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 void blk_mq_poll_stats_start(struct request_queue
*q
);
45 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
);
47 static int blk_mq_poll_stats_bkt(const struct request
*rq
)
49 int ddir
, sectors
, bucket
;
51 ddir
= rq_data_dir(rq
);
52 sectors
= blk_rq_stats_sectors(rq
);
54 bucket
= ddir
+ 2 * ilog2(sectors
);
58 else if (bucket
>= BLK_MQ_POLL_STATS_BKTS
)
59 return ddir
+ BLK_MQ_POLL_STATS_BKTS
- 2;
65 * Check if any of the ctx, dispatch list or elevator
66 * have pending work in this hardware queue.
68 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
70 return !list_empty_careful(&hctx
->dispatch
) ||
71 sbitmap_any_bit_set(&hctx
->ctx_map
) ||
72 blk_mq_sched_has_work(hctx
);
76 * Mark this ctx as having pending work in this hardware queue
78 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
79 struct blk_mq_ctx
*ctx
)
81 const int bit
= ctx
->index_hw
[hctx
->type
];
83 if (!sbitmap_test_bit(&hctx
->ctx_map
, bit
))
84 sbitmap_set_bit(&hctx
->ctx_map
, bit
);
87 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
88 struct blk_mq_ctx
*ctx
)
90 const int bit
= ctx
->index_hw
[hctx
->type
];
92 sbitmap_clear_bit(&hctx
->ctx_map
, bit
);
96 struct hd_struct
*part
;
97 unsigned int inflight
[2];
100 static bool blk_mq_check_inflight(struct blk_mq_hw_ctx
*hctx
,
101 struct request
*rq
, void *priv
,
104 struct mq_inflight
*mi
= priv
;
106 if (rq
->part
== mi
->part
)
107 mi
->inflight
[rq_data_dir(rq
)]++;
112 unsigned int blk_mq_in_flight(struct request_queue
*q
, struct hd_struct
*part
)
114 struct mq_inflight mi
= { .part
= part
};
116 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight
, &mi
);
118 return mi
.inflight
[0] + mi
.inflight
[1];
121 void blk_mq_in_flight_rw(struct request_queue
*q
, struct hd_struct
*part
,
122 unsigned int inflight
[2])
124 struct mq_inflight mi
= { .part
= part
};
126 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight
, &mi
);
127 inflight
[0] = mi
.inflight
[0];
128 inflight
[1] = mi
.inflight
[1];
131 void blk_freeze_queue_start(struct request_queue
*q
)
133 mutex_lock(&q
->mq_freeze_lock
);
134 if (++q
->mq_freeze_depth
== 1) {
135 percpu_ref_kill(&q
->q_usage_counter
);
136 mutex_unlock(&q
->mq_freeze_lock
);
138 blk_mq_run_hw_queues(q
, false);
140 mutex_unlock(&q
->mq_freeze_lock
);
143 EXPORT_SYMBOL_GPL(blk_freeze_queue_start
);
145 void blk_mq_freeze_queue_wait(struct request_queue
*q
)
147 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
149 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait
);
151 int blk_mq_freeze_queue_wait_timeout(struct request_queue
*q
,
152 unsigned long timeout
)
154 return wait_event_timeout(q
->mq_freeze_wq
,
155 percpu_ref_is_zero(&q
->q_usage_counter
),
158 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout
);
161 * Guarantee no request is in use, so we can change any data structure of
162 * the queue afterward.
164 void blk_freeze_queue(struct request_queue
*q
)
167 * In the !blk_mq case we are only calling this to kill the
168 * q_usage_counter, otherwise this increases the freeze depth
169 * and waits for it to return to zero. For this reason there is
170 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
171 * exported to drivers as the only user for unfreeze is blk_mq.
173 blk_freeze_queue_start(q
);
174 blk_mq_freeze_queue_wait(q
);
177 void blk_mq_freeze_queue(struct request_queue
*q
)
180 * ...just an alias to keep freeze and unfreeze actions balanced
181 * in the blk_mq_* namespace
185 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
187 void blk_mq_unfreeze_queue(struct request_queue
*q
)
189 mutex_lock(&q
->mq_freeze_lock
);
190 q
->mq_freeze_depth
--;
191 WARN_ON_ONCE(q
->mq_freeze_depth
< 0);
192 if (!q
->mq_freeze_depth
) {
193 percpu_ref_resurrect(&q
->q_usage_counter
);
194 wake_up_all(&q
->mq_freeze_wq
);
196 mutex_unlock(&q
->mq_freeze_lock
);
198 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
201 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
202 * mpt3sas driver such that this function can be removed.
204 void blk_mq_quiesce_queue_nowait(struct request_queue
*q
)
206 blk_queue_flag_set(QUEUE_FLAG_QUIESCED
, q
);
208 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait
);
211 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
214 * Note: this function does not prevent that the struct request end_io()
215 * callback function is invoked. Once this function is returned, we make
216 * sure no dispatch can happen until the queue is unquiesced via
217 * blk_mq_unquiesce_queue().
219 void blk_mq_quiesce_queue(struct request_queue
*q
)
221 struct blk_mq_hw_ctx
*hctx
;
225 blk_mq_quiesce_queue_nowait(q
);
227 queue_for_each_hw_ctx(q
, hctx
, i
) {
228 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
229 synchronize_srcu(hctx
->srcu
);
236 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
239 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
242 * This function recovers queue into the state before quiescing
243 * which is done by blk_mq_quiesce_queue.
245 void blk_mq_unquiesce_queue(struct request_queue
*q
)
247 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED
, q
);
249 /* dispatch requests which are inserted during quiescing */
250 blk_mq_run_hw_queues(q
, true);
252 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue
);
254 void blk_mq_wake_waiters(struct request_queue
*q
)
256 struct blk_mq_hw_ctx
*hctx
;
259 queue_for_each_hw_ctx(q
, hctx
, i
)
260 if (blk_mq_hw_queue_mapped(hctx
))
261 blk_mq_tag_wakeup_all(hctx
->tags
, true);
265 * Only need start/end time stamping if we have iostat or
266 * blk stats enabled, or using an IO scheduler.
268 static inline bool blk_mq_need_time_stamp(struct request
*rq
)
270 return (rq
->rq_flags
& (RQF_IO_STAT
| RQF_STATS
)) || rq
->q
->elevator
;
273 static struct request
*blk_mq_rq_ctx_init(struct blk_mq_alloc_data
*data
,
274 unsigned int tag
, unsigned int op
, u64 alloc_time_ns
)
276 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
277 struct request
*rq
= tags
->static_rqs
[tag
];
278 req_flags_t rq_flags
= 0;
280 if (data
->flags
& BLK_MQ_REQ_INTERNAL
) {
282 rq
->internal_tag
= tag
;
284 if (data
->hctx
->flags
& BLK_MQ_F_TAG_SHARED
) {
285 rq_flags
= RQF_MQ_INFLIGHT
;
286 atomic_inc(&data
->hctx
->nr_active
);
289 rq
->internal_tag
= -1;
290 data
->hctx
->tags
->rqs
[rq
->tag
] = rq
;
293 /* csd/requeue_work/fifo_time is initialized before use */
295 rq
->mq_ctx
= data
->ctx
;
296 rq
->mq_hctx
= data
->hctx
;
297 rq
->rq_flags
= rq_flags
;
299 if (data
->flags
& BLK_MQ_REQ_PREEMPT
)
300 rq
->rq_flags
|= RQF_PREEMPT
;
301 if (blk_queue_io_stat(data
->q
))
302 rq
->rq_flags
|= RQF_IO_STAT
;
303 INIT_LIST_HEAD(&rq
->queuelist
);
304 INIT_HLIST_NODE(&rq
->hash
);
305 RB_CLEAR_NODE(&rq
->rb_node
);
308 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
309 rq
->alloc_time_ns
= alloc_time_ns
;
311 if (blk_mq_need_time_stamp(rq
))
312 rq
->start_time_ns
= ktime_get_ns();
314 rq
->start_time_ns
= 0;
315 rq
->io_start_time_ns
= 0;
316 rq
->stats_sectors
= 0;
317 rq
->nr_phys_segments
= 0;
318 #if defined(CONFIG_BLK_DEV_INTEGRITY)
319 rq
->nr_integrity_segments
= 0;
321 blk_crypto_rq_set_defaults(rq
);
322 /* tag was already set */
323 WRITE_ONCE(rq
->deadline
, 0);
328 rq
->end_io_data
= NULL
;
330 data
->ctx
->rq_dispatched
[op_is_sync(op
)]++;
331 refcount_set(&rq
->ref
, 1);
335 static struct request
*blk_mq_get_request(struct request_queue
*q
,
337 struct blk_mq_alloc_data
*data
)
339 struct elevator_queue
*e
= q
->elevator
;
342 bool clear_ctx_on_error
= false;
343 u64 alloc_time_ns
= 0;
345 /* alloc_time includes depth and tag waits */
346 if (blk_queue_rq_alloc_time(q
))
347 alloc_time_ns
= ktime_get_ns();
350 if (likely(!data
->ctx
)) {
351 data
->ctx
= blk_mq_get_ctx(q
);
352 clear_ctx_on_error
= true;
354 if (likely(!data
->hctx
))
355 data
->hctx
= blk_mq_map_queue(q
, data
->cmd_flags
,
357 if (data
->cmd_flags
& REQ_NOWAIT
)
358 data
->flags
|= BLK_MQ_REQ_NOWAIT
;
361 data
->flags
|= BLK_MQ_REQ_INTERNAL
;
364 * Flush requests are special and go directly to the
365 * dispatch list. Don't include reserved tags in the
366 * limiting, as it isn't useful.
368 if (!op_is_flush(data
->cmd_flags
) &&
369 e
->type
->ops
.limit_depth
&&
370 !(data
->flags
& BLK_MQ_REQ_RESERVED
))
371 e
->type
->ops
.limit_depth(data
->cmd_flags
, data
);
373 blk_mq_tag_busy(data
->hctx
);
376 tag
= blk_mq_get_tag(data
);
377 if (tag
== BLK_MQ_TAG_FAIL
) {
378 if (clear_ctx_on_error
)
383 rq
= blk_mq_rq_ctx_init(data
, tag
, data
->cmd_flags
, alloc_time_ns
);
384 if (!op_is_flush(data
->cmd_flags
)) {
386 if (e
&& e
->type
->ops
.prepare_request
) {
387 if (e
->type
->icq_cache
)
388 blk_mq_sched_assign_ioc(rq
);
390 e
->type
->ops
.prepare_request(rq
, bio
);
391 rq
->rq_flags
|= RQF_ELVPRIV
;
394 data
->hctx
->queued
++;
398 struct request
*blk_mq_alloc_request(struct request_queue
*q
, unsigned int op
,
399 blk_mq_req_flags_t flags
)
401 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
, .cmd_flags
= op
};
405 ret
= blk_queue_enter(q
, flags
);
409 rq
= blk_mq_get_request(q
, NULL
, &alloc_data
);
413 rq
->__sector
= (sector_t
) -1;
414 rq
->bio
= rq
->biotail
= NULL
;
418 return ERR_PTR(-EWOULDBLOCK
);
420 EXPORT_SYMBOL(blk_mq_alloc_request
);
422 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
,
423 unsigned int op
, blk_mq_req_flags_t flags
, unsigned int hctx_idx
)
425 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
, .cmd_flags
= op
};
431 * If the tag allocator sleeps we could get an allocation for a
432 * different hardware context. No need to complicate the low level
433 * allocator for this for the rare use case of a command tied to
436 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
437 return ERR_PTR(-EINVAL
);
439 if (hctx_idx
>= q
->nr_hw_queues
)
440 return ERR_PTR(-EIO
);
442 ret
= blk_queue_enter(q
, flags
);
447 * Check if the hardware context is actually mapped to anything.
448 * If not tell the caller that it should skip this queue.
