2 * Block multiqueue core code
4 * Copyright (C) 2013-2014 Jens Axboe
5 * Copyright (C) 2013-2014 Christoph Hellwig
7 #include <linux/kernel.h>
8 #include <linux/module.h>
9 #include <linux/backing-dev.h>
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
12 #include <linux/kmemleak.h>
14 #include <linux/init.h>
15 #include <linux/slab.h>
16 #include <linux/workqueue.h>
17 #include <linux/smp.h>
18 #include <linux/llist.h>
19 #include <linux/list_sort.h>
20 #include <linux/cpu.h>
21 #include <linux/cache.h>
22 #include <linux/sched/sysctl.h>
23 #include <linux/sched/topology.h>
24 #include <linux/sched/signal.h>
25 #include <linux/delay.h>
26 #include <linux/crash_dump.h>
27 #include <linux/prefetch.h>
29 #include <trace/events/block.h>
31 #include <linux/blk-mq.h>
34 #include "blk-mq-debugfs.h"
35 #include "blk-mq-tag.h"
38 #include "blk-mq-sched.h"
40 static DEFINE_MUTEX(all_q_mutex
);
41 static LIST_HEAD(all_q_list
);
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
);
45 static void __blk_mq_stop_hw_queues(struct request_queue
*q
, bool sync
);
47 static int blk_mq_poll_stats_bkt(const struct request
*rq
)
49 int ddir
, bytes
, bucket
;
51 ddir
= rq_data_dir(rq
);
52 bytes
= blk_rq_bytes(rq
);
54 bucket
= ddir
+ 2*(ilog2(bytes
) - 9);
58 else if (bucket
>= BLK_MQ_POLL_STATS_BKTS
)
59 return ddir
+ BLK_MQ_POLL_STATS_BKTS
- 2;
65 * Check if any of the ctx's have pending work in this hardware queue
67 bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
69 return sbitmap_any_bit_set(&hctx
->ctx_map
) ||
70 !list_empty_careful(&hctx
->dispatch
) ||
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 if (!sbitmap_test_bit(&hctx
->ctx_map
, ctx
->index_hw
))
81 sbitmap_set_bit(&hctx
->ctx_map
, ctx
->index_hw
);
84 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
85 struct blk_mq_ctx
*ctx
)
87 sbitmap_clear_bit(&hctx
->ctx_map
, ctx
->index_hw
);
90 void blk_freeze_queue_start(struct request_queue
*q
)
94 freeze_depth
= atomic_inc_return(&q
->mq_freeze_depth
);
95 if (freeze_depth
== 1) {
96 percpu_ref_kill(&q
->q_usage_counter
);
97 blk_mq_run_hw_queues(q
, false);
100 EXPORT_SYMBOL_GPL(blk_freeze_queue_start
);
102 void blk_mq_freeze_queue_wait(struct request_queue
*q
)
104 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
106 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait
);
108 int blk_mq_freeze_queue_wait_timeout(struct request_queue
*q
,
109 unsigned long timeout
)
111 return wait_event_timeout(q
->mq_freeze_wq
,
112 percpu_ref_is_zero(&q
->q_usage_counter
),
115 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout
);
118 * Guarantee no request is in use, so we can change any data structure of
119 * the queue afterward.
121 void blk_freeze_queue(struct request_queue
*q
)
124 * In the !blk_mq case we are only calling this to kill the
125 * q_usage_counter, otherwise this increases the freeze depth
126 * and waits for it to return to zero. For this reason there is
127 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
128 * exported to drivers as the only user for unfreeze is blk_mq.
130 blk_freeze_queue_start(q
);
131 blk_mq_freeze_queue_wait(q
);
134 void blk_mq_freeze_queue(struct request_queue
*q
)
137 * ...just an alias to keep freeze and unfreeze actions balanced
138 * in the blk_mq_* namespace
142 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
144 void blk_mq_unfreeze_queue(struct request_queue
*q
)
148 freeze_depth
= atomic_dec_return(&q
->mq_freeze_depth
);
149 WARN_ON_ONCE(freeze_depth
< 0);
151 percpu_ref_reinit(&q
->q_usage_counter
);
152 wake_up_all(&q
->mq_freeze_wq
);
155 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
158 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
161 * Note: this function does not prevent that the struct request end_io()
162 * callback function is invoked. Once this function is returned, we make
163 * sure no dispatch can happen until the queue is unquiesced via
164 * blk_mq_unquiesce_queue().
166 void blk_mq_quiesce_queue(struct request_queue
*q
)
168 struct blk_mq_hw_ctx
*hctx
;
172 blk_mq_quiesce_queue_nowait(q
);
174 queue_for_each_hw_ctx(q
, hctx
, i
) {
175 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
176 synchronize_srcu(&hctx
->queue_rq_srcu
);
183 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
186 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
189 * This function recovers queue into the state before quiescing
190 * which is done by blk_mq_quiesce_queue.
192 void blk_mq_unquiesce_queue(struct request_queue
*q
)
194 spin_lock_irq(q
->queue_lock
);
195 queue_flag_clear(QUEUE_FLAG_QUIESCED
, q
);
196 spin_unlock_irq(q
->queue_lock
);
198 /* dispatch requests which are inserted during quiescing */
199 blk_mq_run_hw_queues(q
, true);
201 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue
);
203 void blk_mq_wake_waiters(struct request_queue
*q
)
205 struct blk_mq_hw_ctx
*hctx
;
208 queue_for_each_hw_ctx(q
, hctx
, i
)
209 if (blk_mq_hw_queue_mapped(hctx
))
210 blk_mq_tag_wakeup_all(hctx
->tags
, true);
213 * If we are called because the queue has now been marked as
214 * dying, we need to ensure that processes currently waiting on
215 * the queue are notified as well.
217 wake_up_all(&q
->mq_freeze_wq
);
220 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
222 return blk_mq_has_free_tags(hctx
->tags
);
224 EXPORT_SYMBOL(blk_mq_can_queue
);
226 static struct request
*blk_mq_rq_ctx_init(struct blk_mq_alloc_data
*data
,
227 unsigned int tag
, unsigned int op
)
229 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
230 struct request
*rq
= tags
->static_rqs
[tag
];
232 if (data
->flags
& BLK_MQ_REQ_INTERNAL
) {
234 rq
->internal_tag
= tag
;
236 if (blk_mq_tag_busy(data
->hctx
)) {
237 rq
->rq_flags
= RQF_MQ_INFLIGHT
;
238 atomic_inc(&data
->hctx
->nr_active
);
241 rq
->internal_tag
= -1;
242 data
->hctx
->tags
->rqs
[rq
->tag
] = rq
;
245 INIT_LIST_HEAD(&rq
->queuelist
);
246 /* csd/requeue_work/fifo_time is initialized before use */
248 rq
->mq_ctx
= data
->ctx
;
250 if (blk_queue_io_stat(data
->q
))
251 rq
->rq_flags
|= RQF_IO_STAT
;
252 /* do not touch atomic flags, it needs atomic ops against the timer */
254 INIT_HLIST_NODE(&rq
->hash
);
255 RB_CLEAR_NODE(&rq
->rb_node
);
258 rq
->start_time
= jiffies
;
259 #ifdef CONFIG_BLK_CGROUP
261 set_start_time_ns(rq
);
262 rq
->io_start_time_ns
= 0;
264 rq
->nr_phys_segments
= 0;
265 #if defined(CONFIG_BLK_DEV_INTEGRITY)
266 rq
->nr_integrity_segments
= 0;
269 /* tag was already set */
272 INIT_LIST_HEAD(&rq
->timeout_list
);
276 rq
->end_io_data
= NULL
;
279 data
->ctx
->rq_dispatched
[op_is_sync(op
)]++;
283 static struct request
*blk_mq_get_request(struct request_queue
*q
,
284 struct bio
*bio
, unsigned int op
,
285 struct blk_mq_alloc_data
*data
)
287 struct elevator_queue
*e
= q
->elevator
;
291 blk_queue_enter_live(q
);
293 if (likely(!data
->ctx
))
294 data
->ctx
= blk_mq_get_ctx(q
);
295 if (likely(!data
->hctx
))
296 data
->hctx
= blk_mq_map_queue(q
, data
->ctx
->cpu
);
299 data
->flags
|= BLK_MQ_REQ_INTERNAL
;
302 * Flush requests are special and go directly to the
305 if (!op_is_flush(op
) && e
->type
->ops
.mq
.limit_depth
)
306 e
->type
->ops
.mq
.limit_depth(op
, data
);
309 tag
= blk_mq_get_tag(data
);
310 if (tag
== BLK_MQ_TAG_FAIL
) {
315 rq
= blk_mq_rq_ctx_init(data
, tag
, op
);
316 if (!op_is_flush(op
)) {
318 if (e
&& e
->type
->ops
.mq
.prepare_request
) {
319 if (e
->type
->icq_cache
&& rq_ioc(bio
))
320 blk_mq_sched_assign_ioc(rq
, bio
);
322 e
->type
->ops
.mq
.prepare_request(rq
, bio
);
323 rq
->rq_flags
|= RQF_ELVPRIV
;
326 data
->hctx
->queued
++;
330 struct request
*blk_mq_alloc_request(struct request_queue
*q
, int rw
,
333 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
337 ret
= blk_queue_enter(q
, flags
& BLK_MQ_REQ_NOWAIT
);
341 rq
= blk_mq_get_request(q
, NULL
, rw
, &alloc_data
);
343 blk_mq_put_ctx(alloc_data
.ctx
);
347 return ERR_PTR(-EWOULDBLOCK
);
350 rq
->__sector
= (sector_t
) -1;
351 rq
->bio
= rq
->biotail
= NULL
;
354 EXPORT_SYMBOL(blk_mq_alloc_request
);
356 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
, int rw
,
357 unsigned int flags
, unsigned int hctx_idx
)
359 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
365 * If the tag allocator sleeps we could get an allocation for a
366 * different hardware context. No need to complicate the low level
367 * allocator for this for the rare use case of a command tied to
370 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
371 return ERR_PTR(-EINVAL
);
373 if (hctx_idx
>= q
->nr_hw_queues
)
374 return ERR_PTR(-EIO
);
376 ret
= blk_queue_enter(q
, true);
381 * Check if the hardware context is actually mapped to anything.
