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/delay.h>
24 #include <linux/crash_dump.h>
25 #include <linux/prefetch.h>
27 #include <trace/events/block.h>
29 #include <linux/blk-mq.h>
32 #include "blk-mq-tag.h"
34 static DEFINE_MUTEX(all_q_mutex
);
35 static LIST_HEAD(all_q_list
);
38 * Check if any of the ctx's have pending work in this hardware queue
40 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
42 return sbitmap_any_bit_set(&hctx
->ctx_map
);
46 * Mark this ctx as having pending work in this hardware queue
48 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
49 struct blk_mq_ctx
*ctx
)
51 if (!sbitmap_test_bit(&hctx
->ctx_map
, ctx
->index_hw
))
52 sbitmap_set_bit(&hctx
->ctx_map
, ctx
->index_hw
);
55 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
56 struct blk_mq_ctx
*ctx
)
58 sbitmap_clear_bit(&hctx
->ctx_map
, ctx
->index_hw
);
61 void blk_mq_freeze_queue_start(struct request_queue
*q
)
65 freeze_depth
= atomic_inc_return(&q
->mq_freeze_depth
);
66 if (freeze_depth
== 1) {
67 percpu_ref_kill(&q
->q_usage_counter
);
68 blk_mq_run_hw_queues(q
, false);
71 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start
);
73 static void blk_mq_freeze_queue_wait(struct request_queue
*q
)
75 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
79 * Guarantee no request is in use, so we can change any data structure of
80 * the queue afterward.
82 void blk_freeze_queue(struct request_queue
*q
)
85 * In the !blk_mq case we are only calling this to kill the
86 * q_usage_counter, otherwise this increases the freeze depth
87 * and waits for it to return to zero. For this reason there is
88 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
89 * exported to drivers as the only user for unfreeze is blk_mq.
91 blk_mq_freeze_queue_start(q
);
92 blk_mq_freeze_queue_wait(q
);
95 void blk_mq_freeze_queue(struct request_queue
*q
)
98 * ...just an alias to keep freeze and unfreeze actions balanced
99 * in the blk_mq_* namespace
103 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
105 void blk_mq_unfreeze_queue(struct request_queue
*q
)
109 freeze_depth
= atomic_dec_return(&q
->mq_freeze_depth
);
110 WARN_ON_ONCE(freeze_depth
< 0);
112 percpu_ref_reinit(&q
->q_usage_counter
);
113 wake_up_all(&q
->mq_freeze_wq
);
116 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
118 void blk_mq_wake_waiters(struct request_queue
*q
)
120 struct blk_mq_hw_ctx
*hctx
;
123 queue_for_each_hw_ctx(q
, hctx
, i
)
124 if (blk_mq_hw_queue_mapped(hctx
))
125 blk_mq_tag_wakeup_all(hctx
->tags
, true);
128 * If we are called because the queue has now been marked as
129 * dying, we need to ensure that processes currently waiting on
130 * the queue are notified as well.
132 wake_up_all(&q
->mq_freeze_wq
);
135 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
137 return blk_mq_has_free_tags(hctx
->tags
);
139 EXPORT_SYMBOL(blk_mq_can_queue
);
141 static void blk_mq_rq_ctx_init(struct request_queue
*q
, struct blk_mq_ctx
*ctx
,
142 struct request
*rq
, unsigned int op
)
144 INIT_LIST_HEAD(&rq
->queuelist
);
145 /* csd/requeue_work/fifo_time is initialized before use */
149 if (blk_queue_io_stat(q
))
150 rq
->rq_flags
|= RQF_IO_STAT
;
151 /* do not touch atomic flags, it needs atomic ops against the timer */
153 INIT_HLIST_NODE(&rq
->hash
);
154 RB_CLEAR_NODE(&rq
->rb_node
);
157 rq
->start_time
= jiffies
;
158 #ifdef CONFIG_BLK_CGROUP
160 set_start_time_ns(rq
);
161 rq
->io_start_time_ns
= 0;
163 rq
->nr_phys_segments
= 0;
164 #if defined(CONFIG_BLK_DEV_INTEGRITY)
165 rq
->nr_integrity_segments
= 0;
168 /* tag was already set */
178 INIT_LIST_HEAD(&rq
->timeout_list
);
182 rq
->end_io_data
= NULL
;
185 ctx
->rq_dispatched
[op_is_sync(op
)]++;
188 static struct request
*
189 __blk_mq_alloc_request(struct blk_mq_alloc_data
*data
, unsigned int op
)
194 tag
= blk_mq_get_tag(data
);
195 if (tag
!= BLK_MQ_TAG_FAIL
) {
196 rq
= data
->hctx
->tags
->rqs
[tag
];
198 if (blk_mq_tag_busy(data
->hctx
)) {
199 rq
->rq_flags
= RQF_MQ_INFLIGHT
;
200 atomic_inc(&data
->hctx
->nr_active
);
204 blk_mq_rq_ctx_init(data
->q
, data
->ctx
, rq
, op
);
211 struct request
*blk_mq_alloc_request(struct request_queue
*q
, int rw
,
214 struct blk_mq_ctx
*ctx
;
215 struct blk_mq_hw_ctx
*hctx
;
217 struct blk_mq_alloc_data alloc_data
;
220 ret
= blk_queue_enter(q
, flags
& BLK_MQ_REQ_NOWAIT
);
224 ctx
= blk_mq_get_ctx(q
);
225 hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
226 blk_mq_set_alloc_data(&alloc_data
, q
, flags
, ctx
, hctx
);
227 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
232 return ERR_PTR(-EWOULDBLOCK
);
236 rq
->__sector
= (sector_t
) -1;
237 rq
->bio
= rq
->biotail
= NULL
;
240 EXPORT_SYMBOL(blk_mq_alloc_request
);
242 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
, int rw
,
243 unsigned int flags
, unsigned int hctx_idx
)
245 struct blk_mq_hw_ctx
*hctx
;
246 struct blk_mq_ctx
*ctx
;
248 struct blk_mq_alloc_data alloc_data
;
252 * If the tag allocator sleeps we could get an allocation for a
253 * different hardware context. No need to complicate the low level
254 * allocator for this for the rare use case of a command tied to
257 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
258 return ERR_PTR(-EINVAL
);
260 if (hctx_idx
>= q
->nr_hw_queues
)
261 return ERR_PTR(-EIO
);
263 ret
= blk_queue_enter(q
, true);
268 * Check if the hardware context is actually mapped to anything.
269 * If not tell the caller that it should skip this queue.
