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>
30 #include <trace/events/block.h>
32 #include <linux/blk-mq.h>
33 #include <linux/t10-pi.h>
36 #include "blk-mq-debugfs.h"
37 #include "blk-mq-tag.h"
40 #include "blk-mq-sched.h"
41 #include "blk-rq-qos.h"
43 static void blk_mq_poll_stats_start(struct request_queue
*q
);
44 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
);
46 static int blk_mq_poll_stats_bkt(const struct request
*rq
)
48 int ddir
, sectors
, bucket
;
50 ddir
= rq_data_dir(rq
);
51 sectors
= blk_rq_stats_sectors(rq
);
53 bucket
= ddir
+ 2 * ilog2(sectors
);
57 else if (bucket
>= BLK_MQ_POLL_STATS_BKTS
)
58 return ddir
+ BLK_MQ_POLL_STATS_BKTS
- 2;
64 * Check if any of the ctx, dispatch list or elevator
65 * have pending work in this hardware queue.
67 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
69 return !list_empty_careful(&hctx
->dispatch
) ||
70 sbitmap_any_bit_set(&hctx
->ctx_map
) ||
71 blk_mq_sched_has_work(hctx
);
75 * Mark this ctx as having pending work in this hardware queue
77 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
78 struct blk_mq_ctx
*ctx
)
80 const int bit
= ctx
->index_hw
[hctx
->type
];
82 if (!sbitmap_test_bit(&hctx
->ctx_map
, bit
))
83 sbitmap_set_bit(&hctx
->ctx_map
, bit
);
86 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
87 struct blk_mq_ctx
*ctx
)
89 const int bit
= ctx
->index_hw
[hctx
->type
];
91 sbitmap_clear_bit(&hctx
->ctx_map
, bit
);
95 struct hd_struct
*part
;
96 unsigned int inflight
[2];
99 static bool blk_mq_check_inflight(struct blk_mq_hw_ctx
*hctx
,
100 struct request
*rq
, void *priv
,
103 struct mq_inflight
*mi
= priv
;
105 if (rq
->part
== mi
->part
)
106 mi
->inflight
[rq_data_dir(rq
)]++;
111 unsigned int blk_mq_in_flight(struct request_queue
*q
, struct hd_struct
*part
)
113 struct mq_inflight mi
= { .part
= part
};
115 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight
, &mi
);
117 return mi
.inflight
[0] + mi
.inflight
[1];
120 void blk_mq_in_flight_rw(struct request_queue
*q
, struct hd_struct
*part
,
121 unsigned int inflight
[2])
123 struct mq_inflight mi
= { .part
= part
};
125 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight
, &mi
);
126 inflight
[0] = mi
.inflight
[0];
127 inflight
[1] = mi
.inflight
[1];
130 void blk_freeze_queue_start(struct request_queue
*q
)
132 mutex_lock(&q
->mq_freeze_lock
);
133 if (++q
->mq_freeze_depth
== 1) {
134 percpu_ref_kill(&q
->q_usage_counter
);
135 mutex_unlock(&q
->mq_freeze_lock
);
137 blk_mq_run_hw_queues(q
, false);
139 mutex_unlock(&q
->mq_freeze_lock
);
142 EXPORT_SYMBOL_GPL(blk_freeze_queue_start
);
144 void blk_mq_freeze_queue_wait(struct request_queue
*q
)
146 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
148 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait
);
150 int blk_mq_freeze_queue_wait_timeout(struct request_queue
*q
,
151 unsigned long timeout
)
153 return wait_event_timeout(q
->mq_freeze_wq
,
154 percpu_ref_is_zero(&q
->q_usage_counter
),
157 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout
);
160 * Guarantee no request is in use, so we can change any data structure of
161 * the queue afterward.
163 void blk_freeze_queue(struct request_queue
*q
)
166 * In the !blk_mq case we are only calling this to kill the
167 * q_usage_counter, otherwise this increases the freeze depth
168 * and waits for it to return to zero. For this reason there is
169 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
170 * exported to drivers as the only user for unfreeze is blk_mq.
172 blk_freeze_queue_start(q
);
173 blk_mq_freeze_queue_wait(q
);
176 void blk_mq_freeze_queue(struct request_queue
*q
)
179 * ...just an alias to keep freeze and unfreeze actions balanced
180 * in the blk_mq_* namespace
184 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
186 void blk_mq_unfreeze_queue(struct request_queue
*q
)
188 mutex_lock(&q
->mq_freeze_lock
);
189 q
->mq_freeze_depth
--;
190 WARN_ON_ONCE(q
->mq_freeze_depth
< 0);
191 if (!q
->mq_freeze_depth
) {
192 percpu_ref_resurrect(&q
->q_usage_counter
);
193 wake_up_all(&q
->mq_freeze_wq
);
195 mutex_unlock(&q
->mq_freeze_lock
);
197 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
200 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
201 * mpt3sas driver such that this function can be removed.
203 void blk_mq_quiesce_queue_nowait(struct request_queue
*q
)
205 blk_queue_flag_set(QUEUE_FLAG_QUIESCED
, q
);
207 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait
);
210 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
213 * Note: this function does not prevent that the struct request end_io()
214 * callback function is invoked. Once this function is returned, we make
215 * sure no dispatch can happen until the queue is unquiesced via
216 * blk_mq_unquiesce_queue().
218 void blk_mq_quiesce_queue(struct request_queue
*q
)
220 struct blk_mq_hw_ctx
*hctx
;
224 blk_mq_quiesce_queue_nowait(q
);
226 queue_for_each_hw_ctx(q
, hctx
, i
) {
227 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
228 synchronize_srcu(hctx
->srcu
);
235 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
238 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
241 * This function recovers queue into the state before quiescing
242 * which is done by blk_mq_quiesce_queue.
244 void blk_mq_unquiesce_queue(struct request_queue
*q
)
246 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED
, q
);
248 /* dispatch requests which are inserted during quiescing */
249 blk_mq_run_hw_queues(q
, true);
251 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue
);
253 void blk_mq_wake_waiters(struct request_queue
*q
)
255 struct blk_mq_hw_ctx
*hctx
;
258 queue_for_each_hw_ctx(q
, hctx
, i
)
259 if (blk_mq_hw_queue_mapped(hctx
))
260 blk_mq_tag_wakeup_all(hctx
->tags
, true);
264 * Only need start/end time stamping if we have iostat or
265 * blk stats enabled, or using an IO scheduler.
267 static inline bool blk_mq_need_time_stamp(struct request
*rq
)
269 return (rq
->rq_flags
& (RQF_IO_STAT
| RQF_STATS
)) || rq
->q
->elevator
;
272 static struct request
*blk_mq_rq_ctx_init(struct blk_mq_alloc_data
*data
,
273 unsigned int tag
, unsigned int op
, u64 alloc_time_ns
)
275 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
276 struct request
*rq
= tags
->static_rqs
[tag
];
277 req_flags_t rq_flags
= 0;
279 if (data
->flags
& BLK_MQ_REQ_INTERNAL
) {
281 rq
->internal_tag
= tag
;
283 if (data
->hctx
->flags
& BLK_MQ_F_TAG_SHARED
) {
284 rq_flags
= RQF_MQ_INFLIGHT
;
285 atomic_inc(&data
->hctx
->nr_active
);
288 rq
->internal_tag
= -1;
289 data
->hctx
->tags
->rqs
[rq
->tag
] = rq
;
292 /* csd/requeue_work/fifo_time is initialized before use */
294 rq
->mq_ctx
= data
->ctx
;
295 rq
->mq_hctx
= data
->hctx
;
296 rq
->rq_flags
= rq_flags
;
298 if (data
->flags
& BLK_MQ_REQ_PREEMPT
)
299 rq
->rq_flags
|= RQF_PREEMPT
;
300 if (blk_queue_io_stat(data
->q
))
301 rq
->rq_flags
|= RQF_IO_STAT
;
302 INIT_LIST_HEAD(&rq
->queuelist
);
303 INIT_HLIST_NODE(&rq
->hash
);
304 RB_CLEAR_NODE(&rq
->rb_node
);
307 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
308 rq
->alloc_time_ns
= alloc_time_ns
;
310 if (blk_mq_need_time_stamp(rq
))
311 rq
->start_time_ns
= ktime_get_ns();
313 rq
->start_time_ns
= 0;
314 rq
->io_start_time_ns
= 0;
315 rq
->stats_sectors
= 0;
316 rq
->nr_phys_segments
= 0;
317 #if defined(CONFIG_BLK_DEV_INTEGRITY)
318 rq
->nr_integrity_segments
= 0;
320 /* tag was already set */
322 WRITE_ONCE(rq
->deadline
, 0);
327 rq
->end_io_data
= NULL
;
329 data
->ctx
->rq_dispatched
[op_is_sync(op
)]++;
330 refcount_set(&rq
->ref
, 1);
334 static struct request
*blk_mq_get_request(struct request_queue
*q
,
336 struct blk_mq_alloc_data
*data
)
338 struct elevator_queue
*e
= q
->elevator
;
341 bool clear_ctx_on_error
= false;
342 u64 alloc_time_ns
= 0;
344 blk_queue_enter_live(q
);
346 /* alloc_time includes depth and tag waits */
347 if (blk_queue_rq_alloc_time(q
))
348 alloc_time_ns
= ktime_get_ns();
351 if (likely(!data
->ctx
)) {
352 data
->ctx
= blk_mq_get_ctx(q
);
353 clear_ctx_on_error
= true;
355 if (likely(!data
->hctx
))
356 data
->hctx
= blk_mq_map_queue(q
, data
->cmd_flags
,
358 if (data
->cmd_flags
& REQ_NOWAIT
)
359 data
->flags
|= BLK_MQ_REQ_NOWAIT
;
362 data
->flags
|= BLK_MQ_REQ_INTERNAL
;
365 * Flush requests are special and go directly to the
366 * dispatch list. Don't include reserved tags in the
367 * limiting, as it isn't useful.
369 if (!op_is_flush(data
->cmd_flags
) &&
370 e
->type
->ops
.limit_depth
&&
371 !(data
->flags
& BLK_MQ_REQ_RESERVED
))
372 e
->type
->ops
.limit_depth(data
->cmd_flags
, data
);
374 blk_mq_tag_busy(data
->hctx
);
377 tag
= blk_mq_get_tag(data
);
378 if (tag
== BLK_MQ_TAG_FAIL
) {
379 if (clear_ctx_on_error
)
385 rq
= blk_mq_rq_ctx_init(data
, tag
, data
->cmd_flags
, alloc_time_ns
);
386 if (!op_is_flush(data
->cmd_flags
)) {
388 if (e
&& e
->type
->ops
.prepare_request
) {
389 if (e
->type
->icq_cache
)
390 blk_mq_sched_assign_ioc(rq
);
392 e
->type
->ops
.prepare_request(rq
, bio
);
393 rq
->rq_flags
|= RQF_ELVPRIV
;
396 data
->hctx
->queued
++;
400 struct request
*blk_mq_alloc_request(struct request_queue
*q
, unsigned int op
,
401 blk_mq_req_flags_t flags
)
403 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
, .cmd_flags
= op
};
407 ret
= blk_queue_enter(q
, flags
);
411 rq
= blk_mq_get_request(q
, NULL
, &alloc_data
);
415 return ERR_PTR(-EWOULDBLOCK
);
418 rq
->__sector
= (sector_t
) -1;
419 rq
->bio
= rq
->biotail
= NULL
;
422 EXPORT_SYMBOL(blk_mq_alloc_request
);
424 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
,
425 unsigned int op
, blk_mq_req_flags_t flags
, unsigned int hctx_idx
)
427 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
, .cmd_flags
= op
};
433 * If the tag allocator sleeps we could get an allocation for a
434 * different hardware context. No need to complicate the low level
435 * allocator for this for the rare use case of a command tied to
438 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
439 return ERR_PTR(-EINVAL
);
441 if (hctx_idx
>= q
->nr_hw_queues
)
442 return ERR_PTR(-EIO
);
444 ret
= blk_queue_enter(q
, flags
);
449 * Check if the hardware context is actually mapped to anything.
