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>
26 #include <trace/events/block.h>
28 #include <linux/blk-mq.h>
31 #include "blk-mq-tag.h"
33 static DEFINE_MUTEX(all_q_mutex
);
34 static LIST_HEAD(all_q_list
);
36 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
);
39 * Check if any of the ctx's have pending work in this hardware queue
41 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
45 for (i
= 0; i
< hctx
->ctx_map
.size
; i
++)
46 if (hctx
->ctx_map
.map
[i
].word
)
52 static inline struct blk_align_bitmap
*get_bm(struct blk_mq_hw_ctx
*hctx
,
53 struct blk_mq_ctx
*ctx
)
55 return &hctx
->ctx_map
.map
[ctx
->index_hw
/ hctx
->ctx_map
.bits_per_word
];
58 #define CTX_TO_BIT(hctx, ctx) \
59 ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
62 * Mark this ctx as having pending work in this hardware queue
64 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
65 struct blk_mq_ctx
*ctx
)
67 struct blk_align_bitmap
*bm
= get_bm(hctx
, ctx
);
69 if (!test_bit(CTX_TO_BIT(hctx
, ctx
), &bm
->word
))
70 set_bit(CTX_TO_BIT(hctx
, ctx
), &bm
->word
);
73 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
74 struct blk_mq_ctx
*ctx
)
76 struct blk_align_bitmap
*bm
= get_bm(hctx
, ctx
);
78 clear_bit(CTX_TO_BIT(hctx
, ctx
), &bm
->word
);
81 void blk_mq_freeze_queue_start(struct request_queue
*q
)
85 freeze_depth
= atomic_inc_return(&q
->mq_freeze_depth
);
86 if (freeze_depth
== 1) {
87 percpu_ref_kill(&q
->q_usage_counter
);
88 blk_mq_run_hw_queues(q
, false);
91 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start
);
93 static void blk_mq_freeze_queue_wait(struct request_queue
*q
)
95 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
99 * Guarantee no request is in use, so we can change any data structure of
100 * the queue afterward.
102 void blk_freeze_queue(struct request_queue
*q
)
105 * In the !blk_mq case we are only calling this to kill the
106 * q_usage_counter, otherwise this increases the freeze depth
107 * and waits for it to return to zero. For this reason there is
108 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
109 * exported to drivers as the only user for unfreeze is blk_mq.
111 blk_mq_freeze_queue_start(q
);
112 blk_mq_freeze_queue_wait(q
);
115 void blk_mq_freeze_queue(struct request_queue
*q
)
118 * ...just an alias to keep freeze and unfreeze actions balanced
119 * in the blk_mq_* namespace
123 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
125 void blk_mq_unfreeze_queue(struct request_queue
*q
)
129 freeze_depth
= atomic_dec_return(&q
->mq_freeze_depth
);
130 WARN_ON_ONCE(freeze_depth
< 0);
132 percpu_ref_reinit(&q
->q_usage_counter
);
133 wake_up_all(&q
->mq_freeze_wq
);
136 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
138 void blk_mq_wake_waiters(struct request_queue
*q
)
140 struct blk_mq_hw_ctx
*hctx
;
143 queue_for_each_hw_ctx(q
, hctx
, i
)
144 if (blk_mq_hw_queue_mapped(hctx
))
145 blk_mq_tag_wakeup_all(hctx
->tags
, true);
148 * If we are called because the queue has now been marked as
149 * dying, we need to ensure that processes currently waiting on
150 * the queue are notified as well.
152 wake_up_all(&q
->mq_freeze_wq
);
155 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
157 return blk_mq_has_free_tags(hctx
->tags
);
159 EXPORT_SYMBOL(blk_mq_can_queue
);
161 static void blk_mq_rq_ctx_init(struct request_queue
*q
, struct blk_mq_ctx
*ctx
,
162 struct request
*rq
, unsigned int rw_flags
)
164 if (blk_queue_io_stat(q
))
165 rw_flags
|= REQ_IO_STAT
;
167 INIT_LIST_HEAD(&rq
->queuelist
);
168 /* csd/requeue_work/fifo_time is initialized before use */
171 rq
->cmd_flags
|= rw_flags
;
172 /* do not touch atomic flags, it needs atomic ops against the timer */
174 INIT_HLIST_NODE(&rq
->hash
);
175 RB_CLEAR_NODE(&rq
->rb_node
);
178 rq
->start_time
= jiffies
;
179 #ifdef CONFIG_BLK_CGROUP
181 set_start_time_ns(rq
);
182 rq
->io_start_time_ns
= 0;
184 rq
->nr_phys_segments
= 0;
185 #if defined(CONFIG_BLK_DEV_INTEGRITY)
186 rq
->nr_integrity_segments
= 0;
189 /* tag was already set */
199 INIT_LIST_HEAD(&rq
->timeout_list
);
203 rq
->end_io_data
= NULL
;
206 ctx
->rq_dispatched
[rw_is_sync(rw_flags
)]++;
209 static struct request
*
210 __blk_mq_alloc_request(struct blk_mq_alloc_data
*data
, int rw
)
215 tag
= blk_mq_get_tag(data
);
216 if (tag
!= BLK_MQ_TAG_FAIL
) {
217 rq
= data
->hctx
->tags
->rqs
[tag
];
219 if (blk_mq_tag_busy(data
->hctx
)) {
220 rq
->cmd_flags
= REQ_MQ_INFLIGHT
;
221 atomic_inc(&data
->hctx
->nr_active
);
225 blk_mq_rq_ctx_init(data
->q
, data
->ctx
, rq
, rw
);
232 struct request
*blk_mq_alloc_request(struct request_queue
*q
, int rw
,
235 struct blk_mq_ctx
*ctx
;
236 struct blk_mq_hw_ctx
*hctx
;
238 struct blk_mq_alloc_data alloc_data
;
241 ret
= blk_queue_enter(q
, flags
& BLK_MQ_REQ_NOWAIT
);
245 ctx
= blk_mq_get_ctx(q
);
246 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
247 blk_mq_set_alloc_data(&alloc_data
, q
, flags
, ctx
, hctx
);
249 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
250 if (!rq
&& !(flags
& BLK_MQ_REQ_NOWAIT
)) {
251 __blk_mq_run_hw_queue(hctx
);
254 ctx
= blk_mq_get_ctx(q
);
255 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
256 blk_mq_set_alloc_data(&alloc_data
, q
, flags
, ctx
, hctx
);
257 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
258 ctx
= alloc_data
.ctx
;
263 return ERR_PTR(-EWOULDBLOCK
);
267 EXPORT_SYMBOL(blk_mq_alloc_request
);
269 static void __blk_mq_free_request(struct blk_mq_hw_ctx
*hctx
,
270 struct blk_mq_ctx
*ctx
, struct request
*rq
)
272 const int tag
= rq
->tag
;
273 struct request_queue
*q
= rq
->q
;
275 if (rq
