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>
35 #include "blk-mq-debugfs.h"
36 #include "blk-mq-tag.h"
39 #include "blk-mq-sched.h"
40 #include "blk-rq-qos.h"
42 static void blk_mq_poll_stats_start(struct request_queue
*q
);
43 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
);
45 static int blk_mq_poll_stats_bkt(const struct request
*rq
)
47 int ddir
, bytes
, bucket
;
49 ddir
= rq_data_dir(rq
);
50 bytes
= blk_rq_bytes(rq
);
52 bucket
= ddir
+ 2*(ilog2(bytes
) - 9);
56 else if (bucket
>= BLK_MQ_POLL_STATS_BKTS
)
57 return ddir
+ BLK_MQ_POLL_STATS_BKTS
- 2;
63 * Check if any of the ctx, dispatch list or elevator
64 * have pending work in this hardware queue.
66 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
68 return !list_empty_careful(&hctx
->dispatch
) ||
69 sbitmap_any_bit_set(&hctx
->ctx_map
) ||
70 blk_mq_sched_has_work(hctx
);
74 * Mark this ctx as having pending work in this hardware queue
76 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
77 struct blk_mq_ctx
*ctx
)
79 const int bit
= ctx
->index_hw
[hctx
->type
];
81 if (!sbitmap_test_bit(&hctx
->ctx_map
, bit
))
82 sbitmap_set_bit(&hctx
->ctx_map
, bit
);
85 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
86 struct blk_mq_ctx
*ctx
)
88 const int bit
= ctx
->index_hw
[hctx
->type
];
90 sbitmap_clear_bit(&hctx
->ctx_map
, bit
);
94 struct hd_struct
*part
;
95 unsigned int *inflight
;
98 static bool blk_mq_check_inflight(struct blk_mq_hw_ctx
*hctx
,
99 struct request
*rq
, void *priv
,
102 struct mq_inflight
*mi
= priv
;
105 * index[0] counts the specific partition that was asked for.
107 if (rq
->part
== mi
->part
)
113 unsigned int blk_mq_in_flight(struct request_queue
*q
, struct hd_struct
*part
)
115 unsigned inflight
[2];
116 struct mq_inflight mi
= { .part
= part
, .inflight
= inflight
, };
118 inflight
[0] = inflight
[1] = 0;
119 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight
, &mi
);
124 static bool blk_mq_check_inflight_rw(struct blk_mq_hw_ctx
*hctx
,
125 struct request
*rq
, void *priv
,
128 struct mq_inflight
*mi
= priv
;
130 if (rq
->part
== mi
->part
)
131 mi
->inflight
[rq_data_dir(rq
)]++;
136 void blk_mq_in_flight_rw(struct request_queue
*q
, struct hd_struct
*part
,
137 unsigned int inflight
[2])
139 struct mq_inflight mi
= { .part
= part
, .inflight
= inflight
, };
141 inflight
[0] = inflight
[1] = 0;
142 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight_rw
, &mi
);
145 void blk_freeze_queue_start(struct request_queue
*q
)
147 mutex_lock(&q
->mq_freeze_lock
);
148 if (++q
->mq_freeze_depth
== 1) {
149 percpu_ref_kill(&q
->q_usage_counter
);
150 mutex_unlock(&q
->mq_freeze_lock
);
152 blk_mq_run_hw_queues(q
, false);
154 mutex_unlock(&q
->mq_freeze_lock
);
157 EXPORT_SYMBOL_GPL(blk_freeze_queue_start
);
159 void blk_mq_freeze_queue_wait(struct request_queue
*q
)
161 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
163 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait
);
165 int blk_mq_freeze_queue_wait_timeout(struct request_queue
*q
,
166 unsigned long timeout
)
168 return wait_event_timeout(q
->mq_freeze_wq
,
169 percpu_ref_is_zero(&q
->q_usage_counter
),
172 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout
);
175 * Guarantee no request is in use, so we can change any data structure of
176 * the queue afterward.
178 void blk_freeze_queue(struct request_queue
*q
)
181 * In the !blk_mq case we are only calling this to kill the
182 * q_usage_counter, otherwise this increases the freeze depth
183 * and waits for it to return to zero. For this reason there is
184 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
185 * exported to drivers as the only user for unfreeze is blk_mq.
187 blk_freeze_queue_start(q
);
188 blk_mq_freeze_queue_wait(q
);
191 void blk_mq_freeze_queue(struct request_queue
*q
)
194 * ...just an alias to keep freeze and unfreeze actions balanced
195 * in the blk_mq_* namespace
199 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
201 void blk_mq_unfreeze_queue(struct request_queue
*q
)
203 mutex_lock(&q
->mq_freeze_lock
);
204 q
->mq_freeze_depth
--;
205 WARN_ON_ONCE(q
->mq_freeze_depth
< 0);
206 if (!q
->mq_freeze_depth
) {
207 percpu_ref_resurrect(&q
->q_usage_counter
);
208 wake_up_all(&q
->mq_freeze_wq
);
210 mutex_unlock(&q
->mq_freeze_lock
);
212 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
215 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
216 * mpt3sas driver such that this function can be removed.
218 void blk_mq_quiesce_queue_nowait(struct request_queue
*q
)
220 blk_queue_flag_set(QUEUE_FLAG_QUIESCED
, q
);
222 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait
);
225 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
228 * Note: this function does not prevent that the struct request end_io()
229 * callback function is invoked. Once this function is returned, we make
230 * sure no dispatch can happen until the queue is unquiesced via
231 * blk_mq_unquiesce_queue().
233 void blk_mq_quiesce_queue(struct request_queue
*q
)
235 struct blk_mq_hw_ctx
*hctx
;
239 blk_mq_quiesce_queue_nowait(q
);
241 queue_for_each_hw_ctx(q
, hctx
, i
) {
242 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
243 synchronize_srcu(hctx
->srcu
);
250 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
253 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
256 * This function recovers queue into the state before quiescing
257 * which is done by blk_mq_quiesce_queue.
259 void blk_mq_unquiesce_queue(struct request_queue
*q
)
261 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED
, q
);
263 /* dispatch requests which are inserted during quiescing */
264 blk_mq_run_hw_queues(q
, true);
266 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue
);
268 void blk_mq_wake_waiters(struct request_queue
*q
)
270 struct blk_mq_hw_ctx
*hctx
;
273 queue_for_each_hw_ctx(q
, hctx
, i
)
274 if (blk_mq_hw_queue_mapped(hctx
))
275 blk_mq_tag_wakeup_all(hctx
->tags
, true);
278 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
280 return blk_mq_has_free_tags(hctx
->tags
);
282 EXPORT_SYMBOL(blk_mq_can_queue
);
285 * Only need start/end time stamping if we have stats enabled, or using
288 static inline bool blk_mq_need_time_stamp(struct request
*rq
)
290 return (rq
->rq_flags
& RQF_IO_STAT
) || rq
->q
->elevator
;
293 static struct request
*blk_mq_rq_ctx_init(struct blk_mq_alloc_data
*data
,
294 unsigned int tag
, unsigned int op
)
296 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
297 struct request
*rq
= tags
->static_rqs
[tag
];
298 req_flags_t rq_flags
= 0;
300 if (data
->flags
& BLK_MQ_REQ_INTERNAL
) {
302 rq
->internal_tag
= tag
;
304 if (data
->hctx
->flags
& BLK_MQ_F_TAG_SHARED
) {
305 rq_flags
= RQF_MQ_INFLIGHT
;
306 atomic_inc(&data
->hctx
->nr_active
);
309 rq
->internal_tag
= -1;
310 data
->hctx
->tags
->rqs
[rq
->tag
] = rq
;
313 /* csd/requeue_work/fifo_time is initialized before use */
315 rq
->mq_ctx
= data
->ctx
;
316 rq
->mq_hctx
= data
->hctx
;
317 rq
->rq_flags
= rq_flags
;
319 if (data
->flags
& BLK_MQ_REQ_PREEMPT
)
320 rq
->rq_flags
|= RQF_PREEMPT
;
321 if (blk_queue_io_stat(data
->q
))
322 rq
->rq_flags
|= RQF_IO_STAT
;
323 INIT_LIST_HEAD(&rq
->queuelist
);
324 INIT_HLIST_NODE(&rq
->hash
);
325 RB_CLEAR_NODE(&rq
->rb_node
);
328 if (blk_mq_need_time_stamp(rq
))
329 rq
->start_time_ns
= ktime_get_ns();
331 rq
->start_time_ns
= 0;
332 rq
->io_start_time_ns
= 0;
333 rq
->nr_phys_segments
= 0;
334 #if defined(CONFIG_BLK_DEV_INTEGRITY)
335 rq
->nr_integrity_segments
= 0;
337 /* tag was already set */
339 WRITE_ONCE(rq
->deadline
, 0);
344 rq
->end_io_data
= NULL
;
346 data
->ctx
->rq_dispatched
[op_is_sync(op
)]++;
347 refcount_set(&rq
->ref
, 1);
351 static struct request
*blk_mq_get_request(struct request_queue
*q
,
353 struct blk_mq_alloc_data
*data
)
355 struct elevator_queue
*e
= q
->elevator
;
358 bool put_ctx_on_error
= false;
360 blk_queue_enter_live(q
);
362 if (likely(!data
->ctx
)) {
363 data
->ctx
= blk_mq_get_ctx(q
);
364 put_ctx_on_error
= true;
366 if (likely(!data
->hctx
))
367 data
->hctx
= blk_mq_map_queue(q
, data
->cmd_flags
,
369 if (data
->cmd_flags
& REQ_NOWAIT
)
370 data
->flags
|= BLK_MQ_REQ_NOWAIT
;
373 data
->flags
|= BLK_MQ_REQ_INTERNAL
;
376 * Flush requests are special and go directly to the
377 * dispatch list. Don't include reserved tags in the
378 * limiting, as it isn't useful.
380 if (!op_is_flush(data
->cmd_flags
) &&
381 e
->type
->ops
.limit_depth
&&
382 !(data
->flags
& BLK_MQ_REQ_RESERVED
))
383 e
->type
->ops
.limit_depth(data
->cmd_flags
, data
);
385 blk_mq_tag_busy(data
->hctx
);
388 tag
= blk_mq_get_tag(data
);
389 if (tag
== BLK_MQ_TAG_FAIL
) {
390 if (put_ctx_on_error
) {
391 blk_mq_put_ctx(data
->ctx
);
398 rq
= blk_mq_rq_ctx_init(data
, tag
, data
->cmd_flags
);
399 if (!op_is_flush(data
->cmd_flags
)) {
401 if (e
&& e
->type
->ops
.prepare_request
) {
402 if (e
->type
->icq_cache
)
403 blk_mq_sched_assign_ioc(rq
);
405 e
->type
->ops
.prepare_request(rq
, bio
);
406 rq
->rq_flags
|= RQF_ELVPRIV
;
409 data
->hctx
->queued
++;
413 struct request
*blk_mq_alloc_request(struct request_queue
*q
, unsigned int op
,
414 blk_mq_req_flags_t flags
)
416 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
, .cmd_flags
= op
};
420 ret
= blk_queue_enter(q
, flags
);
424 rq
= blk_mq_get_request(q
, NULL
, &alloc_data
);
428 return ERR_PTR(-EWOULDBLOCK
);
430 blk_mq_put_ctx(alloc_data
.ctx
);
433 rq
->__sector
= (sector_t
) -1;
434 rq
->bio
= rq
->biotail
= NULL
;
437 EXPORT_SYMBOL(blk_mq_alloc_request
);
439 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
,
440 unsigned int op
, blk_mq_req_flags_t flags
, unsigned int hctx_idx
)
442 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
, .cmd_flags
= op
};
448 * If the tag allocator sleeps we could get an allocation for a
449 * different hardware context. No need to complicate the low level
450 * allocator for this for the rare use case of a command tied to
453 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
454 return ERR_PTR(-EINVAL
);
456 if (hctx_idx
>= q
->nr_hw_queues
)
457 return ERR_PTR(-EIO
);
459 ret
= blk_queue_enter(q
, flags
);
464 * Check if the hardware context is actually mapped to anything.