450 * If not tell the caller that it should skip this queue.
452 alloc_data
.hctx
= q
->queue_hw_ctx
[hctx_idx
];
453 if (!blk_mq_hw_queue_mapped(alloc_data
.hctx
)) {
455 return ERR_PTR(-EXDEV
);
457 cpu
= cpumask_first_and(alloc_data
.hctx
->cpumask
, cpu_online_mask
);
458 alloc_data
.ctx
= __blk_mq_get_ctx(q
, cpu
);
460 rq
= blk_mq_get_request(q
, NULL
, &alloc_data
);
464 return ERR_PTR(-EWOULDBLOCK
);
468 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
470 static void __blk_mq_free_request(struct request
*rq
)
472 struct request_queue
*q
= rq
->q
;
473 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
474 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
475 const int sched_tag
= rq
->internal_tag
;
477 blk_pm_mark_last_busy(rq
);
480 blk_mq_put_tag(hctx
->tags
, ctx
, rq
->tag
);
482 blk_mq_put_tag(hctx
->sched_tags
, ctx
, sched_tag
);
483 blk_mq_sched_restart(hctx
);
487 void blk_mq_free_request(struct request
*rq
)
489 struct request_queue
*q
= rq
->q
;
490 struct elevator_queue
*e
= q
->elevator
;
491 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
492 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
494 if (rq
->rq_flags
& RQF_ELVPRIV
) {
495 if (e
&& e
->type
->ops
.finish_request
)
496 e
->type
->ops
.finish_request(rq
);
498 put_io_context(rq
->elv
.icq
->ioc
);
503 ctx
->rq_completed
[rq_is_sync(rq
)]++;
504 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
505 atomic_dec(&hctx
->nr_active
);
507 if (unlikely(laptop_mode
&& !blk_rq_is_passthrough(rq
)))
508 laptop_io_completion(q
->backing_dev_info
);
512 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
513 if (refcount_dec_and_test(&rq
->ref
))
514 __blk_mq_free_request(rq
);
516 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
518 inline void __blk_mq_end_request(struct request
*rq
, blk_status_t error
)
522 if (blk_mq_need_time_stamp(rq
))
523 now
= ktime_get_ns();
525 if (rq
->rq_flags
& RQF_STATS
) {
526 blk_mq_poll_stats_start(rq
->q
);
527 blk_stat_add(rq
, now
);
530 if (rq
->internal_tag
!= -1)
531 blk_mq_sched_completed_request(rq
, now
);
533 blk_account_io_done(rq
, now
);
536 rq_qos_done(rq
->q
, rq
);
537 rq
->end_io(rq
, error
);
539 blk_mq_free_request(rq
);
542 EXPORT_SYMBOL(__blk_mq_end_request
);
544 void blk_mq_end_request(struct request
*rq
, blk_status_t error
)
546 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
548 __blk_mq_end_request(rq
, error
);
550 EXPORT_SYMBOL(blk_mq_end_request
);
552 static void __blk_mq_complete_request_remote(void *data
)
554 struct request
*rq
= data
;
555 struct request_queue
*q
= rq
->q
;
557 q
->mq_ops
->complete(rq
);
560 static void __blk_mq_complete_request(struct request
*rq
)
562 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
563 struct request_queue
*q
= rq
->q
;
567 WRITE_ONCE(rq
->state
, MQ_RQ_COMPLETE
);
569 * Most of single queue controllers, there is only one irq vector
570 * for handling IO completion, and the only irq's affinity is set
571 * as all possible CPUs. On most of ARCHs, this affinity means the
572 * irq is handled on one specific CPU.
574 * So complete IO reqeust in softirq context in case of single queue
575 * for not degrading IO performance by irqsoff latency.
577 if (q
->nr_hw_queues
== 1) {
578 __blk_complete_request(rq
);
583 * For a polled request, always complete locallly, it's pointless
584 * to redirect the completion.
586 if ((rq
->cmd_flags
& REQ_HIPRI
) ||
587 !test_bit(QUEUE_FLAG_SAME_COMP
, &q
->queue_flags
)) {
588 q
->mq_ops
->complete(rq
);
593 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &q
->queue_flags
))
594 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
596 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
597 rq
->csd
.func
= __blk_mq_complete_request_remote
;
600 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
602 q
->mq_ops
->complete(rq
);
607 static void hctx_unlock(struct blk_mq_hw_ctx
*hctx
, int srcu_idx
)
608 __releases(hctx
->srcu
)
610 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
))
613 srcu_read_unlock(hctx
->srcu
, srcu_idx
);
616 static void hctx_lock(struct blk_mq_hw_ctx
*hctx
, int *srcu_idx
)
617 __acquires(hctx
->srcu
)
619 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
620 /* shut up gcc false positive */
624 *srcu_idx
= srcu_read_lock(hctx
->srcu
);
628 * blk_mq_complete_request - end I/O on a request
629 * @rq: the request being processed
632 * Ends all I/O on a request. It does not handle partial completions.
633 * The actual completion happens out-of-order, through a IPI handler.
635 bool blk_mq_complete_request(struct request
*rq
)
637 if (unlikely(blk_should_fake_timeout(rq
->q
)))
639 __blk_mq_complete_request(rq
);
642 EXPORT_SYMBOL(blk_mq_complete_request
);
645 * blk_mq_start_request - Start processing a request
646 * @rq: Pointer to request to be started
648 * Function used by device drivers to notify the block layer that a request
649 * is going to be processed now, so blk layer can do proper initializations
650 * such as starting the timeout timer.
652 void blk_mq_start_request(struct request
*rq
)
654 struct request_queue
*q
= rq
->q
;
656 trace_block_rq_issue(q
, rq
);
658 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
659 rq
->io_start_time_ns
= ktime_get_ns();
660 rq
->stats_sectors
= blk_rq_sectors(rq
);
661 rq
->rq_flags
|= RQF_STATS
;
665 WARN_ON_ONCE(blk_mq_rq_state(rq
) != MQ_RQ_IDLE
);
668 WRITE_ONCE(rq
->state
, MQ_RQ_IN_FLIGHT
);
670 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
672 * Make sure space for the drain appears. We know we can do
673 * this because max_hw_segments has been adjusted to be one
674 * fewer than the device can handle.
676 rq
->nr_phys_segments
++;
679 #ifdef CONFIG_BLK_DEV_INTEGRITY
680 if (blk_integrity_rq(rq
) && req_op(rq
) == REQ_OP_WRITE
)
681 q
->integrity
.profile
->prepare_fn(rq
);
684 EXPORT_SYMBOL(blk_mq_start_request
);
686 static void __blk_mq_requeue_request(struct request
*rq
)
688 struct request_queue
*q
= rq
->q
;
690 blk_mq_put_driver_tag(rq
);
692 trace_block_rq_requeue(q
, rq
);
693 rq_qos_requeue(q
, rq
);
695 if (blk_mq_request_started(rq
)) {
696 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
697 rq
->rq_flags
&= ~RQF_TIMED_OUT
;
698 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
699 rq
->nr_phys_segments
--;
703 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
705 __blk_mq_requeue_request(rq
);
707 /* this request will be re-inserted to io scheduler queue */
708 blk_mq_sched_requeue_request(rq
);
710 BUG_ON(!list_empty(&rq
->queuelist
));
711 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
713 EXPORT_SYMBOL(blk_mq_requeue_request
);
715 static void blk_mq_requeue_work(struct work_struct
*work
)
717 struct request_queue
*q
=
718 container_of(work
, struct request_queue
, requeue_work
.work
);
720 struct request
*rq
, *next
;
722 spin_lock_irq(&q
->requeue_lock
);
723 list_splice_init(&q
->requeue_list
, &rq_list
);
724 spin_unlock_irq(&q
->requeue_lock
);
726 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
727 if (!(rq
->rq_flags
& (RQF_SOFTBARRIER
| RQF_DONTPREP
)))
730 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
731 list_del_init(&rq
->queuelist
);
733 * If RQF_DONTPREP, rq has contained some driver specific
734 * data, so insert it to hctx dispatch list to avoid any
737 if (rq
->rq_flags
& RQF_DONTPREP
)
738 blk_mq_request_bypass_insert(rq
, false, false);
740 blk_mq_sched_insert_request(rq
, true, false, false);
743 while (!list_empty(&rq_list
)) {
744 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
745 list_del_init(&rq
->queuelist
);
746 blk_mq_sched_insert_request(rq
, false, false, false);
749 blk_mq_run_hw_queues(q
, false);
752 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
753 bool kick_requeue_list
)
755 struct request_queue
*q
= rq
->q
;
759 * We abuse this flag that is otherwise used by the I/O scheduler to
760 * request head insertion from the workqueue.
762 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
764 spin_lock_irqsave(&q
->requeue_lock
, flags
);
766 rq
->rq_flags
|= RQF_SOFTBARRIER
;
767 list_add(&rq
->queuelist
, &q
->requeue_list
);
769 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
771 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
773 if (kick_requeue_list
)
774 blk_mq_kick_requeue_list(q
);
777 void blk_mq_kick_requeue_list(struct request_queue
*q
)
779 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
, 0);
781 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
783 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
786 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
,
787 msecs_to_jiffies(msecs
));
789 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
791 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
793 if (tag
< tags
->nr_tags
) {
794 prefetch(tags
->rqs
[tag
]);
795 return tags
->rqs
[tag
];
800 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
802 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
803 void *priv
, bool reserved
)
806 * If we find a request that is inflight and the queue matches,
807 * we know the queue is busy. Return false to stop the iteration.
809 if (rq
->state
== MQ_RQ_IN_FLIGHT
&& rq
->q
== hctx
->queue
) {
819 bool blk_mq_queue_inflight(struct request_queue
*q
)
823 blk_mq_queue_tag_busy_iter(q
, blk_mq_rq_inflight
, &busy
);
826 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight
);
828 static void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
830 req
->rq_flags
|= RQF_TIMED_OUT
;
831 if (req
->q
->mq_ops
->timeout
) {
832 enum blk_eh_timer_return ret
;
834 ret
= req
->q
->mq_ops
->timeout(req
, reserved
);
835 if (ret
== BLK_EH_DONE
)
837 WARN_ON_ONCE(ret
!= BLK_EH_RESET_TIMER
);
843 static bool blk_mq_req_expired(struct request
*rq
, unsigned long *next
)
845 unsigned long deadline
;
847 if (blk_mq_rq_state(rq
) != MQ_RQ_IN_FLIGHT
)
849 if (rq
->rq_flags
& RQF_TIMED_OUT
)
852 deadline
= READ_ONCE(rq
->deadline
);
853 if (time_after_eq(jiffies
, deadline
))
858 else if (time_after(*next
, deadline
))
863 static bool blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
864 struct request
*rq
, void *priv
, bool reserved
)
866 unsigned long *next
= priv
;
869 * Just do a quick check if it is expired before locking the request in
870 * so we're not unnecessarilly synchronizing across CPUs.
872 if (!blk_mq_req_expired(rq
, next
))
876 * We have reason to believe the request may be expired. Take a
877 * reference on the request to lock this request lifetime into its
878 * currently allocated context to prevent it from being reallocated in
879 * the event the completion by-passes this timeout handler.
881 * If the reference was already released, then the driver beat the
882 * timeout handler to posting a natural completion.
884 if (!refcount_inc_not_zero(&rq
->ref
))
888 * The request is now locked and cannot be reallocated underneath the
889 * timeout handler's processing. Re-verify this exact request is truly
890 * expired; if it is not expired, then the request was completed and
891 * reallocated as a new request.
893 if (blk_mq_req_expired(rq
, next
))
894 blk_mq_rq_timed_out(rq
, reserved
);
896 if (is_flush_rq(rq
, hctx
))
898 else if (refcount_dec_and_test(&rq
->ref
))
899 __blk_mq_free_request(rq
);
904 static void blk_mq_timeout_work(struct work_struct
*work
)
906 struct request_queue
*q
=
907 container_of(work
, struct request_queue
, timeout_work
);
908 unsigned long next
= 0;
909 struct blk_mq_hw_ctx
*hctx
;
912 /* A deadlock might occur if a request is stuck requiring a
913 * timeout at the same time a queue freeze is waiting
914 * completion, since the timeout code would not be able to
915 * acquire the queue reference here.