->cmd_flags
& REQ_MQ_INFLIGHT
)
276 atomic_dec(&hctx
->nr_active
);
279 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
280 blk_mq_put_tag(hctx
, tag
, &ctx
->last_tag
);
284 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
286 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
288 ctx
->rq_completed
[rq_is_sync(rq
)]++;
289 __blk_mq_free_request(hctx
, ctx
, rq
);
292 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request
);
294 void blk_mq_free_request(struct request
*rq
)
296 struct blk_mq_hw_ctx
*hctx
;
297 struct request_queue
*q
= rq
->q
;
299 hctx
= q
->mq_ops
->map_queue(q
, rq
->mq_ctx
->cpu
);
300 blk_mq_free_hctx_request(hctx
, rq
);
302 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
304 inline void __blk_mq_end_request(struct request
*rq
, int error
)
306 blk_account_io_done(rq
);
309 rq
->end_io(rq
, error
);
311 if (unlikely(blk_bidi_rq(rq
)))
312 blk_mq_free_request(rq
->next_rq
);
313 blk_mq_free_request(rq
);
316 EXPORT_SYMBOL(__blk_mq_end_request
);
318 void blk_mq_end_request(struct request
*rq
, int error
)
320 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
322 __blk_mq_end_request(rq
, error
);
324 EXPORT_SYMBOL(blk_mq_end_request
);
326 static void __blk_mq_complete_request_remote(void *data
)
328 struct request
*rq
= data
;
330 rq
->q
->softirq_done_fn(rq
);
333 static void blk_mq_ipi_complete_request(struct request
*rq
)
335 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
339 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
340 rq
->q
->softirq_done_fn(rq
);
345 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
346 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
348 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
349 rq
->csd
.func
= __blk_mq_complete_request_remote
;
352 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
354 rq
->q
->softirq_done_fn(rq
);
359 static void __blk_mq_complete_request(struct request
*rq
)
361 struct request_queue
*q
= rq
->q
;
363 if (!q
->softirq_done_fn
)
364 blk_mq_end_request(rq
, rq
->errors
);
366 blk_mq_ipi_complete_request(rq
);
370 * blk_mq_complete_request - end I/O on a request
371 * @rq: the request being processed
374 * Ends all I/O on a request. It does not handle partial completions.
375 * The actual completion happens out-of-order, through a IPI handler.
377 void blk_mq_complete_request(struct request
*rq
, int error
)
379 struct request_queue
*q
= rq
->q
;
381 if (unlikely(blk_should_fake_timeout(q
)))
383 if (!blk_mark_rq_complete(rq
)) {
385 __blk_mq_complete_request(rq
);
388 EXPORT_SYMBOL(blk_mq_complete_request
);
390 int blk_mq_request_started(struct request
*rq
)
392 return test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
394 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
396 void blk_mq_start_request(struct request
*rq
)
398 struct request_queue
*q
= rq
->q
;
400 trace_block_rq_issue(q
, rq
);
402 rq
->resid_len
= blk_rq_bytes(rq
);
403 if (unlikely(blk_bidi_rq(rq
)))
404 rq
->next_rq
->resid_len
= blk_rq_bytes(rq
->next_rq
);
409 * Ensure that ->deadline is visible before set the started
410 * flag and clear the completed flag.
412 smp_mb__before_atomic();
415 * Mark us as started and clear complete. Complete might have been
416 * set if requeue raced with timeout, which then marked it as
417 * complete. So be sure to clear complete again when we start
418 * the request, otherwise we'll ignore the completion event.
420 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
421 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
422 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
423 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
425 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
427 * Make sure space for the drain appears. We know we can do
428 * this because max_hw_segments has been adjusted to be one
429 * fewer than the device can handle.
431 rq
->nr_phys_segments
++;
434 EXPORT_SYMBOL(blk_mq_start_request
);
436 static void __blk_mq_requeue_request(struct request
*rq
)
438 struct request_queue
*q
= rq
->q
;
440 trace_block_rq_requeue(q
, rq
);
442 if (test_and_clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
443 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
444 rq
->nr_phys_segments
--;
448 void blk_mq_requeue_request(struct request
*rq
)
450 __blk_mq_requeue_request(rq
);
452 BUG_ON(blk_queued_rq(rq
));
453 blk_mq_add_to_requeue_list(rq
, true);
455 EXPORT_SYMBOL(blk_mq_requeue_request
);
457 static void blk_mq_requeue_work(struct work_struct
*work
)
459 struct request_queue
*q
=
460 container_of(work
, struct request_queue
, requeue_work
);
462 struct request
*rq
, *next
;
465 spin_lock_irqsave(&q
->requeue_lock
, flags
);
466 list_splice_init(&q
->requeue_list
, &rq_list
);
467 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
469 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
470 if (!(rq
->cmd_flags
& REQ_SOFTBARRIER
))
473 rq
->cmd_flags
&= ~REQ_SOFTBARRIER
;
474 list_del_init(&rq
->queuelist
);
475 blk_mq_insert_request(rq
, true, false, false);
478 while (!list_empty(&rq_list
)) {
479 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
480 list_del_init(&rq
->queuelist
);
481 blk_mq_insert_request(rq
, false, false, false);
485 * Use the start variant of queue running here, so that running
486 * the requeue work will kick stopped queues.
488 blk_mq_start_hw_queues(q
);
491 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
)
493 struct request_queue
*q
= rq
->q
;
497 * We abuse this flag that is otherwise used by the I/O scheduler to
498 * request head insertation from the workqueue.
