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 BUG_ON(!list_empty(&rq
->queuelist
));
767 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
769 EXPORT_SYMBOL(blk_mq_requeue_request
);
771 static void blk_mq_requeue_work(struct work_struct
*work
)
773 struct request_queue
*q
=
774 container_of(work
, struct request_queue
, requeue_work
.work
);
776 struct request
*rq
, *next
;
778 spin_lock_irq(&q
->requeue_lock
);
779 list_splice_init(&q
->requeue_list
, &rq_list
);
780 spin_unlock_irq(&q
->requeue_lock
);
782 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
783 if (!(rq
->rq_flags
& (RQF_SOFTBARRIER
| RQF_DONTPREP
)))
786 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
787 list_del_init(&rq
->queuelist
);
789 * If RQF_DONTPREP, rq has contained some driver specific
790 * data, so insert it to hctx dispatch list to avoid any
793 if (rq
->rq_flags
& RQF_DONTPREP
)
794 blk_mq_request_bypass_insert(rq
, false, false);
796 blk_mq_sched_insert_request(rq
, true, false, false);
799 while (!list_empty(&rq_list
)) {
800 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
801 list_del_init(&rq
->queuelist
);
802 blk_mq_sched_insert_request(rq
, false, false, false);
805 blk_mq_run_hw_queues(q
, false);
808 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
809 bool kick_requeue_list
)
811 struct request_queue
*q
= rq
->q
;
815 * We abuse this flag that is otherwise used by the I/O scheduler to
816 * request head insertion from the workqueue.
818 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
820 spin_lock_irqsave(&q
->requeue_lock
, flags
);
822 rq
->rq_flags
|= RQF_SOFTBARRIER
;
823 list_add(&rq
->queuelist
, &q
->requeue_list
);
825 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
827 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
829 if (kick_requeue_list
)
830 blk_mq_kick_requeue_list(q
);
833 void blk_mq_kick_requeue_list(struct request_queue
*q
)
835 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
, 0);
837 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
839 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
842 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
,
843 msecs_to_jiffies(msecs
));
845 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
847 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
849 if (tag
< tags
->nr_tags
) {
850 prefetch(tags
->rqs
[tag
]);
851 return tags
->rqs
[tag
];
856 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
858 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
859 void *priv
, bool reserved
)
862 * If we find a request that isn't idle and the queue matches,
863 * we know the queue is busy. Return false to stop the iteration.
865 if (blk_mq_request_started(rq
) && rq
->q
== hctx
->queue
) {
875 bool blk_mq_queue_inflight(struct request_queue
*q
)
879 blk_mq_queue_tag_busy_iter(q
, blk_mq_rq_inflight
, &busy
);
882 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight
);
884 static void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
886 req
->rq_flags
|= RQF_TIMED_OUT
;
887 if (req
->q
->mq_ops
->timeout
) {
888 enum blk_eh_timer_return ret
;
890 ret
= req
->q
->mq_ops
->timeout(req
, reserved
);
891 if (ret
== BLK_EH_DONE
)
893 WARN_ON_ONCE(ret
!= BLK_EH_RESET_TIMER
);
899 static bool blk_mq_req_expired(struct request
*rq
, unsigned long *next
)
901 unsigned long deadline
;
903 if (blk_mq_rq_state(rq
) != MQ_RQ_IN_FLIGHT
)
905 if (rq
->rq_flags
& RQF_TIMED_OUT
)
908 deadline
= READ_ONCE(rq
->deadline
);
909 if (time_after_eq(jiffies
, deadline
))
914 else if (time_after(*next
, deadline
))
919 void blk_mq_put_rq_ref(struct request
*rq
)
923 else if (refcount_dec_and_test(&rq
->ref
))
924 __blk_mq_free_request(rq
);
927 static bool blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
928 struct request
*rq
, void *priv
, bool reserved
)
930 unsigned long *next
= priv
;
933 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
934 * be reallocated underneath the timeout handler's processing, then
935 * the expire check is reliable. If the request is not expired, then
936 * it was completed and reallocated as a new request after returning
937 * from blk_mq_check_expired().
939 if (blk_mq_req_expired(rq
, next
))
940 blk_mq_rq_timed_out(rq
, reserved
);
944 static void blk_mq_timeout_work(struct work_struct
*work
)
946 struct request_queue
*q
=
947 container_of(work
, struct request_queue
, timeout_work
);
948 unsigned long next
= 0;
949 struct blk_mq_hw_ctx
*hctx
;
952 /* A deadlock might occur if a request is stuck requiring a
953 * timeout at the same time a queue freeze is waiting
954 * completion, since the timeout code would not be able to
955 * acquire the queue reference here.
957 * That's why we don't use blk_queue_enter here; instead, we use
958 * percpu_ref_tryget directly, because we need to be able to
959 * obtain a reference even in the short window between the queue
960 * starting to freeze, by dropping the first reference in
961 * blk_freeze_queue_start, and the moment the last request is
962 * consumed, marked by the instant q_usage_counter reaches
965 if (!percpu_ref_tryget(&q
->q_usage_counter
))
968 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &next
);
971 mod_timer(&q
->timeout
, next
);
974 * Request timeouts are handled as a forward rolling timer. If
975 * we end up here it means that no requests are pending and
976 * also that no request has been pending for a while. Mark
979 queue_for_each_hw_ctx(q
, hctx
, i
) {
980 /* the hctx may be unmapped, so check it here */
981 if (blk_mq_hw_queue_mapped(hctx
))
982 blk_mq_tag_idle(hctx
);
988 struct flush_busy_ctx_data
{
989 struct blk_mq_hw_ctx
*hctx
;
990 struct list_head
*list
;
993 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
995 struct flush_busy_ctx_data
*flush_data
= data
;
996 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
997 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
998 enum hctx_type type
= hctx
->type
;
1000 spin_lock(&ctx
->lock
);
1001 list_splice_tail_init(&ctx
->rq_lists
[type
], flush_data
->list
);
1002 sbitmap_clear_bit(sb
, bitnr
);
1003 spin_unlock(&ctx
->lock
);
1008 * Process software queues that have been marked busy, splicing them
1009 * to the for-dispatch
1011 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
1013 struct flush_busy_ctx_data data
= {
1018 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
1020 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
1022 struct dispatch_rq_data
{
1023 struct blk_mq_hw_ctx
*hctx
;
1027 static bool dispatch_rq_from_ctx(struct sbitmap
*sb
, unsigned int bitnr
,
1030 struct dispatch_rq_data
*dispatch_data
= data
;
1031 struct blk_mq_hw_ctx
*hctx
= dispatch_data
->hctx
;
1032 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
1033 enum hctx_type type
= hctx
->type
;
1035 spin_lock(&ctx
->lock
);
1036 if (!list_empty(&ctx
->rq_lists
[type
])) {
1037 dispatch_data
->rq
= list_entry_rq(ctx
->rq_lists
[type
].next
);
1038 list_del_init(&dispatch_data
->rq
->queuelist
);
1039 if (list_empty(&ctx
->rq_lists
[type
]))
1040 sbitmap_clear_bit(sb
, bitnr
);
1042 spin_unlock(&ctx
->lock
);
1044 return !dispatch_data
->rq
;
1047 struct request
*blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx
*hctx
,
1048 struct blk_mq_ctx
*start
)
1050 unsigned off
= start
? start
->index_hw
[hctx
->type
] : 0;
1051 struct dispatch_rq_data data
= {
1056 __sbitmap_for_each_set(&hctx
->ctx_map
, off
,
1057 dispatch_rq_from_ctx
, &data
);
1062 static inline unsigned int queued_to_index(unsigned int queued
)
1067 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
1070 static bool __blk_mq_get_driver_tag(struct request
*rq
)
1072 struct sbitmap_queue
*bt
= rq
->mq_hctx
->tags
->bitmap_tags
;
1073 unsigned int tag_offset
= rq
->mq_hctx
->tags
->nr_reserved_tags
;
1076 blk_mq_tag_busy(rq
->mq_hctx
);
1078 if (blk_mq_tag_is_reserved(rq
->mq_hctx
->sched_tags
, rq
->internal_tag
)) {
1079 bt
= rq
->mq_hctx
->tags
->breserved_tags
;
1082 if (!hctx_may_queue(rq
->mq_hctx
, bt
))
1086 tag
= __sbitmap_queue_get(bt
);
1087 if (tag
== BLK_MQ_NO_TAG
)
1090 rq
->tag
= tag
+ tag_offset
;
1094 bool blk_mq_get_driver_tag(struct request
*rq
)
1096 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1098 if (rq
->tag
== BLK_MQ_NO_TAG
&& !__blk_mq_get_driver_tag(rq
))
1101 if ((hctx
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
) &&
1102 !(rq
->rq_flags
& RQF_MQ_INFLIGHT
)) {
1103 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
1104 __blk_mq_inc_active_requests(hctx
);
1106 hctx
->tags
->rqs
[rq
->tag
] = rq
;
1110 static int blk_mq_dispatch_wake(wait_queue_entry_t
*wait
, unsigned mode
,
1111 int flags
, void *key
)
1113 struct blk_mq_hw_ctx
*hctx
;
1115 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
1117 spin_lock(&hctx
->dispatch_wait_lock
);
1118 if (!list_empty(&wait
->entry
)) {
1119 struct sbitmap_queue
*sbq
;
1121 list_del_init(&wait
->entry
);
1122 sbq
= hctx
->tags
->bitmap_tags
;
1123 atomic_dec(&sbq
->ws_active
);
1125 spin_unlock(&hctx
->dispatch_wait_lock
);
1127 blk_mq_run_hw_queue(hctx
, true);
1132 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1133 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1134 * restart. For both cases, take care to check the condition again after
1135 * marking us as waiting.
1137 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx
*hctx
,
1140 struct sbitmap_queue
*sbq
= hctx
->tags
->bitmap_tags
;
1141 struct wait_queue_head
*wq
;
1142 wait_queue_entry_t
*wait
;
1145 if (!(hctx
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
)) {
1146 blk_mq_sched_mark_restart_hctx(hctx
);
1149 * It's possible that a tag was freed in the window between the
1150 * allocation failure and adding the hardware queue to the wait
1153 * Don't clear RESTART here, someone else could have set it.
1154 * At most this will cost an extra queue run.
1156 return blk_mq_get_driver_tag(rq
);
1159 wait
= &hctx
->dispatch_wait
;
1160 if (!list_empty_careful(&wait
->entry
))
1163 wq
= &bt_wait_ptr(sbq
, hctx
)->wait
;
1165 spin_lock_irq(&wq
->lock
);
1166 spin_lock(&hctx
->dispatch_wait_lock
);
1167 if (!list_empty(&wait
->entry
)) {
1168 spin_unlock(&hctx
->dispatch_wait_lock
);
1169 spin_unlock_irq(&wq
->lock
);
1173 atomic_inc(&sbq
->ws_active
);
1174 wait
->flags
&= ~WQ_FLAG_EXCLUSIVE
;
1175 __add_wait_queue(wq
, wait
);
1178 * It's possible that a tag was freed in the window between the
1179 * allocation failure and adding the hardware queue to the wait
1182 ret
= blk_mq_get_driver_tag(rq
);
1184 spin_unlock(&hctx
->dispatch_wait_lock
);
1185 spin_unlock_irq(&wq
->lock
);
1190 * We got a tag, remove ourselves from the wait queue to ensure
1191 * someone else gets the wakeup.