917 * That's why we don't use blk_queue_enter here; instead, we use
918 * percpu_ref_tryget directly, because we need to be able to
919 * obtain a reference even in the short window between the queue
920 * starting to freeze, by dropping the first reference in
921 * blk_freeze_queue_start, and the moment the last request is
922 * consumed, marked by the instant q_usage_counter reaches
925 if (!percpu_ref_tryget(&q
->q_usage_counter
))
928 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &next
);
931 mod_timer(&q
->timeout
, next
);
934 * Request timeouts are handled as a forward rolling timer. If
935 * we end up here it means that no requests are pending and
936 * also that no request has been pending for a while. Mark
939 queue_for_each_hw_ctx(q
, hctx
, i
) {
940 /* the hctx may be unmapped, so check it here */
941 if (blk_mq_hw_queue_mapped(hctx
))
942 blk_mq_tag_idle(hctx
);
948 struct flush_busy_ctx_data
{
949 struct blk_mq_hw_ctx
*hctx
;
950 struct list_head
*list
;
953 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
955 struct flush_busy_ctx_data
*flush_data
= data
;
956 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
957 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
958 enum hctx_type type
= hctx
->type
;
960 spin_lock(&ctx
->lock
);
961 list_splice_tail_init(&ctx
->rq_lists
[type
], flush_data
->list
);
962 sbitmap_clear_bit(sb
, bitnr
);
963 spin_unlock(&ctx
->lock
);
968 * Process software queues that have been marked busy, splicing them
969 * to the for-dispatch
971 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
973 struct flush_busy_ctx_data data
= {
978 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
980 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
982 struct dispatch_rq_data
{
983 struct blk_mq_hw_ctx
*hctx
;
987 static bool dispatch_rq_from_ctx(struct sbitmap
*sb
, unsigned int bitnr
,
990 struct dispatch_rq_data
*dispatch_data
= data
;
991 struct blk_mq_hw_ctx
*hctx
= dispatch_data
->hctx
;
992 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
993 enum hctx_type type
= hctx
->type
;
995 spin_lock(&ctx
->lock
);
996 if (!list_empty(&ctx
->rq_lists
[type
])) {
997 dispatch_data
->rq
= list_entry_rq(ctx
->rq_lists
[type
].next
);
998 list_del_init(&dispatch_data
->rq
->queuelist
);
999 if (list_empty(&ctx
->rq_lists
[type
]))
1000 sbitmap_clear_bit(sb
, bitnr
);
1002 spin_unlock(&ctx
->lock
);
1004 return !dispatch_data
->rq
;
1007 struct request
*blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx
*hctx
,
1008 struct blk_mq_ctx
*start
)
1010 unsigned off
= start
? start
->index_hw
[hctx
->type
] : 0;
1011 struct dispatch_rq_data data
= {
1016 __sbitmap_for_each_set(&hctx
->ctx_map
, off
,
1017 dispatch_rq_from_ctx
, &data
);
1022 static inline unsigned int queued_to_index(unsigned int queued
)
1027 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
1030 bool blk_mq_get_driver_tag(struct request
*rq
)
1032 struct blk_mq_alloc_data data
= {
1034 .hctx
= rq
->mq_hctx
,
1035 .flags
= BLK_MQ_REQ_NOWAIT
,
1036 .cmd_flags
= rq
->cmd_flags
,
1043 if (blk_mq_tag_is_reserved(data
.hctx
->sched_tags
, rq
->internal_tag
))
1044 data
.flags
|= BLK_MQ_REQ_RESERVED
;
1046 shared
= blk_mq_tag_busy(data
.hctx
);
1047 rq
->tag
= blk_mq_get_tag(&data
);
1050 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
1051 atomic_inc(&data
.hctx
->nr_active
);
1053 data
.hctx
->tags
->rqs
[rq
->tag
] = rq
;
1056 return rq
->tag
!= -1;
1059 static int blk_mq_dispatch_wake(wait_queue_entry_t
*wait
, unsigned mode
,
1060 int flags
, void *key
)
1062 struct blk_mq_hw_ctx
*hctx
;
1064 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
1066 spin_lock(&hctx
->dispatch_wait_lock
);
1067 if (!list_empty(&wait
->entry
)) {
1068 struct sbitmap_queue
*sbq
;
1070 list_del_init(&wait
->entry
);
1071 sbq
= &hctx
->tags
->bitmap_tags
;
1072 atomic_dec(&sbq
->ws_active
);
1074 spin_unlock(&hctx
->dispatch_wait_lock
);
1076 blk_mq_run_hw_queue(hctx
, true);
1081 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1082 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1083 * restart. For both cases, take care to check the condition again after
1084 * marking us as waiting.
1086 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx
*hctx
,
1089 struct sbitmap_queue
*sbq
= &hctx
->tags
->bitmap_tags
;
1090 struct wait_queue_head
*wq
;
1091 wait_queue_entry_t
*wait
;
1094 if (!(hctx
->flags
& BLK_MQ_F_TAG_SHARED
)) {
1095 blk_mq_sched_mark_restart_hctx(hctx
);
1098 * It's possible that a tag was freed in the window between the
1099 * allocation failure and adding the hardware queue to the wait
1102 * Don't clear RESTART here, someone else could have set it.
1103 * At most this will cost an extra queue run.
1105 return blk_mq_get_driver_tag(rq
);
1108 wait
= &hctx
->dispatch_wait
;
1109 if (!list_empty_careful(&wait
->entry
))
1112 wq
= &bt_wait_ptr(sbq
, hctx
)->wait
;
1114 spin_lock_irq(&wq
->lock
);
1115 spin_lock(&hctx
->dispatch_wait_lock
);
1116 if (!list_empty(&wait
->entry
)) {
1117 spin_unlock(&hctx
->dispatch_wait_lock
);
1118 spin_unlock_irq(&wq
->lock
);
1122 atomic_inc(&sbq
->ws_active
);
1123 wait
->flags
&= ~WQ_FLAG_EXCLUSIVE
;
1124 __add_wait_queue(wq
, wait
);
1127 * It's possible that a tag was freed in the window between the
1128 * allocation failure and adding the hardware queue to the wait
1131 ret
= blk_mq_get_driver_tag(rq
);
1133 spin_unlock(&hctx
->dispatch_wait_lock
);
1134 spin_unlock_irq(&wq
->lock
);
1139 * We got a tag, remove ourselves from the wait queue to ensure
1140 * someone else gets the wakeup.
1142 list_del_init(&wait
->entry
);
1143 atomic_dec(&sbq
->ws_active
);
1144 spin_unlock(&hctx
->dispatch_wait_lock
);
1145 spin_unlock_irq(&wq
->lock
);
1150 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1151 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1153 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1154 * - EWMA is one simple way to compute running average value
1155 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1156 * - take 4 as factor for avoiding to get too small(0) result, and this
1157 * factor doesn't matter because EWMA decreases exponentially
1159 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx
*hctx
, bool busy
)
1163 if (hctx
->queue
->elevator
)
1166 ewma
= hctx
->dispatch_busy
;
1171 ewma
*= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
- 1;
1173 ewma
+= 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR
;
1174 ewma
/= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
;
1176 hctx
->dispatch_busy
= ewma
;
1179 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1181 static void blk_mq_handle_dev_resource(struct request
*rq
,
1182 struct list_head
*list
)
1184 struct request
*next
=
1185 list_first_entry_or_null(list
, struct request
, queuelist
);
1188 * If an I/O scheduler has been configured and we got a driver tag for
1189 * the next request already, free it.
1192 blk_mq_put_driver_tag(next
);
1194 list_add(&rq
->queuelist
, list
);
1195 __blk_mq_requeue_request(rq
);
1199 * Returns true if we did some work AND can potentially do more.
1201 bool blk_mq_dispatch_rq_list(struct request_queue
*q
, struct list_head
*list
,
1204 struct blk_mq_hw_ctx
*hctx
;
1205 struct request
*rq
, *nxt
;
1206 bool no_tag
= false;
1208 blk_status_t ret
= BLK_STS_OK
;
1210 if (list_empty(list
))
1213 WARN_ON(!list_is_singular(list
) && got_budget
);
1216 * Now process all the entries, sending them to the driver.
1218 errors
= queued
= 0;
1220 struct blk_mq_queue_data bd
;
1222 rq
= list_first_entry(list
, struct request
, queuelist
);
1225 if (!got_budget
&& !blk_mq_get_dispatch_budget(hctx
))
1228 if (!blk_mq_get_driver_tag(rq
)) {
1230 * The initial allocation attempt failed, so we need to
1231 * rerun the hardware queue when a tag is freed. The
1232 * waitqueue takes care of that. If the queue is run
1233 * before we add this entry back on the dispatch list,
1234 * we'll re-run it below.
1236 if (!blk_mq_mark_tag_wait(hctx
, rq
)) {
1237 blk_mq_put_dispatch_budget(hctx
);
1239 * For non-shared tags, the RESTART check
1242 if (hctx
->flags
& BLK_MQ_F_TAG_SHARED
)
1248 list_del_init(&rq
->queuelist
);
1253 * Flag last if we have no more requests, or if we have more
1254 * but can't assign a driver tag to it.
1256 if (list_empty(list
))
1259 nxt
= list_first_entry(list
, struct request
, queuelist
);
1260 bd
.last
= !blk_mq_get_driver_tag(nxt
);
1263 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1264 if (ret
== BLK_STS_RESOURCE
|| ret
== BLK_STS_DEV_RESOURCE
) {
1265 blk_mq_handle_dev_resource(rq
, list
);
1269 if (unlikely(ret
!= BLK_STS_OK
)) {
1271 blk_mq_end_request(rq
, BLK_STS_IOERR
);
1276 } while (!list_empty(list
));
1278 hctx
->dispatched
[queued_to_index(queued
)]++;
1281 * Any items that need requeuing? Stuff them into hctx->dispatch,
1282 * that is where we will continue on next queue run.
1284 if (!list_empty(list
)) {
1288 * If we didn't flush the entire list, we could have told
1289 * the driver there was more coming, but that turned out to
1292 if (q
->mq_ops
->commit_rqs
)
1293 q
->mq_ops
->commit_rqs(hctx
);
1295 spin_lock(&hctx
->lock
);
1296 list_splice_tail_init(list
, &hctx
->dispatch
);
1297 spin_unlock(&hctx
->lock
);
1300 * If SCHED_RESTART was set by the caller of this function and
1301 * it is no longer set that means that it was cleared by another
1302 * thread and hence that a queue rerun is needed.
1304 * If 'no_tag' is set, that means that we failed getting
1305 * a driver tag with an I/O scheduler attached. If our dispatch
1306 * waitqueue is no longer active, ensure that we run the queue
1307 * AFTER adding our entries back to the list.
1309 * If no I/O scheduler has been configured it is possible that
1310 * the hardware queue got stopped and restarted before requests
1311 * were pushed back onto the dispatch list. Rerun the queue to
1312 * avoid starvation. Notes:
1313 * - blk_mq_run_hw_queue() checks whether or not a queue has
1314 * been stopped before rerunning a queue.
1315 * - Some but not all block drivers stop a queue before
1316 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1319 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1320 * bit is set, run queue after a delay to avoid IO stalls
1321 * that could otherwise occur if the queue is idle.
1323 needs_restart
= blk_mq_sched_needs_restart(hctx
);
1324 if (!needs_restart
||
1325 (no_tag
&& list_empty_careful(&hctx
->dispatch_wait
.entry
)))
1326 blk_mq_run_hw_queue(hctx
, true);
1327 else if (needs_restart
&& (ret
== BLK_STS_RESOURCE
))
1328 blk_mq_delay_run_hw_queue(hctx
, BLK_MQ_RESOURCE_DELAY
);
1330 blk_mq_update_dispatch_busy(hctx
, true);
1333 blk_mq_update_dispatch_busy(hctx
, false);
1336 * If the host/device is unable to accept more work, inform the
1339 if (ret
== BLK_STS_RESOURCE
|| ret
== BLK_STS_DEV_RESOURCE
)
1342 return (queued
+ errors
) != 0;
1346 * __blk_mq_run_hw_queue - Run a hardware queue.