271 hctx
= q
->queue_hw_ctx
[hctx_idx
];
272 if (!blk_mq_hw_queue_mapped(hctx
)) {
276 ctx
= __blk_mq_get_ctx(q
, cpumask_first(hctx
->cpumask
));
278 blk_mq_set_alloc_data(&alloc_data
, q
, flags
, ctx
, hctx
);
279 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
291 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
293 static void __blk_mq_free_request(struct blk_mq_hw_ctx
*hctx
,
294 struct blk_mq_ctx
*ctx
, struct request
*rq
)
296 const int tag
= rq
->tag
;
297 struct request_queue
*q
= rq
->q
;
299 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
300 atomic_dec(&hctx
->nr_active
);
303 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
304 blk_mq_put_tag(hctx
, ctx
, tag
);
308 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
310 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
312 ctx
->rq_completed
[rq_is_sync(rq
)]++;
313 __blk_mq_free_request(hctx
, ctx
, rq
);
316 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request
);
318 void blk_mq_free_request(struct request
*rq
)
320 blk_mq_free_hctx_request(blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
), rq
);
322 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
324 inline void __blk_mq_end_request(struct request
*rq
, int error
)
326 blk_account_io_done(rq
);
329 rq
->end_io(rq
, error
);
331 if (unlikely(blk_bidi_rq(rq
)))
332 blk_mq_free_request(rq
->next_rq
);
333 blk_mq_free_request(rq
);
336 EXPORT_SYMBOL(__blk_mq_end_request
);
338 void blk_mq_end_request(struct request
*rq
, int error
)
340 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
342 __blk_mq_end_request(rq
, error
);
344 EXPORT_SYMBOL(blk_mq_end_request
);
346 static void __blk_mq_complete_request_remote(void *data
)
348 struct request
*rq
= data
;
350 rq
->q
->softirq_done_fn(rq
);
353 static void blk_mq_ipi_complete_request(struct request
*rq
)
355 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
359 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
360 rq
->q
->softirq_done_fn(rq
);
365 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
366 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
368 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
369 rq
->csd
.func
= __blk_mq_complete_request_remote
;
372 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
374 rq
->q
->softirq_done_fn(rq
);
379 static void __blk_mq_complete_request(struct request
*rq
)
381 struct request_queue
*q
= rq
->q
;
383 if (!q
->softirq_done_fn
)
384 blk_mq_end_request(rq
, rq
->errors
);
386 blk_mq_ipi_complete_request(rq
);
390 * blk_mq_complete_request - end I/O on a request
391 * @rq: the request being processed
394 * Ends all I/O on a request. It does not handle partial completions.
395 * The actual completion happens out-of-order, through a IPI handler.
397 void blk_mq_complete_request(struct request
*rq
, int error
)
399 struct request_queue
*q
= rq
->q
;
401 if (unlikely(blk_should_fake_timeout(q
)))
403 if (!blk_mark_rq_complete(rq
)) {
405 __blk_mq_complete_request(rq
);
408 EXPORT_SYMBOL(blk_mq_complete_request
);
410 int blk_mq_request_started(struct request
*rq
)
412 return test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
414 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
416 void blk_mq_start_request(struct request
*rq
)
418 struct request_queue
*q
= rq
->q
;
420 trace_block_rq_issue(q
, rq
);
422 rq
->resid_len
= blk_rq_bytes(rq
);
423 if (unlikely(blk_bidi_rq(rq
)))
424 rq
->next_rq
->resid_len
= blk_rq_bytes(rq
->next_rq
);
429 * Ensure that ->deadline is visible before set the started
430 * flag and clear the completed flag.
432 smp_mb__before_atomic();
435 * Mark us as started and clear complete. Complete might have been
436 * set if requeue raced with timeout, which then marked it as
437 * complete. So be sure to clear complete again when we start
438 * the request, otherwise we'll ignore the completion event.
440 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
441 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
442 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
443 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
445 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
447 * Make sure space for the drain appears. We know we can do
448 * this because max_hw_segments has been adjusted to be one
449 * fewer than the device can handle.
451 rq
->nr_phys_segments
++;
454 EXPORT_SYMBOL(blk_mq_start_request
);
456 static void __blk_mq_requeue_request(struct request
*rq
)
458 struct request_queue
*q
= rq
->q
;
460 trace_block_rq_requeue(q
, rq
);
462 if (test_and_clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
463 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
464 rq
->nr_phys_segments
--;
468 void blk_mq_requeue_request(struct request
*rq
)
470 __blk_mq_requeue_request(rq
);
472 BUG_ON(blk_queued_rq(rq
));
473 blk_mq_add_to_requeue_list(rq
, true);
475 EXPORT_SYMBOL(blk_mq_requeue_request
);
477 static void blk_mq_requeue_work(struct work_struct
*work
)
479 struct request_queue
*q
=
480 container_of(work
, struct request_queue
, requeue_work
.work
);
482 struct request
*rq
, *next
;
485 spin_lock_irqsave(&q
->requeue_lock
, flags
);
486 list_splice_init(&q
->requeue_list
, &rq_list
);
487 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
489 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
490 if (!(rq
->rq_flags
& RQF_SOFTBARRIER
))
493 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
494 list_del_init(&rq
->queuelist
);
495 blk_mq_insert_request(rq
, true, false, false);
498 while (!list_empty(&rq_list
)) {
499 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
500 list_del_init(&rq
->queuelist
);
501 blk_mq_insert_request(rq
, false, false, false);
505 * Use the start variant of queue running here, so that running
506 * the requeue work will kick stopped queues.
508 blk_mq_start_hw_queues(q
);
511 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
)
513 struct request_queue
*q
= rq
->q
;
517 * We abuse this flag that is otherwise used by the I/O scheduler to
518 * request head insertation from the workqueue.
520 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
522 spin_lock_irqsave(&q
->requeue_lock
, flags
);
524 rq
->rq_flags
|= RQF_SOFTBARRIER
;
525 list_add(&rq
->queuelist
, &q
->requeue_list
);
527 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
529 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
531 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
533 void blk_mq_cancel_requeue_work(struct request_queue
*q
)
535 cancel_delayed_work_sync(&q
->requeue_work
);
537 EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work
);
539 void blk_mq_kick_requeue_list(struct request_queue
*q
)
541 kblockd_schedule_delayed_work(&q
->requeue_work
, 0);
543 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
545 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
548 kblockd_schedule_delayed_work(&q
->requeue_work
,
549 msecs_to_jiffies(msecs
));
551 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
553 void blk_mq_abort_requeue_list(struct request_queue
*q
)
558 spin_lock_irqsave(&q
->requeue_lock
, flags
);
559 list_splice_init(&q
->requeue_list
, &rq_list
);
560 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
562 while (!list_empty(&rq_list
)) {
565 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
566 list_del_init(&rq
->queuelist
);
568 blk_mq_end_request(rq
, rq
->errors
);
571 EXPORT_SYMBOL(blk_mq_abort_requeue_list
);
573 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
575 if (tag
< tags
->nr_tags
) {
576 prefetch(tags
->rqs
[tag
]);
577 return tags
->rqs
[tag
];
582 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
584 struct blk_mq_timeout_data
{
586 unsigned int next_set
;
589 void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
591 struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
592 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
595 * We know that complete is set at this point. If STARTED isn't set
596 * anymore, then the request isn't active and the "timeout" should
597 * just be ignored. This can happen due to the bitflag ordering.