1193 list_del_init(&wait
->entry
);
1194 atomic_dec(&sbq
->ws_active
);
1195 spin_unlock(&hctx
->dispatch_wait_lock
);
1196 spin_unlock_irq(&wq
->lock
);
1201 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1202 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1204 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1205 * - EWMA is one simple way to compute running average value
1206 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1207 * - take 4 as factor for avoiding to get too small(0) result, and this
1208 * factor doesn't matter because EWMA decreases exponentially
1210 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx
*hctx
, bool busy
)
1214 ewma
= hctx
->dispatch_busy
;
1219 ewma
*= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
- 1;
1221 ewma
+= 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR
;
1222 ewma
/= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
;
1224 hctx
->dispatch_busy
= ewma
;
1227 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1229 static void blk_mq_handle_dev_resource(struct request
*rq
,
1230 struct list_head
*list
)
1232 struct request
*next
=
1233 list_first_entry_or_null(list
, struct request
, queuelist
);
1236 * If an I/O scheduler has been configured and we got a driver tag for
1237 * the next request already, free it.
1240 blk_mq_put_driver_tag(next
);
1242 list_add(&rq
->queuelist
, list
);
1243 __blk_mq_requeue_request(rq
);
1246 static void blk_mq_handle_zone_resource(struct request
*rq
,
1247 struct list_head
*zone_list
)
1250 * If we end up here it is because we cannot dispatch a request to a
1251 * specific zone due to LLD level zone-write locking or other zone
1252 * related resource not being available. In this case, set the request
1253 * aside in zone_list for retrying it later.
1255 list_add(&rq
->queuelist
, zone_list
);
1256 __blk_mq_requeue_request(rq
);
1259 enum prep_dispatch
{
1261 PREP_DISPATCH_NO_TAG
,
1262 PREP_DISPATCH_NO_BUDGET
,
1265 static enum prep_dispatch
blk_mq_prep_dispatch_rq(struct request
*rq
,
1268 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1269 int budget_token
= -1;
1272 budget_token
= blk_mq_get_dispatch_budget(rq
->q
);
1273 if (budget_token
< 0) {
1274 blk_mq_put_driver_tag(rq
);
1275 return PREP_DISPATCH_NO_BUDGET
;
1277 blk_mq_set_rq_budget_token(rq
, budget_token
);
1280 if (!blk_mq_get_driver_tag(rq
)) {
1282 * The initial allocation attempt failed, so we need to
1283 * rerun the hardware queue when a tag is freed. The
1284 * waitqueue takes care of that. If the queue is run
1285 * before we add this entry back on the dispatch list,
1286 * we'll re-run it below.
1288 if (!blk_mq_mark_tag_wait(hctx
, rq
)) {
1290 * All budgets not got from this function will be put
1291 * together during handling partial dispatch
1294 blk_mq_put_dispatch_budget(rq
->q
, budget_token
);
1295 return PREP_DISPATCH_NO_TAG
;
1299 return PREP_DISPATCH_OK
;
1302 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1303 static void blk_mq_release_budgets(struct request_queue
*q
,
1304 struct list_head
*list
)
1308 list_for_each_entry(rq
, list
, queuelist
) {
1309 int budget_token
= blk_mq_get_rq_budget_token(rq
);
1311 if (budget_token
>= 0)
1312 blk_mq_put_dispatch_budget(q
, budget_token
);
1317 * Returns true if we did some work AND can potentially do more.
1319 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
,
1320 unsigned int nr_budgets
)
1322 enum prep_dispatch prep
;
1323 struct request_queue
*q
= hctx
->queue
;
1324 struct request
*rq
, *nxt
;
1326 blk_status_t ret
= BLK_STS_OK
;
1327 LIST_HEAD(zone_list
);
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 case BLK_STS_DEV_RESOURCE
:
1374 blk_mq_handle_dev_resource(rq
, list
);
1376 case BLK_STS_ZONE_RESOURCE
:
1378 * Move the request to zone_list and keep going through
1379 * the dispatch list to find more requests the drive can
1382 blk_mq_handle_zone_resource(rq
, &zone_list
);
1386 blk_mq_end_request(rq
, ret
);
1388 } while (!list_empty(list
));
1390 if (!list_empty(&zone_list
))
1391 list_splice_tail_init(&zone_list
, list
);
1393 hctx
->dispatched
[queued_to_index(queued
)]++;
1395 /* If we didn't flush the entire list, we could have told the driver
1396 * there was more coming, but that turned out to be a lie.
1398 if ((!list_empty(list
) || errors
) && q
->mq_ops
->commit_rqs
&& queued
)
1399 q
->mq_ops
->commit_rqs(hctx
);
1401 * Any items that need requeuing? Stuff them into hctx->dispatch,
1402 * that is where we will continue on next queue run.
1404 if (!list_empty(list
)) {
1406 /* For non-shared tags, the RESTART check will suffice */
1407 bool no_tag
= prep
== PREP_DISPATCH_NO_TAG
&&
1408 (hctx
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
);
1409 bool no_budget_avail
= prep
== PREP_DISPATCH_NO_BUDGET
;
1412 blk_mq_release_budgets(q
, list
);
1414 spin_lock(&hctx
->lock
);
1415 list_splice_tail_init(list
, &hctx
->dispatch
);
1416 spin_unlock(&hctx
->lock
);
1419 * Order adding requests to hctx->dispatch and checking
1420 * SCHED_RESTART flag. The pair of this smp_mb() is the one
1421 * in blk_mq_sched_restart(). Avoid restart code path to
1422 * miss the new added requests to hctx->dispatch, meantime
1423 * SCHED_RESTART is observed here.
1428 * If SCHED_RESTART was set by the caller of this function and
1429 * it is no longer set that means that it was cleared by another
1430 * thread and hence that a queue rerun is needed.
1432 * If 'no_tag' is set, that means that we failed getting
1433 * a driver tag with an I/O scheduler attached. If our dispatch
1434 * waitqueue is no longer active, ensure that we run the queue
1435 * AFTER adding our entries back to the list.
1437 * If no I/O scheduler has been configured it is possible that
1438 * the hardware queue got stopped and restarted before requests
1439 * were pushed back onto the dispatch list. Rerun the queue to
1440 * avoid starvation. Notes:
1441 * - blk_mq_run_hw_queue() checks whether or not a queue has
1442 * been stopped before rerunning a queue.
1443 * - Some but not all block drivers stop a queue before
1444 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1447 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1448 * bit is set, run queue after a delay to avoid IO stalls
1449 * that could otherwise occur if the queue is idle. We'll do
1450 * similar if we couldn't get budget and SCHED_RESTART is set.
1452 needs_restart
= blk_mq_sched_needs_restart(hctx
);
1453 if (!needs_restart
||
1454 (no_tag
&& list_empty_careful(&hctx
->dispatch_wait
.entry
)))
1455 blk_mq_run_hw_queue(hctx
, true);
1456 else if (needs_restart
&& (ret
== BLK_STS_RESOURCE
||
1458 blk_mq_delay_run_hw_queue(hctx
, BLK_MQ_RESOURCE_DELAY
);
1460 blk_mq_update_dispatch_busy(hctx
, true);
1463 blk_mq_update_dispatch_busy(hctx
, false);
1465 return (queued
+ errors
) != 0;
1469 * __blk_mq_run_hw_queue - Run a hardware queue.
1470 * @hctx: Pointer to the hardware queue to run.
1472 * Send pending requests to the hardware.
1474 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1479 * We can't run the queue inline with ints disabled. Ensure that
1480 * we catch bad users of this early.
1482 WARN_ON_ONCE(in_interrupt());
1484 might_sleep_if(hctx
->flags
& BLK_MQ_F_BLOCKING
);
1486 hctx_lock(hctx
, &srcu_idx
);
1487 blk_mq_sched_dispatch_requests(hctx
);
1488 hctx_unlock(hctx
, srcu_idx
);
1491 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx
*hctx
)
1493 int cpu
= cpumask_first_and(hctx
->cpumask
, cpu_online_mask
);
1495 if (cpu
>= nr_cpu_ids
)
1496 cpu
= cpumask_first(hctx
->cpumask
);
1501 * It'd be great if the workqueue API had a way to pass
1502 * in a mask and had some smarts for more clever placement.
1503 * For now we just round-robin here, switching for every
1504 * BLK_MQ_CPU_WORK_BATCH queued items.
1506 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1509 int next_cpu
= hctx
->next_cpu
;
1511 if (hctx
->queue
->nr_hw_queues
== 1)
1512 return WORK_CPU_UNBOUND
;
1514 if (--hctx
->next_cpu_batch
<= 0) {
1516 next_cpu
= cpumask_next_and(next_cpu
, hctx
->cpumask
,
1518 if (next_cpu
>= nr_cpu_ids
)
1519 next_cpu
= blk_mq_first_mapped_cpu(hctx
);
1520 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1524 * Do unbound schedule if we can't find a online CPU for this hctx,
1525 * and it should only happen in the path of handling CPU DEAD.
1527 if (!cpu_online(next_cpu
)) {
1534 * Make sure to re-select CPU next time once after CPUs
1535 * in hctx->cpumask become online again.
1537 hctx
->next_cpu
= next_cpu
;
1538 hctx
->next_cpu_batch
= 1;
1539 return WORK_CPU_UNBOUND
;
1542 hctx
->next_cpu
= next_cpu
;
1547 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
1548 * @hctx: Pointer to the hardware queue to run.
1549 * @async: If we want to run the queue asynchronously.
1550 * @msecs: Milliseconds of delay to wait before running the queue.
1552 * If !@async, try to run the queue now. Else, run the queue asynchronously and
1553 * with a delay of @msecs.
1555 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
,
1556 unsigned long msecs
)
1558 if (unlikely(blk_mq_hctx_stopped(hctx
)))
1561 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1562 int cpu
= get_cpu();
1563 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1564 __blk_mq_run_hw_queue(hctx
);
1572 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
,
1573 msecs_to_jiffies(msecs
));
1577 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
1578 * @hctx: Pointer to the hardware queue to run.
1579 * @msecs: Milliseconds of delay to wait before running the queue.
1581 * Run a hardware queue asynchronously with a delay of @msecs.
1583 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1585 __blk_mq_delay_run_hw_queue(hctx
, true, msecs
);
1587 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue
);
1590 * blk_mq_run_hw_queue - Start to run a hardware queue.
1591 * @hctx: Pointer to the hardware queue to run.
1592 * @async: If we want to run the queue asynchronously.
1594 * Check if the request queue is not in a quiesced state and if there are
1595 * pending requests to be sent. If this is true, run the queue to send requests
1598 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1604 * When queue is quiesced, we may be switching io scheduler, or
1605 * updating nr_hw_queues, or other things, and we can't run queue
1606 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1608 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1611 hctx_lock(hctx
, &srcu_idx
);
1612 need_run
= !blk_queue_quiesced(hctx
->queue
) &&
1613 blk_mq_hctx_has_pending(hctx
);
1614 hctx_unlock(hctx
, srcu_idx
);
1617 __blk_mq_delay_run_hw_queue(hctx
, async
, 0);
1619 EXPORT_SYMBOL(blk_mq_run_hw_queue
);
1622 * Is the request queue handled by an IO scheduler that does not respect
1623 * hardware queues when dispatching?