451 alloc_data
.hctx
= q
->queue_hw_ctx
[hctx_idx
];
452 if (!blk_mq_hw_queue_mapped(alloc_data
.hctx
))
454 cpu
= cpumask_first_and(alloc_data
.hctx
->cpumask
, cpu_online_mask
);
455 alloc_data
.ctx
= __blk_mq_get_ctx(q
, cpu
);
458 rq
= blk_mq_get_request(q
, NULL
, &alloc_data
);
466 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
468 static void __blk_mq_free_request(struct request
*rq
)
470 struct request_queue
*q
= rq
->q
;
471 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
472 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
473 const int sched_tag
= rq
->internal_tag
;
475 blk_crypto_free_request(rq
);
476 blk_pm_mark_last_busy(rq
);
479 blk_mq_put_tag(hctx
->tags
, ctx
, rq
->tag
);
481 blk_mq_put_tag(hctx
->sched_tags
, ctx
, sched_tag
);
482 blk_mq_sched_restart(hctx
);
486 void blk_mq_free_request(struct request
*rq
)
488 struct request_queue
*q
= rq
->q
;
489 struct elevator_queue
*e
= q
->elevator
;
490 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
491 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
493 if (rq
->rq_flags
& RQF_ELVPRIV
) {
494 if (e
&& e
->type
->ops
.finish_request
)
495 e
->type
->ops
.finish_request(rq
);
497 put_io_context(rq
->elv
.icq
->ioc
);
502 ctx
->rq_completed
[rq_is_sync(rq
)]++;
503 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
504 atomic_dec(&hctx
->nr_active
);
506 if (unlikely(laptop_mode
&& !blk_rq_is_passthrough(rq
)))
507 laptop_io_completion(q
->backing_dev_info
);
511 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
512 if (refcount_dec_and_test(&rq
->ref
))
513 __blk_mq_free_request(rq
);
515 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
517 inline void __blk_mq_end_request(struct request
*rq
, blk_status_t error
)
521 if (blk_mq_need_time_stamp(rq
))
522 now
= ktime_get_ns();
524 if (rq
->rq_flags
& RQF_STATS
) {
525 blk_mq_poll_stats_start(rq
->q
);
526 blk_stat_add(rq
, now
);
529 if (rq
->internal_tag
!= -1)
530 blk_mq_sched_completed_request(rq
, now
);
532 blk_account_io_done(rq
, now
);
535 rq_qos_done(rq
->q
, rq
);
536 rq
->end_io(rq
, error
);
538 blk_mq_free_request(rq
);
541 EXPORT_SYMBOL(__blk_mq_end_request
);
543 void blk_mq_end_request(struct request
*rq
, blk_status_t error
)
545 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
547 __blk_mq_end_request(rq
, error
);
549 EXPORT_SYMBOL(blk_mq_end_request
);
551 static void __blk_mq_complete_request_remote(void *data
)
553 struct request
*rq
= data
;
554 struct request_queue
*q
= rq
->q
;
556 q
->mq_ops
->complete(rq
);
560 * blk_mq_force_complete_rq() - Force complete the request, bypassing any error
561 * injection that could drop the completion.
562 * @rq: Request to be force completed
564 * Drivers should use blk_mq_complete_request() to complete requests in their
565 * normal IO path. For timeout error recovery, drivers may call this forced
566 * completion routine after they've reclaimed timed out requests to bypass
567 * potentially subsequent fake timeouts.
569 void blk_mq_force_complete_rq(struct request
*rq
)
571 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
572 struct request_queue
*q
= rq
->q
;
576 WRITE_ONCE(rq
->state
, MQ_RQ_COMPLETE
);
578 * Most of single queue controllers, there is only one irq vector
579 * for handling IO completion, and the only irq's affinity is set
580 * as all possible CPUs. On most of ARCHs, this affinity means the
581 * irq is handled on one specific CPU.
583 * So complete IO reqeust in softirq context in case of single queue
584 * for not degrading IO performance by irqsoff latency.
586 if (q
->nr_hw_queues
== 1) {
587 __blk_complete_request(rq
);
592 * For a polled request, always complete locallly, it's pointless
593 * to redirect the completion.
595 if ((rq
->cmd_flags
& REQ_HIPRI
) ||
596 !test_bit(QUEUE_FLAG_SAME_COMP
, &q
->queue_flags
)) {
597 q
->mq_ops
->complete(rq
);
602 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &q
->queue_flags
))
603 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
605 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
606 rq
->csd
.func
= __blk_mq_complete_request_remote
;
609 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
611 q
->mq_ops
->complete(rq
);
615 EXPORT_SYMBOL_GPL(blk_mq_force_complete_rq
);
617 static void hctx_unlock(struct blk_mq_hw_ctx
*hctx
, int srcu_idx
)
618 __releases(hctx
->srcu
)
620 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
))
623 srcu_read_unlock(hctx
->srcu
, srcu_idx
);
626 static void hctx_lock(struct blk_mq_hw_ctx
*hctx
, int *srcu_idx
)
627 __acquires(hctx
->srcu
)
629 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
630 /* shut up gcc false positive */
634 *srcu_idx
= srcu_read_lock(hctx
->srcu
);
638 * blk_mq_complete_request - end I/O on a request
639 * @rq: the request being processed
642 * Ends all I/O on a request. It does not handle partial completions.
643 * The actual completion happens out-of-order, through a IPI handler.
645 bool blk_mq_complete_request(struct request
*rq
)
647 if (unlikely(blk_should_fake_timeout(rq
->q
)))
649 blk_mq_force_complete_rq(rq
);
652 EXPORT_SYMBOL(blk_mq_complete_request
);
655 * blk_mq_start_request - Start processing a request
656 * @rq: Pointer to request to be started
658 * Function used by device drivers to notify the block layer that a request
659 * is going to be processed now, so blk layer can do proper initializations
660 * such as starting the timeout timer.
662 void blk_mq_start_request(struct request
*rq
)
664 struct request_queue
*q
= rq
->q
;
666 trace_block_rq_issue(q
, rq
);
668 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
669 rq
->io_start_time_ns
= ktime_get_ns();
670 rq
->stats_sectors
= blk_rq_sectors(rq
);
671 rq
->rq_flags
|= RQF_STATS
;
675 WARN_ON_ONCE(blk_mq_rq_state(rq
) != MQ_RQ_IDLE
);
678 WRITE_ONCE(rq
->state
, MQ_RQ_IN_FLIGHT
);
680 #ifdef CONFIG_BLK_DEV_INTEGRITY
681 if (blk_integrity_rq(rq
) && req_op(rq
) == REQ_OP_WRITE
)
682 q
->integrity
.profile
->prepare_fn(rq
);
685 EXPORT_SYMBOL(blk_mq_start_request
);
687 static void __blk_mq_requeue_request(struct request
*rq
)
689 struct request_queue
*q
= rq
->q
;
691 blk_mq_put_driver_tag(rq
);
693 trace_block_rq_requeue(q
, rq
);
694 rq_qos_requeue(q
, rq
);
696 if (blk_mq_request_started(rq
)) {
697 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
698 rq
->rq_flags
&= ~RQF_TIMED_OUT
;
702 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
704 __blk_mq_requeue_request(rq
);
706 /* this request will be re-inserted to io scheduler queue */
707 blk_mq_sched_requeue_request(rq
);
709 BUG_ON(!list_empty(&rq
->queuelist
));
710 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
712 EXPORT_SYMBOL(blk_mq_requeue_request
);
714 static void blk_mq_requeue_work(struct work_struct
*work
)
716 struct request_queue
*q
=
717 container_of(work
, struct request_queue
, requeue_work
.work
);
719 struct request
*rq
, *next
;
721 spin_lock_irq(&q
->requeue_lock
);
722 list_splice_init(&q
->requeue_list
, &rq_list
);
723 spin_unlock_irq(&q
->requeue_lock
);
725 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
726 if (!(rq
->rq_flags
& (RQF_SOFTBARRIER
| RQF_DONTPREP
)))
729 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
730 list_del_init(&rq
->queuelist
);
732 * If RQF_DONTPREP, rq has contained some driver specific
733 * data, so insert it to hctx dispatch list to avoid any
736 if (rq
->rq_flags
& RQF_DONTPREP
)
737 blk_mq_request_bypass_insert(rq
, false, false);
739 blk_mq_sched_insert_request(rq
, true, false, false);
742 while (!list_empty(&rq_list
)) {
743 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
744 list_del_init(&rq
->queuelist
);
745 blk_mq_sched_insert_request(rq
, false, false, false);
748 blk_mq_run_hw_queues(q
, false);
751 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
752 bool kick_requeue_list
)
754 struct request_queue
*q
= rq
->q
;
758 * We abuse this flag that is otherwise used by the I/O scheduler to
759 * request head insertion from the workqueue.
761 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
763 spin_lock_irqsave(&q
->requeue_lock
, flags
);
765 rq
->rq_flags
|= RQF_SOFTBARRIER
;
766 list_add(&rq
->queuelist
, &q
->requeue_list
);
768 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
770 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
772 if (kick_requeue_list
)
773 blk_mq_kick_requeue_list(q
);
776 void blk_mq_kick_requeue_list(struct request_queue
*q
)
778 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
, 0);
780 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
782 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
785 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
,
786 msecs_to_jiffies(msecs
));
788 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
790 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
792 if (tag
< tags
->nr_tags
) {
793 prefetch(tags
->rqs
[tag
]);
794 return tags
->rqs
[tag
];
799 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
801 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
802 void *priv
, bool reserved
)
805 * If we find a request that is inflight and the queue matches,
806 * we know the queue is busy. Return false to stop the iteration.
808 if (rq
->state
== MQ_RQ_IN_FLIGHT
&& rq
->q
== hctx
->queue
) {
818 bool blk_mq_queue_inflight(struct request_queue
*q
)
822 blk_mq_queue_tag_busy_iter(q
, blk_mq_rq_inflight
, &busy
);
825 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight
);
827 static void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
829 req
->rq_flags
|= RQF_TIMED_OUT
;
830 if (req
->q
->mq_ops
->timeout
) {
831 enum blk_eh_timer_return ret
;
833 ret
= req
->q
->mq_ops
->timeout(req
, reserved
);
834 if (ret
== BLK_EH_DONE
)
836 WARN_ON_ONCE(ret
!= BLK_EH_RESET_TIMER
);
842 static bool blk_mq_req_expired(struct request
*rq
, unsigned long *next
)
844 unsigned long deadline
;
846 if (blk_mq_rq_state(rq
) != MQ_RQ_IN_FLIGHT
)
848 if (rq
->rq_flags
& RQF_TIMED_OUT
)
851 deadline
= READ_ONCE(rq
->deadline
);
852 if (time_after_eq(jiffies
, deadline
))
857 else if (time_after(*next
, deadline
))
862 static bool blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
863 struct request
*rq
, void *priv
, bool reserved
)
865 unsigned long *next
= priv
;
868 * Just do a quick check if it is expired before locking the request in
869 * so we're not unnecessarilly synchronizing across CPUs.
871 if (!blk_mq_req_expired(rq
, next
))
875 * We have reason to believe the request may be expired. Take a
876 * reference on the request to lock this request lifetime into its
877 * currently allocated context to prevent it from being reallocated in
878 * the event the completion by-passes this timeout handler.
880 * If the reference was already released, then the driver beat the
881 * timeout handler to posting a natural completion.
883 if (!refcount_inc_not_zero(&rq
->ref
))
887 * The request is now locked and cannot be reallocated underneath the
888 * timeout handler's processing. Re-verify this exact request is truly
889 * expired; if it is not expired, then the request was completed and
890 * reallocated as a new request.
892 if (blk_mq_req_expired(rq
, next
))
893 blk_mq_rq_timed_out(rq
, reserved
);
895 if (is_flush_rq(rq
, hctx
))
897 else if (refcount_dec_and_test(&rq
->ref
))
898 __blk_mq_free_request(rq
);
903 static void blk_mq_timeout_work(struct work_struct
*work
)
905 struct request_queue
*q
=
906 container_of(work
, struct request_queue
, timeout_work
);
907 unsigned long next
= 0;
908 struct blk_mq_hw_ctx
*hctx
;
911 /* A deadlock might occur if a request is stuck requiring a
912 * timeout at the same time a queue freeze is waiting
913 * completion, since the timeout code would not be able to
914 * acquire the queue reference here.