382 * If not tell the caller that it should skip this queue.
384 alloc_data
.hctx
= q
->queue_hw_ctx
[hctx_idx
];
385 if (!blk_mq_hw_queue_mapped(alloc_data
.hctx
)) {
387 return ERR_PTR(-EXDEV
);
389 cpu
= cpumask_first(alloc_data
.hctx
->cpumask
);
390 alloc_data
.ctx
= __blk_mq_get_ctx(q
, cpu
);
392 rq
= blk_mq_get_request(q
, NULL
, rw
, &alloc_data
);
397 return ERR_PTR(-EWOULDBLOCK
);
401 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
403 void blk_mq_free_request(struct request
*rq
)
405 struct request_queue
*q
= rq
->q
;
406 struct elevator_queue
*e
= q
->elevator
;
407 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
408 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
409 const int sched_tag
= rq
->internal_tag
;
411 if (rq
->rq_flags
& RQF_ELVPRIV
) {
412 if (e
&& e
->type
->ops
.mq
.finish_request
)
413 e
->type
->ops
.mq
.finish_request(rq
);
415 put_io_context(rq
->elv
.icq
->ioc
);
420 ctx
->rq_completed
[rq_is_sync(rq
)]++;
421 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
422 atomic_dec(&hctx
->nr_active
);
424 wbt_done(q
->rq_wb
, &rq
->issue_stat
);
427 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
428 clear_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
430 blk_mq_put_tag(hctx
, hctx
->tags
, ctx
, rq
->tag
);
432 blk_mq_put_tag(hctx
, hctx
->sched_tags
, ctx
, sched_tag
);
433 blk_mq_sched_restart(hctx
);
436 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
438 inline void __blk_mq_end_request(struct request
*rq
, blk_status_t error
)
440 blk_account_io_done(rq
);
443 wbt_done(rq
->q
->rq_wb
, &rq
->issue_stat
);
444 rq
->end_io(rq
, error
);
446 if (unlikely(blk_bidi_rq(rq
)))
447 blk_mq_free_request(rq
->next_rq
);
448 blk_mq_free_request(rq
);
451 EXPORT_SYMBOL(__blk_mq_end_request
);
453 void blk_mq_end_request(struct request
*rq
, blk_status_t error
)
455 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
457 __blk_mq_end_request(rq
, error
);
459 EXPORT_SYMBOL(blk_mq_end_request
);
461 static void __blk_mq_complete_request_remote(void *data
)
463 struct request
*rq
= data
;
465 rq
->q
->softirq_done_fn(rq
);
468 static void __blk_mq_complete_request(struct request
*rq
)
470 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
474 if (rq
->internal_tag
!= -1)
475 blk_mq_sched_completed_request(rq
);
476 if (rq
->rq_flags
& RQF_STATS
) {
477 blk_mq_poll_stats_start(rq
->q
);
481 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
482 rq
->q
->softirq_done_fn(rq
);
487 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
488 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
490 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
491 rq
->csd
.func
= __blk_mq_complete_request_remote
;
494 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
496 rq
->q
->softirq_done_fn(rq
);
502 * blk_mq_complete_request - end I/O on a request
503 * @rq: the request being processed
506 * Ends all I/O on a request. It does not handle partial completions.
507 * The actual completion happens out-of-order, through a IPI handler.
509 void blk_mq_complete_request(struct request
*rq
)
511 struct request_queue
*q
= rq
->q
;
513 if (unlikely(blk_should_fake_timeout(q
)))
515 if (!blk_mark_rq_complete(rq
))
516 __blk_mq_complete_request(rq
);
518 EXPORT_SYMBOL(blk_mq_complete_request
);
520 int blk_mq_request_started(struct request
*rq
)
522 return test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
524 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
526 void blk_mq_start_request(struct request
*rq
)
528 struct request_queue
*q
= rq
->q
;
530 blk_mq_sched_started_request(rq
);
532 trace_block_rq_issue(q
, rq
);
534 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
535 blk_stat_set_issue(&rq
->issue_stat
, blk_rq_sectors(rq
));
536 rq
->rq_flags
|= RQF_STATS
;
537 wbt_issue(q
->rq_wb
, &rq
->issue_stat
);
543 * Ensure that ->deadline is visible before set the started
544 * flag and clear the completed flag.
546 smp_mb__before_atomic();
549 * Mark us as started and clear complete. Complete might have been
550 * set if requeue raced with timeout, which then marked it as
551 * complete. So be sure to clear complete again when we start
552 * the request, otherwise we'll ignore the completion event.
554 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
555 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
556 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
557 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
559 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
561 * Make sure space for the drain appears. We know we can do
562 * this because max_hw_segments has been adjusted to be one
563 * fewer than the device can handle.
565 rq
->nr_phys_segments
++;
568 EXPORT_SYMBOL(blk_mq_start_request
);
571 * When we reach here because queue is busy, REQ_ATOM_COMPLETE
572 * flag isn't set yet, so there may be race with timeout handler,
573 * but given rq->deadline is just set in .queue_rq() under
574 * this situation, the race won't be possible in reality because
575 * rq->timeout should be set as big enough to cover the window
576 * between blk_mq_start_request() called from .queue_rq() and
577 * clearing REQ_ATOM_STARTED here.
579 static void __blk_mq_requeue_request(struct request
*rq
)
581 struct request_queue
*q
= rq
->q
;
583 trace_block_rq_requeue(q
, rq
);
584 wbt_requeue(q
->rq_wb
, &rq
->issue_stat
);
585 blk_mq_sched_requeue_request(rq
);
587 if (test_and_clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
588 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
589 rq
->nr_phys_segments
--;
593 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
595 __blk_mq_requeue_request(rq
);
597 BUG_ON(blk_queued_rq(rq
));
598 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
600 EXPORT_SYMBOL(blk_mq_requeue_request
);
602 static void blk_mq_requeue_work(struct work_struct
*work
)
604 struct request_queue
*q
=
605 container_of(work
, struct request_queue
, requeue_work
.work
);
607 struct request
*rq
, *next
;
610 spin_lock_irqsave(&q
->requeue_lock
, flags
);
611 list_splice_init(&q
->requeue_list
, &rq_list
);
612 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
614 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
615 if (!(rq
->rq_flags
& RQF_SOFTBARRIER
))
618 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
619 list_del_init(&rq
->queuelist
);
620 blk_mq_sched_insert_request(rq
, true, false, false, true);
623 while (!list_empty(&rq_list
)) {
624 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
625 list_del_init(&rq
->queuelist
);
626 blk_mq_sched_insert_request(rq
, false, false, false, true);
629 blk_mq_run_hw_queues(q
, false);
632 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
633 bool kick_requeue_list
)
635 struct request_queue
*q
= rq
->q
;
639 * We abuse this flag that is otherwise used by the I/O scheduler to
640 * request head insertation from the workqueue.