598 * Timeout first checks if STARTED is set, and if it is, assumes
599 * the request is active. But if we race with completion, then
600 * we both flags will get cleared. So check here again, and ignore
601 * a timeout event with a request that isn't active.
603 if (!test_bit(REQ_ATOM_STARTED
, &req
->atomic_flags
))
607 ret
= ops
->timeout(req
, reserved
);
611 __blk_mq_complete_request(req
);
613 case BLK_EH_RESET_TIMER
:
615 blk_clear_rq_complete(req
);
617 case BLK_EH_NOT_HANDLED
:
620 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
625 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
626 struct request
*rq
, void *priv
, bool reserved
)
628 struct blk_mq_timeout_data
*data
= priv
;
630 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
632 * If a request wasn't started before the queue was
633 * marked dying, kill it here or it'll go unnoticed.
635 if (unlikely(blk_queue_dying(rq
->q
))) {
637 blk_mq_end_request(rq
, rq
->errors
);
642 if (time_after_eq(jiffies
, rq
->deadline
)) {
643 if (!blk_mark_rq_complete(rq
))
644 blk_mq_rq_timed_out(rq
, reserved
);
645 } else if (!data
->next_set
|| time_after(data
->next
, rq
->deadline
)) {
646 data
->next
= rq
->deadline
;
651 static void blk_mq_timeout_work(struct work_struct
*work
)
653 struct request_queue
*q
=
654 container_of(work
, struct request_queue
, timeout_work
);
655 struct blk_mq_timeout_data data
= {
661 /* A deadlock might occur if a request is stuck requiring a
662 * timeout at the same time a queue freeze is waiting
663 * completion, since the timeout code would not be able to
664 * acquire the queue reference here.
666 * That's why we don't use blk_queue_enter here; instead, we use
667 * percpu_ref_tryget directly, because we need to be able to
668 * obtain a reference even in the short window between the queue
669 * starting to freeze, by dropping the first reference in
670 * blk_mq_freeze_queue_start, and the moment the last request is
671 * consumed, marked by the instant q_usage_counter reaches
674 if (!percpu_ref_tryget(&q
->q_usage_counter
))
677 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &data
);
680 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
681 mod_timer(&q
->timeout
, data
.next
);
683 struct blk_mq_hw_ctx
*hctx
;
685 queue_for_each_hw_ctx(q
, hctx
, i
) {
686 /* the hctx may be unmapped, so check it here */
687 if (blk_mq_hw_queue_mapped(hctx
))
688 blk_mq_tag_idle(hctx
);
695 * Reverse check our software queue for entries that we could potentially
696 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
697 * too much time checking for merges.
699 static bool blk_mq_attempt_merge(struct request_queue
*q
,
700 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
705 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
711 if (!blk_rq_merge_ok(rq
, bio
))
714 el_ret
= blk_try_merge(rq
, bio
);
715 if (el_ret
== ELEVATOR_BACK_MERGE
) {
716 if (bio_attempt_back_merge(q
, rq
, bio
)) {
721 } else if (el_ret
== ELEVATOR_FRONT_MERGE
) {
722 if (bio_attempt_front_merge(q
, rq
, bio
)) {
733 struct flush_busy_ctx_data
{
734 struct blk_mq_hw_ctx
*hctx
;
735 struct list_head
*list
;
738 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
740 struct flush_busy_ctx_data
*flush_data
= data
;
741 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
742 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
744 sbitmap_clear_bit(sb
, bitnr
);
745 spin_lock(&ctx
->lock
);
746 list_splice_tail_init(&ctx
->rq_list
, flush_data
->list
);
747 spin_unlock(&ctx
->lock
);
752 * Process software queues that have been marked busy, splicing them
753 * to the for-dispatch
755 static void flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
757 struct flush_busy_ctx_data data
= {
762 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
765 static inline unsigned int queued_to_index(unsigned int queued
)
770 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
774 * Run this hardware queue, pulling any software queues mapped to it in.
775 * Note that this function currently has various problems around ordering
776 * of IO. In particular, we'd like FIFO behaviour on handling existing
777 * items on the hctx->dispatch list. Ignore that for now.
779 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
781 struct request_queue
*q
= hctx
->queue
;
784 LIST_HEAD(driver_list
);
785 struct list_head
*dptr
;
788 if (unlikely(blk_mq_hctx_stopped(hctx
)))
791 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
792 cpu_online(hctx
->next_cpu
));
797 * Touch any software queue that has pending entries.
799 flush_busy_ctxs(hctx
, &rq_list
);
802 * If we have previous entries on our dispatch list, grab them
803 * and stuff them at the front for more fair dispatch.
805 if (!list_empty_careful(&hctx
->dispatch
)) {
806 spin_lock(&hctx
->lock
);
807 if (!list_empty(&hctx
->dispatch
))
808 list_splice_init(&hctx
->dispatch
, &rq_list
);
809 spin_unlock(&hctx
->lock
);
813 * Start off with dptr being NULL, so we start the first request
814 * immediately, even if we have more pending.
819 * Now process all the entries, sending them to the driver.
822 while (!list_empty(&rq_list
)) {
823 struct blk_mq_queue_data bd
;
826 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
827 list_del_init(&rq
->queuelist
);
831 bd
.last
= list_empty(&rq_list
);
833 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
835 case BLK_MQ_RQ_QUEUE_OK
:
838 case BLK_MQ_RQ_QUEUE_BUSY
:
839 list_add(&rq
->queuelist
, &rq_list
);
840 __blk_mq_requeue_request(rq
);
843 pr_err("blk-mq: bad return on queue: %d\n", ret
);
844 case BLK_MQ_RQ_QUEUE_ERROR
:
846 blk_mq_end_request(rq
, rq
->errors
);
850 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
854 * We've done the first request. If we have more than 1
855 * left in the list, set dptr to defer issue.
857 if (!dptr
&& rq_list
.next
!= rq_list
.prev
)
861 hctx
->dispatched
[queued_to_index(queued
)]++;
864 * Any items that need requeuing? Stuff them into hctx->dispatch,
865 * that is where we will continue on next queue run.
867 if (!list_empty(&rq_list
)) {
868 spin_lock(&hctx
->lock
);
869 list_splice(&rq_list
, &hctx
->dispatch
);
870 spin_unlock(&hctx
->lock
);
872 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
873 * it's possible the queue is stopped and restarted again
874 * before this. Queue restart will dispatch requests. And since
875 * requests in rq_list aren't added into hctx->dispatch yet,
876 * the requests in rq_list might get lost.
878 * blk_mq_run_hw_queue() already checks the STOPPED bit
880 blk_mq_run_hw_queue(hctx
, true);
885 * It'd be great if the workqueue API had a way to pass
886 * in a mask and had some smarts for more clever placement.
887 * For now we just round-robin here, switching for every
888 * BLK_MQ_CPU_WORK_BATCH queued items.