1625 static bool blk_mq_has_sqsched(struct request_queue
*q
)
1627 struct elevator_queue
*e
= q
->elevator
;
1629 if (e
&& e
->type
->ops
.dispatch_request
&&
1630 !(e
->type
->elevator_features
& ELEVATOR_F_MQ_AWARE
))
1636 * Return prefered queue to dispatch from (if any) for non-mq aware IO
1639 static struct blk_mq_hw_ctx
*blk_mq_get_sq_hctx(struct request_queue
*q
)
1641 struct blk_mq_hw_ctx
*hctx
;
1644 * If the IO scheduler does not respect hardware queues when
1645 * dispatching, we just don't bother with multiple HW queues and
1646 * dispatch from hctx for the current CPU since running multiple queues
1647 * just causes lock contention inside the scheduler and pointless cache
1650 hctx
= blk_mq_map_queue_type(q
, HCTX_TYPE_DEFAULT
,
1651 raw_smp_processor_id());
1652 if (!blk_mq_hctx_stopped(hctx
))
1658 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
1659 * @q: Pointer to the request queue to run.
1660 * @async: If we want to run the queue asynchronously.
1662 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1664 struct blk_mq_hw_ctx
*hctx
, *sq_hctx
;
1668 if (blk_mq_has_sqsched(q
))
1669 sq_hctx
= blk_mq_get_sq_hctx(q
);
1670 queue_for_each_hw_ctx(q
, hctx
, i
) {
1671 if (blk_mq_hctx_stopped(hctx
))
1674 * Dispatch from this hctx either if there's no hctx preferred
1675 * by IO scheduler or if it has requests that bypass the
1678 if (!sq_hctx
|| sq_hctx
== hctx
||
1679 !list_empty_careful(&hctx
->dispatch
))
1680 blk_mq_run_hw_queue(hctx
, async
);
1683 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1686 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
1687 * @q: Pointer to the request queue to run.
1688 * @msecs: Milliseconds of delay to wait before running the queues.
1690 void blk_mq_delay_run_hw_queues(struct request_queue
*q
, unsigned long msecs
)
1692 struct blk_mq_hw_ctx
*hctx
, *sq_hctx
;
1696 if (blk_mq_has_sqsched(q
))
1697 sq_hctx
= blk_mq_get_sq_hctx(q
);
1698 queue_for_each_hw_ctx(q
, hctx
, i
) {
1699 if (blk_mq_hctx_stopped(hctx
))
1702 * Dispatch from this hctx either if there's no hctx preferred
1703 * by IO scheduler or if it has requests that bypass the
1706 if (!sq_hctx
|| sq_hctx
== hctx
||
1707 !list_empty_careful(&hctx
->dispatch
))
1708 blk_mq_delay_run_hw_queue(hctx
, msecs
);
1711 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues
);
1714 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1715 * @q: request queue.
1717 * The caller is responsible for serializing this function against
1718 * blk_mq_{start,stop}_hw_queue().
1720 bool blk_mq_queue_stopped(struct request_queue
*q
)
1722 struct blk_mq_hw_ctx
*hctx
;
1725 queue_for_each_hw_ctx(q
, hctx
, i
)
1726 if (blk_mq_hctx_stopped(hctx
))
1731 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1734 * This function is often used for pausing .queue_rq() by driver when
1735 * there isn't enough resource or some conditions aren't satisfied, and
1736 * BLK_STS_RESOURCE is usually returned.
1738 * We do not guarantee that dispatch can be drained or blocked
1739 * after blk_mq_stop_hw_queue() returns. Please use
1740 * blk_mq_quiesce_queue() for that requirement.
1742 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1744 cancel_delayed_work(&hctx
->run_work
);
1746 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1748 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1751 * This function is often used for pausing .queue_rq() by driver when
1752 * there isn't enough resource or some conditions aren't satisfied, and
1753 * BLK_STS_RESOURCE is usually returned.
1755 * We do not guarantee that dispatch can be drained or blocked
1756 * after blk_mq_stop_hw_queues() returns. Please use
1757 * blk_mq_quiesce_queue() for that requirement.
1759 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1761 struct blk_mq_hw_ctx
*hctx
;
1764 queue_for_each_hw_ctx(q
, hctx
, i
)
1765 blk_mq_stop_hw_queue(hctx
);
1767 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1769 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1771 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1773 blk_mq_run_hw_queue(hctx
, false);
1775 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1777 void blk_mq_start_hw_queues(struct request_queue
*q
)
1779 struct blk_mq_hw_ctx
*hctx
;
1782 queue_for_each_hw_ctx(q
, hctx
, i
)
1783 blk_mq_start_hw_queue(hctx
);
1785 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1787 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1789 if (!blk_mq_hctx_stopped(hctx
))
1792 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1793 blk_mq_run_hw_queue(hctx
, async
);
1795 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1797 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1799 struct blk_mq_hw_ctx
*hctx
;
1802 queue_for_each_hw_ctx(q
, hctx
, i
)
1803 blk_mq_start_stopped_hw_queue(hctx
, async
);
1805 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1807 static void blk_mq_run_work_fn(struct work_struct
*work
)
1809 struct blk_mq_hw_ctx
*hctx
;
1811 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
1814 * If we are stopped, don't run the queue.
1816 if (blk_mq_hctx_stopped(hctx
))
1819 __blk_mq_run_hw_queue(hctx
);
1822 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1826 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1827 enum hctx_type type
= hctx
->type
;
1829 lockdep_assert_held(&ctx
->lock
);
1831 trace_block_rq_insert(rq
);
1834 list_add(&rq
->queuelist
, &ctx
->rq_lists
[type
]);
1836 list_add_tail(&rq
->queuelist
, &ctx
->rq_lists
[type
]);
1839 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1842 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1844 lockdep_assert_held(&ctx
->lock
);
1846 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1847 blk_mq_hctx_mark_pending(hctx
, ctx
);
1851 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
1852 * @rq: Pointer to request to be inserted.
1853 * @at_head: true if the request should be inserted at the head of the list.
1854 * @run_queue: If we should run the hardware queue after inserting the request.
1856 * Should only be used carefully, when the caller knows we want to
1857 * bypass a potential IO scheduler on the target device.
1859 void blk_mq_request_bypass_insert(struct request
*rq
, bool at_head
,
1862 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1864 spin_lock(&hctx
->lock
);
1866 list_add(&rq
->queuelist
, &hctx
->dispatch
);
1868 list_add_tail(&rq
->queuelist
, &hctx
->dispatch
);
1869 spin_unlock(&hctx
->lock
);
1872 blk_mq_run_hw_queue(hctx
, false);
1875 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1876 struct list_head
*list
)
1880 enum hctx_type type
= hctx
->type
;
1883 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1886 list_for_each_entry(rq
, list
, queuelist
) {
1887 BUG_ON(rq
->mq_ctx
!= ctx
);
1888 trace_block_rq_insert(rq
);
1891 spin_lock(&ctx
->lock
);
1892 list_splice_tail_init(list
, &ctx
->rq_lists
[type
]);
1893 blk_mq_hctx_mark_pending(hctx
, ctx
);
1894 spin_unlock(&ctx
->lock
);
1897 static int plug_rq_cmp(void *priv
, const struct list_head
*a
,
1898 const struct list_head
*b
)
1900 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1901 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1903 if (rqa
->mq_ctx
!= rqb
->mq_ctx
)
1904 return rqa
->mq_ctx
> rqb
->mq_ctx
;
1905 if (rqa
->mq_hctx
!= rqb
->mq_hctx
)
1906 return rqa
->mq_hctx
> rqb
->mq_hctx
;
1908 return blk_rq_pos(rqa
) > blk_rq_pos(rqb
);
1911 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1915 if (list_empty(&plug
->mq_list
))
1917 list_splice_init(&plug
->mq_list
, &list
);
1919 if (plug
->rq_count
> 2 && plug
->multiple_queues
)
1920 list_sort(NULL
, &list
, plug_rq_cmp
);
1925 struct list_head rq_list
;
1926 struct request
*rq
, *head_rq
= list_entry_rq(list
.next
);
1927 struct list_head
*pos
= &head_rq
->queuelist
; /* skip first */
1928 struct blk_mq_hw_ctx
*this_hctx
= head_rq
->mq_hctx
;
1929 struct blk_mq_ctx
*this_ctx
= head_rq
->mq_ctx
;
1930 unsigned int depth
= 1;
1932 list_for_each_continue(pos
, &list
) {
1933 rq
= list_entry_rq(pos
);
1935 if (rq
->mq_hctx
!= this_hctx
|| rq
->mq_ctx
!= this_ctx
)
1940 list_cut_before(&rq_list
, &list
, pos
);
1941 trace_block_unplug(head_rq
->q
, depth
, !from_schedule
);
1942 blk_mq_sched_insert_requests(this_hctx
, this_ctx
, &rq_list
,
1944 } while(!list_empty(&list
));
1947 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
,
1948 unsigned int nr_segs
)
1952 if (bio
->bi_opf
& REQ_RAHEAD
)
1953 rq
->cmd_flags
|= REQ_FAILFAST_MASK
;
1955 rq
->__sector
= bio
->bi_iter
.bi_sector
;
1956 rq
->write_hint
= bio
->bi_write_hint
;
1957 blk_rq_bio_prep(rq
, bio
, nr_segs
);
1959 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
1960 err
= blk_crypto_rq_bio_prep(rq
, bio
, GFP_NOIO
);
1963 blk_account_io_start(rq
);
1966 static blk_status_t
__blk_mq_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1968 blk_qc_t
*cookie
, bool last
)
1970 struct request_queue
*q
= rq
->q
;
1971 struct blk_mq_queue_data bd
= {
1975 blk_qc_t new_cookie
;
1978 new_cookie
= request_to_qc_t(hctx
, rq
);
1981 * For OK queue, we are done. For error, caller may kill it.
1982 * Any other error (busy), just add it to our list as we
1983 * previously would have done.
1985 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1988 blk_mq_update_dispatch_busy(hctx
, false);
1989 *cookie
= new_cookie
;
1991 case BLK_STS_RESOURCE
:
1992 case BLK_STS_DEV_RESOURCE
:
1993 blk_mq_update_dispatch_busy(hctx
, true);
1994 __blk_mq_requeue_request(rq
);
1997 blk_mq_update_dispatch_busy(hctx
, false);
1998 *cookie
= BLK_QC_T_NONE
;
2005 static blk_status_t
__blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
2008 bool bypass_insert
, bool last
)
2010 struct request_queue
*q
= rq
->q
;
2011 bool run_queue
= true;
2015 * RCU or SRCU read lock is needed before checking quiesced flag.
2017 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2018 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2019 * and avoid driver to try to dispatch again.
