block_plug(q
);
2219 last
= list_entry_rq(plug
->mq_list
.prev
);
2221 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
2222 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
2223 blk_flush_plug_list(plug
, false);
2224 trace_block_plug(q
);
2227 blk_add_rq_to_plug(plug
, rq
);
2228 } else if (q
->elevator
) {
2229 /* Insert the request at the IO scheduler queue */
2230 blk_mq_sched_insert_request(rq
, false, true, true);
2231 } else if (plug
&& !blk_queue_nomerges(q
)) {
2233 * We do limited plugging. If the bio can be merged, do that.
2234 * Otherwise the existing request in the plug list will be
2235 * issued. So the plug list will have one request at most
2236 * The plug list might get flushed before this. If that happens,
2237 * the plug list is empty, and same_queue_rq is invalid.
2239 if (list_empty(&plug
->mq_list
))
2240 same_queue_rq
= NULL
;
2241 if (same_queue_rq
) {
2242 list_del_init(&same_queue_rq
->queuelist
);
2245 blk_add_rq_to_plug(plug
, rq
);
2246 trace_block_plug(q
);
2248 if (same_queue_rq
) {
2249 data
.hctx
= same_queue_rq
->mq_hctx
;
2250 trace_block_unplug(q
, 1, true);
2251 blk_mq_try_issue_directly(data
.hctx
, same_queue_rq
,
2254 } else if ((q
->nr_hw_queues
> 1 && is_sync
) ||
2255 !data
.hctx
->dispatch_busy
) {
2257 * There is no scheduler and we can try to send directly
2260 blk_mq_try_issue_directly(data
.hctx
, rq
, &cookie
);
2263 blk_mq_sched_insert_request(rq
, false, true, true);
2269 return BLK_QC_T_NONE
;
2272 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
2273 unsigned int hctx_idx
)
2277 if (tags
->rqs
&& set
->ops
->exit_request
) {
2280 for (i
= 0; i
< tags
->nr_tags
; i
++) {
2281 struct request
*rq
= tags
->static_rqs
[i
];
2285 set
->ops
->exit_request(set
, rq
, hctx_idx
);
2286 tags
->static_rqs
[i
] = NULL
;
2290 while (!list_empty(&tags
->page_list
)) {
2291 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
2292 list_del_init(&page
->lru
);
2294 * Remove kmemleak object previously allocated in
2295 * blk_mq_alloc_rqs().
2297 kmemleak_free(page_address(page
));
2298 __free_pages(page
, page
->private);
2302 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
, unsigned int flags
)
2306 kfree(tags
->static_rqs
);
2307 tags
->static_rqs
= NULL
;
2309 blk_mq_free_tags(tags
, flags
);
2312 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
2313 unsigned int hctx_idx
,
2314 unsigned int nr_tags
,
2315 unsigned int reserved_tags
,
2318 struct blk_mq_tags
*tags
;
2321 node
= blk_mq_hw_queue_to_node(&set
->map
[HCTX_TYPE_DEFAULT
], hctx_idx
);
2322 if (node
== NUMA_NO_NODE
)
2323 node
= set
->numa_node
;
2325 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
, flags
);
2329 tags
->rqs
= kcalloc_node(nr_tags
, sizeof(struct request
*),
2330 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
2333 blk_mq_free_tags(tags
, flags
);
2337 tags
->static_rqs
= kcalloc_node(nr_tags
, sizeof(struct request
*),
2338 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
2340 if (!tags
->static_rqs
) {
2342 blk_mq_free_tags(tags
, flags
);
2349 static size_t order_to_size(unsigned int order
)
2351 return (size_t)PAGE_SIZE
<< order
;
2354 static int blk_mq_init_request(struct blk_mq_tag_set
*set
, struct request
*rq
,
2355 unsigned int hctx_idx
, int node
)
2359 if (set
->ops
->init_request
) {
2360 ret
= set
->ops
->init_request(set
, rq
, hctx_idx
, node
);
2365 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
2369 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
2370 unsigned int hctx_idx
, unsigned int depth
)
2372 unsigned int i
, j
, entries_per_page
, max_order
= 4;
2373 size_t rq_size
, left
;
2376 node
= blk_mq_hw_queue_to_node(&set
->map
[HCTX_TYPE_DEFAULT
], hctx_idx
);
2377 if (node
== NUMA_NO_NODE
)
2378 node
= set
->numa_node
;
2380 INIT_LIST_HEAD(&tags
->page_list
);
2383 * rq_size is the size of the request plus driver payload, rounded
2384 * to the cacheline size
2386 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
2388 left
= rq_size
* depth
;
2390 for (i
= 0; i
< depth
; ) {
2391 int this_order
= max_order
;
2396 while (this_order
&& left
< order_to_size(this_order
- 1))
2400 page
= alloc_pages_node(node
,
2401 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
2407 if (order_to_size(this_order
) < rq_size
)
2414 page
->private = this_order
;
2415 list_add_tail(&page
->lru
, &tags
->page_list
);
2417 p
= page_address(page
);
2419 * Allow kmemleak to scan these pages as they contain pointers
2420 * to additional allocations like via ops->init_request().
2422 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
2423 entries_per_page
= order_to_size(this_order
) / rq_size
;
2424 to_do
= min(entries_per_page
, depth
- i
);
2425 left
-= to_do
* rq_size
;
2426 for (j
= 0; j
< to_do
; j
++) {
2427 struct request
*rq
= p
;
2429 tags
->static_rqs
[i
] = rq
;
2430 if (blk_mq_init_request(set
, rq
, hctx_idx
, node
)) {
2431 tags
->static_rqs
[i
] = NULL
;
2442 blk_mq_free_rqs(set
, tags
, hctx_idx
);
2446 struct rq_iter_data
{
2447 struct blk_mq_hw_ctx
*hctx
;
2451 static bool blk_mq_has_request(struct request
*rq
, void *data
, bool reserved
)
2453 struct rq_iter_data
*iter_data
= data
;
2455 if (rq
->mq_hctx
!= iter_data
->hctx
)
2457 iter_data
->has_rq
= true;
2461 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx
*hctx
)
2463 struct blk_mq_tags
*tags
= hctx
->sched_tags
?