890 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
892 if (hctx
->queue
->nr_hw_queues
== 1)
893 return WORK_CPU_UNBOUND
;
895 if (--hctx
->next_cpu_batch
<= 0) {
896 int cpu
= hctx
->next_cpu
, next_cpu
;
898 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
899 if (next_cpu
>= nr_cpu_ids
)
900 next_cpu
= cpumask_first(hctx
->cpumask
);
902 hctx
->next_cpu
= next_cpu
;
903 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
908 return hctx
->next_cpu
;
911 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
913 if (unlikely(blk_mq_hctx_stopped(hctx
) ||
914 !blk_mq_hw_queue_mapped(hctx
)))
917 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
919 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
920 __blk_mq_run_hw_queue(hctx
);
928 kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
);
931 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
933 struct blk_mq_hw_ctx
*hctx
;
936 queue_for_each_hw_ctx(q
, hctx
, i
) {
937 if ((!blk_mq_hctx_has_pending(hctx
) &&
938 list_empty_careful(&hctx
->dispatch
)) ||
939 blk_mq_hctx_stopped(hctx
))
942 blk_mq_run_hw_queue(hctx
, async
);
945 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
948 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
951 * The caller is responsible for serializing this function against
952 * blk_mq_{start,stop}_hw_queue().
954 bool blk_mq_queue_stopped(struct request_queue
*q
)
956 struct blk_mq_hw_ctx
*hctx
;
959 queue_for_each_hw_ctx(q
, hctx
, i
)
960 if (blk_mq_hctx_stopped(hctx
))
965 EXPORT_SYMBOL(blk_mq_queue_stopped
);
967 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
969 cancel_work(&hctx
->run_work
);
970 cancel_delayed_work(&hctx
->delay_work
);
971 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
973 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
975 void blk_mq_stop_hw_queues(struct request_queue
*q
)
977 struct blk_mq_hw_ctx
*hctx
;
980 queue_for_each_hw_ctx(q
, hctx
, i
)
981 blk_mq_stop_hw_queue(hctx
);
983 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
985 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
987 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
989 blk_mq_run_hw_queue(hctx
, false);
991 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
993 void blk_mq_start_hw_queues(struct request_queue
*q
)
995 struct blk_mq_hw_ctx
*hctx
;
998 queue_for_each_hw_ctx(q
, hctx
, i
)
999 blk_mq_start_hw_queue(hctx
);
1001 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1003 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1005 struct blk_mq_hw_ctx
*hctx
;
1008 queue_for_each_hw_ctx(q
, hctx
, i
) {
1009 if (!blk_mq_hctx_stopped(hctx
))
1012 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1013 blk_mq_run_hw_queue(hctx
, async
);
1016 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1018 static void blk_mq_run_work_fn(struct work_struct
*work
)
1020 struct blk_mq_hw_ctx
*hctx
;
1022 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
);
1024 __blk_mq_run_hw_queue(hctx
);
1027 static void blk_mq_delay_work_fn(struct work_struct
*work
)
1029 struct blk_mq_hw_ctx
*hctx
;
1031 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
1033 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
1034 __blk_mq_run_hw_queue(hctx
);
1037 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1039 if (unlikely(!blk_mq_hw_queue_mapped(hctx
)))
1042 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1043 &hctx
->delay_work
, msecs_to_jiffies(msecs
));
1045 EXPORT_SYMBOL(blk_mq_delay_queue
);
1047 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1051 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1053 trace_block_rq_insert(hctx
->queue
, rq
);
1056 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1058 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1061 static void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
,
1062 struct request
*rq
, bool at_head
)
1064 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1066 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1067 blk_mq_hctx_mark_pending(hctx
, ctx
);
1070 void blk_mq_insert_request(struct request
*rq
, bool at_head
, bool run_queue
,
1073 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1074 struct request_queue
*q
= rq
->q
;
1075 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
1077 spin_lock(&ctx
->lock
);
1078 __blk_mq_insert_request(hctx
, rq
, at_head
);
1079 spin_unlock(&ctx
->lock
);
1082 blk_mq_run_hw_queue(hctx
, async
);
1085 static void blk_mq_insert_requests(struct request_queue
*q
,
1086 struct blk_mq_ctx
*ctx
,
1087 struct list_head
*list
,
1092 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
1094 trace_block_unplug(q
, depth
, !from_schedule
);
1097 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1100 spin_lock(&ctx
->lock
);
1101 while (!list_empty(list
)) {
1104 rq
= list_first_entry(list
, struct request
, queuelist
);
1105 BUG_ON(rq
->mq_ctx
!= ctx
);
1106 list_del_init(&rq
->queuelist
);
1107 __blk_mq_insert_req_list(hctx
, rq
, false);
1109 blk_mq_hctx_mark_pending(hctx
, ctx
);
1110 spin_unlock(&ctx
->lock
);
1112 blk_mq_run_hw_queue(hctx
, from_schedule
);
1115 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1117 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1118 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1120 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1121 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1122 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1125 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1127 struct blk_mq_ctx
*this_ctx
;
1128 struct request_queue
*this_q
;
1131 LIST_HEAD(ctx_list
);
1134 list_splice_init(&plug
->mq_list
, &list
);
1136 list_sort(NULL
, &list
, plug_ctx_cmp
);
1142 while (!list_empty(&list
)) {
1143 rq
= list_entry_rq(list
.next
);
1144 list_del_init(&rq
->queuelist
);
1146 if (rq
->mq_ctx
!= this_ctx
) {
1148 blk_mq_insert_requests(this_q
, this_ctx
,
1153 this_ctx
= rq
->mq_ctx
;
1159 list_add_tail(&rq
->queuelist
, &ctx_list
);
1163 * If 'this_ctx' is set, we know we have entries to complete
1164 * on 'ctx_list'. Do those.