2021 if (blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(q
)) {
2023 bypass_insert
= false;
2027 if (q
->elevator
&& !bypass_insert
)
2030 budget_token
= blk_mq_get_dispatch_budget(q
);
2031 if (budget_token
< 0)
2034 blk_mq_set_rq_budget_token(rq
, budget_token
);
2036 if (!blk_mq_get_driver_tag(rq
)) {
2037 blk_mq_put_dispatch_budget(q
, budget_token
);
2041 return __blk_mq_issue_directly(hctx
, rq
, cookie
, last
);
2044 return BLK_STS_RESOURCE
;
2046 blk_mq_sched_insert_request(rq
, false, run_queue
, false);
2052 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2053 * @hctx: Pointer of the associated hardware queue.
2054 * @rq: Pointer to request to be sent.
2055 * @cookie: Request queue cookie.
2057 * If the device has enough resources to accept a new request now, send the
2058 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2059 * we can try send it another time in the future. Requests inserted at this
2060 * queue have higher priority.
2062 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
2063 struct request
*rq
, blk_qc_t
*cookie
)
2068 might_sleep_if(hctx
->flags
& BLK_MQ_F_BLOCKING
);
2070 hctx_lock(hctx
, &srcu_idx
);
2072 ret
= __blk_mq_try_issue_directly(hctx
, rq
, cookie
, false, true);
2073 if (ret
== BLK_STS_RESOURCE
|| ret
== BLK_STS_DEV_RESOURCE
)
2074 blk_mq_request_bypass_insert(rq
, false, true);
2075 else if (ret
!= BLK_STS_OK
)
2076 blk_mq_end_request(rq
, ret
);
2078 hctx_unlock(hctx
, srcu_idx
);
2081 blk_status_t
blk_mq_request_issue_directly(struct request
*rq
, bool last
)
2085 blk_qc_t unused_cookie
;
2086 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
2088 hctx_lock(hctx
, &srcu_idx
);
2089 ret
= __blk_mq_try_issue_directly(hctx
, rq
, &unused_cookie
, true, last
);
2090 hctx_unlock(hctx
, srcu_idx
);
2095 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx
*hctx
,
2096 struct list_head
*list
)
2101 while (!list_empty(list
)) {
2103 struct request
*rq
= list_first_entry(list
, struct request
,
2106 list_del_init(&rq
->queuelist
);
2107 ret
= blk_mq_request_issue_directly(rq
, list_empty(list
));
2108 if (ret
!= BLK_STS_OK
) {
2109 if (ret
== BLK_STS_RESOURCE
||
2110 ret
== BLK_STS_DEV_RESOURCE
) {
2111 blk_mq_request_bypass_insert(rq
, false,
2115 blk_mq_end_request(rq
, ret
);
2122 * If we didn't flush the entire list, we could have told
2123 * the driver there was more coming, but that turned out to
2126 if ((!list_empty(list
) || errors
) &&
2127 hctx
->queue
->mq_ops
->commit_rqs
&& queued
)
2128 hctx
->queue
->mq_ops
->commit_rqs(hctx
);
2131 static void blk_add_rq_to_plug(struct blk_plug
*plug
, struct request
*rq
)
2133 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
2135 if (!plug
->multiple_queues
&& !list_is_singular(&plug
->mq_list
)) {
2136 struct request
*tmp
;
2138 tmp
= list_first_entry(&plug
->mq_list
, struct request
,
2140 if (tmp
->q
!= rq
->q
)
2141 plug
->multiple_queues
= true;
2146 * Allow 4x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
2147 * queues. This is important for md arrays to benefit from merging
2150 static inline unsigned short blk_plug_max_rq_count(struct blk_plug
*plug
)
2152 if (plug
->multiple_queues
)
2153 return BLK_MAX_REQUEST_COUNT
* 4;
2154 return BLK_MAX_REQUEST_COUNT
;
2158 * blk_mq_submit_bio - Create and send a request to block device.
2159 * @bio: Bio pointer.
2161 * Builds up a request structure from @q and @bio and send to the device. The
2162 * request may not be queued directly to hardware if:
2163 * * This request can be merged with another one
2164 * * We want to place request at plug queue for possible future merging
2165 * * There is an IO scheduler active at this queue
2167 * It will not queue the request if there is an error with the bio, or at the
2170 * Returns: Request queue cookie.
2172 blk_qc_t
blk_mq_submit_bio(struct bio
*bio
)
2174 struct request_queue
*q
= bio
->bi_bdev
->bd_disk
->queue
;
2175 const int is_sync
= op_is_sync(bio
->bi_opf
);
2176 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
2177 struct blk_mq_alloc_data data
= {
2181 struct blk_plug
*plug
;
2182 struct request
*same_queue_rq
= NULL
;
2183 unsigned int nr_segs
;
2188 blk_queue_bounce(q
, &bio
);
2189 __blk_queue_split(&bio
, &nr_segs
);
2191 if (!bio_integrity_prep(bio
))
2194 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
2195 blk_attempt_plug_merge(q
, bio
, nr_segs
, &same_queue_rq
))
2198 if (blk_mq_sched_bio_merge(q
, bio
, nr_segs
))
2201 rq_qos_throttle(q
, bio
);
2203 hipri
= bio
->bi_opf
& REQ_HIPRI
;
2205 data
.cmd_flags
= bio
->bi_opf
;
2206 rq
= __blk_mq_alloc_request(&data
);
2207 if (unlikely(!rq
)) {
2208 rq_qos_cleanup(q
, bio
);
2209 if (bio
->bi_opf
& REQ_NOWAIT
)
2210 bio_wouldblock_error(bio
);
2214 trace_block_getrq(bio
);
2216 rq_qos_track(q
, rq
, bio
);
2218 cookie
= request_to_qc_t(data
.hctx
, rq
);
2220 blk_mq_bio_to_request(rq
, bio
, nr_segs
);
2222 ret
= blk_crypto_init_request(rq
);
2223 if (ret
!= BLK_STS_OK
) {
2224 bio
->bi_status
= ret
;
2226 blk_mq_free_request(rq
);
2227 return BLK_QC_T_NONE
;
2230 plug
= blk_mq_plug(q
, bio
);
2231 if (unlikely(is_flush_fua
)) {
2232 /* Bypass scheduler for flush requests */
2233 blk_insert_flush(rq
);
2234 blk_mq_run_hw_queue(data
.hctx
, true);
2235 } else if (plug
&& (q
->nr_hw_queues
== 1 ||
2236 blk_mq_is_sbitmap_shared(rq
->mq_hctx
->flags
) ||
2237 q
->mq_ops
->commit_rqs
|| !blk_queue_nonrot(q
))) {
2239 * Use plugging if we have a ->commit_rqs() hook as well, as
2240 * we know the driver uses bd->last in a smart fashion.
2242 * Use normal plugging if this disk is slow HDD, as sequential
2243 * IO may benefit a lot from plug merging.
2245 unsigned int request_count
= plug
->rq_count
;
2246 struct request
*last
= NULL
;
2249 trace_block_plug(q
);
2251 last
= list_entry_rq(plug
->mq_list
.prev
);
2253 if (request_count
>= blk_plug_max_rq_count(plug
) || (last
&&
2254 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
2255 blk_flush_plug_list(plug
, false);
2256 trace_block_plug(q
);
2259 blk_add_rq_to_plug(plug
, rq
);
2260 } else if (q
->elevator
) {
2261 /* Insert the request at the IO scheduler queue */
2262 blk_mq_sched_insert_request(rq
, false, true, true);
2263 } else if (plug
&& !blk_queue_nomerges(q
)) {
2265 * We do limited plugging. If the bio can be merged, do that.
2266 * Otherwise the existing request in the plug list will be
2267 * issued. So the plug list will have one request at most
2268 * The plug list might get flushed before this. If that happens,
2269 * the plug list is empty, and same_queue_rq is invalid.
2271 if (list_empty(&plug
->mq_list
))
2272 same_queue_rq
= NULL
;
2273 if (same_queue_rq
) {
2274 list_del_init(&same_queue_rq
->queuelist
);
2277 blk_add_rq_to_plug(plug
, rq
);
2278 trace_block_plug(q
);
2280 if (same_queue_rq
) {
2281 data
.hctx
= same_queue_rq
->mq_hctx
;
2282 trace_block_unplug(q
, 1, true);
2283 blk_mq_try_issue_directly(data
.hctx
, same_queue_rq
,
2286 } else if ((q
->nr_hw_queues
> 1 && is_sync
) ||
2287 !data
.hctx
->dispatch_busy
) {
2289 * There is no scheduler and we can try to send directly
2292 blk_mq_try_issue_directly(data
.hctx
, rq
, &cookie
);
2295 blk_mq_sched_insert_request(rq
, false, true, true);
2299 return BLK_QC_T_NONE
;
2303 return BLK_QC_T_NONE
;
2306 static size_t order_to_size(unsigned int order
)
2308 return (size_t)PAGE_SIZE
<< order
;
2311 /* called before freeing request pool in @tags */
2312 static void blk_mq_clear_rq_mapping(struct blk_mq_tag_set
*set
,
2313 struct blk_mq_tags
*tags
, unsigned int hctx_idx
)
2315 struct blk_mq_tags
*drv_tags
= set
->tags
[hctx_idx
];
2317 unsigned long flags
;
2319 list_for_each_entry(page
, &tags
->page_list
, lru
) {
2320 unsigned long start
= (unsigned long)page_address(page
);
2321 unsigned long end
= start
+ order_to_size(page
->private);
2324 for (i
= 0; i
< set
->queue_depth
; i
++) {
2325 struct request
*rq
= drv_tags
->rqs
[i
];
2326 unsigned long rq_addr
= (unsigned long)rq
;
2328 if (rq_addr
>= start
&& rq_addr
< end
) {
2329 WARN_ON_ONCE(refcount_read(&rq
->ref
) != 0);
2330 cmpxchg(&drv_tags
->rqs
[i
], rq
, NULL
);
2336 * Wait until all pending iteration is done.
2338 * Request reference is cleared and it is guaranteed to be observed
2339 * after the ->lock is released.
2341 spin_lock_irqsave(&drv_tags
->lock
, flags
);
2342 spin_unlock_irqrestore(&drv_tags
->lock
, flags
);
2345 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
2346 unsigned int hctx_idx
)
2350 if (tags
->rqs
&& set
->ops
->exit_request
) {
2353 for (i
= 0; i
< tags
->nr_tags
; i
++) {
2354 struct request
*rq
= tags
->static_rqs
[i
];
2358 set
->ops
->exit_request(set
, rq
, hctx_idx
);
2359 tags
->static_rqs
[i
] = NULL
;
2363 blk_mq_clear_rq_mapping(set
, tags
, hctx_idx
);
2365 while (!list_empty(&tags
->page_list
)) {
2366 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
2367 list_del_init(&page
->lru
);
2369 * Remove kmemleak object previously allocated in
2370 * blk_mq_alloc_rqs().
