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
, bool force_atomic
)
193 mutex_lock(&q
->mq_freeze_lock
);
195 q
->q_usage_counter
.data
->force_atomic
= true;
196 q
->mq_freeze_depth
--;
197 WARN_ON_ONCE(q
->mq_freeze_depth
< 0);
198 if (!q
->mq_freeze_depth
) {
199 percpu_ref_resurrect(&q
->q_usage_counter
);
200 wake_up_all(&q
->mq_freeze_wq
);
202 mutex_unlock(&q
->mq_freeze_lock
);
205 void blk_mq_unfreeze_queue(struct request_queue
*q
)
207 __blk_mq_unfreeze_queue(q
, false);
209 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
212 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
213 * mpt3sas driver such that this function can be removed.
215 void blk_mq_quiesce_queue_nowait(struct request_queue
*q
)
217 blk_queue_flag_set(QUEUE_FLAG_QUIESCED
, q
);
219 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait
);
222 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
225 * Note: this function does not prevent that the struct request end_io()
226 * callback function is invoked. Once this function is returned, we make
227 * sure no dispatch can happen until the queue is unquiesced via
228 * blk_mq_unquiesce_queue().
230 void blk_mq_quiesce_queue(struct request_queue
*q
)
232 struct blk_mq_hw_ctx
*hctx
;
236 blk_mq_quiesce_queue_nowait(q
);
238 queue_for_each_hw_ctx(q
, hctx
, i
) {
239 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
240 synchronize_srcu(hctx
->srcu
);
247 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
250 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
253 * This function recovers queue into the state before quiescing
254 * which is done by blk_mq_quiesce_queue.
256 void blk_mq_unquiesce_queue(struct request_queue
*q
)
258 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED
, q
);
260 /* dispatch requests which are inserted during quiescing */
261 blk_mq_run_hw_queues(q
, true);
263 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue
);
265 void blk_mq_wake_waiters(struct request_queue
*q
)
267 struct blk_mq_hw_ctx
*hctx
;
270 queue_for_each_hw_ctx(q
, hctx
, i
)
271 if (blk_mq_hw_queue_mapped(hctx
))
272 blk_mq_tag_wakeup_all(hctx
->tags
, true);
276 * Only need start/end time stamping if we have iostat or
277 * blk stats enabled, or using an IO scheduler.
279 static inline bool blk_mq_need_time_stamp(struct request
*rq
)
281 return (rq
->rq_flags
& (RQF_IO_STAT
| RQF_STATS
)) || rq
->q
->elevator
;
284 static struct request
*blk_mq_rq_ctx_init(struct blk_mq_alloc_data
*data
,
285 unsigned int tag
, u64 alloc_time_ns
)
287 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
288 struct request
*rq
= tags
->static_rqs
[tag
];
290 if (data
->q
->elevator
) {
291 rq
->tag
= BLK_MQ_NO_TAG
;
292 rq
->internal_tag
= tag
;
295 rq
->internal_tag
= BLK_MQ_NO_TAG
;
298 /* csd/requeue_work/fifo_time is initialized before use */
300 rq
->mq_ctx
= data
->ctx
;
301 rq
->mq_hctx
= data
->hctx
;
303 rq
->cmd_flags
= data
->cmd_flags
;
304 if (data
->flags
& BLK_MQ_REQ_PM
)
305 rq
->rq_flags
|= RQF_PM
;
306 if (blk_queue_io_stat(data
->q
))
307 rq
->rq_flags
|= RQF_IO_STAT
;
308 INIT_LIST_HEAD(&rq
->queuelist
);
309 INIT_HLIST_NODE(&rq
->hash
);
310 RB_CLEAR_NODE(&rq
->rb_node
);
313 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
314 rq
->alloc_time_ns
= alloc_time_ns
;
316 if (blk_mq_need_time_stamp(rq
))
317 rq
->start_time_ns
= ktime_get_ns();
319 rq
->start_time_ns
= 0;
320 rq
->io_start_time_ns
= 0;
321 rq
->stats_sectors
= 0;
322 rq
->nr_phys_segments
= 0;
323 #if defined(CONFIG_BLK_DEV_INTEGRITY)
324 rq
->nr_integrity_segments
= 0;
326 blk_crypto_rq_set_defaults(rq
);
327 /* tag was already set */
328 WRITE_ONCE(rq
->deadline
, 0);
333 rq
->end_io_data
= NULL
;
335 data
->ctx
->rq_dispatched
[op_is_sync(data
->cmd_flags
)]++;
336 refcount_set(&rq
->ref
, 1);
338 if (!op_is_flush(data
->cmd_flags
)) {
339 struct elevator_queue
*e
= data
->q
->elevator
;
342 if (e
&& e
->type
->ops
.prepare_request
) {
343 if (e
->type
->icq_cache
)
344 blk_mq_sched_assign_ioc(rq
);
346 e
->type
->ops
.prepare_request(rq
);
347 rq
->rq_flags
|= RQF_ELVPRIV
;
351 data
->hctx
->queued
++;
355 static struct request
*__blk_mq_alloc_request(struct blk_mq_alloc_data
*data
)
357 struct request_queue
*q
= data
->q
;
358 struct elevator_queue
*e
= q
->elevator
;
359 u64 alloc_time_ns
= 0;
362 /* alloc_time includes depth and tag waits */
363 if (blk_queue_rq_alloc_time(q
))
364 alloc_time_ns
= ktime_get_ns();
366 if (data
->cmd_flags
& REQ_NOWAIT
)
367 data
->flags
|= BLK_MQ_REQ_NOWAIT
;
371 * Flush/passthrough requests are special and go directly to the
372 * dispatch list. Don't include reserved tags in the
373 * limiting, as it isn't useful.
375 if (!op_is_flush(data
->cmd_flags
) &&
376 !blk_op_is_passthrough(data
->cmd_flags
) &&
377 e
->type
->ops
.limit_depth
&&
378 !(data
->flags
& BLK_MQ_REQ_RESERVED
))
379 e
->type
->ops
.limit_depth(data
->cmd_flags
, data
);
383 data
->ctx
= blk_mq_get_ctx(q
);
384 data
->hctx
= blk_mq_map_queue(q
, data
->cmd_flags
, data
->ctx
);
386 blk_mq_tag_busy(data
->hctx
);
389 * Waiting allocations only fail because of an inactive hctx. In that
390 * case just retry the hctx assignment and tag allocation as CPU hotplug
391 * should have migrated us to an online CPU by now.
393 tag
= blk_mq_get_tag(data
);
394 if (tag
== BLK_MQ_NO_TAG
) {
395 if (data
->flags
& BLK_MQ_REQ_NOWAIT
)
399 * Give up the CPU and sleep for a random short time to ensure
400 * that thread using a realtime scheduling class are migrated
401 * off the CPU, and thus off the hctx that is going away.
406 return blk_mq_rq_ctx_init(data
, tag
, alloc_time_ns
);
409 struct request
*blk_mq_alloc_request(struct request_queue
*q
, unsigned int op
,
410 blk_mq_req_flags_t flags
)
412 struct blk_mq_alloc_data data
= {
420 ret
= blk_queue_enter(q
, flags
);
424 rq
= __blk_mq_alloc_request(&data
);
428 rq
->__sector
= (sector_t
) -1;
429 rq
->bio
= rq
->biotail
= NULL
;
433 return ERR_PTR(-EWOULDBLOCK
);
435 EXPORT_SYMBOL(blk_mq_alloc_request
);
437 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
,
438 unsigned int op
, blk_mq_req_flags_t flags
, unsigned int hctx_idx
)
440 struct blk_mq_alloc_data data
= {
445 u64 alloc_time_ns
= 0;
450 /* alloc_time includes depth and tag waits */
451 if (blk_queue_rq_alloc_time(q
))
452 alloc_time_ns
= ktime_get_ns();
455 * If the tag allocator sleeps we could get an allocation for a
456 * different hardware context. No need to complicate the low level
457 * allocator for this for the rare use case of a command tied to
460 if (WARN_ON_ONCE(!(flags
& (BLK_MQ_REQ_NOWAIT
| BLK_MQ_REQ_RESERVED
))))
461 return ERR_PTR(-EINVAL
);
463 if (hctx_idx
>= q
->nr_hw_queues
)
464 return ERR_PTR(-EIO
);
466 ret
= blk_queue_enter(q
, flags
);
471 * Check if the hardware context is actually mapped to anything.
472 * If not tell the caller that it should skip this queue.
475 data
.hctx
= q
->queue_hw_ctx
[hctx_idx
];
476 if (!blk_mq_hw_queue_mapped(data
.hctx
))
478 cpu
= cpumask_first_and(data
.hctx
->cpumask
, cpu_online_mask
);
479 data
.ctx
= __blk_mq_get_ctx(q
, cpu
);
482 blk_mq_tag_busy(data
.hctx
);
485 tag
= blk_mq_get_tag(&data
);
486 if (tag
== BLK_MQ_NO_TAG
)
488 return blk_mq_rq_ctx_init(&data
, tag
, alloc_time_ns
);
494 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
496 static void __blk_mq_free_request(struct request
*rq
)
498 struct request_queue
*q
= rq
->q
;
499 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
500 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
501 const int sched_tag
= rq
->internal_tag
;
503 blk_crypto_free_request(rq
);
504 blk_pm_mark_last_busy(rq
);
506 if (rq
->tag
!= BLK_MQ_NO_TAG
)
507 blk_mq_put_tag(hctx
->tags
, ctx
, rq
->tag
);
508 if (sched_tag
!= BLK_MQ_NO_TAG
)
509 blk_mq_put_tag(hctx
->sched_tags
, ctx
, sched_tag
);
510 blk_mq_sched_restart(hctx
);
514 void blk_mq_free_request(struct request
*rq
)
516 struct request_queue
*q
= rq
->q
;
517 struct elevator_queue
*e
= q
->elevator
;
518 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
519 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
521 if (rq
->rq_flags
& RQF_ELVPRIV
) {
522 if (e
&& e
->type
->ops
.finish_request
)
523 e
->type
->ops
.finish_request(rq
);
525 put_io_context(rq
->elv
.icq
->ioc
);
530 ctx
->rq_completed
[rq_is_sync(rq
)]++;
531 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
532 __blk_mq_dec_active_requests(hctx
);
534 if (unlikely(laptop_mode
&& !blk_rq_is_passthrough(rq
)))
535 laptop_io_completion(q
->disk
->bdi
);
539 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
540 if (refcount_dec_and_test(&rq
->ref
))
541 __blk_mq_free_request(rq
);
543 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
545 inline void __blk_mq_end_request(struct request
*rq
, blk_status_t error
)
549 if (blk_mq_need_time_stamp(rq
))
550 now
= ktime_get_ns();
552 if (rq
->rq_flags
& RQF_STATS
) {
553 blk_mq_poll_stats_start(rq
->q
);
554 blk_stat_add(rq
, now
);
557 blk_mq_sched_completed_request(rq
, now
);
559 blk_account_io_done(rq
, now
);
562 rq_qos_done(rq
->q
, rq
);
563 rq
->end_io(rq
, error
);
565 blk_mq_free_request(rq
);
568 EXPORT_SYMBOL(__blk_mq_end_request
);
570 void blk_mq_end_request(struct request
*rq
, blk_status_t error
)
572 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
574 __blk_mq_end_request(rq
, error
);
576 EXPORT_SYMBOL(blk_mq_end_request
);
578 static void blk_complete_reqs(struct llist_head
*list
)
580 struct llist_node
*entry
= llist_reverse_order(llist_del_all(list
));
581 struct request
*rq
, *next
;
583 llist_for_each_entry_safe(rq
, next
, entry
, ipi_list
)
584 rq
->q
->mq_ops
->complete(rq
);
587 static __latent_entropy
void blk_done_softirq(struct softirq_action
*h
)
589 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done
));
592 static int blk_softirq_cpu_dead(unsigned int cpu
)
594 blk_complete_reqs(&per_cpu(blk_cpu_done
, cpu
));
598 static void __blk_mq_complete_request_remote(void *data
)
600 __raise_softirq_irqoff(BLOCK_SOFTIRQ
);
603 static inline bool blk_mq_complete_need_ipi(struct request
*rq
)
605 int cpu
= raw_smp_processor_id();
607 if (!IS_ENABLED(CONFIG_SMP
) ||
608 !test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
))
611 * With force threaded interrupts enabled, raising softirq from an SMP
612 * function call will always result in waking the ksoftirqd thread.
613 * This is probably worse than completing the request on a different
616 if (force_irqthreads())
619 /* same CPU or cache domain? Complete locally */
620 if (cpu
== rq
->mq_ctx
->cpu
||
621 (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
) &&
622 cpus_share_cache(cpu
, rq
->mq_ctx
->cpu
)))
625 /* don't try to IPI to an offline CPU */
626 return cpu_online(rq
->mq_ctx
->cpu
);
629 static void blk_mq_complete_send_ipi(struct request
*rq
)
631 struct llist_head
*list
;
634 cpu
= rq
->mq_ctx
->cpu
;
635 list
= &per_cpu(blk_cpu_done
, cpu
);
636 if (llist_add(&rq
->ipi_list
, list
)) {
637 INIT_CSD(&rq
->csd
, __blk_mq_complete_request_remote
, rq
);
638 smp_call_function_single_async(cpu
, &rq
->csd
);
642 static void blk_mq_raise_softirq(struct request
*rq
)
644 struct llist_head
*list
;
647 list
= this_cpu_ptr(&blk_cpu_done
);
648 if (llist_add(&rq
->ipi_list
, list
))
649 raise_softirq(BLOCK_SOFTIRQ
);
653 bool blk_mq_complete_request_remote(struct request
*rq
)
655 WRITE_ONCE(rq
->state
, MQ_RQ_COMPLETE
);
658 * For a polled request, always complete locallly, it's pointless
659 * to redirect the completion.
