1 // SPDX-License-Identifier: GPL-2.0
3 * Block multiqueue core code
5 * Copyright (C) 2013-2014 Jens Axboe
6 * Copyright (C) 2013-2014 Christoph Hellwig
8 #include <linux/kernel.h>
9 #include <linux/module.h>
10 #include <linux/backing-dev.h>
11 #include <linux/bio.h>
12 #include <linux/blkdev.h>
13 #include <linux/kmemleak.h>
15 #include <linux/init.h>
16 #include <linux/slab.h>
17 #include <linux/workqueue.h>
18 #include <linux/smp.h>
19 #include <linux/llist.h>
20 #include <linux/list_sort.h>
21 #include <linux/cpu.h>
22 #include <linux/cache.h>
23 #include <linux/sched/sysctl.h>
24 #include <linux/sched/topology.h>
25 #include <linux/sched/signal.h>
26 #include <linux/delay.h>
27 #include <linux/crash_dump.h>
28 #include <linux/prefetch.h>
29 #include <linux/blk-crypto.h>
31 #include <trace/events/block.h>
33 #include <linux/blk-mq.h>
34 #include <linux/t10-pi.h>
37 #include "blk-mq-debugfs.h"
38 #include "blk-mq-tag.h"
41 #include "blk-mq-sched.h"
42 #include "blk-rq-qos.h"
44 static DEFINE_PER_CPU(struct llist_head
, blk_cpu_done
);
46 static void blk_mq_poll_stats_start(struct request_queue
*q
);
47 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
);
49 static int blk_mq_poll_stats_bkt(const struct request
*rq
)
51 int ddir
, sectors
, bucket
;
53 ddir
= rq_data_dir(rq
);
54 sectors
= blk_rq_stats_sectors(rq
);
56 bucket
= ddir
+ 2 * ilog2(sectors
);
60 else if (bucket
>= BLK_MQ_POLL_STATS_BKTS
)
61 return ddir
+ BLK_MQ_POLL_STATS_BKTS
- 2;
67 * Check if any of the ctx, dispatch list or elevator
68 * have pending work in this hardware queue.
70 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
72 return !list_empty_careful(&hctx
->dispatch
) ||
73 sbitmap_any_bit_set(&hctx
->ctx_map
) ||
74 blk_mq_sched_has_work(hctx
);
78 * Mark this ctx as having pending work in this hardware queue
80 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
81 struct blk_mq_ctx
*ctx
)
83 const int bit
= ctx
->index_hw
[hctx
->type
];
85 if (!sbitmap_test_bit(&hctx
->ctx_map
, bit
))
86 sbitmap_set_bit(&hctx
->ctx_map
, bit
);
89 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
90 struct blk_mq_ctx
*ctx
)
92 const int bit
= ctx
->index_hw
[hctx
->type
];
94 sbitmap_clear_bit(&hctx
->ctx_map
, bit
);
98 struct block_device
*part
;
99 unsigned int inflight
[2];
102 static bool blk_mq_check_inflight(struct blk_mq_hw_ctx
*hctx
,
103 struct request
*rq
, void *priv
,
106 struct mq_inflight
*mi
= priv
;
108 if ((!mi
->part
->bd_partno
|| rq
->part
== mi
->part
) &&
109 blk_mq_rq_state(rq
) == MQ_RQ_IN_FLIGHT
)
110 mi
->inflight
[rq_data_dir(rq
)]++;
115 unsigned int blk_mq_in_flight(struct request_queue
*q
,
116 struct block_device
*part
)
118 struct mq_inflight mi
= { .part
= part
};
120 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight
, &mi
);
122 return mi
.inflight
[0] + mi
.inflight
[1];
125 void blk_mq_in_flight_rw(struct request_queue
*q
, struct block_device
*part
,
126 unsigned int inflight
[2])
128 struct mq_inflight mi
= { .part
= part
};
130 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight
, &mi
);
131 inflight
[0] = mi
.inflight
[0];
132 inflight
[1] = mi
.inflight
[1];
135 void blk_freeze_queue_start(struct request_queue
*q
)
137 mutex_lock(&q
->mq_freeze_lock
);
138 if (++q
->mq_freeze_depth
== 1) {
139 percpu_ref_kill(&q
->q_usage_counter
);
140 mutex_unlock(&q
->mq_freeze_lock
);
142 blk_mq_run_hw_queues(q
, false);
144 mutex_unlock(&q
->mq_freeze_lock
);
147 EXPORT_SYMBOL_GPL(blk_freeze_queue_start
);
149 void blk_mq_freeze_queue_wait(struct request_queue
*q
)
151 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
153 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait
);
155 int blk_mq_freeze_queue_wait_timeout(struct request_queue
*q
,
156 unsigned long timeout
)
158 return wait_event_timeout(q
->mq_freeze_wq
,
159 percpu_ref_is_zero(&q
->q_usage_counter
),
162 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout
);
165 * Guarantee no request is in use, so we can change any data structure of
166 * the queue afterward.
168 void blk_freeze_queue(struct request_queue
*q
)
171 * In the !blk_mq case we are only calling this to kill the
172 * q_usage_counter, otherwise this increases the freeze depth
173 * and waits for it to return to zero. For this reason there is
174 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
175 * exported to drivers as the only user for unfreeze is blk_mq.
177 blk_freeze_queue_start(q
);
178 blk_mq_freeze_queue_wait(q
);
181 void blk_mq_freeze_queue(struct request_queue
*q
)
184 * ...just an alias to keep freeze and unfreeze actions balanced
185 * in the blk_mq_* namespace
189 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
191 void blk_mq_unfreeze_queue(struct request_queue
*q
)
193 mutex_lock(&q
->mq_freeze_lock
);
194 q
->mq_freeze_depth
--;
195 WARN_ON_ONCE(q
->mq_freeze_depth
< 0);
196 if (!q
->mq_freeze_depth
) {
197 percpu_ref_resurrect(&q
->q_usage_counter
);
198 wake_up_all(&q
->mq_freeze_wq
);
200 mutex_unlock(&q
->mq_freeze_lock
);
202 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
205 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
206 * mpt3sas driver such that this function can be removed.
208 void blk_mq_quiesce_queue_nowait(struct request_queue
*q
)
210 blk_queue_flag_set(QUEUE_FLAG_QUIESCED
, q
);
212 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait
);
215 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
218 * Note: this function does not prevent that the struct request end_io()
219 * callback function is invoked. Once this function is returned, we make
220 * sure no dispatch can happen until the queue is unquiesced via
221 * blk_mq_unquiesce_queue().
223 void blk_mq_quiesce_queue(struct request_queue
*q
)
225 struct blk_mq_hw_ctx
*hctx
;
229 blk_mq_quiesce_queue_nowait(q
);
231 queue_for_each_hw_ctx(q
, hctx
, i
) {
232 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
233 synchronize_srcu(hctx
->srcu
);
240 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
243 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
246 * This function recovers queue into the state before quiescing
247 * which is done by blk_mq_quiesce_queue.
249 void blk_mq_unquiesce_queue(struct request_queue
*q
)
251 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED
, q
);
253 /* dispatch requests which are inserted during quiescing */
254 blk_mq_run_hw_queues(q
, true);
256 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue
);
258 void blk_mq_wake_waiters(struct request_queue
*q
)
260 struct blk_mq_hw_ctx
*hctx
;
263 queue_for_each_hw_ctx(q
, hctx
, i
)
264 if (blk_mq_hw_queue_mapped(hctx
))
265 blk_mq_tag_wakeup_all(hctx
->tags
, true);
269 * Only need start/end time stamping if we have iostat or
270 * blk stats enabled, or using an IO scheduler.
272 static inline bool blk_mq_need_time_stamp(struct request
*rq
)
274 return (rq
->rq_flags
& (RQF_IO_STAT
| RQF_STATS
)) || rq
->q
->elevator
;
277 static struct request
*blk_mq_rq_ctx_init(struct blk_mq_alloc_data
*data
,
278 unsigned int tag
, u64 alloc_time_ns
)
280 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
281 struct request
*rq
= tags
->static_rqs
[tag
];
283 if (data
->q
->elevator
) {
284 rq
->tag
= BLK_MQ_NO_TAG
;
285 rq
->internal_tag
= tag
;
288 rq
->internal_tag
= BLK_MQ_NO_TAG
;
291 /* csd/requeue_work/fifo_time is initialized before use */
293 rq
->mq_ctx
= data
->ctx
;
294 rq
->mq_hctx
= data
->hctx
;
296 rq
->cmd_flags
= data
->cmd_flags
;
297 if (data
->flags
& BLK_MQ_REQ_PM
)
298 rq
->rq_flags
|= RQF_PM
;
299 if (blk_queue_io_stat(data
->q
))
300 rq
->rq_flags
|= RQF_IO_STAT
;
301 INIT_LIST_HEAD(&rq
->queuelist
);
302 INIT_HLIST_NODE(&rq
->hash
);
303 RB_CLEAR_NODE(&rq
->rb_node
);
306 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
307 rq
->alloc_time_ns
= alloc_time_ns
;
309 if (blk_mq_need_time_stamp(rq
))
310 rq
->start_time_ns
= ktime_get_ns();
312 rq
->start_time_ns
= 0;
313 rq
->io_start_time_ns
= 0;
314 rq
->stats_sectors
= 0;
315 rq
->nr_phys_segments
= 0;
316 #if defined(CONFIG_BLK_DEV_INTEGRITY)
317 rq
->nr_integrity_segments
= 0;
319 blk_crypto_rq_set_defaults(rq
);
320 /* tag was already set */
321 WRITE_ONCE(rq
->deadline
, 0);
326 rq
->end_io_data
= NULL
;
328 data
->ctx
->rq_dispatched
[op_is_sync(data
->cmd_flags
)]++;
329 refcount_set(&rq
->ref
, 1);
331 if (!op_is_flush(data
->cmd_flags
)) {
332 struct elevator_queue
*e
= data
->q
->elevator
;
335 if (e
&& e
->type
->ops
.prepare_request
) {
336 if (e
->type
->icq_cache
)
337 blk_mq_sched_assign_ioc(rq
);
339 e
->type
->ops
.prepare_request(rq
);
340 rq
->rq_flags
|= RQF_ELVPRIV
;
344 data
->hctx
->queued
++;
348 static struct request
*__blk_mq_alloc_request(struct blk_mq_alloc_data
*data
)
350 struct request_queue
*q
= data
->q
;
351 struct elevator_queue
*e
= q
->elevator
;
352 u64 alloc_time_ns
= 0;
355 /* alloc_time includes depth and tag waits */
356 if (blk_queue_rq_alloc_time(q
))
357 alloc_time_ns
= ktime_get_ns();
359 if (data
->cmd_flags
& REQ_NOWAIT
)
360 data
->flags
|= BLK_MQ_REQ_NOWAIT
;
364 * Flush/passthrough requests are special and go directly to the
365 * dispatch list. Don't include reserved tags in the
366 * limiting, as it isn't useful.
368 if (!op_is_flush(data
->cmd_flags
) &&
369 !blk_op_is_passthrough(data
->cmd_flags
) &&
370 e
->type
->ops
.limit_depth
&&
371 !(data
->flags
& BLK_MQ_REQ_RESERVED
))
372 e
->type
->ops
.limit_depth(data
->cmd_flags
, data
);
376 data
->ctx
= blk_mq_get_ctx(q
);
377 data
->hctx
= blk_mq_map_queue(q
, data
->cmd_flags
, data
->ctx
);
379 blk_mq_tag_busy(data
->hctx
);
382 * Waiting allocations only fail because of an inactive hctx. In that
383 * case just retry the hctx assignment and tag allocation as CPU hotplug
384 * should have migrated us to an online CPU by now.
386 tag
= blk_mq_get_tag(data
);
387 if (tag
== BLK_MQ_NO_TAG
) {
388 if (data
->flags
& BLK_MQ_REQ_NOWAIT
)
392 * Give up the CPU and sleep for a random short time to ensure
393 * that thread using a realtime scheduling class are migrated
394 * off the CPU, and thus off the hctx that is going away.
399 return blk_mq_rq_ctx_init(data
, tag
, alloc_time_ns
);
402 struct request
*blk_mq_alloc_request(struct request_queue
*q
, unsigned int op
,
403 blk_mq_req_flags_t flags
)
405 struct blk_mq_alloc_data data
= {
413 ret
= blk_queue_enter(q
, flags
);
417 rq
= __blk_mq_alloc_request(&data
);
421 rq
->__sector
= (sector_t
) -1;
422 rq
->bio
= rq
->biotail
= NULL
;
426 return ERR_PTR(-EWOULDBLOCK
);
428 EXPORT_SYMBOL(blk_mq_alloc_request
);
430 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
,
431 unsigned int op
, blk_mq_req_flags_t flags
, unsigned int hctx_idx
)
433 struct blk_mq_alloc_data data
= {
438 u64 alloc_time_ns
= 0;
443 /* alloc_time includes depth and tag waits */
444 if (blk_queue_rq_alloc_time(q
))
445 alloc_time_ns
= ktime_get_ns();
448 * If the tag allocator sleeps we could get an allocation for a
449 * different hardware context. No need to complicate the low level
450 * allocator for this for the rare use case of a command tied to
453 if (WARN_ON_ONCE(!(flags
& (BLK_MQ_REQ_NOWAIT
| BLK_MQ_REQ_RESERVED
))))
454 return ERR_PTR(-EINVAL
);
456 if (hctx_idx
>= q
->nr_hw_queues
)
457 return ERR_PTR(-EIO
);
459 ret
= blk_queue_enter(q
, flags
);
464 * Check if the hardware context is actually mapped to anything.
465 * If not tell the caller that it should skip this queue.