500 BUG_ON(rq
->cmd_flags
& REQ_SOFTBARRIER
);
502 spin_lock_irqsave(&q
->requeue_lock
, flags
);
504 rq
->cmd_flags
|= REQ_SOFTBARRIER
;
505 list_add(&rq
->queuelist
, &q
->requeue_list
);
507 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
509 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
511 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
513 void blk_mq_cancel_requeue_work(struct request_queue
*q
)
515 cancel_work_sync(&q
->requeue_work
);
517 EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work
);
519 void blk_mq_kick_requeue_list(struct request_queue
*q
)
521 kblockd_schedule_work(&q
->requeue_work
);
523 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
525 void blk_mq_abort_requeue_list(struct request_queue
*q
)
530 spin_lock_irqsave(&q
->requeue_lock
, flags
);
531 list_splice_init(&q
->requeue_list
, &rq_list
);
532 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
534 while (!list_empty(&rq_list
)) {
537 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
538 list_del_init(&rq
->queuelist
);
540 blk_mq_end_request(rq
, rq
->errors
);
543 EXPORT_SYMBOL(blk_mq_abort_requeue_list
);
545 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
547 return tags
->rqs
[tag
];
549 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
551 struct blk_mq_timeout_data
{
553 unsigned int next_set
;
556 void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
558 struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
559 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
562 * We know that complete is set at this point. If STARTED isn't set
563 * anymore, then the request isn't active and the "timeout" should
564 * just be ignored. This can happen due to the bitflag ordering.
565 * Timeout first checks if STARTED is set, and if it is, assumes
566 * the request is active. But if we race with completion, then
567 * we both flags will get cleared. So check here again, and ignore
568 * a timeout event with a request that isn't active.
570 if (!test_bit(REQ_ATOM_STARTED
, &req
->atomic_flags
))
574 ret
= ops
->timeout(req
, reserved
);
578 __blk_mq_complete_request(req
);
580 case BLK_EH_RESET_TIMER
:
582 blk_clear_rq_complete(req
);
584 case BLK_EH_NOT_HANDLED
:
587 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
592 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
593 struct request
*rq
, void *priv
, bool reserved
)
595 struct blk_mq_timeout_data
*data
= priv
;
597 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
599 * If a request wasn't started before the queue was
600 * marked dying, kill it here or it'll go unnoticed.
602 if (unlikely(blk_queue_dying(rq
->q
)))
603 blk_mq_complete_request(rq
, -EIO
);
607 if (time_after_eq(jiffies
, rq
->deadline
)) {
608 if (!blk_mark_rq_complete(rq
))
609 blk_mq_rq_timed_out(rq
, reserved
);
610 } else if (!data
->next_set
|| time_after(data
->next
, rq
->deadline
)) {
611 data
->next
= rq
->deadline
;
616 static void blk_mq_timeout_work(struct work_struct
*work
)
618 struct request_queue
*q
=
619 container_of(work
, struct request_queue
, timeout_work
);
620 struct blk_mq_timeout_data data
= {
626 if (blk_queue_enter(q
, true))
629 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &data
);
632 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
633 mod_timer(&q
->timeout
, data
.next
);
635 struct blk_mq_hw_ctx
*hctx
;
637 queue_for_each_hw_ctx(q
, hctx
, i
) {
638 /* the hctx may be unmapped, so check it here */
639 if (blk_mq_hw_queue_mapped(hctx
))
640 blk_mq_tag_idle(hctx
);
647 * Reverse check our software queue for entries that we could potentially
648 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
649 * too much time checking for merges.
651 static bool blk_mq_attempt_merge(struct request_queue
*q
,
652 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
657 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
663 if (!blk_rq_merge_ok(rq
, bio
))
666 el_ret
= blk_try_merge(rq
, bio
);
667 if (el_ret
== ELEVATOR_BACK_MERGE
) {
668 if (bio_attempt_back_merge(q
, rq
, bio
)) {
673 } else if (el_ret
== ELEVATOR_FRONT_MERGE
) {
674 if (bio_attempt_front_merge(q
, rq
, bio
)) {
686 * Process software queues that have been marked busy, splicing them
687 * to the for-dispatch
689 static void flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
691 struct blk_mq_ctx
*ctx
;
694 for (i
= 0; i
< hctx
->ctx_map
.size
; i
++) {
695 struct blk_align_bitmap
*bm
= &hctx
->ctx_map
.map
[i
];
696 unsigned int off
, bit
;
702 off
= i
* hctx
->ctx_map
.bits_per_word
;
704 bit
= find_next_bit(&bm
->word
, bm
->depth
, bit
);
705 if (bit
>= bm
->depth
)
708 ctx
= hctx
->ctxs
[bit
+ off
];
709 clear_bit(bit
, &bm
->word
);
710 spin_lock(&ctx
->lock
);
711 list_splice_tail_init(&ctx
->rq_list
, list
);
712 spin_unlock(&ctx
->lock
);
720 * Run this hardware queue, pulling any software queues mapped to it in.
721 * Note that this function currently has various problems around ordering
722 * of IO. In particular, we'd like FIFO behaviour on handling existing
723 * items on the hctx->dispatch list. Ignore that for now.
725 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
727 struct request_queue
*q
= hctx
->queue
;
730 LIST_HEAD(driver_list
);
731 struct list_head
*dptr
;
734 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
));
736 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)))
742 * Touch any software queue that has pending entries.
744 flush_busy_ctxs(hctx
, &rq_list
);
747 * If we have previous entries on our dispatch list, grab them
748 * and stuff them at the front for more fair dispatch.
750 if (!list_empty_careful(&hctx
->dispatch
)) {
751 spin_lock(&hctx
->lock
);
752 if (!list_empty(&hctx
->dispatch
))
753 list_splice_init(&hctx
->dispatch
, &rq_list
);
754 spin_unlock(&hctx
->lock
);
758 * Start off with dptr being NULL, so we start the first request
759 * immediately, even if we have more pending.
764 * Now process all the entries, sending them to the driver.
767 while (!list_empty(&rq_list
)) {
768 struct blk_mq_queue_data bd
;
771 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
772 list_del_init(&rq
->queuelist
);
776 bd
.last
= list_empty(&rq_list
);
778 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
780 case BLK_MQ_RQ_QUEUE_OK
:
783 case BLK_MQ_RQ_QUEUE_BUSY
:
784 list_add(&rq
->queuelist
, &rq_list
);
785 __blk_mq_requeue_request(rq
);
788 pr_err("blk-mq: bad return on queue: %d\n", ret
);
789 case BLK_MQ_RQ_QUEUE_ERROR
:
791 blk_mq_end_request(rq
, rq
->errors
);
795 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
799 * We've done the first request. If we have more than 1
800 * left in the list, set dptr to defer issue.
802 if (!dptr
&& rq_list
.next
!= rq_list
.prev
)
807 hctx
->dispatched
[0]++;
808 else if (queued
< (1 << (BLK_MQ_MAX_DISPATCH_ORDER
- 1)))
809 hctx
->dispatched
[ilog2(queued
) + 1]++;
812 * Any items that need requeuing? Stuff them into hctx->dispatch,
813 * that is where we will continue on next queue run.