661 if (rq
->cmd_flags
& REQ_HIPRI
)
664 if (blk_mq_complete_need_ipi(rq
)) {
665 blk_mq_complete_send_ipi(rq
);
669 if (rq
->q
->nr_hw_queues
== 1) {
670 blk_mq_raise_softirq(rq
);
675 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote
);
678 * blk_mq_complete_request - end I/O on a request
679 * @rq: the request being processed
682 * Complete a request by scheduling the ->complete_rq operation.
684 void blk_mq_complete_request(struct request
*rq
)
686 if (!blk_mq_complete_request_remote(rq
))
687 rq
->q
->mq_ops
->complete(rq
);
689 EXPORT_SYMBOL(blk_mq_complete_request
);
691 static void hctx_unlock(struct blk_mq_hw_ctx
*hctx
, int srcu_idx
)
692 __releases(hctx
->srcu
)
694 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
))
697 srcu_read_unlock(hctx
->srcu
, srcu_idx
);
700 static void hctx_lock(struct blk_mq_hw_ctx
*hctx
, int *srcu_idx
)
701 __acquires(hctx
->srcu
)
703 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
704 /* shut up gcc false positive */
708 *srcu_idx
= srcu_read_lock(hctx
->srcu
);
712 * blk_mq_start_request - Start processing a request
713 * @rq: Pointer to request to be started
715 * Function used by device drivers to notify the block layer that a request
716 * is going to be processed now, so blk layer can do proper initializations
717 * such as starting the timeout timer.
719 void blk_mq_start_request(struct request
*rq
)
721 struct request_queue
*q
= rq
->q
;
723 trace_block_rq_issue(rq
);
725 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
726 rq
->io_start_time_ns
= ktime_get_ns();
727 rq
->stats_sectors
= blk_rq_sectors(rq
);
728 rq
->rq_flags
|= RQF_STATS
;
732 WARN_ON_ONCE(blk_mq_rq_state(rq
) != MQ_RQ_IDLE
);
735 WRITE_ONCE(rq
->state
, MQ_RQ_IN_FLIGHT
);
737 #ifdef CONFIG_BLK_DEV_INTEGRITY
738 if (blk_integrity_rq(rq
) && req_op(rq
) == REQ_OP_WRITE
)
739 q
->integrity
.profile
->prepare_fn(rq
);
742 EXPORT_SYMBOL(blk_mq_start_request
);
744 static void __blk_mq_requeue_request(struct request
*rq
)
746 struct request_queue
*q
= rq
->q
;
748 blk_mq_put_driver_tag(rq
);
750 trace_block_rq_requeue(rq
);
751 rq_qos_requeue(q
, rq
);
753 if (blk_mq_request_started(rq
)) {
754 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
755 rq
->rq_flags
&= ~RQF_TIMED_OUT
;
759 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
761 __blk_mq_requeue_request(rq
);
763 /* this request will be re-inserted to io scheduler queue */
764 blk_mq_sched_requeue_request(rq
);
766 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
768 EXPORT_SYMBOL(blk_mq_requeue_request
);
770 static void blk_mq_requeue_work(struct work_struct
*work
)
772 struct request_queue
*q
=
773 container_of(work
, struct request_queue
, requeue_work
.work
);
775 struct request
*rq
, *next
;
777 spin_lock_irq(&q
->requeue_lock
);
778 list_splice_init(&q
->requeue_list
, &rq_list
);
779 spin_unlock_irq(&q
->requeue_lock
);
781 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
782 if (!(rq
->rq_flags
& (RQF_SOFTBARRIER
| RQF_DONTPREP
)))
785 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
786 list_del_init(&rq
->queuelist
);
788 * If RQF_DONTPREP, rq has contained some driver specific
789 * data, so insert it to hctx dispatch list to avoid any
792 if (rq
->rq_flags
& RQF_DONTPREP
)
793 blk_mq_request_bypass_insert(rq
, false, false);
795 blk_mq_sched_insert_request(rq
, true, false, false);
798 while (!list_empty(&rq_list
)) {
799 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
800 list_del_init(&rq
->queuelist
);
801 blk_mq_sched_insert_request(rq
, false, false, false);
804 blk_mq_run_hw_queues(q
, false);
807 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
808 bool kick_requeue_list
)
810 struct request_queue
*q
= rq
->q
;
814 * We abuse this flag that is otherwise used by the I/O scheduler to
815 * request head insertion from the workqueue.
817 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
819 spin_lock_irqsave(&q
->requeue_lock
, flags
);
821 rq
->rq_flags
|= RQF_SOFTBARRIER
;
822 list_add(&rq
->queuelist
, &q
->requeue_list
);
824 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
826 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
828 if (kick_requeue_list
)
829 blk_mq_kick_requeue_list(q
);
832 void blk_mq_kick_requeue_list(struct request_queue
*q
)
834 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
, 0);
836 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
838 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
841 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
,
842 msecs_to_jiffies(msecs
));
844 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
846 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
848 if (tag
< tags
->nr_tags
) {
849 prefetch(tags
->rqs
[tag
]);
850 return tags
->rqs
[tag
];
855 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
857 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
858 void *priv
, bool reserved
)
861 * If we find a request that isn't idle and the queue matches,
862 * we know the queue is busy. Return false to stop the iteration.
864 if (blk_mq_request_started(rq
) && rq
->q
== hctx
->queue
) {
874 bool blk_mq_queue_inflight(struct request_queue
*q
)
878 blk_mq_queue_tag_busy_iter(q
, blk_mq_rq_inflight
, &busy
);
881 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight
);
883 static void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
885 req
->rq_flags
|= RQF_TIMED_OUT
;
886 if (req
->q
->mq_ops
->timeout
) {
887 enum blk_eh_timer_return ret
;
889 ret
= req
->q
->mq_ops
->timeout(req
, reserved
);
890 if (ret
== BLK_EH_DONE
)
892 WARN_ON_ONCE(ret
!= BLK_EH_RESET_TIMER
);
898 static bool blk_mq_req_expired(struct request
*rq
, unsigned long *next
)
900 unsigned long deadline
;
902 if (blk_mq_rq_state(rq
) != MQ_RQ_IN_FLIGHT
)
904 if (rq
->rq_flags
& RQF_TIMED_OUT
)
907 deadline
= READ_ONCE(rq
->deadline
);
908 if (time_after_eq(jiffies
, deadline
))
913 else if (time_after(*next
, deadline
))
918 void blk_mq_put_rq_ref(struct request
*rq
)
922 else if (refcount_dec_and_test(&rq
->ref
))
923 __blk_mq_free_request(rq
);
926 static bool blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
927 struct request
*rq
, void *priv
, bool reserved
)
929 unsigned long *next
= priv
;
932 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
933 * be reallocated underneath the timeout handler's processing, then
934 * the expire check is reliable. If the request is not expired, then
935 * it was completed and reallocated as a new request after returning
936 * from blk_mq_check_expired().
938 if (blk_mq_req_expired(rq
, next
))
939 blk_mq_rq_timed_out(rq
, reserved
);
943 static void blk_mq_timeout_work(struct work_struct
*work
)
945 struct request_queue
*q
=
946 container_of(work
, struct request_queue
, timeout_work
);
947 unsigned long next
= 0;
948 struct blk_mq_hw_ctx
*hctx
;
951 /* A deadlock might occur if a request is stuck requiring a
952 * timeout at the same time a queue freeze is waiting
953 * completion, since the timeout code would not be able to
954 * acquire the queue reference here.
956 * That's why we don't use blk_queue_enter here; instead, we use
957 * percpu_ref_tryget directly, because we need to be able to
958 * obtain a reference even in the short window between the queue
959 * starting to freeze, by dropping the first reference in
960 * blk_freeze_queue_start, and the moment the last request is
961 * consumed, marked by the instant q_usage_counter reaches
964 if (!percpu_ref_tryget(&q
->q_usage_counter
))
967 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &next
);
970 mod_timer(&q
->timeout
, next
);
973 * Request timeouts are handled as a forward rolling timer. If
974 * we end up here it means that no requests are pending and
975 * also that no request has been pending for a while. Mark
978 queue_for_each_hw_ctx(q
, hctx
, i
) {
979 /* the hctx may be unmapped, so check it here */
980 if (blk_mq_hw_queue_mapped(hctx
))
981 blk_mq_tag_idle(hctx
);
987 struct flush_busy_ctx_data
{
988 struct blk_mq_hw_ctx
*hctx
;
989 struct list_head
*list
;
992 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
994 struct flush_busy_ctx_data
*flush_data
= data
;
995 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
996 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
997 enum hctx_type type
= hctx
->type
;
999 spin_lock(&ctx
->lock
);
1000 list_splice_tail_init(&ctx
->rq_lists
[type
], flush_data
->list
);
1001 sbitmap_clear_bit(sb
, bitnr
);
1002 spin_unlock(&ctx
->lock
);
1007 * Process software queues that have been marked busy, splicing them
1008 * to the for-dispatch
1010 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
1012 struct flush_busy_ctx_data data
= {
1017 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
1019 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
1021 struct dispatch_rq_data
{
1022 struct blk_mq_hw_ctx
*hctx
;
1026 static bool dispatch_rq_from_ctx(struct sbitmap
*sb
, unsigned int bitnr
,
1029 struct dispatch_rq_data
*dispatch_data
= data
;
1030 struct blk_mq_hw_ctx
*hctx
= dispatch_data
->hctx
;
1031 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
1032 enum hctx_type type
= hctx
->type
;
1034 spin_lock(&ctx
->lock
);
1035 if (!list_empty(&ctx
->rq_lists
[type
])) {
1036 dispatch_data
->rq
= list_entry_rq(ctx
->rq_lists
[type
].next
);
1037 list_del_init(&dispatch_data
->rq
->queuelist
);
1038 if (list_empty(&ctx
->rq_lists
[type
]))
1039 sbitmap_clear_bit(sb
, bitnr
);
1041 spin_unlock(&ctx
->lock
);
1043 return !dispatch_data
->rq
;
1046 struct request
*blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx
*hctx
,
1047 struct blk_mq_ctx
*start
)
1049 unsigned off
= start
? start
->index_hw
[hctx
->type
] : 0;
1050 struct dispatch_rq_data data
= {
1055 __sbitmap_for_each_set(&hctx
->ctx_map
, off
,
1056 dispatch_rq_from_ctx
, &data
);
1061 static inline unsigned int queued_to_index(unsigned int queued
)
1066 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
1069 static bool __blk_mq_get_driver_tag(struct request
*rq
)
1071 struct sbitmap_queue
*bt
= rq
->mq_hctx
->tags
->bitmap_tags
;
1072 unsigned int tag_offset
= rq
->mq_hctx
->tags
->nr_reserved_tags
;
1075 blk_mq_tag_busy(rq
->mq_hctx
);
1077 if (blk_mq_tag_is_reserved(rq
->mq_hctx
->sched_tags
, rq
->internal_tag
)) {
1078 bt
= rq
->mq_hctx
->tags
->breserved_tags
;
1081 if (!hctx_may_queue(rq
->mq_hctx
, bt
))
1085 tag
= __sbitmap_queue_get(bt
);
1086 if (tag
== BLK_MQ_NO_TAG
)
1089 rq
->tag
= tag
+ tag_offset
;
1093 bool blk_mq_get_driver_tag(struct request
*rq
)
1095 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1097 if (rq
->tag
== BLK_MQ_NO_TAG
&& !__blk_mq_get_driver_tag(rq
))
1100 if ((hctx
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
) &&
1101 !(rq
->rq_flags
& RQF_MQ_INFLIGHT
)) {
1102 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
1103 __blk_mq_inc_active_requests(hctx
);
1105 hctx
->tags
->rqs
[rq
->tag
] = rq
;
1109 static int blk_mq_dispatch_wake(wait_queue_entry_t
*wait
, unsigned mode
,
1110 int flags
, void *key
)
1112 struct blk_mq_hw_ctx
*hctx
;
1114 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
1116 spin_lock(&hctx
->dispatch_wait_lock
);
1117 if (!list_empty(&wait
->entry
)) {
1118 struct sbitmap_queue
*sbq
;
1120 list_del_init(&wait
->entry
);
1121 sbq
= hctx
->tags
->bitmap_tags
;
1122 atomic_dec(&sbq
->ws_active
);
1124 spin_unlock(&hctx
->dispatch_wait_lock
);
1126 blk_mq_run_hw_queue(hctx
, true);
1131 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1132 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1133 * restart. For both cases, take care to check the condition again after
1134 * marking us as waiting.
1136 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx
*hctx
,
1139 struct sbitmap_queue
*sbq
= hctx
->tags
->bitmap_tags
;
1140 struct wait_queue_head
*wq
;
1141 wait_queue_entry_t
*wait
;
1144 if (!(hctx
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
)) {
1145 blk_mq_sched_mark_restart_hctx(hctx
);
1148 * It's possible that a tag was freed in the window between the
1149 * allocation failure and adding the hardware queue to the wait
1152 * Don't clear RESTART here, someone else could have set it.