2372 kmemleak_free(page_address(page
));
2373 __free_pages(page
, page
->private);
2377 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
, unsigned int flags
)
2381 kfree(tags
->static_rqs
);
2382 tags
->static_rqs
= NULL
;
2384 blk_mq_free_tags(tags
, flags
);
2387 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
2388 unsigned int hctx_idx
,
2389 unsigned int nr_tags
,
2390 unsigned int reserved_tags
,
2393 struct blk_mq_tags
*tags
;
2396 node
= blk_mq_hw_queue_to_node(&set
->map
[HCTX_TYPE_DEFAULT
], hctx_idx
);
2397 if (node
== NUMA_NO_NODE
)
2398 node
= set
->numa_node
;
2400 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
, flags
);
2404 tags
->rqs
= kcalloc_node(nr_tags
, sizeof(struct request
*),
2405 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
2408 blk_mq_free_tags(tags
, flags
);
2412 tags
->static_rqs
= kcalloc_node(nr_tags
, sizeof(struct request
*),
2413 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
2415 if (!tags
->static_rqs
) {
2417 blk_mq_free_tags(tags
, flags
);
2424 static int blk_mq_init_request(struct blk_mq_tag_set
*set
, struct request
*rq
,
2425 unsigned int hctx_idx
, int node
)
2429 if (set
->ops
->init_request
) {
2430 ret
= set
->ops
->init_request(set
, rq
, hctx_idx
, node
);
2435 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
2439 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
2440 unsigned int hctx_idx
, unsigned int depth
)
2442 unsigned int i
, j
, entries_per_page
, max_order
= 4;
2443 size_t rq_size
, left
;
2446 node
= blk_mq_hw_queue_to_node(&set
->map
[HCTX_TYPE_DEFAULT
], hctx_idx
);
2447 if (node
== NUMA_NO_NODE
)
2448 node
= set
->numa_node
;
2450 INIT_LIST_HEAD(&tags
->page_list
);
2453 * rq_size is the size of the request plus driver payload, rounded
2454 * to the cacheline size
2456 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
2458 left
= rq_size
* depth
;
2460 for (i
= 0; i
< depth
; ) {
2461 int this_order
= max_order
;
2466 while (this_order
&& left
< order_to_size(this_order
- 1))
2470 page
= alloc_pages_node(node
,
2471 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
2477 if (order_to_size(this_order
) < rq_size
)
2484 page
->private = this_order
;
2485 list_add_tail(&page
->lru
, &tags
->page_list
);
2487 p
= page_address(page
);
2489 * Allow kmemleak to scan these pages as they contain pointers
2490 * to additional allocations like via ops->init_request().
2492 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
2493 entries_per_page
= order_to_size(this_order
) / rq_size
;
2494 to_do
= min(entries_per_page
, depth
- i
);
2495 left
-= to_do
* rq_size
;
2496 for (j
= 0; j
< to_do
; j
++) {
2497 struct request
*rq
= p
;
2499 tags
->static_rqs
[i
] = rq
;
2500 if (blk_mq_init_request(set
, rq
, hctx_idx
, node
)) {
2501 tags
->static_rqs
[i
] = NULL
;
2512 blk_mq_free_rqs(set
, tags
, hctx_idx
);
2516 struct rq_iter_data
{
2517 struct blk_mq_hw_ctx
*hctx
;
2521 static bool blk_mq_has_request(struct request
*rq
, void *data
, bool reserved
)
2523 struct rq_iter_data
*iter_data
= data
;
2525 if (rq
->mq_hctx
!= iter_data
->hctx
)
2527 iter_data
->has_rq
= true;
2531 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx
*hctx
)
2533 struct blk_mq_tags
*tags
= hctx
->sched_tags
?
2534 hctx
->sched_tags
: hctx
->tags
;
2535 struct rq_iter_data data
= {
2539 blk_mq_all_tag_iter(tags
, blk_mq_has_request
, &data
);
2543 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu
,
2544 struct blk_mq_hw_ctx
*hctx
)
2546 if (cpumask_next_and(-1, hctx
->cpumask
, cpu_online_mask
) != cpu
)
2548 if (cpumask_next_and(cpu
, hctx
->cpumask
, cpu_online_mask
) < nr_cpu_ids
)
2553 static int blk_mq_hctx_notify_offline(unsigned int cpu
, struct hlist_node
*node
)
2555 struct blk_mq_hw_ctx
*hctx
= hlist_entry_safe(node
,
2556 struct blk_mq_hw_ctx
, cpuhp_online
);
2558 if (!cpumask_test_cpu(cpu
, hctx
->cpumask
) ||
2559 !blk_mq_last_cpu_in_hctx(cpu
, hctx
))
2563 * Prevent new request from being allocated on the current hctx.
2565 * The smp_mb__after_atomic() Pairs with the implied barrier in
2566 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
2567 * seen once we return from the tag allocator.
2569 set_bit(BLK_MQ_S_INACTIVE
, &hctx
->state
);
2570 smp_mb__after_atomic();
2573 * Try to grab a reference to the queue and wait for any outstanding
2574 * requests. If we could not grab a reference the queue has been
2575 * frozen and there are no requests.
2577 if (percpu_ref_tryget(&hctx
->queue
->q_usage_counter
)) {
2578 while (blk_mq_hctx_has_requests(hctx
))
2580 percpu_ref_put(&hctx
->queue
->q_usage_counter
);
2586 static int blk_mq_hctx_notify_online(unsigned int cpu
, struct hlist_node
*node
)
2588 struct blk_mq_hw_ctx
*hctx
= hlist_entry_safe(node
,
2589 struct blk_mq_hw_ctx
, cpuhp_online
);
2591 if (cpumask_test_cpu(cpu
, hctx
->cpumask
))
2592 clear_bit(BLK_MQ_S_INACTIVE
, &hctx
->state
);
2597 * 'cpu' is going away. splice any existing rq_list entries from this
2598 * software queue to the hw queue dispatch list, and ensure that it
2601 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
2603 struct blk_mq_hw_ctx
*hctx
;
2604 struct blk_mq_ctx
*ctx
;
2606 enum hctx_type type
;
2608 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
2609 if (!cpumask_test_cpu(cpu
, hctx
->cpumask
))
2612 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
2615 spin_lock(&ctx
->lock
);
2616 if (!list_empty(&ctx
->rq_lists
[type
])) {
2617 list_splice_init(&ctx
->rq_lists
[type
], &tmp
);
2618 blk_mq_hctx_clear_pending(hctx
, ctx
);
2620 spin_unlock(&ctx
->lock
);
2622 if (list_empty(&tmp
))
2625 spin_lock(&hctx
->lock
);
2626 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
2627 spin_unlock(&hctx
->lock
);
2629 blk_mq_run_hw_queue(hctx
, true);
2633 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
2635 if (!(hctx
->flags
& BLK_MQ_F_STACKING
))
2636 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE
,
2637 &hctx
->cpuhp_online
);
2638 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
2643 * Before freeing hw queue, clearing the flush request reference in
2644 * tags->rqs[] for avoiding potential UAF.
2646 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags
*tags
,
2647 unsigned int queue_depth
, struct request
*flush_rq
)
2650 unsigned long flags
;
2652 /* The hw queue may not be mapped yet */
2656 WARN_ON_ONCE(refcount_read(&flush_rq
->ref
) != 0);
2658 for (i
= 0; i
< queue_depth
; i
++)
2659 cmpxchg(&tags
->rqs
[i
], flush_rq
, NULL
);
2662 * Wait until all pending iteration is done.
2664 * Request reference is cleared and it is guaranteed to be observed
2665 * after the ->lock is released.
2667 spin_lock_irqsave(&tags
->lock
, flags
);
2668 spin_unlock_irqrestore(&tags
->lock
, flags
);
2671 /* hctx->ctxs will be freed in queue's release handler */
2672 static void blk_mq_exit_hctx(struct request_queue
*q
,
2673 struct blk_mq_tag_set
*set
,
2674 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
2676 struct request
*flush_rq
= hctx
->fq
->flush_rq
;
2678 if (blk_mq_hw_queue_mapped(hctx
))
2679 blk_mq_tag_idle(hctx
);
2681 blk_mq_clear_flush_rq_mapping(set
->tags
[hctx_idx
],
2682 set
->queue_depth
, flush_rq
);
2683 if (set
->ops
->exit_request
)
2684 set
->ops
->exit_request(set
, flush_rq
, hctx_idx
);
2686 if (set
->ops
->exit_hctx
)
2687 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2689 blk_mq_remove_cpuhp(hctx
);
2691 spin_lock(&q
->unused_hctx_lock
);
2692 list_add(&hctx
->hctx_list
, &q
->unused_hctx_list
);
2693 spin_unlock(&q
->unused_hctx_lock
);
2696 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
2697 struct blk_mq_tag_set
*set
, int nr_queue
)
2699 struct blk_mq_hw_ctx
*hctx
;
2702 queue_for_each_hw_ctx(q
, hctx
, i
) {
2705 blk_mq_debugfs_unregister_hctx(hctx
);
2706 blk_mq_exit_hctx(q
, set
, hctx
, i
);
2710 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set
*tag_set
)
2712 int hw_ctx_size
= sizeof(struct blk_mq_hw_ctx
);
2714 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx
, srcu
),
2715 __alignof__(struct blk_mq_hw_ctx
)) !=
2716 sizeof(struct blk_mq_hw_ctx
));
2718 if (tag_set
->flags
& BLK_MQ_F_BLOCKING
)
2719 hw_ctx_size
+= sizeof(struct srcu_struct
);
2724 static int blk_mq_init_hctx(struct request_queue
*q
,
2725 struct blk_mq_tag_set
*set
,
2726 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
2728 hctx
->queue_num
= hctx_idx
;
2730 if (!(hctx
->flags
& BLK_MQ_F_STACKING
))
2731 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE
,
2732 &hctx
->cpuhp_online
);
2733 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
2735 hctx
->tags
= set
->tags
[hctx_idx
];
2737 if (set
->ops
->init_hctx
&&
2738 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
2739 goto unregister_cpu_notifier
;
2741 if (blk_mq_init_request(set
, hctx
->fq
->flush_rq
, hctx_idx
,
2747 if (set
->ops
->exit_hctx
)
2748 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2749 unregister_cpu_notifier
:
2750 blk_mq_remove_cpuhp(hctx
);
2754 static struct blk_mq_hw_ctx
*
2755 blk_mq_alloc_hctx(struct request_queue
*q
, struct blk_mq_tag_set
*set
,
2758 struct blk_mq_hw_ctx
*hctx
;
2759 gfp_t gfp
= GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
;
2761 hctx
= kzalloc_node(blk_mq_hw_ctx_size(set
), gfp
, node
);
2763 goto fail_alloc_hctx
;
2765 if (!zalloc_cpumask_var_node(&hctx
->cpumask
, gfp
, node
))
2768 atomic_set(&hctx
->nr_active
, 0);
2769 if (node
== NUMA_NO_NODE
)
2770 node
= set
->numa_node
;
2771 hctx
->numa_node
= node
;
2773 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
2774 spin_lock_init(&hctx
->lock
);
2775 INIT_LIST_HEAD(&hctx
->dispatch
);
2777 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_QUEUE_SHARED
;
2779 INIT_LIST_HEAD(&hctx
->hctx_list
);
2782 * Allocate space for all possible cpus to avoid allocation at
2785 hctx
->ctxs
= kmalloc_array_node(nr_cpu_ids
, sizeof(void *),
2790 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8),
2791 gfp
, node
, false, false))
2795 spin_lock_init(&hctx
->dispatch_wait_lock
);
2796 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
2797 INIT_LIST_HEAD(&hctx
->dispatch_wait
.entry
);
2799 hctx
->fq
= blk_alloc_flush_queue(hctx
->numa_node
, set
->cmd_size
, gfp
);
2803 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
2804 init_srcu_struct(hctx
->srcu
);
2805 blk_mq_hctx_kobj_init(hctx
);
2810 sbitmap_free(&hctx
->ctx_map
);
2814 free_cpumask_var(hctx
->cpumask
);
2821 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
2822 unsigned int nr_hw_queues
)
2824 struct blk_mq_tag_set
*set
= q
->tag_set
;
2827 for_each_possible_cpu(i
) {
2828 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2829 struct blk_mq_hw_ctx
*hctx
;
2833 spin_lock_init(&__ctx
->lock
);
2834 for (k
= HCTX_TYPE_DEFAULT
; k
< HCTX_MAX_TYPES
; k
++)
2835 INIT_LIST_HEAD(&__ctx
->rq_lists
[k
]);
2840 * Set local node, IFF we have more than one hw queue. If
2841 * not, we remain on the home node of the device
2843 for (j
= 0; j
< set
->nr_maps
; j
++) {
2844 hctx
= blk_mq_map_queue_type(q
, j
, i
);
2845 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
2846 hctx
->numa_node
= cpu_to_node(i
);
2851 static bool __blk_mq_alloc_map_and_request(struct blk_mq_tag_set
*set
,
2854 unsigned int flags
= set
->flags
;
2857 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
2858 set
->queue_depth
, set
->reserved_tags
, flags
);
2859 if (!set
->tags
[hctx_idx
])
2862 ret
= blk_mq_alloc_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
,
2867 blk_mq_free_rq_map(set
->tags
[hctx_idx
], flags
);
2868 set
->tags
[hctx_idx
] = NULL
;
2872 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set
*set
,
2873 unsigned int hctx_idx
)
2875 unsigned int flags
= set
->flags
;
2877 if (set
->tags
&& set
->tags
[hctx_idx
]) {
2878 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
2879 blk_mq_free_rq_map(set
->tags
[hctx_idx
], flags
);
2880 set
->tags
[hctx_idx
] = NULL
;
2884 static void blk_mq_map_swqueue(struct request_queue
*q
)
2886 unsigned int i
, j
, hctx_idx
;
2887 struct blk_mq_hw_ctx
*hctx
;
2888 struct blk_mq_ctx
*ctx
;
2889 struct blk_mq_tag_set
*set
= q
->tag_set
;
2891 queue_for_each_hw_ctx(q
, hctx
, i
) {
2892 cpumask_clear(hctx
->cpumask
);
2894 hctx
->dispatch_from
= NULL
;
2898 * Map software to hardware queues.