468 data
.hctx
= q
->queue_hw_ctx
[hctx_idx
];
469 if (!blk_mq_hw_queue_mapped(data
.hctx
))
471 cpu
= cpumask_first_and(data
.hctx
->cpumask
, cpu_online_mask
);
472 data
.ctx
= __blk_mq_get_ctx(q
, cpu
);
475 blk_mq_tag_busy(data
.hctx
);
478 tag
= blk_mq_get_tag(&data
);
479 if (tag
== BLK_MQ_NO_TAG
)
481 return blk_mq_rq_ctx_init(&data
, tag
, alloc_time_ns
);
487 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
489 static void __blk_mq_free_request(struct request
*rq
)
491 struct request_queue
*q
= rq
->q
;
492 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
493 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
494 const int sched_tag
= rq
->internal_tag
;
496 blk_crypto_free_request(rq
);
497 blk_pm_mark_last_busy(rq
);
499 if (rq
->tag
!= BLK_MQ_NO_TAG
)
500 blk_mq_put_tag(hctx
->tags
, ctx
, rq
->tag
);
501 if (sched_tag
!= BLK_MQ_NO_TAG
)
502 blk_mq_put_tag(hctx
->sched_tags
, ctx
, sched_tag
);
503 blk_mq_sched_restart(hctx
);
507 void blk_mq_free_request(struct request
*rq
)
509 struct request_queue
*q
= rq
->q
;
510 struct elevator_queue
*e
= q
->elevator
;
511 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
512 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
514 if (rq
->rq_flags
& RQF_ELVPRIV
) {
515 if (e
&& e
->type
->ops
.finish_request
)
516 e
->type
->ops
.finish_request(rq
);
518 put_io_context(rq
->elv
.icq
->ioc
);
523 ctx
->rq_completed
[rq_is_sync(rq
)]++;
524 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
525 __blk_mq_dec_active_requests(hctx
);
527 if (unlikely(laptop_mode
&& !blk_rq_is_passthrough(rq
)))
528 laptop_io_completion(q
->disk
->bdi
);
532 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
533 if (refcount_dec_and_test(&rq
->ref
))
534 __blk_mq_free_request(rq
);
536 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
538 inline void __blk_mq_end_request(struct request
*rq
, blk_status_t error
)
542 if (blk_mq_need_time_stamp(rq
))
543 now
= ktime_get_ns();
545 if (rq
->rq_flags
& RQF_STATS
) {
546 blk_mq_poll_stats_start(rq
->q
);
547 blk_stat_add(rq
, now
);
550 blk_mq_sched_completed_request(rq
, now
);
552 blk_account_io_done(rq
, now
);
555 rq_qos_done(rq
->q
, rq
);
556 rq
->end_io(rq
, error
);
558 blk_mq_free_request(rq
);
561 EXPORT_SYMBOL(__blk_mq_end_request
);
563 void blk_mq_end_request(struct request
*rq
, blk_status_t error
)
565 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
567 __blk_mq_end_request(rq
, error
);
569 EXPORT_SYMBOL(blk_mq_end_request
);
571 static void blk_complete_reqs(struct llist_head
*list
)
573 struct llist_node
*entry
= llist_reverse_order(llist_del_all(list
));
574 struct request
*rq
, *next
;
576 llist_for_each_entry_safe(rq
, next
, entry
, ipi_list
)
577 rq
->q
->mq_ops
->complete(rq
);
580 static __latent_entropy
void blk_done_softirq(struct softirq_action
*h
)
582 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done
));
585 static int blk_softirq_cpu_dead(unsigned int cpu
)
587 blk_complete_reqs(&per_cpu(blk_cpu_done
, cpu
));
591 static void __blk_mq_complete_request_remote(void *data
)
593 __raise_softirq_irqoff(BLOCK_SOFTIRQ
);
596 static inline bool blk_mq_complete_need_ipi(struct request
*rq
)
598 int cpu
= raw_smp_processor_id();
600 if (!IS_ENABLED(CONFIG_SMP
) ||
601 !test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
))
604 * With force threaded interrupts enabled, raising softirq from an SMP
605 * function call will always result in waking the ksoftirqd thread.
606 * This is probably worse than completing the request on a different
609 if (force_irqthreads())
612 /* same CPU or cache domain? Complete locally */
613 if (cpu
== rq
->mq_ctx
->cpu
||
614 (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
) &&
615 cpus_share_cache(cpu
, rq
->mq_ctx
->cpu
)))
618 /* don't try to IPI to an offline CPU */
619 return cpu_online(rq
->mq_ctx
->cpu
);
622 static void blk_mq_complete_send_ipi(struct request
*rq
)
624 struct llist_head
*list
;
627 cpu
= rq
->mq_ctx
->cpu
;
628 list
= &per_cpu(blk_cpu_done
, cpu
);
629 if (llist_add(&rq
->ipi_list
, list
)) {
630 INIT_CSD(&rq
->csd
, __blk_mq_complete_request_remote
, rq
);
631 smp_call_function_single_async(cpu
, &rq
->csd
);
635 static void blk_mq_raise_softirq(struct request
*rq
)
637 struct llist_head
*list
;
640 list
= this_cpu_ptr(&blk_cpu_done
);
641 if (llist_add(&rq
->ipi_list
, list
))
642 raise_softirq(BLOCK_SOFTIRQ
);
646 bool blk_mq_complete_request_remote(struct request
*rq
)
648 WRITE_ONCE(rq
->state
, MQ_RQ_COMPLETE
);
651 * For a polled request, always complete locallly, it's pointless
652 * to redirect the completion.
654 if (rq
->cmd_flags
& REQ_HIPRI
)
657 if (blk_mq_complete_need_ipi(rq
)) {
658 blk_mq_complete_send_ipi(rq
);
662 if (rq
->q
->nr_hw_queues
== 1) {
663 blk_mq_raise_softirq(rq
);
668 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote
);
671 * blk_mq_complete_request - end I/O on a request
672 * @rq: the request being processed
675 * Complete a request by scheduling the ->complete_rq operation.
677 void blk_mq_complete_request(struct request
*rq
)
679 if (!blk_mq_complete_request_remote(rq
))
680 rq
->q
->mq_ops
->complete(rq
);
682 EXPORT_SYMBOL(blk_mq_complete_request
);
684 static void hctx_unlock(struct blk_mq_hw_ctx
*hctx
, int srcu_idx
)
685 __releases(hctx
->srcu
)
687 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
))
690 srcu_read_unlock(hctx
->srcu
, srcu_idx
);
693 static void hctx_lock(struct blk_mq_hw_ctx
*hctx
, int *srcu_idx
)
694 __acquires(hctx
->srcu
)
696 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
697 /* shut up gcc false positive */
701 *srcu_idx
= srcu_read_lock(hctx
->srcu
);
705 * blk_mq_start_request - Start processing a request
706 * @rq: Pointer to request to be started
708 * Function used by device drivers to notify the block layer that a request
709 * is going to be processed now, so blk layer can do proper initializations
710 * such as starting the timeout timer.
712 void blk_mq_start_request(struct request
*rq
)
714 struct request_queue
*q
= rq
->q
;
716 trace_block_rq_issue(rq
);
718 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
719 rq
->io_start_time_ns
= ktime_get_ns();
720 rq
->stats_sectors
= blk_rq_sectors(rq
);
721 rq
->rq_flags
|= RQF_STATS
;
725 WARN_ON_ONCE(blk_mq_rq_state(rq
) != MQ_RQ_IDLE
);
728 WRITE_ONCE(rq
->state
, MQ_RQ_IN_FLIGHT
);
730 #ifdef CONFIG_BLK_DEV_INTEGRITY
731 if (blk_integrity_rq(rq
) && req_op(rq
) == REQ_OP_WRITE
)
732 q
->integrity
.profile
->prepare_fn(rq
);
735 EXPORT_SYMBOL(blk_mq_start_request
);
737 static void __blk_mq_requeue_request(struct request
*rq
)
739 struct request_queue
*q
= rq
->q
;
741 blk_mq_put_driver_tag(rq
);
743 trace_block_rq_requeue(rq
);
744 rq_qos_requeue(q
, rq
);
746 if (blk_mq_request_started(rq
)) {
747 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
748 rq
->rq_flags
&= ~RQF_TIMED_OUT
;
752 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
754 __blk_mq_requeue_request(rq
);
756 /* this request will be re-inserted to io scheduler queue */
757 blk_mq_sched_requeue_request(rq
);
759 BUG_ON(!list_empty(&rq
->queuelist
));
760 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
762 EXPORT_SYMBOL(blk_mq_requeue_request
);
764 static void blk_mq_requeue_work(struct work_struct
*work
)
766 struct request_queue
*q
=
767 container_of(work
, struct request_queue
, requeue_work
.work
);
769 struct request
*rq
, *next
;
771 spin_lock_irq(&q
->requeue_lock
);
772 list_splice_init(&q
->requeue_list
, &rq_list
);
773 spin_unlock_irq(&q
->requeue_lock
);
775 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
776 if (!(rq
->rq_flags
& (RQF_SOFTBARRIER
| RQF_DONTPREP
)))
779 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
780 list_del_init(&rq
->queuelist
);
782 * If RQF_DONTPREP, rq has contained some driver specific
783 * data, so insert it to hctx dispatch list to avoid any
786 if (rq
->rq_flags
& RQF_DONTPREP
)
787 blk_mq_request_bypass_insert(rq
, false, false);
789 blk_mq_sched_insert_request(rq
, true, false, false);
792 while (!list_empty(&rq_list
)) {
793 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
794 list_del_init(&rq
->queuelist
);
795 blk_mq_sched_insert_request(rq
, false, false, false);
798 blk_mq_run_hw_queues(q
, false);
801 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
802 bool kick_requeue_list
)
804 struct request_queue
*q
= rq
->q
;
808 * We abuse this flag that is otherwise used by the I/O scheduler to
809 * request head insertion from the workqueue.
811 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
813 spin_lock_irqsave(&q
->requeue_lock
, flags
);
815 rq
->rq_flags
|= RQF_SOFTBARRIER
;
816 list_add(&rq
->queuelist
, &q
->requeue_list
);
818 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
820 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
822 if (kick_requeue_list
)
823 blk_mq_kick_requeue_list(q
);
826 void blk_mq_kick_requeue_list(struct request_queue
*q
)
828 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
, 0);
830 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
832 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
835 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
,
836 msecs_to_jiffies(msecs
));
838 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
840 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
842 if (tag
< tags
->nr_tags
) {
843 prefetch(tags
->rqs
[tag
]);
844 return tags
->rqs
[tag
];
849 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
851 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
852 void *priv
, bool reserved
)
855 * If we find a request that isn't idle and the queue matches,
856 * we know the queue is busy. Return false to stop the iteration.
858 if (blk_mq_request_started(rq
) && rq
->q
== hctx
->queue
) {
868 bool blk_mq_queue_inflight(struct request_queue
*q
)
872 blk_mq_queue_tag_busy_iter(q
, blk_mq_rq_inflight
, &busy
);
875 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight
);
877 static void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
879 req
->rq_flags
|= RQF_TIMED_OUT
;
880 if (req
->q
->mq_ops
->timeout
) {
881 enum blk_eh_timer_return ret
;
883 ret
= req
->q
->mq_ops
->timeout(req
, reserved
);
884 if (ret
== BLK_EH_DONE
)
886 WARN_ON_ONCE(ret
!= BLK_EH_RESET_TIMER
);
892 static bool blk_mq_req_expired(struct request
*rq
, unsigned long *next
)
894 unsigned long deadline
;
896 if (blk_mq_rq_state(rq
) != MQ_RQ_IN_FLIGHT
)
898 if (rq
->rq_flags
& RQF_TIMED_OUT
)
901 deadline
= READ_ONCE(rq
->deadline
);
902 if (time_after_eq(jiffies
, deadline
))
907 else if (time_after(*next
, deadline
))
912 void blk_mq_put_rq_ref(struct request
*rq
)
916 else if (refcount_dec_and_test(&rq
->ref
))
917 __blk_mq_free_request(rq
);
920 static bool blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
921 struct request
*rq
, void *priv
, bool reserved
)
923 unsigned long *next
= priv
;
926 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
927 * be reallocated underneath the timeout handler's processing, then
928 * the expire check is reliable. If the request is not expired, then
929 * it was completed and reallocated as a new request after returning
930 * from blk_mq_check_expired().
932 if (blk_mq_req_expired(rq
, next
))
933 blk_mq_rq_timed_out(rq
, reserved
);
937 static void blk_mq_timeout_work(struct work_struct
*work
)
939 struct request_queue
*q
=
940 container_of(work
, struct request_queue
, timeout_work
);
941 unsigned long next
= 0;
942 struct blk_mq_hw_ctx
*hctx
;
945 /* A deadlock might occur if a request is stuck requiring a
946 * timeout at the same time a queue freeze is waiting
947 * completion, since the timeout code would not be able to
948 * acquire the queue reference here.