1347 * @hctx: Pointer to the hardware queue to run.
1349 * Send pending requests to the hardware.
1351 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1356 * We should be running this queue from one of the CPUs that
1359 * There are at least two related races now between setting
1360 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1361 * __blk_mq_run_hw_queue():
1363 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1364 * but later it becomes online, then this warning is harmless
1367 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1368 * but later it becomes offline, then the warning can't be
1369 * triggered, and we depend on blk-mq timeout handler to
1370 * handle dispatched requests to this hctx
1372 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
1373 cpu_online(hctx
->next_cpu
)) {
1374 printk(KERN_WARNING
"run queue from wrong CPU %d, hctx %s\n",
1375 raw_smp_processor_id(),
1376 cpumask_empty(hctx
->cpumask
) ? "inactive": "active");
1381 * We can't run the queue inline with ints disabled. Ensure that
1382 * we catch bad users of this early.
1384 WARN_ON_ONCE(in_interrupt());
1386 might_sleep_if(hctx
->flags
& BLK_MQ_F_BLOCKING
);
1388 hctx_lock(hctx
, &srcu_idx
);
1389 blk_mq_sched_dispatch_requests(hctx
);
1390 hctx_unlock(hctx
, srcu_idx
);
1393 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx
*hctx
)
1395 int cpu
= cpumask_first_and(hctx
->cpumask
, cpu_online_mask
);
1397 if (cpu
>= nr_cpu_ids
)
1398 cpu
= cpumask_first(hctx
->cpumask
);
1403 * It'd be great if the workqueue API had a way to pass
1404 * in a mask and had some smarts for more clever placement.
1405 * For now we just round-robin here, switching for every
1406 * BLK_MQ_CPU_WORK_BATCH queued items.
1408 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1411 int next_cpu
= hctx
->next_cpu
;
1413 if (hctx
->queue
->nr_hw_queues
== 1)
1414 return WORK_CPU_UNBOUND
;
1416 if (--hctx
->next_cpu_batch
<= 0) {
1418 next_cpu
= cpumask_next_and(next_cpu
, hctx
->cpumask
,
1420 if (next_cpu
>= nr_cpu_ids
)
1421 next_cpu
= blk_mq_first_mapped_cpu(hctx
);
1422 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1426 * Do unbound schedule if we can't find a online CPU for this hctx,
1427 * and it should only happen in the path of handling CPU DEAD.
1429 if (!cpu_online(next_cpu
)) {
1436 * Make sure to re-select CPU next time once after CPUs
1437 * in hctx->cpumask become online again.
1439 hctx
->next_cpu
= next_cpu
;
1440 hctx
->next_cpu_batch
= 1;
1441 return WORK_CPU_UNBOUND
;
1444 hctx
->next_cpu
= next_cpu
;
1449 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
1450 * @hctx: Pointer to the hardware queue to run.
1451 * @async: If we want to run the queue asynchronously.
1452 * @msecs: Microseconds of delay to wait before running the queue.
1454 * If !@async, try to run the queue now. Else, run the queue asynchronously and
1455 * with a delay of @msecs.
1457 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
,
1458 unsigned long msecs
)
1460 if (unlikely(blk_mq_hctx_stopped(hctx
)))
1463 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1464 int cpu
= get_cpu();
1465 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1466 __blk_mq_run_hw_queue(hctx
);
1474 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
,
1475 msecs_to_jiffies(msecs
));
1479 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
1480 * @hctx: Pointer to the hardware queue to run.
1481 * @msecs: Microseconds of delay to wait before running the queue.
1483 * Run a hardware queue asynchronously with a delay of @msecs.
1485 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1487 __blk_mq_delay_run_hw_queue(hctx
, true, msecs
);
1489 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue
);
1492 * blk_mq_run_hw_queue - Start to run a hardware queue.
1493 * @hctx: Pointer to the hardware queue to run.
1494 * @async: If we want to run the queue asynchronously.
1496 * Check if the request queue is not in a quiesced state and if there are
1497 * pending requests to be sent. If this is true, run the queue to send requests
1500 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1506 * When queue is quiesced, we may be switching io scheduler, or
1507 * updating nr_hw_queues, or other things, and we can't run queue
1508 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1510 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1513 hctx_lock(hctx
, &srcu_idx
);
1514 need_run
= !blk_queue_quiesced(hctx
->queue
) &&
1515 blk_mq_hctx_has_pending(hctx
);
1516 hctx_unlock(hctx
, srcu_idx
);
1519 __blk_mq_delay_run_hw_queue(hctx
, async
, 0);
1521 EXPORT_SYMBOL(blk_mq_run_hw_queue
);
1524 * blk_mq_run_hw_queue - Run all hardware queues in a request queue.
1525 * @q: Pointer to the request queue to run.
1526 * @async: If we want to run the queue asynchronously.
1528 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1530 struct blk_mq_hw_ctx
*hctx
;
1533 queue_for_each_hw_ctx(q
, hctx
, i
) {
1534 if (blk_mq_hctx_stopped(hctx
))
1537 blk_mq_run_hw_queue(hctx
, async
);
1540 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1543 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1544 * @q: request queue.
1546 * The caller is responsible for serializing this function against
1547 * blk_mq_{start,stop}_hw_queue().
1549 bool blk_mq_queue_stopped(struct request_queue
*q
)
1551 struct blk_mq_hw_ctx
*hctx
;
1554 queue_for_each_hw_ctx(q
, hctx
, i
)
1555 if (blk_mq_hctx_stopped(hctx
))
1560 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1563 * This function is often used for pausing .queue_rq() by driver when
1564 * there isn't enough resource or some conditions aren't satisfied, and
1565 * BLK_STS_RESOURCE is usually returned.
1567 * We do not guarantee that dispatch can be drained or blocked
1568 * after blk_mq_stop_hw_queue() returns. Please use
1569 * blk_mq_quiesce_queue() for that requirement.
1571 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1573 cancel_delayed_work(&hctx
->run_work
);
1575 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1577 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1580 * This function is often used for pausing .queue_rq() by driver when
1581 * there isn't enough resource or some conditions aren't satisfied, and
1582 * BLK_STS_RESOURCE is usually returned.
1584 * We do not guarantee that dispatch can be drained or blocked
1585 * after blk_mq_stop_hw_queues() returns. Please use
1586 * blk_mq_quiesce_queue() for that requirement.
1588 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1590 struct blk_mq_hw_ctx
*hctx
;
1593 queue_for_each_hw_ctx(q
, hctx
, i
)
1594 blk_mq_stop_hw_queue(hctx
);
1596 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1598 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1600 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1602 blk_mq_run_hw_queue(hctx
, false);
1604 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1606 void blk_mq_start_hw_queues(struct request_queue
*q
)
1608 struct blk_mq_hw_ctx
*hctx
;
1611 queue_for_each_hw_ctx(q
, hctx
, i
)
1612 blk_mq_start_hw_queue(hctx
);
1614 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1616 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1618 if (!blk_mq_hctx_stopped(hctx
))
1621 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1622 blk_mq_run_hw_queue(hctx
, async
);
1624 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1626 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1628 struct blk_mq_hw_ctx
*hctx
;
1631 queue_for_each_hw_ctx(q
, hctx
, i
)
1632 blk_mq_start_stopped_hw_queue(hctx
, async
);
1634 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1636 static void blk_mq_run_work_fn(struct work_struct
*work
)
1638 struct blk_mq_hw_ctx
*hctx
;
1640 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
1643 * If we are stopped, don't run the queue.
1645 if (test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
1648 __blk_mq_run_hw_queue(hctx
);
1651 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1655 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1656 enum hctx_type type
= hctx
->type
;
1658 lockdep_assert_held(&ctx
->lock
);
1660 trace_block_rq_insert(hctx
->queue
, rq
);
1663 list_add(&rq
->queuelist
, &ctx
->rq_lists
[type
]);
1665 list_add_tail(&rq
->queuelist
, &ctx
->rq_lists
[type
]);
1668 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1671 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1673 lockdep_assert_held(&ctx
->lock
);
1675 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1676 blk_mq_hctx_mark_pending(hctx
, ctx
);
1680 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
1681 * @rq: Pointer to request to be inserted.
1682 * @run_queue: If we should run the hardware queue after inserting the request.
1684 * Should only be used carefully, when the caller knows we want to
1685 * bypass a potential IO scheduler on the target device.
1687 void blk_mq_request_bypass_insert(struct request
*rq
, bool at_head
,
1690 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1692 spin_lock(&hctx
->lock
);
1694 list_add(&rq
->queuelist
, &hctx
->dispatch
);
1696 list_add_tail(&rq
->queuelist
, &hctx
->dispatch
);
1697 spin_unlock(&hctx
->lock
);
1700 blk_mq_run_hw_queue(hctx
, false);
1703 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1704 struct list_head
*list
)
1708 enum hctx_type type
= hctx
->type
;
1711 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1714 list_for_each_entry(rq
, list
, queuelist
) {
1715 BUG_ON(rq
->mq_ctx
!= ctx
);
1716 trace_block_rq_insert(hctx
->queue
, rq
);
1719 spin_lock(&ctx
->lock
);
1720 list_splice_tail_init(list
, &ctx
->rq_lists
[type
]);
1721 blk_mq_hctx_mark_pending(hctx
, ctx
);
1722 spin_unlock(&ctx
->lock
);
1725 static int plug_rq_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1727 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1728 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1730 if (rqa
->mq_ctx
!= rqb
->mq_ctx
)
1731 return rqa
->mq_ctx
> rqb
->mq_ctx
;
1732 if (rqa
->mq_hctx
!= rqb
->mq_hctx
)
1733 return rqa
->mq_hctx
> rqb
->mq_hctx
;
1735 return blk_rq_pos(rqa
) > blk_rq_pos(rqb
);
1738 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1742 if (list_empty(&plug
->mq_list
))
1744 list_splice_init(&plug
->mq_list
, &list
);
1746 if (plug
->rq_count
> 2 && plug
->multiple_queues
)
1747 list_sort(NULL
, &list
, plug_rq_cmp
);
1752 struct list_head rq_list
;
1753 struct request
*rq
, *head_rq
= list_entry_rq(list
.next
);
1754 struct list_head
*pos
= &head_rq
->queuelist
; /* skip first */
1755 struct blk_mq_hw_ctx
*this_hctx
= head_rq
->mq_hctx
;
1756 struct blk_mq_ctx
*this_ctx
= head_rq
->mq_ctx
;
1757 unsigned int depth
= 1;
1759 list_for_each_continue(pos
, &list
) {
1760 rq
= list_entry_rq(pos
);
1762 if (rq
->mq_hctx
!= this_hctx
|| rq
->mq_ctx
!= this_ctx
)
1767 list_cut_before(&rq_list
, &list
, pos
);
1768 trace_block_unplug(head_rq
->q
, depth
, !from_schedule
);
1769 blk_mq_sched_insert_requests(this_hctx
, this_ctx
, &rq_list
,
1771 } while(!list_empty(&list
));
1774 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
,
1775 unsigned int nr_segs
)
1777 if (bio
->bi_opf
& REQ_RAHEAD
)
1778 rq
->cmd_flags
|= REQ_FAILFAST_MASK
;
1780 rq
->__sector
= bio
->bi_iter
.bi_sector
;
1781 rq
->write_hint
= bio
->bi_write_hint
;
1782 blk_rq_bio_prep(rq
, bio
, nr_segs
);
1784 blk_account_io_start(rq
, true);
1787 static blk_status_t
__blk_mq_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1789 blk_qc_t
*cookie
, bool last
)
1791 struct request_queue
*q
= rq
->q
;
1792 struct blk_mq_queue_data bd
= {
1796 blk_qc_t new_cookie
;
1799 new_cookie
= request_to_qc_t(hctx
, rq
);
1802 * For OK queue, we are done. For error, caller may kill it.