2900 * If the cpu isn't present, the cpu is mapped to first hctx.
2902 for_each_possible_cpu(i
) {
2904 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2905 for (j
= 0; j
< set
->nr_maps
; j
++) {
2906 if (!set
->map
[j
].nr_queues
) {
2907 ctx
->hctxs
[j
] = blk_mq_map_queue_type(q
,
2908 HCTX_TYPE_DEFAULT
, i
);
2911 hctx_idx
= set
->map
[j
].mq_map
[i
];
2912 /* unmapped hw queue can be remapped after CPU topo changed */
2913 if (!set
->tags
[hctx_idx
] &&
2914 !__blk_mq_alloc_map_and_request(set
, hctx_idx
)) {
2916 * If tags initialization fail for some hctx,
2917 * that hctx won't be brought online. In this
2918 * case, remap the current ctx to hctx[0] which
2919 * is guaranteed to always have tags allocated
2921 set
->map
[j
].mq_map
[i
] = 0;
2924 hctx
= blk_mq_map_queue_type(q
, j
, i
);
2925 ctx
->hctxs
[j
] = hctx
;
2927 * If the CPU is already set in the mask, then we've
2928 * mapped this one already. This can happen if
2929 * devices share queues across queue maps.
2931 if (cpumask_test_cpu(i
, hctx
->cpumask
))
2934 cpumask_set_cpu(i
, hctx
->cpumask
);
2936 ctx
->index_hw
[hctx
->type
] = hctx
->nr_ctx
;
2937 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2940 * If the nr_ctx type overflows, we have exceeded the
2941 * amount of sw queues we can support.
2943 BUG_ON(!hctx
->nr_ctx
);
2946 for (; j
< HCTX_MAX_TYPES
; j
++)
2947 ctx
->hctxs
[j
] = blk_mq_map_queue_type(q
,
2948 HCTX_TYPE_DEFAULT
, i
);
2951 queue_for_each_hw_ctx(q
, hctx
, i
) {
2953 * If no software queues are mapped to this hardware queue,
2954 * disable it and free the request entries.
2956 if (!hctx
->nr_ctx
) {
2957 /* Never unmap queue 0. We need it as a
2958 * fallback in case of a new remap fails
2961 if (i
&& set
->tags
[i
])
2962 blk_mq_free_map_and_requests(set
, i
);
2968 hctx
->tags
= set
->tags
[i
];
2969 WARN_ON(!hctx
->tags
);
2972 * Set the map size to the number of mapped software queues.
2973 * This is more accurate and more efficient than looping
2974 * over all possibly mapped software queues.
2976 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2979 * Initialize batch roundrobin counts
2981 hctx
->next_cpu
= blk_mq_first_mapped_cpu(hctx
);
2982 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2987 * Caller needs to ensure that we're either frozen/quiesced, or that
2988 * the queue isn't live yet.
2990 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2992 struct blk_mq_hw_ctx
*hctx
;
2995 queue_for_each_hw_ctx(q
, hctx
, i
) {
2997 hctx
->flags
|= BLK_MQ_F_TAG_QUEUE_SHARED
;
2999 blk_mq_tag_idle(hctx
);
3000 hctx
->flags
&= ~BLK_MQ_F_TAG_QUEUE_SHARED
;
3005 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set
*set
,
3008 struct request_queue
*q
;
3010 lockdep_assert_held(&set
->tag_list_lock
);
3012 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3013 blk_mq_freeze_queue(q
);
3014 queue_set_hctx_shared(q
, shared
);
3015 blk_mq_unfreeze_queue(q
);
3019 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
3021 struct blk_mq_tag_set
*set
= q
->tag_set
;
3023 mutex_lock(&set
->tag_list_lock
);
3024 list_del(&q
->tag_set_list
);
3025 if (list_is_singular(&set
->tag_list
)) {
3026 /* just transitioned to unshared */
3027 set
->flags
&= ~BLK_MQ_F_TAG_QUEUE_SHARED
;
3028 /* update existing queue */
3029 blk_mq_update_tag_set_shared(set
, false);
3031 mutex_unlock(&set
->tag_list_lock
);
3032 INIT_LIST_HEAD(&q
->tag_set_list
);
3035 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
3036 struct request_queue
*q
)
3038 mutex_lock(&set
->tag_list_lock
);
3041 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3043 if (!list_empty(&set
->tag_list
) &&
3044 !(set
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
)) {
3045 set
->flags
|= BLK_MQ_F_TAG_QUEUE_SHARED
;
3046 /* update existing queue */
3047 blk_mq_update_tag_set_shared(set
, true);
3049 if (set
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
)
3050 queue_set_hctx_shared(q
, true);
3051 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
3053 mutex_unlock(&set
->tag_list_lock
);
3056 /* All allocations will be freed in release handler of q->mq_kobj */
3057 static int blk_mq_alloc_ctxs(struct request_queue
*q
)
3059 struct blk_mq_ctxs
*ctxs
;
3062 ctxs
= kzalloc(sizeof(*ctxs
), GFP_KERNEL
);
3066 ctxs
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
3067 if (!ctxs
->queue_ctx
)
3070 for_each_possible_cpu(cpu
) {
3071 struct blk_mq_ctx
*ctx
= per_cpu_ptr(ctxs
->queue_ctx
, cpu
);
3075 q
->mq_kobj
= &ctxs
->kobj
;
3076 q
->queue_ctx
= ctxs
->queue_ctx
;
3085 * It is the actual release handler for mq, but we do it from
3086 * request queue's release handler for avoiding use-after-free
3087 * and headache because q->mq_kobj shouldn't have been introduced,
3088 * but we can't group ctx/kctx kobj without it.
3090 void blk_mq_release(struct request_queue
*q
)
3092 struct blk_mq_hw_ctx
*hctx
, *next
;
3095 queue_for_each_hw_ctx(q
, hctx
, i
)
3096 WARN_ON_ONCE(hctx
&& list_empty(&hctx
->hctx_list
));
3098 /* all hctx are in .unused_hctx_list now */
3099 list_for_each_entry_safe(hctx
, next
, &q
->unused_hctx_list
, hctx_list
) {
3100 list_del_init(&hctx
->hctx_list
);
3101 kobject_put(&hctx
->kobj
);
3104 kfree(q
->queue_hw_ctx
);
3107 * release .mq_kobj and sw queue's kobject now because
3108 * both share lifetime with request queue.
3110 blk_mq_sysfs_deinit(q
);
3113 static struct request_queue
*blk_mq_init_queue_data(struct blk_mq_tag_set
*set
,
3116 struct request_queue
*q
;
3119 q
= blk_alloc_queue(set
->numa_node
);
3121 return ERR_PTR(-ENOMEM
);
3122 q
->queuedata
= queuedata
;
3123 ret
= blk_mq_init_allocated_queue(set
, q
);
3125 blk_cleanup_queue(q
);
3126 return ERR_PTR(ret
);
3131 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
3133 return blk_mq_init_queue_data(set
, NULL
);
3135 EXPORT_SYMBOL(blk_mq_init_queue
);
3137 struct gendisk
*__blk_mq_alloc_disk(struct blk_mq_tag_set
*set
, void *queuedata
,
3138 struct lock_class_key
*lkclass
)
3140 struct request_queue
*q
;
3141 struct gendisk
*disk
;
3143 q
= blk_mq_init_queue_data(set
, queuedata
);
3147 disk
= __alloc_disk_node(q
, set
->numa_node
, lkclass
);
3149 blk_cleanup_queue(q
);
3150 return ERR_PTR(-ENOMEM
);
3154 EXPORT_SYMBOL(__blk_mq_alloc_disk
);
3156 static struct blk_mq_hw_ctx
*blk_mq_alloc_and_init_hctx(
3157 struct blk_mq_tag_set
*set
, struct request_queue
*q
,
3158 int hctx_idx
, int node
)
3160 struct blk_mq_hw_ctx
*hctx
= NULL
, *tmp
;
3162 /* reuse dead hctx first */
3163 spin_lock(&q
->unused_hctx_lock
);
3164 list_for_each_entry(tmp
, &q
->unused_hctx_list
, hctx_list
) {
3165 if (tmp
->numa_node
== node
) {
3171 list_del_init(&hctx
->hctx_list
);
3172 spin_unlock(&q
->unused_hctx_lock
);
3175 hctx
= blk_mq_alloc_hctx(q
, set
, node
);
3179 if (blk_mq_init_hctx(q
, set
, hctx
, hctx_idx
))
3185 kobject_put(&hctx
->kobj
);
3190 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
3191 struct request_queue
*q
)
3194 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
3196 if (q
->nr_hw_queues
< set
->nr_hw_queues
) {
3197 struct blk_mq_hw_ctx
**new_hctxs
;
3199 new_hctxs
= kcalloc_node(set
->nr_hw_queues
,
3200 sizeof(*new_hctxs
), GFP_KERNEL
,
3205 memcpy(new_hctxs
, hctxs
, q
->nr_hw_queues
*
3207 q
->queue_hw_ctx
= new_hctxs
;
3212 /* protect against switching io scheduler */
3213 mutex_lock(&q
->sysfs_lock
);
3214 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
3216 struct blk_mq_hw_ctx
*hctx
;
3218 node
= blk_mq_hw_queue_to_node(&set
->map
[HCTX_TYPE_DEFAULT
], i
);
3220 * If the hw queue has been mapped to another numa node,
3221 * we need to realloc the hctx. If allocation fails, fallback
3222 * to use the previous one.