950 * That's why we don't use blk_queue_enter here; instead, we use
951 * percpu_ref_tryget directly, because we need to be able to
952 * obtain a reference even in the short window between the queue
953 * starting to freeze, by dropping the first reference in
954 * blk_freeze_queue_start, and the moment the last request is
955 * consumed, marked by the instant q_usage_counter reaches
958 if (!percpu_ref_tryget(&q
->q_usage_counter
))
961 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &next
);
964 mod_timer(&q
->timeout
, next
);
967 * Request timeouts are handled as a forward rolling timer. If
968 * we end up here it means that no requests are pending and
969 * also that no request has been pending for a while. Mark
972 queue_for_each_hw_ctx(q
, hctx
, i
) {
973 /* the hctx may be unmapped, so check it here */
974 if (blk_mq_hw_queue_mapped(hctx
))
975 blk_mq_tag_idle(hctx
);
981 struct flush_busy_ctx_data
{
982 struct blk_mq_hw_ctx
*hctx
;
983 struct list_head
*list
;
986 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
988 struct flush_busy_ctx_data
*flush_data
= data
;
989 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
990 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
991 enum hctx_type type
= hctx
->type
;
993 spin_lock(&ctx
->lock
);
994 list_splice_tail_init(&ctx
->rq_lists
[type
], flush_data
->list
);
995 sbitmap_clear_bit(sb
, bitnr
);
996 spin_unlock(&ctx
->lock
);
1001 * Process software queues that have been marked busy, splicing them
1002 * to the for-dispatch
1004 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
1006 struct flush_busy_ctx_data data
= {
1011 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
1013 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
1015 struct dispatch_rq_data
{
1016 struct blk_mq_hw_ctx
*hctx
;
1020 static bool dispatch_rq_from_ctx(struct sbitmap
*sb
, unsigned int bitnr
,
1023 struct dispatch_rq_data
*dispatch_data
= data
;
1024 struct blk_mq_hw_ctx
*hctx
= dispatch_data
->hctx
;
1025 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
1026 enum hctx_type type
= hctx
->type
;
1028 spin_lock(&ctx
->lock
);
1029 if (!list_empty(&ctx
->rq_lists
[type
])) {
1030 dispatch_data
->rq
= list_entry_rq(ctx
->rq_lists
[type
].next
);
1031 list_del_init(&dispatch_data
->rq
->queuelist
);
1032 if (list_empty(&ctx
->rq_lists
[type
]))
1033 sbitmap_clear_bit(sb
, bitnr
);
1035 spin_unlock(&ctx
->lock
);
1037 return !dispatch_data
->rq
;
1040 struct request
*blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx
*hctx
,
1041 struct blk_mq_ctx
*start
)
1043 unsigned off
= start
? start
->index_hw
[hctx
->type
] : 0;
1044 struct dispatch_rq_data data
= {
1049 __sbitmap_for_each_set(&hctx
->ctx_map
, off
,
1050 dispatch_rq_from_ctx
, &data
);
1055 static inline unsigned int queued_to_index(unsigned int queued
)
1060 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
1063 static bool __blk_mq_get_driver_tag(struct request
*rq
)
1065 struct sbitmap_queue
*bt
= rq
->mq_hctx
->tags
->bitmap_tags
;
1066 unsigned int tag_offset
= rq
->mq_hctx
->tags
->nr_reserved_tags
;
1069 blk_mq_tag_busy(rq
->mq_hctx
);
1071 if (blk_mq_tag_is_reserved(rq
->mq_hctx
->sched_tags
, rq
->internal_tag
)) {
1072 bt
= rq
->mq_hctx
->tags
->breserved_tags
;
1075 if (!hctx_may_queue(rq
->mq_hctx
, bt
))
1079 tag
= __sbitmap_queue_get(bt
);
1080 if (tag
== BLK_MQ_NO_TAG
)
1083 rq
->tag
= tag
+ tag_offset
;
1087 bool blk_mq_get_driver_tag(struct request
*rq
)
1089 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1091 if (rq
->tag
== BLK_MQ_NO_TAG
&& !__blk_mq_get_driver_tag(rq
))
1094 if ((hctx
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
) &&
1095 !(rq
->rq_flags
& RQF_MQ_INFLIGHT
)) {
1096 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
1097 __blk_mq_inc_active_requests(hctx
);
1099 hctx
->tags
->rqs
[rq
->tag
] = rq
;
1103 static int blk_mq_dispatch_wake(wait_queue_entry_t
*wait
, unsigned mode
,
1104 int flags
, void *key
)
1106 struct blk_mq_hw_ctx
*hctx
;
1108 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
1110 spin_lock(&hctx
->dispatch_wait_lock
);
1111 if (!list_empty(&wait
->entry
)) {
1112 struct sbitmap_queue
*sbq
;
1114 list_del_init(&wait
->entry
);
1115 sbq
= hctx
->tags
->bitmap_tags
;
1116 atomic_dec(&sbq
->ws_active
);
1118 spin_unlock(&hctx
->dispatch_wait_lock
);
1120 blk_mq_run_hw_queue(hctx
, true);
1125 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1126 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1127 * restart. For both cases, take care to check the condition again after
1128 * marking us as waiting.
1130 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx
*hctx
,
1133 struct sbitmap_queue
*sbq
= hctx
->tags
->bitmap_tags
;
1134 struct wait_queue_head
*wq
;
1135 wait_queue_entry_t
*wait
;
1138 if (!(hctx
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
)) {
1139 blk_mq_sched_mark_restart_hctx(hctx
);
1142 * It's possible that a tag was freed in the window between the
1143 * allocation failure and adding the hardware queue to the wait
1146 * Don't clear RESTART here, someone else could have set it.
1147 * At most this will cost an extra queue run.
1149 return blk_mq_get_driver_tag(rq
);
1152 wait
= &hctx
->dispatch_wait
;
1153 if (!list_empty_careful(&wait
->entry
))
1156 wq
= &bt_wait_ptr(sbq
, hctx
)->wait
;
1158 spin_lock_irq(&wq
->lock
);
1159 spin_lock(&hctx
->dispatch_wait_lock
);
1160 if (!list_empty(&wait
->entry
)) {
1161 spin_unlock(&hctx
->dispatch_wait_lock
);
1162 spin_unlock_irq(&wq
->lock
);
1166 atomic_inc(&sbq
->ws_active
);
1167 wait
->flags
&= ~WQ_FLAG_EXCLUSIVE
;
1168 __add_wait_queue(wq
, wait
);
1171 * It's possible that a tag was freed in the window between the
1172 * allocation failure and adding the hardware queue to the wait
1175 ret
= blk_mq_get_driver_tag(rq
);
1177 spin_unlock(&hctx
->dispatch_wait_lock
);
1178 spin_unlock_irq(&wq
->lock
);
1183 * We got a tag, remove ourselves from the wait queue to ensure
1184 * someone else gets the wakeup.
1186 list_del_init(&wait
->entry
);
1187 atomic_dec(&sbq
->ws_active
);
1188 spin_unlock(&hctx
->dispatch_wait_lock
);
1189 spin_unlock_irq(&wq
->lock
);
1194 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1195 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1197 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1198 * - EWMA is one simple way to compute running average value
1199 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1200 * - take 4 as factor for avoiding to get too small(0) result, and this
1201 * factor doesn't matter because EWMA decreases exponentially
1203 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx
*hctx
, bool busy
)
1207 ewma
= hctx
->dispatch_busy
;
1212 ewma
*= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
- 1;
1214 ewma
+= 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR
;
1215 ewma
/= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
;
1217 hctx
->dispatch_busy
= ewma
;
1220 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1222 static void blk_mq_handle_dev_resource(struct request
*rq
,
1223 struct list_head
*list
)
1225 struct request
*next
=
1226 list_first_entry_or_null(list
, struct request
, queuelist
);
1229 * If an I/O scheduler has been configured and we got a driver tag for
1230 * the next request already, free it.
1233 blk_mq_put_driver_tag(next
);
1235 list_add(&rq
->queuelist
, list
);
1236 __blk_mq_requeue_request(rq
);
1239 static void blk_mq_handle_zone_resource(struct request
*rq
,
1240 struct list_head
*zone_list
)
1243 * If we end up here it is because we cannot dispatch a request to a
1244 * specific zone due to LLD level zone-write locking or other zone
1245 * related resource not being available. In this case, set the request
1246 * aside in zone_list for retrying it later.
1248 list_add(&rq
->queuelist
, zone_list
);
1249 __blk_mq_requeue_request(rq
);
1252 enum prep_dispatch
{
1254 PREP_DISPATCH_NO_TAG
,
1255 PREP_DISPATCH_NO_BUDGET
,
1258 static enum prep_dispatch
blk_mq_prep_dispatch_rq(struct request
*rq
,
1261 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1262 int budget_token
= -1;
1265 budget_token
= blk_mq_get_dispatch_budget(rq
->q
);
1266 if (budget_token
< 0) {
1267 blk_mq_put_driver_tag(rq
);
1268 return PREP_DISPATCH_NO_BUDGET
;
1270 blk_mq_set_rq_budget_token(rq
, budget_token
);
1273 if (!blk_mq_get_driver_tag(rq
)) {
1275 * The initial allocation attempt failed, so we need to
1276 * rerun the hardware queue when a tag is freed. The
1277 * waitqueue takes care of that. If the queue is run
1278 * before we add this entry back on the dispatch list,
1279 * we'll re-run it below.
1281 if (!blk_mq_mark_tag_wait(hctx
, rq
)) {
1283 * All budgets not got from this function will be put
1284 * together during handling partial dispatch
1287 blk_mq_put_dispatch_budget(rq
->q
, budget_token
);
1288 return PREP_DISPATCH_NO_TAG
;
1292 return PREP_DISPATCH_OK
;
1295 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1296 static void blk_mq_release_budgets(struct request_queue
*q
,
1297 struct list_head
*list
)
1301 list_for_each_entry(rq
, list
, queuelist
) {
1302 int budget_token
= blk_mq_get_rq_budget_token(rq
);
1304 if (budget_token
>= 0)
1305 blk_mq_put_dispatch_budget(q
, budget_token
);
1310 * Returns true if we did some work AND can potentially do more.
1312 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
,
1313 unsigned int nr_budgets
)
1315 enum prep_dispatch prep
;
1316 struct request_queue
*q
= hctx
->queue
;
1317 struct request
*rq
, *nxt
;
1319 blk_status_t ret
= BLK_STS_OK
;
1320 LIST_HEAD(zone_list
);
1322 if (list_empty(list
))
1326 * Now process all the entries, sending them to the driver.
1328 errors
= queued
= 0;
1330 struct blk_mq_queue_data bd
;
1332 rq
= list_first_entry(list
, struct request
, queuelist
);
1334 WARN_ON_ONCE(hctx
!= rq
->mq_hctx
);
1335 prep
= blk_mq_prep_dispatch_rq(rq
, !nr_budgets
);
1336 if (prep
!= PREP_DISPATCH_OK
)
1339 list_del_init(&rq
->queuelist
);
1344 * Flag last if we have no more requests, or if we have more
1345 * but can't assign a driver tag to it.
1347 if (list_empty(list
))
1350 nxt
= list_first_entry(list
, struct request
, queuelist
);
1351 bd
.last
= !blk_mq_get_driver_tag(nxt
);
1355 * once the request is queued to lld, no need to cover the
1360 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1365 case BLK_STS_RESOURCE
:
1366 case BLK_STS_DEV_RESOURCE
:
1367 blk_mq_handle_dev_resource(rq
, list
);
1369 case BLK_STS_ZONE_RESOURCE
:
1371 * Move the request to zone_list and keep going through
1372 * the dispatch list to find more requests the drive can
1375 blk_mq_handle_zone_resource(rq
, &zone_list
);
1379 blk_mq_end_request(rq
, ret
);
1381 } while (!list_empty(list
));
1383 if (!list_empty(&zone_list
))
1384 list_splice_tail_init(&zone_list
, list
);
1386 hctx
->dispatched
[queued_to_index(queued
)]++;
1388 /* If we didn't flush the entire list, we could have told the driver
1389 * there was more coming, but that turned out to be a lie.
1391 if ((!list_empty(list
) || errors
) && q
->mq_ops
->commit_rqs
&& queued
)
1392 q
->mq_ops
->commit_rqs(hctx
);
1394 * Any items that need requeuing? Stuff them into hctx->dispatch,
1395 * that is where we will continue on next queue run.
1397 if (!list_empty(list
)) {
1399 /* For non-shared tags, the RESTART check will suffice */
1400 bool no_tag
= prep
== PREP_DISPATCH_NO_TAG
&&
1401 (hctx
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
);
1402 bool no_budget_avail
= prep
== PREP_DISPATCH_NO_BUDGET
;
1405 blk_mq_release_budgets(q
, list
);
1407 spin_lock(&hctx
->lock
);
1408 list_splice_tail_init(list
, &hctx
->dispatch
);
1409 spin_unlock(&hctx
->lock
);
1412 * Order adding requests to hctx->dispatch and checking
1413 * SCHED_RESTART flag. The pair of this smp_mb() is the one
1414 * in blk_mq_sched_restart(). Avoid restart code path to
1415 * miss the new added requests to hctx->dispatch, meantime
1416 * SCHED_RESTART is observed here.
1421 * If SCHED_RESTART was set by the caller of this function and
1422 * it is no longer set that means that it was cleared by another
1423 * thread and hence that a queue rerun is needed.
1425 * If 'no_tag' is set, that means that we failed getting
1426 * a driver tag with an I/O scheduler attached. If our dispatch
1427 * waitqueue is no longer active, ensure that we run the queue
1428 * AFTER adding our entries back to the list.
1430 * If no I/O scheduler has been configured it is possible that
1431 * the hardware queue got stopped and restarted before requests
1432 * were pushed back onto the dispatch list. Rerun the queue to
1433 * avoid starvation. Notes:
1434 * - blk_mq_run_hw_queue() checks whether or not a queue has
1435 * been stopped before rerunning a queue.
1436 * - Some but not all block drivers stop a queue before
1437 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1440 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1441 * bit is set, run queue after a delay to avoid IO stalls
1442 * that could otherwise occur if the queue is idle. We'll do
1443 * similar if we couldn't get budget and SCHED_RESTART is set.
1445 needs_restart
= blk_mq_sched_needs_restart(hctx
);
1446 if (!needs_restart
||
1447 (no_tag
&& list_empty_careful(&hctx
->dispatch_wait
.entry
)))
1448 blk_mq_run_hw_queue(hctx
, true);
1449 else if (needs_restart
&& (ret
== BLK_STS_RESOURCE
||
1451 blk_mq_delay_run_hw_queue(hctx
, BLK_MQ_RESOURCE_DELAY
);
1453 blk_mq_update_dispatch_busy(hctx
, true);
1456 blk_mq_update_dispatch_busy(hctx
, false);
1458 return (queued
+ errors
) != 0;
1462 * __blk_mq_run_hw_queue - Run a hardware queue.
1463 * @hctx: Pointer to the hardware queue to run.
1465 * Send pending requests to the hardware.
1467 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1472 * We can't run the queue inline with ints disabled. Ensure that
1473 * we catch bad users of this early.