465 * If not tell the caller that it should skip this queue.
467 alloc_data
.hctx
= q
->queue_hw_ctx
[hctx_idx
];
468 if (!blk_mq_hw_queue_mapped(alloc_data
.hctx
)) {
470 return ERR_PTR(-EXDEV
);
472 cpu
= cpumask_first_and(alloc_data
.hctx
->cpumask
, cpu_online_mask
);
473 alloc_data
.ctx
= __blk_mq_get_ctx(q
, cpu
);
475 rq
= blk_mq_get_request(q
, NULL
, &alloc_data
);
479 return ERR_PTR(-EWOULDBLOCK
);
483 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
485 static void __blk_mq_free_request(struct request
*rq
)
487 struct request_queue
*q
= rq
->q
;
488 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
489 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
490 const int sched_tag
= rq
->internal_tag
;
492 blk_pm_mark_last_busy(rq
);
495 blk_mq_put_tag(hctx
, hctx
->tags
, ctx
, rq
->tag
);
497 blk_mq_put_tag(hctx
, hctx
->sched_tags
, ctx
, sched_tag
);
498 blk_mq_sched_restart(hctx
);
502 void blk_mq_free_request(struct request
*rq
)
504 struct request_queue
*q
= rq
->q
;
505 struct elevator_queue
*e
= q
->elevator
;
506 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
507 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
509 if (rq
->rq_flags
& RQF_ELVPRIV
) {
510 if (e
&& e
->type
->ops
.finish_request
)
511 e
->type
->ops
.finish_request(rq
);
513 put_io_context(rq
->elv
.icq
->ioc
);
518 ctx
->rq_completed
[rq_is_sync(rq
)]++;
519 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
520 atomic_dec(&hctx
->nr_active
);
522 if (unlikely(laptop_mode
&& !blk_rq_is_passthrough(rq
)))
523 laptop_io_completion(q
->backing_dev_info
);
527 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
528 if (refcount_dec_and_test(&rq
->ref
))
529 __blk_mq_free_request(rq
);
531 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
533 inline void __blk_mq_end_request(struct request
*rq
, blk_status_t error
)
537 if (blk_mq_need_time_stamp(rq
))
538 now
= ktime_get_ns();
540 if (rq
->rq_flags
& RQF_STATS
) {
541 blk_mq_poll_stats_start(rq
->q
);
542 blk_stat_add(rq
, now
);
545 if (rq
->internal_tag
!= -1)
546 blk_mq_sched_completed_request(rq
, now
);
548 blk_account_io_done(rq
, now
);
551 rq_qos_done(rq
->q
, rq
);
552 rq
->end_io(rq
, error
);
554 blk_mq_free_request(rq
);
557 EXPORT_SYMBOL(__blk_mq_end_request
);
559 void blk_mq_end_request(struct request
*rq
, blk_status_t error
)
561 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
563 __blk_mq_end_request(rq
, error
);
565 EXPORT_SYMBOL(blk_mq_end_request
);
567 static void __blk_mq_complete_request_remote(void *data
)
569 struct request
*rq
= data
;
570 struct request_queue
*q
= rq
->q
;
572 q
->mq_ops
->complete(rq
);
575 static void __blk_mq_complete_request(struct request
*rq
)
577 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
578 struct request_queue
*q
= rq
->q
;
582 WRITE_ONCE(rq
->state
, MQ_RQ_COMPLETE
);
584 * Most of single queue controllers, there is only one irq vector
585 * for handling IO completion, and the only irq's affinity is set
586 * as all possible CPUs. On most of ARCHs, this affinity means the
587 * irq is handled on one specific CPU.
589 * So complete IO reqeust in softirq context in case of single queue
590 * for not degrading IO performance by irqsoff latency.
592 if (q
->nr_hw_queues
== 1) {
593 __blk_complete_request(rq
);
598 * For a polled request, always complete locallly, it's pointless
599 * to redirect the completion.
601 if ((rq
->cmd_flags
& REQ_HIPRI
) ||
602 !test_bit(QUEUE_FLAG_SAME_COMP
, &q
->queue_flags
)) {
603 q
->mq_ops
->complete(rq
);
608 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &q
->queue_flags
))
609 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
611 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
612 rq
->csd
.func
= __blk_mq_complete_request_remote
;
615 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
617 q
->mq_ops
->complete(rq
);
622 static void hctx_unlock(struct blk_mq_hw_ctx
*hctx
, int srcu_idx
)
623 __releases(hctx
->srcu
)
625 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
))
628 srcu_read_unlock(hctx
->srcu
, srcu_idx
);
631 static void hctx_lock(struct blk_mq_hw_ctx
*hctx
, int *srcu_idx
)
632 __acquires(hctx
->srcu
)
634 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
635 /* shut up gcc false positive */
639 *srcu_idx
= srcu_read_lock(hctx
->srcu
);
643 * blk_mq_complete_request - end I/O on a request
644 * @rq: the request being processed
647 * Ends all I/O on a request. It does not handle partial completions.
648 * The actual completion happens out-of-order, through a IPI handler.
650 bool blk_mq_complete_request(struct request
*rq
)
652 if (unlikely(blk_should_fake_timeout(rq
->q
)))
654 __blk_mq_complete_request(rq
);
657 EXPORT_SYMBOL(blk_mq_complete_request
);
659 void blk_mq_complete_request_sync(struct request
*rq
)
661 WRITE_ONCE(rq
->state
, MQ_RQ_COMPLETE
);
662 rq
->q
->mq_ops
->complete(rq
);
664 EXPORT_SYMBOL_GPL(blk_mq_complete_request_sync
);
666 int blk_mq_request_started(struct request
*rq
)
668 return blk_mq_rq_state(rq
) != MQ_RQ_IDLE
;
670 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
672 void blk_mq_start_request(struct request
*rq
)
674 struct request_queue
*q
= rq
->q
;
676 blk_mq_sched_started_request(rq
);
678 trace_block_rq_issue(q
, rq
);
680 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
681 rq
->io_start_time_ns
= ktime_get_ns();
682 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
683 rq
->throtl_size
= blk_rq_sectors(rq
);
685 rq
->rq_flags
|= RQF_STATS
;
689 WARN_ON_ONCE(blk_mq_rq_state(rq
) != MQ_RQ_IDLE
);
692 WRITE_ONCE(rq
->state
, MQ_RQ_IN_FLIGHT
);
694 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
696 * Make sure space for the drain appears. We know we can do
697 * this because max_hw_segments has been adjusted to be one
698 * fewer than the device can handle.
700 rq
->nr_phys_segments
++;
703 EXPORT_SYMBOL(blk_mq_start_request
);
705 static void __blk_mq_requeue_request(struct request
*rq
)
707 struct request_queue
*q
= rq
->q
;
709 blk_mq_put_driver_tag(rq
);
711 trace_block_rq_requeue(q
, rq
);
712 rq_qos_requeue(q
, rq
);
714 if (blk_mq_request_started(rq
)) {
715 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
716 rq
->rq_flags
&= ~RQF_TIMED_OUT
;
717 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
718 rq
->nr_phys_segments
--;
722 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
724 __blk_mq_requeue_request(rq
);
726 /* this request will be re-inserted to io scheduler queue */
727 blk_mq_sched_requeue_request(rq
);
729 BUG_ON(!list_empty(&rq
->queuelist
));
730 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
732 EXPORT_SYMBOL(blk_mq_requeue_request
);
734 static void blk_mq_requeue_work(struct work_struct
*work
)
736 struct request_queue
*q
=
737 container_of(work
, struct request_queue
, requeue_work
.work
);
739 struct request
*rq
, *next
;
741 spin_lock_irq(&q
->requeue_lock
);
742 list_splice_init(&q
->requeue_list
, &rq_list
);
743 spin_unlock_irq(&q
->requeue_lock
);
745 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
746 if (!(rq
->rq_flags
& (RQF_SOFTBARRIER
| RQF_DONTPREP
)))
749 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
750 list_del_init(&rq
->queuelist
);
752 * If RQF_DONTPREP, rq has contained some driver specific
753 * data, so insert it to hctx dispatch list to avoid any
756 if (rq
->rq_flags
& RQF_DONTPREP
)
757 blk_mq_request_bypass_insert(rq
, false);
759 blk_mq_sched_insert_request(rq
, true, false, false);
762 while (!list_empty(&rq_list
)) {
763 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
764 list_del_init(&rq
->queuelist
);
765 blk_mq_sched_insert_request(rq
, false, false, false);
768 blk_mq_run_hw_queues(q
, false);
771 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
772 bool kick_requeue_list
)
774 struct request_queue
*q
= rq
->q
;
778 * We abuse this flag that is otherwise used by the I/O scheduler to
779 * request head insertion from the workqueue.
781 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
783 spin_lock_irqsave(&q
->requeue_lock
, flags
);
785 rq
->rq_flags
|= RQF_SOFTBARRIER
;
786 list_add(&rq
->queuelist
, &q
->requeue_list
);
788 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
790 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
792 if (kick_requeue_list
)
793 blk_mq_kick_requeue_list(q
);
796 void blk_mq_kick_requeue_list(struct request_queue
*q
)
798 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
, 0);
800 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
802 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
805 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
,
806 msecs_to_jiffies(msecs
));
808 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
810 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
812 if (tag
< tags
->nr_tags
) {
813 prefetch(tags
->rqs
[tag
]);
814 return tags
->rqs
[tag
];
819 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
821 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
822 void *priv
, bool reserved
)
825 * If we find a request that is inflight and the queue matches,
826 * we know the queue is busy. Return false to stop the iteration.
828 if (rq
->state
== MQ_RQ_IN_FLIGHT
&& rq
->q
== hctx
->queue
) {
838 bool blk_mq_queue_inflight(struct request_queue
*q
)
842 blk_mq_queue_tag_busy_iter(q
, blk_mq_rq_inflight
, &busy
);
845 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight
);
847 static void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
849 req
->rq_flags
|= RQF_TIMED_OUT
;
850 if (req
->q
->mq_ops
->timeout
) {
851 enum blk_eh_timer_return ret
;
853 ret
= req
->q
->mq_ops
->timeout(req
, reserved
);
854 if (ret
== BLK_EH_DONE
)
856 WARN_ON_ONCE(ret
!= BLK_EH_RESET_TIMER
);
862 static bool blk_mq_req_expired(struct request
*rq
, unsigned long *next
)
864 unsigned long deadline
;
866 if (blk_mq_rq_state(rq
) != MQ_RQ_IN_FLIGHT
)
868 if (rq
->rq_flags
& RQF_TIMED_OUT
)
871 deadline
= READ_ONCE(rq
->deadline
);
872 if (time_after_eq(jiffies
, deadline
))
877 else if (time_after(*next
, deadline
))
882 static bool blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
883 struct request
*rq
, void *priv
, bool reserved
)
885 unsigned long *next
= priv
;
888 * Just do a quick check if it is expired before locking the request in
889 * so we're not unnecessarilly synchronizing across CPUs.