1167 blk_mq_insert_requests(this_q
, this_ctx
, &ctx_list
, depth
,
1172 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1174 init_request_from_bio(rq
, bio
);
1176 blk_account_io_start(rq
, 1);
1179 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1181 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1182 !blk_queue_nomerges(hctx
->queue
);
1185 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx
*hctx
,
1186 struct blk_mq_ctx
*ctx
,
1187 struct request
*rq
, struct bio
*bio
)
1189 if (!hctx_allow_merges(hctx
) || !bio_mergeable(bio
)) {
1190 blk_mq_bio_to_request(rq
, bio
);
1191 spin_lock(&ctx
->lock
);
1193 __blk_mq_insert_request(hctx
, rq
, false);
1194 spin_unlock(&ctx
->lock
);
1197 struct request_queue
*q
= hctx
->queue
;
1199 spin_lock(&ctx
->lock
);
1200 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1201 blk_mq_bio_to_request(rq
, bio
);
1205 spin_unlock(&ctx
->lock
);
1206 __blk_mq_free_request(hctx
, ctx
, rq
);
1211 static struct request
*blk_mq_map_request(struct request_queue
*q
,
1213 struct blk_mq_alloc_data
*data
)
1215 struct blk_mq_hw_ctx
*hctx
;
1216 struct blk_mq_ctx
*ctx
;
1219 blk_queue_enter_live(q
);
1220 ctx
= blk_mq_get_ctx(q
);
1221 hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
1223 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1224 blk_mq_set_alloc_data(data
, q
, 0, ctx
, hctx
);
1225 rq
= __blk_mq_alloc_request(data
, bio
->bi_opf
);
1227 data
->hctx
->queued
++;
1231 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1232 struct request
*rq
, blk_qc_t
*cookie
)
1235 struct request_queue
*q
= rq
->q
;
1236 struct blk_mq_queue_data bd
= {
1241 blk_qc_t new_cookie
= blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
);
1243 if (blk_mq_hctx_stopped(hctx
))
1247 * For OK queue, we are done. For error, kill it. Any other
1248 * error (busy), just add it to our list as we previously
1251 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1252 if (ret
== BLK_MQ_RQ_QUEUE_OK
) {
1253 *cookie
= new_cookie
;
1257 __blk_mq_requeue_request(rq
);
1259 if (ret
== BLK_MQ_RQ_QUEUE_ERROR
) {
1260 *cookie
= BLK_QC_T_NONE
;
1262 blk_mq_end_request(rq
, rq
->errors
);
1267 blk_mq_insert_request(rq
, false, true, true);
1271 * Multiple hardware queue variant. This will not use per-process plugs,
1272 * but will attempt to bypass the hctx queueing if we can go straight to
1273 * hardware for SYNC IO.
1275 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1277 const int is_sync
= op_is_sync(bio
->bi_opf
);
1278 const int is_flush_fua
= bio
->bi_opf
& (REQ_PREFLUSH
| REQ_FUA
);
1279 struct blk_mq_alloc_data data
;
1281 unsigned int request_count
= 0;
1282 struct blk_plug
*plug
;
1283 struct request
*same_queue_rq
= NULL
;
1286 blk_queue_bounce(q
, &bio
);
1288 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1290 return BLK_QC_T_NONE
;
1293 blk_queue_split(q
, &bio
, q
->bio_split
);
1295 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1296 blk_attempt_plug_merge(q
, bio
, &request_count
, &same_queue_rq
))
1297 return BLK_QC_T_NONE
;
1299 rq
= blk_mq_map_request(q
, bio
, &data
);
1301 return BLK_QC_T_NONE
;
1303 cookie
= blk_tag_to_qc_t(rq
->tag
, data
.hctx
->queue_num
);
1305 if (unlikely(is_flush_fua
)) {
1306 blk_mq_bio_to_request(rq
, bio
);
1307 blk_insert_flush(rq
);
1311 plug
= current
->plug
;
1313 * If the driver supports defer issued based on 'last', then
1314 * queue it up like normal since we can potentially save some
1317 if (((plug
&& !blk_queue_nomerges(q
)) || is_sync
) &&
1318 !(data
.hctx
->flags
& BLK_MQ_F_DEFER_ISSUE
)) {
1319 struct request
*old_rq
= NULL
;
1321 blk_mq_bio_to_request(rq
, bio
);
1324 * We do limited pluging. If the bio can be merged, do that.
1325 * Otherwise the existing request in the plug list will be
1326 * issued. So the plug list will have one request at most
1330 * The plug list might get flushed before this. If that
1331 * happens, same_queue_rq is invalid and plug list is
1334 if (same_queue_rq
&& !list_empty(&plug
->mq_list
)) {
1335 old_rq
= same_queue_rq
;
1336 list_del_init(&old_rq
->queuelist
);
1338 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1339 } else /* is_sync */
1341 blk_mq_put_ctx(data
.ctx
);
1344 blk_mq_try_issue_directly(data
.hctx
, old_rq
, &cookie
);
1348 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1350 * For a SYNC request, send it to the hardware immediately. For
1351 * an ASYNC request, just ensure that we run it later on. The
1352 * latter allows for merging opportunities and more efficient
1356 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1358 blk_mq_put_ctx(data
.ctx
);
1364 * Single hardware queue variant. This will attempt to use any per-process
1365 * plug for merging and IO deferral.
1367 static blk_qc_t
blk_sq_make_request(struct request_queue
*q
, struct bio
*bio
)
1369 const int is_sync
= op_is_sync(bio
->bi_opf
);
1370 const int is_flush_fua
= bio
->bi_opf
& (REQ_PREFLUSH
| REQ_FUA
);
1371 struct blk_plug
*plug
;
1372 unsigned int request_count
= 0;
1373 struct blk_mq_alloc_data data
;
1377 blk_queue_bounce(q
, &bio
);
1379 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1381 return BLK_QC_T_NONE
;
1384 blk_queue_split(q
, &bio
, q
->bio_split
);
1386 if (!is_flush_fua
&& !blk_queue_nomerges(q
)) {
1387 if (blk_attempt_plug_merge(q
, bio
, &request_count
, NULL
))
1388 return BLK_QC_T_NONE
;
1390 request_count
= blk_plug_queued_count(q
);
1392 rq
= blk_mq_map_request(q
, bio
, &data
);
1394 return BLK_QC_T_NONE
;
1396 cookie
= blk_tag_to_qc_t(rq
->tag
, data
.hctx
->queue_num
);
1398 if (unlikely(is_flush_fua
)) {
1399 blk_mq_bio_to_request(rq
, bio
);
1400 blk_insert_flush(rq
);
1405 * A task plug currently exists. Since this is completely lockless,
1406 * utilize that to temporarily store requests until the task is
1407 * either done or scheduled away.
1409 plug
= current
->plug
;
1411 blk_mq_bio_to_request(rq
, bio
);
1413 trace_block_plug(q
);
1415 blk_mq_put_ctx(data
.ctx
);
1417 if (request_count
>= BLK_MAX_REQUEST_COUNT
) {
1418 blk_flush_plug_list(plug
, false);
1419 trace_block_plug(q
);
1422 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1426 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1428 * For a SYNC request, send it to the hardware immediately. For
1429 * an ASYNC request, just ensure that we run it later on. The
1430 * latter allows for merging opportunities and more efficient
1434 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1437 blk_mq_put_ctx(data
.ctx
);
1441 static void blk_mq_free_rq_map(struct blk_mq_tag_set
*set
,
1442 struct blk_mq_tags
*tags
, unsigned int hctx_idx
)
1446 if (tags
->rqs
&& set
->ops
->exit_request
) {
1449 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1452 set
->ops
->exit_request(set
->driver_data
, tags
->rqs
[i
],
1454 tags
->rqs
[i
] = NULL
;
1458 while (!list_empty(&tags
->page_list
)) {
1459 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1460 list_del_init(&page
->lru
);
1462 * Remove kmemleak object previously allocated in
1463 * blk_mq_init_rq_map().