642 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
644 spin_lock_irqsave(&q
->requeue_lock
, flags
);
646 rq
->rq_flags
|= RQF_SOFTBARRIER
;
647 list_add(&rq
->queuelist
, &q
->requeue_list
);
649 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
651 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
653 if (kick_requeue_list
)
654 blk_mq_kick_requeue_list(q
);
656 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
658 void blk_mq_kick_requeue_list(struct request_queue
*q
)
660 kblockd_schedule_delayed_work(&q
->requeue_work
, 0);
662 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
664 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
667 kblockd_schedule_delayed_work(&q
->requeue_work
,
668 msecs_to_jiffies(msecs
));
670 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
672 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
674 if (tag
< tags
->nr_tags
) {
675 prefetch(tags
->rqs
[tag
]);
676 return tags
->rqs
[tag
];
681 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
683 struct blk_mq_timeout_data
{
685 unsigned int next_set
;
688 void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
690 const struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
691 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
694 * We know that complete is set at this point. If STARTED isn't set
695 * anymore, then the request isn't active and the "timeout" should
696 * just be ignored. This can happen due to the bitflag ordering.
697 * Timeout first checks if STARTED is set, and if it is, assumes
698 * the request is active. But if we race with completion, then
699 * both flags will get cleared. So check here again, and ignore
700 * a timeout event with a request that isn't active.
702 if (!test_bit(REQ_ATOM_STARTED
, &req
->atomic_flags
))
706 ret
= ops
->timeout(req
, reserved
);
710 __blk_mq_complete_request(req
);
712 case BLK_EH_RESET_TIMER
:
714 blk_clear_rq_complete(req
);
716 case BLK_EH_NOT_HANDLED
:
719 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
724 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
725 struct request
*rq
, void *priv
, bool reserved
)
727 struct blk_mq_timeout_data
*data
= priv
;
729 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
733 * The rq being checked may have been freed and reallocated
734 * out already here, we avoid this race by checking rq->deadline
735 * and REQ_ATOM_COMPLETE flag together:
737 * - if rq->deadline is observed as new value because of
738 * reusing, the rq won't be timed out because of timing.
739 * - if rq->deadline is observed as previous value,
740 * REQ_ATOM_COMPLETE flag won't be cleared in reuse path
741 * because we put a barrier between setting rq->deadline
742 * and clearing the flag in blk_mq_start_request(), so
743 * this rq won't be timed out too.
745 if (time_after_eq(jiffies
, rq
->deadline
)) {
746 if (!blk_mark_rq_complete(rq
))
747 blk_mq_rq_timed_out(rq
, reserved
);
748 } else if (!data
->next_set
|| time_after(data
->next
, rq
->deadline
)) {
749 data
->next
= rq
->deadline
;
754 static void blk_mq_timeout_work(struct work_struct
*work
)
756 struct request_queue
*q
=
757 container_of(work
, struct request_queue
, timeout_work
);
758 struct blk_mq_timeout_data data
= {
764 /* A deadlock might occur if a request is stuck requiring a
765 * timeout at the same time a queue freeze is waiting
766 * completion, since the timeout code would not be able to
767 * acquire the queue reference here.
769 * That's why we don't use blk_queue_enter here; instead, we use
770 * percpu_ref_tryget directly, because we need to be able to
771 * obtain a reference even in the short window between the queue
772 * starting to freeze, by dropping the first reference in
773 * blk_freeze_queue_start, and the moment the last request is
774 * consumed, marked by the instant q_usage_counter reaches
777 if (!percpu_ref_tryget(&q
->q_usage_counter
))
780 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &data
);
783 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
784 mod_timer(&q
->timeout
, data
.next
);
786 struct blk_mq_hw_ctx
*hctx
;
788 queue_for_each_hw_ctx(q
, hctx
, i
) {
789 /* the hctx may be unmapped, so check it here */
790 if (blk_mq_hw_queue_mapped(hctx
))
791 blk_mq_tag_idle(hctx
);
797 struct flush_busy_ctx_data
{
798 struct blk_mq_hw_ctx
*hctx
;
799 struct list_head
*list
;
802 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
804 struct flush_busy_ctx_data
*flush_data
= data
;
805 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
806 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
808 sbitmap_clear_bit(sb
, bitnr
);
809 spin_lock(&ctx
->lock
);
810 list_splice_tail_init(&ctx
->rq_list
, flush_data
->list
);
811 spin_unlock(&ctx
->lock
);
816 * Process software queues that have been marked busy, splicing them
817 * to the for-dispatch
819 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
821 struct flush_busy_ctx_data data
= {
826 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
828 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
830 static inline unsigned int queued_to_index(unsigned int queued
)
835 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
838 bool blk_mq_get_driver_tag(struct request
*rq
, struct blk_mq_hw_ctx
**hctx
,
841 struct blk_mq_alloc_data data
= {
843 .hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
),
844 .flags
= wait
? 0 : BLK_MQ_REQ_NOWAIT
,
847 might_sleep_if(wait
);
852 if (blk_mq_tag_is_reserved(data
.hctx
->sched_tags
, rq
->internal_tag
))
853 data
.flags
|= BLK_MQ_REQ_RESERVED
;
855 rq
->tag
= blk_mq_get_tag(&data
);
857 if (blk_mq_tag_busy(data
.hctx
)) {
858 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
859 atomic_inc(&data
.hctx
->nr_active
);
861 data
.hctx
->tags
->rqs
[rq
->tag
] = rq
;
867 return rq
->tag
!= -1;
870 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx
*hctx
,
873 blk_mq_put_tag(hctx
, hctx
->tags
, rq
->mq_ctx
, rq
->tag
);
876 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
) {
877 rq
->rq_flags
&= ~RQF_MQ_INFLIGHT
;
878 atomic_dec(&hctx
->nr_active
);
882 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx
*hctx
,
885 if (rq
->tag
== -1 || rq
->internal_tag
== -1)
888 __blk_mq_put_driver_tag(hctx
, rq
);
891 static void blk_mq_put_driver_tag(struct request
*rq
)
893 struct blk_mq_hw_ctx
*hctx
;
895 if (rq
->tag
== -1 || rq
->internal_tag
== -1)
898 hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
);
899 __blk_mq_put_driver_tag(hctx
, rq
);
903 * If we fail getting a driver tag because all the driver tags are already
904 * assigned and on the dispatch list, BUT the first entry does not have a
905 * tag, then we could deadlock. For that case, move entries with assigned
906 * driver tags to the front, leaving the set of tagged requests in the
907 * same order, and the untagged set in the same order.
909 static bool reorder_tags_to_front(struct list_head
*list
)
911 struct request
*rq
, *tmp
, *first
= NULL
;
913 list_for_each_entry_safe_reverse(rq
, tmp
, list
, queuelist
) {
917 list_move(&rq
->queuelist
, list
);
923 return first
!= NULL
;
926 static int blk_mq_dispatch_wake(wait_queue_t
*wait
, unsigned mode
, int flags
,
929 struct blk_mq_hw_ctx
*hctx
;
931 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
933 list_del(&wait
->task_list
);
934 clear_bit_unlock(BLK_MQ_S_TAG_WAITING
, &hctx
->state
);
935 blk_mq_run_hw_queue(hctx
, true);
939 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx
*hctx
)
941 struct sbq_wait_state
*ws
;
944 * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
945 * The thread which wins the race to grab this bit adds the hardware
946 * queue to the wait queue.
948 if (test_bit(BLK_MQ_S_TAG_WAITING
, &hctx
->state
) ||
949 test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING
, &hctx
->state
))
952 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
953 ws
= bt_wait_ptr(&hctx
->tags
->bitmap_tags
, hctx
);
956 * As soon as this returns, it's no longer safe to fiddle with
957 * hctx->dispatch_wait, since a completion can wake up the wait queue
958 * and unlock the bit.
960 add_wait_queue(&ws
->wait
, &hctx
->dispatch_wait
);
964 bool blk_mq_dispatch_rq_list(struct request_queue
*q
, struct list_head
*list
)
966 struct blk_mq_hw_ctx
*hctx
;
970 if (list_empty(list
))
974 * Now process all the entries, sending them to the driver.
978 struct blk_mq_queue_data bd
;
981 rq
= list_first_entry(list
, struct request
, queuelist
);
982 if (!blk_mq_get_driver_tag(rq
, &hctx
, false)) {
983 if (!queued
&& reorder_tags_to_front(list
))
987 * The initial allocation attempt failed, so we need to
988 * rerun the hardware queue when a tag is freed.
990 if (!blk_mq_dispatch_wait_add(hctx
))
994 * It's possible that a tag was freed in the window
995 * between the allocation failure and adding the
996 * hardware queue to the wait queue.
998 if (!blk_mq_get_driver_tag(rq
, &hctx
, false))
1002 list_del_init(&rq
->queuelist
);
1007 * Flag last if we have no more requests, or if we have more
1008 * but can't assign a driver tag to it.