891 if (!blk_mq_req_expired(rq
, next
))
895 * We have reason to believe the request may be expired. Take a
896 * reference on the request to lock this request lifetime into its
897 * currently allocated context to prevent it from being reallocated in
898 * the event the completion by-passes this timeout handler.
900 * If the reference was already released, then the driver beat the
901 * timeout handler to posting a natural completion.
903 if (!refcount_inc_not_zero(&rq
->ref
))
907 * The request is now locked and cannot be reallocated underneath the
908 * timeout handler's processing. Re-verify this exact request is truly
909 * expired; if it is not expired, then the request was completed and
910 * reallocated as a new request.
912 if (blk_mq_req_expired(rq
, next
))
913 blk_mq_rq_timed_out(rq
, reserved
);
914 if (refcount_dec_and_test(&rq
->ref
))
915 __blk_mq_free_request(rq
);
920 static void blk_mq_timeout_work(struct work_struct
*work
)
922 struct request_queue
*q
=
923 container_of(work
, struct request_queue
, timeout_work
);
924 unsigned long next
= 0;
925 struct blk_mq_hw_ctx
*hctx
;
928 /* A deadlock might occur if a request is stuck requiring a
929 * timeout at the same time a queue freeze is waiting
930 * completion, since the timeout code would not be able to
931 * acquire the queue reference here.
933 * That's why we don't use blk_queue_enter here; instead, we use
934 * percpu_ref_tryget directly, because we need to be able to
935 * obtain a reference even in the short window between the queue
936 * starting to freeze, by dropping the first reference in
937 * blk_freeze_queue_start, and the moment the last request is
938 * consumed, marked by the instant q_usage_counter reaches
941 if (!percpu_ref_tryget(&q
->q_usage_counter
))
944 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &next
);
947 mod_timer(&q
->timeout
, next
);
950 * Request timeouts are handled as a forward rolling timer. If
951 * we end up here it means that no requests are pending and
952 * also that no request has been pending for a while. Mark
955 queue_for_each_hw_ctx(q
, hctx
, i
) {
956 /* the hctx may be unmapped, so check it here */
957 if (blk_mq_hw_queue_mapped(hctx
))
958 blk_mq_tag_idle(hctx
);
964 struct flush_busy_ctx_data
{
965 struct blk_mq_hw_ctx
*hctx
;
966 struct list_head
*list
;
969 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
971 struct flush_busy_ctx_data
*flush_data
= data
;
972 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
973 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
974 enum hctx_type type
= hctx
->type
;
976 spin_lock(&ctx
->lock
);
977 list_splice_tail_init(&ctx
->rq_lists
[type
], flush_data
->list
);
978 sbitmap_clear_bit(sb
, bitnr
);
979 spin_unlock(&ctx
->lock
);
984 * Process software queues that have been marked busy, splicing them
985 * to the for-dispatch
987 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
989 struct flush_busy_ctx_data data
= {
994 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
996 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
998 struct dispatch_rq_data
{
999 struct blk_mq_hw_ctx
*hctx
;
1003 static bool dispatch_rq_from_ctx(struct sbitmap
*sb
, unsigned int bitnr
,
1006 struct dispatch_rq_data
*dispatch_data
= data
;
1007 struct blk_mq_hw_ctx
*hctx
= dispatch_data
->hctx
;
1008 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
1009 enum hctx_type type
= hctx
->type
;
1011 spin_lock(&ctx
->lock
);
1012 if (!list_empty(&ctx
->rq_lists
[type
])) {
1013 dispatch_data
->rq
= list_entry_rq(ctx
->rq_lists
[type
].next
);
1014 list_del_init(&dispatch_data
->rq
->queuelist
);
1015 if (list_empty(&ctx
->rq_lists
[type
]))
1016 sbitmap_clear_bit(sb
, bitnr
);
1018 spin_unlock(&ctx
->lock
);
1020 return !dispatch_data
->rq
;
1023 struct request
*blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx
*hctx
,
1024 struct blk_mq_ctx
*start
)
1026 unsigned off
= start
? start
->index_hw
[hctx
->type
] : 0;
1027 struct dispatch_rq_data data
= {
1032 __sbitmap_for_each_set(&hctx
->ctx_map
, off
,
1033 dispatch_rq_from_ctx
, &data
);
1038 static inline unsigned int queued_to_index(unsigned int queued
)
1043 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
1046 bool blk_mq_get_driver_tag(struct request
*rq
)
1048 struct blk_mq_alloc_data data
= {
1050 .hctx
= rq
->mq_hctx
,
1051 .flags
= BLK_MQ_REQ_NOWAIT
,
1052 .cmd_flags
= rq
->cmd_flags
,
1059 if (blk_mq_tag_is_reserved(data
.hctx
->sched_tags
, rq
->internal_tag
))
1060 data
.flags
|= BLK_MQ_REQ_RESERVED
;
1062 shared
= blk_mq_tag_busy(data
.hctx
);
1063 rq
->tag
= blk_mq_get_tag(&data
);
1066 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
1067 atomic_inc(&data
.hctx
->nr_active
);
1069 data
.hctx
->tags
->rqs
[rq
->tag
] = rq
;
1073 return rq
->tag
!= -1;
1076 static int blk_mq_dispatch_wake(wait_queue_entry_t
*wait
, unsigned mode
,
1077 int flags
, void *key
)
1079 struct blk_mq_hw_ctx
*hctx
;
1081 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
1083 spin_lock(&hctx
->dispatch_wait_lock
);
1084 if (!list_empty(&wait
->entry
)) {
1085 struct sbitmap_queue
*sbq
;
1087 list_del_init(&wait
->entry
);
1088 sbq
= &hctx
->tags
->bitmap_tags
;
1089 atomic_dec(&sbq
->ws_active
);
1091 spin_unlock(&hctx
->dispatch_wait_lock
);
1093 blk_mq_run_hw_queue(hctx
, true);
1098 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1099 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1100 * restart. For both cases, take care to check the condition again after
1101 * marking us as waiting.
1103 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx
*hctx
,
1106 struct sbitmap_queue
*sbq
= &hctx
->tags
->bitmap_tags
;
1107 struct wait_queue_head
*wq
;
1108 wait_queue_entry_t
*wait
;
1111 if (!(hctx
->flags
& BLK_MQ_F_TAG_SHARED
)) {
1112 blk_mq_sched_mark_restart_hctx(hctx
);
1115 * It's possible that a tag was freed in the window between the
1116 * allocation failure and adding the hardware queue to the wait
1119 * Don't clear RESTART here, someone else could have set it.
1120 * At most this will cost an extra queue run.
1122 return blk_mq_get_driver_tag(rq
);
1125 wait
= &hctx
->dispatch_wait
;
1126 if (!list_empty_careful(&wait
->entry
))
1129 wq
= &bt_wait_ptr(sbq
, hctx
)->wait
;
1131 spin_lock_irq(&wq
->lock
);
1132 spin_lock(&hctx
->dispatch_wait_lock
);
1133 if (!list_empty(&wait
->entry
)) {
1134 spin_unlock(&hctx
->dispatch_wait_lock
);
1135 spin_unlock_irq(&wq
->lock
);
1139 atomic_inc(&sbq
->ws_active
);
1140 wait
->flags
&= ~WQ_FLAG_EXCLUSIVE
;
1141 __add_wait_queue(wq
, wait
);
1144 * It's possible that a tag was freed in the window between the
1145 * allocation failure and adding the hardware queue to the wait
1148 ret
= blk_mq_get_driver_tag(rq
);
1150 spin_unlock(&hctx
->dispatch_wait_lock
);
1151 spin_unlock_irq(&wq
->lock
);
1156 * We got a tag, remove ourselves from the wait queue to ensure
1157 * someone else gets the wakeup.
1159 list_del_init(&wait
->entry
);
1160 atomic_dec(&sbq
->ws_active
);
1161 spin_unlock(&hctx
->dispatch_wait_lock
);
1162 spin_unlock_irq(&wq
->lock
);
1167 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1168 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1170 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1171 * - EWMA is one simple way to compute running average value
1172 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1173 * - take 4 as factor for avoiding to get too small(0) result, and this
1174 * factor doesn't matter because EWMA decreases exponentially
1176 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx
*hctx
, bool busy
)
1180 if (hctx
->queue
->elevator
)
1183 ewma
= hctx
->dispatch_busy
;
1188 ewma
*= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
- 1;
1190 ewma
+= 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR
;
1191 ewma
/= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
;
1193 hctx
->dispatch_busy
= ewma
;
1196 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
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
) {
1266 * If an I/O scheduler has been configured and we got a
1267 * driver tag for the next request already, free it
1270 if (!list_empty(list
)) {
1271 nxt
= list_first_entry(list
, struct request
, queuelist
);
1272 blk_mq_put_driver_tag(nxt
);
1274 list_add(&rq
->queuelist
, list
);
1275 __blk_mq_requeue_request(rq
);
1279 if (unlikely(ret
!= BLK_STS_OK
)) {
1281 blk_mq_end_request(rq
, BLK_STS_IOERR
);
1286 } while (!list_empty(list
));
1288 hctx
->dispatched
[queued_to_index(queued
)]++;
1291 * Any items that need requeuing? Stuff them into hctx->dispatch,
1292 * that is where we will continue on next queue run.
1294 if (!list_empty(list
)) {
1298 * If we didn't flush the entire list, we could have told
1299 * the driver there was more coming, but that turned out to
1302 if (q
->mq_ops
->commit_rqs
)
1303 q
->mq_ops
->commit_rqs(hctx
);
1305 spin_lock(&hctx
->lock
);
1306 list_splice_init(list
, &hctx
->dispatch
);
1307 spin_unlock(&hctx
->lock
);
1310 * If SCHED_RESTART was set by the caller of this function and
1311 * it is no longer set that means that it was cleared by another
1312 * thread and hence that a queue rerun is needed.
1314 * If 'no_tag' is set, that means that we failed getting
1315 * a driver tag with an I/O scheduler attached. If our dispatch
1316 * waitqueue is no longer active, ensure that we run the queue
1317 * AFTER adding our entries back to the list.
1319 * If no I/O scheduler has been configured it is possible that
1320 * the hardware queue got stopped and restarted before requests
1321 * were pushed back onto the dispatch list. Rerun the queue to
1322 * avoid starvation. Notes:
1323 * - blk_mq_run_hw_queue() checks whether or not a queue has
1324 * been stopped before rerunning a queue.
1325 * - Some but not all block drivers stop a queue before
1326 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1329 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1330 * bit is set, run queue after a delay to avoid IO stalls
1331 * that could otherwise occur if the queue is idle.