1475 WARN_ON_ONCE(in_interrupt());
1477 might_sleep_if(hctx
->flags
& BLK_MQ_F_BLOCKING
);
1479 hctx_lock(hctx
, &srcu_idx
);
1480 blk_mq_sched_dispatch_requests(hctx
);
1481 hctx_unlock(hctx
, srcu_idx
);
1484 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx
*hctx
)
1486 int cpu
= cpumask_first_and(hctx
->cpumask
, cpu_online_mask
);
1488 if (cpu
>= nr_cpu_ids
)
1489 cpu
= cpumask_first(hctx
->cpumask
);
1494 * It'd be great if the workqueue API had a way to pass
1495 * in a mask and had some smarts for more clever placement.
1496 * For now we just round-robin here, switching for every
1497 * BLK_MQ_CPU_WORK_BATCH queued items.
1499 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1502 int next_cpu
= hctx
->next_cpu
;
1504 if (hctx
->queue
->nr_hw_queues
== 1)
1505 return WORK_CPU_UNBOUND
;
1507 if (--hctx
->next_cpu_batch
<= 0) {
1509 next_cpu
= cpumask_next_and(next_cpu
, hctx
->cpumask
,
1511 if (next_cpu
>= nr_cpu_ids
)
1512 next_cpu
= blk_mq_first_mapped_cpu(hctx
);
1513 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1517 * Do unbound schedule if we can't find a online CPU for this hctx,
1518 * and it should only happen in the path of handling CPU DEAD.
1520 if (!cpu_online(next_cpu
)) {
1527 * Make sure to re-select CPU next time once after CPUs
1528 * in hctx->cpumask become online again.
1530 hctx
->next_cpu
= next_cpu
;
1531 hctx
->next_cpu_batch
= 1;
1532 return WORK_CPU_UNBOUND
;
1535 hctx
->next_cpu
= next_cpu
;
1540 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
1541 * @hctx: Pointer to the hardware queue to run.
1542 * @async: If we want to run the queue asynchronously.
1543 * @msecs: Milliseconds of delay to wait before running the queue.
1545 * If !@async, try to run the queue now. Else, run the queue asynchronously and
1546 * with a delay of @msecs.
1548 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
,
1549 unsigned long msecs
)
1551 if (unlikely(blk_mq_hctx_stopped(hctx
)))
1554 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1555 int cpu
= get_cpu();
1556 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1557 __blk_mq_run_hw_queue(hctx
);
1565 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
,
1566 msecs_to_jiffies(msecs
));
1570 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
1571 * @hctx: Pointer to the hardware queue to run.
1572 * @msecs: Milliseconds of delay to wait before running the queue.
1574 * Run a hardware queue asynchronously with a delay of @msecs.
1576 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1578 __blk_mq_delay_run_hw_queue(hctx
, true, msecs
);
1580 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue
);
1583 * blk_mq_run_hw_queue - Start to run a hardware queue.
1584 * @hctx: Pointer to the hardware queue to run.
1585 * @async: If we want to run the queue asynchronously.
1587 * Check if the request queue is not in a quiesced state and if there are
1588 * pending requests to be sent. If this is true, run the queue to send requests
1591 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1597 * When queue is quiesced, we may be switching io scheduler, or
1598 * updating nr_hw_queues, or other things, and we can't run queue
1599 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1601 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1604 hctx_lock(hctx
, &srcu_idx
);
1605 need_run
= !blk_queue_quiesced(hctx
->queue
) &&
1606 blk_mq_hctx_has_pending(hctx
);
1607 hctx_unlock(hctx
, srcu_idx
);
1610 __blk_mq_delay_run_hw_queue(hctx
, async
, 0);
1612 EXPORT_SYMBOL(blk_mq_run_hw_queue
);
1615 * Is the request queue handled by an IO scheduler that does not respect
1616 * hardware queues when dispatching?
1618 static bool blk_mq_has_sqsched(struct request_queue
*q
)
1620 struct elevator_queue
*e
= q
->elevator
;
1622 if (e
&& e
->type
->ops
.dispatch_request
&&
1623 !(e
->type
->elevator_features
& ELEVATOR_F_MQ_AWARE
))
1629 * Return prefered queue to dispatch from (if any) for non-mq aware IO
1632 static struct blk_mq_hw_ctx
*blk_mq_get_sq_hctx(struct request_queue
*q
)
1634 struct blk_mq_hw_ctx
*hctx
;
1637 * If the IO scheduler does not respect hardware queues when
1638 * dispatching, we just don't bother with multiple HW queues and
1639 * dispatch from hctx for the current CPU since running multiple queues
1640 * just causes lock contention inside the scheduler and pointless cache
1643 hctx
= blk_mq_map_queue_type(q
, HCTX_TYPE_DEFAULT
,
1644 raw_smp_processor_id());
1645 if (!blk_mq_hctx_stopped(hctx
))
1651 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
1652 * @q: Pointer to the request queue to run.
1653 * @async: If we want to run the queue asynchronously.
1655 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1657 struct blk_mq_hw_ctx
*hctx
, *sq_hctx
;
1661 if (blk_mq_has_sqsched(q
))
1662 sq_hctx
= blk_mq_get_sq_hctx(q
);
1663 queue_for_each_hw_ctx(q
, hctx
, i
) {
1664 if (blk_mq_hctx_stopped(hctx
))
1667 * Dispatch from this hctx either if there's no hctx preferred
1668 * by IO scheduler or if it has requests that bypass the
1671 if (!sq_hctx
|| sq_hctx
== hctx
||
1672 !list_empty_careful(&hctx
->dispatch
))
1673 blk_mq_run_hw_queue(hctx
, async
);
1676 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1679 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
1680 * @q: Pointer to the request queue to run.
1681 * @msecs: Milliseconds of delay to wait before running the queues.
1683 void blk_mq_delay_run_hw_queues(struct request_queue
*q
, unsigned long msecs
)
1685 struct blk_mq_hw_ctx
*hctx
, *sq_hctx
;
1689 if (blk_mq_has_sqsched(q
))
1690 sq_hctx
= blk_mq_get_sq_hctx(q
);
1691 queue_for_each_hw_ctx(q
, hctx
, i
) {
1692 if (blk_mq_hctx_stopped(hctx
))
1695 * Dispatch from this hctx either if there's no hctx preferred
1696 * by IO scheduler or if it has requests that bypass the
1699 if (!sq_hctx
|| sq_hctx
== hctx
||
1700 !list_empty_careful(&hctx
->dispatch
))
1701 blk_mq_delay_run_hw_queue(hctx
, msecs
);
1704 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues
);
1707 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1708 * @q: request queue.
1710 * The caller is responsible for serializing this function against
1711 * blk_mq_{start,stop}_hw_queue().
1713 bool blk_mq_queue_stopped(struct request_queue
*q
)
1715 struct blk_mq_hw_ctx
*hctx
;
1718 queue_for_each_hw_ctx(q
, hctx
, i
)
1719 if (blk_mq_hctx_stopped(hctx
))
1724 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1727 * This function is often used for pausing .queue_rq() by driver when
1728 * there isn't enough resource or some conditions aren't satisfied, and
1729 * BLK_STS_RESOURCE is usually returned.
1731 * We do not guarantee that dispatch can be drained or blocked
1732 * after blk_mq_stop_hw_queue() returns. Please use
1733 * blk_mq_quiesce_queue() for that requirement.
1735 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1737 cancel_delayed_work(&hctx
->run_work
);
1739 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1741 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1744 * This function is often used for pausing .queue_rq() by driver when
1745 * there isn't enough resource or some conditions aren't satisfied, and
1746 * BLK_STS_RESOURCE is usually returned.
1748 * We do not guarantee that dispatch can be drained or blocked
1749 * after blk_mq_stop_hw_queues() returns. Please use
1750 * blk_mq_quiesce_queue() for that requirement.
1752 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1754 struct blk_mq_hw_ctx
*hctx
;
1757 queue_for_each_hw_ctx(q
, hctx
, i
)
1758 blk_mq_stop_hw_queue(hctx
);
1760 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1762 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1764 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1766 blk_mq_run_hw_queue(hctx
, false);
1768 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1770 void blk_mq_start_hw_queues(struct request_queue
*q
)
1772 struct blk_mq_hw_ctx
*hctx
;
1775 queue_for_each_hw_ctx(q
, hctx
, i
)
1776 blk_mq_start_hw_queue(hctx
);
1778 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1780 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1782 if (!blk_mq_hctx_stopped(hctx
))
1785 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1786 blk_mq_run_hw_queue(hctx
, async
);
1788 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1790 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1792 struct blk_mq_hw_ctx
*hctx
;
1795 queue_for_each_hw_ctx(q
, hctx
, i
)
1796 blk_mq_start_stopped_hw_queue(hctx
, async
);
1798 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1800 static void blk_mq_run_work_fn(struct work_struct
*work
)
1802 struct blk_mq_hw_ctx
*hctx
;
1804 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
1807 * If we are stopped, don't run the queue.
1809 if (blk_mq_hctx_stopped(hctx
))
1812 __blk_mq_run_hw_queue(hctx
);
1815 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1819 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1820 enum hctx_type type
= hctx
->type
;
1822 lockdep_assert_held(&ctx
->lock
);
1824 trace_block_rq_insert(rq
);
1827 list_add(&rq
->queuelist
, &ctx
->rq_lists
[type
]);
1829 list_add_tail(&rq
->queuelist
, &ctx
->rq_lists
[type
]);
1832 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1835 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1837 lockdep_assert_held(&ctx
->lock
);
1839 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1840 blk_mq_hctx_mark_pending(hctx
, ctx
);
1844 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
1845 * @rq: Pointer to request to be inserted.
1846 * @at_head: true if the request should be inserted at the head of the list.
1847 * @run_queue: If we should run the hardware queue after inserting the request.
1849 * Should only be used carefully, when the caller knows we want to
1850 * bypass a potential IO scheduler on the target device.
1852 void blk_mq_request_bypass_insert(struct request
*rq
, bool at_head
,
1855 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1857 spin_lock(&hctx
->lock
);
1859 list_add(&rq
->queuelist
, &hctx
->dispatch
);
1861 list_add_tail(&rq
->queuelist
, &hctx
->dispatch
);
1862 spin_unlock(&hctx
->lock
);
1865 blk_mq_run_hw_queue(hctx
, false);
1868 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1869 struct list_head
*list
)
1873 enum hctx_type type
= hctx
->type
;
1876 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1879 list_for_each_entry(rq
, list
, queuelist
) {
1880 BUG_ON(rq
->mq_ctx
!= ctx
);
1881 trace_block_rq_insert(rq
);
1884 spin_lock(&ctx
->lock
);
1885 list_splice_tail_init(list
, &ctx
->rq_lists
[type
]);
1886 blk_mq_hctx_mark_pending(hctx
, ctx
);
1887 spin_unlock(&ctx
->lock
);
1890 static int plug_rq_cmp(void *priv
, const struct list_head
*a
,
1891 const struct list_head
*b
)
1893 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1894 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1896 if (rqa
->mq_ctx
!= rqb
->mq_ctx
)
1897 return rqa
->mq_ctx
> rqb
->mq_ctx
;
1898 if (rqa
->mq_hctx
!= rqb
->mq_hctx
)
1899 return rqa
->mq_hctx
> rqb
->mq_hctx
;
1901 return blk_rq_pos(rqa
) > blk_rq_pos(rqb
);
1904 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1908 if (list_empty(&plug
->mq_list
))
1910 list_splice_init(&plug
->mq_list
, &list
);
1912 if (plug
->rq_count
> 2 && plug
->multiple_queues
)
1913 list_sort(NULL
, &list
, plug_rq_cmp
);
1918 struct list_head rq_list
;
1919 struct request
*rq
, *head_rq
= list_entry_rq(list
.next
);
1920 struct list_head
*pos
= &head_rq
->queuelist
; /* skip first */
1921 struct blk_mq_hw_ctx
*this_hctx
= head_rq
->mq_hctx
;
1922 struct blk_mq_ctx
*this_ctx
= head_rq
->mq_ctx
;
1923 unsigned int depth
= 1;
1925 list_for_each_continue(pos
, &list
) {
1926 rq
= list_entry_rq(pos
);
1928 if (rq
->mq_hctx
!= this_hctx
|| rq
->mq_ctx
!= this_ctx
)
1933 list_cut_before(&rq_list
, &list
, pos
);
1934 trace_block_unplug(head_rq
->q
, depth
, !from_schedule
);
1935 blk_mq_sched_insert_requests(this_hctx
, this_ctx
, &rq_list
,
1937 } while(!list_empty(&list
));
1940 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
,
1941 unsigned int nr_segs
)
1945 if (bio
->bi_opf
& REQ_RAHEAD
)
1946 rq
->cmd_flags
|= REQ_FAILFAST_MASK
;
1948 rq
->__sector
= bio
->bi_iter
.bi_sector
;
1949 rq
->write_hint
= bio
->bi_write_hint
;
1950 blk_rq_bio_prep(rq
, bio
, nr_segs
);
1952 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
1953 err
= blk_crypto_rq_bio_prep(rq
, bio
, GFP_NOIO
);
1956 blk_account_io_start(rq
);
1959 static blk_status_t
__blk_mq_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1961 blk_qc_t
*cookie
, bool last
)
1963 struct request_queue
*q
= rq
->q
;
1964 struct blk_mq_queue_data bd
= {
1968 blk_qc_t new_cookie
;
1971 new_cookie
= request_to_qc_t(hctx
, rq
);
1974 * For OK queue, we are done. For error, caller may kill it.
1975 * Any other error (busy), just add it to our list as we
1976 * previously would have done.
1978 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1981 blk_mq_update_dispatch_busy(hctx
, false);
1982 *cookie
= new_cookie
;
1984 case BLK_STS_RESOURCE
:
1985 case BLK_STS_DEV_RESOURCE
:
1986 blk_mq_update_dispatch_busy(hctx
, true);
1987 __blk_mq_requeue_request(rq
);
1990 blk_mq_update_dispatch_busy(hctx
, false);
1991 *cookie
= BLK_QC_T_NONE
;
1998 static blk_status_t
__blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
2001 bool bypass_insert
, bool last
)
2003 struct request_queue
*q
= rq
->q
;
2004 bool run_queue
= true;
2008 * RCU or SRCU read lock is needed before checking quiesced flag.