1010 if (list_empty(list
))
1013 struct request
*nxt
;
1015 nxt
= list_first_entry(list
, struct request
, queuelist
);
1016 bd
.last
= !blk_mq_get_driver_tag(nxt
, NULL
, false);
1019 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1020 if (ret
== BLK_STS_RESOURCE
) {
1021 blk_mq_put_driver_tag_hctx(hctx
, rq
);
1022 list_add(&rq
->queuelist
, list
);
1023 __blk_mq_requeue_request(rq
);
1027 if (unlikely(ret
!= BLK_STS_OK
)) {
1029 blk_mq_end_request(rq
, BLK_STS_IOERR
);
1034 } while (!list_empty(list
));
1036 hctx
->dispatched
[queued_to_index(queued
)]++;
1039 * Any items that need requeuing? Stuff them into hctx->dispatch,
1040 * that is where we will continue on next queue run.
1042 if (!list_empty(list
)) {
1044 * If an I/O scheduler has been configured and we got a driver
1045 * tag for the next request already, free it again.
1047 rq
= list_first_entry(list
, struct request
, queuelist
);
1048 blk_mq_put_driver_tag(rq
);
1050 spin_lock(&hctx
->lock
);
1051 list_splice_init(list
, &hctx
->dispatch
);
1052 spin_unlock(&hctx
->lock
);
1055 * If SCHED_RESTART was set by the caller of this function and
1056 * it is no longer set that means that it was cleared by another
1057 * thread and hence that a queue rerun is needed.
1059 * If TAG_WAITING is set that means that an I/O scheduler has
1060 * been configured and another thread is waiting for a driver
1061 * tag. To guarantee fairness, do not rerun this hardware queue
1062 * but let the other thread grab the driver tag.
1064 * If no I/O scheduler has been configured it is possible that
1065 * the hardware queue got stopped and restarted before requests
1066 * were pushed back onto the dispatch list. Rerun the queue to
1067 * avoid starvation. Notes:
1068 * - blk_mq_run_hw_queue() checks whether or not a queue has
1069 * been stopped before rerunning a queue.
1070 * - Some but not all block drivers stop a queue before
1071 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1074 if (!blk_mq_sched_needs_restart(hctx
) &&
1075 !test_bit(BLK_MQ_S_TAG_WAITING
, &hctx
->state
))
1076 blk_mq_run_hw_queue(hctx
, true);
1079 return (queued
+ errors
) != 0;
1082 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1086 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
1087 cpu_online(hctx
->next_cpu
));
1089 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1091 blk_mq_sched_dispatch_requests(hctx
);
1096 srcu_idx
= srcu_read_lock(&hctx
->queue_rq_srcu
);
1097 blk_mq_sched_dispatch_requests(hctx
);
1098 srcu_read_unlock(&hctx
->queue_rq_srcu
, srcu_idx
);
1103 * It'd be great if the workqueue API had a way to pass
1104 * in a mask and had some smarts for more clever placement.
1105 * For now we just round-robin here, switching for every
1106 * BLK_MQ_CPU_WORK_BATCH queued items.
1108 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1110 if (hctx
->queue
->nr_hw_queues
== 1)
1111 return WORK_CPU_UNBOUND
;
1113 if (--hctx
->next_cpu_batch
<= 0) {
1116 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
1117 if (next_cpu
>= nr_cpu_ids
)
1118 next_cpu
= cpumask_first(hctx
->cpumask
);
1120 hctx
->next_cpu
= next_cpu
;
1121 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1124 return hctx
->next_cpu
;
1127 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
,
1128 unsigned long msecs
)
1130 if (unlikely(blk_mq_hctx_stopped(hctx
) ||
1131 !blk_mq_hw_queue_mapped(hctx
)))
1134 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1135 int cpu
= get_cpu();
1136 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1137 __blk_mq_run_hw_queue(hctx
);
1145 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1147 msecs_to_jiffies(msecs
));
1150 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1152 __blk_mq_delay_run_hw_queue(hctx
, true, msecs
);
1154 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue
);
1156 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1158 __blk_mq_delay_run_hw_queue(hctx
, async
, 0);
1160 EXPORT_SYMBOL(blk_mq_run_hw_queue
);
1162 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1164 struct blk_mq_hw_ctx
*hctx
;
1167 queue_for_each_hw_ctx(q
, hctx
, i
) {
1168 if (!blk_mq_hctx_has_pending(hctx
) ||
1169 blk_mq_hctx_stopped(hctx
))
1172 blk_mq_run_hw_queue(hctx
, async
);
1175 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1178 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1179 * @q: request queue.
1181 * The caller is responsible for serializing this function against
1182 * blk_mq_{start,stop}_hw_queue().
1184 bool blk_mq_queue_stopped(struct request_queue
*q
)
1186 struct blk_mq_hw_ctx
*hctx
;
1189 queue_for_each_hw_ctx(q
, hctx
, i
)
1190 if (blk_mq_hctx_stopped(hctx
))
1195 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1197 static void __blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool sync
)
1200 cancel_delayed_work_sync(&hctx
->run_work
);
1202 cancel_delayed_work(&hctx
->run_work
);
1204 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1207 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1209 __blk_mq_stop_hw_queue(hctx
, false);
1211 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1213 static void __blk_mq_stop_hw_queues(struct request_queue
*q
, bool sync
)
1215 struct blk_mq_hw_ctx
*hctx
;
1218 queue_for_each_hw_ctx(q
, hctx
, i
)
1219 __blk_mq_stop_hw_queue(hctx
, sync
);
1222 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1224 __blk_mq_stop_hw_queues(q
, false);
1226 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1228 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1230 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1232 blk_mq_run_hw_queue(hctx
, false);
1234 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1236 void blk_mq_start_hw_queues(struct request_queue
*q
)
1238 struct blk_mq_hw_ctx
*hctx
;
1241 queue_for_each_hw_ctx(q
, hctx
, i
)
1242 blk_mq_start_hw_queue(hctx
);
1244 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1246 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1248 if (!blk_mq_hctx_stopped(hctx
))
1251 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1252 blk_mq_run_hw_queue(hctx
, async
);
1254 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1256 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1258 struct blk_mq_hw_ctx
*hctx
;
1261 queue_for_each_hw_ctx(q
, hctx
, i
)
1262 blk_mq_start_stopped_hw_queue(hctx
, async
);
1264 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1266 static void blk_mq_run_work_fn(struct work_struct
*work
)
1268 struct blk_mq_hw_ctx
*hctx
;
1270 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
1273 * If we are stopped, don't run the queue. The exception is if
1274 * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1275 * the STOPPED bit and run it.
1277 if (test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)) {
1278 if (!test_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
))
1281 clear_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
);
1282 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1285 __blk_mq_run_hw_queue(hctx
);
1289 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1291 if (unlikely(!blk_mq_hw_queue_mapped(hctx
)))
1295 * Stop the hw queue, then modify currently delayed work.
1296 * This should prevent us from running the queue prematurely.
1297 * Mark the queue as auto-clearing STOPPED when it runs.
1299 blk_mq_stop_hw_queue(hctx
);
1300 set_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
);
1301 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1303 msecs_to_jiffies(msecs
));
1305 EXPORT_SYMBOL(blk_mq_delay_queue
);
1307 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1311 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1313 trace_block_rq_insert(hctx
->queue
, rq
);
1316 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1318 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1321 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1324 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1326 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1327 blk_mq_hctx_mark_pending(hctx
, ctx
);
1330 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1331 struct list_head
*list
)
1335 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1338 spin_lock(&ctx
->lock
);
1339 while (!list_empty(list
)) {
1342 rq
= list_first_entry(list
, struct request
, queuelist
);
1343 BUG_ON(rq
->mq_ctx
!= ctx
);
1344 list_del_init(&rq
->queuelist
);
1345 __blk_mq_insert_req_list(hctx
, rq
, false);
1347 blk_mq_hctx_mark_pending(hctx
, ctx
);
1348 spin_unlock(&ctx
->lock
);
1351 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1353 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1354 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1356 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1357 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1358 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1361 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1363 struct blk_mq_ctx
*this_ctx
;
1364 struct request_queue
*this_q
;
1367 LIST_HEAD(ctx_list
);
1370 list_splice_init(&plug
->mq_list
, &list
);
1372 list_sort(NULL
, &list
, plug_ctx_cmp
);
1378 while (!list_empty(&list
)) {
1379 rq
= list_entry_rq(list
.next
);
1380 list_del_init(&rq
->queuelist
);
1382 if (rq
->mq_ctx
!= this_ctx
) {
1384 trace_block_unplug(this_q
, depth
, from_schedule
);
1385 blk_mq_sched_insert_requests(this_q
, this_ctx
,
1390 this_ctx
= rq
->mq_ctx
;
1396 list_add_tail(&rq
->queuelist
, &ctx_list
);
1400 * If 'this_ctx' is set, we know we have entries to complete
1401 * on 'ctx_list'. Do those.