1333 needs_restart
= blk_mq_sched_needs_restart(hctx
);
1334 if (!needs_restart
||
1335 (no_tag
&& list_empty_careful(&hctx
->dispatch_wait
.entry
)))
1336 blk_mq_run_hw_queue(hctx
, true);
1337 else if (needs_restart
&& (ret
== BLK_STS_RESOURCE
))
1338 blk_mq_delay_run_hw_queue(hctx
, BLK_MQ_RESOURCE_DELAY
);
1340 blk_mq_update_dispatch_busy(hctx
, true);
1343 blk_mq_update_dispatch_busy(hctx
, false);
1346 * If the host/device is unable to accept more work, inform the
1349 if (ret
== BLK_STS_RESOURCE
|| ret
== BLK_STS_DEV_RESOURCE
)
1352 return (queued
+ errors
) != 0;
1355 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1360 * We should be running this queue from one of the CPUs that
1363 * There are at least two related races now between setting
1364 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1365 * __blk_mq_run_hw_queue():
1367 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1368 * but later it becomes online, then this warning is harmless
1371 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1372 * but later it becomes offline, then the warning can't be
1373 * triggered, and we depend on blk-mq timeout handler to
1374 * handle dispatched requests to this hctx
1376 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
1377 cpu_online(hctx
->next_cpu
)) {
1378 printk(KERN_WARNING
"run queue from wrong CPU %d, hctx %s\n",
1379 raw_smp_processor_id(),
1380 cpumask_empty(hctx
->cpumask
) ? "inactive": "active");
1385 * We can't run the queue inline with ints disabled. Ensure that
1386 * we catch bad users of this early.
1388 WARN_ON_ONCE(in_interrupt());
1390 might_sleep_if(hctx
->flags
& BLK_MQ_F_BLOCKING
);
1392 hctx_lock(hctx
, &srcu_idx
);
1393 blk_mq_sched_dispatch_requests(hctx
);
1394 hctx_unlock(hctx
, srcu_idx
);
1397 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx
*hctx
)
1399 int cpu
= cpumask_first_and(hctx
->cpumask
, cpu_online_mask
);
1401 if (cpu
>= nr_cpu_ids
)
1402 cpu
= cpumask_first(hctx
->cpumask
);
1407 * It'd be great if the workqueue API had a way to pass
1408 * in a mask and had some smarts for more clever placement.
1409 * For now we just round-robin here, switching for every
1410 * BLK_MQ_CPU_WORK_BATCH queued items.
1412 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1415 int next_cpu
= hctx
->next_cpu
;
1417 if (hctx
->queue
->nr_hw_queues
== 1)
1418 return WORK_CPU_UNBOUND
;
1420 if (--hctx
->next_cpu_batch
<= 0) {
1422 next_cpu
= cpumask_next_and(next_cpu
, hctx
->cpumask
,
1424 if (next_cpu
>= nr_cpu_ids
)
1425 next_cpu
= blk_mq_first_mapped_cpu(hctx
);
1426 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1430 * Do unbound schedule if we can't find a online CPU for this hctx,
1431 * and it should only happen in the path of handling CPU DEAD.
1433 if (!cpu_online(next_cpu
)) {
1440 * Make sure to re-select CPU next time once after CPUs
1441 * in hctx->cpumask become online again.
1443 hctx
->next_cpu
= next_cpu
;
1444 hctx
->next_cpu_batch
= 1;
1445 return WORK_CPU_UNBOUND
;
1448 hctx
->next_cpu
= next_cpu
;
1452 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
,
1453 unsigned long msecs
)
1455 if (unlikely(blk_mq_hctx_stopped(hctx
)))
1458 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1459 int cpu
= get_cpu();
1460 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1461 __blk_mq_run_hw_queue(hctx
);
1469 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
,
1470 msecs_to_jiffies(msecs
));
1473 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1475 __blk_mq_delay_run_hw_queue(hctx
, true, msecs
);
1477 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue
);
1479 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1485 * When queue is quiesced, we may be switching io scheduler, or
1486 * updating nr_hw_queues, or other things, and we can't run queue
1487 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1489 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1492 hctx_lock(hctx
, &srcu_idx
);
1493 need_run
= !blk_queue_quiesced(hctx
->queue
) &&
1494 blk_mq_hctx_has_pending(hctx
);
1495 hctx_unlock(hctx
, srcu_idx
);
1498 __blk_mq_delay_run_hw_queue(hctx
, async
, 0);
1504 EXPORT_SYMBOL(blk_mq_run_hw_queue
);
1506 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1508 struct blk_mq_hw_ctx
*hctx
;
1511 queue_for_each_hw_ctx(q
, hctx
, i
) {
1512 if (blk_mq_hctx_stopped(hctx
))
1515 blk_mq_run_hw_queue(hctx
, async
);
1518 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1521 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1522 * @q: request queue.
1524 * The caller is responsible for serializing this function against
1525 * blk_mq_{start,stop}_hw_queue().
1527 bool blk_mq_queue_stopped(struct request_queue
*q
)
1529 struct blk_mq_hw_ctx
*hctx
;
1532 queue_for_each_hw_ctx(q
, hctx
, i
)
1533 if (blk_mq_hctx_stopped(hctx
))
1538 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1541 * This function is often used for pausing .queue_rq() by driver when
1542 * there isn't enough resource or some conditions aren't satisfied, and
1543 * BLK_STS_RESOURCE is usually returned.
1545 * We do not guarantee that dispatch can be drained or blocked
1546 * after blk_mq_stop_hw_queue() returns. Please use
1547 * blk_mq_quiesce_queue() for that requirement.
1549 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1551 cancel_delayed_work(&hctx
->run_work
);
1553 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1555 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1558 * This function is often used for pausing .queue_rq() by driver when
1559 * there isn't enough resource or some conditions aren't satisfied, and
1560 * BLK_STS_RESOURCE is usually returned.
1562 * We do not guarantee that dispatch can be drained or blocked
1563 * after blk_mq_stop_hw_queues() returns. Please use
1564 * blk_mq_quiesce_queue() for that requirement.
1566 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1568 struct blk_mq_hw_ctx
*hctx
;
1571 queue_for_each_hw_ctx(q
, hctx
, i
)
1572 blk_mq_stop_hw_queue(hctx
);
1574 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1576 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1578 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1580 blk_mq_run_hw_queue(hctx
, false);
1582 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1584 void blk_mq_start_hw_queues(struct request_queue
*q
)
1586 struct blk_mq_hw_ctx
*hctx
;
1589 queue_for_each_hw_ctx(q
, hctx
, i
)
1590 blk_mq_start_hw_queue(hctx
);
1592 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1594 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1596 if (!blk_mq_hctx_stopped(hctx
))
1599 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1600 blk_mq_run_hw_queue(hctx
, async
);
1602 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1604 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1606 struct blk_mq_hw_ctx
*hctx
;
1609 queue_for_each_hw_ctx(q
, hctx
, i
)
1610 blk_mq_start_stopped_hw_queue(hctx
, async
);
1612 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1614 static void blk_mq_run_work_fn(struct work_struct
*work
)
1616 struct blk_mq_hw_ctx
*hctx
;
1618 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
1621 * If we are stopped, don't run the queue.
1623 if (test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
1626 __blk_mq_run_hw_queue(hctx
);
1629 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1633 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1634 enum hctx_type type
= hctx
->type
;
1636 lockdep_assert_held(&ctx
->lock
);
1638 trace_block_rq_insert(hctx
->queue
, rq
);
1641 list_add(&rq
->queuelist
, &ctx
->rq_lists
[type
]);
1643 list_add_tail(&rq
->queuelist
, &ctx
->rq_lists
[type
]);
1646 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1649 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1651 lockdep_assert_held(&ctx
->lock
);
1653 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1654 blk_mq_hctx_mark_pending(hctx
, ctx
);
1658 * Should only be used carefully, when the caller knows we want to
1659 * bypass a potential IO scheduler on the target device.
1661 void blk_mq_request_bypass_insert(struct request
*rq
, bool run_queue
)
1663 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1665 spin_lock(&hctx
->lock
);
1666 list_add_tail(&rq
->queuelist
, &hctx
->dispatch
);
1667 spin_unlock(&hctx
->lock
);
1670 blk_mq_run_hw_queue(hctx
, false);
1673 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1674 struct list_head
*list
)
1678 enum hctx_type type
= hctx
->type
;
1681 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1684 list_for_each_entry(rq
, list
, queuelist
) {
1685 BUG_ON(rq
->mq_ctx
!= ctx
);
1686 trace_block_rq_insert(hctx
->queue
, rq
);
1689 spin_lock(&ctx
->lock
);
1690 list_splice_tail_init(list
, &ctx
->rq_lists
[type
]);
1691 blk_mq_hctx_mark_pending(hctx
, ctx
);
1692 spin_unlock(&ctx
->lock
);
1695 static int plug_rq_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1697 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1698 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1700 if (rqa
->mq_ctx
< rqb
->mq_ctx
)
1702 else if (rqa
->mq_ctx
> rqb
->mq_ctx
)
1704 else if (rqa
->mq_hctx
< rqb
->mq_hctx
)
1706 else if (rqa
->mq_hctx
> rqb
->mq_hctx
)
1709 return blk_rq_pos(rqa
) > blk_rq_pos(rqb
);
1712 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1714 struct blk_mq_hw_ctx
*this_hctx
;
1715 struct blk_mq_ctx
*this_ctx
;
1716 struct request_queue
*this_q
;
1722 list_splice_init(&plug
->mq_list
, &list
);
1724 if (plug
->rq_count
> 2 && plug
->multiple_queues
)
1725 list_sort(NULL
, &list
, plug_rq_cmp
);
1734 while (!list_empty(&list
)) {
1735 rq
= list_entry_rq(list
.next
);
1736 list_del_init(&rq
->queuelist
);
1738 if (rq
->mq_hctx
!= this_hctx
|| rq
->mq_ctx
!= this_ctx
) {
1740 trace_block_unplug(this_q
, depth
, !from_schedule
);
1741 blk_mq_sched_insert_requests(this_hctx
, this_ctx
,
1747 this_ctx
= rq
->mq_ctx
;
1748 this_hctx
= rq
->mq_hctx
;
1753 list_add_tail(&rq
->queuelist
, &rq_list
);
1757 * If 'this_hctx' is set, we know we have entries to complete
1758 * on 'rq_list'. Do those.
1761 trace_block_unplug(this_q
, depth
, !from_schedule
);
1762 blk_mq_sched_insert_requests(this_hctx
, this_ctx
, &rq_list
,
1767 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1769 blk_init_request_from_bio(rq
, bio
);
1771 blk_account_io_start(rq
, true);
1774 static blk_status_t
__blk_mq_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1776 blk_qc_t
*cookie
, bool last
)
1778 struct request_queue
*q
= rq
->q
;
1779 struct blk_mq_queue_data bd
= {
1783 blk_qc_t new_cookie
;
1786 new_cookie
= request_to_qc_t(hctx
, rq
);
1789 * For OK queue, we are done. For error, caller may kill it.
1790 * Any other error (busy), just add it to our list as we
1791 * previously would have done.
1793 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1796 blk_mq_update_dispatch_busy(hctx
, false);
1797 *cookie
= new_cookie
;
1799 case BLK_STS_RESOURCE
:
1800 case BLK_STS_DEV_RESOURCE
:
1801 blk_mq_update_dispatch_busy(hctx
, true);
1802 __blk_mq_requeue_request(rq
);
1805 blk_mq_update_dispatch_busy(hctx
, false);
1806 *cookie
= BLK_QC_T_NONE
;
1813 static blk_status_t
__blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1816 bool bypass_insert
, bool last
)
1818 struct request_queue
*q
= rq
->q
;
1819 bool run_queue
= true;
1822 * RCU or SRCU read lock is needed before checking quiesced flag.
1824 * When queue is stopped or quiesced, ignore 'bypass_insert' from
1825 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1826 * and avoid driver to try to dispatch again.