916 * That's why we don't use blk_queue_enter here; instead, we use
917 * percpu_ref_tryget directly, because we need to be able to
918 * obtain a reference even in the short window between the queue
919 * starting to freeze, by dropping the first reference in
920 * blk_freeze_queue_start, and the moment the last request is
921 * consumed, marked by the instant q_usage_counter reaches
924 if (!percpu_ref_tryget(&q
->q_usage_counter
))
927 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &next
);
930 mod_timer(&q
->timeout
, next
);
933 * Request timeouts are handled as a forward rolling timer. If
934 * we end up here it means that no requests are pending and
935 * also that no request has been pending for a while. Mark
938 queue_for_each_hw_ctx(q
, hctx
, i
) {
939 /* the hctx may be unmapped, so check it here */
940 if (blk_mq_hw_queue_mapped(hctx
))
941 blk_mq_tag_idle(hctx
);
947 struct flush_busy_ctx_data
{
948 struct blk_mq_hw_ctx
*hctx
;
949 struct list_head
*list
;
952 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
954 struct flush_busy_ctx_data
*flush_data
= data
;
955 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
956 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
957 enum hctx_type type
= hctx
->type
;
959 spin_lock(&ctx
->lock
);
960 list_splice_tail_init(&ctx
->rq_lists
[type
], flush_data
->list
);
961 sbitmap_clear_bit(sb
, bitnr
);
962 spin_unlock(&ctx
->lock
);
967 * Process software queues that have been marked busy, splicing them
968 * to the for-dispatch
970 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
972 struct flush_busy_ctx_data data
= {
977 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
979 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
981 struct dispatch_rq_data
{
982 struct blk_mq_hw_ctx
*hctx
;
986 static bool dispatch_rq_from_ctx(struct sbitmap
*sb
, unsigned int bitnr
,
989 struct dispatch_rq_data
*dispatch_data
= data
;
990 struct blk_mq_hw_ctx
*hctx
= dispatch_data
->hctx
;
991 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
992 enum hctx_type type
= hctx
->type
;
994 spin_lock(&ctx
->lock
);
995 if (!list_empty(&ctx
->rq_lists
[type
])) {
996 dispatch_data
->rq
= list_entry_rq(ctx
->rq_lists
[type
].next
);
997 list_del_init(&dispatch_data
->rq
->queuelist
);
998 if (list_empty(&ctx
->rq_lists
[type
]))
999 sbitmap_clear_bit(sb
, bitnr
);
1001 spin_unlock(&ctx
->lock
);
1003 return !dispatch_data
->rq
;
1006 struct request
*blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx
*hctx
,
1007 struct blk_mq_ctx
*start
)
1009 unsigned off
= start
? start
->index_hw
[hctx
->type
] : 0;
1010 struct dispatch_rq_data data
= {
1015 __sbitmap_for_each_set(&hctx
->ctx_map
, off
,
1016 dispatch_rq_from_ctx
, &data
);
1021 static inline unsigned int queued_to_index(unsigned int queued
)
1026 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
1029 bool blk_mq_get_driver_tag(struct request
*rq
)
1031 struct blk_mq_alloc_data data
= {
1033 .hctx
= rq
->mq_hctx
,
1034 .flags
= BLK_MQ_REQ_NOWAIT
,
1035 .cmd_flags
= rq
->cmd_flags
,
1042 if (blk_mq_tag_is_reserved(data
.hctx
->sched_tags
, rq
->internal_tag
))
1043 data
.flags
|= BLK_MQ_REQ_RESERVED
;
1045 shared
= blk_mq_tag_busy(data
.hctx
);
1046 rq
->tag
= blk_mq_get_tag(&data
);
1049 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
1050 atomic_inc(&data
.hctx
->nr_active
);
1052 data
.hctx
->tags
->rqs
[rq
->tag
] = rq
;
1055 return rq
->tag
!= -1;
1058 static int blk_mq_dispatch_wake(wait_queue_entry_t
*wait
, unsigned mode
,
1059 int flags
, void *key
)
1061 struct blk_mq_hw_ctx
*hctx
;
1063 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
1065 spin_lock(&hctx
->dispatch_wait_lock
);
1066 if (!list_empty(&wait
->entry
)) {
1067 struct sbitmap_queue
*sbq
;
1069 list_del_init(&wait
->entry
);
1070 sbq
= &hctx
->tags
->bitmap_tags
;
1071 atomic_dec(&sbq
->ws_active
);
1073 spin_unlock(&hctx
->dispatch_wait_lock
);
1075 blk_mq_run_hw_queue(hctx
, true);
1080 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1081 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1082 * restart. For both cases, take care to check the condition again after
1083 * marking us as waiting.
1085 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx
*hctx
,
1088 struct sbitmap_queue
*sbq
= &hctx
->tags
->bitmap_tags
;
1089 struct wait_queue_head
*wq
;
1090 wait_queue_entry_t
*wait
;
1093 if (!(hctx
->flags
& BLK_MQ_F_TAG_SHARED
)) {
1094 blk_mq_sched_mark_restart_hctx(hctx
);
1097 * It's possible that a tag was freed in the window between the
1098 * allocation failure and adding the hardware queue to the wait
1101 * Don't clear RESTART here, someone else could have set it.
1102 * At most this will cost an extra queue run.
1104 return blk_mq_get_driver_tag(rq
);
1107 wait
= &hctx
->dispatch_wait
;
1108 if (!list_empty_careful(&wait
->entry
))
1111 wq
= &bt_wait_ptr(sbq
, hctx
)->wait
;
1113 spin_lock_irq(&wq
->lock
);
1114 spin_lock(&hctx
->dispatch_wait_lock
);
1115 if (!list_empty(&wait
->entry
)) {
1116 spin_unlock(&hctx
->dispatch_wait_lock
);
1117 spin_unlock_irq(&wq
->lock
);
1121 atomic_inc(&sbq
->ws_active
);
1122 wait
->flags
&= ~WQ_FLAG_EXCLUSIVE
;
1123 __add_wait_queue(wq
, wait
);
1126 * It's possible that a tag was freed in the window between the
1127 * allocation failure and adding the hardware queue to the wait
1130 ret
= blk_mq_get_driver_tag(rq
);
1132 spin_unlock(&hctx
->dispatch_wait_lock
);
1133 spin_unlock_irq(&wq
->lock
);
1138 * We got a tag, remove ourselves from the wait queue to ensure
1139 * someone else gets the wakeup.
1141 list_del_init(&wait
->entry
);
1142 atomic_dec(&sbq
->ws_active
);
1143 spin_unlock(&hctx
->dispatch_wait_lock
);
1144 spin_unlock_irq(&wq
->lock
);
1149 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1150 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1152 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1153 * - EWMA is one simple way to compute running average value
1154 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1155 * - take 4 as factor for avoiding to get too small(0) result, and this
1156 * factor doesn't matter because EWMA decreases exponentially
1158 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx
*hctx
, bool busy
)
1162 if (hctx
->queue
->elevator
)
1165 ewma
= hctx
->dispatch_busy
;
1170 ewma
*= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
- 1;
1172 ewma
+= 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR
;
1173 ewma
/= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
;
1175 hctx
->dispatch_busy
= ewma
;
1178 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1180 static void blk_mq_handle_dev_resource(struct request
*rq
,
1181 struct list_head
*list
)
1183 struct request
*next
=
1184 list_first_entry_or_null(list
, struct request
, queuelist
);
1187 * If an I/O scheduler has been configured and we got a driver tag for
1188 * the next request already, free it.
1191 blk_mq_put_driver_tag(next
);
1193 list_add(&rq
->queuelist
, list
);
1194 __blk_mq_requeue_request(rq
);
1197 static void blk_mq_handle_zone_resource(struct request
*rq
,
1198 struct list_head
*zone_list
)
1201 * If we end up here it is because we cannot dispatch a request to a
1202 * specific zone due to LLD level zone-write locking or other zone
1203 * related resource not being available. In this case, set the request
1204 * aside in zone_list for retrying it later.
1206 list_add(&rq
->queuelist
, zone_list
);
1207 __blk_mq_requeue_request(rq
);
1211 * Returns true if we did some work AND can potentially do more.
1213 bool blk_mq_dispatch_rq_list(struct request_queue
*q
, struct list_head
*list
,
1216 struct blk_mq_hw_ctx
*hctx
;
1217 struct request
*rq
, *nxt
;
1218 bool no_tag
= false;
1220 blk_status_t ret
= BLK_STS_OK
;
1221 bool no_budget_avail
= false;
1222 LIST_HEAD(zone_list
);
1224 if (list_empty(list
))
1227 WARN_ON(!list_is_singular(list
) && got_budget
);
1230 * Now process all the entries, sending them to the driver.
1232 errors
= queued
= 0;
1234 struct blk_mq_queue_data bd
;
1236 rq
= list_first_entry(list
, struct request
, queuelist
);
1239 if (!got_budget
&& !blk_mq_get_dispatch_budget(hctx
)) {
1240 blk_mq_put_driver_tag(rq
);
1241 no_budget_avail
= true;
1245 if (!blk_mq_get_driver_tag(rq
)) {
1247 * The initial allocation attempt failed, so we need to
1248 * rerun the hardware queue when a tag is freed. The
1249 * waitqueue takes care of that. If the queue is run
1250 * before we add this entry back on the dispatch list,
1251 * we'll re-run it below.
1253 if (!blk_mq_mark_tag_wait(hctx
, rq
)) {
1254 blk_mq_put_dispatch_budget(hctx
);
1256 * For non-shared tags, the RESTART check
1259 if (hctx
->flags
& BLK_MQ_F_TAG_SHARED
)
1265 list_del_init(&rq
->queuelist
);
1270 * Flag last if we have no more requests, or if we have more
1271 * but can't assign a driver tag to it.
1273 if (list_empty(list
))
1276 nxt
= list_first_entry(list
, struct request
, queuelist
);
1277 bd
.last
= !blk_mq_get_driver_tag(nxt
);
1280 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1281 if (ret
== BLK_STS_RESOURCE
|| ret
== BLK_STS_DEV_RESOURCE
) {
1282 blk_mq_handle_dev_resource(rq
, list
);
1284 } else if (ret
== BLK_STS_ZONE_RESOURCE
) {
1286 * Move the request to zone_list and keep going through
1287 * the dispatch list to find more requests the drive can
1290 blk_mq_handle_zone_resource(rq
, &zone_list
);
1291 if (list_empty(list
))
1296 if (unlikely(ret
!= BLK_STS_OK
)) {
1298 blk_mq_end_request(rq
, BLK_STS_IOERR
);
1303 } while (!list_empty(list
));
1305 if (!list_empty(&zone_list
))
1306 list_splice_tail_init(&zone_list
, list
);
1308 hctx
->dispatched
[queued_to_index(queued
)]++;
1311 * Any items that need requeuing? Stuff them into hctx->dispatch,
1312 * that is where we will continue on next queue run.
1314 if (!list_empty(list
)) {
1318 * If we didn't flush the entire list, we could have told
1319 * the driver there was more coming, but that turned out to
1322 if (q
->mq_ops
->commit_rqs
&& queued
)
1323 q
->mq_ops
->commit_rqs(hctx
);
1325 spin_lock(&hctx
->lock
);
1326 list_splice_tail_init(list
, &hctx
->dispatch
);
1327 spin_unlock(&hctx
->lock
);
1330 * If SCHED_RESTART was set by the caller of this function and
1331 * it is no longer set that means that it was cleared by another
1332 * thread and hence that a queue rerun is needed.
1334 * If 'no_tag' is set, that means that we failed getting
1335 * a driver tag with an I/O scheduler attached. If our dispatch
1336 * waitqueue is no longer active, ensure that we run the queue
1337 * AFTER adding our entries back to the list.
1339 * If no I/O scheduler has been configured it is possible that
1340 * the hardware queue got stopped and restarted before requests
1341 * were pushed back onto the dispatch list. Rerun the queue to
1342 * avoid starvation. Notes:
1343 * - blk_mq_run_hw_queue() checks whether or not a queue has
1344 * been stopped before rerunning a queue.
1345 * - Some but not all block drivers stop a queue before
1346 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1349 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1350 * bit is set, run queue after a delay to avoid IO stalls
1351 * that could otherwise occur if the queue is idle. We'll do
1352 * similar if we couldn't get budget and SCHED_RESTART is set.