815 if (!list_empty(&rq_list
)) {
816 spin_lock(&hctx
->lock
);
817 list_splice(&rq_list
, &hctx
->dispatch
);
818 spin_unlock(&hctx
->lock
);
820 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
821 * it's possible the queue is stopped and restarted again
822 * before this. Queue restart will dispatch requests. And since
823 * requests in rq_list aren't added into hctx->dispatch yet,
824 * the requests in rq_list might get lost.
826 * blk_mq_run_hw_queue() already checks the STOPPED bit
828 blk_mq_run_hw_queue(hctx
, true);
833 * It'd be great if the workqueue API had a way to pass
834 * in a mask and had some smarts for more clever placement.
835 * For now we just round-robin here, switching for every
836 * BLK_MQ_CPU_WORK_BATCH queued items.
838 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
840 if (hctx
->queue
->nr_hw_queues
== 1)
841 return WORK_CPU_UNBOUND
;
843 if (--hctx
->next_cpu_batch
<= 0) {
844 int cpu
= hctx
->next_cpu
, next_cpu
;
846 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
847 if (next_cpu
>= nr_cpu_ids
)
848 next_cpu
= cpumask_first(hctx
->cpumask
);
850 hctx
->next_cpu
= next_cpu
;
851 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
856 return hctx
->next_cpu
;
859 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
861 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
) ||
862 !blk_mq_hw_queue_mapped(hctx
)))
867 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
868 __blk_mq_run_hw_queue(hctx
);
876 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
880 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
882 struct blk_mq_hw_ctx
*hctx
;
885 queue_for_each_hw_ctx(q
, hctx
, i
) {
886 if ((!blk_mq_hctx_has_pending(hctx
) &&
887 list_empty_careful(&hctx
->dispatch
)) ||
888 test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
891 blk_mq_run_hw_queue(hctx
, async
);
894 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
896 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
898 cancel_delayed_work(&hctx
->run_work
);
899 cancel_delayed_work(&hctx
->delay_work
);
900 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
902 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
904 void blk_mq_stop_hw_queues(struct request_queue
*q
)
906 struct blk_mq_hw_ctx
*hctx
;
909 queue_for_each_hw_ctx(q
, hctx
, i
)
910 blk_mq_stop_hw_queue(hctx
);
912 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
914 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
916 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
918 blk_mq_run_hw_queue(hctx
, false);
920 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
922 void blk_mq_start_hw_queues(struct request_queue
*q
)
924 struct blk_mq_hw_ctx
*hctx
;
927 queue_for_each_hw_ctx(q
, hctx
, i
)
928 blk_mq_start_hw_queue(hctx
);
930 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
932 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
934 struct blk_mq_hw_ctx
*hctx
;
937 queue_for_each_hw_ctx(q
, hctx
, i
) {
938 if (!test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
941 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
942 blk_mq_run_hw_queue(hctx
, async
);
945 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
947 static void blk_mq_run_work_fn(struct work_struct
*work
)
949 struct blk_mq_hw_ctx
*hctx
;
951 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
953 __blk_mq_run_hw_queue(hctx
);
956 static void blk_mq_delay_work_fn(struct work_struct
*work
)
958 struct blk_mq_hw_ctx
*hctx
;
960 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
962 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
963 __blk_mq_run_hw_queue(hctx
);
966 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
968 if (unlikely(!blk_mq_hw_queue_mapped(hctx
)))
971 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
972 &hctx
->delay_work
, msecs_to_jiffies(msecs
));
974 EXPORT_SYMBOL(blk_mq_delay_queue
);
976 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
977 struct blk_mq_ctx
*ctx
,
981 trace_block_rq_insert(hctx
->queue
, rq
);
984 list_add(&rq
->queuelist
, &ctx
->rq_list
);
986 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
989 static void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
,
990 struct request
*rq
, bool at_head
)
992 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
994 __blk_mq_insert_req_list(hctx
, ctx
, rq
, at_head
);
995 blk_mq_hctx_mark_pending(hctx
, ctx
);
998 void blk_mq_insert_request(struct request
*rq
, bool at_head
, bool run_queue
,
1001 struct request_queue
*q
= rq
->q
;
1002 struct blk_mq_hw_ctx
*hctx
;
1003 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
, *current_ctx
;
1005 current_ctx
= blk_mq_get_ctx(q
);
1006 if (!cpu_online(ctx
->cpu
))
1007 rq
->mq_ctx
= ctx
= current_ctx
;
1009 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1011 spin_lock(&ctx
->lock
);
1012 __blk_mq_insert_request(hctx
, rq
, at_head
);
1013 spin_unlock(&ctx
->lock
);
1016 blk_mq_run_hw_queue(hctx
, async
);
1018 blk_mq_put_ctx(current_ctx
);
1021 static void blk_mq_insert_requests(struct request_queue
*q
,
1022 struct blk_mq_ctx
*ctx
,
1023 struct list_head
*list
,
1028 struct blk_mq_hw_ctx
*hctx
;
1029 struct blk_mq_ctx
*current_ctx
;
1031 trace_block_unplug(q
, depth
, !from_schedule
);
1033 current_ctx
= blk_mq_get_ctx(q
);
1035 if (!cpu_online(ctx
->cpu
))
1037 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1040 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1043 spin_lock(&ctx
->lock
);
1044 while (!list_empty(list
)) {
1047 rq
= list_first_entry(list
, struct request
, queuelist
);
1048 list_del_init(&rq
->queuelist
);
1050 __blk_mq_insert_req_list(hctx
, ctx
, rq
, false);
1052 blk_mq_hctx_mark_pending(hctx
, ctx
);
1053 spin_unlock(&ctx
->lock
);
1055 blk_mq_run_hw_queue(hctx
, from_schedule
);
1056 blk_mq_put_ctx(current_ctx
);
1059 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1061 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1062 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1064 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1065 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1066 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1069 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1071 struct blk_mq_ctx
*this_ctx
;
1072 struct request_queue
*this_q
;
1075 LIST_HEAD(ctx_list
);
1078 list_splice_init(&plug
->mq_list
, &list
);
1080 list_sort(NULL
, &list
, plug_ctx_cmp
);
1086 while (!list_empty(&list
)) {
1087 rq
= list_entry_rq(list
.next
);
1088 list_del_init(&rq
->queuelist
);
1090 if (rq
->mq_ctx
!= this_ctx
) {
1092 blk_mq_insert_requests(this_q
, this_ctx
,
1097 this_ctx
= rq
->mq_ctx
;
1103 list_add_tail(&rq
->queuelist
, &ctx_list
);
1107 * If 'this_ctx' is set, we know we have entries to complete
1108 * on 'ctx_list'. Do those.