1828 if (blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(q
)) {
1830 bypass_insert
= false;
1834 if (q
->elevator
&& !bypass_insert
)
1837 if (!blk_mq_get_dispatch_budget(hctx
))
1840 if (!blk_mq_get_driver_tag(rq
)) {
1841 blk_mq_put_dispatch_budget(hctx
);
1845 return __blk_mq_issue_directly(hctx
, rq
, cookie
, last
);
1848 return BLK_STS_RESOURCE
;
1850 blk_mq_request_bypass_insert(rq
, run_queue
);
1854 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1855 struct request
*rq
, blk_qc_t
*cookie
)
1860 might_sleep_if(hctx
->flags
& BLK_MQ_F_BLOCKING
);
1862 hctx_lock(hctx
, &srcu_idx
);
1864 ret
= __blk_mq_try_issue_directly(hctx
, rq
, cookie
, false, true);
1865 if (ret
== BLK_STS_RESOURCE
|| ret
== BLK_STS_DEV_RESOURCE
)
1866 blk_mq_request_bypass_insert(rq
, true);
1867 else if (ret
!= BLK_STS_OK
)
1868 blk_mq_end_request(rq
, ret
);
1870 hctx_unlock(hctx
, srcu_idx
);
1873 blk_status_t
blk_mq_request_issue_directly(struct request
*rq
, bool last
)
1877 blk_qc_t unused_cookie
;
1878 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1880 hctx_lock(hctx
, &srcu_idx
);
1881 ret
= __blk_mq_try_issue_directly(hctx
, rq
, &unused_cookie
, true, last
);
1882 hctx_unlock(hctx
, srcu_idx
);
1887 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx
*hctx
,
1888 struct list_head
*list
)
1890 while (!list_empty(list
)) {
1892 struct request
*rq
= list_first_entry(list
, struct request
,
1895 list_del_init(&rq
->queuelist
);
1896 ret
= blk_mq_request_issue_directly(rq
, list_empty(list
));
1897 if (ret
!= BLK_STS_OK
) {
1898 if (ret
== BLK_STS_RESOURCE
||
1899 ret
== BLK_STS_DEV_RESOURCE
) {
1900 blk_mq_request_bypass_insert(rq
,
1904 blk_mq_end_request(rq
, ret
);
1909 * If we didn't flush the entire list, we could have told
1910 * the driver there was more coming, but that turned out to
1913 if (!list_empty(list
) && hctx
->queue
->mq_ops
->commit_rqs
)
1914 hctx
->queue
->mq_ops
->commit_rqs(hctx
);
1917 static void blk_add_rq_to_plug(struct blk_plug
*plug
, struct request
*rq
)
1919 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1921 if (!plug
->multiple_queues
&& !list_is_singular(&plug
->mq_list
)) {
1922 struct request
*tmp
;
1924 tmp
= list_first_entry(&plug
->mq_list
, struct request
,
1926 if (tmp
->q
!= rq
->q
)
1927 plug
->multiple_queues
= true;
1931 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1933 const int is_sync
= op_is_sync(bio
->bi_opf
);
1934 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1935 struct blk_mq_alloc_data data
= { .flags
= 0};
1937 struct blk_plug
*plug
;
1938 struct request
*same_queue_rq
= NULL
;
1941 blk_queue_bounce(q
, &bio
);
1943 blk_queue_split(q
, &bio
);
1945 if (!bio_integrity_prep(bio
))
1946 return BLK_QC_T_NONE
;
1948 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1949 blk_attempt_plug_merge(q
, bio
, &same_queue_rq
))
1950 return BLK_QC_T_NONE
;
1952 if (blk_mq_sched_bio_merge(q
, bio
))
1953 return BLK_QC_T_NONE
;
1955 rq_qos_throttle(q
, bio
);
1957 data
.cmd_flags
= bio
->bi_opf
;
1958 rq
= blk_mq_get_request(q
, bio
, &data
);
1959 if (unlikely(!rq
)) {
1960 rq_qos_cleanup(q
, bio
);
1961 if (bio
->bi_opf
& REQ_NOWAIT
)
1962 bio_wouldblock_error(bio
);
1963 return BLK_QC_T_NONE
;
1966 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1968 rq_qos_track(q
, rq
, bio
);
1970 cookie
= request_to_qc_t(data
.hctx
, rq
);
1972 plug
= current
->plug
;
1973 if (unlikely(is_flush_fua
)) {
1974 blk_mq_put_ctx(data
.ctx
);
1975 blk_mq_bio_to_request(rq
, bio
);
1977 /* bypass scheduler for flush rq */
1978 blk_insert_flush(rq
);
1979 blk_mq_run_hw_queue(data
.hctx
, true);
1980 } else if (plug
&& (q
->nr_hw_queues
== 1 || q
->mq_ops
->commit_rqs
)) {
1982 * Use plugging if we have a ->commit_rqs() hook as well, as
1983 * we know the driver uses bd->last in a smart fashion.
1985 unsigned int request_count
= plug
->rq_count
;
1986 struct request
*last
= NULL
;
1988 blk_mq_put_ctx(data
.ctx
);
1989 blk_mq_bio_to_request(rq
, bio
);
1992 trace_block_plug(q
);
1994 last
= list_entry_rq(plug
->mq_list
.prev
);
1996 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
1997 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1998 blk_flush_plug_list(plug
, false);
1999 trace_block_plug(q
);
2002 blk_add_rq_to_plug(plug
, rq
);
2003 } else if (plug
&& !blk_queue_nomerges(q
)) {
2004 blk_mq_bio_to_request(rq
, bio
);
2007 * We do limited plugging. If the bio can be merged, do that.
2008 * Otherwise the existing request in the plug list will be
2009 * issued. So the plug list will have one request at most
2010 * The plug list might get flushed before this. If that happens,
2011 * the plug list is empty, and same_queue_rq is invalid.
2013 if (list_empty(&plug
->mq_list
))
2014 same_queue_rq
= NULL
;
2015 if (same_queue_rq
) {
2016 list_del_init(&same_queue_rq
->queuelist
);
2019 blk_add_rq_to_plug(plug
, rq
);
2020 trace_block_plug(q
);
2022 blk_mq_put_ctx(data
.ctx
);
2024 if (same_queue_rq
) {
2025 data
.hctx
= same_queue_rq
->mq_hctx
;
2026 trace_block_unplug(q
, 1, true);
2027 blk_mq_try_issue_directly(data
.hctx
, same_queue_rq
,
2030 } else if ((q
->nr_hw_queues
> 1 && is_sync
) || (!q
->elevator
&&
2031 !data
.hctx
->dispatch_busy
)) {
2032 blk_mq_put_ctx(data
.ctx
);
2033 blk_mq_bio_to_request(rq
, bio
);
2034 blk_mq_try_issue_directly(data
.hctx
, rq
, &cookie
);
2036 blk_mq_put_ctx(data
.ctx
);
2037 blk_mq_bio_to_request(rq
, bio
);
2038 blk_mq_sched_insert_request(rq
, false, true, true);
2044 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
2045 unsigned int hctx_idx
)
2049 if (tags
->rqs
&& set
->ops
->exit_request
) {
2052 for (i
= 0; i
< tags
->nr_tags
; i
++) {
2053 struct request
*rq
= tags
->static_rqs
[i
];
2057 set
->ops
->exit_request(set
, rq
, hctx_idx
);
2058 tags
->static_rqs
[i
] = NULL
;
2062 while (!list_empty(&tags
->page_list
)) {
2063 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
2064 list_del_init(&page
->lru
);
2066 * Remove kmemleak object previously allocated in
2067 * blk_mq_alloc_rqs().
2069 kmemleak_free(page_address(page
));
2070 __free_pages(page
, page
->private);
2074 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
2078 kfree(tags
->static_rqs
);
2079 tags
->static_rqs
= NULL
;
2081 blk_mq_free_tags(tags
);
2084 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
2085 unsigned int hctx_idx
,
2086 unsigned int nr_tags
,
2087 unsigned int reserved_tags
)
2089 struct blk_mq_tags
*tags
;
2092 node
= blk_mq_hw_queue_to_node(&set
->map
[HCTX_TYPE_DEFAULT
], hctx_idx
);
2093 if (node
== NUMA_NO_NODE
)
2094 node
= set
->numa_node
;
2096 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
,
2097 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
2101 tags
->rqs
= kcalloc_node(nr_tags
, sizeof(struct request
*),
2102 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
2105 blk_mq_free_tags(tags
);
2109 tags
->static_rqs
= kcalloc_node(nr_tags
, sizeof(struct request
*),
2110 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
2112 if (!tags
->static_rqs
) {
2114 blk_mq_free_tags(tags
);
2121 static size_t order_to_size(unsigned int order
)
2123 return (size_t)PAGE_SIZE
<< order
;
2126 static int blk_mq_init_request(struct blk_mq_tag_set
*set
, struct request
*rq
,
2127 unsigned int hctx_idx
, int node
)
2131 if (set
->ops
->init_request
) {
2132 ret
= set
->ops
->init_request(set
, rq
, hctx_idx
, node
);
2137 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
2141 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
2142 unsigned int hctx_idx
, unsigned int depth
)
2144 unsigned int i
, j
, entries_per_page
, max_order
= 4;
2145 size_t rq_size
, left
;
2148 node
= blk_mq_hw_queue_to_node(&set
->map
[HCTX_TYPE_DEFAULT
], hctx_idx
);
2149 if (node
== NUMA_NO_NODE
)
2150 node
= set
->numa_node
;
2152 INIT_LIST_HEAD(&tags
->page_list
);
2155 * rq_size is the size of the request plus driver payload, rounded
2156 * to the cacheline size
2158 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
2160 left
= rq_size
* depth
;
2162 for (i
= 0; i
< depth
; ) {
2163 int this_order
= max_order
;
2168 while (this_order
&& left
< order_to_size(this_order
- 1))
2172 page
= alloc_pages_node(node
,
2173 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
2179 if (order_to_size(this_order
) < rq_size
)
2186 page
->private = this_order
;
2187 list_add_tail(&page
->lru
, &tags
->page_list
);
2189 p
= page_address(page
);
2191 * Allow kmemleak to scan these pages as they contain pointers
2192 * to additional allocations like via ops->init_request().