1803 * Any other error (busy), just add it to our list as we
1804 * previously would have done.
1806 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1809 blk_mq_update_dispatch_busy(hctx
, false);
1810 *cookie
= new_cookie
;
1812 case BLK_STS_RESOURCE
:
1813 case BLK_STS_DEV_RESOURCE
:
1814 blk_mq_update_dispatch_busy(hctx
, true);
1815 __blk_mq_requeue_request(rq
);
1818 blk_mq_update_dispatch_busy(hctx
, false);
1819 *cookie
= BLK_QC_T_NONE
;
1826 static blk_status_t
__blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1829 bool bypass_insert
, bool last
)
1831 struct request_queue
*q
= rq
->q
;
1832 bool run_queue
= true;
1835 * RCU or SRCU read lock is needed before checking quiesced flag.
1837 * When queue is stopped or quiesced, ignore 'bypass_insert' from
1838 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1839 * and avoid driver to try to dispatch again.
1841 if (blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(q
)) {
1843 bypass_insert
= false;
1847 if (q
->elevator
&& !bypass_insert
)
1850 if (!blk_mq_get_dispatch_budget(hctx
))
1853 if (!blk_mq_get_driver_tag(rq
)) {
1854 blk_mq_put_dispatch_budget(hctx
);
1858 return __blk_mq_issue_directly(hctx
, rq
, cookie
, last
);
1861 return BLK_STS_RESOURCE
;
1863 blk_mq_request_bypass_insert(rq
, false, run_queue
);
1868 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
1869 * @hctx: Pointer of the associated hardware queue.
1870 * @rq: Pointer to request to be sent.
1871 * @cookie: Request queue cookie.
1873 * If the device has enough resources to accept a new request now, send the
1874 * request directly to device driver. Else, insert at hctx->dispatch queue, so
1875 * we can try send it another time in the future. Requests inserted at this
1876 * queue have higher priority.
1878 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1879 struct request
*rq
, blk_qc_t
*cookie
)
1884 might_sleep_if(hctx
->flags
& BLK_MQ_F_BLOCKING
);
1886 hctx_lock(hctx
, &srcu_idx
);
1888 ret
= __blk_mq_try_issue_directly(hctx
, rq
, cookie
, false, true);
1889 if (ret
== BLK_STS_RESOURCE
|| ret
== BLK_STS_DEV_RESOURCE
)
1890 blk_mq_request_bypass_insert(rq
, false, true);
1891 else if (ret
!= BLK_STS_OK
)
1892 blk_mq_end_request(rq
, ret
);
1894 hctx_unlock(hctx
, srcu_idx
);
1897 blk_status_t
blk_mq_request_issue_directly(struct request
*rq
, bool last
)
1901 blk_qc_t unused_cookie
;
1902 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1904 hctx_lock(hctx
, &srcu_idx
);
1905 ret
= __blk_mq_try_issue_directly(hctx
, rq
, &unused_cookie
, true, last
);
1906 hctx_unlock(hctx
, srcu_idx
);
1911 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx
*hctx
,
1912 struct list_head
*list
)
1914 while (!list_empty(list
)) {
1916 struct request
*rq
= list_first_entry(list
, struct request
,
1919 list_del_init(&rq
->queuelist
);
1920 ret
= blk_mq_request_issue_directly(rq
, list_empty(list
));
1921 if (ret
!= BLK_STS_OK
) {
1922 if (ret
== BLK_STS_RESOURCE
||
1923 ret
== BLK_STS_DEV_RESOURCE
) {
1924 blk_mq_request_bypass_insert(rq
, false,
1928 blk_mq_end_request(rq
, ret
);
1933 * If we didn't flush the entire list, we could have told
1934 * the driver there was more coming, but that turned out to
1937 if (!list_empty(list
) && hctx
->queue
->mq_ops
->commit_rqs
)
1938 hctx
->queue
->mq_ops
->commit_rqs(hctx
);
1941 static void blk_add_rq_to_plug(struct blk_plug
*plug
, struct request
*rq
)
1943 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1945 if (!plug
->multiple_queues
&& !list_is_singular(&plug
->mq_list
)) {
1946 struct request
*tmp
;
1948 tmp
= list_first_entry(&plug
->mq_list
, struct request
,
1950 if (tmp
->q
!= rq
->q
)
1951 plug
->multiple_queues
= true;
1956 * blk_mq_make_request - Create and send a request to block device.
1957 * @q: Request queue pointer.
1958 * @bio: Bio pointer.
1960 * Builds up a request structure from @q and @bio and send to the device. The
1961 * request may not be queued directly to hardware if:
1962 * * This request can be merged with another one
1963 * * We want to place request at plug queue for possible future merging
1964 * * There is an IO scheduler active at this queue
1966 * It will not queue the request if there is an error with the bio, or at the
1969 * Returns: Request queue cookie.
1971 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1973 const int is_sync
= op_is_sync(bio
->bi_opf
);
1974 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1975 struct blk_mq_alloc_data data
= { .flags
= 0};
1977 struct blk_plug
*plug
;
1978 struct request
*same_queue_rq
= NULL
;
1979 unsigned int nr_segs
;
1982 blk_queue_bounce(q
, &bio
);
1983 __blk_queue_split(q
, &bio
, &nr_segs
);
1985 if (!bio_integrity_prep(bio
))
1986 return BLK_QC_T_NONE
;
1988 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1989 blk_attempt_plug_merge(q
, bio
, nr_segs
, &same_queue_rq
))
1990 return BLK_QC_T_NONE
;
1992 if (blk_mq_sched_bio_merge(q
, bio
, nr_segs
))
1993 return BLK_QC_T_NONE
;
1995 rq_qos_throttle(q
, bio
);
1997 data
.cmd_flags
= bio
->bi_opf
;
1998 rq
= blk_mq_get_request(q
, bio
, &data
);
1999 if (unlikely(!rq
)) {
2000 rq_qos_cleanup(q
, bio
);
2001 if (bio
->bi_opf
& REQ_NOWAIT
)
2002 bio_wouldblock_error(bio
);
2003 return BLK_QC_T_NONE
;
2006 trace_block_getrq(q
, bio
, bio
->bi_opf
);
2008 rq_qos_track(q
, rq
, bio
);
2010 cookie
= request_to_qc_t(data
.hctx
, rq
);
2012 blk_mq_bio_to_request(rq
, bio
, nr_segs
);
2014 plug
= blk_mq_plug(q
, bio
);
2015 if (unlikely(is_flush_fua
)) {
2016 /* Bypass scheduler for flush requests */
2017 blk_insert_flush(rq
);
2018 blk_mq_run_hw_queue(data
.hctx
, true);
2019 } else if (plug
&& (q
->nr_hw_queues
== 1 || q
->mq_ops
->commit_rqs
||
2020 !blk_queue_nonrot(q
))) {
2022 * Use plugging if we have a ->commit_rqs() hook as well, as
2023 * we know the driver uses bd->last in a smart fashion.
2025 * Use normal plugging if this disk is slow HDD, as sequential
2026 * IO may benefit a lot from plug merging.
2028 unsigned int request_count
= plug
->rq_count
;
2029 struct request
*last
= NULL
;
2032 trace_block_plug(q
);
2034 last
= list_entry_rq(plug
->mq_list
.prev
);
2036 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
2037 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
2038 blk_flush_plug_list(plug
, false);
2039 trace_block_plug(q
);
2042 blk_add_rq_to_plug(plug
, rq
);
2043 } else if (q
->elevator
) {
2044 /* Insert the request at the IO scheduler queue */
2045 blk_mq_sched_insert_request(rq
, false, true, true);
2046 } else if (plug
&& !blk_queue_nomerges(q
)) {
2048 * We do limited plugging. If the bio can be merged, do that.
2049 * Otherwise the existing request in the plug list will be
2050 * issued. So the plug list will have one request at most
2051 * The plug list might get flushed before this. If that happens,
2052 * the plug list is empty, and same_queue_rq is invalid.
2054 if (list_empty(&plug
->mq_list
))
2055 same_queue_rq
= NULL
;
2056 if (same_queue_rq
) {
2057 list_del_init(&same_queue_rq
->queuelist
);
2060 blk_add_rq_to_plug(plug
, rq
);
2061 trace_block_plug(q
);
2063 if (same_queue_rq
) {
2064 data
.hctx
= same_queue_rq
->mq_hctx
;
2065 trace_block_unplug(q
, 1, true);
2066 blk_mq_try_issue_directly(data
.hctx
, same_queue_rq
,
2069 } else if ((q
->nr_hw_queues
> 1 && is_sync
) ||
2070 !data
.hctx
->dispatch_busy
) {
2072 * There is no scheduler and we can try to send directly
2075 blk_mq_try_issue_directly(data
.hctx
, rq
, &cookie
);
2078 blk_mq_sched_insert_request(rq
, false, true, true);
2084 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
2085 unsigned int hctx_idx
)
2089 if (tags
->rqs
&& set
->ops
->exit_request
) {
2092 for (i
= 0; i
< tags
->nr_tags
; i
++) {
2093 struct request
*rq
= tags
->static_rqs
[i
];
2097 set
->ops
->exit_request(set
, rq
, hctx_idx
);
2098 tags
->static_rqs
[i
] = NULL
;
2102 while (!list_empty(&tags
->page_list
)) {
2103 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
2104 list_del_init(&page
->lru
);
2106 * Remove kmemleak object previously allocated in
2107 * blk_mq_alloc_rqs().
2109 kmemleak_free(page_address(page
));
2110 __free_pages(page
, page
->private);
2114 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
2118 kfree(tags
->static_rqs
);
2119 tags
->static_rqs
= NULL
;
2121 blk_mq_free_tags(tags
);
2124 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
2125 unsigned int hctx_idx
,
2126 unsigned int nr_tags
,
2127 unsigned int reserved_tags
)
2129 struct blk_mq_tags
*tags
;
2132 node
= blk_mq_hw_queue_to_node(&set
->map
[HCTX_TYPE_DEFAULT
], hctx_idx
);
2133 if (node
== NUMA_NO_NODE
)
2134 node
= set
->numa_node
;
2136 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
,
2137 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
2141 tags
->rqs
= kcalloc_node(nr_tags
, sizeof(struct request
*),
2142 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
2145 blk_mq_free_tags(tags
);
2149 tags
->static_rqs
= kcalloc_node(nr_tags
, sizeof(struct request
*),
2150 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
2152 if (!tags
->static_rqs
) {
2154 blk_mq_free_tags(tags
);
2161 static size_t order_to_size(unsigned int order
)
2163 return (size_t)PAGE_SIZE
<< order
;
2166 static int blk_mq_init_request(struct blk_mq_tag_set
*set
, struct request
*rq
,
2167 unsigned int hctx_idx
, int node
)
2171 if (set
->ops
->init_request
) {
2172 ret
= set
->ops
->init_request(set
, rq
, hctx_idx
, node
);
2177 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
2181 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
2182 unsigned int hctx_idx
, unsigned int depth
)
2184 unsigned int i
, j
, entries_per_page
, max_order
= 4;
2185 size_t rq_size
, left
;
2188 node
= blk_mq_hw_queue_to_node(&set
->map
[HCTX_TYPE_DEFAULT
], hctx_idx
);
2189 if (node
== NUMA_NO_NODE
)
2190 node
= set
->numa_node
;
2192 INIT_LIST_HEAD(&tags
->page_list
);
2195 * rq_size is the size of the request plus driver payload, rounded
2196 * to the cacheline size
2198 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
2200 left
= rq_size
* depth
;
2202 for (i
= 0; i
< depth
; ) {
2203 int this_order
= max_order
;
2208 while (this_order
&& left
< order_to_size(this_order
- 1))
2212 page
= alloc_pages_node(node
,
2213 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
2219 if (order_to_size(this_order
) < rq_size
)
2226 page
->private = this_order
;
2227 list_add_tail(&page
->lru
, &tags
->page_list
);
2229 p
= page_address(page
);
2231 * Allow kmemleak to scan these pages as they contain pointers
2232 * to additional allocations like via ops->init_request().