3224 if (hctxs
[i
] && (hctxs
[i
]->numa_node
== node
))
3227 hctx
= blk_mq_alloc_and_init_hctx(set
, q
, i
, node
);
3230 blk_mq_exit_hctx(q
, set
, hctxs
[i
], i
);
3234 pr_warn("Allocate new hctx on node %d fails,\
3235 fallback to previous one on node %d\n",
3236 node
, hctxs
[i
]->numa_node
);
3242 * Increasing nr_hw_queues fails. Free the newly allocated
3243 * hctxs and keep the previous q->nr_hw_queues.
3245 if (i
!= set
->nr_hw_queues
) {
3246 j
= q
->nr_hw_queues
;
3250 end
= q
->nr_hw_queues
;
3251 q
->nr_hw_queues
= set
->nr_hw_queues
;
3254 for (; j
< end
; j
++) {
3255 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
3259 blk_mq_free_map_and_requests(set
, j
);
3260 blk_mq_exit_hctx(q
, set
, hctx
, j
);
3264 mutex_unlock(&q
->sysfs_lock
);
3267 int blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
3268 struct request_queue
*q
)
3270 /* mark the queue as mq asap */
3271 q
->mq_ops
= set
->ops
;
3273 q
->poll_cb
= blk_stat_alloc_callback(blk_mq_poll_stats_fn
,
3274 blk_mq_poll_stats_bkt
,
3275 BLK_MQ_POLL_STATS_BKTS
, q
);
3279 if (blk_mq_alloc_ctxs(q
))
3282 /* init q->mq_kobj and sw queues' kobjects */
3283 blk_mq_sysfs_init(q
);
3285 INIT_LIST_HEAD(&q
->unused_hctx_list
);
3286 spin_lock_init(&q
->unused_hctx_lock
);
3288 blk_mq_realloc_hw_ctxs(set
, q
);
3289 if (!q
->nr_hw_queues
)
3292 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
3293 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
3297 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
3298 if (set
->nr_maps
> HCTX_TYPE_POLL
&&
3299 set
->map
[HCTX_TYPE_POLL
].nr_queues
)
3300 blk_queue_flag_set(QUEUE_FLAG_POLL
, q
);
3302 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
3303 INIT_LIST_HEAD(&q
->requeue_list
);
3304 spin_lock_init(&q
->requeue_lock
);
3306 q
->nr_requests
= set
->queue_depth
;
3309 * Default to classic polling
3311 q
->poll_nsec
= BLK_MQ_POLL_CLASSIC
;
3313 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
3314 blk_mq_add_queue_tag_set(set
, q
);
3315 blk_mq_map_swqueue(q
);
3319 kfree(q
->queue_hw_ctx
);
3320 q
->nr_hw_queues
= 0;
3321 blk_mq_sysfs_deinit(q
);
3323 blk_stat_free_callback(q
->poll_cb
);
3329 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
3331 /* tags can _not_ be used after returning from blk_mq_exit_queue */
3332 void blk_mq_exit_queue(struct request_queue
*q
)
3334 struct blk_mq_tag_set
*set
= q
->tag_set
;
3336 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
3337 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
3338 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
3339 blk_mq_del_queue_tag_set(q
);
3342 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
3346 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
3347 if (!__blk_mq_alloc_map_and_request(set
, i
))
3356 blk_mq_free_map_and_requests(set
, i
);
3362 * Allocate the request maps associated with this tag_set. Note that this
3363 * may reduce the depth asked for, if memory is tight. set->queue_depth
3364 * will be updated to reflect the allocated depth.
3366 static int blk_mq_alloc_map_and_requests(struct blk_mq_tag_set
*set
)
3371 depth
= set
->queue_depth
;
3373 err
= __blk_mq_alloc_rq_maps(set
);
3377 set
->queue_depth
>>= 1;
3378 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
3382 } while (set
->queue_depth
);
3384 if (!set
->queue_depth
|| err
) {
3385 pr_err("blk-mq: failed to allocate request map\n");
3389 if (depth
!= set
->queue_depth
)
3390 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3391 depth
, set
->queue_depth
);
3396 static int blk_mq_update_queue_map(struct blk_mq_tag_set
*set
)
3399 * blk_mq_map_queues() and multiple .map_queues() implementations
3400 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
3401 * number of hardware queues.
3403 if (set
->nr_maps
== 1)
3404 set
->map
[HCTX_TYPE_DEFAULT
].nr_queues
= set
->nr_hw_queues
;
3406 if (set
->ops
->map_queues
&& !is_kdump_kernel()) {
3410 * transport .map_queues is usually done in the following
3413 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3414 * mask = get_cpu_mask(queue)
3415 * for_each_cpu(cpu, mask)
3416 * set->map[x].mq_map[cpu] = queue;
3419 * When we need to remap, the table has to be cleared for
3420 * killing stale mapping since one CPU may not be mapped
3423 for (i
= 0; i
< set
->nr_maps
; i
++)
3424 blk_mq_clear_mq_map(&set
->map
[i
]);
3426 return set
->ops
->map_queues(set
);
3428 BUG_ON(set
->nr_maps
> 1);
3429 return blk_mq_map_queues(&set
->map
[HCTX_TYPE_DEFAULT
]);
3433 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set
*set
,
3434 int cur_nr_hw_queues
, int new_nr_hw_queues
)
3436 struct blk_mq_tags
**new_tags
;
3438 if (cur_nr_hw_queues
>= new_nr_hw_queues
)
3441 new_tags
= kcalloc_node(new_nr_hw_queues
, sizeof(struct blk_mq_tags
*),
3442 GFP_KERNEL
, set
->numa_node
);
3447 memcpy(new_tags
, set
->tags
, cur_nr_hw_queues
*
3448 sizeof(*set
->tags
));
3450 set
->tags
= new_tags
;
3451 set
->nr_hw_queues
= new_nr_hw_queues
;
3456 static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set
*set
,
3457 int new_nr_hw_queues
)
3459 return blk_mq_realloc_tag_set_tags(set
, 0, new_nr_hw_queues
);
3463 * Alloc a tag set to be associated with one or more request queues.
3464 * May fail with EINVAL for various error conditions. May adjust the
3465 * requested depth down, if it's too large. In that case, the set
3466 * value will be stored in set->queue_depth.
3468 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
3472 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
3474 if (!set
->nr_hw_queues
)
3476 if (!set
->queue_depth
)
3478 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
3481 if (!set
->ops
->queue_rq
)
3484 if (!set
->ops
->get_budget
^ !set
->ops
->put_budget
)
3487 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
3488 pr_info("blk-mq: reduced tag depth to %u\n",
3490 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
3495 else if (set
->nr_maps
> HCTX_MAX_TYPES
)
3499 * If a crashdump is active, then we are potentially in a very
3500 * memory constrained environment. Limit us to 1 queue and
3501 * 64 tags to prevent using too much memory.
3503 if (is_kdump_kernel()) {
3504 set
->nr_hw_queues
= 1;
3506 set
->queue_depth
= min(64U, set
->queue_depth
);
3509 * There is no use for more h/w queues than cpus if we just have
3512 if (set
->nr_maps
== 1 && set
->nr_hw_queues
> nr_cpu_ids
)
3513 set
->nr_hw_queues
= nr_cpu_ids
;
3515 if (blk_mq_alloc_tag_set_tags(set
, set
->nr_hw_queues
) < 0)
3519 for (i
= 0; i
< set
->nr_maps
; i
++) {
3520 set
->map
[i
].mq_map
= kcalloc_node(nr_cpu_ids
,
3521 sizeof(set
->map
[i
].mq_map
[0]),
3522 GFP_KERNEL
, set
->numa_node
);
3523 if (!set
->map
[i
].mq_map
)
3524 goto out_free_mq_map
;
3525 set
->map
[i
].nr_queues
= is_kdump_kernel() ? 1 : set
->nr_hw_queues
;
3528 ret
= blk_mq_update_queue_map(set
);
3530 goto out_free_mq_map
;
3532 ret
= blk_mq_alloc_map_and_requests(set
);
3534 goto out_free_mq_map
;
3536 if (blk_mq_is_sbitmap_shared(set
->flags
)) {
3537 atomic_set(&set
->active_queues_shared_sbitmap
, 0);
3539 if (blk_mq_init_shared_sbitmap(set
)) {
3541 goto out_free_mq_rq_maps
;
3545 mutex_init(&set
->tag_list_lock
);
3546 INIT_LIST_HEAD(&set
->tag_list
);
3550 out_free_mq_rq_maps
:
3551 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
3552 blk_mq_free_map_and_requests(set
, i
);
3554 for (i
= 0; i
< set
->nr_maps
; i
++) {
3555 kfree(set
->map
[i
].mq_map
);
3556 set
->map
[i
].mq_map
= NULL
;
3562 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
3564 /* allocate and initialize a tagset for a simple single-queue device */
3565 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set
*set
,
3566 const struct blk_mq_ops
*ops
, unsigned int queue_depth
,
3567 unsigned int set_flags
)
3569 memset(set
, 0, sizeof(*set
));
3571 set
->nr_hw_queues
= 1;
3573 set
->queue_depth
= queue_depth
;
3574 set
->numa_node
= NUMA_NO_NODE
;
3575 set
->flags
= set_flags
;
3576 return blk_mq_alloc_tag_set(set
);
3578 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set
);
3580 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
3584 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
3585 blk_mq_free_map_and_requests(set
, i
);
3587 if (blk_mq_is_sbitmap_shared(set
->flags
))
3588 blk_mq_exit_shared_sbitmap(set
);
3590 for (j
= 0; j
< set
->nr_maps
; j
++) {
3591 kfree(set
->map
[j
].mq_map
);
3592 set
->map
[j
].mq_map
= NULL
;
3598 EXPORT_SYMBOL(blk_mq_free_tag_set
);
3600 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
3602 struct blk_mq_tag_set
*set
= q
->tag_set
;
3603 struct blk_mq_hw_ctx
*hctx
;
3609 if (q
->nr_requests
== nr
)
3612 blk_mq_freeze_queue(q
);
3613 blk_mq_quiesce_queue(q
);
3616 queue_for_each_hw_ctx(q
, hctx
, i
) {
3620 * If we're using an MQ scheduler, just update the scheduler
3621 * queue depth. This is similar to what the old code would do.