2464 hctx
->sched_tags
: hctx
->tags
;
2465 struct rq_iter_data data
= {
2469 blk_mq_all_tag_iter(tags
, blk_mq_has_request
, &data
);
2473 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu
,
2474 struct blk_mq_hw_ctx
*hctx
)
2476 if (cpumask_next_and(-1, hctx
->cpumask
, cpu_online_mask
) != cpu
)
2478 if (cpumask_next_and(cpu
, hctx
->cpumask
, cpu_online_mask
) < nr_cpu_ids
)
2483 static int blk_mq_hctx_notify_offline(unsigned int cpu
, struct hlist_node
*node
)
2485 struct blk_mq_hw_ctx
*hctx
= hlist_entry_safe(node
,
2486 struct blk_mq_hw_ctx
, cpuhp_online
);
2488 if (!cpumask_test_cpu(cpu
, hctx
->cpumask
) ||
2489 !blk_mq_last_cpu_in_hctx(cpu
, hctx
))
2493 * Prevent new request from being allocated on the current hctx.
2495 * The smp_mb__after_atomic() Pairs with the implied barrier in
2496 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
2497 * seen once we return from the tag allocator.
2499 set_bit(BLK_MQ_S_INACTIVE
, &hctx
->state
);
2500 smp_mb__after_atomic();
2503 * Try to grab a reference to the queue and wait for any outstanding
2504 * requests. If we could not grab a reference the queue has been
2505 * frozen and there are no requests.
2507 if (percpu_ref_tryget(&hctx
->queue
->q_usage_counter
)) {
2508 while (blk_mq_hctx_has_requests(hctx
))
2510 percpu_ref_put(&hctx
->queue
->q_usage_counter
);
2516 static int blk_mq_hctx_notify_online(unsigned int cpu
, struct hlist_node
*node
)
2518 struct blk_mq_hw_ctx
*hctx
= hlist_entry_safe(node
,
2519 struct blk_mq_hw_ctx
, cpuhp_online
);
2521 if (cpumask_test_cpu(cpu
, hctx
->cpumask
))
2522 clear_bit(BLK_MQ_S_INACTIVE
, &hctx
->state
);
2527 * 'cpu' is going away. splice any existing rq_list entries from this
2528 * software queue to the hw queue dispatch list, and ensure that it
2531 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
2533 struct blk_mq_hw_ctx
*hctx
;
2534 struct blk_mq_ctx
*ctx
;
2536 enum hctx_type type
;
2538 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
2539 if (!cpumask_test_cpu(cpu
, hctx
->cpumask
))
2542 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
2545 spin_lock(&ctx
->lock
);
2546 if (!list_empty(&ctx
->rq_lists
[type
])) {
2547 list_splice_init(&ctx
->rq_lists
[type
], &tmp
);
2548 blk_mq_hctx_clear_pending(hctx
, ctx
);
2550 spin_unlock(&ctx
->lock
);
2552 if (list_empty(&tmp
))
2555 spin_lock(&hctx
->lock
);
2556 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
2557 spin_unlock(&hctx
->lock
);
2559 blk_mq_run_hw_queue(hctx
, true);
2563 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
2565 if (!(hctx
->flags
& BLK_MQ_F_STACKING
))
2566 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE
,
2567 &hctx
->cpuhp_online
);
2568 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
2572 /* hctx->ctxs will be freed in queue's release handler */
2573 static void blk_mq_exit_hctx(struct request_queue
*q
,
2574 struct blk_mq_tag_set
*set
,
2575 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
2577 if (blk_mq_hw_queue_mapped(hctx
))
2578 blk_mq_tag_idle(hctx
);
2580 if (set
->ops
->exit_request
)
2581 set
->ops
->exit_request(set
, hctx
->fq
->flush_rq
, hctx_idx
);
2583 if (set
->ops
->exit_hctx
)
2584 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2586 blk_mq_remove_cpuhp(hctx
);
2588 spin_lock(&q
->unused_hctx_lock
);
2589 list_add(&hctx
->hctx_list
, &q
->unused_hctx_list
);
2590 spin_unlock(&q
->unused_hctx_lock
);
2593 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
2594 struct blk_mq_tag_set
*set
, int nr_queue
)
2596 struct blk_mq_hw_ctx
*hctx
;
2599 queue_for_each_hw_ctx(q
, hctx
, i
) {
2602 blk_mq_debugfs_unregister_hctx(hctx
);
2603 blk_mq_exit_hctx(q
, set
, hctx
, i
);
2607 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set
*tag_set
)
2609 int hw_ctx_size
= sizeof(struct blk_mq_hw_ctx
);
2611 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx
, srcu
),
2612 __alignof__(struct blk_mq_hw_ctx
)) !=
2613 sizeof(struct blk_mq_hw_ctx
));
2615 if (tag_set
->flags
& BLK_MQ_F_BLOCKING
)
2616 hw_ctx_size
+= sizeof(struct srcu_struct
);
2621 static int blk_mq_init_hctx(struct request_queue
*q
,
2622 struct blk_mq_tag_set
*set
,
2623 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
2625 hctx
->queue_num
= hctx_idx
;
2627 if (!(hctx
->flags
& BLK_MQ_F_STACKING
))
2628 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE
,
2629 &hctx
->cpuhp_online
);
2630 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
2632 hctx
->tags
= set
->tags
[hctx_idx
];
2634 if (set
->ops
->init_hctx
&&
2635 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
2636 goto unregister_cpu_notifier
;
2638 if (blk_mq_init_request(set
, hctx
->fq
->flush_rq
, hctx_idx
,
2644 if (set
->ops
->exit_hctx
)