2234 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
2235 entries_per_page
= order_to_size(this_order
) / rq_size
;
2236 to_do
= min(entries_per_page
, depth
- i
);
2237 left
-= to_do
* rq_size
;
2238 for (j
= 0; j
< to_do
; j
++) {
2239 struct request
*rq
= p
;
2241 tags
->static_rqs
[i
] = rq
;
2242 if (blk_mq_init_request(set
, rq
, hctx_idx
, node
)) {
2243 tags
->static_rqs
[i
] = NULL
;
2254 blk_mq_free_rqs(set
, tags
, hctx_idx
);
2259 * 'cpu' is going away. splice any existing rq_list entries from this
2260 * software queue to the hw queue dispatch list, and ensure that it
2263 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
2265 struct blk_mq_hw_ctx
*hctx
;
2266 struct blk_mq_ctx
*ctx
;
2268 enum hctx_type type
;
2270 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
2271 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
2274 spin_lock(&ctx
->lock
);
2275 if (!list_empty(&ctx
->rq_lists
[type
])) {
2276 list_splice_init(&ctx
->rq_lists
[type
], &tmp
);
2277 blk_mq_hctx_clear_pending(hctx
, ctx
);
2279 spin_unlock(&ctx
->lock
);
2281 if (list_empty(&tmp
))
2284 spin_lock(&hctx
->lock
);
2285 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
2286 spin_unlock(&hctx
->lock
);
2288 blk_mq_run_hw_queue(hctx
, true);
2292 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
2294 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
2298 /* hctx->ctxs will be freed in queue's release handler */
2299 static void blk_mq_exit_hctx(struct request_queue
*q
,
2300 struct blk_mq_tag_set
*set
,
2301 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
2303 if (blk_mq_hw_queue_mapped(hctx
))
2304 blk_mq_tag_idle(hctx
);
2306 if (set
->ops
->exit_request
)
2307 set
->ops
->exit_request(set
, hctx
->fq
->flush_rq
, hctx_idx
);
2309 if (set
->ops
->exit_hctx
)
2310 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2312 blk_mq_remove_cpuhp(hctx
);
2314 spin_lock(&q
->unused_hctx_lock
);
2315 list_add(&hctx
->hctx_list
, &q
->unused_hctx_list
);
2316 spin_unlock(&q
->unused_hctx_lock
);
2319 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
2320 struct blk_mq_tag_set
*set
, int nr_queue
)
2322 struct blk_mq_hw_ctx
*hctx
;
2325 queue_for_each_hw_ctx(q
, hctx
, i
) {
2328 blk_mq_debugfs_unregister_hctx(hctx
);
2329 blk_mq_exit_hctx(q
, set
, hctx
, i
);
2333 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set
*tag_set
)
2335 int hw_ctx_size
= sizeof(struct blk_mq_hw_ctx
);
2337 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx
, srcu
),
2338 __alignof__(struct blk_mq_hw_ctx
)) !=
2339 sizeof(struct blk_mq_hw_ctx
));
2341 if (tag_set
->flags
& BLK_MQ_F_BLOCKING
)
2342 hw_ctx_size
+= sizeof(struct srcu_struct
);
2347 static int blk_mq_init_hctx(struct request_queue
*q
,
2348 struct blk_mq_tag_set
*set
,
2349 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
2351 hctx
->queue_num
= hctx_idx
;
2353 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
2355 hctx
->tags
= set
->tags
[hctx_idx
];
2357 if (set
->ops
->init_hctx
&&
2358 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
2359 goto unregister_cpu_notifier
;
2361 if (blk_mq_init_request(set
, hctx
->fq
->flush_rq
, hctx_idx
,
2367 if (set
->ops
->exit_hctx
)
2368 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2369 unregister_cpu_notifier
:
2370 blk_mq_remove_cpuhp(hctx
);
2374 static struct blk_mq_hw_ctx
*
2375 blk_mq_alloc_hctx(struct request_queue
*q
, struct blk_mq_tag_set
*set
,
2378 struct blk_mq_hw_ctx
*hctx
;
2379 gfp_t gfp
= GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
;
2381 hctx
= kzalloc_node(blk_mq_hw_ctx_size(set
), gfp
, node
);
2383 goto fail_alloc_hctx
;
2385 if (!zalloc_cpumask_var_node(&hctx
->cpumask
, gfp
, node
))
2388 atomic_set(&hctx
->nr_active
, 0);
2389 if (node
== NUMA_NO_NODE
)
2390 node
= set
->numa_node
;
2391 hctx
->numa_node
= node
;
2393 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
2394 spin_lock_init(&hctx
->lock
);
2395 INIT_LIST_HEAD(&hctx
->dispatch
);
2397 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
2399 INIT_LIST_HEAD(&hctx
->hctx_list
);
2402 * Allocate space for all possible cpus to avoid allocation at
2405 hctx
->ctxs
= kmalloc_array_node(nr_cpu_ids
, sizeof(void *),
2410 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8),
2415 spin_lock_init(&hctx
->dispatch_wait_lock
);
2416 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
2417 INIT_LIST_HEAD(&hctx
->dispatch_wait
.entry
);
2419 hctx
->fq
= blk_alloc_flush_queue(hctx
->numa_node
, set
->cmd_size
, gfp
);
2423 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
2424 init_srcu_struct(hctx
->srcu
);
2425 blk_mq_hctx_kobj_init(hctx
);
2430 sbitmap_free(&hctx
->ctx_map
);
2434 free_cpumask_var(hctx
->cpumask
);
2441 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
2442 unsigned int nr_hw_queues
)
2444 struct blk_mq_tag_set
*set
= q
->tag_set
;
2447 for_each_possible_cpu(i
) {
2448 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2449 struct blk_mq_hw_ctx
*hctx
;
2453 spin_lock_init(&__ctx
->lock
);
2454 for (k
= HCTX_TYPE_DEFAULT
; k
< HCTX_MAX_TYPES
; k
++)
2455 INIT_LIST_HEAD(&__ctx
->rq_lists
[k
]);
2460 * Set local node, IFF we have more than one hw queue. If
2461 * not, we remain on the home node of the device
2463 for (j
= 0; j
< set
->nr_maps
; j
++) {
2464 hctx
= blk_mq_map_queue_type(q
, j
, i
);
2465 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
2466 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
2471 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
, int hctx_idx
)
2475 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
2476 set
->queue_depth
, set
->reserved_tags
);
2477 if (!set
->tags
[hctx_idx
])
2480 ret
= blk_mq_alloc_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
,
2485 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2486 set
->tags
[hctx_idx
] = NULL
;
2490 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set
*set
,
2491 unsigned int hctx_idx
)
2493 if (set
->tags
&& set
->tags
[hctx_idx
]) {
2494 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
2495 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2496 set
->tags
[hctx_idx
] = NULL
;
2500 static void blk_mq_map_swqueue(struct request_queue
*q
)
2502 unsigned int i
, j
, hctx_idx
;
2503 struct blk_mq_hw_ctx
*hctx
;
2504 struct blk_mq_ctx
*ctx
;
2505 struct blk_mq_tag_set
*set
= q
->tag_set
;
2507 queue_for_each_hw_ctx(q
, hctx
, i
) {
2508 cpumask_clear(hctx
->cpumask
);
2510 hctx
->dispatch_from
= NULL
;
2514 * Map software to hardware queues.
2516 * If the cpu isn't present, the cpu is mapped to first hctx.
2518 for_each_possible_cpu(i
) {
2519 hctx_idx
= set
->map
[HCTX_TYPE_DEFAULT
].mq_map
[i
];
2520 /* unmapped hw queue can be remapped after CPU topo changed */
2521 if (!set
->tags
[hctx_idx
] &&
2522 !__blk_mq_alloc_rq_map(set
, hctx_idx
)) {
2524 * If tags initialization fail for some hctx,
2525 * that hctx won't be brought online. In this
2526 * case, remap the current ctx to hctx[0] which
2527 * is guaranteed to always have tags allocated
2529 set
->map
[HCTX_TYPE_DEFAULT
].mq_map
[i
] = 0;
2532 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2533 for (j
= 0; j
< set
->nr_maps
; j
++) {
2534 if (!set
->map
[j
].nr_queues
) {
2535 ctx
->hctxs
[j
] = blk_mq_map_queue_type(q
,
2536 HCTX_TYPE_DEFAULT
, i
);
2540 hctx
= blk_mq_map_queue_type(q
, j
, i
);
2541 ctx
->hctxs
[j
] = hctx
;
2543 * If the CPU is already set in the mask, then we've
2544 * mapped this one already. This can happen if
2545 * devices share queues across queue maps.
2547 if (cpumask_test_cpu(i
, hctx
->cpumask
))
2550 cpumask_set_cpu(i
, hctx
->cpumask
);
2552 ctx
->index_hw
[hctx
->type
] = hctx
->nr_ctx
;
2553 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2556 * If the nr_ctx type overflows, we have exceeded the
2557 * amount of sw queues we can support.
2559 BUG_ON(!hctx
->nr_ctx
);
2562 for (; j
< HCTX_MAX_TYPES
; j
++)
2563 ctx
->hctxs
[j
] = blk_mq_map_queue_type(q
,
2564 HCTX_TYPE_DEFAULT
, i
);
2567 queue_for_each_hw_ctx(q
, hctx
, i
) {
2569 * If no software queues are mapped to this hardware queue,
2570 * disable it and free the request entries.
2572 if (!hctx
->nr_ctx
) {
2573 /* Never unmap queue 0. We need it as a
2574 * fallback in case of a new remap fails
2577 if (i
&& set
->tags
[i
])
2578 blk_mq_free_map_and_requests(set
, i
);
2584 hctx
->tags
= set
->tags
[i
];
2585 WARN_ON(!hctx
->tags
);
2588 * Set the map size to the number of mapped software queues.
2589 * This is more accurate and more efficient than looping
2590 * over all possibly mapped software queues.
2592 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2595 * Initialize batch roundrobin counts
2597 hctx
->next_cpu
= blk_mq_first_mapped_cpu(hctx
);
2598 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2603 * Caller needs to ensure that we're either frozen/quiesced, or that
2604 * the queue isn't live yet.
2606 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2608 struct blk_mq_hw_ctx
*hctx
;
2611 queue_for_each_hw_ctx(q
, hctx
, i
) {
2613 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
2615 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2619 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
,
2622 struct request_queue
*q
;
2624 lockdep_assert_held(&set
->tag_list_lock
);
2626 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2627 blk_mq_freeze_queue(q
);
2628 queue_set_hctx_shared(q
, shared
);
2629 blk_mq_unfreeze_queue(q
);
2633 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
2635 struct blk_mq_tag_set
*set
= q
->tag_set
;
2637 mutex_lock(&set
->tag_list_lock
);
2638 list_del_rcu(&q
->tag_set_list
);
2639 if (list_is_singular(&set
->tag_list
)) {
2640 /* just transitioned to unshared */
2641 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2642 /* update existing queue */
2643 blk_mq_update_tag_set_depth(set
, false);
2645 mutex_unlock(&set
->tag_list_lock
);
2646 INIT_LIST_HEAD(&q
->tag_set_list
);
2649 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
2650 struct request_queue
*q
)
2652 mutex_lock(&set
->tag_list_lock
);
2655 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2657 if (!list_empty(&set
->tag_list
) &&
2658 !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
2659 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
2660 /* update existing queue */
2661 blk_mq_update_tag_set_depth(set
, true);
2663 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
2664 queue_set_hctx_shared(q
, true);
2665 list_add_tail_rcu(&q
->tag_set_list
, &set
->tag_list
);
2667 mutex_unlock(&set
->tag_list_lock
);
2670 /* All allocations will be freed in release handler of q->mq_kobj */
2671 static int blk_mq_alloc_ctxs(struct request_queue
*q
)
2673 struct blk_mq_ctxs
*ctxs
;
2676 ctxs
= kzalloc(sizeof(*ctxs
), GFP_KERNEL
);
2680 ctxs
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2681 if (!ctxs
->queue_ctx
)
2684 for_each_possible_cpu(cpu
) {
2685 struct blk_mq_ctx
*ctx
= per_cpu_ptr(ctxs
->queue_ctx
, cpu
);
2689 q
->mq_kobj
= &ctxs
->kobj
;
2690 q
->queue_ctx
= ctxs
->queue_ctx
;
2699 * It is the actual release handler for mq, but we do it from
2700 * request queue's release handler for avoiding use-after-free
2701 * and headache because q->mq_kobj shouldn't have been introduced,
2702 * but we can't group ctx/kctx kobj without it.