3623 if (!hctx
->sched_tags
) {
3624 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
, nr
,
3626 if (!ret
&& blk_mq_is_sbitmap_shared(set
->flags
))
3627 blk_mq_tag_resize_shared_sbitmap(set
, nr
);
3629 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
3631 if (blk_mq_is_sbitmap_shared(set
->flags
)) {
3632 hctx
->sched_tags
->bitmap_tags
=
3633 &q
->sched_bitmap_tags
;
3634 hctx
->sched_tags
->breserved_tags
=
3635 &q
->sched_breserved_tags
;
3640 if (q
->elevator
&& q
->elevator
->type
->ops
.depth_updated
)
3641 q
->elevator
->type
->ops
.depth_updated(hctx
);
3644 q
->nr_requests
= nr
;
3645 if (q
->elevator
&& blk_mq_is_sbitmap_shared(set
->flags
))
3646 sbitmap_queue_resize(&q
->sched_bitmap_tags
,
3647 nr
- set
->reserved_tags
);
3650 blk_mq_unquiesce_queue(q
);
3651 blk_mq_unfreeze_queue(q
);
3657 * request_queue and elevator_type pair.
3658 * It is just used by __blk_mq_update_nr_hw_queues to cache
3659 * the elevator_type associated with a request_queue.
3661 struct blk_mq_qe_pair
{
3662 struct list_head node
;
3663 struct request_queue
*q
;
3664 struct elevator_type
*type
;
3668 * Cache the elevator_type in qe pair list and switch the
3669 * io scheduler to 'none'
3671 static bool blk_mq_elv_switch_none(struct list_head
*head
,
3672 struct request_queue
*q
)
3674 struct blk_mq_qe_pair
*qe
;
3679 qe
= kmalloc(sizeof(*qe
), GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
);
3683 INIT_LIST_HEAD(&qe
->node
);
3685 qe
->type
= q
->elevator
->type
;
3686 list_add(&qe
->node
, head
);
3688 mutex_lock(&q
->sysfs_lock
);
3690 * After elevator_switch_mq, the previous elevator_queue will be
3691 * released by elevator_release. The reference of the io scheduler
3692 * module get by elevator_get will also be put. So we need to get
3693 * a reference of the io scheduler module here to prevent it to be
3696 __module_get(qe
->type
->elevator_owner
);
3697 elevator_switch_mq(q
, NULL
);
3698 mutex_unlock(&q
->sysfs_lock
);
3703 static void blk_mq_elv_switch_back(struct list_head
*head
,
3704 struct request_queue
*q
)
3706 struct blk_mq_qe_pair
*qe
;
3707 struct elevator_type
*t
= NULL
;
3709 list_for_each_entry(qe
, head
, node
)
3718 list_del(&qe
->node
);
3721 mutex_lock(&q
->sysfs_lock
);
3722 elevator_switch_mq(q
, t
);
3723 mutex_unlock(&q
->sysfs_lock
);
3726 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
,
3729 struct request_queue
*q
;
3731 int prev_nr_hw_queues
;
3733 lockdep_assert_held(&set
->tag_list_lock
);
3735 if (set
->nr_maps
== 1 && nr_hw_queues
> nr_cpu_ids
)
3736 nr_hw_queues
= nr_cpu_ids
;
3737 if (nr_hw_queues
< 1)
3739 if (set
->nr_maps
== 1 && nr_hw_queues
== set
->nr_hw_queues
)
3742 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3743 blk_mq_freeze_queue(q
);
3745 * Switch IO scheduler to 'none', cleaning up the data associated
3746 * with the previous scheduler. We will switch back once we are done
3747 * updating the new sw to hw queue mappings.
3749 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3750 if (!blk_mq_elv_switch_none(&head
, q
))
3753 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3754 blk_mq_debugfs_unregister_hctxs(q
);
3755 blk_mq_sysfs_unregister(q
);
3758 prev_nr_hw_queues
= set
->nr_hw_queues
;
3759 if (blk_mq_realloc_tag_set_tags(set
, set
->nr_hw_queues
, nr_hw_queues
) <
3763 set
->nr_hw_queues
= nr_hw_queues
;
3765 blk_mq_update_queue_map(set
);
3766 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3767 blk_mq_realloc_hw_ctxs(set
, q
);
3768 if (q
->nr_hw_queues
!= set
->nr_hw_queues
) {
3769 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3770 nr_hw_queues
, prev_nr_hw_queues
);
3771 set
->nr_hw_queues
= prev_nr_hw_queues
;
3772 blk_mq_map_queues(&set
->map
[HCTX_TYPE_DEFAULT
]);
3775 blk_mq_map_swqueue(q
);
3779 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3780 blk_mq_sysfs_register(q
);
3781 blk_mq_debugfs_register_hctxs(q
);
3785 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3786 blk_mq_elv_switch_back(&head
, q
);
3788 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3789 blk_mq_unfreeze_queue(q
);
3792 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
3794 mutex_lock(&set
->tag_list_lock
);
3795 __blk_mq_update_nr_hw_queues(set
, nr_hw_queues
);
3796 mutex_unlock(&set
->tag_list_lock
);
3798 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
3800 /* Enable polling stats and return whether they were already enabled. */
3801 static bool blk_poll_stats_enable(struct request_queue
*q
)
3803 if (test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
3804 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS
, q
))
3806 blk_stat_add_callback(q
, q
->poll_cb
);
3810 static void blk_mq_poll_stats_start(struct request_queue
*q
)
3813 * We don't arm the callback if polling stats are not enabled or the
3814 * callback is already active.
3816 if (!test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
3817 blk_stat_is_active(q
->poll_cb
))
3820 blk_stat_activate_msecs(q
->poll_cb
, 100);
3823 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
)
3825 struct request_queue
*q
= cb
->data
;
3828 for (bucket
= 0; bucket
< BLK_MQ_POLL_STATS_BKTS
; bucket
++) {
3829 if (cb
->stat
[bucket
].nr_samples
)
3830 q
->poll_stat
[bucket
] = cb
->stat
[bucket
];
3834 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
3837 unsigned long ret
= 0;
3841 * If stats collection isn't on, don't sleep but turn it on for
3844 if (!blk_poll_stats_enable(q
))
3848 * As an optimistic guess, use half of the mean service time
3849 * for this type of request. We can (and should) make this smarter.
3850 * For instance, if the completion latencies are tight, we can
3851 * get closer than just half the mean. This is especially
3852 * important on devices where the completion latencies are longer
3853 * than ~10 usec. We do use the stats for the relevant IO size
3854 * if available which does lead to better estimates.
3856 bucket
= blk_mq_poll_stats_bkt(rq
);
3860 if (q
->poll_stat
[bucket
].nr_samples
)
3861 ret
= (q
->poll_stat
[bucket
].mean
+ 1) / 2;
3866 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
3869 struct hrtimer_sleeper hs
;
3870 enum hrtimer_mode mode
;
3874 if (rq
->rq_flags
& RQF_MQ_POLL_SLEPT
)
3878 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3880 * 0: use half of prev avg
3881 * >0: use this specific value
3883 if (q
->poll_nsec
> 0)
3884 nsecs
= q
->poll_nsec
;
3886 nsecs
= blk_mq_poll_nsecs(q
, rq
);
3891 rq
->rq_flags
|= RQF_MQ_POLL_SLEPT
;
3894 * This will be replaced with the stats tracking code, using
3895 * 'avg_completion_time / 2' as the pre-sleep target.
3899 mode
= HRTIMER_MODE_REL
;
3900 hrtimer_init_sleeper_on_stack(&hs
, CLOCK_MONOTONIC
, mode
);
3901 hrtimer_set_expires(&hs
.timer
, kt
);
3904 if (blk_mq_rq_state(rq
) == MQ_RQ_COMPLETE
)
3906 set_current_state(TASK_UNINTERRUPTIBLE
);
3907 hrtimer_sleeper_start_expires(&hs
, mode
);
3910 hrtimer_cancel(&hs
.timer
);
3911 mode
= HRTIMER_MODE_ABS
;
3912 } while (hs
.task
&& !signal_pending(current
));
3914 __set_current_state(TASK_RUNNING
);
3915 destroy_hrtimer_on_stack(&hs
.timer
);
3919 static bool blk_mq_poll_hybrid(struct request_queue
*q
,
3920 struct blk_mq_hw_ctx
*hctx
, blk_qc_t cookie
)
3924 if (q
->poll_nsec
== BLK_MQ_POLL_CLASSIC
)
3927 if (!blk_qc_t_is_internal(cookie
))
3928 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
3930 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
3932 * With scheduling, if the request has completed, we'll
3933 * get a NULL return here, as we clear the sched tag when
3934 * that happens. The request still remains valid, like always,
3935 * so we should be safe with just the NULL check.
3941 return blk_mq_poll_hybrid_sleep(q
, rq
);
3945 * blk_poll - poll for IO completions
3947 * @cookie: cookie passed back at IO submission time
3948 * @spin: whether to spin for completions
3951 * Poll for completions on the passed in queue. Returns number of
3952 * completed entries found. If @spin is true, then blk_poll will continue
3953 * looping until at least one completion is found, unless the task is
3954 * otherwise marked running (or we need to reschedule).
3956 int blk_poll(struct request_queue
*q
, blk_qc_t cookie
, bool spin
)
3958 struct blk_mq_hw_ctx
*hctx
;
3961 if (!blk_qc_t_valid(cookie
) ||
3962 !test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
3966 blk_flush_plug_list(current
->plug
, false);
3968 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
3971 * If we sleep, have the caller restart the poll loop to reset
3972 * the state. Like for the other success return cases, the
3973 * caller is responsible for checking if the IO completed. If
3974 * the IO isn't complete, we'll get called again and will go
3975 * straight to the busy poll loop. If specified not to spin,
3976 * we also should not sleep.
3978 if (spin
&& blk_mq_poll_hybrid(q
, hctx
, cookie
))
3981 hctx
->poll_considered
++;
3983 state
= get_current_state();
3987 hctx
->poll_invoked
++;
3989 ret
= q
->mq_ops
->poll(hctx
);
3991 hctx
->poll_success
++;
3992 __set_current_state(TASK_RUNNING
);
3996 if (signal_pending_state(state
, current
))
3997 __set_current_state(TASK_RUNNING
);
3999 if (task_is_running(current
))
4001 if (ret
< 0 || !spin
)
4004 } while (!need_resched());
4006 __set_current_state(TASK_RUNNING
);
4009 EXPORT_SYMBOL_GPL(blk_poll
);
4011 unsigned int blk_mq_rq_cpu(struct request
*rq
)
4013 return rq
->mq_ctx
->cpu
;
4015 EXPORT_SYMBOL(blk_mq_rq_cpu
);
4017 static int __init
blk_mq_init(void)
4021 for_each_possible_cpu(i
)
4022 init_llist_head(&per_cpu(blk_cpu_done
, i
));
4023 open_softirq(BLOCK_SOFTIRQ
, blk_done_softirq
);
4025 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD
,
4026 "block/softirq:dead", NULL
,
4027 blk_softirq_cpu_dead
);
4028 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
4029 blk_mq_hctx_notify_dead
);
4030 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE
, "block/mq:online",
4031 blk_mq_hctx_notify_online
,
4032 blk_mq_hctx_notify_offline
);
4035 subsys_initcall(blk_mq_init
);