2194 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
2195 entries_per_page
= order_to_size(this_order
) / rq_size
;
2196 to_do
= min(entries_per_page
, depth
- i
);
2197 left
-= to_do
* rq_size
;
2198 for (j
= 0; j
< to_do
; j
++) {
2199 struct request
*rq
= p
;
2201 tags
->static_rqs
[i
] = rq
;
2202 if (blk_mq_init_request(set
, rq
, hctx_idx
, node
)) {
2203 tags
->static_rqs
[i
] = NULL
;
2214 blk_mq_free_rqs(set
, tags
, hctx_idx
);
2219 * 'cpu' is going away. splice any existing rq_list entries from this
2220 * software queue to the hw queue dispatch list, and ensure that it
2223 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
2225 struct blk_mq_hw_ctx
*hctx
;
2226 struct blk_mq_ctx
*ctx
;
2228 enum hctx_type type
;
2230 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
2231 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
2234 spin_lock(&ctx
->lock
);
2235 if (!list_empty(&ctx
->rq_lists
[type
])) {
2236 list_splice_init(&ctx
->rq_lists
[type
], &tmp
);
2237 blk_mq_hctx_clear_pending(hctx
, ctx
);
2239 spin_unlock(&ctx
->lock
);
2241 if (list_empty(&tmp
))
2244 spin_lock(&hctx
->lock
);
2245 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
2246 spin_unlock(&hctx
->lock
);
2248 blk_mq_run_hw_queue(hctx
, true);
2252 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
2254 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
2258 /* hctx->ctxs will be freed in queue's release handler */
2259 static void blk_mq_exit_hctx(struct request_queue
*q
,
2260 struct blk_mq_tag_set
*set
,
2261 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
2263 if (blk_mq_hw_queue_mapped(hctx
))
2264 blk_mq_tag_idle(hctx
);
2266 if (set
->ops
->exit_request
)
2267 set
->ops
->exit_request(set
, hctx
->fq
->flush_rq
, hctx_idx
);
2269 if (set
->ops
->exit_hctx
)
2270 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2272 blk_mq_remove_cpuhp(hctx
);
2274 spin_lock(&q
->unused_hctx_lock
);
2275 list_add(&hctx
->hctx_list
, &q
->unused_hctx_list
);
2276 spin_unlock(&q
->unused_hctx_lock
);
2279 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
2280 struct blk_mq_tag_set
*set
, int nr_queue
)
2282 struct blk_mq_hw_ctx
*hctx
;
2285 queue_for_each_hw_ctx(q
, hctx
, i
) {
2288 blk_mq_debugfs_unregister_hctx(hctx
);
2289 blk_mq_exit_hctx(q
, set
, hctx
, i
);
2293 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set
*tag_set
)
2295 int hw_ctx_size
= sizeof(struct blk_mq_hw_ctx
);
2297 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx
, srcu
),
2298 __alignof__(struct blk_mq_hw_ctx
)) !=
2299 sizeof(struct blk_mq_hw_ctx
));
2301 if (tag_set
->flags
& BLK_MQ_F_BLOCKING
)
2302 hw_ctx_size
+= sizeof(struct srcu_struct
);
2307 static int blk_mq_init_hctx(struct request_queue
*q
,
2308 struct blk_mq_tag_set
*set
,
2309 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
2311 hctx
->queue_num
= hctx_idx
;
2313 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
2315 hctx
->tags
= set
->tags
[hctx_idx
];
2317 if (set
->ops
->init_hctx
&&
2318 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
2319 goto unregister_cpu_notifier
;
2321 if (blk_mq_init_request(set
, hctx
->fq
->flush_rq
, hctx_idx
,
2327 if (set
->ops
->exit_hctx
)
2328 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2329 unregister_cpu_notifier
:
2330 blk_mq_remove_cpuhp(hctx
);
2334 static struct blk_mq_hw_ctx
*
2335 blk_mq_alloc_hctx(struct request_queue
*q
, struct blk_mq_tag_set
*set
,
2338 struct blk_mq_hw_ctx
*hctx
;
2339 gfp_t gfp
= GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
;
2341 hctx
= kzalloc_node(blk_mq_hw_ctx_size(set
), gfp
, node
);
2343 goto fail_alloc_hctx
;
2345 if (!zalloc_cpumask_var_node(&hctx
->cpumask
, gfp
, node
))
2348 atomic_set(&hctx
->nr_active
, 0);
2349 if (node
== NUMA_NO_NODE
)
2350 node
= set
->numa_node
;
2351 hctx
->numa_node
= node
;
2353 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
2354 spin_lock_init(&hctx
->lock
);
2355 INIT_LIST_HEAD(&hctx
->dispatch
);
2357 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
2359 INIT_LIST_HEAD(&hctx
->hctx_list
);
2362 * Allocate space for all possible cpus to avoid allocation at
2365 hctx
->ctxs
= kmalloc_array_node(nr_cpu_ids
, sizeof(void *),
2370 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8),
2375 spin_lock_init(&hctx
->dispatch_wait_lock
);
2376 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
2377 INIT_LIST_HEAD(&hctx
->dispatch_wait
.entry
);
2379 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
,
2384 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
2385 init_srcu_struct(hctx
->srcu
);
2386 blk_mq_hctx_kobj_init(hctx
);
2391 sbitmap_free(&hctx
->ctx_map
);
2395 free_cpumask_var(hctx
->cpumask
);
2402 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
2403 unsigned int nr_hw_queues
)
2405 struct blk_mq_tag_set
*set
= q
->tag_set
;
2408 for_each_possible_cpu(i
) {
2409 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2410 struct blk_mq_hw_ctx
*hctx
;
2414 spin_lock_init(&__ctx
->lock
);
2415 for (k
= HCTX_TYPE_DEFAULT
; k
< HCTX_MAX_TYPES
; k
++)
2416 INIT_LIST_HEAD(&__ctx
->rq_lists
[k
]);
2421 * Set local node, IFF we have more than one hw queue. If
2422 * not, we remain on the home node of the device
2424 for (j
= 0; j
< set
->nr_maps
; j
++) {
2425 hctx
= blk_mq_map_queue_type(q
, j
, i
);
2426 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
2427 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
2432 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
, int hctx_idx
)
2436 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
2437 set
->queue_depth
, set
->reserved_tags
);
2438 if (!set
->tags
[hctx_idx
])
2441 ret
= blk_mq_alloc_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
,
2446 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2447 set
->tags
[hctx_idx
] = NULL
;
2451 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set
*set
,
2452 unsigned int hctx_idx
)
2454 if (set
->tags
&& set
->tags
[hctx_idx
]) {
2455 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
2456 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2457 set
->tags
[hctx_idx
] = NULL
;
2461 static void blk_mq_map_swqueue(struct request_queue
*q
)
2463 unsigned int i
, j
, hctx_idx
;
2464 struct blk_mq_hw_ctx
*hctx
;
2465 struct blk_mq_ctx
*ctx
;
2466 struct blk_mq_tag_set
*set
= q
->tag_set
;
2469 * Avoid others reading imcomplete hctx->cpumask through sysfs
2471 mutex_lock(&q
->sysfs_lock
);
2473 queue_for_each_hw_ctx(q
, hctx
, i
) {
2474 cpumask_clear(hctx
->cpumask
);
2476 hctx
->dispatch_from
= NULL
;
2480 * Map software to hardware queues.
2482 * If the cpu isn't present, the cpu is mapped to first hctx.
2484 for_each_possible_cpu(i
) {
2485 hctx_idx
= set
->map
[HCTX_TYPE_DEFAULT
].mq_map
[i
];
2486 /* unmapped hw queue can be remapped after CPU topo changed */
2487 if (!set
->tags
[hctx_idx
] &&
2488 !__blk_mq_alloc_rq_map(set
, hctx_idx
)) {
2490 * If tags initialization fail for some hctx,
2491 * that hctx won't be brought online. In this
2492 * case, remap the current ctx to hctx[0] which
2493 * is guaranteed to always have tags allocated
2495 set
->map
[HCTX_TYPE_DEFAULT
].mq_map
[i
] = 0;
2498 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2499 for (j
= 0; j
< set
->nr_maps
; j
++) {
2500 if (!set
->map
[j
].nr_queues
) {
2501 ctx
->hctxs
[j
] = blk_mq_map_queue_type(q
,
2502 HCTX_TYPE_DEFAULT
, i
);
2506 hctx
= blk_mq_map_queue_type(q
, j
, i
);
2507 ctx
->hctxs
[j
] = hctx
;
2509 * If the CPU is already set in the mask, then we've
2510 * mapped this one already. This can happen if
2511 * devices share queues across queue maps.
2513 if (cpumask_test_cpu(i
, hctx
->cpumask
))
2516 cpumask_set_cpu(i
, hctx
->cpumask
);
2518 ctx
->index_hw
[hctx
->type
] = hctx
->nr_ctx
;
2519 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2522 * If the nr_ctx type overflows, we have exceeded the
2523 * amount of sw queues we can support.
2525 BUG_ON(!hctx
->nr_ctx
);
2528 for (; j
< HCTX_MAX_TYPES
; j
++)
2529 ctx
->hctxs
[j
] = blk_mq_map_queue_type(q
,
2530 HCTX_TYPE_DEFAULT
, i
);
2533 mutex_unlock(&q
->sysfs_lock
);
2535 queue_for_each_hw_ctx(q
, hctx
, i
) {
2537 * If no software queues are mapped to this hardware queue,
2538 * disable it and free the request entries.
2540 if (!hctx
->nr_ctx
) {
2541 /* Never unmap queue 0. We need it as a
2542 * fallback in case of a new remap fails
2545 if (i
&& set
->tags
[i
])
2546 blk_mq_free_map_and_requests(set
, i
);
2552 hctx
->tags
= set
->tags
[i
];
2553 WARN_ON(!hctx
->tags
);
2556 * Set the map size to the number of mapped software queues.
2557 * This is more accurate and more efficient than looping
2558 * over all possibly mapped software queues.
2560 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2563 * Initialize batch roundrobin counts
2565 hctx
->next_cpu
= blk_mq_first_mapped_cpu(hctx
);
2566 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2571 * Caller needs to ensure that we're either frozen/quiesced, or that
2572 * the queue isn't live yet.
2574 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2576 struct blk_mq_hw_ctx
*hctx
;
2579 queue_for_each_hw_ctx(q
, hctx
, i
) {
2581 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
2583 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2587 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
,
2590 struct request_queue
*q
;
2592 lockdep_assert_held(&set
->tag_list_lock
);
2594 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2595 blk_mq_freeze_queue(q
);
2596 queue_set_hctx_shared(q
, shared
);
2597 blk_mq_unfreeze_queue(q
);
2601 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
2603 struct blk_mq_tag_set
*set
= q
->tag_set
;
2605 mutex_lock(&set
->tag_list_lock
);
2606 list_del_rcu(&q
->tag_set_list
);
2607 if (list_is_singular(&set
->tag_list
)) {
2608 /* just transitioned to unshared */
2609 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2610 /* update existing queue */
2611 blk_mq_update_tag_set_depth(set
, false);
2613 mutex_unlock(&set
->tag_list_lock
);
2614 INIT_LIST_HEAD(&q
->tag_set_list
);
2617 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
2618 struct request_queue
*q
)
2620 mutex_lock(&set
->tag_list_lock
);
2623 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2625 if (!list_empty(&set
->tag_list
) &&
2626 !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
2627 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
2628 /* update existing queue */
2629 blk_mq_update_tag_set_depth(set
, true);
2631 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
2632 queue_set_hctx_shared(q
, true);
2633 list_add_tail_rcu(&q
->tag_set_list
, &set
->tag_list
);
2635 mutex_unlock(&set
->tag_list_lock
);
2638 /* All allocations will be freed in release handler of q->mq_kobj */
2639 static int blk_mq_alloc_ctxs(struct request_queue
*q
)
2641 struct blk_mq_ctxs
*ctxs
;
2644 ctxs
= kzalloc(sizeof(*ctxs
), GFP_KERNEL
);
2648 ctxs
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2649 if (!ctxs
->queue_ctx
)
2652 for_each_possible_cpu(cpu
) {
2653 struct blk_mq_ctx
*ctx
= per_cpu_ptr(ctxs
->queue_ctx
, cpu
);
2657 q
->mq_kobj
= &ctxs
->kobj
;
2658 q
->queue_ctx
= ctxs
->queue_ctx
;
2667 * It is the actual release handler for mq, but we do it from
2668 * request queue's release handler for avoiding use-after-free
2669 * and headache because q->mq_kobj shouldn't have been introduced,
2670 * but we can't group ctx/kctx kobj without it.