2704 void blk_mq_release(struct request_queue
*q
)
2706 struct blk_mq_hw_ctx
*hctx
, *next
;
2709 queue_for_each_hw_ctx(q
, hctx
, i
)
2710 WARN_ON_ONCE(hctx
&& list_empty(&hctx
->hctx_list
));
2712 /* all hctx are in .unused_hctx_list now */
2713 list_for_each_entry_safe(hctx
, next
, &q
->unused_hctx_list
, hctx_list
) {
2714 list_del_init(&hctx
->hctx_list
);
2715 kobject_put(&hctx
->kobj
);
2718 kfree(q
->queue_hw_ctx
);
2721 * release .mq_kobj and sw queue's kobject now because
2722 * both share lifetime with request queue.
2724 blk_mq_sysfs_deinit(q
);
2727 struct request_queue
*blk_mq_init_queue_data(struct blk_mq_tag_set
*set
,
2730 struct request_queue
*uninit_q
, *q
;
2732 uninit_q
= __blk_alloc_queue(set
->numa_node
);
2734 return ERR_PTR(-ENOMEM
);
2735 uninit_q
->queuedata
= queuedata
;
2738 * Initialize the queue without an elevator. device_add_disk() will do
2739 * the initialization.
2741 q
= blk_mq_init_allocated_queue(set
, uninit_q
, false);
2743 blk_cleanup_queue(uninit_q
);
2747 EXPORT_SYMBOL_GPL(blk_mq_init_queue_data
);
2749 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2751 return blk_mq_init_queue_data(set
, NULL
);
2753 EXPORT_SYMBOL(blk_mq_init_queue
);
2756 * Helper for setting up a queue with mq ops, given queue depth, and
2757 * the passed in mq ops flags.
2759 struct request_queue
*blk_mq_init_sq_queue(struct blk_mq_tag_set
*set
,
2760 const struct blk_mq_ops
*ops
,
2761 unsigned int queue_depth
,
2762 unsigned int set_flags
)
2764 struct request_queue
*q
;
2767 memset(set
, 0, sizeof(*set
));
2769 set
->nr_hw_queues
= 1;
2771 set
->queue_depth
= queue_depth
;
2772 set
->numa_node
= NUMA_NO_NODE
;
2773 set
->flags
= set_flags
;
2775 ret
= blk_mq_alloc_tag_set(set
);
2777 return ERR_PTR(ret
);
2779 q
= blk_mq_init_queue(set
);
2781 blk_mq_free_tag_set(set
);
2787 EXPORT_SYMBOL(blk_mq_init_sq_queue
);
2789 static struct blk_mq_hw_ctx
*blk_mq_alloc_and_init_hctx(
2790 struct blk_mq_tag_set
*set
, struct request_queue
*q
,
2791 int hctx_idx
, int node
)
2793 struct blk_mq_hw_ctx
*hctx
= NULL
, *tmp
;
2795 /* reuse dead hctx first */
2796 spin_lock(&q
->unused_hctx_lock
);
2797 list_for_each_entry(tmp
, &q
->unused_hctx_list
, hctx_list
) {
2798 if (tmp
->numa_node
== node
) {
2804 list_del_init(&hctx
->hctx_list
);
2805 spin_unlock(&q
->unused_hctx_lock
);
2808 hctx
= blk_mq_alloc_hctx(q
, set
, node
);
2812 if (blk_mq_init_hctx(q
, set
, hctx
, hctx_idx
))
2818 kobject_put(&hctx
->kobj
);
2823 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2824 struct request_queue
*q
)
2827 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2829 if (q
->nr_hw_queues
< set
->nr_hw_queues
) {
2830 struct blk_mq_hw_ctx
**new_hctxs
;
2832 new_hctxs
= kcalloc_node(set
->nr_hw_queues
,
2833 sizeof(*new_hctxs
), GFP_KERNEL
,
2838 memcpy(new_hctxs
, hctxs
, q
->nr_hw_queues
*
2840 q
->queue_hw_ctx
= new_hctxs
;
2845 /* protect against switching io scheduler */
2846 mutex_lock(&q
->sysfs_lock
);
2847 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2849 struct blk_mq_hw_ctx
*hctx
;
2851 node
= blk_mq_hw_queue_to_node(&set
->map
[HCTX_TYPE_DEFAULT
], i
);
2853 * If the hw queue has been mapped to another numa node,
2854 * we need to realloc the hctx. If allocation fails, fallback
2855 * to use the previous one.
2857 if (hctxs
[i
] && (hctxs
[i
]->numa_node
== node
))
2860 hctx
= blk_mq_alloc_and_init_hctx(set
, q
, i
, node
);
2863 blk_mq_exit_hctx(q
, set
, hctxs
[i
], i
);
2867 pr_warn("Allocate new hctx on node %d fails,\
2868 fallback to previous one on node %d\n",
2869 node
, hctxs
[i
]->numa_node
);
2875 * Increasing nr_hw_queues fails. Free the newly allocated
2876 * hctxs and keep the previous q->nr_hw_queues.
2878 if (i
!= set
->nr_hw_queues
) {
2879 j
= q
->nr_hw_queues
;
2883 end
= q
->nr_hw_queues
;
2884 q
->nr_hw_queues
= set
->nr_hw_queues
;
2887 for (; j
< end
; j
++) {
2888 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2892 blk_mq_free_map_and_requests(set
, j
);
2893 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2897 mutex_unlock(&q
->sysfs_lock
);
2900 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2901 struct request_queue
*q
,
2904 /* mark the queue as mq asap */
2905 q
->mq_ops
= set
->ops
;
2907 q
->poll_cb
= blk_stat_alloc_callback(blk_mq_poll_stats_fn
,
2908 blk_mq_poll_stats_bkt
,
2909 BLK_MQ_POLL_STATS_BKTS
, q
);
2913 if (blk_mq_alloc_ctxs(q
))
2916 /* init q->mq_kobj and sw queues' kobjects */
2917 blk_mq_sysfs_init(q
);
2919 INIT_LIST_HEAD(&q
->unused_hctx_list
);
2920 spin_lock_init(&q
->unused_hctx_lock
);
2922 blk_mq_realloc_hw_ctxs(set
, q
);
2923 if (!q
->nr_hw_queues
)
2926 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2927 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2931 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2932 if (set
->nr_maps
> HCTX_TYPE_POLL
&&
2933 set
->map
[HCTX_TYPE_POLL
].nr_queues
)
2934 blk_queue_flag_set(QUEUE_FLAG_POLL
, q
);
2936 q
->sg_reserved_size
= INT_MAX
;
2938 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2939 INIT_LIST_HEAD(&q
->requeue_list
);
2940 spin_lock_init(&q
->requeue_lock
);
2942 q
->make_request_fn
= blk_mq_make_request
;
2943 q
->nr_requests
= set
->queue_depth
;
2946 * Default to classic polling
2948 q
->poll_nsec
= BLK_MQ_POLL_CLASSIC
;
2950 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2951 blk_mq_add_queue_tag_set(set
, q
);
2952 blk_mq_map_swqueue(q
);
2955 elevator_init_mq(q
);
2960 kfree(q
->queue_hw_ctx
);
2961 q
->nr_hw_queues
= 0;
2962 blk_mq_sysfs_deinit(q
);
2964 blk_stat_free_callback(q
->poll_cb
);
2968 return ERR_PTR(-ENOMEM
);
2970 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2972 /* tags can _not_ be used after returning from blk_mq_exit_queue */
2973 void blk_mq_exit_queue(struct request_queue
*q
)
2975 struct blk_mq_tag_set
*set
= q
->tag_set
;
2977 blk_mq_del_queue_tag_set(q
);
2978 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2981 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2985 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
2986 if (!__blk_mq_alloc_rq_map(set
, i
))
2993 blk_mq_free_rq_map(set
->tags
[i
]);
2999 * Allocate the request maps associated with this tag_set. Note that this
3000 * may reduce the depth asked for, if memory is tight. set->queue_depth
3001 * will be updated to reflect the allocated depth.
3003 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
3008 depth
= set
->queue_depth
;
3010 err
= __blk_mq_alloc_rq_maps(set
);
3014 set
->queue_depth
>>= 1;
3015 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
3019 } while (set
->queue_depth
);
3021 if (!set
->queue_depth
|| err
) {
3022 pr_err("blk-mq: failed to allocate request map\n");
3026 if (depth
!= set
->queue_depth
)
3027 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3028 depth
, set
->queue_depth
);
3033 static int blk_mq_update_queue_map(struct blk_mq_tag_set
*set
)
3036 * blk_mq_map_queues() and multiple .map_queues() implementations
3037 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
3038 * number of hardware queues.
3040 if (set
->nr_maps
== 1)
3041 set
->map
[HCTX_TYPE_DEFAULT
].nr_queues
= set
->nr_hw_queues
;
3043 if (set
->ops
->map_queues
&& !is_kdump_kernel()) {
3047 * transport .map_queues is usually done in the following
3050 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3051 * mask = get_cpu_mask(queue)
3052 * for_each_cpu(cpu, mask)
3053 * set->map[x].mq_map[cpu] = queue;
3056 * When we need to remap, the table has to be cleared for
3057 * killing stale mapping since one CPU may not be mapped
3060 for (i
= 0; i
< set
->nr_maps
; i
++)
3061 blk_mq_clear_mq_map(&set
->map
[i
]);
3063 return set
->ops
->map_queues(set
);
3065 BUG_ON(set
->nr_maps
> 1);
3066 return blk_mq_map_queues(&set
->map
[HCTX_TYPE_DEFAULT
]);
3070 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set
*set
,
3071 int cur_nr_hw_queues
, int new_nr_hw_queues
)
3073 struct blk_mq_tags
**new_tags
;
3075 if (cur_nr_hw_queues
>= new_nr_hw_queues
)
3078 new_tags
= kcalloc_node(new_nr_hw_queues
, sizeof(struct blk_mq_tags
*),
3079 GFP_KERNEL
, set
->numa_node
);
3084 memcpy(new_tags
, set
->tags
, cur_nr_hw_queues
*
3085 sizeof(*set
->tags
));
3087 set
->tags
= new_tags
;
3088 set
->nr_hw_queues
= new_nr_hw_queues
;
3094 * Alloc a tag set to be associated with one or more request queues.
3095 * May fail with EINVAL for various error conditions. May adjust the
3096 * requested depth down, if it's too large. In that case, the set
3097 * value will be stored in set->queue_depth.
3099 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
3103 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
3105 if (!set
->nr_hw_queues
)
3107 if (!set
->queue_depth
)
3109 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
3112 if (!set
->ops
->queue_rq
)
3115 if (!set
->ops
->get_budget
^ !set
->ops
->put_budget
)
3118 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
3119 pr_info("blk-mq: reduced tag depth to %u\n",
3121 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
3126 else if (set
->nr_maps
> HCTX_MAX_TYPES
)
3130 * If a crashdump is active, then we are potentially in a very
3131 * memory constrained environment. Limit us to 1 queue and
3132 * 64 tags to prevent using too much memory.