1111 blk_mq_insert_requests(this_q
, this_ctx
, &ctx_list
, depth
,
1116 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1118 init_request_from_bio(rq
, bio
);
1120 if (blk_do_io_stat(rq
))
1121 blk_account_io_start(rq
, 1);
1124 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1126 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1127 !blk_queue_nomerges(hctx
->queue
);
1130 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx
*hctx
,
1131 struct blk_mq_ctx
*ctx
,
1132 struct request
*rq
, struct bio
*bio
)
1134 if (!hctx_allow_merges(hctx
) || !bio_mergeable(bio
)) {
1135 blk_mq_bio_to_request(rq
, bio
);
1136 spin_lock(&ctx
->lock
);
1138 __blk_mq_insert_request(hctx
, rq
, false);
1139 spin_unlock(&ctx
->lock
);
1142 struct request_queue
*q
= hctx
->queue
;
1144 spin_lock(&ctx
->lock
);
1145 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1146 blk_mq_bio_to_request(rq
, bio
);
1150 spin_unlock(&ctx
->lock
);
1151 __blk_mq_free_request(hctx
, ctx
, rq
);
1156 struct blk_map_ctx
{
1157 struct blk_mq_hw_ctx
*hctx
;
1158 struct blk_mq_ctx
*ctx
;
1161 static struct request
*blk_mq_map_request(struct request_queue
*q
,
1163 struct blk_map_ctx
*data
)
1165 struct blk_mq_hw_ctx
*hctx
;
1166 struct blk_mq_ctx
*ctx
;
1168 int rw
= bio_data_dir(bio
);
1169 struct blk_mq_alloc_data alloc_data
;
1171 blk_queue_enter_live(q
);
1172 ctx
= blk_mq_get_ctx(q
);
1173 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1175 if (rw_is_sync(bio
->bi_rw
))
1178 trace_block_getrq(q
, bio
, rw
);
1179 blk_mq_set_alloc_data(&alloc_data
, q
, BLK_MQ_REQ_NOWAIT
, ctx
, hctx
);
1180 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
1181 if (unlikely(!rq
)) {
1182 __blk_mq_run_hw_queue(hctx
);
1183 blk_mq_put_ctx(ctx
);
1184 trace_block_sleeprq(q
, bio
, rw
);
1186 ctx
= blk_mq_get_ctx(q
);
1187 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1188 blk_mq_set_alloc_data(&alloc_data
, q
, 0, ctx
, hctx
);
1189 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
1190 ctx
= alloc_data
.ctx
;
1191 hctx
= alloc_data
.hctx
;
1200 static int blk_mq_direct_issue_request(struct request
*rq
, blk_qc_t
*cookie
)
1203 struct request_queue
*q
= rq
->q
;
1204 struct blk_mq_hw_ctx
*hctx
= q
->mq_ops
->map_queue(q
,
1206 struct blk_mq_queue_data bd
= {
1211 blk_qc_t new_cookie
= blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
);
1214 * For OK queue, we are done. For error, kill it. Any other
1215 * error (busy), just add it to our list as we previously
1218 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1219 if (ret
== BLK_MQ_RQ_QUEUE_OK
) {
1220 *cookie
= new_cookie
;
1224 __blk_mq_requeue_request(rq
);
1226 if (ret
== BLK_MQ_RQ_QUEUE_ERROR
) {
1227 *cookie
= BLK_QC_T_NONE
;
1229 blk_mq_end_request(rq
, rq
->errors
);
1237 * Multiple hardware queue variant. This will not use per-process plugs,
1238 * but will attempt to bypass the hctx queueing if we can go straight to
1239 * hardware for SYNC IO.
1241 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1243 const int is_sync
= rw_is_sync(bio
->bi_rw
);
1244 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
1245 struct blk_map_ctx data
;
1247 unsigned int request_count
= 0;
1248 struct blk_plug
*plug
;
1249 struct request
*same_queue_rq
= NULL
;
1252 blk_queue_bounce(q
, &bio
);
1254 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1256 return BLK_QC_T_NONE
;
1259 blk_queue_split(q
, &bio
, q
->bio_split
);
1261 if (!is_flush_fua
&& !blk_queue_nomerges(q
)) {
1262 if (blk_attempt_plug_merge(q
, bio
, &request_count
,
1264 return BLK_QC_T_NONE
;
1266 request_count
= blk_plug_queued_count(q
);
1268 rq
= blk_mq_map_request(q
, bio
, &data
);
1270 return BLK_QC_T_NONE
;
1272 cookie
= blk_tag_to_qc_t(rq
->tag
, data
.hctx
->queue_num
);
1274 if (unlikely(is_flush_fua
)) {
1275 blk_mq_bio_to_request(rq
, bio
);
1276 blk_insert_flush(rq
);
1280 plug
= current
->plug
;
1282 * If the driver supports defer issued based on 'last', then
1283 * queue it up like normal since we can potentially save some
1286 if (((plug
&& !blk_queue_nomerges(q
)) || is_sync
) &&
1287 !(data
.hctx
->flags
& BLK_MQ_F_DEFER_ISSUE
)) {
1288 struct request
*old_rq
= NULL
;
1290 blk_mq_bio_to_request(rq
, bio
);
1293 * We do limited pluging. If the bio can be merged, do that.
1294 * Otherwise the existing request in the plug list will be
1295 * issued. So the plug list will have one request at most
1299 * The plug list might get flushed before this. If that
1300 * happens, same_queue_rq is invalid and plug list is
1303 if (same_queue_rq
&& !list_empty(&plug
->mq_list
)) {
1304 old_rq
= same_queue_rq
;
1305 list_del_init(&old_rq
->queuelist
);
1307 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1308 } else /* is_sync */
1310 blk_mq_put_ctx(data
.ctx
);
1313 if (!blk_mq_direct_issue_request(old_rq
, &cookie
))
1315 blk_mq_insert_request(old_rq
, false, true, true);
1319 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1321 * For a SYNC request, send it to the hardware immediately. For
1322 * an ASYNC request, just ensure that we run it later on. The
1323 * latter allows for merging opportunities and more efficient
1327 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1329 blk_mq_put_ctx(data
.ctx
);
1335 * Single hardware queue variant. This will attempt to use any per-process
1336 * plug for merging and IO deferral.