1354 needs_restart
= blk_mq_sched_needs_restart(hctx
);
1355 if (!needs_restart
||
1356 (no_tag
&& list_empty_careful(&hctx
->dispatch_wait
.entry
)))
1357 blk_mq_run_hw_queue(hctx
, true);
1358 else if (needs_restart
&& (ret
== BLK_STS_RESOURCE
||
1360 blk_mq_delay_run_hw_queue(hctx
, BLK_MQ_RESOURCE_DELAY
);
1362 blk_mq_update_dispatch_busy(hctx
, true);
1365 blk_mq_update_dispatch_busy(hctx
, false);
1368 * If the host/device is unable to accept more work, inform the
1371 if (ret
== BLK_STS_RESOURCE
|| ret
== BLK_STS_DEV_RESOURCE
)
1374 return (queued
+ errors
) != 0;
1378 * __blk_mq_run_hw_queue - Run a hardware queue.
1379 * @hctx: Pointer to the hardware queue to run.
1381 * Send pending requests to the hardware.
1383 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1388 * We should be running this queue from one of the CPUs that
1391 * There are at least two related races now between setting
1392 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1393 * __blk_mq_run_hw_queue():
1395 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1396 * but later it becomes online, then this warning is harmless
1399 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1400 * but later it becomes offline, then the warning can't be
1401 * triggered, and we depend on blk-mq timeout handler to
1402 * handle dispatched requests to this hctx
1404 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
1405 cpu_online(hctx
->next_cpu
)) {
1406 printk(KERN_WARNING
"run queue from wrong CPU %d, hctx %s\n",
1407 raw_smp_processor_id(),
1408 cpumask_empty(hctx
->cpumask
) ? "inactive": "active");
1413 * We can't run the queue inline with ints disabled. Ensure that
1414 * we catch bad users of this early.
1416 WARN_ON_ONCE(in_interrupt());
1418 might_sleep_if(hctx
->flags
& BLK_MQ_F_BLOCKING
);
1420 hctx_lock(hctx
, &srcu_idx
);
1421 blk_mq_sched_dispatch_requests(hctx
);
1422 hctx_unlock(hctx
, srcu_idx
);
1425 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx
*hctx
)
1427 int cpu
= cpumask_first_and(hctx
->cpumask
, cpu_online_mask
);
1429 if (cpu
>= nr_cpu_ids
)
1430 cpu
= cpumask_first(hctx
->cpumask
);
1435 * It'd be great if the workqueue API had a way to pass
1436 * in a mask and had some smarts for more clever placement.
1437 * For now we just round-robin here, switching for every
1438 * BLK_MQ_CPU_WORK_BATCH queued items.
1440 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1443 int next_cpu
= hctx
->next_cpu
;
1445 if (hctx
->queue
->nr_hw_queues
== 1)
1446 return WORK_CPU_UNBOUND
;
1448 if (--hctx
->next_cpu_batch
<= 0) {
1450 next_cpu
= cpumask_next_and(next_cpu
, hctx
->cpumask
,
1452 if (next_cpu
>= nr_cpu_ids
)
1453 next_cpu
= blk_mq_first_mapped_cpu(hctx
);
1454 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1458 * Do unbound schedule if we can't find a online CPU for this hctx,
1459 * and it should only happen in the path of handling CPU DEAD.
1461 if (!cpu_online(next_cpu
)) {
1468 * Make sure to re-select CPU next time once after CPUs
1469 * in hctx->cpumask become online again.
1471 hctx
->next_cpu
= next_cpu
;
1472 hctx
->next_cpu_batch
= 1;
1473 return WORK_CPU_UNBOUND
;
1476 hctx
->next_cpu
= next_cpu
;
1481 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
1482 * @hctx: Pointer to the hardware queue to run.
1483 * @async: If we want to run the queue asynchronously.
1484 * @msecs: Microseconds of delay to wait before running the queue.
1486 * If !@async, try to run the queue now. Else, run the queue asynchronously and
1487 * with a delay of @msecs.
1489 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
,
1490 unsigned long msecs
)
1492 if (unlikely(blk_mq_hctx_stopped(hctx
)))
1495 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1496 int cpu
= get_cpu();
1497 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1498 __blk_mq_run_hw_queue(hctx
);
1506 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
,
1507 msecs_to_jiffies(msecs
));
1511 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
1512 * @hctx: Pointer to the hardware queue to run.
1513 * @msecs: Microseconds of delay to wait before running the queue.
1515 * Run a hardware queue asynchronously with a delay of @msecs.
1517 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1519 __blk_mq_delay_run_hw_queue(hctx
, true, msecs
);
1521 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue
);
1524 * blk_mq_run_hw_queue - Start to run a hardware queue.
1525 * @hctx: Pointer to the hardware queue to run.
1526 * @async: If we want to run the queue asynchronously.
1528 * Check if the request queue is not in a quiesced state and if there are
1529 * pending requests to be sent. If this is true, run the queue to send requests
1532 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1538 * When queue is quiesced, we may be switching io scheduler, or
1539 * updating nr_hw_queues, or other things, and we can't run queue
1540 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1542 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1545 hctx_lock(hctx
, &srcu_idx
);
1546 need_run
= !blk_queue_quiesced(hctx
->queue
) &&
1547 blk_mq_hctx_has_pending(hctx
);
1548 hctx_unlock(hctx
, srcu_idx
);
1551 __blk_mq_delay_run_hw_queue(hctx
, async
, 0);
1553 EXPORT_SYMBOL(blk_mq_run_hw_queue
);
1556 * blk_mq_run_hw_queue - Run all hardware queues in a request queue.
1557 * @q: Pointer to the request queue to run.
1558 * @async: If we want to run the queue asynchronously.
1560 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1562 struct blk_mq_hw_ctx
*hctx
;
1565 queue_for_each_hw_ctx(q
, hctx
, i
) {
1566 if (blk_mq_hctx_stopped(hctx
))
1569 blk_mq_run_hw_queue(hctx
, async
);
1572 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1575 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
1576 * @q: Pointer to the request queue to run.
1577 * @msecs: Microseconds of delay to wait before running the queues.
1579 void blk_mq_delay_run_hw_queues(struct request_queue
*q
, unsigned long msecs
)
1581 struct blk_mq_hw_ctx
*hctx
;
1584 queue_for_each_hw_ctx(q
, hctx
, i
) {
1585 if (blk_mq_hctx_stopped(hctx
))
1588 blk_mq_delay_run_hw_queue(hctx
, msecs
);
1591 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues
);
1594 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1595 * @q: request queue.
1597 * The caller is responsible for serializing this function against
1598 * blk_mq_{start,stop}_hw_queue().
1600 bool blk_mq_queue_stopped(struct request_queue
*q
)
1602 struct blk_mq_hw_ctx
*hctx
;
1605 queue_for_each_hw_ctx(q
, hctx
, i
)
1606 if (blk_mq_hctx_stopped(hctx
))
1611 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1614 * This function is often used for pausing .queue_rq() by driver when
1615 * there isn't enough resource or some conditions aren't satisfied, and
1616 * BLK_STS_RESOURCE is usually returned.
1618 * We do not guarantee that dispatch can be drained or blocked
1619 * after blk_mq_stop_hw_queue() returns. Please use
1620 * blk_mq_quiesce_queue() for that requirement.
1622 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1624 cancel_delayed_work(&hctx
->run_work
);
1626 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1628 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1631 * This function is often used for pausing .queue_rq() by driver when
1632 * there isn't enough resource or some conditions aren't satisfied, and
1633 * BLK_STS_RESOURCE is usually returned.
1635 * We do not guarantee that dispatch can be drained or blocked
1636 * after blk_mq_stop_hw_queues() returns. Please use
1637 * blk_mq_quiesce_queue() for that requirement.
1639 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1641 struct blk_mq_hw_ctx
*hctx
;
1644 queue_for_each_hw_ctx(q
, hctx
, i
)
1645 blk_mq_stop_hw_queue(hctx
);
1647 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1649 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1651 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1653 blk_mq_run_hw_queue(hctx
, false);
1655 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1657 void blk_mq_start_hw_queues(struct request_queue
*q
)
1659 struct blk_mq_hw_ctx
*hctx
;
1662 queue_for_each_hw_ctx(q
, hctx
, i
)
1663 blk_mq_start_hw_queue(hctx
);
1665 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1667 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1669 if (!blk_mq_hctx_stopped(hctx
))
1672 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1673 blk_mq_run_hw_queue(hctx
, async
);
1675 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1677 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1679 struct blk_mq_hw_ctx
*hctx
;
1682 queue_for_each_hw_ctx(q
, hctx
, i
)
1683 blk_mq_start_stopped_hw_queue(hctx
, async
);
1685 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1687 static void blk_mq_run_work_fn(struct work_struct
*work
)
1689 struct blk_mq_hw_ctx
*hctx
;
1691 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
1694 * If we are stopped, don't run the queue.
1696 if (test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
1699 __blk_mq_run_hw_queue(hctx
);
1702 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1706 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1707 enum hctx_type type
= hctx
->type
;
1709 lockdep_assert_held(&ctx
->lock
);
1711 trace_block_rq_insert(hctx
->queue
, rq
);
1714 list_add(&rq
->queuelist
, &ctx
->rq_lists
[type
]);
1716 list_add_tail(&rq
->queuelist
, &ctx
->rq_lists
[type
]);
1719 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1722 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1724 lockdep_assert_held(&ctx
->lock
);
1726 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1727 blk_mq_hctx_mark_pending(hctx
, ctx
);
1731 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
1732 * @rq: Pointer to request to be inserted.
1733 * @run_queue: If we should run the hardware queue after inserting the request.
1735 * Should only be used carefully, when the caller knows we want to
1736 * bypass a potential IO scheduler on the target device.
1738 void blk_mq_request_bypass_insert(struct request
*rq
, bool at_head
,
1741 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1743 spin_lock(&hctx
->lock
);
1745 list_add(&rq
->queuelist
, &hctx
->dispatch
);
1747 list_add_tail(&rq
->queuelist
, &hctx
->dispatch
);
1748 spin_unlock(&hctx
->lock
);
1751 blk_mq_run_hw_queue(hctx
, false);
1754 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1755 struct list_head
*list
)
1759 enum hctx_type type
= hctx
->type
;
1762 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1765 list_for_each_entry(rq
, list
, queuelist
) {
1766 BUG_ON(rq
->mq_ctx
!= ctx
);
1767 trace_block_rq_insert(hctx
->queue
, rq
);
1770 spin_lock(&ctx
->lock
);
1771 list_splice_tail_init(list
, &ctx
->rq_lists
[type
]);
1772 blk_mq_hctx_mark_pending(hctx
, ctx
);
1773 spin_unlock(&ctx
->lock
);
1776 static int plug_rq_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1778 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1779 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1781 if (rqa
->mq_ctx
!= rqb
->mq_ctx
)
1782 return rqa
->mq_ctx
> rqb
->mq_ctx
;
1783 if (rqa
->mq_hctx
!= rqb
->mq_hctx
)
1784 return rqa
->mq_hctx
> rqb
->mq_hctx
;
1786 return blk_rq_pos(rqa
) > blk_rq_pos(rqb
);
1789 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1793 if (list_empty(&plug
->mq_list
))
1795 list_splice_init(&plug
->mq_list
, &list
);
1797 if (plug
->rq_count
> 2 && plug
->multiple_queues
)
1798 list_sort(NULL
, &list
, plug_rq_cmp
);
1803 struct list_head rq_list
;
1804 struct request
*rq
, *head_rq
= list_entry_rq(list
.next
);
1805 struct list_head
*pos
= &head_rq
->queuelist
; /* skip first */
1806 struct blk_mq_hw_ctx
*this_hctx
= head_rq
->mq_hctx
;
1807 struct blk_mq_ctx
*this_ctx
= head_rq
->mq_ctx
;
1808 unsigned int depth
= 1;
1810 list_for_each_continue(pos
, &list
) {
1811 rq
= list_entry_rq(pos
);
1813 if (rq
->mq_hctx
!= this_hctx
|| rq
->mq_ctx
!= this_ctx
)
1818 list_cut_before(&rq_list
, &list
, pos
);
1819 trace_block_unplug(head_rq
->q
, depth
, !from_schedule
);
1820 blk_mq_sched_insert_requests(this_hctx
, this_ctx
, &rq_list
,
1822 } while(!list_empty(&list
));
1825 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
,
1826 unsigned int nr_segs
)
1828 if (bio
->bi_opf
& REQ_RAHEAD
)
1829 rq
->cmd_flags
|= REQ_FAILFAST_MASK
;
1831 rq
->__sector
= bio
->bi_iter
.bi_sector
;
1832 rq
->write_hint
= bio
->bi_write_hint
;
1833 blk_rq_bio_prep(rq
, bio
, nr_segs
);
1834 blk_crypto_rq_bio_prep(rq
, bio
, GFP_NOIO
);
1836 blk_account_io_start(rq
);
1839 static blk_status_t
__blk_mq_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1841 blk_qc_t
*cookie
, bool last
)
1843 struct request_queue
*q
= rq
->q
;
1844 struct blk_mq_queue_data bd
= {
1848 blk_qc_t new_cookie
;
1851 new_cookie
= request_to_qc_t(hctx
, rq
);
1854 * For OK queue, we are done. For error, caller may kill it.