1404 trace_block_unplug(this_q
, depth
, from_schedule
);
1405 blk_mq_sched_insert_requests(this_q
, this_ctx
, &ctx_list
,
1410 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1412 blk_init_request_from_bio(rq
, bio
);
1414 blk_account_io_start(rq
, true);
1417 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1419 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1420 !blk_queue_nomerges(hctx
->queue
);
1423 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx
*hctx
,
1424 struct blk_mq_ctx
*ctx
,
1427 spin_lock(&ctx
->lock
);
1428 __blk_mq_insert_request(hctx
, rq
, false);
1429 spin_unlock(&ctx
->lock
);
1432 static blk_qc_t
request_to_qc_t(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
1435 return blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
, false);
1437 return blk_tag_to_qc_t(rq
->internal_tag
, hctx
->queue_num
, true);
1440 static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1442 blk_qc_t
*cookie
, bool may_sleep
)
1444 struct request_queue
*q
= rq
->q
;
1445 struct blk_mq_queue_data bd
= {
1449 blk_qc_t new_cookie
;
1451 bool run_queue
= true;
1453 /* RCU or SRCU read lock is needed before checking quiesced flag */
1454 if (blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(q
)) {
1462 if (!blk_mq_get_driver_tag(rq
, NULL
, false))
1465 new_cookie
= request_to_qc_t(hctx
, rq
);
1468 * For OK queue, we are done. For error, kill it. Any other
1469 * error (busy), just add it to our list as we previously
1472 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1475 *cookie
= new_cookie
;
1477 case BLK_STS_RESOURCE
:
1478 __blk_mq_requeue_request(rq
);
1481 *cookie
= BLK_QC_T_NONE
;
1482 blk_mq_end_request(rq
, ret
);
1487 blk_mq_sched_insert_request(rq
, false, run_queue
, false, may_sleep
);
1490 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1491 struct request
*rq
, blk_qc_t
*cookie
)
1493 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1495 __blk_mq_try_issue_directly(hctx
, rq
, cookie
, false);
1498 unsigned int srcu_idx
;
1502 srcu_idx
= srcu_read_lock(&hctx
->queue_rq_srcu
);
1503 __blk_mq_try_issue_directly(hctx
, rq
, cookie
, true);
1504 srcu_read_unlock(&hctx
->queue_rq_srcu
, srcu_idx
);
1508 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1510 const int is_sync
= op_is_sync(bio
->bi_opf
);
1511 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1512 struct blk_mq_alloc_data data
= { .flags
= 0 };
1514 unsigned int request_count
= 0;
1515 struct blk_plug
*plug
;
1516 struct request
*same_queue_rq
= NULL
;
1518 unsigned int wb_acct
;
1520 blk_queue_bounce(q
, &bio
);
1522 blk_queue_split(q
, &bio
);
1524 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1526 return BLK_QC_T_NONE
;
1529 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1530 blk_attempt_plug_merge(q
, bio
, &request_count
, &same_queue_rq
))
1531 return BLK_QC_T_NONE
;
1533 if (blk_mq_sched_bio_merge(q
, bio
))
1534 return BLK_QC_T_NONE
;
1536 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1538 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1540 rq
= blk_mq_get_request(q
, bio
, bio
->bi_opf
, &data
);
1541 if (unlikely(!rq
)) {
1542 __wbt_done(q
->rq_wb
, wb_acct
);
1543 return BLK_QC_T_NONE
;
1546 wbt_track(&rq
->issue_stat
, wb_acct
);
1548 cookie
= request_to_qc_t(data
.hctx
, rq
);
1550 plug
= current
->plug
;
1551 if (unlikely(is_flush_fua
)) {
1552 blk_mq_put_ctx(data
.ctx
);
1553 blk_mq_bio_to_request(rq
, bio
);
1555 blk_mq_sched_insert_request(rq
, false, true, true,
1558 blk_insert_flush(rq
);
1559 blk_mq_run_hw_queue(data
.hctx
, true);
1561 } else if (plug
&& q
->nr_hw_queues
== 1) {
1562 struct request
*last
= NULL
;
1564 blk_mq_put_ctx(data
.ctx
);
1565 blk_mq_bio_to_request(rq
, bio
);
1568 * @request_count may become stale because of schedule
1569 * out, so check the list again.
1571 if (list_empty(&plug
->mq_list
))
1573 else if (blk_queue_nomerges(q
))
1574 request_count
= blk_plug_queued_count(q
);
1577 trace_block_plug(q
);
1579 last
= list_entry_rq(plug
->mq_list
.prev
);
1581 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
1582 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1583 blk_flush_plug_list(plug
, false);
1584 trace_block_plug(q
);
1587 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1588 } else if (plug
&& !blk_queue_nomerges(q
)) {
1589 blk_mq_bio_to_request(rq
, bio
);
1592 * We do limited plugging. If the bio can be merged, do that.
1593 * Otherwise the existing request in the plug list will be
1594 * issued. So the plug list will have one request at most
1595 * The plug list might get flushed before this. If that happens,
1596 * the plug list is empty, and same_queue_rq is invalid.
1598 if (list_empty(&plug
->mq_list
))
1599 same_queue_rq
= NULL
;
1601 list_del_init(&same_queue_rq
->queuelist
);
1602 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1604 blk_mq_put_ctx(data
.ctx
);
1606 if (same_queue_rq
) {
1607 data
.hctx
= blk_mq_map_queue(q
,
1608 same_queue_rq
->mq_ctx
->cpu
);
1609 blk_mq_try_issue_directly(data
.hctx
, same_queue_rq
,
1612 } else if (q
->nr_hw_queues
> 1 && is_sync
) {
1613 blk_mq_put_ctx(data
.ctx
);
1614 blk_mq_bio_to_request(rq
, bio
);
1615 blk_mq_try_issue_directly(data
.hctx
, rq
, &cookie
);
1616 } else if (q
->elevator
) {
1617 blk_mq_put_ctx(data
.ctx
);
1618 blk_mq_bio_to_request(rq
, bio
);
1619 blk_mq_sched_insert_request(rq
, false, true, true, true);
1621 blk_mq_put_ctx(data
.ctx
);
1622 blk_mq_bio_to_request(rq
, bio
);
1623 blk_mq_queue_io(data
.hctx
, data
.ctx
, rq
);
1624 blk_mq_run_hw_queue(data
.hctx
, true);
1630 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1631 unsigned int hctx_idx
)
1635 if (tags
->rqs
&& set
->ops
->exit_request
) {
1638 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1639 struct request
*rq
= tags
->static_rqs
[i
];
1643 set
->ops
->exit_request(set
, rq
, hctx_idx
);
1644 tags
->static_rqs
[i
] = NULL
;
1648 while (!list_empty(&tags
->page_list
)) {
1649 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1650 list_del_init(&page
->lru
);
1652 * Remove kmemleak object previously allocated in
1653 * blk_mq_init_rq_map().
1655 kmemleak_free(page_address(page
));
1656 __free_pages(page
, page
->private);
1660 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
1664 kfree(tags
->static_rqs
);
1665 tags
->static_rqs
= NULL
;
1667 blk_mq_free_tags(tags
);
1670 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
1671 unsigned int hctx_idx
,
1672 unsigned int nr_tags
,
1673 unsigned int reserved_tags
)
1675 struct blk_mq_tags
*tags
;
1678 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1679 if (node
== NUMA_NO_NODE
)
1680 node
= set
->numa_node
;
1682 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
,
1683 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1687 tags
->rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1688 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1691 blk_mq_free_tags(tags
);
1695 tags
->static_rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1696 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1698 if (!tags
->static_rqs
) {
1700 blk_mq_free_tags(tags
);
1707 static size_t order_to_size(unsigned int order
)
1709 return (size_t)PAGE_SIZE
<< order
;
1712 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1713 unsigned int hctx_idx
, unsigned int depth
)
1715 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1716 size_t rq_size
, left
;
1719 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1720 if (node
== NUMA_NO_NODE
)
1721 node
= set
->numa_node
;
1723 INIT_LIST_HEAD(&tags
->page_list
);
1726 * rq_size is the size of the request plus driver payload, rounded
1727 * to the cacheline size
1729 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1731 left
= rq_size
* depth
;
1733 for (i
= 0; i
< depth
; ) {
1734 int this_order
= max_order
;
1739 while (this_order
&& left
< order_to_size(this_order
- 1))
1743 page
= alloc_pages_node(node
,
1744 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
1750 if (order_to_size(this_order
) < rq_size
)
1757 page
->private = this_order
;
1758 list_add_tail(&page
->lru
, &tags
->page_list
);
1760 p
= page_address(page
);
1762 * Allow kmemleak to scan these pages as they contain pointers
1763 * to additional allocations like via ops->init_request().