1153 * At most this will cost an extra queue run.
1155 return blk_mq_get_driver_tag(rq
);
1158 wait
= &hctx
->dispatch_wait
;
1159 if (!list_empty_careful(&wait
->entry
))
1162 wq
= &bt_wait_ptr(sbq
, hctx
)->wait
;
1164 spin_lock_irq(&wq
->lock
);
1165 spin_lock(&hctx
->dispatch_wait_lock
);
1166 if (!list_empty(&wait
->entry
)) {
1167 spin_unlock(&hctx
->dispatch_wait_lock
);
1168 spin_unlock_irq(&wq
->lock
);
1172 atomic_inc(&sbq
->ws_active
);
1173 wait
->flags
&= ~WQ_FLAG_EXCLUSIVE
;
1174 __add_wait_queue(wq
, wait
);
1177 * It's possible that a tag was freed in the window between the
1178 * allocation failure and adding the hardware queue to the wait
1181 ret
= blk_mq_get_driver_tag(rq
);
1183 spin_unlock(&hctx
->dispatch_wait_lock
);
1184 spin_unlock_irq(&wq
->lock
);
1189 * We got a tag, remove ourselves from the wait queue to ensure
1190 * someone else gets the wakeup.
1192 list_del_init(&wait
->entry
);
1193 atomic_dec(&sbq
->ws_active
);
1194 spin_unlock(&hctx
->dispatch_wait_lock
);
1195 spin_unlock_irq(&wq
->lock
);
1200 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1201 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1203 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1204 * - EWMA is one simple way to compute running average value
1205 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1206 * - take 4 as factor for avoiding to get too small(0) result, and this
1207 * factor doesn't matter because EWMA decreases exponentially
1209 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx
*hctx
, bool busy
)
1213 ewma
= hctx
->dispatch_busy
;
1218 ewma
*= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
- 1;
1220 ewma
+= 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR
;
1221 ewma
/= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
;
1223 hctx
->dispatch_busy
= ewma
;
1226 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1228 static void blk_mq_handle_dev_resource(struct request
*rq
,
1229 struct list_head
*list
)
1231 struct request
*next
=
1232 list_first_entry_or_null(list
, struct request
, queuelist
);
1235 * If an I/O scheduler has been configured and we got a driver tag for
1236 * the next request already, free it.
1239 blk_mq_put_driver_tag(next
);
1241 list_add(&rq
->queuelist
, list
);
1242 __blk_mq_requeue_request(rq
);
1245 static void blk_mq_handle_zone_resource(struct request
*rq
,
1246 struct list_head
*zone_list
)
1249 * If we end up here it is because we cannot dispatch a request to a
1250 * specific zone due to LLD level zone-write locking or other zone
1251 * related resource not being available. In this case, set the request
1252 * aside in zone_list for retrying it later.
1254 list_add(&rq
->queuelist
, zone_list
);
1255 __blk_mq_requeue_request(rq
);
1258 enum prep_dispatch
{
1260 PREP_DISPATCH_NO_TAG
,
1261 PREP_DISPATCH_NO_BUDGET
,
1264 static enum prep_dispatch
blk_mq_prep_dispatch_rq(struct request
*rq
,
1267 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1268 int budget_token
= -1;
1271 budget_token
= blk_mq_get_dispatch_budget(rq
->q
);
1272 if (budget_token
< 0) {
1273 blk_mq_put_driver_tag(rq
);
1274 return PREP_DISPATCH_NO_BUDGET
;
1276 blk_mq_set_rq_budget_token(rq
, budget_token
);
1279 if (!blk_mq_get_driver_tag(rq
)) {
1281 * The initial allocation attempt failed, so we need to
1282 * rerun the hardware queue when a tag is freed. The
1283 * waitqueue takes care of that. If the queue is run
1284 * before we add this entry back on the dispatch list,
1285 * we'll re-run it below.
1287 if (!blk_mq_mark_tag_wait(hctx
, rq
)) {
1289 * All budgets not got from this function will be put
1290 * together during handling partial dispatch
1293 blk_mq_put_dispatch_budget(rq
->q
, budget_token
);
1294 return PREP_DISPATCH_NO_TAG
;
1298 return PREP_DISPATCH_OK
;
1301 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1302 static void blk_mq_release_budgets(struct request_queue
*q
,
1303 struct list_head
*list
)
1307 list_for_each_entry(rq
, list
, queuelist
) {
1308 int budget_token
= blk_mq_get_rq_budget_token(rq
);
1310 if (budget_token
>= 0)
1311 blk_mq_put_dispatch_budget(q
, budget_token
);
1316 * Returns true if we did some work AND can potentially do more.
1318 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
,
1319 unsigned int nr_budgets
)
1321 enum prep_dispatch prep
;
1322 struct request_queue
*q
= hctx
->queue
;
1323 struct request
*rq
, *nxt
;
1325 blk_status_t ret
= BLK_STS_OK
;
1326 LIST_HEAD(zone_list
);
1327 bool needs_resource
= false;
1329 if (list_empty(list
))
1333 * Now process all the entries, sending them to the driver.
1335 errors
= queued
= 0;
1337 struct blk_mq_queue_data bd
;
1339 rq
= list_first_entry(list
, struct request
, queuelist
);
1341 WARN_ON_ONCE(hctx
!= rq
->mq_hctx
);
1342 prep
= blk_mq_prep_dispatch_rq(rq
, !nr_budgets
);
1343 if (prep
!= PREP_DISPATCH_OK
)
1346 list_del_init(&rq
->queuelist
);
1351 * Flag last if we have no more requests, or if we have more
1352 * but can't assign a driver tag to it.
1354 if (list_empty(list
))
1357 nxt
= list_first_entry(list
, struct request
, queuelist
);
1358 bd
.last
= !blk_mq_get_driver_tag(nxt
);
1362 * once the request is queued to lld, no need to cover the
1367 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1372 case BLK_STS_RESOURCE
:
1373 needs_resource
= true;
1375 case BLK_STS_DEV_RESOURCE
:
1376 blk_mq_handle_dev_resource(rq
, list
);
1378 case BLK_STS_ZONE_RESOURCE
:
1380 * Move the request to zone_list and keep going through
1381 * the dispatch list to find more requests the drive can
1384 blk_mq_handle_zone_resource(rq
, &zone_list
);
1385 needs_resource
= true;
1389 blk_mq_end_request(rq
, ret
);
1391 } while (!list_empty(list
));
1393 if (!list_empty(&zone_list
))
1394 list_splice_tail_init(&zone_list
, list
);
1396 hctx
->dispatched
[queued_to_index(queued
)]++;
1398 /* If we didn't flush the entire list, we could have told the driver
1399 * there was more coming, but that turned out to be a lie.
1401 if ((!list_empty(list
) || errors
) && q
->mq_ops
->commit_rqs
&& queued
)
1402 q
->mq_ops
->commit_rqs(hctx
);
1404 * Any items that need requeuing? Stuff them into hctx->dispatch,
1405 * that is where we will continue on next queue run.
1407 if (!list_empty(list
)) {
1409 /* For non-shared tags, the RESTART check will suffice */
1410 bool no_tag
= prep
== PREP_DISPATCH_NO_TAG
&&
1411 (hctx
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
);
1414 blk_mq_release_budgets(q
, list
);
1416 spin_lock(&hctx
->lock
);
1417 list_splice_tail_init(list
, &hctx
->dispatch
);
1418 spin_unlock(&hctx
->lock
);
1421 * Order adding requests to hctx->dispatch and checking
1422 * SCHED_RESTART flag. The pair of this smp_mb() is the one
1423 * in blk_mq_sched_restart(). Avoid restart code path to
1424 * miss the new added requests to hctx->dispatch, meantime
1425 * SCHED_RESTART is observed here.
1430 * If SCHED_RESTART was set by the caller of this function and
1431 * it is no longer set that means that it was cleared by another
1432 * thread and hence that a queue rerun is needed.
1434 * If 'no_tag' is set, that means that we failed getting
1435 * a driver tag with an I/O scheduler attached. If our dispatch
1436 * waitqueue is no longer active, ensure that we run the queue
1437 * AFTER adding our entries back to the list.
1439 * If no I/O scheduler has been configured it is possible that
1440 * the hardware queue got stopped and restarted before requests
1441 * were pushed back onto the dispatch list. Rerun the queue to
1442 * avoid starvation. Notes:
1443 * - blk_mq_run_hw_queue() checks whether or not a queue has
1444 * been stopped before rerunning a queue.
1445 * - Some but not all block drivers stop a queue before
1446 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1449 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1450 * bit is set, run queue after a delay to avoid IO stalls
1451 * that could otherwise occur if the queue is idle. We'll do
1452 * similar if we couldn't get budget or couldn't lock a zone
1453 * and SCHED_RESTART is set.
1455 needs_restart
= blk_mq_sched_needs_restart(hctx
);
1456 if (prep
== PREP_DISPATCH_NO_BUDGET
)
1457 needs_resource
= true;
1458 if (!needs_restart
||
1459 (no_tag
&& list_empty_careful(&hctx
->dispatch_wait
.entry
)))
1460 blk_mq_run_hw_queue(hctx
, true);
1461 else if (needs_restart
&& needs_resource
)
1462 blk_mq_delay_run_hw_queue(hctx
, BLK_MQ_RESOURCE_DELAY
);
1464 blk_mq_update_dispatch_busy(hctx
, true);
1467 blk_mq_update_dispatch_busy(hctx
, false);
1469 return (queued
+ errors
) != 0;
1473 * __blk_mq_run_hw_queue - Run a hardware queue.
1474 * @hctx: Pointer to the hardware queue to run.
1476 * Send pending requests to the hardware.
1478 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1483 * We can't run the queue inline with ints disabled. Ensure that
1484 * we catch bad users of this early.
1486 WARN_ON_ONCE(in_interrupt());
1488 might_sleep_if(hctx
->flags
& BLK_MQ_F_BLOCKING
);
1490 hctx_lock(hctx
, &srcu_idx
);
1491 blk_mq_sched_dispatch_requests(hctx
);
1492 hctx_unlock(hctx
, srcu_idx
);
1495 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx
*hctx
)
1497 int cpu
= cpumask_first_and(hctx
->cpumask
, cpu_online_mask
);
1499 if (cpu
>= nr_cpu_ids
)
1500 cpu
= cpumask_first(hctx
->cpumask
);
1505 * It'd be great if the workqueue API had a way to pass
1506 * in a mask and had some smarts for more clever placement.
1507 * For now we just round-robin here, switching for every
1508 * BLK_MQ_CPU_WORK_BATCH queued items.
1510 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1513 int next_cpu
= hctx
->next_cpu
;
1515 if (hctx
->queue
->nr_hw_queues
== 1)
1516 return WORK_CPU_UNBOUND
;
1518 if (--hctx
->next_cpu_batch
<= 0) {
1520 next_cpu
= cpumask_next_and(next_cpu
, hctx
->cpumask
,
1522 if (next_cpu
>= nr_cpu_ids
)
1523 next_cpu
= blk_mq_first_mapped_cpu(hctx
);
1524 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1528 * Do unbound schedule if we can't find a online CPU for this hctx,
1529 * and it should only happen in the path of handling CPU DEAD.
1531 if (!cpu_online(next_cpu
)) {
1538 * Make sure to re-select CPU next time once after CPUs
1539 * in hctx->cpumask become online again.
1541 hctx
->next_cpu
= next_cpu
;
1542 hctx
->next_cpu_batch
= 1;
1543 return WORK_CPU_UNBOUND
;
1546 hctx
->next_cpu
= next_cpu
;
1551 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
1552 * @hctx: Pointer to the hardware queue to run.
1553 * @async: If we want to run the queue asynchronously.
1554 * @msecs: Milliseconds of delay to wait before running the queue.
1556 * If !@async, try to run the queue now. Else, run the queue asynchronously and
1557 * with a delay of @msecs.
1559 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
,
1560 unsigned long msecs
)
1562 if (unlikely(blk_mq_hctx_stopped(hctx
)))
1565 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1566 int cpu
= get_cpu();
1567 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1568 __blk_mq_run_hw_queue(hctx
);
1576 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
,
1577 msecs_to_jiffies(msecs
));
1581 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
1582 * @hctx: Pointer to the hardware queue to run.
1583 * @msecs: Milliseconds of delay to wait before running the queue.
1585 * Run a hardware queue asynchronously with a delay of @msecs.
1587 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1589 __blk_mq_delay_run_hw_queue(hctx
, true, msecs
);
1591 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue
);
1594 * blk_mq_run_hw_queue - Start to run a hardware queue.
1595 * @hctx: Pointer to the hardware queue to run.
1596 * @async: If we want to run the queue asynchronously.
1598 * Check if the request queue is not in a quiesced state and if there are
1599 * pending requests to be sent. If this is true, run the queue to send requests
1602 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1608 * When queue is quiesced, we may be switching io scheduler, or
1609 * updating nr_hw_queues, or other things, and we can't run queue
1610 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1612 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1615 hctx_lock(hctx
, &srcu_idx
);
1616 need_run
= !blk_queue_quiesced(hctx
->queue
) &&
1617 blk_mq_hctx_has_pending(hctx
);
1618 hctx_unlock(hctx
, srcu_idx
);
1621 __blk_mq_delay_run_hw_queue(hctx
, async
, 0);
1623 EXPORT_SYMBOL(blk_mq_run_hw_queue
);
1626 * Is the request queue handled by an IO scheduler that does not respect
1627 * hardware queues when dispatching?