1465 kmemleak_free(page_address(page
));
1466 __free_pages(page
, page
->private);
1471 blk_mq_free_tags(tags
);
1474 static size_t order_to_size(unsigned int order
)
1476 return (size_t)PAGE_SIZE
<< order
;
1479 static struct blk_mq_tags
*blk_mq_init_rq_map(struct blk_mq_tag_set
*set
,
1480 unsigned int hctx_idx
)
1482 struct blk_mq_tags
*tags
;
1483 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1484 size_t rq_size
, left
;
1486 tags
= blk_mq_init_tags(set
->queue_depth
, set
->reserved_tags
,
1488 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1492 INIT_LIST_HEAD(&tags
->page_list
);
1494 tags
->rqs
= kzalloc_node(set
->queue_depth
* sizeof(struct request
*),
1495 GFP_KERNEL
| __GFP_NOWARN
| __GFP_NORETRY
,
1498 blk_mq_free_tags(tags
);
1503 * rq_size is the size of the request plus driver payload, rounded
1504 * to the cacheline size
1506 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1508 left
= rq_size
* set
->queue_depth
;
1510 for (i
= 0; i
< set
->queue_depth
; ) {
1511 int this_order
= max_order
;
1516 while (this_order
&& left
< order_to_size(this_order
- 1))
1520 page
= alloc_pages_node(set
->numa_node
,
1521 GFP_KERNEL
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
1527 if (order_to_size(this_order
) < rq_size
)
1534 page
->private = this_order
;
1535 list_add_tail(&page
->lru
, &tags
->page_list
);
1537 p
= page_address(page
);
1539 * Allow kmemleak to scan these pages as they contain pointers
1540 * to additional allocations like via ops->init_request().
1542 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_KERNEL
);
1543 entries_per_page
= order_to_size(this_order
) / rq_size
;
1544 to_do
= min(entries_per_page
, set
->queue_depth
- i
);
1545 left
-= to_do
* rq_size
;
1546 for (j
= 0; j
< to_do
; j
++) {
1548 if (set
->ops
->init_request
) {
1549 if (set
->ops
->init_request(set
->driver_data
,
1550 tags
->rqs
[i
], hctx_idx
, i
,
1552 tags
->rqs
[i
] = NULL
;
1564 blk_mq_free_rq_map(set
, tags
, hctx_idx
);
1569 * 'cpu' is going away. splice any existing rq_list entries from this
1570 * software queue to the hw queue dispatch list, and ensure that it
1573 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
1575 struct blk_mq_hw_ctx
*hctx
;
1576 struct blk_mq_ctx
*ctx
;
1579 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
1580 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
1582 spin_lock(&ctx
->lock
);
1583 if (!list_empty(&ctx
->rq_list
)) {
1584 list_splice_init(&ctx
->rq_list
, &tmp
);
1585 blk_mq_hctx_clear_pending(hctx
, ctx
);
1587 spin_unlock(&ctx
->lock
);
1589 if (list_empty(&tmp
))
1592 spin_lock(&hctx
->lock
);
1593 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
1594 spin_unlock(&hctx
->lock
);
1596 blk_mq_run_hw_queue(hctx
, true);
1600 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
1602 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
1606 /* hctx->ctxs will be freed in queue's release handler */
1607 static void blk_mq_exit_hctx(struct request_queue
*q
,
1608 struct blk_mq_tag_set
*set
,
1609 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
1611 unsigned flush_start_tag
= set
->queue_depth
;
1613 blk_mq_tag_idle(hctx
);
1615 if (set
->ops
->exit_request
)
1616 set
->ops
->exit_request(set
->driver_data
,
1617 hctx
->fq
->flush_rq
, hctx_idx
,
1618 flush_start_tag
+ hctx_idx
);
1620 if (set
->ops
->exit_hctx
)
1621 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1623 blk_mq_remove_cpuhp(hctx
);
1624 blk_free_flush_queue(hctx
->fq
);
1625 sbitmap_free(&hctx
->ctx_map
);
1628 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1629 struct blk_mq_tag_set
*set
, int nr_queue
)
1631 struct blk_mq_hw_ctx
*hctx
;
1634 queue_for_each_hw_ctx(q
, hctx
, i
) {
1637 blk_mq_exit_hctx(q
, set
, hctx
, i
);
1641 static void blk_mq_free_hw_queues(struct request_queue
*q
,
1642 struct blk_mq_tag_set
*set
)
1644 struct blk_mq_hw_ctx
*hctx
;
1647 queue_for_each_hw_ctx(q
, hctx
, i
)
1648 free_cpumask_var(hctx
->cpumask
);
1651 static int blk_mq_init_hctx(struct request_queue
*q
,
1652 struct blk_mq_tag_set
*set
,
1653 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
1656 unsigned flush_start_tag
= set
->queue_depth
;
1658 node
= hctx
->numa_node
;
1659 if (node
== NUMA_NO_NODE
)
1660 node
= hctx
->numa_node
= set
->numa_node
;
1662 INIT_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1663 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1664 spin_lock_init(&hctx
->lock
);
1665 INIT_LIST_HEAD(&hctx
->dispatch
);
1667 hctx
->queue_num
= hctx_idx
;
1668 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
1670 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
1672 hctx
->tags
= set
->tags
[hctx_idx
];
1675 * Allocate space for all possible cpus to avoid allocation at
1678 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1681 goto unregister_cpu_notifier
;
1683 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8), GFP_KERNEL
,
1689 if (set
->ops
->init_hctx
&&
1690 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
1693 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
1697 if (set
->ops
->init_request
&&
1698 set
->ops
->init_request(set
->driver_data
,
1699 hctx
->fq
->flush_rq
, hctx_idx
,
1700 flush_start_tag
+ hctx_idx
, node
))
1708 if (set
->ops
->exit_hctx
)
1709 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1711 sbitmap_free(&hctx
->ctx_map
);
1714 unregister_cpu_notifier
:
1715 blk_mq_remove_cpuhp(hctx
);
1719 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1720 unsigned int nr_hw_queues
)
1724 for_each_possible_cpu(i
) {
1725 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1726 struct blk_mq_hw_ctx
*hctx
;
1728 memset(__ctx
, 0, sizeof(*__ctx
));
1730 spin_lock_init(&__ctx
->lock
);
1731 INIT_LIST_HEAD(&__ctx
->rq_list
);
1734 /* If the cpu isn't online, the cpu is mapped to first hctx */
1738 hctx
= blk_mq_map_queue(q
, i
);
1741 * Set local node, IFF we have more than one hw queue. If
1742 * not, we remain on the home node of the device
1744 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1745 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
1749 static void blk_mq_map_swqueue(struct request_queue
*q
,
1750 const struct cpumask
*online_mask
)
1753 struct blk_mq_hw_ctx
*hctx
;
1754 struct blk_mq_ctx
*ctx
;
1755 struct blk_mq_tag_set
*set
= q
->tag_set
;
1758 * Avoid others reading imcomplete hctx->cpumask through sysfs
1760 mutex_lock(&q
->sysfs_lock
);
1762 queue_for_each_hw_ctx(q
, hctx
, i
) {
1763 cpumask_clear(hctx
->cpumask
);
1768 * Map software to hardware queues
1770 for_each_possible_cpu(i
) {
1771 /* If the cpu isn't online, the cpu is mapped to first hctx */
1772 if (!cpumask_test_cpu(i
, online_mask
))
1775 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1776 hctx
= blk_mq_map_queue(q
, i
);
1778 cpumask_set_cpu(i
, hctx
->cpumask
);
1779 ctx
->index_hw
= hctx
->nr_ctx
;
1780 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
1783 mutex_unlock(&q
->sysfs_lock
);
1785 queue_for_each_hw_ctx(q
, hctx
, i
) {
1787 * If no software queues are mapped to this hardware queue,
1788 * disable it and free the request entries.