1338 static blk_qc_t
blk_sq_make_request(struct request_queue
*q
, struct bio
*bio
)
1340 const int is_sync
= rw_is_sync(bio
->bi_rw
);
1341 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
1342 struct blk_plug
*plug
;
1343 unsigned int request_count
= 0;
1344 struct blk_map_ctx data
;
1348 blk_queue_bounce(q
, &bio
);
1350 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1352 return BLK_QC_T_NONE
;
1355 blk_queue_split(q
, &bio
, q
->bio_split
);
1357 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1358 blk_attempt_plug_merge(q
, bio
, &request_count
, NULL
))
1359 return BLK_QC_T_NONE
;
1361 rq
= blk_mq_map_request(q
, bio
, &data
);
1363 return BLK_QC_T_NONE
;
1365 cookie
= blk_tag_to_qc_t(rq
->tag
, data
.hctx
->queue_num
);
1367 if (unlikely(is_flush_fua
)) {
1368 blk_mq_bio_to_request(rq
, bio
);
1369 blk_insert_flush(rq
);
1374 * A task plug currently exists. Since this is completely lockless,
1375 * utilize that to temporarily store requests until the task is
1376 * either done or scheduled away.
1378 plug
= current
->plug
;
1380 blk_mq_bio_to_request(rq
, bio
);
1382 trace_block_plug(q
);
1384 blk_mq_put_ctx(data
.ctx
);
1386 if (request_count
>= BLK_MAX_REQUEST_COUNT
) {
1387 blk_flush_plug_list(plug
, false);
1388 trace_block_plug(q
);
1391 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1395 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1397 * For a SYNC request, send it to the hardware immediately. For
1398 * an ASYNC request, just ensure that we run it later on. The
1399 * latter allows for merging opportunities and more efficient
1403 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1406 blk_mq_put_ctx(data
.ctx
);
1411 * Default mapping to a software queue, since we use one per CPU.
1413 struct blk_mq_hw_ctx
*blk_mq_map_queue(struct request_queue
*q
, const int cpu
)
1415 return q
->queue_hw_ctx
[q
->mq_map
[cpu
]];
1417 EXPORT_SYMBOL(blk_mq_map_queue
);
1419 static void blk_mq_free_rq_map(struct blk_mq_tag_set
*set
,
1420 struct blk_mq_tags
*tags
, unsigned int hctx_idx
)
1424 if (tags
->rqs
&& set
->ops
->exit_request
) {
1427 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1430 set
->ops
->exit_request(set
->driver_data
, tags
->rqs
[i
],
1432 tags
->rqs
[i
] = NULL
;
1436 while (!list_empty(&tags
->page_list
)) {
1437 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1438 list_del_init(&page
->lru
);
1440 * Remove kmemleak object previously allocated in
1441 * blk_mq_init_rq_map().
1443 kmemleak_free(page_address(page
));
1444 __free_pages(page
, page
->private);
1449 blk_mq_free_tags(tags
);
1452 static size_t order_to_size(unsigned int order
)
1454 return (size_t)PAGE_SIZE
<< order
;
1457 static struct blk_mq_tags
*blk_mq_init_rq_map(struct blk_mq_tag_set
*set
,
1458 unsigned int hctx_idx
)
1460 struct blk_mq_tags
*tags
;
1461 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1462 size_t rq_size
, left
;
1464 tags
= blk_mq_init_tags(set
->queue_depth
, set
->reserved_tags
,
1466 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1470 INIT_LIST_HEAD(&tags
->page_list
);
1472 tags
->rqs
= kzalloc_node(set
->queue_depth
* sizeof(struct request
*),
1473 GFP_KERNEL
| __GFP_NOWARN
| __GFP_NORETRY
,
1476 blk_mq_free_tags(tags
);
1481 * rq_size is the size of the request plus driver payload, rounded
1482 * to the cacheline size
1484 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1486 left
= rq_size
* set
->queue_depth
;
1488 for (i
= 0; i
< set
->queue_depth
; ) {
1489 int this_order
= max_order
;
1494 while (left
< order_to_size(this_order
- 1) && this_order
)
1498 page
= alloc_pages_node(set
->numa_node
,
1499 GFP_KERNEL
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
1505 if (order_to_size(this_order
) < rq_size
)
1512 page
->private = this_order
;
1513 list_add_tail(&page
->lru
, &tags
->page_list
);
1515 p
= page_address(page
);
1517 * Allow kmemleak to scan these pages as they contain pointers
1518 * to additional allocations like via ops->init_request().
1520 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_KERNEL
);
1521 entries_per_page
= order_to_size(this_order
) / rq_size
;
1522 to_do
= min(entries_per_page
, set
->queue_depth
- i
);
1523 left
-= to_do
* rq_size
;
1524 for (j
= 0; j
< to_do
; j
++) {
1526 if (set
->ops
->init_request
) {
1527 if (set
->ops
->init_request(set
->driver_data
,
1528 tags
->rqs
[i
], hctx_idx
, i
,
1530 tags
->rqs
[i
] = NULL
;
1542 blk_mq_free_rq_map(set
, tags
, hctx_idx
);
1546 static void blk_mq_free_bitmap(struct blk_mq_ctxmap
*bitmap
)
1551 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap
*bitmap
, int node
)
1553 unsigned int bpw
= 8, total
, num_maps
, i
;
1555 bitmap
->bits_per_word
= bpw
;
1557 num_maps
= ALIGN(nr_cpu_ids
, bpw
) / bpw
;
1558 bitmap
->map
= kzalloc_node(num_maps
* sizeof(struct blk_align_bitmap
),
1564 for (i
= 0; i
< num_maps
; i
++) {
1565 bitmap
->map
[i
].depth
= min(total
, bitmap
->bits_per_word
);
1566 total
-= bitmap
->map
[i
].depth
;
1572 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx
*hctx
, int cpu
)
1574 struct request_queue
*q
= hctx
->queue
;
1575 struct blk_mq_ctx
*ctx
;
1579 * Move ctx entries to new CPU, if this one is going away.
1581 ctx
= __blk_mq_get_ctx(q
, cpu
);
1583 spin_lock(&ctx
->lock
);
1584 if (!list_empty(&ctx
->rq_list
)) {
1585 list_splice_init(&ctx
->rq_list
, &tmp
);
1586 blk_mq_hctx_clear_pending(hctx
, ctx
);
1588 spin_unlock(&ctx
->lock
);
1590 if (list_empty(&tmp
))
1593 ctx
= blk_mq_get_ctx(q
);
1594 spin_lock(&ctx
->lock
);
1596 while (!list_empty(&tmp
)) {
1599 rq
= list_first_entry(&tmp
, struct request
, queuelist
);
1601 list_move_tail(&rq
->queuelist
, &ctx
->rq_list
);
1604 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1605 blk_mq_hctx_mark_pending(hctx
, ctx
);
1607 spin_unlock(&ctx
->lock
);
1609 blk_mq_run_hw_queue(hctx
, true);
1610 blk_mq_put_ctx(ctx
);
1614 static int blk_mq_hctx_notify(void *data
, unsigned long action
,
1617 struct blk_mq_hw_ctx
*hctx
= data
;
1619 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
)
1620 return blk_mq_hctx_cpu_offline(hctx
, cpu
);
1623 * In case of CPU online, tags may be reallocated
1624 * in blk_mq_map_swqueue() after mapping is updated.