2645 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2646 unregister_cpu_notifier
:
2647 blk_mq_remove_cpuhp(hctx
);
2651 static struct blk_mq_hw_ctx
*
2652 blk_mq_alloc_hctx(struct request_queue
*q
, struct blk_mq_tag_set
*set
,
2655 struct blk_mq_hw_ctx
*hctx
;
2656 gfp_t gfp
= GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
;
2658 hctx
= kzalloc_node(blk_mq_hw_ctx_size(set
), gfp
, node
);
2660 goto fail_alloc_hctx
;
2662 if (!zalloc_cpumask_var_node(&hctx
->cpumask
, gfp
, node
))
2665 atomic_set(&hctx
->nr_active
, 0);
2666 atomic_set(&hctx
->elevator_queued
, 0);
2667 if (node
== NUMA_NO_NODE
)
2668 node
= set
->numa_node
;
2669 hctx
->numa_node
= node
;
2671 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
2672 spin_lock_init(&hctx
->lock
);
2673 INIT_LIST_HEAD(&hctx
->dispatch
);
2675 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_QUEUE_SHARED
;
2677 INIT_LIST_HEAD(&hctx
->hctx_list
);
2680 * Allocate space for all possible cpus to avoid allocation at
2683 hctx
->ctxs
= kmalloc_array_node(nr_cpu_ids
, sizeof(void *),
2688 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8),
2693 spin_lock_init(&hctx
->dispatch_wait_lock
);
2694 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
2695 INIT_LIST_HEAD(&hctx
->dispatch_wait
.entry
);
2697 hctx
->fq
= blk_alloc_flush_queue(hctx
->numa_node
, set
->cmd_size
, gfp
);
2701 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
2702 init_srcu_struct(hctx
->srcu
);
2703 blk_mq_hctx_kobj_init(hctx
);
2708 sbitmap_free(&hctx
->ctx_map
);
2712 free_cpumask_var(hctx
->cpumask
);
2719 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
2720 unsigned int nr_hw_queues
)
2722 struct blk_mq_tag_set
*set
= q
->tag_set
;
2725 for_each_possible_cpu(i
) {
2726 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2727 struct blk_mq_hw_ctx
*hctx
;
2731 spin_lock_init(&__ctx
->lock
);
2732 for (k
= HCTX_TYPE_DEFAULT
; k
< HCTX_MAX_TYPES
; k
++)
2733 INIT_LIST_HEAD(&__ctx
->rq_lists
[k
]);
2738 * Set local node, IFF we have more than one hw queue. If
2739 * not, we remain on the home node of the device
2741 for (j
= 0; j
< set
->nr_maps
; j
++) {
2742 hctx
= blk_mq_map_queue_type(q
, j
, i
);
2743 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
2744 hctx
->numa_node
= cpu_to_node(i
);
2749 static bool __blk_mq_alloc_map_and_request(struct blk_mq_tag_set
*set
,
2752 unsigned int flags
= set
->flags
;
2755 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
2756 set
->queue_depth
, set
->reserved_tags
, flags
);
2757 if (!set
->tags
[hctx_idx
])
2760 ret
= blk_mq_alloc_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
,
2765 blk_mq_free_rq_map(set
->tags
[hctx_idx
], flags
);
2766 set
->tags
[hctx_idx
] = NULL
;
2770 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set
*set
,
2771 unsigned int hctx_idx
)
2773 unsigned int flags
= set
->flags
;
2775 if (set
->tags
&& set
->tags
[hctx_idx
]) {
2776 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
2777 blk_mq_free_rq_map(set
->tags
[hctx_idx
], flags
);
2778 set
->tags
[hctx_idx
] = NULL
;
2782 static void blk_mq_map_swqueue(struct request_queue
*q
)
2784 unsigned int i
, j
, hctx_idx
;
2785 struct blk_mq_hw_ctx
*hctx
;
2786 struct blk_mq_ctx
*ctx
;
2787 struct blk_mq_tag_set
*set
= q
->tag_set
;
2789 queue_for_each_hw_ctx(q
, hctx
, i
) {
2790 cpumask_clear(hctx
->cpumask
);
2792 hctx
->dispatch_from
= NULL
;
2796 * Map software to hardware queues.
2798 * If the cpu isn't present, the cpu is mapped to first hctx.
2800 for_each_possible_cpu(i
) {
2802 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2803 for (j
= 0; j
< set
->nr_maps
; j
++) {
2804 if (!set
->map
[j
].nr_queues
) {
2805 ctx
->hctxs
[j
] = blk_mq_map_queue_type(q
,
2806 HCTX_TYPE_DEFAULT
, i
);
2809 hctx_idx
= set
->map
[j
].mq_map
[i
];
2810 /* unmapped hw queue can be remapped after CPU topo changed */
2811 if (!set
->tags
[hctx_idx
] &&
2812 !__blk_mq_alloc_map_and_request(set
, hctx_idx
)) {
2814 * If tags initialization fail for some hctx,
2815 * that hctx won't be brought online. In this
2816 * case, remap the current ctx to hctx[0] which
2817 * is guaranteed to always have tags allocated
2819 set
->map
[j
].mq_map
[i
] = 0;
2822 hctx
= blk_mq_map_queue_type(q
, j
, i
);
2823 ctx
->hctxs
[j
] = hctx
;
2825 * If the CPU is already set in the mask, then we've
2826 * mapped this one already. This can happen if
2827 * devices share queues across queue maps.
2829 if (cpumask_test_cpu(i
, hctx
->cpumask
))
2832 cpumask_set_cpu(i
, hctx
->cpumask
);
2834 ctx
->index_hw
[hctx
->type
] = hctx
->nr_ctx
;
2835 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2838 * If the nr_ctx type overflows, we have exceeded the
2839 * amount of sw queues we can support.