1790 if (!hctx
->nr_ctx
) {
1792 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1793 set
->tags
[i
] = NULL
;
1799 /* unmapped hw queue can be remapped after CPU topo changed */
1801 set
->tags
[i
] = blk_mq_init_rq_map(set
, i
);
1802 hctx
->tags
= set
->tags
[i
];
1803 WARN_ON(!hctx
->tags
);
1806 * Set the map size to the number of mapped software queues.
1807 * This is more accurate and more efficient than looping
1808 * over all possibly mapped software queues.
1810 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
1813 * Initialize batch roundrobin counts
1815 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
1816 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1820 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
1822 struct blk_mq_hw_ctx
*hctx
;
1825 queue_for_each_hw_ctx(q
, hctx
, i
) {
1827 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
1829 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
1833 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
, bool shared
)
1835 struct request_queue
*q
;
1837 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
1838 blk_mq_freeze_queue(q
);
1839 queue_set_hctx_shared(q
, shared
);
1840 blk_mq_unfreeze_queue(q
);
1844 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
1846 struct blk_mq_tag_set
*set
= q
->tag_set
;
1848 mutex_lock(&set
->tag_list_lock
);
1849 list_del_init(&q
->tag_set_list
);
1850 if (list_is_singular(&set
->tag_list
)) {
1851 /* just transitioned to unshared */
1852 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
1853 /* update existing queue */
1854 blk_mq_update_tag_set_depth(set
, false);
1856 mutex_unlock(&set
->tag_list_lock
);
1859 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
1860 struct request_queue
*q
)
1864 mutex_lock(&set
->tag_list_lock
);
1866 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
1867 if (!list_empty(&set
->tag_list
) && !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
1868 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
1869 /* update existing queue */
1870 blk_mq_update_tag_set_depth(set
, true);
1872 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
1873 queue_set_hctx_shared(q
, true);
1874 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
1876 mutex_unlock(&set
->tag_list_lock
);
1880 * It is the actual release handler for mq, but we do it from
1881 * request queue's release handler for avoiding use-after-free
1882 * and headache because q->mq_kobj shouldn't have been introduced,
1883 * but we can't group ctx/kctx kobj without it.
1885 void blk_mq_release(struct request_queue
*q
)
1887 struct blk_mq_hw_ctx
*hctx
;
1890 /* hctx kobj stays in hctx */
1891 queue_for_each_hw_ctx(q
, hctx
, i
) {
1900 kfree(q
->queue_hw_ctx
);
1902 /* ctx kobj stays in queue_ctx */
1903 free_percpu(q
->queue_ctx
);
1906 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
1908 struct request_queue
*uninit_q
, *q
;
1910 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
1912 return ERR_PTR(-ENOMEM
);
1914 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
1916 blk_cleanup_queue(uninit_q
);
1920 EXPORT_SYMBOL(blk_mq_init_queue
);
1922 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
1923 struct request_queue
*q
)
1926 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
1928 blk_mq_sysfs_unregister(q
);
1929 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1935 node
= blk_mq_hw_queue_to_node(q
->mq_map
, i
);
1936 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
1941 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
1948 atomic_set(&hctxs
[i
]->nr_active
, 0);
1949 hctxs
[i
]->numa_node
= node
;
1950 hctxs
[i
]->queue_num
= i
;
1952 if (blk_mq_init_hctx(q
, set
, hctxs
[i
], i
)) {
1953 free_cpumask_var(hctxs
[i
]->cpumask
);
1958 blk_mq_hctx_kobj_init(hctxs
[i
]);
1960 for (j
= i
; j
< q
->nr_hw_queues
; j
++) {
1961 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
1965 blk_mq_free_rq_map(set
, hctx
->tags
, j
);
1966 set
->tags
[j
] = NULL
;
1968 blk_mq_exit_hctx(q
, set
, hctx
, j
);
1969 free_cpumask_var(hctx
->cpumask
);
1970 kobject_put(&hctx
->kobj
);
1977 q
->nr_hw_queues
= i
;
1978 blk_mq_sysfs_register(q
);
1981 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
1982 struct request_queue
*q
)
1984 /* mark the queue as mq asap */
1985 q
->mq_ops
= set
->ops
;
1987 q
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
1991 q
->queue_hw_ctx
= kzalloc_node(nr_cpu_ids
* sizeof(*(q
->queue_hw_ctx
)),
1992 GFP_KERNEL
, set
->numa_node
);
1993 if (!q
->queue_hw_ctx
)
1996 q
->mq_map
= set
->mq_map
;
1998 blk_mq_realloc_hw_ctxs(set
, q
);
1999 if (!q
->nr_hw_queues
)
2002 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2003 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2005 q
->nr_queues
= nr_cpu_ids
;
2007 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2009 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2010 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
2012 q
->sg_reserved_size
= INT_MAX
;
2014 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2015 INIT_LIST_HEAD(&q
->requeue_list
);
2016 spin_lock_init(&q
->requeue_lock
);
2018 if (q
->nr_hw_queues
> 1)
2019 blk_queue_make_request(q
, blk_mq_make_request
);
2021 blk_queue_make_request(q
, blk_sq_make_request
);
2024 * Do this after blk_queue_make_request() overrides it...
2026 q
->nr_requests
= set
->queue_depth
;
2028 if (set
->ops
->complete
)
2029 blk_queue_softirq_done(q
, set
->ops
->complete
);
2031 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2034 mutex_lock(&all_q_mutex
);
2036 list_add_tail(&q
->all_q_node
, &all_q_list
);
2037 blk_mq_add_queue_tag_set(set
, q
);
2038 blk_mq_map_swqueue(q
, cpu_online_mask
);
2040 mutex_unlock(&all_q_mutex
);
2046 kfree(q
->queue_hw_ctx
);
2048 free_percpu(q
->queue_ctx
);
2051 return ERR_PTR(-ENOMEM
);
2053 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2055 void blk_mq_free_queue(struct request_queue
*q
)
2057 struct blk_mq_tag_set
*set
= q
->tag_set
;
2059 mutex_lock(&all_q_mutex
);
2060 list_del_init(&q
->all_q_node
);
2061 mutex_unlock(&all_q_mutex
);
2063 blk_mq_del_queue_tag_set(q
);
2065 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2066 blk_mq_free_hw_queues(q
, set
);
2069 /* Basically redo blk_mq_init_queue with queue frozen */
2070 static void blk_mq_queue_reinit(struct request_queue
*q
,
2071 const struct cpumask
*online_mask
)
2073 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2075 blk_mq_sysfs_unregister(q
);
2078 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2079 * we should change hctx numa_node according to new topology (this
2080 * involves free and re-allocate memory, worthy doing?)