1630 /* hctx->ctxs will be freed in queue's release handler */
1631 static void blk_mq_exit_hctx(struct request_queue
*q
,
1632 struct blk_mq_tag_set
*set
,
1633 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
1635 unsigned flush_start_tag
= set
->queue_depth
;
1637 blk_mq_tag_idle(hctx
);
1639 if (set
->ops
->exit_request
)
1640 set
->ops
->exit_request(set
->driver_data
,
1641 hctx
->fq
->flush_rq
, hctx_idx
,
1642 flush_start_tag
+ hctx_idx
);
1644 if (set
->ops
->exit_hctx
)
1645 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1647 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1648 blk_free_flush_queue(hctx
->fq
);
1649 blk_mq_free_bitmap(&hctx
->ctx_map
);
1652 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1653 struct blk_mq_tag_set
*set
, int nr_queue
)
1655 struct blk_mq_hw_ctx
*hctx
;
1658 queue_for_each_hw_ctx(q
, hctx
, i
) {
1661 blk_mq_exit_hctx(q
, set
, hctx
, i
);
1665 static void blk_mq_free_hw_queues(struct request_queue
*q
,
1666 struct blk_mq_tag_set
*set
)
1668 struct blk_mq_hw_ctx
*hctx
;
1671 queue_for_each_hw_ctx(q
, hctx
, i
)
1672 free_cpumask_var(hctx
->cpumask
);
1675 static int blk_mq_init_hctx(struct request_queue
*q
,
1676 struct blk_mq_tag_set
*set
,
1677 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
1680 unsigned flush_start_tag
= set
->queue_depth
;
1682 node
= hctx
->numa_node
;
1683 if (node
== NUMA_NO_NODE
)
1684 node
= hctx
->numa_node
= set
->numa_node
;
1686 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1687 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1688 spin_lock_init(&hctx
->lock
);
1689 INIT_LIST_HEAD(&hctx
->dispatch
);
1691 hctx
->queue_num
= hctx_idx
;
1692 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
1694 blk_mq_init_cpu_notifier(&hctx
->cpu_notifier
,
1695 blk_mq_hctx_notify
, hctx
);
1696 blk_mq_register_cpu_notifier(&hctx
->cpu_notifier
);
1698 hctx
->tags
= set
->tags
[hctx_idx
];
1701 * Allocate space for all possible cpus to avoid allocation at
1704 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1707 goto unregister_cpu_notifier
;
1709 if (blk_mq_alloc_bitmap(&hctx
->ctx_map
, node
))
1714 if (set
->ops
->init_hctx
&&
1715 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
1718 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
1722 if (set
->ops
->init_request
&&
1723 set
->ops
->init_request(set
->driver_data
,
1724 hctx
->fq
->flush_rq
, hctx_idx
,
1725 flush_start_tag
+ hctx_idx
, node
))
1733 if (set
->ops
->exit_hctx
)
1734 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1736 blk_mq_free_bitmap(&hctx
->ctx_map
);
1739 unregister_cpu_notifier
:
1740 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1745 static int blk_mq_init_hw_queues(struct request_queue
*q
,
1746 struct blk_mq_tag_set
*set
)
1748 struct blk_mq_hw_ctx
*hctx
;
1752 * Initialize hardware queues
1754 queue_for_each_hw_ctx(q
, hctx
, i
) {
1755 if (blk_mq_init_hctx(q
, set
, hctx
, i
))
1759 if (i
== q
->nr_hw_queues
)
1765 blk_mq_exit_hw_queues(q
, set
, i
);
1770 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1771 unsigned int nr_hw_queues
)
1775 for_each_possible_cpu(i
) {
1776 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1777 struct blk_mq_hw_ctx
*hctx
;
1779 memset(__ctx
, 0, sizeof(*__ctx
));
1781 spin_lock_init(&__ctx
->lock
);
1782 INIT_LIST_HEAD(&__ctx
->rq_list
);
1785 /* If the cpu isn't online, the cpu is mapped to first hctx */
1789 hctx
= q
->mq_ops
->map_queue(q
, i
);
1792 * Set local node, IFF we have more than one hw queue. If
1793 * not, we remain on the home node of the device
1795 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1796 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
1800 static void blk_mq_map_swqueue(struct request_queue
*q
,
1801 const struct cpumask
*online_mask
)
1804 struct blk_mq_hw_ctx
*hctx
;
1805 struct blk_mq_ctx
*ctx
;
1806 struct blk_mq_tag_set
*set
= q
->tag_set
;
1809 * Avoid others reading imcomplete hctx->cpumask through sysfs
1811 mutex_lock(&q
->sysfs_lock
);
1813 queue_for_each_hw_ctx(q
, hctx
, i
) {
1814 cpumask_clear(hctx
->cpumask
);
1819 * Map software to hardware queues
1821 queue_for_each_ctx(q
, ctx
, i
) {
1822 /* If the cpu isn't online, the cpu is mapped to first hctx */
1823 if (!cpumask_test_cpu(i
, online_mask
))
1826 hctx
= q
->mq_ops
->map_queue(q
, i
);
1827 cpumask_set_cpu(i
, hctx
->cpumask
);
1828 ctx
->index_hw
= hctx
->nr_ctx
;
1829 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
1832 mutex_unlock(&q
->sysfs_lock
);
1834 queue_for_each_hw_ctx(q
, hctx
, i
) {
1835 struct blk_mq_ctxmap
*map
= &hctx
->ctx_map
;
1838 * If no software queues are mapped to this hardware queue,
1839 * disable it and free the request entries.
1841 if (!hctx
->nr_ctx
) {
1843 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1844 set
->tags
[i
] = NULL
;
1850 /* unmapped hw queue can be remapped after CPU topo changed */
1852 set
->tags
[i
] = blk_mq_init_rq_map(set
, i
);
1853 hctx
->tags
= set
->tags
[i
];
1854 WARN_ON(!hctx
->tags
);
1856 cpumask_copy(hctx
->tags
->cpumask
, hctx
->cpumask
);
1858 * Set the map size to the number of mapped software queues.