1629 static bool blk_mq_has_sqsched(struct request_queue
*q
)
1631 struct elevator_queue
*e
= q
->elevator
;
1633 if (e
&& e
->type
->ops
.dispatch_request
&&
1634 !(e
->type
->elevator_features
& ELEVATOR_F_MQ_AWARE
))
1640 * Return prefered queue to dispatch from (if any) for non-mq aware IO
1643 static struct blk_mq_hw_ctx
*blk_mq_get_sq_hctx(struct request_queue
*q
)
1645 struct blk_mq_hw_ctx
*hctx
;
1648 * If the IO scheduler does not respect hardware queues when
1649 * dispatching, we just don't bother with multiple HW queues and
1650 * dispatch from hctx for the current CPU since running multiple queues
1651 * just causes lock contention inside the scheduler and pointless cache
1654 hctx
= blk_mq_map_queue_type(q
, HCTX_TYPE_DEFAULT
,
1655 raw_smp_processor_id());
1656 if (!blk_mq_hctx_stopped(hctx
))
1662 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
1663 * @q: Pointer to the request queue to run.
1664 * @async: If we want to run the queue asynchronously.
1666 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1668 struct blk_mq_hw_ctx
*hctx
, *sq_hctx
;
1672 if (blk_mq_has_sqsched(q
))
1673 sq_hctx
= blk_mq_get_sq_hctx(q
);
1674 queue_for_each_hw_ctx(q
, hctx
, i
) {
1675 if (blk_mq_hctx_stopped(hctx
))
1678 * Dispatch from this hctx either if there's no hctx preferred
1679 * by IO scheduler or if it has requests that bypass the
1682 if (!sq_hctx
|| sq_hctx
== hctx
||
1683 !list_empty_careful(&hctx
->dispatch
))
1684 blk_mq_run_hw_queue(hctx
, async
);
1687 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1690 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
1691 * @q: Pointer to the request queue to run.
1692 * @msecs: Milliseconds of delay to wait before running the queues.
1694 void blk_mq_delay_run_hw_queues(struct request_queue
*q
, unsigned long msecs
)
1696 struct blk_mq_hw_ctx
*hctx
, *sq_hctx
;
1700 if (blk_mq_has_sqsched(q
))
1701 sq_hctx
= blk_mq_get_sq_hctx(q
);
1702 queue_for_each_hw_ctx(q
, hctx
, i
) {
1703 if (blk_mq_hctx_stopped(hctx
))
1706 * Dispatch from this hctx either if there's no hctx preferred
1707 * by IO scheduler or if it has requests that bypass the
1710 if (!sq_hctx
|| sq_hctx
== hctx
||
1711 !list_empty_careful(&hctx
->dispatch
))
1712 blk_mq_delay_run_hw_queue(hctx
, msecs
);
1715 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues
);
1718 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1719 * @q: request queue.
1721 * The caller is responsible for serializing this function against
1722 * blk_mq_{start,stop}_hw_queue().
1724 bool blk_mq_queue_stopped(struct request_queue
*q
)
1726 struct blk_mq_hw_ctx
*hctx
;
1729 queue_for_each_hw_ctx(q
, hctx
, i
)
1730 if (blk_mq_hctx_stopped(hctx
))
1735 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1738 * This function is often used for pausing .queue_rq() by driver when
1739 * there isn't enough resource or some conditions aren't satisfied, and
1740 * BLK_STS_RESOURCE is usually returned.
1742 * We do not guarantee that dispatch can be drained or blocked
1743 * after blk_mq_stop_hw_queue() returns. Please use
1744 * blk_mq_quiesce_queue() for that requirement.
1746 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1748 cancel_delayed_work(&hctx
->run_work
);
1750 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1752 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1755 * This function is often used for pausing .queue_rq() by driver when
1756 * there isn't enough resource or some conditions aren't satisfied, and
1757 * BLK_STS_RESOURCE is usually returned.
1759 * We do not guarantee that dispatch can be drained or blocked
1760 * after blk_mq_stop_hw_queues() returns. Please use
1761 * blk_mq_quiesce_queue() for that requirement.
1763 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1765 struct blk_mq_hw_ctx
*hctx
;
1768 queue_for_each_hw_ctx(q
, hctx
, i
)
1769 blk_mq_stop_hw_queue(hctx
);
1771 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1773 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1775 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1777 blk_mq_run_hw_queue(hctx
, false);
1779 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1781 void blk_mq_start_hw_queues(struct request_queue
*q
)
1783 struct blk_mq_hw_ctx
*hctx
;
1786 queue_for_each_hw_ctx(q
, hctx
, i
)
1787 blk_mq_start_hw_queue(hctx
);
1789 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1791 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1793 if (!blk_mq_hctx_stopped(hctx
))
1796 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1797 blk_mq_run_hw_queue(hctx
, async
);
1799 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1801 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1803 struct blk_mq_hw_ctx
*hctx
;
1806 queue_for_each_hw_ctx(q
, hctx
, i
)
1807 blk_mq_start_stopped_hw_queue(hctx
, async
);
1809 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1811 static void blk_mq_run_work_fn(struct work_struct
*work
)
1813 struct blk_mq_hw_ctx
*hctx
;
1815 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
1818 * If we are stopped, don't run the queue.
1820 if (blk_mq_hctx_stopped(hctx
))
1823 __blk_mq_run_hw_queue(hctx
);
1826 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1830 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1831 enum hctx_type type
= hctx
->type
;
1833 lockdep_assert_held(&ctx
->lock
);
1835 trace_block_rq_insert(rq
);
1838 list_add(&rq
->queuelist
, &ctx
->rq_lists
[type
]);
1840 list_add_tail(&rq
->queuelist
, &ctx
->rq_lists
[type
]);
1843 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1846 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1848 lockdep_assert_held(&ctx
->lock
);
1850 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1851 blk_mq_hctx_mark_pending(hctx
, ctx
);
1855 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
1856 * @rq: Pointer to request to be inserted.
1857 * @at_head: true if the request should be inserted at the head of the list.
1858 * @run_queue: If we should run the hardware queue after inserting the request.
1860 * Should only be used carefully, when the caller knows we want to
1861 * bypass a potential IO scheduler on the target device.
1863 void blk_mq_request_bypass_insert(struct request
*rq
, bool at_head
,
1866 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1868 spin_lock(&hctx
->lock
);
1870 list_add(&rq
->queuelist
, &hctx
->dispatch
);
1872 list_add_tail(&rq
->queuelist
, &hctx
->dispatch
);
1873 spin_unlock(&hctx
->lock
);
1876 blk_mq_run_hw_queue(hctx
, false);
1879 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1880 struct list_head
*list
)
1884 enum hctx_type type
= hctx
->type
;
1887 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1890 list_for_each_entry(rq
, list
, queuelist
) {
1891 BUG_ON(rq
->mq_ctx
!= ctx
);
1892 trace_block_rq_insert(rq
);
1895 spin_lock(&ctx
->lock
);
1896 list_splice_tail_init(list
, &ctx
->rq_lists
[type
]);
1897 blk_mq_hctx_mark_pending(hctx
, ctx
);
1898 spin_unlock(&ctx
->lock
);
1901 static int plug_rq_cmp(void *priv
, const struct list_head
*a
,
1902 const struct list_head
*b
)
1904 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1905 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1907 if (rqa
->mq_ctx
!= rqb
->mq_ctx
)
1908 return rqa
->mq_ctx
> rqb
->mq_ctx
;
1909 if (rqa
->mq_hctx
!= rqb
->mq_hctx
)
1910 return rqa
->mq_hctx
> rqb
->mq_hctx
;
1912 return blk_rq_pos(rqa
) > blk_rq_pos(rqb
);
1915 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1919 if (list_empty(&plug
->mq_list
))
1921 list_splice_init(&plug
->mq_list
, &list
);
1923 if (plug
->rq_count
> 2 && plug
->multiple_queues
)
1924 list_sort(NULL
, &list
, plug_rq_cmp
);
1929 struct list_head rq_list
;
1930 struct request
*rq
, *head_rq
= list_entry_rq(list
.next
);
1931 struct list_head
*pos
= &head_rq
->queuelist
; /* skip first */
1932 struct blk_mq_hw_ctx
*this_hctx
= head_rq
->mq_hctx
;
1933 struct blk_mq_ctx
*this_ctx
= head_rq
->mq_ctx
;
1934 unsigned int depth
= 1;
1936 list_for_each_continue(pos
, &list
) {
1937 rq
= list_entry_rq(pos
);
1939 if (rq
->mq_hctx
!= this_hctx
|| rq
->mq_ctx
!= this_ctx
)
1944 list_cut_before(&rq_list
, &list
, pos
);
1945 trace_block_unplug(head_rq
->q
, depth
, !from_schedule
);
1946 blk_mq_sched_insert_requests(this_hctx
, this_ctx
, &rq_list
,
1948 } while(!list_empty(&list
));
1951 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
,
1952 unsigned int nr_segs
)
1956 if (bio
->bi_opf
& REQ_RAHEAD
)
1957 rq
->cmd_flags
|= REQ_FAILFAST_MASK
;
1959 rq
->__sector
= bio
->bi_iter
.bi_sector
;
1960 rq
->write_hint
= bio
->bi_write_hint
;
1961 blk_rq_bio_prep(rq
, bio
, nr_segs
);
1963 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
1964 err
= blk_crypto_rq_bio_prep(rq
, bio
, GFP_NOIO
);
1967 blk_account_io_start(rq
);
1970 static blk_status_t
__blk_mq_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1972 blk_qc_t
*cookie
, bool last
)
1974 struct request_queue
*q
= rq
->q
;
1975 struct blk_mq_queue_data bd
= {
1979 blk_qc_t new_cookie
;
1982 new_cookie
= request_to_qc_t(hctx
, rq
);
1985 * For OK queue, we are done. For error, caller may kill it.
1986 * Any other error (busy), just add it to our list as we
1987 * previously would have done.
1989 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1992 blk_mq_update_dispatch_busy(hctx
, false);
1993 *cookie
= new_cookie
;
1995 case BLK_STS_RESOURCE
:
1996 case BLK_STS_DEV_RESOURCE
:
1997 blk_mq_update_dispatch_busy(hctx
, true);
1998 __blk_mq_requeue_request(rq
);
2001 blk_mq_update_dispatch_busy(hctx
, false);
2002 *cookie
= BLK_QC_T_NONE
;
2009 static blk_status_t
__blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
2012 bool bypass_insert
, bool last
)
2014 struct request_queue
*q
= rq
->q
;
2015 bool run_queue
= true;
2019 * RCU or SRCU read lock is needed before checking quiesced flag.
2021 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2022 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2023 * and avoid driver to try to dispatch again.
2025 if (blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(q
)) {
2027 bypass_insert
= false;
2031 if (q
->elevator
&& !bypass_insert
)
2034 budget_token
= blk_mq_get_dispatch_budget(q
);
2035 if (budget_token
< 0)
2038 blk_mq_set_rq_budget_token(rq
, budget_token
);
2040 if (!blk_mq_get_driver_tag(rq
)) {
2041 blk_mq_put_dispatch_budget(q
, budget_token
);
2045 return __blk_mq_issue_directly(hctx
, rq
, cookie
, last
);
2048 return BLK_STS_RESOURCE
;
2050 blk_mq_sched_insert_request(rq
, false, run_queue
, false);
2056 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2057 * @hctx: Pointer of the associated hardware queue.
2058 * @rq: Pointer to request to be sent.
2059 * @cookie: Request queue cookie.
2061 * If the device has enough resources to accept a new request now, send the
2062 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2063 * we can try send it another time in the future. Requests inserted at this
2064 * queue have higher priority.
2066 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
2067 struct request
*rq
, blk_qc_t
*cookie
)
2072 might_sleep_if(hctx
->flags
& BLK_MQ_F_BLOCKING
);
2074 hctx_lock(hctx
, &srcu_idx
);
2076 ret
= __blk_mq_try_issue_directly(hctx
, rq
, cookie
, false, true);
2077 if (ret
== BLK_STS_RESOURCE
|| ret
== BLK_STS_DEV_RESOURCE
)
2078 blk_mq_request_bypass_insert(rq
, false, true);
2079 else if (ret
!= BLK_STS_OK
)
2080 blk_mq_end_request(rq
, ret
);
2082 hctx_unlock(hctx
, srcu_idx
);
2085 blk_status_t
blk_mq_request_issue_directly(struct request
*rq
, bool last
)
2089 blk_qc_t unused_cookie
;
2090 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
2092 hctx_lock(hctx
, &srcu_idx
);
2093 ret
= __blk_mq_try_issue_directly(hctx
, rq
, &unused_cookie
, true, last
);
2094 hctx_unlock(hctx
, srcu_idx
);
2099 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx
*hctx
,
2100 struct list_head
*list
)
2105 while (!list_empty(list
)) {
2107 struct request
*rq
= list_first_entry(list
, struct request
,
2110 list_del_init(&rq
->queuelist
);
2111 ret
= blk_mq_request_issue_directly(rq
, list_empty(list
));
2112 if (ret
!= BLK_STS_OK
) {
2113 if (ret
== BLK_STS_RESOURCE
||
2114 ret
== BLK_STS_DEV_RESOURCE
) {
2115 blk_mq_request_bypass_insert(rq
, false,
2119 blk_mq_end_request(rq
, ret
);
2126 * If we didn't flush the entire list, we could have told
2127 * the driver there was more coming, but that turned out to
2130 if ((!list_empty(list
) || errors
) &&
2131 hctx
->queue
->mq_ops
->commit_rqs
&& queued
)
2132 hctx
->queue
->mq_ops
->commit_rqs(hctx
);
2135 static void blk_add_rq_to_plug(struct blk_plug
*plug
, struct request
*rq
)
2137 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
2139 if (!plug
->multiple_queues
&& !list_is_singular(&plug
->mq_list
)) {
2140 struct request
*tmp
;
2142 tmp
= list_first_entry(&plug
->mq_list
, struct request
,
2144 if (tmp
->q
!= rq
->q
)
2145 plug
->multiple_queues
= true;
2150 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
2151 * queues. This is important for md arrays to benefit from merging
2154 static inline unsigned short blk_plug_max_rq_count(struct blk_plug
*plug
)
2156 if (plug
->multiple_queues
)
2157 return BLK_MAX_REQUEST_COUNT
* 2;
2158 return BLK_MAX_REQUEST_COUNT
;
2162 * blk_mq_submit_bio - Create and send a request to block device.