1855 * Any other error (busy), just add it to our list as we
1856 * previously would have done.
1858 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1861 blk_mq_update_dispatch_busy(hctx
, false);
1862 *cookie
= new_cookie
;
1864 case BLK_STS_RESOURCE
:
1865 case BLK_STS_DEV_RESOURCE
:
1866 blk_mq_update_dispatch_busy(hctx
, true);
1867 __blk_mq_requeue_request(rq
);
1870 blk_mq_update_dispatch_busy(hctx
, false);
1871 *cookie
= BLK_QC_T_NONE
;
1878 static blk_status_t
__blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1881 bool bypass_insert
, bool last
)
1883 struct request_queue
*q
= rq
->q
;
1884 bool run_queue
= true;
1887 * RCU or SRCU read lock is needed before checking quiesced flag.
1889 * When queue is stopped or quiesced, ignore 'bypass_insert' from
1890 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1891 * and avoid driver to try to dispatch again.
1893 if (blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(q
)) {
1895 bypass_insert
= false;
1899 if (q
->elevator
&& !bypass_insert
)
1902 if (!blk_mq_get_dispatch_budget(hctx
))
1905 if (!blk_mq_get_driver_tag(rq
)) {
1906 blk_mq_put_dispatch_budget(hctx
);
1910 return __blk_mq_issue_directly(hctx
, rq
, cookie
, last
);
1913 return BLK_STS_RESOURCE
;
1915 blk_mq_request_bypass_insert(rq
, false, run_queue
);
1920 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
1921 * @hctx: Pointer of the associated hardware queue.
1922 * @rq: Pointer to request to be sent.
1923 * @cookie: Request queue cookie.
1925 * If the device has enough resources to accept a new request now, send the
1926 * request directly to device driver. Else, insert at hctx->dispatch queue, so
1927 * we can try send it another time in the future. Requests inserted at this
1928 * queue have higher priority.
1930 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1931 struct request
*rq
, blk_qc_t
*cookie
)
1936 might_sleep_if(hctx
->flags
& BLK_MQ_F_BLOCKING
);
1938 hctx_lock(hctx
, &srcu_idx
);
1940 ret
= __blk_mq_try_issue_directly(hctx
, rq
, cookie
, false, true);
1941 if (ret
== BLK_STS_RESOURCE
|| ret
== BLK_STS_DEV_RESOURCE
)
1942 blk_mq_request_bypass_insert(rq
, false, true);
1943 else if (ret
!= BLK_STS_OK
)
1944 blk_mq_end_request(rq
, ret
);
1946 hctx_unlock(hctx
, srcu_idx
);
1949 blk_status_t
blk_mq_request_issue_directly(struct request
*rq
, bool last
)
1953 blk_qc_t unused_cookie
;
1954 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1956 hctx_lock(hctx
, &srcu_idx
);
1957 ret
= __blk_mq_try_issue_directly(hctx
, rq
, &unused_cookie
, true, last
);
1958 hctx_unlock(hctx
, srcu_idx
);
1963 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx
*hctx
,
1964 struct list_head
*list
)
1968 while (!list_empty(list
)) {
1970 struct request
*rq
= list_first_entry(list
, struct request
,
1973 list_del_init(&rq
->queuelist
);
1974 ret
= blk_mq_request_issue_directly(rq
, list_empty(list
));
1975 if (ret
!= BLK_STS_OK
) {
1976 if (ret
== BLK_STS_RESOURCE
||
1977 ret
== BLK_STS_DEV_RESOURCE
) {
1978 blk_mq_request_bypass_insert(rq
, false,
1982 blk_mq_end_request(rq
, ret
);
1988 * If we didn't flush the entire list, we could have told
1989 * the driver there was more coming, but that turned out to
1992 if (!list_empty(list
) && hctx
->queue
->mq_ops
->commit_rqs
&& queued
)
1993 hctx
->queue
->mq_ops
->commit_rqs(hctx
);
1996 static void blk_add_rq_to_plug(struct blk_plug
*plug
, struct request
*rq
)
1998 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
2000 if (!plug
->multiple_queues
&& !list_is_singular(&plug
->mq_list
)) {
2001 struct request
*tmp
;
2003 tmp
= list_first_entry(&plug
->mq_list
, struct request
,
2005 if (tmp
->q
!= rq
->q
)
2006 plug
->multiple_queues
= true;
2011 * blk_mq_make_request - Create and send a request to block device.
2012 * @q: Request queue pointer.
2013 * @bio: Bio pointer.
2015 * Builds up a request structure from @q and @bio and send to the device. The
2016 * request may not be queued directly to hardware if:
2017 * * This request can be merged with another one
2018 * * We want to place request at plug queue for possible future merging
2019 * * There is an IO scheduler active at this queue
2021 * It will not queue the request if there is an error with the bio, or at the
2024 * Returns: Request queue cookie.
2026 blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
2028 const int is_sync
= op_is_sync(bio
->bi_opf
);
2029 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
2030 struct blk_mq_alloc_data data
= { .flags
= 0};
2032 struct blk_plug
*plug
;
2033 struct request
*same_queue_rq
= NULL
;
2034 unsigned int nr_segs
;
2038 blk_queue_bounce(q
, &bio
);
2039 __blk_queue_split(q
, &bio
, &nr_segs
);
2041 if (!bio_integrity_prep(bio
))
2044 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
2045 blk_attempt_plug_merge(q
, bio
, nr_segs
, &same_queue_rq
))
2048 if (blk_mq_sched_bio_merge(q
, bio
, nr_segs
))
2051 rq_qos_throttle(q
, bio
);
2053 data
.cmd_flags
= bio
->bi_opf
;
2054 rq
= blk_mq_get_request(q
, bio
, &data
);
2055 if (unlikely(!rq
)) {
2056 rq_qos_cleanup(q
, bio
);
2057 if (bio
->bi_opf
& REQ_NOWAIT
)
2058 bio_wouldblock_error(bio
);
2062 trace_block_getrq(q
, bio
, bio
->bi_opf
);
2064 rq_qos_track(q
, rq
, bio
);
2066 cookie
= request_to_qc_t(data
.hctx
, rq
);
2068 blk_mq_bio_to_request(rq
, bio
, nr_segs
);
2070 ret
= blk_crypto_init_request(rq
);
2071 if (ret
!= BLK_STS_OK
) {
2072 bio
->bi_status
= ret
;
2074 blk_mq_free_request(rq
);
2075 return BLK_QC_T_NONE
;
2078 plug
= blk_mq_plug(q
, bio
);
2079 if (unlikely(is_flush_fua
)) {
2080 /* Bypass scheduler for flush requests */
2081 blk_insert_flush(rq
);
2082 blk_mq_run_hw_queue(data
.hctx
, true);
2083 } else if (plug
&& (q
->nr_hw_queues
== 1 || q
->mq_ops
->commit_rqs
||
2084 !blk_queue_nonrot(q
))) {
2086 * Use plugging if we have a ->commit_rqs() hook as well, as
2087 * we know the driver uses bd->last in a smart fashion.
2089 * Use normal plugging if this disk is slow HDD, as sequential
2090 * IO may benefit a lot from plug merging.
2092 unsigned int request_count
= plug
->rq_count
;
2093 struct request
*last
= NULL
;
2096 trace_block_plug(q
);
2098 last
= list_entry_rq(plug
->mq_list
.prev
);
2100 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
2101 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
2102 blk_flush_plug_list(plug
, false);
2103 trace_block_plug(q
);
2106 blk_add_rq_to_plug(plug
, rq
);
2107 } else if (q
->elevator
) {
2108 /* Insert the request at the IO scheduler queue */
2109 blk_mq_sched_insert_request(rq
, false, true, true);
2110 } else if (plug
&& !blk_queue_nomerges(q
)) {
2112 * We do limited plugging. If the bio can be merged, do that.
2113 * Otherwise the existing request in the plug list will be
2114 * issued. So the plug list will have one request at most
2115 * The plug list might get flushed before this. If that happens,
2116 * the plug list is empty, and same_queue_rq is invalid.
2118 if (list_empty(&plug
->mq_list
))
2119 same_queue_rq
= NULL
;
2120 if (same_queue_rq
) {
2121 list_del_init(&same_queue_rq
->queuelist
);
2124 blk_add_rq_to_plug(plug
, rq
);
2125 trace_block_plug(q
);
2127 if (same_queue_rq
) {
2128 data
.hctx
= same_queue_rq
->mq_hctx
;
2129 trace_block_unplug(q
, 1, true);
2130 blk_mq_try_issue_directly(data
.hctx
, same_queue_rq
,
2133 } else if ((q
->nr_hw_queues
> 1 && is_sync
) ||
2134 !data
.hctx
->dispatch_busy
) {
2136 * There is no scheduler and we can try to send directly
2139 blk_mq_try_issue_directly(data
.hctx
, rq
, &cookie
);
2142 blk_mq_sched_insert_request(rq
, false, true, true);
2148 return BLK_QC_T_NONE
;
2150 EXPORT_SYMBOL_GPL(blk_mq_make_request
); /* only for request based dm */
2152 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
2153 unsigned int hctx_idx
)
2157 if (tags
->rqs
&& set
->ops
->exit_request
) {
2160 for (i
= 0; i
< tags
->nr_tags
; i
++) {
2161 struct request
*rq
= tags
->static_rqs
[i
];
2165 set
->ops
->exit_request(set
, rq
, hctx_idx
);
2166 tags
->static_rqs
[i
] = NULL
;
2170 while (!list_empty(&tags
->page_list
)) {
2171 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
2172 list_del_init(&page
->lru
);
2174 * Remove kmemleak object previously allocated in
2175 * blk_mq_alloc_rqs().
2177 kmemleak_free(page_address(page
));
2178 __free_pages(page
, page
->private);
2182 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
2186 kfree(tags
->static_rqs
);
2187 tags
->static_rqs
= NULL
;
2189 blk_mq_free_tags(tags
);
2192 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
2193 unsigned int hctx_idx
,
2194 unsigned int nr_tags
,
2195 unsigned int reserved_tags
)
2197 struct blk_mq_tags
*tags
;
2200 node
= blk_mq_hw_queue_to_node(&set
->map
[HCTX_TYPE_DEFAULT
], hctx_idx
);
2201 if (node
== NUMA_NO_NODE
)
2202 node
= set
->numa_node
;
2204 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
,
2205 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
2209 tags
->rqs
= kcalloc_node(nr_tags
, sizeof(struct request
*),
2210 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
2213 blk_mq_free_tags(tags
);
2217 tags
->static_rqs
= kcalloc_node(nr_tags
, sizeof(struct request
*),
2218 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
2220 if (!tags
->static_rqs
) {
2222 blk_mq_free_tags(tags
);
2229 static size_t order_to_size(unsigned int order
)
2231 return (size_t)PAGE_SIZE
<< order
;
2234 static int blk_mq_init_request(struct blk_mq_tag_set
*set
, struct request
*rq
,
2235 unsigned int hctx_idx
, int node
)
2239 if (set
->ops
->init_request
) {
2240 ret
= set
->ops
->init_request(set
, rq
, hctx_idx
, node
);
2245 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
2249 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
2250 unsigned int hctx_idx
, unsigned int depth
)
2252 unsigned int i
, j
, entries_per_page
, max_order
= 4;
2253 size_t rq_size
, left
;
2256 node
= blk_mq_hw_queue_to_node(&set
->map
[HCTX_TYPE_DEFAULT
], hctx_idx
);
2257 if (node
== NUMA_NO_NODE
)
2258 node
= set
->numa_node
;
2260 INIT_LIST_HEAD(&tags
->page_list
);
2263 * rq_size is the size of the request plus driver payload, rounded
2264 * to the cacheline size
2266 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
2268 left
= rq_size
* depth
;
2270 for (i
= 0; i
< depth
; ) {
2271 int this_order
= max_order
;
2276 while (this_order
&& left
< order_to_size(this_order
- 1))
2280 page
= alloc_pages_node(node
,
2281 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
2287 if (order_to_size(this_order
) < rq_size
)
2294 page
->private = this_order
;
2295 list_add_tail(&page
->lru
, &tags
->page_list
);
2297 p
= page_address(page
);
2299 * Allow kmemleak to scan these pages as they contain pointers
2300 * to additional allocations like via ops->init_request().