2841 BUG_ON(!hctx
->nr_ctx
);
2844 for (; j
< HCTX_MAX_TYPES
; j
++)
2845 ctx
->hctxs
[j
] = blk_mq_map_queue_type(q
,
2846 HCTX_TYPE_DEFAULT
, i
);
2849 queue_for_each_hw_ctx(q
, hctx
, i
) {
2851 * If no software queues are mapped to this hardware queue,
2852 * disable it and free the request entries.
2854 if (!hctx
->nr_ctx
) {
2855 /* Never unmap queue 0. We need it as a
2856 * fallback in case of a new remap fails
2859 if (i
&& set
->tags
[i
])
2860 blk_mq_free_map_and_requests(set
, i
);
2866 hctx
->tags
= set
->tags
[i
];
2867 WARN_ON(!hctx
->tags
);
2870 * Set the map size to the number of mapped software queues.
2871 * This is more accurate and more efficient than looping
2872 * over all possibly mapped software queues.
2874 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2877 * Initialize batch roundrobin counts
2879 hctx
->next_cpu
= blk_mq_first_mapped_cpu(hctx
);
2880 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2885 * Caller needs to ensure that we're either frozen/quiesced, or that
2886 * the queue isn't live yet.
2888 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2890 struct blk_mq_hw_ctx
*hctx
;
2893 queue_for_each_hw_ctx(q
, hctx
, i
) {
2895 hctx
->flags
|= BLK_MQ_F_TAG_QUEUE_SHARED
;
2897 hctx
->flags
&= ~BLK_MQ_F_TAG_QUEUE_SHARED
;
2901 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set
*set
,
2904 struct request_queue
*q
;
2906 lockdep_assert_held(&set
->tag_list_lock
);
2908 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2909 blk_mq_freeze_queue(q
);
2910 queue_set_hctx_shared(q
, shared
);
2911 blk_mq_unfreeze_queue(q
);
2915 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
2917 struct blk_mq_tag_set
*set
= q
->tag_set
;
2919 mutex_lock(&set
->tag_list_lock
);
2920 list_del(&q
->tag_set_list
);
2921 if (list_is_singular(&set
->tag_list
)) {
2922 /* just transitioned to unshared */
2923 set
->flags
&= ~BLK_MQ_F_TAG_QUEUE_SHARED
;
2924 /* update existing queue */
2925 blk_mq_update_tag_set_shared(set
, false);
2927 mutex_unlock(&set
->tag_list_lock
);
2928 INIT_LIST_HEAD(&q
->tag_set_list
);
2931 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
2932 struct request_queue
*q
)
2934 mutex_lock(&set
->tag_list_lock
);
2937 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2939 if (!list_empty(&set
->tag_list
) &&
2940 !(set
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
)) {
2941 set
->flags
|= BLK_MQ_F_TAG_QUEUE_SHARED
;
2942 /* update existing queue */
2943 blk_mq_update_tag_set_shared(set
, true);
2945 if (set
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
)
2946 queue_set_hctx_shared(q
, true);
2947 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
2949 mutex_unlock(&set
->tag_list_lock
);
2952 /* All allocations will be freed in release handler of q->mq_kobj */
2953 static int blk_mq_alloc_ctxs(struct request_queue
*q
)
2955 struct blk_mq_ctxs
*ctxs
;
2958 ctxs
= kzalloc(sizeof(*ctxs
), GFP_KERNEL
);
2962 ctxs
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2963 if (!ctxs
->queue_ctx
)
2966 for_each_possible_cpu(cpu
) {
2967 struct blk_mq_ctx
*ctx
= per_cpu_ptr(ctxs
->queue_ctx
, cpu
);
2971 q
->mq_kobj
= &ctxs
->kobj
;
2972 q
->queue_ctx
= ctxs
->queue_ctx
;
2981 * It is the actual release handler for mq, but we do it from
2982 * request queue's release handler for avoiding use-after-free
2983 * and headache because q->mq_kobj shouldn't have been introduced,
2984 * but we can't group ctx/kctx kobj without it.
2986 void blk_mq_release(struct request_queue
*q
)
2988 struct blk_mq_hw_ctx
*hctx
, *next
;
2991 queue_for_each_hw_ctx(q
, hctx
, i
)
2992 WARN_ON_ONCE(hctx
&& list_empty(&hctx
->hctx_list
));
2994 /* all hctx are in .unused_hctx_list now */
2995 list_for_each_entry_safe(hctx
, next
, &q
->unused_hctx_list
, hctx_list
) {
2996 list_del_init(&hctx
->hctx_list
);
2997 kobject_put(&hctx
->kobj
);
3000 kfree(q
->queue_hw_ctx
);
3003 * release .mq_kobj and sw queue's kobject now because
3004 * both share lifetime with request queue.
3006 blk_mq_sysfs_deinit(q
);
3009 struct request_queue
*blk_mq_init_queue_data(struct blk_mq_tag_set
*set
,
3012 struct request_queue
*uninit_q
, *q
;
3014 uninit_q
= blk_alloc_queue(set
->numa_node
);
3016 return ERR_PTR(-ENOMEM
);
3017 uninit_q
->queuedata
= queuedata
;
3020 * Initialize the queue without an elevator. device_add_disk() will do
3021 * the initialization.
3023 q
= blk_mq_init_allocated_queue(set
, uninit_q
, false);
3025 blk_cleanup_queue(uninit_q
);
3029 EXPORT_SYMBOL_GPL(blk_mq_init_queue_data
);
3031 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
3033 return blk_mq_init_queue_data(set
, NULL
);
3035 EXPORT_SYMBOL(blk_mq_init_queue
);
3038 * Helper for setting up a queue with mq ops, given queue depth, and
3039 * the passed in mq ops flags.