2163 * @bio: Bio pointer.
2165 * Builds up a request structure from @q and @bio and send to the device. The
2166 * request may not be queued directly to hardware if:
2167 * * This request can be merged with another one
2168 * * We want to place request at plug queue for possible future merging
2169 * * There is an IO scheduler active at this queue
2171 * It will not queue the request if there is an error with the bio, or at the
2174 * Returns: Request queue cookie.
2176 blk_qc_t
blk_mq_submit_bio(struct bio
*bio
)
2178 struct request_queue
*q
= bio
->bi_bdev
->bd_disk
->queue
;
2179 const int is_sync
= op_is_sync(bio
->bi_opf
);
2180 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
2181 struct blk_mq_alloc_data data
= {
2185 struct blk_plug
*plug
;
2186 struct request
*same_queue_rq
= NULL
;
2187 unsigned int nr_segs
;
2192 blk_queue_bounce(q
, &bio
);
2193 __blk_queue_split(&bio
, &nr_segs
);
2195 if (!bio_integrity_prep(bio
))
2198 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
2199 blk_attempt_plug_merge(q
, bio
, nr_segs
, &same_queue_rq
))
2202 if (blk_mq_sched_bio_merge(q
, bio
, nr_segs
))
2205 rq_qos_throttle(q
, bio
);
2207 hipri
= bio
->bi_opf
& REQ_HIPRI
;
2209 data
.cmd_flags
= bio
->bi_opf
;
2210 rq
= __blk_mq_alloc_request(&data
);
2211 if (unlikely(!rq
)) {
2212 rq_qos_cleanup(q
, bio
);
2213 if (bio
->bi_opf
& REQ_NOWAIT
)
2214 bio_wouldblock_error(bio
);
2218 trace_block_getrq(bio
);
2220 rq_qos_track(q
, rq
, bio
);
2222 cookie
= request_to_qc_t(data
.hctx
, rq
);
2224 blk_mq_bio_to_request(rq
, bio
, nr_segs
);
2226 ret
= blk_crypto_init_request(rq
);
2227 if (ret
!= BLK_STS_OK
) {
2228 bio
->bi_status
= ret
;
2230 blk_mq_free_request(rq
);
2231 return BLK_QC_T_NONE
;
2234 plug
= blk_mq_plug(q
, bio
);
2235 if (unlikely(is_flush_fua
)) {
2236 /* Bypass scheduler for flush requests */
2237 blk_insert_flush(rq
);
2238 blk_mq_run_hw_queue(data
.hctx
, true);
2239 } else if (plug
&& (q
->nr_hw_queues
== 1 ||
2240 blk_mq_is_sbitmap_shared(rq
->mq_hctx
->flags
) ||
2241 q
->mq_ops
->commit_rqs
|| !blk_queue_nonrot(q
))) {
2243 * Use plugging if we have a ->commit_rqs() hook as well, as
2244 * we know the driver uses bd->last in a smart fashion.
2246 * Use normal plugging if this disk is slow HDD, as sequential
2247 * IO may benefit a lot from plug merging.
2249 unsigned int request_count
= plug
->rq_count
;
2250 struct request
*last
= NULL
;
2253 trace_block_plug(q
);
2255 last
= list_entry_rq(plug
->mq_list
.prev
);
2257 if (request_count
>= blk_plug_max_rq_count(plug
) || (last
&&
2258 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
2259 blk_flush_plug_list(plug
, false);
2260 trace_block_plug(q
);
2263 blk_add_rq_to_plug(plug
, rq
);
2264 } else if (q
->elevator
) {
2265 /* Insert the request at the IO scheduler queue */
2266 blk_mq_sched_insert_request(rq
, false, true, true);
2267 } else if (plug
&& !blk_queue_nomerges(q
)) {
2269 * We do limited plugging. If the bio can be merged, do that.
2270 * Otherwise the existing request in the plug list will be
2271 * issued. So the plug list will have one request at most
2272 * The plug list might get flushed before this. If that happens,
2273 * the plug list is empty, and same_queue_rq is invalid.
2275 if (list_empty(&plug
->mq_list
))
2276 same_queue_rq
= NULL
;
2277 if (same_queue_rq
) {
2278 list_del_init(&same_queue_rq
->queuelist
);
2281 blk_add_rq_to_plug(plug
, rq
);
2282 trace_block_plug(q
);
2284 if (same_queue_rq
) {
2285 data
.hctx
= same_queue_rq
->mq_hctx
;
2286 trace_block_unplug(q
, 1, true);
2287 blk_mq_try_issue_directly(data
.hctx
, same_queue_rq
,
2290 } else if ((q
->nr_hw_queues
> 1 && is_sync
) ||
2291 !data
.hctx
->dispatch_busy
) {
2293 * There is no scheduler and we can try to send directly
2296 blk_mq_try_issue_directly(data
.hctx
, rq
, &cookie
);
2299 blk_mq_sched_insert_request(rq
, false, true, true);
2303 return BLK_QC_T_NONE
;
2307 return BLK_QC_T_NONE
;
2310 static size_t order_to_size(unsigned int order
)
2312 return (size_t)PAGE_SIZE
<< order
;
2315 /* called before freeing request pool in @tags */
2316 static void blk_mq_clear_rq_mapping(struct blk_mq_tag_set
*set
,
2317 struct blk_mq_tags
*tags
, unsigned int hctx_idx
)
2319 struct blk_mq_tags
*drv_tags
= set
->tags
[hctx_idx
];
2321 unsigned long flags
;
2323 list_for_each_entry(page
, &tags
->page_list
, lru
) {
2324 unsigned long start
= (unsigned long)page_address(page
);
2325 unsigned long end
= start
+ order_to_size(page
->private);
2328 for (i
= 0; i
< set
->queue_depth
; i
++) {
2329 struct request
*rq
= drv_tags
->rqs
[i
];
2330 unsigned long rq_addr
= (unsigned long)rq
;
2332 if (rq_addr
>= start
&& rq_addr
< end
) {
2333 WARN_ON_ONCE(refcount_read(&rq
->ref
) != 0);
2334 cmpxchg(&drv_tags
->rqs
[i
], rq
, NULL
);
2340 * Wait until all pending iteration is done.
2342 * Request reference is cleared and it is guaranteed to be observed
2343 * after the ->lock is released.
2345 spin_lock_irqsave(&drv_tags
->lock
, flags
);
2346 spin_unlock_irqrestore(&drv_tags
->lock
, flags
);
2349 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
2350 unsigned int hctx_idx
)
2354 if (tags
->rqs
&& set
->ops
->exit_request
) {
2357 for (i
= 0; i
< tags
->nr_tags
; i
++) {
2358 struct request
*rq
= tags
->static_rqs
[i
];
2362 set
->ops
->exit_request(set
, rq
, hctx_idx
);
2363 tags
->static_rqs
[i
] = NULL
;
2367 blk_mq_clear_rq_mapping(set
, tags
, hctx_idx
);
2369 while (!list_empty(&tags
->page_list
)) {
2370 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
2371 list_del_init(&page
->lru
);
2373 * Remove kmemleak object previously allocated in
2374 * blk_mq_alloc_rqs().
2376 kmemleak_free(page_address(page
));
2377 __free_pages(page
, page
->private);
2381 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
, unsigned int flags
)
2385 kfree(tags
->static_rqs
);
2386 tags
->static_rqs
= NULL
;
2388 blk_mq_free_tags(tags
, flags
);
2391 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
2392 unsigned int hctx_idx
,
2393 unsigned int nr_tags
,
2394 unsigned int reserved_tags
,
2397 struct blk_mq_tags
*tags
;
2400 node
= blk_mq_hw_queue_to_node(&set
->map
[HCTX_TYPE_DEFAULT
], hctx_idx
);
2401 if (node
== NUMA_NO_NODE
)
2402 node
= set
->numa_node
;
2404 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
, flags
);
2408 tags
->rqs
= kcalloc_node(nr_tags
, sizeof(struct request
*),
2409 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
2412 blk_mq_free_tags(tags
, flags
);
2416 tags
->static_rqs
= kcalloc_node(nr_tags
, sizeof(struct request
*),
2417 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
2419 if (!tags
->static_rqs
) {
2421 blk_mq_free_tags(tags
, flags
);
2428 static int blk_mq_init_request(struct blk_mq_tag_set
*set
, struct request
*rq
,
2429 unsigned int hctx_idx
, int node
)
2433 if (set
->ops
->init_request
) {
2434 ret
= set
->ops
->init_request(set
, rq
, hctx_idx
, node
);
2439 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
2443 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
2444 unsigned int hctx_idx
, unsigned int depth
)
2446 unsigned int i
, j
, entries_per_page
, max_order
= 4;
2447 size_t rq_size
, left
;
2450 node
= blk_mq_hw_queue_to_node(&set
->map
[HCTX_TYPE_DEFAULT
], hctx_idx
);
2451 if (node
== NUMA_NO_NODE
)
2452 node
= set
->numa_node
;
2454 INIT_LIST_HEAD(&tags
->page_list
);
2457 * rq_size is the size of the request plus driver payload, rounded
2458 * to the cacheline size
2460 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
2462 left
= rq_size
* depth
;
2464 for (i
= 0; i
< depth
; ) {
2465 int this_order
= max_order
;
2470 while (this_order
&& left
< order_to_size(this_order
- 1))
2474 page
= alloc_pages_node(node
,
2475 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
2481 if (order_to_size(this_order
) < rq_size
)
2488 page
->private = this_order
;
2489 list_add_tail(&page
->lru
, &tags
->page_list
);
2491 p
= page_address(page
);
2493 * Allow kmemleak to scan these pages as they contain pointers
2494 * to additional allocations like via ops->init_request().
2496 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
2497 entries_per_page
= order_to_size(this_order
) / rq_size
;
2498 to_do
= min(entries_per_page
, depth
- i
);
2499 left
-= to_do
* rq_size
;
2500 for (j
= 0; j
< to_do
; j
++) {
2501 struct request
*rq
= p
;
2503 tags
->static_rqs
[i
] = rq
;
2504 if (blk_mq_init_request(set
, rq
, hctx_idx
, node
)) {
2505 tags
->static_rqs
[i
] = NULL
;
2516 blk_mq_free_rqs(set
, tags
, hctx_idx
);
2520 struct rq_iter_data
{
2521 struct blk_mq_hw_ctx
*hctx
;
2525 static bool blk_mq_has_request(struct request
*rq
, void *data
, bool reserved
)
2527 struct rq_iter_data
*iter_data
= data
;
2529 if (rq
->mq_hctx
!= iter_data
->hctx
)
2531 iter_data
->has_rq
= true;
2535 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx
*hctx
)
2537 struct blk_mq_tags
*tags
= hctx
->sched_tags
?
2538 hctx
->sched_tags
: hctx
->tags
;
2539 struct rq_iter_data data
= {
2543 blk_mq_all_tag_iter(tags
, blk_mq_has_request
, &data
);
2547 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu
,
2548 struct blk_mq_hw_ctx
*hctx
)
2550 if (cpumask_next_and(-1, hctx
->cpumask
, cpu_online_mask
) != cpu
)
2552 if (cpumask_next_and(cpu
, hctx
->cpumask
, cpu_online_mask
) < nr_cpu_ids
)
2557 static int blk_mq_hctx_notify_offline(unsigned int cpu
, struct hlist_node
*node
)
2559 struct blk_mq_hw_ctx
*hctx
= hlist_entry_safe(node
,
2560 struct blk_mq_hw_ctx
, cpuhp_online
);
2562 if (!cpumask_test_cpu(cpu
, hctx
->cpumask
) ||
2563 !blk_mq_last_cpu_in_hctx(cpu
, hctx
))
2567 * Prevent new request from being allocated on the current hctx.
2569 * The smp_mb__after_atomic() Pairs with the implied barrier in
2570 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
2571 * seen once we return from the tag allocator.
2573 set_bit(BLK_MQ_S_INACTIVE
, &hctx
->state
);
2574 smp_mb__after_atomic();
2577 * Try to grab a reference to the queue and wait for any outstanding
2578 * requests. If we could not grab a reference the queue has been
2579 * frozen and there are no requests.