2302 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
2303 entries_per_page
= order_to_size(this_order
) / rq_size
;
2304 to_do
= min(entries_per_page
, depth
- i
);
2305 left
-= to_do
* rq_size
;
2306 for (j
= 0; j
< to_do
; j
++) {
2307 struct request
*rq
= p
;
2309 tags
->static_rqs
[i
] = rq
;
2310 if (blk_mq_init_request(set
, rq
, hctx_idx
, node
)) {
2311 tags
->static_rqs
[i
] = NULL
;
2322 blk_mq_free_rqs(set
, tags
, hctx_idx
);
2327 * 'cpu' is going away. splice any existing rq_list entries from this
2328 * software queue to the hw queue dispatch list, and ensure that it
2331 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
2333 struct blk_mq_hw_ctx
*hctx
;
2334 struct blk_mq_ctx
*ctx
;
2336 enum hctx_type type
;
2338 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
2339 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
2342 spin_lock(&ctx
->lock
);
2343 if (!list_empty(&ctx
->rq_lists
[type
])) {
2344 list_splice_init(&ctx
->rq_lists
[type
], &tmp
);
2345 blk_mq_hctx_clear_pending(hctx
, ctx
);
2347 spin_unlock(&ctx
->lock
);
2349 if (list_empty(&tmp
))
2352 spin_lock(&hctx
->lock
);
2353 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
2354 spin_unlock(&hctx
->lock
);
2356 blk_mq_run_hw_queue(hctx
, true);
2360 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
2362 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
2366 /* hctx->ctxs will be freed in queue's release handler */
2367 static void blk_mq_exit_hctx(struct request_queue
*q
,
2368 struct blk_mq_tag_set
*set
,
2369 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
2371 if (blk_mq_hw_queue_mapped(hctx
))
2372 blk_mq_tag_idle(hctx
);
2374 if (set
->ops
->exit_request
)
2375 set
->ops
->exit_request(set
, hctx
->fq
->flush_rq
, hctx_idx
);
2377 if (set
->ops
->exit_hctx
)
2378 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2380 blk_mq_remove_cpuhp(hctx
);
2382 spin_lock(&q
->unused_hctx_lock
);
2383 list_add(&hctx
->hctx_list
, &q
->unused_hctx_list
);
2384 spin_unlock(&q
->unused_hctx_lock
);
2387 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
2388 struct blk_mq_tag_set
*set
, int nr_queue
)
2390 struct blk_mq_hw_ctx
*hctx
;
2393 queue_for_each_hw_ctx(q
, hctx
, i
) {
2396 blk_mq_debugfs_unregister_hctx(hctx
);
2397 blk_mq_exit_hctx(q
, set
, hctx
, i
);
2401 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set
*tag_set
)
2403 int hw_ctx_size
= sizeof(struct blk_mq_hw_ctx
);
2405 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx
, srcu
),
2406 __alignof__(struct blk_mq_hw_ctx
)) !=
2407 sizeof(struct blk_mq_hw_ctx
));
2409 if (tag_set
->flags
& BLK_MQ_F_BLOCKING
)
2410 hw_ctx_size
+= sizeof(struct srcu_struct
);
2415 static int blk_mq_init_hctx(struct request_queue
*q
,
2416 struct blk_mq_tag_set
*set
,
2417 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
2419 hctx
->queue_num
= hctx_idx
;
2421 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
2423 hctx
->tags
= set
->tags
[hctx_idx
];
2425 if (set
->ops
->init_hctx
&&
2426 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
2427 goto unregister_cpu_notifier
;
2429 if (blk_mq_init_request(set
, hctx
->fq
->flush_rq
, hctx_idx
,
2435 if (set
->ops
->exit_hctx
)
2436 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2437 unregister_cpu_notifier
:
2438 blk_mq_remove_cpuhp(hctx
);
2442 static struct blk_mq_hw_ctx
*
2443 blk_mq_alloc_hctx(struct request_queue
*q
, struct blk_mq_tag_set
*set
,
2446 struct blk_mq_hw_ctx
*hctx
;
2447 gfp_t gfp
= GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
;
2449 hctx
= kzalloc_node(blk_mq_hw_ctx_size(set
), gfp
, node
);
2451 goto fail_alloc_hctx
;
2453 if (!zalloc_cpumask_var_node(&hctx
->cpumask
, gfp
, node
))
2456 atomic_set(&hctx
->nr_active
, 0);
2457 if (node
== NUMA_NO_NODE
)
2458 node
= set
->numa_node
;
2459 hctx
->numa_node
= node
;
2461 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
2462 spin_lock_init(&hctx
->lock
);
2463 INIT_LIST_HEAD(&hctx
->dispatch
);
2465 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
2467 INIT_LIST_HEAD(&hctx
->hctx_list
);
2470 * Allocate space for all possible cpus to avoid allocation at
2473 hctx
->ctxs
= kmalloc_array_node(nr_cpu_ids
, sizeof(void *),
2478 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8),
2483 spin_lock_init(&hctx
->dispatch_wait_lock
);
2484 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
2485 INIT_LIST_HEAD(&hctx
->dispatch_wait
.entry
);
2487 hctx
->fq
= blk_alloc_flush_queue(hctx
->numa_node
, set
->cmd_size
, gfp
);
2491 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
2492 init_srcu_struct(hctx
->srcu
);
2493 blk_mq_hctx_kobj_init(hctx
);
2498 sbitmap_free(&hctx
->ctx_map
);
2502 free_cpumask_var(hctx
->cpumask
);
2509 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
2510 unsigned int nr_hw_queues
)
2512 struct blk_mq_tag_set
*set
= q
->tag_set
;
2515 for_each_possible_cpu(i
) {
2516 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2517 struct blk_mq_hw_ctx
*hctx
;
2521 spin_lock_init(&__ctx
->lock
);
2522 for (k
= HCTX_TYPE_DEFAULT
; k
< HCTX_MAX_TYPES
; k
++)
2523 INIT_LIST_HEAD(&__ctx
->rq_lists
[k
]);
2528 * Set local node, IFF we have more than one hw queue. If
2529 * not, we remain on the home node of the device
2531 for (j
= 0; j
< set
->nr_maps
; j
++) {
2532 hctx
= blk_mq_map_queue_type(q
, j
, i
);
2533 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
2534 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
2539 static bool __blk_mq_alloc_map_and_request(struct blk_mq_tag_set
*set
,
2544 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
2545 set
->queue_depth
, set
->reserved_tags
);
2546 if (!set
->tags
[hctx_idx
])
2549 ret
= blk_mq_alloc_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
,
2554 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2555 set
->tags
[hctx_idx
] = NULL
;
2559 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set
*set
,
2560 unsigned int hctx_idx
)
2562 if (set
->tags
&& set
->tags
[hctx_idx
]) {
2563 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
2564 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2565 set
->tags
[hctx_idx
] = NULL
;
2569 static void blk_mq_map_swqueue(struct request_queue
*q
)
2571 unsigned int i
, j
, hctx_idx
;
2572 struct blk_mq_hw_ctx
*hctx
;
2573 struct blk_mq_ctx
*ctx
;
2574 struct blk_mq_tag_set
*set
= q
->tag_set
;
2576 queue_for_each_hw_ctx(q
, hctx
, i
) {
2577 cpumask_clear(hctx
->cpumask
);
2579 hctx
->dispatch_from
= NULL
;
2583 * Map software to hardware queues.
2585 * If the cpu isn't present, the cpu is mapped to first hctx.
2587 for_each_possible_cpu(i
) {
2589 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2590 for (j
= 0; j
< set
->nr_maps
; j
++) {
2591 if (!set
->map
[j
].nr_queues
) {
2592 ctx
->hctxs
[j
] = blk_mq_map_queue_type(q
,
2593 HCTX_TYPE_DEFAULT
, i
);
2596 hctx_idx
= set
->map
[j
].mq_map
[i
];
2597 /* unmapped hw queue can be remapped after CPU topo changed */
2598 if (!set
->tags
[hctx_idx
] &&
2599 !__blk_mq_alloc_map_and_request(set
, hctx_idx
)) {
2601 * If tags initialization fail for some hctx,
2602 * that hctx won't be brought online. In this
2603 * case, remap the current ctx to hctx[0] which
2604 * is guaranteed to always have tags allocated
2606 set
->map
[j
].mq_map
[i
] = 0;
2609 hctx
= blk_mq_map_queue_type(q
, j
, i
);
2610 ctx
->hctxs
[j
] = hctx
;
2612 * If the CPU is already set in the mask, then we've
2613 * mapped this one already. This can happen if
2614 * devices share queues across queue maps.
2616 if (cpumask_test_cpu(i
, hctx
->cpumask
))
2619 cpumask_set_cpu(i
, hctx
->cpumask
);
2621 ctx
->index_hw
[hctx
->type
] = hctx
->nr_ctx
;
2622 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2625 * If the nr_ctx type overflows, we have exceeded the
2626 * amount of sw queues we can support.
2628 BUG_ON(!hctx
->nr_ctx
);
2631 for (; j
< HCTX_MAX_TYPES
; j
++)
2632 ctx
->hctxs
[j
] = blk_mq_map_queue_type(q
,
2633 HCTX_TYPE_DEFAULT
, i
);
2636 queue_for_each_hw_ctx(q
, hctx
, i
) {
2638 * If no software queues are mapped to this hardware queue,
2639 * disable it and free the request entries.
2641 if (!hctx
->nr_ctx
) {
2642 /* Never unmap queue 0. We need it as a
2643 * fallback in case of a new remap fails
2646 if (i
&& set
->tags
[i
])
2647 blk_mq_free_map_and_requests(set
, i
);
2653 hctx
->tags
= set
->tags
[i
];
2654 WARN_ON(!hctx
->tags
);
2657 * Set the map size to the number of mapped software queues.
2658 * This is more accurate and more efficient than looping
2659 * over all possibly mapped software queues.
2661 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2664 * Initialize batch roundrobin counts
2666 hctx
->next_cpu
= blk_mq_first_mapped_cpu(hctx
);
2667 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2672 * Caller needs to ensure that we're either frozen/quiesced, or that
2673 * the queue isn't live yet.
2675 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2677 struct blk_mq_hw_ctx
*hctx
;
2680 queue_for_each_hw_ctx(q
, hctx
, i
) {
2682 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
2684 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2688 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
,
2691 struct request_queue
*q
;
2693 lockdep_assert_held(&set
->tag_list_lock
);
2695 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2696 blk_mq_freeze_queue(q
);
2697 queue_set_hctx_shared(q
, shared
);
2698 blk_mq_unfreeze_queue(q
);
2702 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
2704 struct blk_mq_tag_set
*set
= q
->tag_set
;
2706 mutex_lock(&set
->tag_list_lock
);
2707 list_del_rcu(&q
->tag_set_list
);
2708 if (list_is_singular(&set
->tag_list
)) {
2709 /* just transitioned to unshared */
2710 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2711 /* update existing queue */
2712 blk_mq_update_tag_set_depth(set
, false);
2714 mutex_unlock(&set
->tag_list_lock
);
2715 INIT_LIST_HEAD(&q
->tag_set_list
);
2718 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
2719 struct request_queue
*q
)
2721 mutex_lock(&set
->tag_list_lock
);
2724 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2726 if (!list_empty(&set
->tag_list
) &&
2727 !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
2728 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
2729 /* update existing queue */
2730 blk_mq_update_tag_set_depth(set
, true);
2732 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
2733 queue_set_hctx_shared(q
, true);
2734 list_add_tail_rcu(&q
->tag_set_list
, &set
->tag_list
);
2736 mutex_unlock(&set
->tag_list_lock
);
2739 /* All allocations will be freed in release handler of q->mq_kobj */
2740 static int blk_mq_alloc_ctxs(struct request_queue
*q
)
2742 struct blk_mq_ctxs
*ctxs
;
2745 ctxs
= kzalloc(sizeof(*ctxs
), GFP_KERNEL
);
2749 ctxs
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2750 if (!ctxs
->queue_ctx
)
2753 for_each_possible_cpu(cpu
) {
2754 struct blk_mq_ctx
*ctx
= per_cpu_ptr(ctxs
->queue_ctx
, cpu
);
2758 q
->mq_kobj
= &ctxs
->kobj
;
2759 q
->queue_ctx
= ctxs
->queue_ctx
;
2768 * It is the actual release handler for mq, but we do it from
2769 * request queue's release handler for avoiding use-after-free
2770 * and headache because q->mq_kobj shouldn't have been introduced,
2771 * but we can't group ctx/kctx kobj without it.