1765 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
1766 entries_per_page
= order_to_size(this_order
) / rq_size
;
1767 to_do
= min(entries_per_page
, depth
- i
);
1768 left
-= to_do
* rq_size
;
1769 for (j
= 0; j
< to_do
; j
++) {
1770 struct request
*rq
= p
;
1772 tags
->static_rqs
[i
] = rq
;
1773 if (set
->ops
->init_request
) {
1774 if (set
->ops
->init_request(set
, rq
, hctx_idx
,
1776 tags
->static_rqs
[i
] = NULL
;
1788 blk_mq_free_rqs(set
, tags
, hctx_idx
);
1793 * 'cpu' is going away. splice any existing rq_list entries from this
1794 * software queue to the hw queue dispatch list, and ensure that it
1797 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
1799 struct blk_mq_hw_ctx
*hctx
;
1800 struct blk_mq_ctx
*ctx
;
1803 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
1804 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
1806 spin_lock(&ctx
->lock
);
1807 if (!list_empty(&ctx
->rq_list
)) {
1808 list_splice_init(&ctx
->rq_list
, &tmp
);
1809 blk_mq_hctx_clear_pending(hctx
, ctx
);
1811 spin_unlock(&ctx
->lock
);
1813 if (list_empty(&tmp
))
1816 spin_lock(&hctx
->lock
);
1817 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
1818 spin_unlock(&hctx
->lock
);
1820 blk_mq_run_hw_queue(hctx
, true);
1824 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
1826 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
1830 /* hctx->ctxs will be freed in queue's release handler */
1831 static void blk_mq_exit_hctx(struct request_queue
*q
,
1832 struct blk_mq_tag_set
*set
,
1833 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
1835 blk_mq_debugfs_unregister_hctx(hctx
);
1837 blk_mq_tag_idle(hctx
);
1839 if (set
->ops
->exit_request
)
1840 set
->ops
->exit_request(set
, hctx
->fq
->flush_rq
, hctx_idx
);
1842 blk_mq_sched_exit_hctx(q
, hctx
, hctx_idx
);
1844 if (set
->ops
->exit_hctx
)
1845 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1847 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1848 cleanup_srcu_struct(&hctx
->queue_rq_srcu
);
1850 blk_mq_remove_cpuhp(hctx
);
1851 blk_free_flush_queue(hctx
->fq
);
1852 sbitmap_free(&hctx
->ctx_map
);
1855 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1856 struct blk_mq_tag_set
*set
, int nr_queue
)
1858 struct blk_mq_hw_ctx
*hctx
;
1861 queue_for_each_hw_ctx(q
, hctx
, i
) {
1864 blk_mq_exit_hctx(q
, set
, hctx
, i
);
1868 static int blk_mq_init_hctx(struct request_queue
*q
,
1869 struct blk_mq_tag_set
*set
,
1870 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
1874 node
= hctx
->numa_node
;
1875 if (node
== NUMA_NO_NODE
)
1876 node
= hctx
->numa_node
= set
->numa_node
;
1878 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1879 spin_lock_init(&hctx
->lock
);
1880 INIT_LIST_HEAD(&hctx
->dispatch
);
1882 hctx
->queue_num
= hctx_idx
;
1883 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
1885 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
1887 hctx
->tags
= set
->tags
[hctx_idx
];
1890 * Allocate space for all possible cpus to avoid allocation at
1893 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1896 goto unregister_cpu_notifier
;
1898 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8), GFP_KERNEL
,
1904 if (set
->ops
->init_hctx
&&
1905 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
1908 if (blk_mq_sched_init_hctx(q
, hctx
, hctx_idx
))
1911 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
1913 goto sched_exit_hctx
;
1915 if (set
->ops
->init_request
&&
1916 set
->ops
->init_request(set
, hctx
->fq
->flush_rq
, hctx_idx
,
1920 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1921 init_srcu_struct(&hctx
->queue_rq_srcu
);
1923 blk_mq_debugfs_register_hctx(q
, hctx
);
1930 blk_mq_sched_exit_hctx(q
, hctx
, hctx_idx
);
1932 if (set
->ops
->exit_hctx
)
1933 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1935 sbitmap_free(&hctx
->ctx_map
);
1938 unregister_cpu_notifier
:
1939 blk_mq_remove_cpuhp(hctx
);
1943 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1944 unsigned int nr_hw_queues
)
1948 for_each_possible_cpu(i
) {
1949 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1950 struct blk_mq_hw_ctx
*hctx
;
1953 spin_lock_init(&__ctx
->lock
);
1954 INIT_LIST_HEAD(&__ctx
->rq_list
);
1957 /* If the cpu isn't online, the cpu is mapped to first hctx */
1961 hctx
= blk_mq_map_queue(q
, i
);
1964 * Set local node, IFF we have more than one hw queue. If
1965 * not, we remain on the home node of the device
1967 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1968 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
1972 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
, int hctx_idx
)
1976 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
1977 set
->queue_depth
, set
->reserved_tags
);
1978 if (!set
->tags
[hctx_idx
])
1981 ret
= blk_mq_alloc_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
,
1986 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
1987 set
->tags
[hctx_idx
] = NULL
;
1991 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set
*set
,
1992 unsigned int hctx_idx
)
1994 if (set
->tags
[hctx_idx
]) {
1995 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
1996 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
1997 set
->tags
[hctx_idx
] = NULL
;
2001 static void blk_mq_map_swqueue(struct request_queue
*q
,
2002 const struct cpumask
*online_mask
)
2004 unsigned int i
, hctx_idx
;
2005 struct blk_mq_hw_ctx
*hctx
;
2006 struct blk_mq_ctx
*ctx
;
2007 struct blk_mq_tag_set
*set
= q
->tag_set
;
2010 * Avoid others reading imcomplete hctx->cpumask through sysfs
2012 mutex_lock(&q
->sysfs_lock
);
2014 queue_for_each_hw_ctx(q
, hctx
, i
) {
2015 cpumask_clear(hctx
->cpumask
);
2020 * Map software to hardware queues
2022 for_each_possible_cpu(i
) {
2023 /* If the cpu isn't online, the cpu is mapped to first hctx */
2024 if (!cpumask_test_cpu(i
, online_mask
))
2027 hctx_idx
= q
->mq_map
[i
];
2028 /* unmapped hw queue can be remapped after CPU topo changed */
2029 if (!set
->tags
[hctx_idx
] &&
2030 !__blk_mq_alloc_rq_map(set
, hctx_idx
)) {
2032 * If tags initialization fail for some hctx,
2033 * that hctx won't be brought online. In this
2034 * case, remap the current ctx to hctx[0] which
2035 * is guaranteed to always have tags allocated
2040 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2041 hctx
= blk_mq_map_queue(q
, i
);
2043 cpumask_set_cpu(i
, hctx
->cpumask
);
2044 ctx
->index_hw
= hctx
->nr_ctx
;
2045 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2048 mutex_unlock(&q
->sysfs_lock
);
2050 queue_for_each_hw_ctx(q
, hctx
, i
) {
2052 * If no software queues are mapped to this hardware queue,
2053 * disable it and free the request entries.
2055 if (!hctx
->nr_ctx
) {
2056 /* Never unmap queue 0. We need it as a
2057 * fallback in case of a new remap fails
2060 if (i
&& set
->tags
[i
])
2061 blk_mq_free_map_and_requests(set
, i
);
2067 hctx
->tags
= set
->tags
[i
];
2068 WARN_ON(!hctx
->tags
);
2071 * Set the map size to the number of mapped software queues.
2072 * This is more accurate and more efficient than looping
2073 * over all possibly mapped software queues.