2581 if (percpu_ref_tryget(&hctx
->queue
->q_usage_counter
)) {
2582 while (blk_mq_hctx_has_requests(hctx
))
2584 percpu_ref_put(&hctx
->queue
->q_usage_counter
);
2590 static int blk_mq_hctx_notify_online(unsigned int cpu
, struct hlist_node
*node
)
2592 struct blk_mq_hw_ctx
*hctx
= hlist_entry_safe(node
,
2593 struct blk_mq_hw_ctx
, cpuhp_online
);
2595 if (cpumask_test_cpu(cpu
, hctx
->cpumask
))
2596 clear_bit(BLK_MQ_S_INACTIVE
, &hctx
->state
);
2601 * 'cpu' is going away. splice any existing rq_list entries from this
2602 * software queue to the hw queue dispatch list, and ensure that it
2605 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
2607 struct blk_mq_hw_ctx
*hctx
;
2608 struct blk_mq_ctx
*ctx
;
2610 enum hctx_type type
;
2612 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
2613 if (!cpumask_test_cpu(cpu
, hctx
->cpumask
))
2616 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
2619 spin_lock(&ctx
->lock
);
2620 if (!list_empty(&ctx
->rq_lists
[type
])) {
2621 list_splice_init(&ctx
->rq_lists
[type
], &tmp
);
2622 blk_mq_hctx_clear_pending(hctx
, ctx
);
2624 spin_unlock(&ctx
->lock
);
2626 if (list_empty(&tmp
))
2629 spin_lock(&hctx
->lock
);
2630 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
2631 spin_unlock(&hctx
->lock
);
2633 blk_mq_run_hw_queue(hctx
, true);
2637 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
2639 if (!(hctx
->flags
& BLK_MQ_F_STACKING
))
2640 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE
,
2641 &hctx
->cpuhp_online
);
2642 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
2647 * Before freeing hw queue, clearing the flush request reference in
2648 * tags->rqs[] for avoiding potential UAF.
2650 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags
*tags
,
2651 unsigned int queue_depth
, struct request
*flush_rq
)
2654 unsigned long flags
;
2656 /* The hw queue may not be mapped yet */
2660 WARN_ON_ONCE(refcount_read(&flush_rq
->ref
) != 0);
2662 for (i
= 0; i
< queue_depth
; i
++)
2663 cmpxchg(&tags
->rqs
[i
], flush_rq
, NULL
);
2666 * Wait until all pending iteration is done.
2668 * Request reference is cleared and it is guaranteed to be observed
2669 * after the ->lock is released.
2671 spin_lock_irqsave(&tags
->lock
, flags
);
2672 spin_unlock_irqrestore(&tags
->lock
, flags
);
2675 /* hctx->ctxs will be freed in queue's release handler */
2676 static void blk_mq_exit_hctx(struct request_queue
*q
,
2677 struct blk_mq_tag_set
*set
,
2678 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
2680 struct request
*flush_rq
= hctx
->fq
->flush_rq
;
2682 if (blk_mq_hw_queue_mapped(hctx
))
2683 blk_mq_tag_idle(hctx
);
2685 blk_mq_clear_flush_rq_mapping(set
->tags
[hctx_idx
],
2686 set
->queue_depth
, flush_rq
);
2687 if (set
->ops
->exit_request
)
2688 set
->ops
->exit_request(set
, flush_rq
, hctx_idx
);
2690 if (set
->ops
->exit_hctx
)
2691 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2693 blk_mq_remove_cpuhp(hctx
);
2695 spin_lock(&q
->unused_hctx_lock
);
2696 list_add(&hctx
->hctx_list
, &q
->unused_hctx_list
);
2697 spin_unlock(&q
->unused_hctx_lock
);
2700 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
2701 struct blk_mq_tag_set
*set
, int nr_queue
)
2703 struct blk_mq_hw_ctx
*hctx
;
2706 queue_for_each_hw_ctx(q
, hctx
, i
) {
2709 blk_mq_debugfs_unregister_hctx(hctx
);
2710 blk_mq_exit_hctx(q
, set
, hctx
, i
);
2714 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set
*tag_set
)
2716 int hw_ctx_size
= sizeof(struct blk_mq_hw_ctx
);
2718 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx
, srcu
),
2719 __alignof__(struct blk_mq_hw_ctx
)) !=
2720 sizeof(struct blk_mq_hw_ctx
));
2722 if (tag_set
->flags
& BLK_MQ_F_BLOCKING
)
2723 hw_ctx_size
+= sizeof(struct srcu_struct
);
2728 static int blk_mq_init_hctx(struct request_queue
*q
,
2729 struct blk_mq_tag_set
*set
,
2730 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
2732 hctx
->queue_num
= hctx_idx
;
2734 if (!(hctx
->flags
& BLK_MQ_F_STACKING
))
2735 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE
,
2736 &hctx
->cpuhp_online
);
2737 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
2739 hctx
->tags
= set
->tags
[hctx_idx
];
2741 if (set
->ops
->init_hctx
&&
2742 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
2743 goto unregister_cpu_notifier
;
2745 if (blk_mq_init_request(set
, hctx
->fq
->flush_rq
, hctx_idx
,
2751 if (set
->ops
->exit_hctx
)
2752 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2753 unregister_cpu_notifier
:
2754 blk_mq_remove_cpuhp(hctx
);
2758 static struct blk_mq_hw_ctx
*
2759 blk_mq_alloc_hctx(struct request_queue
*q
, struct blk_mq_tag_set
*set
,
2762 struct blk_mq_hw_ctx
*hctx
;
2763 gfp_t gfp
= GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
;
2765 hctx
= kzalloc_node(blk_mq_hw_ctx_size(set
), gfp
, node
);
2767 goto fail_alloc_hctx
;
2769 if (!zalloc_cpumask_var_node(&hctx
->cpumask
, gfp
, node
))
2772 atomic_set(&hctx
->nr_active
, 0);
2773 if (node
== NUMA_NO_NODE
)
2774 node
= set
->numa_node
;
2775 hctx
->numa_node
= node
;
2777 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
2778 spin_lock_init(&hctx
->lock
);
2779 INIT_LIST_HEAD(&hctx
->dispatch
);
2781 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_QUEUE_SHARED
;
2783 INIT_LIST_HEAD(&hctx
->hctx_list
);
2786 * Allocate space for all possible cpus to avoid allocation at
2789 hctx
->ctxs
= kmalloc_array_node(nr_cpu_ids
, sizeof(void *),
2794 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8),
2795 gfp
, node
, false, false))
2799 spin_lock_init(&hctx
->dispatch_wait_lock
);
2800 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
2801 INIT_LIST_HEAD(&hctx
->dispatch_wait
.entry
);
2803 hctx
->fq
= blk_alloc_flush_queue(hctx
->numa_node
, set
->cmd_size
, gfp
);
2807 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
2808 init_srcu_struct(hctx
->srcu
);
2809 blk_mq_hctx_kobj_init(hctx
);
2814 sbitmap_free(&hctx
->ctx_map
);
2818 free_cpumask_var(hctx
->cpumask
);
2825 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
2826 unsigned int nr_hw_queues
)
2828 struct blk_mq_tag_set
*set
= q
->tag_set
;
2831 for_each_possible_cpu(i
) {
2832 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2833 struct blk_mq_hw_ctx
*hctx
;
2837 spin_lock_init(&__ctx
->lock
);
2838 for (k
= HCTX_TYPE_DEFAULT
; k
< HCTX_MAX_TYPES
; k
++)
2839 INIT_LIST_HEAD(&__ctx
->rq_lists
[k
]);
2844 * Set local node, IFF we have more than one hw queue. If
2845 * not, we remain on the home node of the device
2847 for (j
= 0; j
< set
->nr_maps
; j
++) {
2848 hctx
= blk_mq_map_queue_type(q
, j
, i
);
2849 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
2850 hctx
->numa_node
= cpu_to_node(i
);
2855 static bool __blk_mq_alloc_map_and_request(struct blk_mq_tag_set
*set
,
2858 unsigned int flags
= set
->flags
;
2861 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
2862 set
->queue_depth
, set
->reserved_tags
, flags
);
2863 if (!set
->tags
[hctx_idx
])
2866 ret
= blk_mq_alloc_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
,
2871 blk_mq_free_rq_map(set
->tags
[hctx_idx
], flags
);
2872 set
->tags
[hctx_idx
] = NULL
;
2876 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set
*set
,
2877 unsigned int hctx_idx
)
2879 unsigned int flags
= set
->flags
;
2881 if (set
->tags
&& set
->tags
[hctx_idx
]) {
2882 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
2883 blk_mq_free_rq_map(set
->tags
[hctx_idx
], flags
);
2884 set
->tags
[hctx_idx
] = NULL
;
2888 static void blk_mq_map_swqueue(struct request_queue
*q
)
2890 unsigned int i
, j
, hctx_idx
;
2891 struct blk_mq_hw_ctx
*hctx
;
2892 struct blk_mq_ctx
*ctx
;
2893 struct blk_mq_tag_set
*set
= q
->tag_set
;
2895 queue_for_each_hw_ctx(q
, hctx
, i
) {
2896 cpumask_clear(hctx
->cpumask
);
2898 hctx
->dispatch_from
= NULL
;
2902 * Map software to hardware queues.
2904 * If the cpu isn't present, the cpu is mapped to first hctx.
2906 for_each_possible_cpu(i
) {
2908 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2909 for (j
= 0; j
< set
->nr_maps
; j
++) {
2910 if (!set
->map
[j
].nr_queues
) {
2911 ctx
->hctxs
[j
] = blk_mq_map_queue_type(q
,
2912 HCTX_TYPE_DEFAULT
, i
);
2915 hctx_idx
= set
->map
[j
].mq_map
[i
];
2916 /* unmapped hw queue can be remapped after CPU topo changed */
2917 if (!set
->tags
[hctx_idx
] &&
2918 !__blk_mq_alloc_map_and_request(set
, hctx_idx
)) {
2920 * If tags initialization fail for some hctx,
2921 * that hctx won't be brought online. In this
2922 * case, remap the current ctx to hctx[0] which
2923 * is guaranteed to always have tags allocated
2925 set
->map
[j
].mq_map
[i
] = 0;
2928 hctx
= blk_mq_map_queue_type(q
, j
, i
);
2929 ctx
->hctxs
[j
] = hctx
;
2931 * If the CPU is already set in the mask, then we've
2932 * mapped this one already. This can happen if
2933 * devices share queues across queue maps.
2935 if (cpumask_test_cpu(i
, hctx
->cpumask
))
2938 cpumask_set_cpu(i
, hctx
->cpumask
);
2940 ctx
->index_hw
[hctx
->type
] = hctx
->nr_ctx
;
2941 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2944 * If the nr_ctx type overflows, we have exceeded the
2945 * amount of sw queues we can support.
2947 BUG_ON(!hctx
->nr_ctx
);
2950 for (; j
< HCTX_MAX_TYPES
; j
++)
2951 ctx
->hctxs
[j
] = blk_mq_map_queue_type(q
,
2952 HCTX_TYPE_DEFAULT
, i
);
2955 queue_for_each_hw_ctx(q
, hctx
, i
) {
2957 * If no software queues are mapped to this hardware queue,
2958 * disable it and free the request entries.
2960 if (!hctx
->nr_ctx
) {
2961 /* Never unmap queue 0. We need it as a
2962 * fallback in case of a new remap fails
2965 if (i
&& set
->tags
[i
])
2966 blk_mq_free_map_and_requests(set
, i
);
2972 hctx
->tags
= set
->tags
[i
];
2973 WARN_ON(!hctx
->tags
);
2976 * Set the map size to the number of mapped software queues.
2977 * This is more accurate and more efficient than looping
2978 * over all possibly mapped software queues.
2980 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2983 * Initialize batch roundrobin counts
2985 hctx
->next_cpu
= blk_mq_first_mapped_cpu(hctx
);
2986 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2991 * Caller needs to ensure that we're either frozen/quiesced, or that
2992 * the queue isn't live yet.
2994 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2996 struct blk_mq_hw_ctx
*hctx
;
2999 queue_for_each_hw_ctx(q
, hctx
, i
) {
3001 hctx
->flags
|= BLK_MQ_F_TAG_QUEUE_SHARED
;
3003 blk_mq_tag_idle(hctx
);
3004 hctx
->flags
&= ~BLK_MQ_F_TAG_QUEUE_SHARED
;
3009 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set
*set
,
3012 struct request_queue
*q
;
3014 lockdep_assert_held(&set
->tag_list_lock
);
3016 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3017 blk_mq_freeze_queue(q
);
3018 queue_set_hctx_shared(q
, shared
);
3019 blk_mq_unfreeze_queue(q
);
3023 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
3025 struct blk_mq_tag_set
*set
= q
->tag_set
;
3027 mutex_lock(&set
->tag_list_lock
);
3028 list_del(&q
->tag_set_list
);
3029 if (list_is_singular(&set
->tag_list
)) {
3030 /* just transitioned to unshared */
3031 set
->flags
&= ~BLK_MQ_F_TAG_QUEUE_SHARED
;
3032 /* update existing queue */
3033 blk_mq_update_tag_set_shared(set
, false);
3035 mutex_unlock(&set
->tag_list_lock
);
3036 INIT_LIST_HEAD(&q
->tag_set_list
);
3039 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
3040 struct request_queue
*q
)
3042 mutex_lock(&set
->tag_list_lock
);
3045 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3047 if (!list_empty(&set
->tag_list
) &&
3048 !(set
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
)) {
3049 set
->flags
|= BLK_MQ_F_TAG_QUEUE_SHARED
;
3050 /* update existing queue */
3051 blk_mq_update_tag_set_shared(set
, true);
3053 if (set
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
)
3054 queue_set_hctx_shared(q
, true);
3055 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
3057 mutex_unlock(&set
->tag_list_lock
);
3060 /* All allocations will be freed in release handler of q->mq_kobj */
3061 static int blk_mq_alloc_ctxs(struct request_queue
*q
)
3063 struct blk_mq_ctxs
*ctxs
;
3066 ctxs
= kzalloc(sizeof(*ctxs
), GFP_KERNEL
);
3070 ctxs
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
3071 if (!ctxs
->queue_ctx
)
3074 for_each_possible_cpu(cpu
) {
3075 struct blk_mq_ctx
*ctx
= per_cpu_ptr(ctxs
->queue_ctx
, cpu
);
3079 q
->mq_kobj
= &ctxs
->kobj
;
3080 q
->queue_ctx
= ctxs
->queue_ctx
;
3089 * It is the actual release handler for mq, but we do it from
3090 * request queue's release handler for avoiding use-after-free
3091 * and headache because q->mq_kobj shouldn't have been introduced,
3092 * but we can't group ctx/kctx kobj without it.