2672 void blk_mq_release(struct request_queue
*q
)
2674 struct blk_mq_hw_ctx
*hctx
, *next
;
2677 cancel_delayed_work_sync(&q
->requeue_work
);
2679 queue_for_each_hw_ctx(q
, hctx
, i
)
2680 WARN_ON_ONCE(hctx
&& list_empty(&hctx
->hctx_list
));
2682 /* all hctx are in .unused_hctx_list now */
2683 list_for_each_entry_safe(hctx
, next
, &q
->unused_hctx_list
, hctx_list
) {
2684 list_del_init(&hctx
->hctx_list
);
2685 kobject_put(&hctx
->kobj
);
2688 kfree(q
->queue_hw_ctx
);
2691 * release .mq_kobj and sw queue's kobject now because
2692 * both share lifetime with request queue.
2694 blk_mq_sysfs_deinit(q
);
2697 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2699 struct request_queue
*uninit_q
, *q
;
2701 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
2703 return ERR_PTR(-ENOMEM
);
2705 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
2707 blk_cleanup_queue(uninit_q
);
2711 EXPORT_SYMBOL(blk_mq_init_queue
);
2714 * Helper for setting up a queue with mq ops, given queue depth, and
2715 * the passed in mq ops flags.
2717 struct request_queue
*blk_mq_init_sq_queue(struct blk_mq_tag_set
*set
,
2718 const struct blk_mq_ops
*ops
,
2719 unsigned int queue_depth
,
2720 unsigned int set_flags
)
2722 struct request_queue
*q
;
2725 memset(set
, 0, sizeof(*set
));
2727 set
->nr_hw_queues
= 1;
2729 set
->queue_depth
= queue_depth
;
2730 set
->numa_node
= NUMA_NO_NODE
;
2731 set
->flags
= set_flags
;
2733 ret
= blk_mq_alloc_tag_set(set
);
2735 return ERR_PTR(ret
);
2737 q
= blk_mq_init_queue(set
);
2739 blk_mq_free_tag_set(set
);
2745 EXPORT_SYMBOL(blk_mq_init_sq_queue
);
2747 static struct blk_mq_hw_ctx
*blk_mq_alloc_and_init_hctx(
2748 struct blk_mq_tag_set
*set
, struct request_queue
*q
,
2749 int hctx_idx
, int node
)
2751 struct blk_mq_hw_ctx
*hctx
= NULL
, *tmp
;
2753 /* reuse dead hctx first */
2754 spin_lock(&q
->unused_hctx_lock
);
2755 list_for_each_entry(tmp
, &q
->unused_hctx_list
, hctx_list
) {
2756 if (tmp
->numa_node
== node
) {
2762 list_del_init(&hctx
->hctx_list
);
2763 spin_unlock(&q
->unused_hctx_lock
);
2766 hctx
= blk_mq_alloc_hctx(q
, set
, node
);
2770 if (blk_mq_init_hctx(q
, set
, hctx
, hctx_idx
))
2776 kobject_put(&hctx
->kobj
);
2781 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2782 struct request_queue
*q
)
2785 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2787 /* protect against switching io scheduler */
2788 mutex_lock(&q
->sysfs_lock
);
2789 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2791 struct blk_mq_hw_ctx
*hctx
;
2793 node
= blk_mq_hw_queue_to_node(&set
->map
[HCTX_TYPE_DEFAULT
], i
);
2795 * If the hw queue has been mapped to another numa node,
2796 * we need to realloc the hctx. If allocation fails, fallback
2797 * to use the previous one.
2799 if (hctxs
[i
] && (hctxs
[i
]->numa_node
== node
))
2802 hctx
= blk_mq_alloc_and_init_hctx(set
, q
, i
, node
);
2805 blk_mq_exit_hctx(q
, set
, hctxs
[i
], i
);
2809 pr_warn("Allocate new hctx on node %d fails,\
2810 fallback to previous one on node %d\n",
2811 node
, hctxs
[i
]->numa_node
);
2817 * Increasing nr_hw_queues fails. Free the newly allocated
2818 * hctxs and keep the previous q->nr_hw_queues.
2820 if (i
!= set
->nr_hw_queues
) {
2821 j
= q
->nr_hw_queues
;
2825 end
= q
->nr_hw_queues
;
2826 q
->nr_hw_queues
= set
->nr_hw_queues
;
2829 for (; j
< end
; j
++) {
2830 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2834 blk_mq_free_map_and_requests(set
, j
);
2835 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2839 mutex_unlock(&q
->sysfs_lock
);
2843 * Maximum number of hardware queues we support. For single sets, we'll never
2844 * have more than the CPUs (software queues). For multiple sets, the tag_set
2845 * user may have set ->nr_hw_queues larger.
2847 static unsigned int nr_hw_queues(struct blk_mq_tag_set
*set
)
2849 if (set
->nr_maps
== 1)
2852 return max(set
->nr_hw_queues
, nr_cpu_ids
);
2855 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2856 struct request_queue
*q
)
2858 /* mark the queue as mq asap */
2859 q
->mq_ops
= set
->ops
;
2861 q
->poll_cb
= blk_stat_alloc_callback(blk_mq_poll_stats_fn
,
2862 blk_mq_poll_stats_bkt
,
2863 BLK_MQ_POLL_STATS_BKTS
, q
);
2867 if (blk_mq_alloc_ctxs(q
))
2870 /* init q->mq_kobj and sw queues' kobjects */
2871 blk_mq_sysfs_init(q
);
2873 q
->nr_queues
= nr_hw_queues(set
);
2874 q
->queue_hw_ctx
= kcalloc_node(q
->nr_queues
, sizeof(*(q
->queue_hw_ctx
)),
2875 GFP_KERNEL
, set
->numa_node
);
2876 if (!q
->queue_hw_ctx
)
2879 INIT_LIST_HEAD(&q
->unused_hctx_list
);
2880 spin_lock_init(&q
->unused_hctx_lock
);
2882 blk_mq_realloc_hw_ctxs(set
, q
);
2883 if (!q
->nr_hw_queues
)
2886 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2887 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2891 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2892 if (set
->nr_maps
> HCTX_TYPE_POLL
&&
2893 set
->map
[HCTX_TYPE_POLL
].nr_queues
)
2894 blk_queue_flag_set(QUEUE_FLAG_POLL
, q
);
2896 q
->sg_reserved_size
= INT_MAX
;
2898 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2899 INIT_LIST_HEAD(&q
->requeue_list
);
2900 spin_lock_init(&q
->requeue_lock
);
2902 blk_queue_make_request(q
, blk_mq_make_request
);
2905 * Do this after blk_queue_make_request() overrides it...
2907 q
->nr_requests
= set
->queue_depth
;
2910 * Default to classic polling
2912 q
->poll_nsec
= BLK_MQ_POLL_CLASSIC
;
2914 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2915 blk_mq_add_queue_tag_set(set
, q
);
2916 blk_mq_map_swqueue(q
);
2918 if (!(set
->flags
& BLK_MQ_F_NO_SCHED
)) {
2921 ret
= elevator_init_mq(q
);
2923 return ERR_PTR(ret
);
2929 kfree(q
->queue_hw_ctx
);
2931 blk_mq_sysfs_deinit(q
);
2933 blk_stat_free_callback(q
->poll_cb
);
2937 return ERR_PTR(-ENOMEM
);
2939 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2941 /* tags can _not_ be used after returning from blk_mq_exit_queue */
2942 void blk_mq_exit_queue(struct request_queue
*q
)
2944 struct blk_mq_tag_set
*set
= q
->tag_set
;
2946 blk_mq_del_queue_tag_set(q
);
2947 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2950 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2954 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
2955 if (!__blk_mq_alloc_rq_map(set
, i
))
2962 blk_mq_free_rq_map(set
->tags
[i
]);
2968 * Allocate the request maps associated with this tag_set. Note that this
2969 * may reduce the depth asked for, if memory is tight. set->queue_depth
2970 * will be updated to reflect the allocated depth.
2972 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2977 depth
= set
->queue_depth
;
2979 err
= __blk_mq_alloc_rq_maps(set
);
2983 set
->queue_depth
>>= 1;
2984 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2988 } while (set
->queue_depth
);
2990 if (!set
->queue_depth
|| err
) {
2991 pr_err("blk-mq: failed to allocate request map\n");
2995 if (depth
!= set
->queue_depth
)
2996 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2997 depth
, set
->queue_depth
);
3002 static int blk_mq_update_queue_map(struct blk_mq_tag_set
*set
)
3004 if (set
->ops
->map_queues
&& !is_kdump_kernel()) {
3008 * transport .map_queues is usually done in the following
3011 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3012 * mask = get_cpu_mask(queue)
3013 * for_each_cpu(cpu, mask)
3014 * set->map[x].mq_map[cpu] = queue;
3017 * When we need to remap, the table has to be cleared for
3018 * killing stale mapping since one CPU may not be mapped
3021 for (i
= 0; i
< set
->nr_maps
; i
++)
3022 blk_mq_clear_mq_map(&set
->map
[i
]);
3024 return set
->ops
->map_queues(set
);
3026 BUG_ON(set
->nr_maps
> 1);
3027 return blk_mq_map_queues(&set
->map
[HCTX_TYPE_DEFAULT
]);
3032 * Alloc a tag set to be associated with one or more request queues.
3033 * May fail with EINVAL for various error conditions. May adjust the
3034 * requested depth down, if it's too large. In that case, the set
3035 * value will be stored in set->queue_depth.
3037 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
3041 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
3043 if (!set
->nr_hw_queues
)
3045 if (!set
->queue_depth
)
3047 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
3050 if (!set
->ops
->queue_rq
)
3053 if (!set
->ops
->get_budget
^ !set
->ops
->put_budget
)
3056 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
3057 pr_info("blk-mq: reduced tag depth to %u\n",
3059 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
3064 else if (set
->nr_maps
> HCTX_MAX_TYPES
)
3068 * If a crashdump is active, then we are potentially in a very
3069 * memory constrained environment. Limit us to 1 queue and
3070 * 64 tags to prevent using too much memory.