3041 struct request_queue
*blk_mq_init_sq_queue(struct blk_mq_tag_set
*set
,
3042 const struct blk_mq_ops
*ops
,
3043 unsigned int queue_depth
,
3044 unsigned int set_flags
)
3046 struct request_queue
*q
;
3049 memset(set
, 0, sizeof(*set
));
3051 set
->nr_hw_queues
= 1;
3053 set
->queue_depth
= queue_depth
;
3054 set
->numa_node
= NUMA_NO_NODE
;
3055 set
->flags
= set_flags
;
3057 ret
= blk_mq_alloc_tag_set(set
);
3059 return ERR_PTR(ret
);
3061 q
= blk_mq_init_queue(set
);
3063 blk_mq_free_tag_set(set
);
3069 EXPORT_SYMBOL(blk_mq_init_sq_queue
);
3071 static struct blk_mq_hw_ctx
*blk_mq_alloc_and_init_hctx(
3072 struct blk_mq_tag_set
*set
, struct request_queue
*q
,
3073 int hctx_idx
, int node
)
3075 struct blk_mq_hw_ctx
*hctx
= NULL
, *tmp
;
3077 /* reuse dead hctx first */
3078 spin_lock(&q
->unused_hctx_lock
);
3079 list_for_each_entry(tmp
, &q
->unused_hctx_list
, hctx_list
) {
3080 if (tmp
->numa_node
== node
) {
3086 list_del_init(&hctx
->hctx_list
);
3087 spin_unlock(&q
->unused_hctx_lock
);
3090 hctx
= blk_mq_alloc_hctx(q
, set
, node
);
3094 if (blk_mq_init_hctx(q
, set
, hctx
, hctx_idx
))
3100 kobject_put(&hctx
->kobj
);
3105 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
3106 struct request_queue
*q
)
3109 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
3111 if (q
->nr_hw_queues
< set
->nr_hw_queues
) {
3112 struct blk_mq_hw_ctx
**new_hctxs
;
3114 new_hctxs
= kcalloc_node(set
->nr_hw_queues
,
3115 sizeof(*new_hctxs
), GFP_KERNEL
,
3120 memcpy(new_hctxs
, hctxs
, q
->nr_hw_queues
*
3122 q
->queue_hw_ctx
= new_hctxs
;
3127 /* protect against switching io scheduler */
3128 mutex_lock(&q
->sysfs_lock
);
3129 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
3131 struct blk_mq_hw_ctx
*hctx
;
3133 node
= blk_mq_hw_queue_to_node(&set
->map
[HCTX_TYPE_DEFAULT
], i
);
3135 * If the hw queue has been mapped to another numa node,
3136 * we need to realloc the hctx. If allocation fails, fallback
3137 * to use the previous one.
3139 if (hctxs
[i
] && (hctxs
[i
]->numa_node
== node
))
3142 hctx
= blk_mq_alloc_and_init_hctx(set
, q
, i
, node
);
3145 blk_mq_exit_hctx(q
, set
, hctxs
[i
], i
);
3149 pr_warn("Allocate new hctx on node %d fails,\
3150 fallback to previous one on node %d\n",
3151 node
, hctxs
[i
]->numa_node
);
3157 * Increasing nr_hw_queues fails. Free the newly allocated
3158 * hctxs and keep the previous q->nr_hw_queues.
3160 if (i
!= set
->nr_hw_queues
) {
3161 j
= q
->nr_hw_queues
;
3165 end
= q
->nr_hw_queues
;
3166 q
->nr_hw_queues
= set
->nr_hw_queues
;
3169 for (; j
< end
; j
++) {
3170 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
3174 blk_mq_free_map_and_requests(set
, j
);
3175 blk_mq_exit_hctx(q
, set
, hctx
, j
);
3179 mutex_unlock(&q
->sysfs_lock
);
3182 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
3183 struct request_queue
*q
,
3186 /* mark the queue as mq asap */
3187 q
->mq_ops
= set
->ops
;
3189 q
->poll_cb
= blk_stat_alloc_callback(blk_mq_poll_stats_fn
,
3190 blk_mq_poll_stats_bkt
,
3191 BLK_MQ_POLL_STATS_BKTS
, q
);
3195 if (blk_mq_alloc_ctxs(q
))
3198 /* init q->mq_kobj and sw queues' kobjects */
3199 blk_mq_sysfs_init(q
);
3201 INIT_LIST_HEAD(&q
->unused_hctx_list
);
3202 spin_lock_init(&q
->unused_hctx_lock
);
3204 blk_mq_realloc_hw_ctxs(set
, q
);
3205 if (!q
->nr_hw_queues
)
3208 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
3209 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
3213 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
3214 if (set
->nr_maps
> HCTX_TYPE_POLL
&&
3215 set
->map
[HCTX_TYPE_POLL
].nr_queues
)
3216 blk_queue_flag_set(QUEUE_FLAG_POLL
, q
);
3218 q
->sg_reserved_size
= INT_MAX
;
3220 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
3221 INIT_LIST_HEAD(&q
->requeue_list
);
3222 spin_lock_init(&q
->requeue_lock
);
3224 q
->nr_requests
= set
->queue_depth
;
3227 * Default to classic polling
3229 q
->poll_nsec
= BLK_MQ_POLL_CLASSIC
;
3231 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
3232 blk_mq_add_queue_tag_set(set
, q
);
3233 blk_mq_map_swqueue(q
);
3236 elevator_init_mq(q
);
3241 kfree(q
->queue_hw_ctx
);
3242 q
->nr_hw_queues
= 0;
3243 blk_mq_sysfs_deinit(q
);
3245 blk_stat_free_callback(q
->poll_cb
);
3249 return ERR_PTR(-ENOMEM
);
3251 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
3253 /* tags can _not_ be used after returning from blk_mq_exit_queue */
3254 void blk_mq_exit_queue(struct request_queue
*q
)
3256 struct blk_mq_tag_set
*set
= q
->tag_set
;
3258 blk_mq_del_queue_tag_set(q
);
3259 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
3262 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
3266 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
3267 if (!__blk_mq_alloc_map_and_request(set
, i
))
3276 blk_mq_free_map_and_requests(set
, i
);
3282 * Allocate the request maps associated with this tag_set. Note that this
3283 * may reduce the depth asked for, if memory is tight. set->queue_depth
3284 * will be updated to reflect the allocated depth.