3094 void blk_mq_release(struct request_queue
*q
)
3096 struct blk_mq_hw_ctx
*hctx
, *next
;
3099 queue_for_each_hw_ctx(q
, hctx
, i
)
3100 WARN_ON_ONCE(hctx
&& list_empty(&hctx
->hctx_list
));
3102 /* all hctx are in .unused_hctx_list now */
3103 list_for_each_entry_safe(hctx
, next
, &q
->unused_hctx_list
, hctx_list
) {
3104 list_del_init(&hctx
->hctx_list
);
3105 kobject_put(&hctx
->kobj
);
3108 kfree(q
->queue_hw_ctx
);
3111 * release .mq_kobj and sw queue's kobject now because
3112 * both share lifetime with request queue.
3114 blk_mq_sysfs_deinit(q
);
3117 static struct request_queue
*blk_mq_init_queue_data(struct blk_mq_tag_set
*set
,
3120 struct request_queue
*q
;
3123 q
= blk_alloc_queue(set
->numa_node
);
3125 return ERR_PTR(-ENOMEM
);
3126 q
->queuedata
= queuedata
;
3127 ret
= blk_mq_init_allocated_queue(set
, q
);
3129 blk_cleanup_queue(q
);
3130 return ERR_PTR(ret
);
3135 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
3137 return blk_mq_init_queue_data(set
, NULL
);
3139 EXPORT_SYMBOL(blk_mq_init_queue
);
3141 struct gendisk
*__blk_mq_alloc_disk(struct blk_mq_tag_set
*set
, void *queuedata
,
3142 struct lock_class_key
*lkclass
)
3144 struct request_queue
*q
;
3145 struct gendisk
*disk
;
3147 q
= blk_mq_init_queue_data(set
, queuedata
);
3151 disk
= __alloc_disk_node(q
, set
->numa_node
, lkclass
);
3153 blk_cleanup_queue(q
);
3154 return ERR_PTR(-ENOMEM
);
3158 EXPORT_SYMBOL(__blk_mq_alloc_disk
);
3160 static struct blk_mq_hw_ctx
*blk_mq_alloc_and_init_hctx(
3161 struct blk_mq_tag_set
*set
, struct request_queue
*q
,
3162 int hctx_idx
, int node
)
3164 struct blk_mq_hw_ctx
*hctx
= NULL
, *tmp
;
3166 /* reuse dead hctx first */
3167 spin_lock(&q
->unused_hctx_lock
);
3168 list_for_each_entry(tmp
, &q
->unused_hctx_list
, hctx_list
) {
3169 if (tmp
->numa_node
== node
) {
3175 list_del_init(&hctx
->hctx_list
);
3176 spin_unlock(&q
->unused_hctx_lock
);
3179 hctx
= blk_mq_alloc_hctx(q
, set
, node
);
3183 if (blk_mq_init_hctx(q
, set
, hctx
, hctx_idx
))
3189 kobject_put(&hctx
->kobj
);
3194 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
3195 struct request_queue
*q
)
3198 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
3200 if (q
->nr_hw_queues
< set
->nr_hw_queues
) {
3201 struct blk_mq_hw_ctx
**new_hctxs
;
3203 new_hctxs
= kcalloc_node(set
->nr_hw_queues
,
3204 sizeof(*new_hctxs
), GFP_KERNEL
,
3209 memcpy(new_hctxs
, hctxs
, q
->nr_hw_queues
*
3211 q
->queue_hw_ctx
= new_hctxs
;
3216 /* protect against switching io scheduler */
3217 mutex_lock(&q
->sysfs_lock
);
3218 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
3220 struct blk_mq_hw_ctx
*hctx
;
3222 node
= blk_mq_hw_queue_to_node(&set
->map
[HCTX_TYPE_DEFAULT
], i
);
3224 * If the hw queue has been mapped to another numa node,
3225 * we need to realloc the hctx. If allocation fails, fallback
3226 * to use the previous one.
3228 if (hctxs
[i
] && (hctxs
[i
]->numa_node
== node
))
3231 hctx
= blk_mq_alloc_and_init_hctx(set
, q
, i
, node
);
3234 blk_mq_exit_hctx(q
, set
, hctxs
[i
], i
);
3238 pr_warn("Allocate new hctx on node %d fails,\
3239 fallback to previous one on node %d\n",
3240 node
, hctxs
[i
]->numa_node
);
3246 * Increasing nr_hw_queues fails. Free the newly allocated
3247 * hctxs and keep the previous q->nr_hw_queues.
3249 if (i
!= set
->nr_hw_queues
) {
3250 j
= q
->nr_hw_queues
;
3254 end
= q
->nr_hw_queues
;
3255 q
->nr_hw_queues
= set
->nr_hw_queues
;
3258 for (; j
< end
; j
++) {
3259 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
3263 blk_mq_free_map_and_requests(set
, j
);
3264 blk_mq_exit_hctx(q
, set
, hctx
, j
);
3268 mutex_unlock(&q
->sysfs_lock
);
3271 int blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
3272 struct request_queue
*q
)
3274 /* mark the queue as mq asap */
3275 q
->mq_ops
= set
->ops
;
3277 q
->poll_cb
= blk_stat_alloc_callback(blk_mq_poll_stats_fn
,
3278 blk_mq_poll_stats_bkt
,
3279 BLK_MQ_POLL_STATS_BKTS
, q
);
3283 if (blk_mq_alloc_ctxs(q
))
3286 /* init q->mq_kobj and sw queues' kobjects */
3287 blk_mq_sysfs_init(q
);
3289 INIT_LIST_HEAD(&q
->unused_hctx_list
);
3290 spin_lock_init(&q
->unused_hctx_lock
);
3292 blk_mq_realloc_hw_ctxs(set
, q
);
3293 if (!q
->nr_hw_queues
)
3296 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
3297 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
3301 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
3302 if (set
->nr_maps
> HCTX_TYPE_POLL
&&
3303 set
->map
[HCTX_TYPE_POLL
].nr_queues
)
3304 blk_queue_flag_set(QUEUE_FLAG_POLL
, q
);
3306 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
3307 INIT_LIST_HEAD(&q
->requeue_list
);
3308 spin_lock_init(&q
->requeue_lock
);
3310 q
->nr_requests
= set
->queue_depth
;
3313 * Default to classic polling
3315 q
->poll_nsec
= BLK_MQ_POLL_CLASSIC
;
3317 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
3318 blk_mq_add_queue_tag_set(set
, q
);
3319 blk_mq_map_swqueue(q
);
3323 kfree(q
->queue_hw_ctx
);
3324 q
->nr_hw_queues
= 0;
3325 blk_mq_sysfs_deinit(q
);
3327 blk_stat_free_callback(q
->poll_cb
);
3333 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
3335 /* tags can _not_ be used after returning from blk_mq_exit_queue */
3336 void blk_mq_exit_queue(struct request_queue
*q
)
3338 struct blk_mq_tag_set
*set
= q
->tag_set
;
3340 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
3341 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
3342 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
3343 blk_mq_del_queue_tag_set(q
);
3346 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
3350 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
3351 if (!__blk_mq_alloc_map_and_request(set
, i
))
3360 blk_mq_free_map_and_requests(set
, i
);
3366 * Allocate the request maps associated with this tag_set. Note that this
3367 * may reduce the depth asked for, if memory is tight. set->queue_depth
3368 * will be updated to reflect the allocated depth.
3370 static int blk_mq_alloc_map_and_requests(struct blk_mq_tag_set
*set
)
3375 depth
= set
->queue_depth
;
3377 err
= __blk_mq_alloc_rq_maps(set
);
3381 set
->queue_depth
>>= 1;
3382 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
3386 } while (set
->queue_depth
);
3388 if (!set
->queue_depth
|| err
) {
3389 pr_err("blk-mq: failed to allocate request map\n");
3393 if (depth
!= set
->queue_depth
)
3394 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3395 depth
, set
->queue_depth
);
3400 static int blk_mq_update_queue_map(struct blk_mq_tag_set
*set
)
3403 * blk_mq_map_queues() and multiple .map_queues() implementations
3404 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
3405 * number of hardware queues.
3407 if (set
->nr_maps
== 1)
3408 set
->map
[HCTX_TYPE_DEFAULT
].nr_queues
= set
->nr_hw_queues
;
3410 if (set
->ops
->map_queues
&& !is_kdump_kernel()) {
3414 * transport .map_queues is usually done in the following
3417 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3418 * mask = get_cpu_mask(queue)
3419 * for_each_cpu(cpu, mask)
3420 * set->map[x].mq_map[cpu] = queue;
3423 * When we need to remap, the table has to be cleared for
3424 * killing stale mapping since one CPU may not be mapped
3427 for (i
= 0; i
< set
->nr_maps
; i
++)
3428 blk_mq_clear_mq_map(&set
->map
[i
]);
3430 return set
->ops
->map_queues(set
);
3432 BUG_ON(set
->nr_maps
> 1);
3433 return blk_mq_map_queues(&set
->map
[HCTX_TYPE_DEFAULT
]);
3437 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set
*set
,
3438 int cur_nr_hw_queues
, int new_nr_hw_queues
)
3440 struct blk_mq_tags
**new_tags
;
3442 if (cur_nr_hw_queues
>= new_nr_hw_queues
)
3445 new_tags
= kcalloc_node(new_nr_hw_queues
, sizeof(struct blk_mq_tags
*),
3446 GFP_KERNEL
, set
->numa_node
);
3451 memcpy(new_tags
, set
->tags
, cur_nr_hw_queues
*
3452 sizeof(*set
->tags
));
3454 set
->tags
= new_tags
;
3455 set
->nr_hw_queues
= new_nr_hw_queues
;
3460 static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set
*set
,
3461 int new_nr_hw_queues
)
3463 return blk_mq_realloc_tag_set_tags(set
, 0, new_nr_hw_queues
);
3467 * Alloc a tag set to be associated with one or more request queues.
3468 * May fail with EINVAL for various error conditions. May adjust the
3469 * requested depth down, if it's too large. In that case, the set
3470 * value will be stored in set->queue_depth.
3472 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
3476 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
3478 if (!set
->nr_hw_queues
)
3480 if (!set
->queue_depth
)
3482 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
3485 if (!set
->ops
->queue_rq
)
3488 if (!set
->ops
->get_budget
^ !set
->ops
->put_budget
)
3491 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
3492 pr_info("blk-mq: reduced tag depth to %u\n",
3494 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
3499 else if (set
->nr_maps
> HCTX_MAX_TYPES
)
3503 * If a crashdump is active, then we are potentially in a very
3504 * memory constrained environment. Limit us to 1 queue and
3505 * 64 tags to prevent using too much memory.