2010 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2011 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2012 * and avoid driver to try to dispatch again.
2014 if (blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(q
)) {
2016 bypass_insert
= false;
2020 if (q
->elevator
&& !bypass_insert
)
2023 budget_token
= blk_mq_get_dispatch_budget(q
);
2024 if (budget_token
< 0)
2027 blk_mq_set_rq_budget_token(rq
, budget_token
);
2029 if (!blk_mq_get_driver_tag(rq
)) {
2030 blk_mq_put_dispatch_budget(q
, budget_token
);
2034 return __blk_mq_issue_directly(hctx
, rq
, cookie
, last
);
2037 return BLK_STS_RESOURCE
;
2039 blk_mq_sched_insert_request(rq
, false, run_queue
, false);
2045 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2046 * @hctx: Pointer of the associated hardware queue.
2047 * @rq: Pointer to request to be sent.
2048 * @cookie: Request queue cookie.
2050 * If the device has enough resources to accept a new request now, send the
2051 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2052 * we can try send it another time in the future. Requests inserted at this
2053 * queue have higher priority.
2055 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
2056 struct request
*rq
, blk_qc_t
*cookie
)
2061 might_sleep_if(hctx
->flags
& BLK_MQ_F_BLOCKING
);
2063 hctx_lock(hctx
, &srcu_idx
);
2065 ret
= __blk_mq_try_issue_directly(hctx
, rq
, cookie
, false, true);
2066 if (ret
== BLK_STS_RESOURCE
|| ret
== BLK_STS_DEV_RESOURCE
)
2067 blk_mq_request_bypass_insert(rq
, false, true);
2068 else if (ret
!= BLK_STS_OK
)
2069 blk_mq_end_request(rq
, ret
);
2071 hctx_unlock(hctx
, srcu_idx
);
2074 blk_status_t
blk_mq_request_issue_directly(struct request
*rq
, bool last
)
2078 blk_qc_t unused_cookie
;
2079 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
2081 hctx_lock(hctx
, &srcu_idx
);
2082 ret
= __blk_mq_try_issue_directly(hctx
, rq
, &unused_cookie
, true, last
);
2083 hctx_unlock(hctx
, srcu_idx
);
2088 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx
*hctx
,
2089 struct list_head
*list
)
2094 while (!list_empty(list
)) {
2096 struct request
*rq
= list_first_entry(list
, struct request
,
2099 list_del_init(&rq
->queuelist
);
2100 ret
= blk_mq_request_issue_directly(rq
, list_empty(list
));
2101 if (ret
!= BLK_STS_OK
) {
2102 if (ret
== BLK_STS_RESOURCE
||
2103 ret
== BLK_STS_DEV_RESOURCE
) {
2104 blk_mq_request_bypass_insert(rq
, false,
2108 blk_mq_end_request(rq
, ret
);
2115 * If we didn't flush the entire list, we could have told
2116 * the driver there was more coming, but that turned out to
2119 if ((!list_empty(list
) || errors
) &&
2120 hctx
->queue
->mq_ops
->commit_rqs
&& queued
)
2121 hctx
->queue
->mq_ops
->commit_rqs(hctx
);
2124 static void blk_add_rq_to_plug(struct blk_plug
*plug
, struct request
*rq
)
2126 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
2128 if (!plug
->multiple_queues
&& !list_is_singular(&plug
->mq_list
)) {
2129 struct request
*tmp
;
2131 tmp
= list_first_entry(&plug
->mq_list
, struct request
,
2133 if (tmp
->q
!= rq
->q
)
2134 plug
->multiple_queues
= true;
2139 * blk_mq_submit_bio - Create and send a request to block device.
2140 * @bio: Bio pointer.
2142 * Builds up a request structure from @q and @bio and send to the device. The
2143 * request may not be queued directly to hardware if:
2144 * * This request can be merged with another one
2145 * * We want to place request at plug queue for possible future merging
2146 * * There is an IO scheduler active at this queue
2148 * It will not queue the request if there is an error with the bio, or at the
2151 * Returns: Request queue cookie.
2153 blk_qc_t
blk_mq_submit_bio(struct bio
*bio
)
2155 struct request_queue
*q
= bio
->bi_bdev
->bd_disk
->queue
;
2156 const int is_sync
= op_is_sync(bio
->bi_opf
);
2157 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
2158 struct blk_mq_alloc_data data
= {
2162 struct blk_plug
*plug
;
2163 struct request
*same_queue_rq
= NULL
;
2164 unsigned int nr_segs
;
2169 blk_queue_bounce(q
, &bio
);
2170 __blk_queue_split(&bio
, &nr_segs
);
2172 if (!bio_integrity_prep(bio
))
2175 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
2176 blk_attempt_plug_merge(q
, bio
, nr_segs
, &same_queue_rq
))
2179 if (blk_mq_sched_bio_merge(q
, bio
, nr_segs
))
2182 rq_qos_throttle(q
, bio
);
2184 hipri
= bio
->bi_opf
& REQ_HIPRI
;
2186 data
.cmd_flags
= bio
->bi_opf
;
2187 rq
= __blk_mq_alloc_request(&data
);
2188 if (unlikely(!rq
)) {
2189 rq_qos_cleanup(q
, bio
);
2190 if (bio
->bi_opf
& REQ_NOWAIT
)
2191 bio_wouldblock_error(bio
);
2195 trace_block_getrq(bio
);
2197 rq_qos_track(q
, rq
, bio
);
2199 cookie
= request_to_qc_t(data
.hctx
, rq
);
2201 blk_mq_bio_to_request(rq
, bio
, nr_segs
);
2203 ret
= blk_crypto_init_request(rq
);
2204 if (ret
!= BLK_STS_OK
) {
2205 bio
->bi_status
= ret
;
2207 blk_mq_free_request(rq
);
2208 return BLK_QC_T_NONE
;
2211 plug
= blk_mq_plug(q
, bio
);
2212 if (unlikely(is_flush_fua
)) {
2213 /* Bypass scheduler for flush requests */
2214 blk_insert_flush(rq
);
2215 blk_mq_run_hw_queue(data
.hctx
, true);
2216 } else if (plug
&& (q
->nr_hw_queues
== 1 ||
2217 blk_mq_is_sbitmap_shared(rq
->mq_hctx
->flags
) ||
2218 q
->mq_ops
->commit_rqs
|| !blk_queue_nonrot(q
))) {
2220 * Use plugging if we have a ->commit_rqs() hook as well, as
2221 * we know the driver uses bd->last in a smart fashion.
2223 * Use normal plugging if this disk is slow HDD, as sequential
2224 * IO may benefit a lot from plug merging.
2226 unsigned int request_count
= plug
->rq_count
;
2227 struct request
*last
= NULL
;
2230 trace_block_plug(q
);
2232 last
= list_entry_rq(plug
->mq_list
.prev
);
2234 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
2235 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
2236 blk_flush_plug_list(plug
, false);
2237 trace_block_plug(q
);
2240 blk_add_rq_to_plug(plug
, rq
);
2241 } else if (q
->elevator
) {
2242 /* Insert the request at the IO scheduler queue */
2243 blk_mq_sched_insert_request(rq
, false, true, true);
2244 } else if (plug
&& !blk_queue_nomerges(q
)) {
2246 * We do limited plugging. If the bio can be merged, do that.
2247 * Otherwise the existing request in the plug list will be
2248 * issued. So the plug list will have one request at most
2249 * The plug list might get flushed before this. If that happens,
2250 * the plug list is empty, and same_queue_rq is invalid.
2252 if (list_empty(&plug
->mq_list
))
2253 same_queue_rq
= NULL
;
2254 if (same_queue_rq
) {
2255 list_del_init(&same_queue_rq
->queuelist
);
2258 blk_add_rq_to_plug(plug
, rq
);
2259 trace_block_plug(q
);
2261 if (same_queue_rq
) {
2262 data
.hctx
= same_queue_rq
->mq_hctx
;
2263 trace_block_unplug(q
, 1, true);
2264 blk_mq_try_issue_directly(data
.hctx
, same_queue_rq
,
2267 } else if ((q
->nr_hw_queues
> 1 && is_sync
) ||
2268 !data
.hctx
->dispatch_busy
) {
2270 * There is no scheduler and we can try to send directly
2273 blk_mq_try_issue_directly(data
.hctx
, rq
, &cookie
);
2276 blk_mq_sched_insert_request(rq
, false, true, true);
2280 return BLK_QC_T_NONE
;
2284 return BLK_QC_T_NONE
;
2287 static size_t order_to_size(unsigned int order
)
2289 return (size_t)PAGE_SIZE
<< order
;
2292 /* called before freeing request pool in @tags */
2293 static void blk_mq_clear_rq_mapping(struct blk_mq_tag_set
*set
,
2294 struct blk_mq_tags
*tags
, unsigned int hctx_idx
)
2296 struct blk_mq_tags
*drv_tags
= set
->tags
[hctx_idx
];
2298 unsigned long flags
;
2300 list_for_each_entry(page
, &tags
->page_list
, lru
) {
2301 unsigned long start
= (unsigned long)page_address(page
);
2302 unsigned long end
= start
+ order_to_size(page
->private);
2305 for (i
= 0; i
< set
->queue_depth
; i
++) {
2306 struct request
*rq
= drv_tags
->rqs
[i
];
2307 unsigned long rq_addr
= (unsigned long)rq
;
2309 if (rq_addr
>= start
&& rq_addr
< end
) {
2310 WARN_ON_ONCE(refcount_read(&rq
->ref
) != 0);
2311 cmpxchg(&drv_tags
->rqs
[i
], rq
, NULL
);
2317 * Wait until all pending iteration is done.
2319 * Request reference is cleared and it is guaranteed to be observed
2320 * after the ->lock is released.
2322 spin_lock_irqsave(&drv_tags
->lock
, flags
);
2323 spin_unlock_irqrestore(&drv_tags
->lock
, flags
);
2326 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
2327 unsigned int hctx_idx
)
2331 if (tags
->rqs
&& set
->ops
->exit_request
) {
2334 for (i
= 0; i
< tags
->nr_tags
; i
++) {
2335 struct request
*rq
= tags
->static_rqs
[i
];
2339 set
->ops
->exit_request(set
, rq
, hctx_idx
);
2340 tags
->static_rqs
[i
] = NULL
;
2344 blk_mq_clear_rq_mapping(set
, tags
, hctx_idx
);
2346 while (!list_empty(&tags
->page_list
)) {
2347 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
2348 list_del_init(&page
->lru
);
2350 * Remove kmemleak object previously allocated in
2351 * blk_mq_alloc_rqs().
2353 kmemleak_free(page_address(page
));
2354 __free_pages(page
, page
->private);
2358 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
, unsigned int flags
)
2362 kfree(tags
->static_rqs
);
2363 tags
->static_rqs
= NULL
;
2365 blk_mq_free_tags(tags
, flags
);
2368 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
2369 unsigned int hctx_idx
,
2370 unsigned int nr_tags
,
2371 unsigned int reserved_tags
,
2374 struct blk_mq_tags
*tags
;
2377 node
= blk_mq_hw_queue_to_node(&set
->map
[HCTX_TYPE_DEFAULT
], hctx_idx
);
2378 if (node
== NUMA_NO_NODE
)
2379 node
= set
->numa_node
;
2381 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
, flags
);
2385 tags
->rqs
= kcalloc_node(nr_tags
, sizeof(struct request
*),
2386 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
2389 blk_mq_free_tags(tags
, flags
);
2393 tags
->static_rqs
= kcalloc_node(nr_tags
, sizeof(struct request
*),
2394 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
2396 if (!tags
->static_rqs
) {
2398 blk_mq_free_tags(tags
, flags
);
2405 static int blk_mq_init_request(struct blk_mq_tag_set
*set
, struct request
*rq
,
2406 unsigned int hctx_idx
, int node
)
2410 if (set
->ops
->init_request
) {
2411 ret
= set
->ops
->init_request(set
, rq
, hctx_idx
, node
);
2416 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
2420 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
2421 unsigned int hctx_idx
, unsigned int depth
)
2423 unsigned int i
, j
, entries_per_page
, max_order
= 4;
2424 size_t rq_size
, left
;
2427 node
= blk_mq_hw_queue_to_node(&set
->map
[HCTX_TYPE_DEFAULT
], hctx_idx
);
2428 if (node
== NUMA_NO_NODE
)
2429 node
= set
->numa_node
;
2431 INIT_LIST_HEAD(&tags
->page_list
);
2434 * rq_size is the size of the request plus driver payload, rounded
2435 * to the cacheline size
2437 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
2439 left
= rq_size
* depth
;
2441 for (i
= 0; i
< depth
; ) {
2442 int this_order
= max_order
;
2447 while (this_order
&& left
< order_to_size(this_order
- 1))
2451 page
= alloc_pages_node(node
,
2452 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
2458 if (order_to_size(this_order
) < rq_size
)
2465 page
->private = this_order
;
2466 list_add_tail(&page
->lru
, &tags
->page_list
);
2468 p
= page_address(page
);
2470 * Allow kmemleak to scan these pages as they contain pointers
2471 * to additional allocations like via ops->init_request().
2473 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
2474 entries_per_page
= order_to_size(this_order
) / rq_size
;
2475 to_do
= min(entries_per_page
, depth
- i
);
2476 left
-= to_do
* rq_size
;
2477 for (j
= 0; j
< to_do
; j
++) {
2478 struct request
*rq
= p
;
2480 tags
->static_rqs
[i
] = rq
;
2481 if (blk_mq_init_request(set
, rq
, hctx_idx
, node
)) {
2482 tags
->static_rqs
[i
] = NULL
;
2493 blk_mq_free_rqs(set
, tags
, hctx_idx
);
2497 struct rq_iter_data
{
2498 struct blk_mq_hw_ctx
*hctx
;
2502 static bool blk_mq_has_request(struct request
*rq
, void *data
, bool reserved
)
2504 struct rq_iter_data
*iter_data
= data
;
2506 if (rq
->mq_hctx
!= iter_data
->hctx
)
2508 iter_data
->has_rq
= true;
2512 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx
*hctx
)
2514 struct blk_mq_tags
*tags
= hctx
->sched_tags
?