2075 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2078 * Initialize batch roundrobin counts
2080 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
2081 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2085 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2087 struct blk_mq_hw_ctx
*hctx
;
2090 queue_for_each_hw_ctx(q
, hctx
, i
) {
2092 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
2094 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2098 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
, bool shared
)
2100 struct request_queue
*q
;
2102 lockdep_assert_held(&set
->tag_list_lock
);
2104 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2105 blk_mq_freeze_queue(q
);
2106 queue_set_hctx_shared(q
, shared
);
2107 blk_mq_unfreeze_queue(q
);
2111 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
2113 struct blk_mq_tag_set
*set
= q
->tag_set
;
2115 mutex_lock(&set
->tag_list_lock
);
2116 list_del_rcu(&q
->tag_set_list
);
2117 INIT_LIST_HEAD(&q
->tag_set_list
);
2118 if (list_is_singular(&set
->tag_list
)) {
2119 /* just transitioned to unshared */
2120 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2121 /* update existing queue */
2122 blk_mq_update_tag_set_depth(set
, false);
2124 mutex_unlock(&set
->tag_list_lock
);
2129 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
2130 struct request_queue
*q
)
2134 mutex_lock(&set
->tag_list_lock
);
2136 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2137 if (!list_empty(&set
->tag_list
) && !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
2138 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
2139 /* update existing queue */
2140 blk_mq_update_tag_set_depth(set
, true);
2142 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
2143 queue_set_hctx_shared(q
, true);
2144 list_add_tail_rcu(&q
->tag_set_list
, &set
->tag_list
);
2146 mutex_unlock(&set
->tag_list_lock
);
2150 * It is the actual release handler for mq, but we do it from
2151 * request queue's release handler for avoiding use-after-free
2152 * and headache because q->mq_kobj shouldn't have been introduced,
2153 * but we can't group ctx/kctx kobj without it.
2155 void blk_mq_release(struct request_queue
*q
)
2157 struct blk_mq_hw_ctx
*hctx
;
2160 /* hctx kobj stays in hctx */
2161 queue_for_each_hw_ctx(q
, hctx
, i
) {
2164 kobject_put(&hctx
->kobj
);
2169 kfree(q
->queue_hw_ctx
);
2172 * release .mq_kobj and sw queue's kobject now because
2173 * both share lifetime with request queue.
2175 blk_mq_sysfs_deinit(q
);
2177 free_percpu(q
->queue_ctx
);
2180 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2182 struct request_queue
*uninit_q
, *q
;
2184 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
2186 return ERR_PTR(-ENOMEM
);
2188 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
2190 blk_cleanup_queue(uninit_q
);
2194 EXPORT_SYMBOL(blk_mq_init_queue
);
2196 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2197 struct request_queue
*q
)
2200 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2202 blk_mq_sysfs_unregister(q
);
2203 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2209 node
= blk_mq_hw_queue_to_node(q
->mq_map
, i
);
2210 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
2215 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
2222 atomic_set(&hctxs
[i
]->nr_active
, 0);
2223 hctxs
[i
]->numa_node
= node
;
2224 hctxs
[i
]->queue_num
= i
;
2226 if (blk_mq_init_hctx(q
, set
, hctxs
[i
], i
)) {
2227 free_cpumask_var(hctxs
[i
]->cpumask
);
2232 blk_mq_hctx_kobj_init(hctxs
[i
]);
2234 for (j
= i
; j
< q
->nr_hw_queues
; j
++) {
2235 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2239 blk_mq_free_map_and_requests(set
, j
);
2240 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2241 kobject_put(&hctx
->kobj
);
2246 q
->nr_hw_queues
= i
;
2247 blk_mq_sysfs_register(q
);
2250 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2251 struct request_queue
*q
)
2253 /* mark the queue as mq asap */
2254 q
->mq_ops
= set
->ops
;
2256 q
->poll_cb
= blk_stat_alloc_callback(blk_mq_poll_stats_fn
,
2257 blk_mq_poll_stats_bkt
,
2258 BLK_MQ_POLL_STATS_BKTS
, q
);
2262 q
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2266 /* init q->mq_kobj and sw queues' kobjects */
2267 blk_mq_sysfs_init(q
);
2269 q
->queue_hw_ctx
= kzalloc_node(nr_cpu_ids
* sizeof(*(q
->queue_hw_ctx
)),
2270 GFP_KERNEL
, set
->numa_node
);
2271 if (!q
->queue_hw_ctx
)
2274 q
->mq_map
= set
->mq_map
;
2276 blk_mq_realloc_hw_ctxs(set
, q
);
2277 if (!q
->nr_hw_queues
)
2280 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2281 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2283 q
->nr_queues
= nr_cpu_ids
;
2285 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2287 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2288 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
2290 q
->sg_reserved_size
= INT_MAX
;
2292 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2293 INIT_LIST_HEAD(&q
->requeue_list
);
2294 spin_lock_init(&q
->requeue_lock
);
2296 blk_queue_make_request(q
, blk_mq_make_request
);
2299 * Do this after blk_queue_make_request() overrides it...
2301 q
->nr_requests
= set
->queue_depth
;
2304 * Default to classic polling
2308 if (set
->ops
->complete
)
2309 blk_queue_softirq_done(q
, set
->ops
->complete
);
2311 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2314 mutex_lock(&all_q_mutex
);
2316 list_add_tail(&q
->all_q_node
, &all_q_list
);
2317 blk_mq_add_queue_tag_set(set
, q
);
2318 blk_mq_map_swqueue(q
, cpu_online_mask
);
2320 mutex_unlock(&all_q_mutex
);
2323 if (!(set
->flags
& BLK_MQ_F_NO_SCHED
)) {
2326 ret
= blk_mq_sched_init(q
);
2328 return ERR_PTR(ret
);
2334 kfree(q
->queue_hw_ctx
);
2336 free_percpu(q
->queue_ctx
);
2339 return ERR_PTR(-ENOMEM
);
2341 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2343 void blk_mq_free_queue(struct request_queue
*q
)
2345 struct blk_mq_tag_set
*set
= q
->tag_set
;
2347 mutex_lock(&all_q_mutex
);
2348 list_del_init(&q
->all_q_node
);
2349 mutex_unlock(&all_q_mutex
);
2351 blk_mq_del_queue_tag_set(q
);
2353 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2356 /* Basically redo blk_mq_init_queue with queue frozen */
2357 static void blk_mq_queue_reinit(struct request_queue
*q
,
2358 const struct cpumask
*online_mask
)
2360 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2362 blk_mq_debugfs_unregister_hctxs(q
);
2363 blk_mq_sysfs_unregister(q
);
2366 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2367 * we should change hctx numa_node according to new topology (this
2368 * involves free and re-allocate memory, worthy doing?)
2371 blk_mq_map_swqueue(q
, online_mask
);
2373 blk_mq_sysfs_register(q
);
2374 blk_mq_debugfs_register_hctxs(q
);
2378 * New online cpumask which is going to be set in this hotplug event.
2379 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2380 * one-by-one and dynamically allocating this could result in a failure.
2382 static struct cpumask cpuhp_online_new
;
2384 static void blk_mq_queue_reinit_work(void)
2386 struct request_queue
*q
;
2388 mutex_lock(&all_q_mutex
);
2390 * We need to freeze and reinit all existing queues. Freezing
2391 * involves synchronous wait for an RCU grace period and doing it
2392 * one by one may take a long time. Start freezing all queues in
2393 * one swoop and then wait for the completions so that freezing can
2394 * take place in parallel.
2396 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2397 blk_freeze_queue_start(q
);
2398 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2399 blk_mq_freeze_queue_wait(q
);
2401 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2402 blk_mq_queue_reinit(q
, &cpuhp_online_new
);
2404 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2405 blk_mq_unfreeze_queue(q
);
2407 mutex_unlock(&all_q_mutex
);
2410 static int blk_mq_queue_reinit_dead(unsigned int cpu
)
2412 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2413 blk_mq_queue_reinit_work();
2418 * Before hotadded cpu starts handling requests, new mappings must be
2419 * established. Otherwise, these requests in hw queue might never be
2422 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2423 * for CPU0, and ctx1 for CPU1).
2425 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2426 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2428 * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2429 * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2430 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2433 static int blk_mq_queue_reinit_prepare(unsigned int cpu
)
2435 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2436 cpumask_set_cpu(cpu
, &cpuhp_online_new
);
2437 blk_mq_queue_reinit_work();
2441 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2445 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
2446 if (!__blk_mq_alloc_rq_map(set
, i
))
2453 blk_mq_free_rq_map(set
->tags
[i
]);
2459 * Allocate the request maps associated with this tag_set. Note that this
2460 * may reduce the depth asked for, if memory is tight. set->queue_depth
2461 * will be updated to reflect the allocated depth.
2463 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2468 depth
= set
->queue_depth
;
2470 err
= __blk_mq_alloc_rq_maps(set
);
2474 set
->queue_depth
>>= 1;
2475 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2479 } while (set
->queue_depth
);
2481 if (!set
->queue_depth
|| err
) {
2482 pr_err("blk-mq: failed to allocate request map\n");
2486 if (depth
!= set
->queue_depth
)
2487 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2488 depth
, set
->queue_depth
);
2493 static int blk_mq_update_queue_map(struct blk_mq_tag_set
*set
)
2495 if (set
->ops
->map_queues
)
2496 return set
->ops
->map_queues(set
);
2498 return blk_mq_map_queues(set
);
2502 * Alloc a tag set to be associated with one or more request queues.