3134 if (is_kdump_kernel()) {
3135 set
->nr_hw_queues
= 1;
3137 set
->queue_depth
= min(64U, set
->queue_depth
);
3140 * There is no use for more h/w queues than cpus if we just have
3143 if (set
->nr_maps
== 1 && set
->nr_hw_queues
> nr_cpu_ids
)
3144 set
->nr_hw_queues
= nr_cpu_ids
;
3146 if (blk_mq_realloc_tag_set_tags(set
, 0, set
->nr_hw_queues
) < 0)
3150 for (i
= 0; i
< set
->nr_maps
; i
++) {
3151 set
->map
[i
].mq_map
= kcalloc_node(nr_cpu_ids
,
3152 sizeof(set
->map
[i
].mq_map
[0]),
3153 GFP_KERNEL
, set
->numa_node
);
3154 if (!set
->map
[i
].mq_map
)
3155 goto out_free_mq_map
;
3156 set
->map
[i
].nr_queues
= is_kdump_kernel() ? 1 : set
->nr_hw_queues
;
3159 ret
= blk_mq_update_queue_map(set
);
3161 goto out_free_mq_map
;
3163 ret
= blk_mq_alloc_rq_maps(set
);
3165 goto out_free_mq_map
;
3167 mutex_init(&set
->tag_list_lock
);
3168 INIT_LIST_HEAD(&set
->tag_list
);
3173 for (i
= 0; i
< set
->nr_maps
; i
++) {
3174 kfree(set
->map
[i
].mq_map
);
3175 set
->map
[i
].mq_map
= NULL
;
3181 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
3183 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
3187 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
3188 blk_mq_free_map_and_requests(set
, i
);
3190 for (j
= 0; j
< set
->nr_maps
; j
++) {
3191 kfree(set
->map
[j
].mq_map
);
3192 set
->map
[j
].mq_map
= NULL
;
3198 EXPORT_SYMBOL(blk_mq_free_tag_set
);
3200 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
3202 struct blk_mq_tag_set
*set
= q
->tag_set
;
3203 struct blk_mq_hw_ctx
*hctx
;
3209 if (q
->nr_requests
== nr
)
3212 blk_mq_freeze_queue(q
);
3213 blk_mq_quiesce_queue(q
);
3216 queue_for_each_hw_ctx(q
, hctx
, i
) {
3220 * If we're using an MQ scheduler, just update the scheduler
3221 * queue depth. This is similar to what the old code would do.
3223 if (!hctx
->sched_tags
) {
3224 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
, nr
,
3227 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
3232 if (q
->elevator
&& q
->elevator
->type
->ops
.depth_updated
)
3233 q
->elevator
->type
->ops
.depth_updated(hctx
);
3237 q
->nr_requests
= nr
;
3239 blk_mq_unquiesce_queue(q
);
3240 blk_mq_unfreeze_queue(q
);
3246 * request_queue and elevator_type pair.
3247 * It is just used by __blk_mq_update_nr_hw_queues to cache
3248 * the elevator_type associated with a request_queue.
3250 struct blk_mq_qe_pair
{
3251 struct list_head node
;
3252 struct request_queue
*q
;
3253 struct elevator_type
*type
;
3257 * Cache the elevator_type in qe pair list and switch the
3258 * io scheduler to 'none'
3260 static bool blk_mq_elv_switch_none(struct list_head
*head
,
3261 struct request_queue
*q
)
3263 struct blk_mq_qe_pair
*qe
;
3268 qe
= kmalloc(sizeof(*qe
), GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
);
3272 INIT_LIST_HEAD(&qe
->node
);
3274 qe
->type
= q
->elevator
->type
;
3275 list_add(&qe
->node
, head
);
3277 mutex_lock(&q
->sysfs_lock
);
3279 * After elevator_switch_mq, the previous elevator_queue will be
3280 * released by elevator_release. The reference of the io scheduler
3281 * module get by elevator_get will also be put. So we need to get
3282 * a reference of the io scheduler module here to prevent it to be
3285 __module_get(qe
->type
->elevator_owner
);
3286 elevator_switch_mq(q
, NULL
);
3287 mutex_unlock(&q
->sysfs_lock
);
3292 static void blk_mq_elv_switch_back(struct list_head
*head
,
3293 struct request_queue
*q
)
3295 struct blk_mq_qe_pair
*qe
;
3296 struct elevator_type
*t
= NULL
;
3298 list_for_each_entry(qe
, head
, node
)
3307 list_del(&qe
->node
);
3310 mutex_lock(&q
->sysfs_lock
);
3311 elevator_switch_mq(q
, t
);
3312 mutex_unlock(&q
->sysfs_lock
);
3315 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
,
3318 struct request_queue
*q
;
3320 int prev_nr_hw_queues
;
3322 lockdep_assert_held(&set
->tag_list_lock
);
3324 if (set
->nr_maps
== 1 && nr_hw_queues
> nr_cpu_ids
)
3325 nr_hw_queues
= nr_cpu_ids
;
3326 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
3329 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3330 blk_mq_freeze_queue(q
);
3332 * Switch IO scheduler to 'none', cleaning up the data associated
3333 * with the previous scheduler. We will switch back once we are done
3334 * updating the new sw to hw queue mappings.
3336 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3337 if (!blk_mq_elv_switch_none(&head
, q
))
3340 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3341 blk_mq_debugfs_unregister_hctxs(q
);
3342 blk_mq_sysfs_unregister(q
);
3345 if (blk_mq_realloc_tag_set_tags(set
, set
->nr_hw_queues
, nr_hw_queues
) <
3349 prev_nr_hw_queues
= set
->nr_hw_queues
;
3350 set
->nr_hw_queues
= nr_hw_queues
;
3351 blk_mq_update_queue_map(set
);
3353 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3354 blk_mq_realloc_hw_ctxs(set
, q
);
3355 if (q
->nr_hw_queues
!= set
->nr_hw_queues
) {
3356 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3357 nr_hw_queues
, prev_nr_hw_queues
);
3358 set
->nr_hw_queues
= prev_nr_hw_queues
;
3359 blk_mq_map_queues(&set
->map
[HCTX_TYPE_DEFAULT
]);
3362 blk_mq_map_swqueue(q
);
3366 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3367 blk_mq_sysfs_register(q
);
3368 blk_mq_debugfs_register_hctxs(q
);
3372 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3373 blk_mq_elv_switch_back(&head
, q
);
3375 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3376 blk_mq_unfreeze_queue(q
);
3379 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
3381 mutex_lock(&set
->tag_list_lock
);
3382 __blk_mq_update_nr_hw_queues(set
, nr_hw_queues
);
3383 mutex_unlock(&set
->tag_list_lock
);
3385 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
3387 /* Enable polling stats and return whether they were already enabled. */
3388 static bool blk_poll_stats_enable(struct request_queue
*q
)
3390 if (test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
3391 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS
, q
))
3393 blk_stat_add_callback(q
, q
->poll_cb
);
3397 static void blk_mq_poll_stats_start(struct request_queue
*q
)
3400 * We don't arm the callback if polling stats are not enabled or the
3401 * callback is already active.
3403 if (!test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
3404 blk_stat_is_active(q
->poll_cb
))
3407 blk_stat_activate_msecs(q
->poll_cb
, 100);
3410 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
)
3412 struct request_queue
*q
= cb
->data
;
3415 for (bucket
= 0; bucket
< BLK_MQ_POLL_STATS_BKTS
; bucket
++) {
3416 if (cb
->stat
[bucket
].nr_samples
)
3417 q
->poll_stat
[bucket
] = cb
->stat
[bucket
];
3421 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
3424 unsigned long ret
= 0;
3428 * If stats collection isn't on, don't sleep but turn it on for
3431 if (!blk_poll_stats_enable(q
))
3435 * As an optimistic guess, use half of the mean service time
3436 * for this type of request. We can (and should) make this smarter.
3437 * For instance, if the completion latencies are tight, we can
3438 * get closer than just half the mean. This is especially
3439 * important on devices where the completion latencies are longer
3440 * than ~10 usec. We do use the stats for the relevant IO size
3441 * if available which does lead to better estimates.
3443 bucket
= blk_mq_poll_stats_bkt(rq
);
3447 if (q
->poll_stat
[bucket
].nr_samples
)
3448 ret
= (q
->poll_stat
[bucket
].mean
+ 1) / 2;
3453 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
3456 struct hrtimer_sleeper hs
;
3457 enum hrtimer_mode mode
;
3461 if (rq
->rq_flags
& RQF_MQ_POLL_SLEPT
)
3465 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3467 * 0: use half of prev avg
3468 * >0: use this specific value
3470 if (q
->poll_nsec
> 0)
3471 nsecs
= q
->poll_nsec
;
3473 nsecs
= blk_mq_poll_nsecs(q
, rq
);
3478 rq
->rq_flags
|= RQF_MQ_POLL_SLEPT
;
3481 * This will be replaced with the stats tracking code, using
3482 * 'avg_completion_time / 2' as the pre-sleep target.
3486 mode
= HRTIMER_MODE_REL
;
3487 hrtimer_init_sleeper_on_stack(&hs
, CLOCK_MONOTONIC
, mode
);
3488 hrtimer_set_expires(&hs
.timer
, kt
);
3491 if (blk_mq_rq_state(rq
) == MQ_RQ_COMPLETE
)
3493 set_current_state(TASK_UNINTERRUPTIBLE
);
3494 hrtimer_sleeper_start_expires(&hs
, mode
);
3497 hrtimer_cancel(&hs
.timer
);
3498 mode
= HRTIMER_MODE_ABS
;
3499 } while (hs
.task
&& !signal_pending(current
));
3501 __set_current_state(TASK_RUNNING
);
3502 destroy_hrtimer_on_stack(&hs
.timer
);
3506 static bool blk_mq_poll_hybrid(struct request_queue
*q
,
3507 struct blk_mq_hw_ctx
*hctx
, blk_qc_t cookie
)
3511 if (q
->poll_nsec
== BLK_MQ_POLL_CLASSIC
)
3514 if (!blk_qc_t_is_internal(cookie
))
3515 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
3517 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
3519 * With scheduling, if the request has completed, we'll
3520 * get a NULL return here, as we clear the sched tag when
3521 * that happens. The request still remains valid, like always,
3522 * so we should be safe with just the NULL check.
3528 return blk_mq_poll_hybrid_sleep(q
, rq
);
3532 * blk_poll - poll for IO completions
3534 * @cookie: cookie passed back at IO submission time
3535 * @spin: whether to spin for completions
3538 * Poll for completions on the passed in queue. Returns number of
3539 * completed entries found. If @spin is true, then blk_poll will continue
3540 * looping until at least one completion is found, unless the task is
3541 * otherwise marked running (or we need to reschedule).
3543 int blk_poll(struct request_queue
*q
, blk_qc_t cookie
, bool spin
)
3545 struct blk_mq_hw_ctx
*hctx
;
3548 if (!blk_qc_t_valid(cookie
) ||
3549 !test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
3553 blk_flush_plug_list(current
->plug
, false);
3555 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
3558 * If we sleep, have the caller restart the poll loop to reset
3559 * the state. Like for the other success return cases, the
3560 * caller is responsible for checking if the IO completed. If
3561 * the IO isn't complete, we'll get called again and will go
3562 * straight to the busy poll loop.
3564 if (blk_mq_poll_hybrid(q
, hctx
, cookie
))
3567 hctx
->poll_considered
++;
3569 state
= current
->state
;
3573 hctx
->poll_invoked
++;
3575 ret
= q
->mq_ops
->poll(hctx
);
3577 hctx
->poll_success
++;
3578 __set_current_state(TASK_RUNNING
);
3582 if (signal_pending_state(state
, current
))
3583 __set_current_state(TASK_RUNNING
);
3585 if (current
->state
== TASK_RUNNING
)
3587 if (ret
< 0 || !spin
)
3590 } while (!need_resched());
3592 __set_current_state(TASK_RUNNING
);
3595 EXPORT_SYMBOL_GPL(blk_poll
);
3597 unsigned int blk_mq_rq_cpu(struct request
*rq
)
3599 return rq
->mq_ctx
->cpu
;
3601 EXPORT_SYMBOL(blk_mq_rq_cpu
);
3603 static int __init
blk_mq_init(void)
3605 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
3606 blk_mq_hctx_notify_dead
);
3609 subsys_initcall(blk_mq_init
);