2515 hctx
->sched_tags
: hctx
->tags
;
2516 struct rq_iter_data data
= {
2520 blk_mq_all_tag_iter(tags
, blk_mq_has_request
, &data
);
2524 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu
,
2525 struct blk_mq_hw_ctx
*hctx
)
2527 if (cpumask_next_and(-1, hctx
->cpumask
, cpu_online_mask
) != cpu
)
2529 if (cpumask_next_and(cpu
, hctx
->cpumask
, cpu_online_mask
) < nr_cpu_ids
)
2534 static int blk_mq_hctx_notify_offline(unsigned int cpu
, struct hlist_node
*node
)
2536 struct blk_mq_hw_ctx
*hctx
= hlist_entry_safe(node
,
2537 struct blk_mq_hw_ctx
, cpuhp_online
);
2539 if (!cpumask_test_cpu(cpu
, hctx
->cpumask
) ||
2540 !blk_mq_last_cpu_in_hctx(cpu
, hctx
))
2544 * Prevent new request from being allocated on the current hctx.
2546 * The smp_mb__after_atomic() Pairs with the implied barrier in
2547 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
2548 * seen once we return from the tag allocator.
2550 set_bit(BLK_MQ_S_INACTIVE
, &hctx
->state
);
2551 smp_mb__after_atomic();
2554 * Try to grab a reference to the queue and wait for any outstanding
2555 * requests. If we could not grab a reference the queue has been
2556 * frozen and there are no requests.
2558 if (percpu_ref_tryget(&hctx
->queue
->q_usage_counter
)) {
2559 while (blk_mq_hctx_has_requests(hctx
))
2561 percpu_ref_put(&hctx
->queue
->q_usage_counter
);
2567 static int blk_mq_hctx_notify_online(unsigned int cpu
, struct hlist_node
*node
)
2569 struct blk_mq_hw_ctx
*hctx
= hlist_entry_safe(node
,
2570 struct blk_mq_hw_ctx
, cpuhp_online
);
2572 if (cpumask_test_cpu(cpu
, hctx
->cpumask
))
2573 clear_bit(BLK_MQ_S_INACTIVE
, &hctx
->state
);
2578 * 'cpu' is going away. splice any existing rq_list entries from this
2579 * software queue to the hw queue dispatch list, and ensure that it
2582 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
2584 struct blk_mq_hw_ctx
*hctx
;
2585 struct blk_mq_ctx
*ctx
;
2587 enum hctx_type type
;
2589 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
2590 if (!cpumask_test_cpu(cpu
, hctx
->cpumask
))
2593 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
2596 spin_lock(&ctx
->lock
);
2597 if (!list_empty(&ctx
->rq_lists
[type
])) {
2598 list_splice_init(&ctx
->rq_lists
[type
], &tmp
);
2599 blk_mq_hctx_clear_pending(hctx
, ctx
);
2601 spin_unlock(&ctx
->lock
);
2603 if (list_empty(&tmp
))
2606 spin_lock(&hctx
->lock
);
2607 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
2608 spin_unlock(&hctx
->lock
);
2610 blk_mq_run_hw_queue(hctx
, true);
2614 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
2616 if (!(hctx
->flags
& BLK_MQ_F_STACKING
))
2617 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE
,
2618 &hctx
->cpuhp_online
);
2619 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
2624 * Before freeing hw queue, clearing the flush request reference in
2625 * tags->rqs[] for avoiding potential UAF.
2627 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags
*tags
,
2628 unsigned int queue_depth
, struct request
*flush_rq
)
2631 unsigned long flags
;
2633 /* The hw queue may not be mapped yet */
2637 WARN_ON_ONCE(refcount_read(&flush_rq
->ref
) != 0);
2639 for (i
= 0; i
< queue_depth
; i
++)
2640 cmpxchg(&tags
->rqs
[i
], flush_rq
, NULL
);
2643 * Wait until all pending iteration is done.
2645 * Request reference is cleared and it is guaranteed to be observed
2646 * after the ->lock is released.
2648 spin_lock_irqsave(&tags
->lock
, flags
);
2649 spin_unlock_irqrestore(&tags
->lock
, flags
);
2652 /* hctx->ctxs will be freed in queue's release handler */
2653 static void blk_mq_exit_hctx(struct request_queue
*q
,
2654 struct blk_mq_tag_set
*set
,
2655 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
2657 struct request
*flush_rq
= hctx
->fq
->flush_rq
;
2659 if (blk_mq_hw_queue_mapped(hctx
))
2660 blk_mq_tag_idle(hctx
);
2662 blk_mq_clear_flush_rq_mapping(set
->tags
[hctx_idx
],
2663 set
->queue_depth
, flush_rq
);
2664 if (set
->ops
->exit_request
)
2665 set
->ops
->exit_request(set
, flush_rq
, hctx_idx
);
2667 if (set
->ops
->exit_hctx
)
2668 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2670 blk_mq_remove_cpuhp(hctx
);
2672 spin_lock(&q
->unused_hctx_lock
);
2673 list_add(&hctx
->hctx_list
, &q
->unused_hctx_list
);
2674 spin_unlock(&q
->unused_hctx_lock
);
2677 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
2678 struct blk_mq_tag_set
*set
, int nr_queue
)
2680 struct blk_mq_hw_ctx
*hctx
;
2683 queue_for_each_hw_ctx(q
, hctx
, i
) {
2686 blk_mq_debugfs_unregister_hctx(hctx
);
2687 blk_mq_exit_hctx(q
, set
, hctx
, i
);
2691 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set
*tag_set
)
2693 int hw_ctx_size
= sizeof(struct blk_mq_hw_ctx
);
2695 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx
, srcu
),
2696 __alignof__(struct blk_mq_hw_ctx
)) !=
2697 sizeof(struct blk_mq_hw_ctx
));
2699 if (tag_set
->flags
& BLK_MQ_F_BLOCKING
)
2700 hw_ctx_size
+= sizeof(struct srcu_struct
);
2705 static int blk_mq_init_hctx(struct request_queue
*q
,
2706 struct blk_mq_tag_set
*set
,
2707 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
2709 hctx
->queue_num
= hctx_idx
;
2711 if (!(hctx
->flags
& BLK_MQ_F_STACKING
))
2712 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE
,
2713 &hctx
->cpuhp_online
);
2714 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
2716 hctx
->tags
= set
->tags
[hctx_idx
];
2718 if (set
->ops
->init_hctx
&&
2719 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
2720 goto unregister_cpu_notifier
;
2722 if (blk_mq_init_request(set
, hctx
->fq
->flush_rq
, hctx_idx
,
2728 if (set
->ops
->exit_hctx
)
2729 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2730 unregister_cpu_notifier
:
2731 blk_mq_remove_cpuhp(hctx
);
2735 static struct blk_mq_hw_ctx
*
2736 blk_mq_alloc_hctx(struct request_queue
*q
, struct blk_mq_tag_set
*set
,
2739 struct blk_mq_hw_ctx
*hctx
;
2740 gfp_t gfp
= GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
;
2742 hctx
= kzalloc_node(blk_mq_hw_ctx_size(set
), gfp
, node
);
2744 goto fail_alloc_hctx
;
2746 if (!zalloc_cpumask_var_node(&hctx
->cpumask
, gfp
, node
))
2749 atomic_set(&hctx
->nr_active
, 0);
2750 if (node
== NUMA_NO_NODE
)
2751 node
= set
->numa_node
;
2752 hctx
->numa_node
= node
;
2754 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
2755 spin_lock_init(&hctx
->lock
);
2756 INIT_LIST_HEAD(&hctx
->dispatch
);
2758 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_QUEUE_SHARED
;
2760 INIT_LIST_HEAD(&hctx
->hctx_list
);
2763 * Allocate space for all possible cpus to avoid allocation at
2766 hctx
->ctxs
= kmalloc_array_node(nr_cpu_ids
, sizeof(void *),
2771 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8),
2772 gfp
, node
, false, false))
2776 spin_lock_init(&hctx
->dispatch_wait_lock
);
2777 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
2778 INIT_LIST_HEAD(&hctx
->dispatch_wait
.entry
);
2780 hctx
->fq
= blk_alloc_flush_queue(hctx
->numa_node
, set
->cmd_size
, gfp
);
2784 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
2785 init_srcu_struct(hctx
->srcu
);
2786 blk_mq_hctx_kobj_init(hctx
);
2791 sbitmap_free(&hctx
->ctx_map
);
2795 free_cpumask_var(hctx
->cpumask
);
2802 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
2803 unsigned int nr_hw_queues
)
2805 struct blk_mq_tag_set
*set
= q
->tag_set
;
2808 for_each_possible_cpu(i
) {
2809 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2810 struct blk_mq_hw_ctx
*hctx
;
2814 spin_lock_init(&__ctx
->lock
);
2815 for (k
= HCTX_TYPE_DEFAULT
; k
< HCTX_MAX_TYPES
; k
++)
2816 INIT_LIST_HEAD(&__ctx
->rq_lists
[k
]);
2821 * Set local node, IFF we have more than one hw queue. If
2822 * not, we remain on the home node of the device
2824 for (j
= 0; j
< set
->nr_maps
; j
++) {
2825 hctx
= blk_mq_map_queue_type(q
, j
, i
);
2826 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
2827 hctx
->numa_node
= cpu_to_node(i
);
2832 static bool __blk_mq_alloc_map_and_request(struct blk_mq_tag_set
*set
,
2835 unsigned int flags
= set
->flags
;
2838 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
2839 set
->queue_depth
, set
->reserved_tags
, flags
);
2840 if (!set
->tags
[hctx_idx
])
2843 ret
= blk_mq_alloc_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
,
2848 blk_mq_free_rq_map(set
->tags
[hctx_idx
], flags
);
2849 set
->tags
[hctx_idx
] = NULL
;
2853 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set
*set
,
2854 unsigned int hctx_idx
)
2856 unsigned int flags
= set
->flags
;
2858 if (set
->tags
&& set
->tags
[hctx_idx
]) {
2859 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
2860 blk_mq_free_rq_map(set
->tags
[hctx_idx
], flags
);
2861 set
->tags
[hctx_idx
] = NULL
;
2865 static void blk_mq_map_swqueue(struct request_queue
*q
)
2867 unsigned int i
, j
, hctx_idx
;
2868 struct blk_mq_hw_ctx
*hctx
;
2869 struct blk_mq_ctx
*ctx
;
2870 struct blk_mq_tag_set
*set
= q
->tag_set
;
2872 queue_for_each_hw_ctx(q
, hctx
, i
) {
2873 cpumask_clear(hctx
->cpumask
);
2875 hctx
->dispatch_from
= NULL
;
2879 * Map software to hardware queues.
2881 * If the cpu isn't present, the cpu is mapped to first hctx.
2883 for_each_possible_cpu(i
) {
2885 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2886 for (j
= 0; j
< set
->nr_maps
; j
++) {
2887 if (!set
->map
[j
].nr_queues
) {
2888 ctx
->hctxs
[j
] = blk_mq_map_queue_type(q
,
2889 HCTX_TYPE_DEFAULT
, i
);
2892 hctx_idx
= set
->map
[j
].mq_map
[i
];
2893 /* unmapped hw queue can be remapped after CPU topo changed */
2894 if (!set
->tags
[hctx_idx
] &&
2895 !__blk_mq_alloc_map_and_request(set
, hctx_idx
)) {
2897 * If tags initialization fail for some hctx,
2898 * that hctx won't be brought online. In this
2899 * case, remap the current ctx to hctx[0] which
2900 * is guaranteed to always have tags allocated
2902 set
->map
[j
].mq_map
[i
] = 0;
2905 hctx
= blk_mq_map_queue_type(q
, j
, i
);
2906 ctx
->hctxs
[j
] = hctx
;
2908 * If the CPU is already set in the mask, then we've
2909 * mapped this one already. This can happen if
2910 * devices share queues across queue maps.
2912 if (cpumask_test_cpu(i
, hctx
->cpumask
))
2915 cpumask_set_cpu(i
, hctx
->cpumask
);
2917 ctx
->index_hw
[hctx
->type
] = hctx
->nr_ctx
;
2918 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2921 * If the nr_ctx type overflows, we have exceeded the
2922 * amount of sw queues we can support.
2924 BUG_ON(!hctx
->nr_ctx
);
2927 for (; j
< HCTX_MAX_TYPES
; j
++)
2928 ctx
->hctxs
[j
] = blk_mq_map_queue_type(q
,
2929 HCTX_TYPE_DEFAULT
, i
);
2932 queue_for_each_hw_ctx(q
, hctx
, i
) {
2934 * If no software queues are mapped to this hardware queue,
2935 * disable it and free the request entries.
2937 if (!hctx
->nr_ctx
) {
2938 /* Never unmap queue 0. We need it as a
2939 * fallback in case of a new remap fails
2942 if (i
&& set
->tags
[i
])
2943 blk_mq_free_map_and_requests(set
, i
);
2949 hctx
->tags
= set
->tags
[i
];
2950 WARN_ON(!hctx
->tags
);
2953 * Set the map size to the number of mapped software queues.
2954 * This is more accurate and more efficient than looping
2955 * over all possibly mapped software queues.
2957 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2960 * Initialize batch roundrobin counts
2962 hctx
->next_cpu
= blk_mq_first_mapped_cpu(hctx
);
2963 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2968 * Caller needs to ensure that we're either frozen/quiesced, or that
2969 * the queue isn't live yet.