3507 if (is_kdump_kernel()) {
3508 set
->nr_hw_queues
= 1;
3510 set
->queue_depth
= min(64U, set
->queue_depth
);
3513 * There is no use for more h/w queues than cpus if we just have
3516 if (set
->nr_maps
== 1 && set
->nr_hw_queues
> nr_cpu_ids
)
3517 set
->nr_hw_queues
= nr_cpu_ids
;
3519 if (blk_mq_alloc_tag_set_tags(set
, set
->nr_hw_queues
) < 0)
3523 for (i
= 0; i
< set
->nr_maps
; i
++) {
3524 set
->map
[i
].mq_map
= kcalloc_node(nr_cpu_ids
,
3525 sizeof(set
->map
[i
].mq_map
[0]),
3526 GFP_KERNEL
, set
->numa_node
);
3527 if (!set
->map
[i
].mq_map
)
3528 goto out_free_mq_map
;
3529 set
->map
[i
].nr_queues
= is_kdump_kernel() ? 1 : set
->nr_hw_queues
;
3532 ret
= blk_mq_update_queue_map(set
);
3534 goto out_free_mq_map
;
3536 ret
= blk_mq_alloc_map_and_requests(set
);
3538 goto out_free_mq_map
;
3540 if (blk_mq_is_sbitmap_shared(set
->flags
)) {
3541 atomic_set(&set
->active_queues_shared_sbitmap
, 0);
3543 if (blk_mq_init_shared_sbitmap(set
)) {
3545 goto out_free_mq_rq_maps
;
3549 mutex_init(&set
->tag_list_lock
);
3550 INIT_LIST_HEAD(&set
->tag_list
);
3554 out_free_mq_rq_maps
:
3555 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
3556 blk_mq_free_map_and_requests(set
, i
);
3558 for (i
= 0; i
< set
->nr_maps
; i
++) {
3559 kfree(set
->map
[i
].mq_map
);
3560 set
->map
[i
].mq_map
= NULL
;
3566 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
3568 /* allocate and initialize a tagset for a simple single-queue device */
3569 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set
*set
,
3570 const struct blk_mq_ops
*ops
, unsigned int queue_depth
,
3571 unsigned int set_flags
)
3573 memset(set
, 0, sizeof(*set
));
3575 set
->nr_hw_queues
= 1;
3577 set
->queue_depth
= queue_depth
;
3578 set
->numa_node
= NUMA_NO_NODE
;
3579 set
->flags
= set_flags
;
3580 return blk_mq_alloc_tag_set(set
);
3582 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set
);
3584 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
3588 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
3589 blk_mq_free_map_and_requests(set
, i
);
3591 if (blk_mq_is_sbitmap_shared(set
->flags
))
3592 blk_mq_exit_shared_sbitmap(set
);
3594 for (j
= 0; j
< set
->nr_maps
; j
++) {
3595 kfree(set
->map
[j
].mq_map
);
3596 set
->map
[j
].mq_map
= NULL
;
3602 EXPORT_SYMBOL(blk_mq_free_tag_set
);
3604 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
3606 struct blk_mq_tag_set
*set
= q
->tag_set
;
3607 struct blk_mq_hw_ctx
*hctx
;
3613 if (q
->nr_requests
== nr
)
3616 blk_mq_freeze_queue(q
);
3617 blk_mq_quiesce_queue(q
);
3620 queue_for_each_hw_ctx(q
, hctx
, i
) {
3624 * If we're using an MQ scheduler, just update the scheduler
3625 * queue depth. This is similar to what the old code would do.
3627 if (!hctx
->sched_tags
) {
3628 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
, nr
,
3630 if (!ret
&& blk_mq_is_sbitmap_shared(set
->flags
))
3631 blk_mq_tag_resize_shared_sbitmap(set
, nr
);
3633 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
3635 if (blk_mq_is_sbitmap_shared(set
->flags
)) {
3636 hctx
->sched_tags
->bitmap_tags
=
3637 &q
->sched_bitmap_tags
;
3638 hctx
->sched_tags
->breserved_tags
=
3639 &q
->sched_breserved_tags
;
3644 if (q
->elevator
&& q
->elevator
->type
->ops
.depth_updated
)
3645 q
->elevator
->type
->ops
.depth_updated(hctx
);
3648 q
->nr_requests
= nr
;
3649 if (q
->elevator
&& blk_mq_is_sbitmap_shared(set
->flags
))
3650 sbitmap_queue_resize(&q
->sched_bitmap_tags
,
3651 nr
- set
->reserved_tags
);
3654 blk_mq_unquiesce_queue(q
);
3655 blk_mq_unfreeze_queue(q
);
3661 * request_queue and elevator_type pair.
3662 * It is just used by __blk_mq_update_nr_hw_queues to cache
3663 * the elevator_type associated with a request_queue.
3665 struct blk_mq_qe_pair
{
3666 struct list_head node
;
3667 struct request_queue
*q
;
3668 struct elevator_type
*type
;
3672 * Cache the elevator_type in qe pair list and switch the
3673 * io scheduler to 'none'
3675 static bool blk_mq_elv_switch_none(struct list_head
*head
,
3676 struct request_queue
*q
)
3678 struct blk_mq_qe_pair
*qe
;
3683 qe
= kmalloc(sizeof(*qe
), GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
);
3687 INIT_LIST_HEAD(&qe
->node
);
3689 qe
->type
= q
->elevator
->type
;
3690 list_add(&qe
->node
, head
);
3692 mutex_lock(&q
->sysfs_lock
);
3694 * After elevator_switch_mq, the previous elevator_queue will be
3695 * released by elevator_release. The reference of the io scheduler
3696 * module get by elevator_get will also be put. So we need to get
3697 * a reference of the io scheduler module here to prevent it to be
3700 __module_get(qe
->type
->elevator_owner
);
3701 elevator_switch_mq(q
, NULL
);
3702 mutex_unlock(&q
->sysfs_lock
);
3707 static void blk_mq_elv_switch_back(struct list_head
*head
,
3708 struct request_queue
*q
)
3710 struct blk_mq_qe_pair
*qe
;
3711 struct elevator_type
*t
= NULL
;
3713 list_for_each_entry(qe
, head
, node
)
3722 list_del(&qe
->node
);
3725 mutex_lock(&q
->sysfs_lock
);
3726 elevator_switch_mq(q
, t
);
3727 mutex_unlock(&q
->sysfs_lock
);
3730 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
,
3733 struct request_queue
*q
;
3735 int prev_nr_hw_queues
;
3737 lockdep_assert_held(&set
->tag_list_lock
);
3739 if (set
->nr_maps
== 1 && nr_hw_queues
> nr_cpu_ids
)
3740 nr_hw_queues
= nr_cpu_ids
;
3741 if (nr_hw_queues
< 1)
3743 if (set
->nr_maps
== 1 && nr_hw_queues
== set
->nr_hw_queues
)
3746 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3747 blk_mq_freeze_queue(q
);
3749 * Switch IO scheduler to 'none', cleaning up the data associated
3750 * with the previous scheduler. We will switch back once we are done
3751 * updating the new sw to hw queue mappings.
3753 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3754 if (!blk_mq_elv_switch_none(&head
, q
))
3757 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3758 blk_mq_debugfs_unregister_hctxs(q
);
3759 blk_mq_sysfs_unregister(q
);
3762 prev_nr_hw_queues
= set
->nr_hw_queues
;
3763 if (blk_mq_realloc_tag_set_tags(set
, set
->nr_hw_queues
, nr_hw_queues
) <
3767 set
->nr_hw_queues
= nr_hw_queues
;
3769 blk_mq_update_queue_map(set
);
3770 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3771 blk_mq_realloc_hw_ctxs(set
, q
);
3772 if (q
->nr_hw_queues
!= set
->nr_hw_queues
) {
3773 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3774 nr_hw_queues
, prev_nr_hw_queues
);
3775 set
->nr_hw_queues
= prev_nr_hw_queues
;
3776 blk_mq_map_queues(&set
->map
[HCTX_TYPE_DEFAULT
]);
3779 blk_mq_map_swqueue(q
);
3783 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3784 blk_mq_sysfs_register(q
);
3785 blk_mq_debugfs_register_hctxs(q
);
3789 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3790 blk_mq_elv_switch_back(&head
, q
);
3792 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3793 blk_mq_unfreeze_queue(q
);
3796 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
3798 mutex_lock(&set
->tag_list_lock
);
3799 __blk_mq_update_nr_hw_queues(set
, nr_hw_queues
);
3800 mutex_unlock(&set
->tag_list_lock
);
3802 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
3804 /* Enable polling stats and return whether they were already enabled. */
3805 static bool blk_poll_stats_enable(struct request_queue
*q
)
3807 if (test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
3808 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS
, q
))
3810 blk_stat_add_callback(q
, q
->poll_cb
);
3814 static void blk_mq_poll_stats_start(struct request_queue
*q
)
3817 * We don't arm the callback if polling stats are not enabled or the
3818 * callback is already active.
3820 if (!test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
3821 blk_stat_is_active(q
->poll_cb
))
3824 blk_stat_activate_msecs(q
->poll_cb
, 100);
3827 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
)
3829 struct request_queue
*q
= cb
->data
;
3832 for (bucket
= 0; bucket
< BLK_MQ_POLL_STATS_BKTS
; bucket
++) {
3833 if (cb
->stat
[bucket
].nr_samples
)
3834 q
->poll_stat
[bucket
] = cb
->stat
[bucket
];
3838 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
3841 unsigned long ret
= 0;
3845 * If stats collection isn't on, don't sleep but turn it on for
3848 if (!blk_poll_stats_enable(q
))
3852 * As an optimistic guess, use half of the mean service time
3853 * for this type of request. We can (and should) make this smarter.
3854 * For instance, if the completion latencies are tight, we can
3855 * get closer than just half the mean. This is especially
3856 * important on devices where the completion latencies are longer
3857 * than ~10 usec. We do use the stats for the relevant IO size
3858 * if available which does lead to better estimates.
3860 bucket
= blk_mq_poll_stats_bkt(rq
);
3864 if (q
->poll_stat
[bucket
].nr_samples
)
3865 ret
= (q
->poll_stat
[bucket
].mean
+ 1) / 2;
3870 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
3873 struct hrtimer_sleeper hs
;
3874 enum hrtimer_mode mode
;
3878 if (rq
->rq_flags
& RQF_MQ_POLL_SLEPT
)
3882 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3884 * 0: use half of prev avg
3885 * >0: use this specific value
3887 if (q
->poll_nsec
> 0)
3888 nsecs
= q
->poll_nsec
;
3890 nsecs
= blk_mq_poll_nsecs(q
, rq
);
3895 rq
->rq_flags
|= RQF_MQ_POLL_SLEPT
;
3898 * This will be replaced with the stats tracking code, using
3899 * 'avg_completion_time / 2' as the pre-sleep target.
3903 mode
= HRTIMER_MODE_REL
;
3904 hrtimer_init_sleeper_on_stack(&hs
, CLOCK_MONOTONIC
, mode
);
3905 hrtimer_set_expires(&hs
.timer
, kt
);
3908 if (blk_mq_rq_state(rq
) == MQ_RQ_COMPLETE
)
3910 set_current_state(TASK_UNINTERRUPTIBLE
);
3911 hrtimer_sleeper_start_expires(&hs
, mode
);
3914 hrtimer_cancel(&hs
.timer
);
3915 mode
= HRTIMER_MODE_ABS
;
3916 } while (hs
.task
&& !signal_pending(current
));
3918 __set_current_state(TASK_RUNNING
);
3919 destroy_hrtimer_on_stack(&hs
.timer
);
3923 static bool blk_mq_poll_hybrid(struct request_queue
*q
,
3924 struct blk_mq_hw_ctx
*hctx
, blk_qc_t cookie
)
3928 if (q
->poll_nsec
== BLK_MQ_POLL_CLASSIC
)
3931 if (!blk_qc_t_is_internal(cookie
))
3932 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
3934 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
3936 * With scheduling, if the request has completed, we'll
3937 * get a NULL return here, as we clear the sched tag when
3938 * that happens. The request still remains valid, like always,
3939 * so we should be safe with just the NULL check.
3945 return blk_mq_poll_hybrid_sleep(q
, rq
);
3949 * blk_poll - poll for IO completions
3951 * @cookie: cookie passed back at IO submission time
3952 * @spin: whether to spin for completions
3955 * Poll for completions on the passed in queue. Returns number of
3956 * completed entries found. If @spin is true, then blk_poll will continue
3957 * looping until at least one completion is found, unless the task is
3958 * otherwise marked running (or we need to reschedule).
3960 int blk_poll(struct request_queue
*q
, blk_qc_t cookie
, bool spin
)
3962 struct blk_mq_hw_ctx
*hctx
;
3965 if (!blk_qc_t_valid(cookie
) ||
3966 !test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
3970 blk_flush_plug_list(current
->plug
, false);
3972 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
3975 * If we sleep, have the caller restart the poll loop to reset
3976 * the state. Like for the other success return cases, the
3977 * caller is responsible for checking if the IO completed. If
3978 * the IO isn't complete, we'll get called again and will go
3979 * straight to the busy poll loop. If specified not to spin,
3980 * we also should not sleep.
3982 if (spin
&& blk_mq_poll_hybrid(q
, hctx
, cookie
))
3985 hctx
->poll_considered
++;
3987 state
= get_current_state();
3991 hctx
->poll_invoked
++;
3993 ret
= q
->mq_ops
->poll(hctx
);
3995 hctx
->poll_success
++;
3996 __set_current_state(TASK_RUNNING
);
4000 if (signal_pending_state(state
, current
))
4001 __set_current_state(TASK_RUNNING
);
4003 if (task_is_running(current
))
4005 if (ret
< 0 || !spin
)
4008 } while (!need_resched());
4010 __set_current_state(TASK_RUNNING
);
4013 EXPORT_SYMBOL_GPL(blk_poll
);
4015 unsigned int blk_mq_rq_cpu(struct request
*rq
)
4017 return rq
->mq_ctx
->cpu
;
4019 EXPORT_SYMBOL(blk_mq_rq_cpu
);
4021 void blk_mq_cancel_work_sync(struct request_queue
*q
)
4023 if (queue_is_mq(q
)) {
4024 struct blk_mq_hw_ctx
*hctx
;
4027 cancel_delayed_work_sync(&q
->requeue_work
);
4029 queue_for_each_hw_ctx(q
, hctx
, i
)
4030 cancel_delayed_work_sync(&hctx
->run_work
);
4034 static int __init
blk_mq_init(void)
4038 for_each_possible_cpu(i
)
4039 init_llist_head(&per_cpu(blk_cpu_done
, i
));
4040 open_softirq(BLOCK_SOFTIRQ
, blk_done_softirq
);
4042 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD
,
4043 "block/softirq:dead", NULL
,
4044 blk_softirq_cpu_dead
);
4045 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
4046 blk_mq_hctx_notify_dead
);
4047 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE
, "block/mq:online",
4048 blk_mq_hctx_notify_online
,
4049 blk_mq_hctx_notify_offline
);
4052 subsys_initcall(blk_mq_init
);