2083 blk_mq_map_swqueue(q
, online_mask
);
2085 blk_mq_sysfs_register(q
);
2089 * New online cpumask which is going to be set in this hotplug event.
2090 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2091 * one-by-one and dynamically allocating this could result in a failure.
2093 static struct cpumask cpuhp_online_new
;
2095 static void blk_mq_queue_reinit_work(void)
2097 struct request_queue
*q
;
2099 mutex_lock(&all_q_mutex
);
2101 * We need to freeze and reinit all existing queues. Freezing
2102 * involves synchronous wait for an RCU grace period and doing it
2103 * one by one may take a long time. Start freezing all queues in
2104 * one swoop and then wait for the completions so that freezing can
2105 * take place in parallel.
2107 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2108 blk_mq_freeze_queue_start(q
);
2109 list_for_each_entry(q
, &all_q_list
, all_q_node
) {
2110 blk_mq_freeze_queue_wait(q
);
2113 * timeout handler can't touch hw queue during the
2116 del_timer_sync(&q
->timeout
);
2119 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2120 blk_mq_queue_reinit(q
, &cpuhp_online_new
);
2122 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2123 blk_mq_unfreeze_queue(q
);
2125 mutex_unlock(&all_q_mutex
);
2128 static int blk_mq_queue_reinit_dead(unsigned int cpu
)
2130 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2131 blk_mq_queue_reinit_work();
2136 * Before hotadded cpu starts handling requests, new mappings must be
2137 * established. Otherwise, these requests in hw queue might never be
2140 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2141 * for CPU0, and ctx1 for CPU1).
2143 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2144 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2146 * And then while running hw queue, flush_busy_ctxs() finds bit0 is set in
2147 * pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2148 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list
2151 static int blk_mq_queue_reinit_prepare(unsigned int cpu
)
2153 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2154 cpumask_set_cpu(cpu
, &cpuhp_online_new
);
2155 blk_mq_queue_reinit_work();
2159 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2163 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2164 set
->tags
[i
] = blk_mq_init_rq_map(set
, i
);
2173 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2179 * Allocate the request maps associated with this tag_set. Note that this
2180 * may reduce the depth asked for, if memory is tight. set->queue_depth
2181 * will be updated to reflect the allocated depth.
2183 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2188 depth
= set
->queue_depth
;
2190 err
= __blk_mq_alloc_rq_maps(set
);
2194 set
->queue_depth
>>= 1;
2195 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2199 } while (set
->queue_depth
);
2201 if (!set
->queue_depth
|| err
) {
2202 pr_err("blk-mq: failed to allocate request map\n");
2206 if (depth
!= set
->queue_depth
)
2207 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2208 depth
, set
->queue_depth
);
2214 * Alloc a tag set to be associated with one or more request queues.
2215 * May fail with EINVAL for various error conditions. May adjust the
2216 * requested depth down, if if it too large. In that case, the set
2217 * value will be stored in set->queue_depth.
2219 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2223 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2225 if (!set
->nr_hw_queues
)
2227 if (!set
->queue_depth
)
2229 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2232 if (!set
->ops
->queue_rq
)
2235 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2236 pr_info("blk-mq: reduced tag depth to %u\n",
2238 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2242 * If a crashdump is active, then we are potentially in a very
2243 * memory constrained environment. Limit us to 1 queue and
2244 * 64 tags to prevent using too much memory.
2246 if (is_kdump_kernel()) {
2247 set
->nr_hw_queues
= 1;
2248 set
->queue_depth
= min(64U, set
->queue_depth
);
2251 * There is no use for more h/w queues than cpus.
2253 if (set
->nr_hw_queues
> nr_cpu_ids
)
2254 set
->nr_hw_queues
= nr_cpu_ids
;
2256 set
->tags
= kzalloc_node(nr_cpu_ids
* sizeof(struct blk_mq_tags
*),
2257 GFP_KERNEL
, set
->numa_node
);
2262 set
->mq_map
= kzalloc_node(sizeof(*set
->mq_map
) * nr_cpu_ids
,
2263 GFP_KERNEL
, set
->numa_node
);
2267 if (set
->ops
->map_queues
)
2268 ret
= set
->ops
->map_queues(set
);
2270 ret
= blk_mq_map_queues(set
);
2272 goto out_free_mq_map
;
2274 ret
= blk_mq_alloc_rq_maps(set
);
2276 goto out_free_mq_map
;
2278 mutex_init(&set
->tag_list_lock
);
2279 INIT_LIST_HEAD(&set
->tag_list
);
2291 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2293 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2297 for (i
= 0; i
< nr_cpu_ids
; i
++) {
2299 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2308 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2310 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2312 struct blk_mq_tag_set
*set
= q
->tag_set
;
2313 struct blk_mq_hw_ctx
*hctx
;
2316 if (!set
|| nr
> set
->queue_depth
)
2320 queue_for_each_hw_ctx(q
, hctx
, i
) {
2323 ret
= blk_mq_tag_update_depth(hctx
->tags
, nr
);
2329 q
->nr_requests
= nr
;
2334 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
2336 struct request_queue
*q
;
2338 if (nr_hw_queues
> nr_cpu_ids
)
2339 nr_hw_queues
= nr_cpu_ids
;
2340 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
2343 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2344 blk_mq_freeze_queue(q
);
2346 set
->nr_hw_queues
= nr_hw_queues
;
2347 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2348 blk_mq_realloc_hw_ctxs(set
, q
);
2350 if (q
->nr_hw_queues
> 1)
2351 blk_queue_make_request(q
, blk_mq_make_request
);
2353 blk_queue_make_request(q
, blk_sq_make_request
);
2355 blk_mq_queue_reinit(q
, cpu_online_mask
);
2358 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2359 blk_mq_unfreeze_queue(q
);
2361 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
2363 void blk_mq_disable_hotplug(void)
2365 mutex_lock(&all_q_mutex
);
2368 void blk_mq_enable_hotplug(void)
2370 mutex_unlock(&all_q_mutex
);
2373 static int __init
blk_mq_init(void)
2375 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
2376 blk_mq_hctx_notify_dead
);
2378 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE
, "block/mq:prepare",
2379 blk_mq_queue_reinit_prepare
,
2380 blk_mq_queue_reinit_dead
);
2383 subsys_initcall(blk_mq_init
);