3286 static int blk_mq_alloc_map_and_requests(struct blk_mq_tag_set
*set
)
3291 depth
= set
->queue_depth
;
3293 err
= __blk_mq_alloc_rq_maps(set
);
3297 set
->queue_depth
>>= 1;
3298 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
3302 } while (set
->queue_depth
);
3304 if (!set
->queue_depth
|| err
) {
3305 pr_err("blk-mq: failed to allocate request map\n");
3309 if (depth
!= set
->queue_depth
)
3310 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3311 depth
, set
->queue_depth
);
3316 static int blk_mq_update_queue_map(struct blk_mq_tag_set
*set
)
3319 * blk_mq_map_queues() and multiple .map_queues() implementations
3320 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
3321 * number of hardware queues.
3323 if (set
->nr_maps
== 1)
3324 set
->map
[HCTX_TYPE_DEFAULT
].nr_queues
= set
->nr_hw_queues
;
3326 if (set
->ops
->map_queues
&& !is_kdump_kernel()) {
3330 * transport .map_queues is usually done in the following
3333 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3334 * mask = get_cpu_mask(queue)
3335 * for_each_cpu(cpu, mask)
3336 * set->map[x].mq_map[cpu] = queue;
3339 * When we need to remap, the table has to be cleared for
3340 * killing stale mapping since one CPU may not be mapped
3343 for (i
= 0; i
< set
->nr_maps
; i
++)
3344 blk_mq_clear_mq_map(&set
->map
[i
]);
3346 return set
->ops
->map_queues(set
);
3348 BUG_ON(set
->nr_maps
> 1);
3349 return blk_mq_map_queues(&set
->map
[HCTX_TYPE_DEFAULT
]);
3353 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set
*set
,
3354 int cur_nr_hw_queues
, int new_nr_hw_queues
)
3356 struct blk_mq_tags
**new_tags
;
3358 if (cur_nr_hw_queues
>= new_nr_hw_queues
)
3361 new_tags
= kcalloc_node(new_nr_hw_queues
, sizeof(struct blk_mq_tags
*),
3362 GFP_KERNEL
, set
->numa_node
);
3367 memcpy(new_tags
, set
->tags
, cur_nr_hw_queues
*
3368 sizeof(*set
->tags
));
3370 set
->tags
= new_tags
;
3371 set
->nr_hw_queues
= new_nr_hw_queues
;
3377 * Alloc a tag set to be associated with one or more request queues.
3378 * May fail with EINVAL for various error conditions. May adjust the
3379 * requested depth down, if it's too large. In that case, the set
3380 * value will be stored in set->queue_depth.
3382 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
3386 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
3388 if (!set
->nr_hw_queues
)
3390 if (!set
->queue_depth
)
3392 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
3395 if (!set
->ops
->queue_rq
)
3398 if (!set
->ops
->get_budget
^ !set
->ops
->put_budget
)
3401 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
3402 pr_info("blk-mq: reduced tag depth to %u\n",
3404 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
3409 else if (set
->nr_maps
> HCTX_MAX_TYPES
)
3413 * If a crashdump is active, then we are potentially in a very
3414 * memory constrained environment. Limit us to 1 queue and
3415 * 64 tags to prevent using too much memory.
3417 if (is_kdump_kernel()) {
3418 set
->nr_hw_queues
= 1;
3420 set
->queue_depth
= min(64U, set
->queue_depth
);
3423 * There is no use for more h/w queues than cpus if we just have
3426 if (set
->nr_maps
== 1 && set
->nr_hw_queues
> nr_cpu_ids
)
3427 set
->nr_hw_queues
= nr_cpu_ids
;
3429 if (blk_mq_realloc_tag_set_tags(set
, 0, set
->nr_hw_queues
) < 0)
3433 for (i
= 0; i
< set
->nr_maps
; i
++) {
3434 set
->map
[i
].mq_map
= kcalloc_node(nr_cpu_ids
,
3435 sizeof(set
->map
[i
].mq_map
[0]),
3436 GFP_KERNEL
, set
->numa_node
);
3437 if (!set
->map
[i
].mq_map
)
3438 goto out_free_mq_map
;
3439 set
->map
[i
].nr_queues
= is_kdump_kernel() ? 1 : set
->nr_hw_queues
;
3442 ret
= blk_mq_update_queue_map(set
);
3444 goto out_free_mq_map
;
3446 ret
= blk_mq_alloc_map_and_requests(set
);
3448 goto out_free_mq_map
;
3450 if (blk_mq_is_sbitmap_shared(set
->flags
)) {
3451 atomic_set(&set
->active_queues_shared_sbitmap
, 0);
3453 if (blk_mq_init_shared_sbitmap(set
, set
->flags
)) {
3455 goto out_free_mq_rq_maps
;
3459 mutex_init(&set
->tag_list_lock
);
3460 INIT_LIST_HEAD(&set
->tag_list
);
3464 out_free_mq_rq_maps
:
3465 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
3466 blk_mq_free_map_and_requests(set
, i
);
3468 for (i
= 0; i
< set
->nr_maps
; i
++) {
3469 kfree(set
->map
[i
].mq_map
);
3470 set
->map
[i
].mq_map
= NULL
;
3476 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
3478 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
3482 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
3483 blk_mq_free_map_and_requests(set
, i
);
3485 if (blk_mq_is_sbitmap_shared(set
->flags
))
3486 blk_mq_exit_shared_sbitmap(set
);
3488 for (j
= 0; j
< set
->nr_maps
; j
++) {
3489 kfree(set
->map
[j
].mq_map
);
3490 set
->map
[j
].mq_map
= NULL
;
3496 EXPORT_SYMBOL(blk_mq_free_tag_set
);
3498 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
3500 struct blk_mq_tag_set
*set
= q
->tag_set
;
3501 struct blk_mq_hw_ctx
*hctx
;
3507 if (q
->nr_requests
== nr
)
3510 blk_mq_freeze_queue(q
);
3511 blk_mq_quiesce_queue(q
);
3514 queue_for_each_hw_ctx(q
, hctx
, i
) {
3518 * If we're using an MQ scheduler, just update the scheduler
3519 * queue depth. This is similar to what the old code would do.