3072 if (is_kdump_kernel()) {
3073 set
->nr_hw_queues
= 1;
3075 set
->queue_depth
= min(64U, set
->queue_depth
);
3078 * There is no use for more h/w queues than cpus if we just have
3081 if (set
->nr_maps
== 1 && set
->nr_hw_queues
> nr_cpu_ids
)
3082 set
->nr_hw_queues
= nr_cpu_ids
;
3084 set
->tags
= kcalloc_node(nr_hw_queues(set
), sizeof(struct blk_mq_tags
*),
3085 GFP_KERNEL
, set
->numa_node
);
3090 for (i
= 0; i
< set
->nr_maps
; i
++) {
3091 set
->map
[i
].mq_map
= kcalloc_node(nr_cpu_ids
,
3092 sizeof(set
->map
[i
].mq_map
[0]),
3093 GFP_KERNEL
, set
->numa_node
);
3094 if (!set
->map
[i
].mq_map
)
3095 goto out_free_mq_map
;
3096 set
->map
[i
].nr_queues
= is_kdump_kernel() ? 1 : set
->nr_hw_queues
;
3099 ret
= blk_mq_update_queue_map(set
);
3101 goto out_free_mq_map
;
3103 ret
= blk_mq_alloc_rq_maps(set
);
3105 goto out_free_mq_map
;
3107 mutex_init(&set
->tag_list_lock
);
3108 INIT_LIST_HEAD(&set
->tag_list
);
3113 for (i
= 0; i
< set
->nr_maps
; i
++) {
3114 kfree(set
->map
[i
].mq_map
);
3115 set
->map
[i
].mq_map
= NULL
;
3121 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
3123 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
3127 for (i
= 0; i
< nr_hw_queues(set
); i
++)
3128 blk_mq_free_map_and_requests(set
, i
);
3130 for (j
= 0; j
< set
->nr_maps
; j
++) {
3131 kfree(set
->map
[j
].mq_map
);
3132 set
->map
[j
].mq_map
= NULL
;
3138 EXPORT_SYMBOL(blk_mq_free_tag_set
);
3140 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
3142 struct blk_mq_tag_set
*set
= q
->tag_set
;
3143 struct blk_mq_hw_ctx
*hctx
;
3149 if (q
->nr_requests
== nr
)
3152 blk_mq_freeze_queue(q
);
3153 blk_mq_quiesce_queue(q
);
3156 queue_for_each_hw_ctx(q
, hctx
, i
) {
3160 * If we're using an MQ scheduler, just update the scheduler
3161 * queue depth. This is similar to what the old code would do.
3163 if (!hctx
->sched_tags
) {
3164 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
, nr
,
3167 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
3172 if (q
->elevator
&& q
->elevator
->type
->ops
.depth_updated
)
3173 q
->elevator
->type
->ops
.depth_updated(hctx
);
3177 q
->nr_requests
= nr
;
3179 blk_mq_unquiesce_queue(q
);
3180 blk_mq_unfreeze_queue(q
);
3186 * request_queue and elevator_type pair.
3187 * It is just used by __blk_mq_update_nr_hw_queues to cache
3188 * the elevator_type associated with a request_queue.
3190 struct blk_mq_qe_pair
{
3191 struct list_head node
;
3192 struct request_queue
*q
;
3193 struct elevator_type
*type
;
3197 * Cache the elevator_type in qe pair list and switch the
3198 * io scheduler to 'none'
3200 static bool blk_mq_elv_switch_none(struct list_head
*head
,
3201 struct request_queue
*q
)
3203 struct blk_mq_qe_pair
*qe
;
3208 qe
= kmalloc(sizeof(*qe
), GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
);
3212 INIT_LIST_HEAD(&qe
->node
);
3214 qe
->type
= q
->elevator
->type
;
3215 list_add(&qe
->node
, head
);
3217 mutex_lock(&q
->sysfs_lock
);
3219 * After elevator_switch_mq, the previous elevator_queue will be
3220 * released by elevator_release. The reference of the io scheduler
3221 * module get by elevator_get will also be put. So we need to get
3222 * a reference of the io scheduler module here to prevent it to be
3225 __module_get(qe
->type
->elevator_owner
);
3226 elevator_switch_mq(q
, NULL
);
3227 mutex_unlock(&q
->sysfs_lock
);
3232 static void blk_mq_elv_switch_back(struct list_head
*head
,
3233 struct request_queue
*q
)
3235 struct blk_mq_qe_pair
*qe
;
3236 struct elevator_type
*t
= NULL
;
3238 list_for_each_entry(qe
, head
, node
)
3247 list_del(&qe
->node
);
3250 mutex_lock(&q
->sysfs_lock
);
3251 elevator_switch_mq(q
, t
);
3252 mutex_unlock(&q
->sysfs_lock
);
3255 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
,
3258 struct request_queue
*q
;
3260 int prev_nr_hw_queues
;
3262 lockdep_assert_held(&set
->tag_list_lock
);
3264 if (set
->nr_maps
== 1 && nr_hw_queues
> nr_cpu_ids
)
3265 nr_hw_queues
= nr_cpu_ids
;
3266 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
3269 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3270 blk_mq_freeze_queue(q
);
3272 * Sync with blk_mq_queue_tag_busy_iter.
3276 * Switch IO scheduler to 'none', cleaning up the data associated
3277 * with the previous scheduler. We will switch back once we are done
3278 * updating the new sw to hw queue mappings.
3280 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3281 if (!blk_mq_elv_switch_none(&head
, q
))
3284 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3285 blk_mq_debugfs_unregister_hctxs(q
);
3286 blk_mq_sysfs_unregister(q
);
3289 prev_nr_hw_queues
= set
->nr_hw_queues
;
3290 set
->nr_hw_queues
= nr_hw_queues
;
3291 blk_mq_update_queue_map(set
);
3293 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3294 blk_mq_realloc_hw_ctxs(set
, q
);
3295 if (q
->nr_hw_queues
!= set
->nr_hw_queues
) {
3296 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3297 nr_hw_queues
, prev_nr_hw_queues
);
3298 set
->nr_hw_queues
= prev_nr_hw_queues
;
3299 blk_mq_map_queues(&set
->map
[HCTX_TYPE_DEFAULT
]);
3302 blk_mq_map_swqueue(q
);
3305 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3306 blk_mq_sysfs_register(q
);
3307 blk_mq_debugfs_register_hctxs(q
);
3311 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3312 blk_mq_elv_switch_back(&head
, q
);
3314 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3315 blk_mq_unfreeze_queue(q
);
3318 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
3320 mutex_lock(&set
->tag_list_lock
);
3321 __blk_mq_update_nr_hw_queues(set
, nr_hw_queues
);
3322 mutex_unlock(&set
->tag_list_lock
);
3324 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
3326 /* Enable polling stats and return whether they were already enabled. */
3327 static bool blk_poll_stats_enable(struct request_queue
*q
)
3329 if (test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
3330 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS
, q
))
3332 blk_stat_add_callback(q
, q
->poll_cb
);
3336 static void blk_mq_poll_stats_start(struct request_queue
*q
)
3339 * We don't arm the callback if polling stats are not enabled or the
3340 * callback is already active.
3342 if (!test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
3343 blk_stat_is_active(q
->poll_cb
))
3346 blk_stat_activate_msecs(q
->poll_cb
, 100);
3349 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
)
3351 struct request_queue
*q
= cb
->data
;
3354 for (bucket
= 0; bucket
< BLK_MQ_POLL_STATS_BKTS
; bucket
++) {
3355 if (cb
->stat
[bucket
].nr_samples
)
3356 q
->poll_stat
[bucket
] = cb
->stat
[bucket
];
3360 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
3361 struct blk_mq_hw_ctx
*hctx
,
3364 unsigned long ret
= 0;
3368 * If stats collection isn't on, don't sleep but turn it on for
3371 if (!blk_poll_stats_enable(q
))
3375 * As an optimistic guess, use half of the mean service time
3376 * for this type of request. We can (and should) make this smarter.
3377 * For instance, if the completion latencies are tight, we can
3378 * get closer than just half the mean. This is especially
3379 * important on devices where the completion latencies are longer
3380 * than ~10 usec. We do use the stats for the relevant IO size
3381 * if available which does lead to better estimates.
3383 bucket
= blk_mq_poll_stats_bkt(rq
);
3387 if (q
->poll_stat
[bucket
].nr_samples
)
3388 ret
= (q
->poll_stat
[bucket
].mean
+ 1) / 2;
3393 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
3394 struct blk_mq_hw_ctx
*hctx
,
3397 struct hrtimer_sleeper hs
;
3398 enum hrtimer_mode mode
;
3402 if (rq
->rq_flags
& RQF_MQ_POLL_SLEPT
)
3406 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3408 * 0: use half of prev avg
3409 * >0: use this specific value
3411 if (q
->poll_nsec
> 0)
3412 nsecs
= q
->poll_nsec
;
3414 nsecs
= blk_mq_poll_nsecs(q
, hctx
, rq
);
3419 rq
->rq_flags
|= RQF_MQ_POLL_SLEPT
;
3422 * This will be replaced with the stats tracking code, using
3423 * 'avg_completion_time / 2' as the pre-sleep target.
3427 mode
= HRTIMER_MODE_REL
;
3428 hrtimer_init_on_stack(&hs
.timer
, CLOCK_MONOTONIC
, mode
);
3429 hrtimer_set_expires(&hs
.timer
, kt
);
3431 hrtimer_init_sleeper(&hs
, current
);
3433 if (blk_mq_rq_state(rq
) == MQ_RQ_COMPLETE
)
3435 set_current_state(TASK_UNINTERRUPTIBLE
);
3436 hrtimer_start_expires(&hs
.timer
, mode
);
3439 hrtimer_cancel(&hs
.timer
);
3440 mode
= HRTIMER_MODE_ABS
;
3441 } while (hs
.task
&& !signal_pending(current
));
3443 __set_current_state(TASK_RUNNING
);
3444 destroy_hrtimer_on_stack(&hs
.timer
);
3448 static bool blk_mq_poll_hybrid(struct request_queue
*q
,
3449 struct blk_mq_hw_ctx
*hctx
, blk_qc_t cookie
)
3453 if (q
->poll_nsec
== BLK_MQ_POLL_CLASSIC
)
3456 if (!blk_qc_t_is_internal(cookie
))
3457 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
3459 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
3461 * With scheduling, if the request has completed, we'll
3462 * get a NULL return here, as we clear the sched tag when
3463 * that happens. The request still remains valid, like always,
3464 * so we should be safe with just the NULL check.
3470 return blk_mq_poll_hybrid_sleep(q
, hctx
, rq
);
3474 * blk_poll - poll for IO completions
3476 * @cookie: cookie passed back at IO submission time
3477 * @spin: whether to spin for completions
3480 * Poll for completions on the passed in queue. Returns number of
3481 * completed entries found. If @spin is true, then blk_poll will continue
3482 * looping until at least one completion is found, unless the task is
3483 * otherwise marked running (or we need to reschedule).
3485 int blk_poll(struct request_queue
*q
, blk_qc_t cookie
, bool spin
)
3487 struct blk_mq_hw_ctx
*hctx
;
3490 if (!blk_qc_t_valid(cookie
) ||
3491 !test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
3495 blk_flush_plug_list(current
->plug
, false);
3497 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
3500 * If we sleep, have the caller restart the poll loop to reset
3501 * the state. Like for the other success return cases, the
3502 * caller is responsible for checking if the IO completed. If
3503 * the IO isn't complete, we'll get called again and will go
3504 * straight to the busy poll loop.
3506 if (blk_mq_poll_hybrid(q
, hctx
, cookie
))
3509 hctx
->poll_considered
++;
3511 state
= current
->state
;
3515 hctx
->poll_invoked
++;
3517 ret
= q
->mq_ops
->poll(hctx
);
3519 hctx
->poll_success
++;
3520 __set_current_state(TASK_RUNNING
);
3524 if (signal_pending_state(state
, current
))
3525 __set_current_state(TASK_RUNNING
);
3527 if (current
->state
== TASK_RUNNING
)
3529 if (ret
< 0 || !spin
)
3532 } while (!need_resched());
3534 __set_current_state(TASK_RUNNING
);
3537 EXPORT_SYMBOL_GPL(blk_poll
);
3539 unsigned int blk_mq_rq_cpu(struct request
*rq
)
3541 return rq
->mq_ctx
->cpu
;
3543 EXPORT_SYMBOL(blk_mq_rq_cpu
);
3545 static int __init
blk_mq_init(void)
3547 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
3548 blk_mq_hctx_notify_dead
);
3551 subsys_initcall(blk_mq_init
);