2971 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2973 struct blk_mq_hw_ctx
*hctx
;
2976 queue_for_each_hw_ctx(q
, hctx
, i
) {
2978 hctx
->flags
|= BLK_MQ_F_TAG_QUEUE_SHARED
;
2980 blk_mq_tag_idle(hctx
);
2981 hctx
->flags
&= ~BLK_MQ_F_TAG_QUEUE_SHARED
;
2986 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set
*set
,
2989 struct request_queue
*q
;
2991 lockdep_assert_held(&set
->tag_list_lock
);
2993 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2994 blk_mq_freeze_queue(q
);
2995 queue_set_hctx_shared(q
, shared
);
2996 blk_mq_unfreeze_queue(q
);
3000 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
3002 struct blk_mq_tag_set
*set
= q
->tag_set
;
3004 mutex_lock(&set
->tag_list_lock
);
3005 list_del(&q
->tag_set_list
);
3006 if (list_is_singular(&set
->tag_list
)) {
3007 /* just transitioned to unshared */
3008 set
->flags
&= ~BLK_MQ_F_TAG_QUEUE_SHARED
;
3009 /* update existing queue */
3010 blk_mq_update_tag_set_shared(set
, false);
3012 mutex_unlock(&set
->tag_list_lock
);
3013 INIT_LIST_HEAD(&q
->tag_set_list
);
3016 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
3017 struct request_queue
*q
)
3019 mutex_lock(&set
->tag_list_lock
);
3022 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3024 if (!list_empty(&set
->tag_list
) &&
3025 !(set
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
)) {
3026 set
->flags
|= BLK_MQ_F_TAG_QUEUE_SHARED
;
3027 /* update existing queue */
3028 blk_mq_update_tag_set_shared(set
, true);
3030 if (set
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
)
3031 queue_set_hctx_shared(q
, true);
3032 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
3034 mutex_unlock(&set
->tag_list_lock
);
3037 /* All allocations will be freed in release handler of q->mq_kobj */
3038 static int blk_mq_alloc_ctxs(struct request_queue
*q
)
3040 struct blk_mq_ctxs
*ctxs
;
3043 ctxs
= kzalloc(sizeof(*ctxs
), GFP_KERNEL
);
3047 ctxs
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
3048 if (!ctxs
->queue_ctx
)
3051 for_each_possible_cpu(cpu
) {
3052 struct blk_mq_ctx
*ctx
= per_cpu_ptr(ctxs
->queue_ctx
, cpu
);
3056 q
->mq_kobj
= &ctxs
->kobj
;
3057 q
->queue_ctx
= ctxs
->queue_ctx
;
3066 * It is the actual release handler for mq, but we do it from
3067 * request queue's release handler for avoiding use-after-free
3068 * and headache because q->mq_kobj shouldn't have been introduced,
3069 * but we can't group ctx/kctx kobj without it.
3071 void blk_mq_release(struct request_queue
*q
)
3073 struct blk_mq_hw_ctx
*hctx
, *next
;
3076 queue_for_each_hw_ctx(q
, hctx
, i
)
3077 WARN_ON_ONCE(hctx
&& list_empty(&hctx
->hctx_list
));
3079 /* all hctx are in .unused_hctx_list now */
3080 list_for_each_entry_safe(hctx
, next
, &q
->unused_hctx_list
, hctx_list
) {
3081 list_del_init(&hctx
->hctx_list
);
3082 kobject_put(&hctx
->kobj
);
3085 kfree(q
->queue_hw_ctx
);
3088 * release .mq_kobj and sw queue's kobject now because
3089 * both share lifetime with request queue.
3091 blk_mq_sysfs_deinit(q
);
3094 static struct request_queue
*blk_mq_init_queue_data(struct blk_mq_tag_set
*set
,
3097 struct request_queue
*q
;
3100 q
= blk_alloc_queue(set
->numa_node
);
3102 return ERR_PTR(-ENOMEM
);
3103 q
->queuedata
= queuedata
;
3104 ret
= blk_mq_init_allocated_queue(set
, q
);
3106 blk_cleanup_queue(q
);
3107 return ERR_PTR(ret
);
3112 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
3114 return blk_mq_init_queue_data(set
, NULL
);
3116 EXPORT_SYMBOL(blk_mq_init_queue
);
3118 struct gendisk
*__blk_mq_alloc_disk(struct blk_mq_tag_set
*set
, void *queuedata
,
3119 struct lock_class_key
*lkclass
)
3121 struct request_queue
*q
;
3122 struct gendisk
*disk
;
3124 q
= blk_mq_init_queue_data(set
, queuedata
);
3128 disk
= __alloc_disk_node(q
, set
->numa_node
, lkclass
);
3130 blk_cleanup_queue(q
);
3131 return ERR_PTR(-ENOMEM
);
3135 EXPORT_SYMBOL(__blk_mq_alloc_disk
);
3137 static struct blk_mq_hw_ctx
*blk_mq_alloc_and_init_hctx(
3138 struct blk_mq_tag_set
*set
, struct request_queue
*q
,
3139 int hctx_idx
, int node
)
3141 struct blk_mq_hw_ctx
*hctx
= NULL
, *tmp
;
3143 /* reuse dead hctx first */
3144 spin_lock(&q
->unused_hctx_lock
);
3145 list_for_each_entry(tmp
, &q
->unused_hctx_list
, hctx_list
) {
3146 if (tmp
->numa_node
== node
) {
3152 list_del_init(&hctx
->hctx_list
);
3153 spin_unlock(&q
->unused_hctx_lock
);
3156 hctx
= blk_mq_alloc_hctx(q
, set
, node
);
3160 if (blk_mq_init_hctx(q
, set
, hctx
, hctx_idx
))
3166 kobject_put(&hctx
->kobj
);
3171 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
3172 struct request_queue
*q
)
3175 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
3177 if (q
->nr_hw_queues
< set
->nr_hw_queues
) {
3178 struct blk_mq_hw_ctx
**new_hctxs
;
3180 new_hctxs
= kcalloc_node(set
->nr_hw_queues
,
3181 sizeof(*new_hctxs
), GFP_KERNEL
,
3186 memcpy(new_hctxs
, hctxs
, q
->nr_hw_queues
*
3188 q
->queue_hw_ctx
= new_hctxs
;
3193 /* protect against switching io scheduler */
3194 mutex_lock(&q
->sysfs_lock
);
3195 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
3197 struct blk_mq_hw_ctx
*hctx
;
3199 node
= blk_mq_hw_queue_to_node(&set
->map
[HCTX_TYPE_DEFAULT
], i
);
3201 * If the hw queue has been mapped to another numa node,
3202 * we need to realloc the hctx. If allocation fails, fallback
3203 * to use the previous one.
3205 if (hctxs
[i
] && (hctxs
[i
]->numa_node
== node
))
3208 hctx
= blk_mq_alloc_and_init_hctx(set
, q
, i
, node
);
3211 blk_mq_exit_hctx(q
, set
, hctxs
[i
], i
);
3215 pr_warn("Allocate new hctx on node %d fails,\
3216 fallback to previous one on node %d\n",
3217 node
, hctxs
[i
]->numa_node
);
3223 * Increasing nr_hw_queues fails. Free the newly allocated
3224 * hctxs and keep the previous q->nr_hw_queues.
3226 if (i
!= set
->nr_hw_queues
) {
3227 j
= q
->nr_hw_queues
;
3231 end
= q
->nr_hw_queues
;
3232 q
->nr_hw_queues
= set
->nr_hw_queues
;
3235 for (; j
< end
; j
++) {
3236 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
3240 blk_mq_free_map_and_requests(set
, j
);
3241 blk_mq_exit_hctx(q
, set
, hctx
, j
);
3245 mutex_unlock(&q
->sysfs_lock
);
3248 int blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
3249 struct request_queue
*q
)
3251 /* mark the queue as mq asap */
3252 q
->mq_ops
= set
->ops
;
3254 q
->poll_cb
= blk_stat_alloc_callback(blk_mq_poll_stats_fn
,
3255 blk_mq_poll_stats_bkt
,
3256 BLK_MQ_POLL_STATS_BKTS
, q
);
3260 if (blk_mq_alloc_ctxs(q
))
3263 /* init q->mq_kobj and sw queues' kobjects */
3264 blk_mq_sysfs_init(q
);
3266 INIT_LIST_HEAD(&q
->unused_hctx_list
);
3267 spin_lock_init(&q
->unused_hctx_lock
);
3269 blk_mq_realloc_hw_ctxs(set
, q
);
3270 if (!q
->nr_hw_queues
)
3273 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
3274 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
3278 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
3279 if (set
->nr_maps
> HCTX_TYPE_POLL
&&
3280 set
->map
[HCTX_TYPE_POLL
].nr_queues
)
3281 blk_queue_flag_set(QUEUE_FLAG_POLL
, q
);
3283 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
3284 INIT_LIST_HEAD(&q
->requeue_list
);
3285 spin_lock_init(&q
->requeue_lock
);
3287 q
->nr_requests
= set
->queue_depth
;
3290 * Default to classic polling
3292 q
->poll_nsec
= BLK_MQ_POLL_CLASSIC
;
3294 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
3295 blk_mq_add_queue_tag_set(set
, q
);
3296 blk_mq_map_swqueue(q
);
3300 kfree(q
->queue_hw_ctx
);
3301 q
->nr_hw_queues
= 0;
3302 blk_mq_sysfs_deinit(q
);
3304 blk_stat_free_callback(q
->poll_cb
);
3310 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
3312 /* tags can _not_ be used after returning from blk_mq_exit_queue */
3313 void blk_mq_exit_queue(struct request_queue
*q
)
3315 struct blk_mq_tag_set
*set
= q
->tag_set
;
3317 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
3318 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
3319 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
3320 blk_mq_del_queue_tag_set(q
);
3323 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
3327 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
3328 if (!__blk_mq_alloc_map_and_request(set
, i
))
3337 blk_mq_free_map_and_requests(set
, i
);
3343 * Allocate the request maps associated with this tag_set. Note that this
3344 * may reduce the depth asked for, if memory is tight. set->queue_depth
3345 * will be updated to reflect the allocated depth.
3347 static int blk_mq_alloc_map_and_requests(struct blk_mq_tag_set
*set
)
3352 depth
= set
->queue_depth
;
3354 err
= __blk_mq_alloc_rq_maps(set
);
3358 set
->queue_depth
>>= 1;
3359 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
3363 } while (set
->queue_depth
);
3365 if (!set
->queue_depth
|| err
) {
3366 pr_err("blk-mq: failed to allocate request map\n");
3370 if (depth
!= set
->queue_depth
)
3371 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3372 depth
, set
->queue_depth
);
3377 static int blk_mq_update_queue_map(struct blk_mq_tag_set
*set
)
3380 * blk_mq_map_queues() and multiple .map_queues() implementations
3381 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
3382 * number of hardware queues.
3384 if (set
->nr_maps
== 1)
3385 set
->map
[HCTX_TYPE_DEFAULT
].nr_queues
= set
->nr_hw_queues
;
3387 if (set
->ops
->map_queues
&& !is_kdump_kernel()) {
3391 * transport .map_queues is usually done in the following
3394 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3395 * mask = get_cpu_mask(queue)
3396 * for_each_cpu(cpu, mask)
3397 * set->map[x].mq_map[cpu] = queue;
3400 * When we need to remap, the table has to be cleared for
3401 * killing stale mapping since one CPU may not be mapped
3404 for (i
= 0; i
< set
->nr_maps
; i
++)
3405 blk_mq_clear_mq_map(&set
->map
[i
]);
3407 return set
->ops
->map_queues(set
);
3409 BUG_ON(set
->nr_maps
> 1);
3410 return blk_mq_map_queues(&set
->map
[HCTX_TYPE_DEFAULT
]);
3414 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set
*set
,
3415 int cur_nr_hw_queues
, int new_nr_hw_queues
)
3417 struct blk_mq_tags
**new_tags
;
3419 if (cur_nr_hw_queues
>= new_nr_hw_queues
)
3422 new_tags
= kcalloc_node(new_nr_hw_queues
, sizeof(struct blk_mq_tags
*),
3423 GFP_KERNEL
, set
->numa_node
);
3428 memcpy(new_tags
, set
->tags
, cur_nr_hw_queues
*
3429 sizeof(*set
->tags
));
3431 set
->tags
= new_tags
;
3432 set
->nr_hw_queues
= new_nr_hw_queues
;
3437 static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set
*set
,
3438 int new_nr_hw_queues
)
3440 return blk_mq_realloc_tag_set_tags(set
, 0, new_nr_hw_queues
);
3444 * Alloc a tag set to be associated with one or more request queues.
3445 * May fail with EINVAL for various error conditions. May adjust the
3446 * requested depth down, if it's too large. In that case, the set
3447 * value will be stored in set->queue_depth.
3449 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
3453 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
3455 if (!set
->nr_hw_queues
)
3457 if (!set
->queue_depth
)
3459 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
3462 if (!set
->ops
->queue_rq
)
3465 if (!set
->ops
->get_budget
^ !set
->ops
->put_budget
)
3468 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
3469 pr_info("blk-mq: reduced tag depth to %u\n",
3471 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
3476 else if (set
->nr_maps
> HCTX_MAX_TYPES
)
3480 * If a crashdump is active, then we are potentially in a very
3481 * memory constrained environment. Limit us to 1 queue and
3482 * 64 tags to prevent using too much memory.