2503 * May fail with EINVAL for various error conditions. May adjust the
2504 * requested depth down, if if it too large. In that case, the set
2505 * value will be stored in set->queue_depth.
2507 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2511 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2513 if (!set
->nr_hw_queues
)
2515 if (!set
->queue_depth
)
2517 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2520 if (!set
->ops
->queue_rq
)
2523 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2524 pr_info("blk-mq: reduced tag depth to %u\n",
2526 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2530 * If a crashdump is active, then we are potentially in a very
2531 * memory constrained environment. Limit us to 1 queue and
2532 * 64 tags to prevent using too much memory.
2534 if (is_kdump_kernel()) {
2535 set
->nr_hw_queues
= 1;
2536 set
->queue_depth
= min(64U, set
->queue_depth
);
2539 * There is no use for more h/w queues than cpus.
2541 if (set
->nr_hw_queues
> nr_cpu_ids
)
2542 set
->nr_hw_queues
= nr_cpu_ids
;
2544 set
->tags
= kzalloc_node(nr_cpu_ids
* sizeof(struct blk_mq_tags
*),
2545 GFP_KERNEL
, set
->numa_node
);
2550 set
->mq_map
= kzalloc_node(sizeof(*set
->mq_map
) * nr_cpu_ids
,
2551 GFP_KERNEL
, set
->numa_node
);
2555 ret
= blk_mq_update_queue_map(set
);
2557 goto out_free_mq_map
;
2559 ret
= blk_mq_alloc_rq_maps(set
);
2561 goto out_free_mq_map
;
2563 mutex_init(&set
->tag_list_lock
);
2564 INIT_LIST_HEAD(&set
->tag_list
);
2576 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2578 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2582 for (i
= 0; i
< nr_cpu_ids
; i
++)
2583 blk_mq_free_map_and_requests(set
, i
);
2591 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2593 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2595 struct blk_mq_tag_set
*set
= q
->tag_set
;
2596 struct blk_mq_hw_ctx
*hctx
;
2602 blk_mq_freeze_queue(q
);
2605 queue_for_each_hw_ctx(q
, hctx
, i
) {
2609 * If we're using an MQ scheduler, just update the scheduler
2610 * queue depth. This is similar to what the old code would do.
2612 if (!hctx
->sched_tags
) {
2613 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
,
2614 min(nr
, set
->queue_depth
),
2617 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
2625 q
->nr_requests
= nr
;
2627 blk_mq_unfreeze_queue(q
);
2632 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
,
2635 struct request_queue
*q
;
2637 lockdep_assert_held(&set
->tag_list_lock
);
2639 if (nr_hw_queues
> nr_cpu_ids
)
2640 nr_hw_queues
= nr_cpu_ids
;
2641 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
2644 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2645 blk_mq_freeze_queue(q
);
2647 set
->nr_hw_queues
= nr_hw_queues
;
2648 blk_mq_update_queue_map(set
);
2649 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2650 blk_mq_realloc_hw_ctxs(set
, q
);
2651 blk_mq_queue_reinit(q
, cpu_online_mask
);
2654 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2655 blk_mq_unfreeze_queue(q
);
2658 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
2660 mutex_lock(&set
->tag_list_lock
);
2661 __blk_mq_update_nr_hw_queues(set
, nr_hw_queues
);
2662 mutex_unlock(&set
->tag_list_lock
);
2664 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
2666 /* Enable polling stats and return whether they were already enabled. */
2667 static bool blk_poll_stats_enable(struct request_queue
*q
)
2669 if (test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
2670 test_and_set_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
))
2672 blk_stat_add_callback(q
, q
->poll_cb
);
2676 static void blk_mq_poll_stats_start(struct request_queue
*q
)
2679 * We don't arm the callback if polling stats are not enabled or the
2680 * callback is already active.
2682 if (!test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
2683 blk_stat_is_active(q
->poll_cb
))
2686 blk_stat_activate_msecs(q
->poll_cb
, 100);
2689 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
)
2691 struct request_queue
*q
= cb
->data
;
2694 for (bucket
= 0; bucket
< BLK_MQ_POLL_STATS_BKTS
; bucket
++) {
2695 if (cb
->stat
[bucket
].nr_samples
)
2696 q
->poll_stat
[bucket
] = cb
->stat
[bucket
];
2700 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
2701 struct blk_mq_hw_ctx
*hctx
,
2704 unsigned long ret
= 0;
2708 * If stats collection isn't on, don't sleep but turn it on for
2711 if (!blk_poll_stats_enable(q
))
2715 * As an optimistic guess, use half of the mean service time
2716 * for this type of request. We can (and should) make this smarter.
2717 * For instance, if the completion latencies are tight, we can
2718 * get closer than just half the mean. This is especially
2719 * important on devices where the completion latencies are longer
2720 * than ~10 usec. We do use the stats for the relevant IO size
2721 * if available which does lead to better estimates.
2723 bucket
= blk_mq_poll_stats_bkt(rq
);
2727 if (q
->poll_stat
[bucket
].nr_samples
)
2728 ret
= (q
->poll_stat
[bucket
].mean
+ 1) / 2;
2733 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
2734 struct blk_mq_hw_ctx
*hctx
,
2737 struct hrtimer_sleeper hs
;
2738 enum hrtimer_mode mode
;
2742 if (test_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
))
2748 * -1: don't ever hybrid sleep
2749 * 0: use half of prev avg
2750 * >0: use this specific value
2752 if (q
->poll_nsec
== -1)
2754 else if (q
->poll_nsec
> 0)
2755 nsecs
= q
->poll_nsec
;
2757 nsecs
= blk_mq_poll_nsecs(q
, hctx
, rq
);
2762 set_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
2765 * This will be replaced with the stats tracking code, using
2766 * 'avg_completion_time / 2' as the pre-sleep target.
2770 mode
= HRTIMER_MODE_REL
;
2771 hrtimer_init_on_stack(&hs
.timer
, CLOCK_MONOTONIC
, mode
);
2772 hrtimer_set_expires(&hs
.timer
, kt
);
2774 hrtimer_init_sleeper(&hs
, current
);
2776 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
2778 set_current_state(TASK_UNINTERRUPTIBLE
);
2779 hrtimer_start_expires(&hs
.timer
, mode
);
2782 hrtimer_cancel(&hs
.timer
);
2783 mode
= HRTIMER_MODE_ABS
;
2784 } while (hs
.task
&& !signal_pending(current
));
2786 __set_current_state(TASK_RUNNING
);
2787 destroy_hrtimer_on_stack(&hs
.timer
);
2791 static bool __blk_mq_poll(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
2793 struct request_queue
*q
= hctx
->queue
;
2797 * If we sleep, have the caller restart the poll loop to reset
2798 * the state. Like for the other success return cases, the
2799 * caller is responsible for checking if the IO completed. If
2800 * the IO isn't complete, we'll get called again and will go
2801 * straight to the busy poll loop.
2803 if (blk_mq_poll_hybrid_sleep(q
, hctx
, rq
))
2806 hctx
->poll_considered
++;
2808 state
= current
->state
;
2809 while (!need_resched()) {
2812 hctx
->poll_invoked
++;
2814 ret
= q
->mq_ops
->poll(hctx
, rq
->tag
);
2816 hctx
->poll_success
++;
2817 set_current_state(TASK_RUNNING
);
2821 if (signal_pending_state(state
, current
))
2822 set_current_state(TASK_RUNNING
);
2824 if (current
->state
== TASK_RUNNING
)
2834 bool blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
)
2836 struct blk_mq_hw_ctx
*hctx
;
2837 struct blk_plug
*plug
;
2840 if (!q
->mq_ops
|| !q
->mq_ops
->poll
|| !blk_qc_t_valid(cookie
) ||
2841 !test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
2844 plug
= current
->plug
;
2846 blk_flush_plug_list(plug
, false);
2848 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
2849 if (!blk_qc_t_is_internal(cookie
))
2850 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
2852 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
2854 * With scheduling, if the request has completed, we'll
2855 * get a NULL return here, as we clear the sched tag when
2856 * that happens. The request still remains valid, like always,
2857 * so we should be safe with just the NULL check.
2863 return __blk_mq_poll(hctx
, rq
);
2865 EXPORT_SYMBOL_GPL(blk_mq_poll
);
2867 void blk_mq_disable_hotplug(void)
2869 mutex_lock(&all_q_mutex
);
2872 void blk_mq_enable_hotplug(void)
2874 mutex_unlock(&all_q_mutex
);
2877 static int __init
blk_mq_init(void)
2879 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
2880 blk_mq_hctx_notify_dead
);
2882 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE
, "block/mq:prepare",
2883 blk_mq_queue_reinit_prepare
,
2884 blk_mq_queue_reinit_dead
);
2887 subsys_initcall(blk_mq_init
);