2773 void blk_mq_release(struct request_queue
*q
)
2775 struct blk_mq_hw_ctx
*hctx
, *next
;
2778 queue_for_each_hw_ctx(q
, hctx
, i
)
2779 WARN_ON_ONCE(hctx
&& list_empty(&hctx
->hctx_list
));
2781 /* all hctx are in .unused_hctx_list now */
2782 list_for_each_entry_safe(hctx
, next
, &q
->unused_hctx_list
, hctx_list
) {
2783 list_del_init(&hctx
->hctx_list
);
2784 kobject_put(&hctx
->kobj
);
2787 kfree(q
->queue_hw_ctx
);
2790 * release .mq_kobj and sw queue's kobject now because
2791 * both share lifetime with request queue.
2793 blk_mq_sysfs_deinit(q
);
2796 struct request_queue
*blk_mq_init_queue_data(struct blk_mq_tag_set
*set
,
2799 struct request_queue
*uninit_q
, *q
;
2801 uninit_q
= __blk_alloc_queue(set
->numa_node
);
2803 return ERR_PTR(-ENOMEM
);
2804 uninit_q
->queuedata
= queuedata
;
2807 * Initialize the queue without an elevator. device_add_disk() will do
2808 * the initialization.
2810 q
= blk_mq_init_allocated_queue(set
, uninit_q
, false);
2812 blk_cleanup_queue(uninit_q
);
2816 EXPORT_SYMBOL_GPL(blk_mq_init_queue_data
);
2818 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2820 return blk_mq_init_queue_data(set
, NULL
);
2822 EXPORT_SYMBOL(blk_mq_init_queue
);
2825 * Helper for setting up a queue with mq ops, given queue depth, and
2826 * the passed in mq ops flags.
2828 struct request_queue
*blk_mq_init_sq_queue(struct blk_mq_tag_set
*set
,
2829 const struct blk_mq_ops
*ops
,
2830 unsigned int queue_depth
,
2831 unsigned int set_flags
)
2833 struct request_queue
*q
;
2836 memset(set
, 0, sizeof(*set
));
2838 set
->nr_hw_queues
= 1;
2840 set
->queue_depth
= queue_depth
;
2841 set
->numa_node
= NUMA_NO_NODE
;
2842 set
->flags
= set_flags
;
2844 ret
= blk_mq_alloc_tag_set(set
);
2846 return ERR_PTR(ret
);
2848 q
= blk_mq_init_queue(set
);
2850 blk_mq_free_tag_set(set
);
2856 EXPORT_SYMBOL(blk_mq_init_sq_queue
);
2858 static struct blk_mq_hw_ctx
*blk_mq_alloc_and_init_hctx(
2859 struct blk_mq_tag_set
*set
, struct request_queue
*q
,
2860 int hctx_idx
, int node
)
2862 struct blk_mq_hw_ctx
*hctx
= NULL
, *tmp
;
2864 /* reuse dead hctx first */
2865 spin_lock(&q
->unused_hctx_lock
);
2866 list_for_each_entry(tmp
, &q
->unused_hctx_list
, hctx_list
) {
2867 if (tmp
->numa_node
== node
) {
2873 list_del_init(&hctx
->hctx_list
);
2874 spin_unlock(&q
->unused_hctx_lock
);
2877 hctx
= blk_mq_alloc_hctx(q
, set
, node
);
2881 if (blk_mq_init_hctx(q
, set
, hctx
, hctx_idx
))
2887 kobject_put(&hctx
->kobj
);
2892 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2893 struct request_queue
*q
)
2896 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2898 if (q
->nr_hw_queues
< set
->nr_hw_queues
) {
2899 struct blk_mq_hw_ctx
**new_hctxs
;
2901 new_hctxs
= kcalloc_node(set
->nr_hw_queues
,
2902 sizeof(*new_hctxs
), GFP_KERNEL
,
2907 memcpy(new_hctxs
, hctxs
, q
->nr_hw_queues
*
2909 q
->queue_hw_ctx
= new_hctxs
;
2914 /* protect against switching io scheduler */
2915 mutex_lock(&q
->sysfs_lock
);
2916 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2918 struct blk_mq_hw_ctx
*hctx
;
2920 node
= blk_mq_hw_queue_to_node(&set
->map
[HCTX_TYPE_DEFAULT
], i
);
2922 * If the hw queue has been mapped to another numa node,
2923 * we need to realloc the hctx. If allocation fails, fallback
2924 * to use the previous one.
2926 if (hctxs
[i
] && (hctxs
[i
]->numa_node
== node
))
2929 hctx
= blk_mq_alloc_and_init_hctx(set
, q
, i
, node
);
2932 blk_mq_exit_hctx(q
, set
, hctxs
[i
], i
);
2936 pr_warn("Allocate new hctx on node %d fails,\
2937 fallback to previous one on node %d\n",
2938 node
, hctxs
[i
]->numa_node
);
2944 * Increasing nr_hw_queues fails. Free the newly allocated
2945 * hctxs and keep the previous q->nr_hw_queues.
2947 if (i
!= set
->nr_hw_queues
) {
2948 j
= q
->nr_hw_queues
;
2952 end
= q
->nr_hw_queues
;
2953 q
->nr_hw_queues
= set
->nr_hw_queues
;
2956 for (; j
< end
; j
++) {
2957 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2961 blk_mq_free_map_and_requests(set
, j
);
2962 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2966 mutex_unlock(&q
->sysfs_lock
);
2969 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2970 struct request_queue
*q
,
2973 /* mark the queue as mq asap */
2974 q
->mq_ops
= set
->ops
;
2976 q
->poll_cb
= blk_stat_alloc_callback(blk_mq_poll_stats_fn
,
2977 blk_mq_poll_stats_bkt
,
2978 BLK_MQ_POLL_STATS_BKTS
, q
);
2982 if (blk_mq_alloc_ctxs(q
))
2985 /* init q->mq_kobj and sw queues' kobjects */
2986 blk_mq_sysfs_init(q
);
2988 INIT_LIST_HEAD(&q
->unused_hctx_list
);
2989 spin_lock_init(&q
->unused_hctx_lock
);
2991 blk_mq_realloc_hw_ctxs(set
, q
);
2992 if (!q
->nr_hw_queues
)
2995 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2996 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
3000 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
3001 if (set
->nr_maps
> HCTX_TYPE_POLL
&&
3002 set
->map
[HCTX_TYPE_POLL
].nr_queues
)
3003 blk_queue_flag_set(QUEUE_FLAG_POLL
, q
);
3005 q
->sg_reserved_size
= INT_MAX
;
3007 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
3008 INIT_LIST_HEAD(&q
->requeue_list
);
3009 spin_lock_init(&q
->requeue_lock
);
3011 q
->nr_requests
= set
->queue_depth
;
3014 * Default to classic polling
3016 q
->poll_nsec
= BLK_MQ_POLL_CLASSIC
;
3018 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
3019 blk_mq_add_queue_tag_set(set
, q
);
3020 blk_mq_map_swqueue(q
);
3023 elevator_init_mq(q
);
3028 kfree(q
->queue_hw_ctx
);
3029 q
->nr_hw_queues
= 0;
3030 blk_mq_sysfs_deinit(q
);
3032 blk_stat_free_callback(q
->poll_cb
);
3036 return ERR_PTR(-ENOMEM
);
3038 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
3040 /* tags can _not_ be used after returning from blk_mq_exit_queue */
3041 void blk_mq_exit_queue(struct request_queue
*q
)
3043 struct blk_mq_tag_set
*set
= q
->tag_set
;
3045 blk_mq_del_queue_tag_set(q
);
3046 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
3049 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
3053 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
3054 if (!__blk_mq_alloc_map_and_request(set
, i
))
3061 blk_mq_free_map_and_requests(set
, i
);
3067 * Allocate the request maps associated with this tag_set. Note that this
3068 * may reduce the depth asked for, if memory is tight. set->queue_depth
3069 * will be updated to reflect the allocated depth.
3071 static int blk_mq_alloc_map_and_requests(struct blk_mq_tag_set
*set
)
3076 depth
= set
->queue_depth
;
3078 err
= __blk_mq_alloc_rq_maps(set
);
3082 set
->queue_depth
>>= 1;
3083 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
3087 } while (set
->queue_depth
);
3089 if (!set
->queue_depth
|| err
) {
3090 pr_err("blk-mq: failed to allocate request map\n");
3094 if (depth
!= set
->queue_depth
)
3095 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3096 depth
, set
->queue_depth
);
3101 static int blk_mq_update_queue_map(struct blk_mq_tag_set
*set
)
3104 * blk_mq_map_queues() and multiple .map_queues() implementations
3105 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
3106 * number of hardware queues.
3108 if (set
->nr_maps
== 1)
3109 set
->map
[HCTX_TYPE_DEFAULT
].nr_queues
= set
->nr_hw_queues
;
3111 if (set
->ops
->map_queues
&& !is_kdump_kernel()) {
3115 * transport .map_queues is usually done in the following
3118 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3119 * mask = get_cpu_mask(queue)
3120 * for_each_cpu(cpu, mask)
3121 * set->map[x].mq_map[cpu] = queue;
3124 * When we need to remap, the table has to be cleared for
3125 * killing stale mapping since one CPU may not be mapped
3128 for (i
= 0; i
< set
->nr_maps
; i
++)
3129 blk_mq_clear_mq_map(&set
->map
[i
]);
3131 return set
->ops
->map_queues(set
);
3133 BUG_ON(set
->nr_maps
> 1);
3134 return blk_mq_map_queues(&set
->map
[HCTX_TYPE_DEFAULT
]);
3138 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set
*set
,
3139 int cur_nr_hw_queues
, int new_nr_hw_queues
)
3141 struct blk_mq_tags
**new_tags
;
3143 if (cur_nr_hw_queues
>= new_nr_hw_queues
)
3146 new_tags
= kcalloc_node(new_nr_hw_queues
, sizeof(struct blk_mq_tags
*),
3147 GFP_KERNEL
, set
->numa_node
);
3152 memcpy(new_tags
, set
->tags
, cur_nr_hw_queues
*
3153 sizeof(*set
->tags
));
3155 set
->tags
= new_tags
;
3156 set
->nr_hw_queues
= new_nr_hw_queues
;
3162 * Alloc a tag set to be associated with one or more request queues.
3163 * May fail with EINVAL for various error conditions. May adjust the
3164 * requested depth down, if it's too large. In that case, the set
3165 * value will be stored in set->queue_depth.
3167 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
3171 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
3173 if (!set
->nr_hw_queues
)
3175 if (!set
->queue_depth
)
3177 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
3180 if (!set
->ops
->queue_rq
)
3183 if (!set
->ops
->get_budget
^ !set
->ops
->put_budget
)
3186 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
3187 pr_info("blk-mq: reduced tag depth to %u\n",
3189 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
3194 else if (set
->nr_maps
> HCTX_MAX_TYPES
)
3198 * If a crashdump is active, then we are potentially in a very
3199 * memory constrained environment. Limit us to 1 queue and
3200 * 64 tags to prevent using too much memory.