3484 if (is_kdump_kernel()) {
3485 set
->nr_hw_queues
= 1;
3487 set
->queue_depth
= min(64U, set
->queue_depth
);
3490 * There is no use for more h/w queues than cpus if we just have
3493 if (set
->nr_maps
== 1 && set
->nr_hw_queues
> nr_cpu_ids
)
3494 set
->nr_hw_queues
= nr_cpu_ids
;
3496 if (blk_mq_alloc_tag_set_tags(set
, set
->nr_hw_queues
) < 0)
3500 for (i
= 0; i
< set
->nr_maps
; i
++) {
3501 set
->map
[i
].mq_map
= kcalloc_node(nr_cpu_ids
,
3502 sizeof(set
->map
[i
].mq_map
[0]),
3503 GFP_KERNEL
, set
->numa_node
);
3504 if (!set
->map
[i
].mq_map
)
3505 goto out_free_mq_map
;
3506 set
->map
[i
].nr_queues
= is_kdump_kernel() ? 1 : set
->nr_hw_queues
;
3509 ret
= blk_mq_update_queue_map(set
);
3511 goto out_free_mq_map
;
3513 ret
= blk_mq_alloc_map_and_requests(set
);
3515 goto out_free_mq_map
;
3517 if (blk_mq_is_sbitmap_shared(set
->flags
)) {
3518 atomic_set(&set
->active_queues_shared_sbitmap
, 0);
3520 if (blk_mq_init_shared_sbitmap(set
)) {
3522 goto out_free_mq_rq_maps
;
3526 mutex_init(&set
->tag_list_lock
);
3527 INIT_LIST_HEAD(&set
->tag_list
);
3531 out_free_mq_rq_maps
:
3532 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
3533 blk_mq_free_map_and_requests(set
, i
);
3535 for (i
= 0; i
< set
->nr_maps
; i
++) {
3536 kfree(set
->map
[i
].mq_map
);
3537 set
->map
[i
].mq_map
= NULL
;
3543 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
3545 /* allocate and initialize a tagset for a simple single-queue device */
3546 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set
*set
,
3547 const struct blk_mq_ops
*ops
, unsigned int queue_depth
,
3548 unsigned int set_flags
)
3550 memset(set
, 0, sizeof(*set
));
3552 set
->nr_hw_queues
= 1;
3554 set
->queue_depth
= queue_depth
;
3555 set
->numa_node
= NUMA_NO_NODE
;
3556 set
->flags
= set_flags
;
3557 return blk_mq_alloc_tag_set(set
);
3559 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set
);
3561 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
3565 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
3566 blk_mq_free_map_and_requests(set
, i
);
3568 if (blk_mq_is_sbitmap_shared(set
->flags
))
3569 blk_mq_exit_shared_sbitmap(set
);
3571 for (j
= 0; j
< set
->nr_maps
; j
++) {
3572 kfree(set
->map
[j
].mq_map
);
3573 set
->map
[j
].mq_map
= NULL
;
3579 EXPORT_SYMBOL(blk_mq_free_tag_set
);
3581 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
3583 struct blk_mq_tag_set
*set
= q
->tag_set
;
3584 struct blk_mq_hw_ctx
*hctx
;
3590 if (q
->nr_requests
== nr
)
3593 blk_mq_freeze_queue(q
);
3594 blk_mq_quiesce_queue(q
);
3597 queue_for_each_hw_ctx(q
, hctx
, i
) {
3601 * If we're using an MQ scheduler, just update the scheduler
3602 * queue depth. This is similar to what the old code would do.
3604 if (!hctx
->sched_tags
) {
3605 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
, nr
,
3607 if (!ret
&& blk_mq_is_sbitmap_shared(set
->flags
))
3608 blk_mq_tag_resize_shared_sbitmap(set
, nr
);
3610 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
3612 if (blk_mq_is_sbitmap_shared(set
->flags
)) {
3613 hctx
->sched_tags
->bitmap_tags
=
3614 &q
->sched_bitmap_tags
;
3615 hctx
->sched_tags
->breserved_tags
=
3616 &q
->sched_breserved_tags
;
3621 if (q
->elevator
&& q
->elevator
->type
->ops
.depth_updated
)
3622 q
->elevator
->type
->ops
.depth_updated(hctx
);
3625 q
->nr_requests
= nr
;
3626 if (q
->elevator
&& blk_mq_is_sbitmap_shared(set
->flags
))
3627 sbitmap_queue_resize(&q
->sched_bitmap_tags
,
3628 nr
- set
->reserved_tags
);
3631 blk_mq_unquiesce_queue(q
);
3632 blk_mq_unfreeze_queue(q
);
3638 * request_queue and elevator_type pair.
3639 * It is just used by __blk_mq_update_nr_hw_queues to cache
3640 * the elevator_type associated with a request_queue.
3642 struct blk_mq_qe_pair
{
3643 struct list_head node
;
3644 struct request_queue
*q
;
3645 struct elevator_type
*type
;
3649 * Cache the elevator_type in qe pair list and switch the
3650 * io scheduler to 'none'
3652 static bool blk_mq_elv_switch_none(struct list_head
*head
,
3653 struct request_queue
*q
)
3655 struct blk_mq_qe_pair
*qe
;
3660 qe
= kmalloc(sizeof(*qe
), GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
);
3664 INIT_LIST_HEAD(&qe
->node
);
3666 qe
->type
= q
->elevator
->type
;
3667 list_add(&qe
->node
, head
);
3669 mutex_lock(&q
->sysfs_lock
);
3671 * After elevator_switch_mq, the previous elevator_queue will be
3672 * released by elevator_release. The reference of the io scheduler
3673 * module get by elevator_get will also be put. So we need to get
3674 * a reference of the io scheduler module here to prevent it to be
3677 __module_get(qe
->type
->elevator_owner
);
3678 elevator_switch_mq(q
, NULL
);
3679 mutex_unlock(&q
->sysfs_lock
);
3684 static void blk_mq_elv_switch_back(struct list_head
*head
,
3685 struct request_queue
*q
)
3687 struct blk_mq_qe_pair
*qe
;
3688 struct elevator_type
*t
= NULL
;
3690 list_for_each_entry(qe
, head
, node
)
3699 list_del(&qe
->node
);
3702 mutex_lock(&q
->sysfs_lock
);
3703 elevator_switch_mq(q
, t
);
3704 mutex_unlock(&q
->sysfs_lock
);
3707 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
,
3710 struct request_queue
*q
;
3712 int prev_nr_hw_queues
;
3714 lockdep_assert_held(&set
->tag_list_lock
);
3716 if (set
->nr_maps
== 1 && nr_hw_queues
> nr_cpu_ids
)
3717 nr_hw_queues
= nr_cpu_ids
;
3718 if (nr_hw_queues
< 1)
3720 if (set
->nr_maps
== 1 && nr_hw_queues
== set
->nr_hw_queues
)
3723 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3724 blk_mq_freeze_queue(q
);
3726 * Switch IO scheduler to 'none', cleaning up the data associated
3727 * with the previous scheduler. We will switch back once we are done
3728 * updating the new sw to hw queue mappings.
3730 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3731 if (!blk_mq_elv_switch_none(&head
, q
))
3734 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3735 blk_mq_debugfs_unregister_hctxs(q
);
3736 blk_mq_sysfs_unregister(q
);
3739 prev_nr_hw_queues
= set
->nr_hw_queues
;
3740 if (blk_mq_realloc_tag_set_tags(set
, set
->nr_hw_queues
, nr_hw_queues
) <
3744 set
->nr_hw_queues
= nr_hw_queues
;
3746 blk_mq_update_queue_map(set
);
3747 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3748 blk_mq_realloc_hw_ctxs(set
, q
);
3749 if (q
->nr_hw_queues
!= set
->nr_hw_queues
) {
3750 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3751 nr_hw_queues
, prev_nr_hw_queues
);
3752 set
->nr_hw_queues
= prev_nr_hw_queues
;
3753 blk_mq_map_queues(&set
->map
[HCTX_TYPE_DEFAULT
]);
3756 blk_mq_map_swqueue(q
);
3760 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3761 blk_mq_sysfs_register(q
);
3762 blk_mq_debugfs_register_hctxs(q
);
3766 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3767 blk_mq_elv_switch_back(&head
, q
);
3769 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3770 blk_mq_unfreeze_queue(q
);
3773 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
3775 mutex_lock(&set
->tag_list_lock
);
3776 __blk_mq_update_nr_hw_queues(set
, nr_hw_queues
);
3777 mutex_unlock(&set
->tag_list_lock
);
3779 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
3781 /* Enable polling stats and return whether they were already enabled. */
3782 static bool blk_poll_stats_enable(struct request_queue
*q
)
3784 if (test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
3785 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS
, q
))
3787 blk_stat_add_callback(q
, q
->poll_cb
);
3791 static void blk_mq_poll_stats_start(struct request_queue
*q
)
3794 * We don't arm the callback if polling stats are not enabled or the
3795 * callback is already active.
3797 if (!test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
3798 blk_stat_is_active(q
->poll_cb
))
3801 blk_stat_activate_msecs(q
->poll_cb
, 100);
3804 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
)
3806 struct request_queue
*q
= cb
->data
;
3809 for (bucket
= 0; bucket
< BLK_MQ_POLL_STATS_BKTS
; bucket
++) {
3810 if (cb
->stat
[bucket
].nr_samples
)
3811 q
->poll_stat
[bucket
] = cb
->stat
[bucket
];
3815 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
3818 unsigned long ret
= 0;
3822 * If stats collection isn't on, don't sleep but turn it on for
3825 if (!blk_poll_stats_enable(q
))
3829 * As an optimistic guess, use half of the mean service time
3830 * for this type of request. We can (and should) make this smarter.
3831 * For instance, if the completion latencies are tight, we can
3832 * get closer than just half the mean. This is especially
3833 * important on devices where the completion latencies are longer
3834 * than ~10 usec. We do use the stats for the relevant IO size
3835 * if available which does lead to better estimates.
3837 bucket
= blk_mq_poll_stats_bkt(rq
);
3841 if (q
->poll_stat
[bucket
].nr_samples
)
3842 ret
= (q
->poll_stat
[bucket
].mean
+ 1) / 2;
3847 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
3850 struct hrtimer_sleeper hs
;
3851 enum hrtimer_mode mode
;
3855 if (rq
->rq_flags
& RQF_MQ_POLL_SLEPT
)
3859 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3861 * 0: use half of prev avg
3862 * >0: use this specific value
3864 if (q
->poll_nsec
> 0)
3865 nsecs
= q
->poll_nsec
;
3867 nsecs
= blk_mq_poll_nsecs(q
, rq
);
3872 rq
->rq_flags
|= RQF_MQ_POLL_SLEPT
;
3875 * This will be replaced with the stats tracking code, using
3876 * 'avg_completion_time / 2' as the pre-sleep target.
3880 mode
= HRTIMER_MODE_REL
;
3881 hrtimer_init_sleeper_on_stack(&hs
, CLOCK_MONOTONIC
, mode
);
3882 hrtimer_set_expires(&hs
.timer
, kt
);
3885 if (blk_mq_rq_state(rq
) == MQ_RQ_COMPLETE
)
3887 set_current_state(TASK_UNINTERRUPTIBLE
);
3888 hrtimer_sleeper_start_expires(&hs
, mode
);
3891 hrtimer_cancel(&hs
.timer
);
3892 mode
= HRTIMER_MODE_ABS
;
3893 } while (hs
.task
&& !signal_pending(current
));
3895 __set_current_state(TASK_RUNNING
);
3896 destroy_hrtimer_on_stack(&hs
.timer
);
3900 static bool blk_mq_poll_hybrid(struct request_queue
*q
,
3901 struct blk_mq_hw_ctx
*hctx
, blk_qc_t cookie
)
3905 if (q
->poll_nsec
== BLK_MQ_POLL_CLASSIC
)
3908 if (!blk_qc_t_is_internal(cookie
))
3909 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
3911 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
3913 * With scheduling, if the request has completed, we'll
3914 * get a NULL return here, as we clear the sched tag when
3915 * that happens. The request still remains valid, like always,
3916 * so we should be safe with just the NULL check.
3922 return blk_mq_poll_hybrid_sleep(q
, rq
);
3926 * blk_poll - poll for IO completions
3928 * @cookie: cookie passed back at IO submission time
3929 * @spin: whether to spin for completions
3932 * Poll for completions on the passed in queue. Returns number of
3933 * completed entries found. If @spin is true, then blk_poll will continue
3934 * looping until at least one completion is found, unless the task is
3935 * otherwise marked running (or we need to reschedule).
3937 int blk_poll(struct request_queue
*q
, blk_qc_t cookie
, bool spin
)
3939 struct blk_mq_hw_ctx
*hctx
;
3942 if (!blk_qc_t_valid(cookie
) ||
3943 !test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
3947 blk_flush_plug_list(current
->plug
, false);
3949 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
3952 * If we sleep, have the caller restart the poll loop to reset
3953 * the state. Like for the other success return cases, the
3954 * caller is responsible for checking if the IO completed. If
3955 * the IO isn't complete, we'll get called again and will go
3956 * straight to the busy poll loop. If specified not to spin,
3957 * we also should not sleep.
3959 if (spin
&& blk_mq_poll_hybrid(q
, hctx
, cookie
))
3962 hctx
->poll_considered
++;
3964 state
= get_current_state();
3968 hctx
->poll_invoked
++;
3970 ret
= q
->mq_ops
->poll(hctx
);
3972 hctx
->poll_success
++;
3973 __set_current_state(TASK_RUNNING
);
3977 if (signal_pending_state(state
, current
))
3978 __set_current_state(TASK_RUNNING
);
3980 if (task_is_running(current
))
3982 if (ret
< 0 || !spin
)
3985 } while (!need_resched());
3987 __set_current_state(TASK_RUNNING
);
3990 EXPORT_SYMBOL_GPL(blk_poll
);
3992 unsigned int blk_mq_rq_cpu(struct request
*rq
)
3994 return rq
->mq_ctx
->cpu
;
3996 EXPORT_SYMBOL(blk_mq_rq_cpu
);
3998 static int __init
blk_mq_init(void)
4002 for_each_possible_cpu(i
)
4003 init_llist_head(&per_cpu(blk_cpu_done
, i
));
4004 open_softirq(BLOCK_SOFTIRQ
, blk_done_softirq
);
4006 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD
,
4007 "block/softirq:dead", NULL
,
4008 blk_softirq_cpu_dead
);
4009 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
4010 blk_mq_hctx_notify_dead
);
4011 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE
, "block/mq:online",
4012 blk_mq_hctx_notify_online
,
4013 blk_mq_hctx_notify_offline
);
4016 subsys_initcall(blk_mq_init
);