3521 if (!hctx
->sched_tags
) {
3522 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
, nr
,
3524 if (!ret
&& blk_mq_is_sbitmap_shared(set
->flags
))
3525 blk_mq_tag_resize_shared_sbitmap(set
, nr
);
3527 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
3532 if (q
->elevator
&& q
->elevator
->type
->ops
.depth_updated
)
3533 q
->elevator
->type
->ops
.depth_updated(hctx
);
3537 q
->nr_requests
= nr
;
3539 blk_mq_unquiesce_queue(q
);
3540 blk_mq_unfreeze_queue(q
);
3546 * request_queue and elevator_type pair.
3547 * It is just used by __blk_mq_update_nr_hw_queues to cache
3548 * the elevator_type associated with a request_queue.
3550 struct blk_mq_qe_pair
{
3551 struct list_head node
;
3552 struct request_queue
*q
;
3553 struct elevator_type
*type
;
3557 * Cache the elevator_type in qe pair list and switch the
3558 * io scheduler to 'none'
3560 static bool blk_mq_elv_switch_none(struct list_head
*head
,
3561 struct request_queue
*q
)
3563 struct blk_mq_qe_pair
*qe
;
3568 qe
= kmalloc(sizeof(*qe
), GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
);
3572 INIT_LIST_HEAD(&qe
->node
);
3574 qe
->type
= q
->elevator
->type
;
3575 list_add(&qe
->node
, head
);
3577 mutex_lock(&q
->sysfs_lock
);
3579 * After elevator_switch_mq, the previous elevator_queue will be
3580 * released by elevator_release. The reference of the io scheduler
3581 * module get by elevator_get will also be put. So we need to get
3582 * a reference of the io scheduler module here to prevent it to be
3585 __module_get(qe
->type
->elevator_owner
);
3586 elevator_switch_mq(q
, NULL
);
3587 mutex_unlock(&q
->sysfs_lock
);
3592 static void blk_mq_elv_switch_back(struct list_head
*head
,
3593 struct request_queue
*q
)
3595 struct blk_mq_qe_pair
*qe
;
3596 struct elevator_type
*t
= NULL
;
3598 list_for_each_entry(qe
, head
, node
)
3607 list_del(&qe
->node
);
3610 mutex_lock(&q
->sysfs_lock
);
3611 elevator_switch_mq(q
, t
);
3612 mutex_unlock(&q
->sysfs_lock
);
3615 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
,
3618 struct request_queue
*q
;
3620 int prev_nr_hw_queues
;
3622 lockdep_assert_held(&set
->tag_list_lock
);
3624 if (set
->nr_maps
== 1 && nr_hw_queues
> nr_cpu_ids
)
3625 nr_hw_queues
= nr_cpu_ids
;
3626 if (nr_hw_queues
< 1)
3628 if (set
->nr_maps
== 1 && nr_hw_queues
== set
->nr_hw_queues
)
3631 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3632 blk_mq_freeze_queue(q
);
3634 * Switch IO scheduler to 'none', cleaning up the data associated
3635 * with the previous scheduler. We will switch back once we are done
3636 * updating the new sw to hw queue mappings.
3638 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3639 if (!blk_mq_elv_switch_none(&head
, q
))
3642 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3643 blk_mq_debugfs_unregister_hctxs(q
);
3644 blk_mq_sysfs_unregister(q
);
3647 prev_nr_hw_queues
= set
->nr_hw_queues
;
3648 if (blk_mq_realloc_tag_set_tags(set
, set
->nr_hw_queues
, nr_hw_queues
) <
3652 set
->nr_hw_queues
= nr_hw_queues
;
3654 blk_mq_update_queue_map(set
);
3655 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3656 blk_mq_realloc_hw_ctxs(set
, q
);
3657 if (q
->nr_hw_queues
!= set
->nr_hw_queues
) {
3658 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3659 nr_hw_queues
, prev_nr_hw_queues
);
3660 set
->nr_hw_queues
= prev_nr_hw_queues
;
3661 blk_mq_map_queues(&set
->map
[HCTX_TYPE_DEFAULT
]);
3664 blk_mq_map_swqueue(q
);
3668 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3669 blk_mq_sysfs_register(q
);
3670 blk_mq_debugfs_register_hctxs(q
);
3674 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3675 blk_mq_elv_switch_back(&head
, q
);
3677 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3678 blk_mq_unfreeze_queue(q
);
3681 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
3683 mutex_lock(&set
->tag_list_lock
);
3684 __blk_mq_update_nr_hw_queues(set
, nr_hw_queues
);
3685 mutex_unlock(&set
->tag_list_lock
);
3687 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
3689 /* Enable polling stats and return whether they were already enabled. */
3690 static bool blk_poll_stats_enable(struct request_queue
*q
)
3692 if (test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
3693 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS
, q
))
3695 blk_stat_add_callback(q
, q
->poll_cb
);
3699 static void blk_mq_poll_stats_start(struct request_queue
*q
)
3702 * We don't arm the callback if polling stats are not enabled or the
3703 * callback is already active.