3202 if (is_kdump_kernel()) {
3203 set
->nr_hw_queues
= 1;
3205 set
->queue_depth
= min(64U, set
->queue_depth
);
3208 * There is no use for more h/w queues than cpus if we just have
3211 if (set
->nr_maps
== 1 && set
->nr_hw_queues
> nr_cpu_ids
)
3212 set
->nr_hw_queues
= nr_cpu_ids
;
3214 if (blk_mq_realloc_tag_set_tags(set
, 0, set
->nr_hw_queues
) < 0)
3218 for (i
= 0; i
< set
->nr_maps
; i
++) {
3219 set
->map
[i
].mq_map
= kcalloc_node(nr_cpu_ids
,
3220 sizeof(set
->map
[i
].mq_map
[0]),
3221 GFP_KERNEL
, set
->numa_node
);
3222 if (!set
->map
[i
].mq_map
)
3223 goto out_free_mq_map
;
3224 set
->map
[i
].nr_queues
= is_kdump_kernel() ? 1 : set
->nr_hw_queues
;
3227 ret
= blk_mq_update_queue_map(set
);
3229 goto out_free_mq_map
;
3231 ret
= blk_mq_alloc_map_and_requests(set
);
3233 goto out_free_mq_map
;
3235 mutex_init(&set
->tag_list_lock
);
3236 INIT_LIST_HEAD(&set
->tag_list
);
3241 for (i
= 0; i
< set
->nr_maps
; i
++) {
3242 kfree(set
->map
[i
].mq_map
);
3243 set
->map
[i
].mq_map
= NULL
;
3249 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
3251 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
3255 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
3256 blk_mq_free_map_and_requests(set
, i
);
3258 for (j
= 0; j
< set
->nr_maps
; j
++) {
3259 kfree(set
->map
[j
].mq_map
);
3260 set
->map
[j
].mq_map
= NULL
;
3266 EXPORT_SYMBOL(blk_mq_free_tag_set
);
3268 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
3270 struct blk_mq_tag_set
*set
= q
->tag_set
;
3271 struct blk_mq_hw_ctx
*hctx
;
3277 if (q
->nr_requests
== nr
)
3280 blk_mq_freeze_queue(q
);
3281 blk_mq_quiesce_queue(q
);
3284 queue_for_each_hw_ctx(q
, hctx
, i
) {
3288 * If we're using an MQ scheduler, just update the scheduler
3289 * queue depth. This is similar to what the old code would do.
3291 if (!hctx
->sched_tags
) {
3292 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
, nr
,
3295 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
3300 if (q
->elevator
&& q
->elevator
->type
->ops
.depth_updated
)
3301 q
->elevator
->type
->ops
.depth_updated(hctx
);
3305 q
->nr_requests
= nr
;
3307 blk_mq_unquiesce_queue(q
);
3308 blk_mq_unfreeze_queue(q
);
3314 * request_queue and elevator_type pair.
3315 * It is just used by __blk_mq_update_nr_hw_queues to cache
3316 * the elevator_type associated with a request_queue.
3318 struct blk_mq_qe_pair
{
3319 struct list_head node
;
3320 struct request_queue
*q
;
3321 struct elevator_type
*type
;
3325 * Cache the elevator_type in qe pair list and switch the
3326 * io scheduler to 'none'
3328 static bool blk_mq_elv_switch_none(struct list_head
*head
,
3329 struct request_queue
*q
)
3331 struct blk_mq_qe_pair
*qe
;
3336 qe
= kmalloc(sizeof(*qe
), GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
);
3340 INIT_LIST_HEAD(&qe
->node
);
3342 qe
->type
= q
->elevator
->type
;
3343 list_add(&qe
->node
, head
);
3345 mutex_lock(&q
->sysfs_lock
);
3347 * After elevator_switch_mq, the previous elevator_queue will be
3348 * released by elevator_release. The reference of the io scheduler
3349 * module get by elevator_get will also be put. So we need to get
3350 * a reference of the io scheduler module here to prevent it to be
3353 __module_get(qe
->type
->elevator_owner
);
3354 elevator_switch_mq(q
, NULL
);
3355 mutex_unlock(&q
->sysfs_lock
);
3360 static void blk_mq_elv_switch_back(struct list_head
*head
,
3361 struct request_queue
*q
)
3363 struct blk_mq_qe_pair
*qe
;
3364 struct elevator_type
*t
= NULL
;
3366 list_for_each_entry(qe
, head
, node
)
3375 list_del(&qe
->node
);
3378 mutex_lock(&q
->sysfs_lock
);
3379 elevator_switch_mq(q
, t
);
3380 mutex_unlock(&q
->sysfs_lock
);
3383 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
,
3386 struct request_queue
*q
;
3388 int prev_nr_hw_queues
;
3390 lockdep_assert_held(&set
->tag_list_lock
);
3392 if (set
->nr_maps
== 1 && nr_hw_queues
> nr_cpu_ids
)
3393 nr_hw_queues
= nr_cpu_ids
;
3394 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
3397 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3398 blk_mq_freeze_queue(q
);
3400 * Switch IO scheduler to 'none', cleaning up the data associated
3401 * with the previous scheduler. We will switch back once we are done
3402 * updating the new sw to hw queue mappings.
3404 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3405 if (!blk_mq_elv_switch_none(&head
, q
))
3408 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3409 blk_mq_debugfs_unregister_hctxs(q
);
3410 blk_mq_sysfs_unregister(q
);
3413 prev_nr_hw_queues
= set
->nr_hw_queues
;
3414 if (blk_mq_realloc_tag_set_tags(set
, set
->nr_hw_queues
, nr_hw_queues
) <
3418 set
->nr_hw_queues
= nr_hw_queues
;
3420 blk_mq_update_queue_map(set
);
3421 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3422 blk_mq_realloc_hw_ctxs(set
, q
);
3423 if (q
->nr_hw_queues
!= set
->nr_hw_queues
) {
3424 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3425 nr_hw_queues
, prev_nr_hw_queues
);
3426 set
->nr_hw_queues
= prev_nr_hw_queues
;
3427 blk_mq_map_queues(&set
->map
[HCTX_TYPE_DEFAULT
]);
3430 blk_mq_map_swqueue(q
);
3434 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3435 blk_mq_sysfs_register(q
);
3436 blk_mq_debugfs_register_hctxs(q
);
3440 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3441 blk_mq_elv_switch_back(&head
, q
);
3443 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3444 blk_mq_unfreeze_queue(q
);
3447 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
3449 mutex_lock(&set
->tag_list_lock
);
3450 __blk_mq_update_nr_hw_queues(set
, nr_hw_queues
);
3451 mutex_unlock(&set
->tag_list_lock
);
3453 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
3455 /* Enable polling stats and return whether they were already enabled. */
3456 static bool blk_poll_stats_enable(struct request_queue
*q
)
3458 if (test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
3459 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS
, q
))
3461 blk_stat_add_callback(q
, q
->poll_cb
);
3465 static void blk_mq_poll_stats_start(struct request_queue
*q
)
3468 * We don't arm the callback if polling stats are not enabled or the
3469 * callback is already active.
3471 if (!test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
3472 blk_stat_is_active(q
->poll_cb
))
3475 blk_stat_activate_msecs(q
->poll_cb
, 100);
3478 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
)
3480 struct request_queue
*q
= cb
->data
;
3483 for (bucket
= 0; bucket
< BLK_MQ_POLL_STATS_BKTS
; bucket
++) {
3484 if (cb
->stat
[bucket
].nr_samples
)
3485 q
->poll_stat
[bucket
] = cb
->stat
[bucket
];
3489 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
3492 unsigned long ret
= 0;
3496 * If stats collection isn't on, don't sleep but turn it on for
3499 if (!blk_poll_stats_enable(q
))
3503 * As an optimistic guess, use half of the mean service time
3504 * for this type of request. We can (and should) make this smarter.
3505 * For instance, if the completion latencies are tight, we can
3506 * get closer than just half the mean. This is especially
3507 * important on devices where the completion latencies are longer
3508 * than ~10 usec. We do use the stats for the relevant IO size
3509 * if available which does lead to better estimates.
3511 bucket
= blk_mq_poll_stats_bkt(rq
);
3515 if (q
->poll_stat
[bucket
].nr_samples
)
3516 ret
= (q
->poll_stat
[bucket
].mean
+ 1) / 2;
3521 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
3524 struct hrtimer_sleeper hs
;
3525 enum hrtimer_mode mode
;
3529 if (rq
->rq_flags
& RQF_MQ_POLL_SLEPT
)
3533 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3535 * 0: use half of prev avg
3536 * >0: use this specific value
3538 if (q
->poll_nsec
> 0)
3539 nsecs
= q
->poll_nsec
;
3541 nsecs
= blk_mq_poll_nsecs(q
, rq
);
3546 rq
->rq_flags
|= RQF_MQ_POLL_SLEPT
;
3549 * This will be replaced with the stats tracking code, using
3550 * 'avg_completion_time / 2' as the pre-sleep target.
3554 mode
= HRTIMER_MODE_REL
;
3555 hrtimer_init_sleeper_on_stack(&hs
, CLOCK_MONOTONIC
, mode
);
3556 hrtimer_set_expires(&hs
.timer
, kt
);
3559 if (blk_mq_rq_state(rq
) == MQ_RQ_COMPLETE
)
3561 set_current_state(TASK_UNINTERRUPTIBLE
);
3562 hrtimer_sleeper_start_expires(&hs
, mode
);
3565 hrtimer_cancel(&hs
.timer
);
3566 mode
= HRTIMER_MODE_ABS
;
3567 } while (hs
.task
&& !signal_pending(current
));
3569 __set_current_state(TASK_RUNNING
);
3570 destroy_hrtimer_on_stack(&hs
.timer
);
3574 static bool blk_mq_poll_hybrid(struct request_queue
*q
,
3575 struct blk_mq_hw_ctx
*hctx
, blk_qc_t cookie
)
3579 if (q
->poll_nsec
== BLK_MQ_POLL_CLASSIC
)
3582 if (!blk_qc_t_is_internal(cookie
))
3583 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
3585 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
3587 * With scheduling, if the request has completed, we'll
3588 * get a NULL return here, as we clear the sched tag when
3589 * that happens. The request still remains valid, like always,
3590 * so we should be safe with just the NULL check.
3596 return blk_mq_poll_hybrid_sleep(q
, rq
);
3600 * blk_poll - poll for IO completions
3602 * @cookie: cookie passed back at IO submission time
3603 * @spin: whether to spin for completions
3606 * Poll for completions on the passed in queue. Returns number of
3607 * completed entries found. If @spin is true, then blk_poll will continue
3608 * looping until at least one completion is found, unless the task is
3609 * otherwise marked running (or we need to reschedule).
3611 int blk_poll(struct request_queue
*q
, blk_qc_t cookie
, bool spin
)
3613 struct blk_mq_hw_ctx
*hctx
;
3616 if (!blk_qc_t_valid(cookie
) ||
3617 !test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
3621 blk_flush_plug_list(current
->plug
, false);
3623 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
3626 * If we sleep, have the caller restart the poll loop to reset
3627 * the state. Like for the other success return cases, the
3628 * caller is responsible for checking if the IO completed. If
3629 * the IO isn't complete, we'll get called again and will go
3630 * straight to the busy poll loop.
3632 if (blk_mq_poll_hybrid(q
, hctx
, cookie
))
3635 hctx
->poll_considered
++;
3637 state
= current
->state
;
3641 hctx
->poll_invoked
++;
3643 ret
= q
->mq_ops
->poll(hctx
);
3645 hctx
->poll_success
++;
3646 __set_current_state(TASK_RUNNING
);
3650 if (signal_pending_state(state
, current
))
3651 __set_current_state(TASK_RUNNING
);
3653 if (current
->state
== TASK_RUNNING
)
3655 if (ret
< 0 || !spin
)
3658 } while (!need_resched());
3660 __set_current_state(TASK_RUNNING
);
3663 EXPORT_SYMBOL_GPL(blk_poll
);
3665 unsigned int blk_mq_rq_cpu(struct request
*rq
)
3667 return rq
->mq_ctx
->cpu
;
3669 EXPORT_SYMBOL(blk_mq_rq_cpu
);
3671 static int __init
blk_mq_init(void)
3673 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
3674 blk_mq_hctx_notify_dead
);
3677 subsys_initcall(blk_mq_init
);