3705 if (!test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
3706 blk_stat_is_active(q
->poll_cb
))
3709 blk_stat_activate_msecs(q
->poll_cb
, 100);
3712 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
)
3714 struct request_queue
*q
= cb
->data
;
3717 for (bucket
= 0; bucket
< BLK_MQ_POLL_STATS_BKTS
; bucket
++) {
3718 if (cb
->stat
[bucket
].nr_samples
)
3719 q
->poll_stat
[bucket
] = cb
->stat
[bucket
];
3723 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
3726 unsigned long ret
= 0;
3730 * If stats collection isn't on, don't sleep but turn it on for
3733 if (!blk_poll_stats_enable(q
))
3737 * As an optimistic guess, use half of the mean service time
3738 * for this type of request. We can (and should) make this smarter.
3739 * For instance, if the completion latencies are tight, we can
3740 * get closer than just half the mean. This is especially
3741 * important on devices where the completion latencies are longer
3742 * than ~10 usec. We do use the stats for the relevant IO size
3743 * if available which does lead to better estimates.
3745 bucket
= blk_mq_poll_stats_bkt(rq
);
3749 if (q
->poll_stat
[bucket
].nr_samples
)
3750 ret
= (q
->poll_stat
[bucket
].mean
+ 1) / 2;
3755 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
3758 struct hrtimer_sleeper hs
;
3759 enum hrtimer_mode mode
;
3763 if (rq
->rq_flags
& RQF_MQ_POLL_SLEPT
)
3767 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3769 * 0: use half of prev avg
3770 * >0: use this specific value
3772 if (q
->poll_nsec
> 0)
3773 nsecs
= q
->poll_nsec
;
3775 nsecs
= blk_mq_poll_nsecs(q
, rq
);
3780 rq
->rq_flags
|= RQF_MQ_POLL_SLEPT
;
3783 * This will be replaced with the stats tracking code, using
3784 * 'avg_completion_time / 2' as the pre-sleep target.
3788 mode
= HRTIMER_MODE_REL
;
3789 hrtimer_init_sleeper_on_stack(&hs
, CLOCK_MONOTONIC
, mode
);
3790 hrtimer_set_expires(&hs
.timer
, kt
);
3793 if (blk_mq_rq_state(rq
) == MQ_RQ_COMPLETE
)
3795 set_current_state(TASK_UNINTERRUPTIBLE
);
3796 hrtimer_sleeper_start_expires(&hs
, mode
);
3799 hrtimer_cancel(&hs
.timer
);
3800 mode
= HRTIMER_MODE_ABS
;
3801 } while (hs
.task
&& !signal_pending(current
));
3803 __set_current_state(TASK_RUNNING
);
3804 destroy_hrtimer_on_stack(&hs
.timer
);
3808 static bool blk_mq_poll_hybrid(struct request_queue
*q
,
3809 struct blk_mq_hw_ctx
*hctx
, blk_qc_t cookie
)
3813 if (q
->poll_nsec
== BLK_MQ_POLL_CLASSIC
)
3816 if (!blk_qc_t_is_internal(cookie
))
3817 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
3819 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
3821 * With scheduling, if the request has completed, we'll
3822 * get a NULL return here, as we clear the sched tag when
3823 * that happens. The request still remains valid, like always,
3824 * so we should be safe with just the NULL check.
3830 return blk_mq_poll_hybrid_sleep(q
, rq
);
3834 * blk_poll - poll for IO completions
3836 * @cookie: cookie passed back at IO submission time
3837 * @spin: whether to spin for completions
3840 * Poll for completions on the passed in queue. Returns number of
3841 * completed entries found. If @spin is true, then blk_poll will continue
3842 * looping until at least one completion is found, unless the task is
3843 * otherwise marked running (or we need to reschedule).
3845 int blk_poll(struct request_queue
*q
, blk_qc_t cookie
, bool spin
)
3847 struct blk_mq_hw_ctx
*hctx
;
3850 if (!blk_qc_t_valid(cookie
) ||
3851 !test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
3855 blk_flush_plug_list(current
->plug
, false);
3857 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
3860 * If we sleep, have the caller restart the poll loop to reset
3861 * the state. Like for the other success return cases, the
3862 * caller is responsible for checking if the IO completed. If
3863 * the IO isn't complete, we'll get called again and will go
3864 * straight to the busy poll loop.
3866 if (blk_mq_poll_hybrid(q
, hctx
, cookie
))
3869 hctx
->poll_considered
++;
3871 state
= current
->state
;
3875 hctx
->poll_invoked
++;
3877 ret
= q
->mq_ops
->poll(hctx
);
3879 hctx
->poll_success
++;
3880 __set_current_state(TASK_RUNNING
);
3884 if (signal_pending_state(state
, current
))
3885 __set_current_state(TASK_RUNNING
);
3887 if (current
->state
== TASK_RUNNING
)
3889 if (ret
< 0 || !spin
)
3892 } while (!need_resched());
3894 __set_current_state(TASK_RUNNING
);
3897 EXPORT_SYMBOL_GPL(blk_poll
);
3899 unsigned int blk_mq_rq_cpu(struct request
*rq
)
3901 return rq
->mq_ctx
->cpu
;
3903 EXPORT_SYMBOL(blk_mq_rq_cpu
);
3905 static int __init
blk_mq_init(void)
3909 for_each_possible_cpu(i
)
3910 INIT_LIST_HEAD(&per_cpu(blk_cpu_done
, i
));
3911 open_softirq(BLOCK_SOFTIRQ
, blk_done_softirq
);
3913 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD
,
3914 "block/softirq:dead", NULL
,
3915 blk_softirq_cpu_dead
);
3916 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
3917 blk_mq_hctx_notify_dead
);
3918 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE
, "block/mq:online",
3919 blk_mq_hctx_notify_online
,
3920 blk_mq_hctx_notify_offline
);
3923 subsys_initcall(blk_mq_init
);