1859 * This is more accurate and more efficient than looping
1860 * over all possibly mapped software queues.
1862 map
->size
= DIV_ROUND_UP(hctx
->nr_ctx
, map
->bits_per_word
);
1865 * Initialize batch roundrobin counts
1867 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
1868 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1872 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
1874 struct blk_mq_hw_ctx
*hctx
;
1877 queue_for_each_hw_ctx(q
, hctx
, i
) {
1879 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
1881 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
1885 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
, bool shared
)
1887 struct request_queue
*q
;
1889 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
1890 blk_mq_freeze_queue(q
);
1891 queue_set_hctx_shared(q
, shared
);
1892 blk_mq_unfreeze_queue(q
);
1896 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
1898 struct blk_mq_tag_set
*set
= q
->tag_set
;
1900 mutex_lock(&set
->tag_list_lock
);
1901 list_del_init(&q
->tag_set_list
);
1902 if (list_is_singular(&set
->tag_list
)) {
1903 /* just transitioned to unshared */
1904 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
1905 /* update existing queue */
1906 blk_mq_update_tag_set_depth(set
, false);
1908 mutex_unlock(&set
->tag_list_lock
);
1911 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
1912 struct request_queue
*q
)
1916 mutex_lock(&set
->tag_list_lock
);
1918 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
1919 if (!list_empty(&set
->tag_list
) && !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
1920 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
1921 /* update existing queue */
1922 blk_mq_update_tag_set_depth(set
, true);
1924 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
1925 queue_set_hctx_shared(q
, true);
1926 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
1928 mutex_unlock(&set
->tag_list_lock
);
1932 * It is the actual release handler for mq, but we do it from
1933 * request queue's release handler for avoiding use-after-free
1934 * and headache because q->mq_kobj shouldn't have been introduced,
1935 * but we can't group ctx/kctx kobj without it.
1937 void blk_mq_release(struct request_queue
*q
)
1939 struct blk_mq_hw_ctx
*hctx
;
1942 /* hctx kobj stays in hctx */
1943 queue_for_each_hw_ctx(q
, hctx
, i
) {
1953 kfree(q
->queue_hw_ctx
);
1955 /* ctx kobj stays in queue_ctx */
1956 free_percpu(q
->queue_ctx
);
1959 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
1961 struct request_queue
*uninit_q
, *q
;
1963 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
1965 return ERR_PTR(-ENOMEM
);
1967 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
1969 blk_cleanup_queue(uninit_q
);
1973 EXPORT_SYMBOL(blk_mq_init_queue
);
1975 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
1976 struct request_queue
*q
)
1978 struct blk_mq_hw_ctx
**hctxs
;
1979 struct blk_mq_ctx __percpu
*ctx
;
1983 ctx
= alloc_percpu(struct blk_mq_ctx
);
1985 return ERR_PTR(-ENOMEM
);
1987 hctxs
= kmalloc_node(set
->nr_hw_queues
* sizeof(*hctxs
), GFP_KERNEL
,
1993 map
= blk_mq_make_queue_map(set
);
1997 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1998 int node
= blk_mq_hw_queue_to_node(map
, i
);
2000 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
2005 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
2009 atomic_set(&hctxs
[i
]->nr_active
, 0);
2010 hctxs
[i
]->numa_node
= node
;
2011 hctxs
[i
]->queue_num
= i
;
2014 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2015 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2017 q
->nr_queues
= nr_cpu_ids
;
2018 q
->nr_hw_queues
= set
->nr_hw_queues
;
2022 q
->queue_hw_ctx
= hctxs
;
2024 q
->mq_ops
= set
->ops
;
2025 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2027 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2028 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
2030 q
->sg_reserved_size
= INT_MAX
;
2032 INIT_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2033 INIT_LIST_HEAD(&q
->requeue_list
);
2034 spin_lock_init(&q
->requeue_lock
);
2036 if (q
->nr_hw_queues
> 1)
2037 blk_queue_make_request(q
, blk_mq_make_request
);
2039 blk_queue_make_request(q
, blk_sq_make_request
);
2042 * Do this after blk_queue_make_request() overrides it...
2044 q
->nr_requests
= set
->queue_depth
;
2046 if (set
->ops
->complete
)
2047 blk_queue_softirq_done(q
, set
->ops
->complete
);
2049 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2051 if (blk_mq_init_hw_queues(q
, set
))
2055 mutex_lock(&all_q_mutex
);
2057 list_add_tail(&q
->all_q_node
, &all_q_list
);
2058 blk_mq_add_queue_tag_set(set
, q
);
2059 blk_mq_map_swqueue(q
, cpu_online_mask
);
2061 mutex_unlock(&all_q_mutex
);
2068 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2071 free_cpumask_var(hctxs
[i
]->cpumask
);
2078 return ERR_PTR(-ENOMEM
);
2080 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2082 void blk_mq_free_queue(struct request_queue
*q
)
2084 struct blk_mq_tag_set
*set
= q
->tag_set
;
2086 mutex_lock(&all_q_mutex
);
2087 list_del_init(&q
->all_q_node
);
2088 mutex_unlock(&all_q_mutex
);
2090 blk_mq_del_queue_tag_set(q
);
2092 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2093 blk_mq_free_hw_queues(q
, set
);
2096 /* Basically redo blk_mq_init_queue with queue frozen */
2097 static void blk_mq_queue_reinit(struct request_queue
*q
,
2098 const struct cpumask
*online_mask
)
2100 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2102 blk_mq_sysfs_unregister(q
);
2104 blk_mq_update_queue_map(q
->mq_map
, q
->nr_hw_queues
, online_mask
);
2107 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2108 * we should change hctx numa_node according to new topology (this
2109 * involves free and re-allocate memory, worthy doing?)
2112 blk_mq_map_swqueue(q
, online_mask
);
2114 blk_mq_sysfs_register(q
);
2117 static int blk_mq_queue_reinit_notify(struct notifier_block
*nb
,
2118 unsigned long action
, void *hcpu
)
2120 struct request_queue
*q
;
2121 int cpu
= (unsigned long)hcpu
;
2123 * New online cpumask which is going to be set in this hotplug event.
2124 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2125 * one-by-one and dynamically allocating this could result in a failure.
2127 static struct cpumask online_new
;
2130 * Before hotadded cpu starts handling requests, new mappings must
2131 * be established. Otherwise, these requests in hw queue might
2132 * never be dispatched.
2134 * For example, there is a single hw queue (hctx) and two CPU queues
2135 * (ctx0 for CPU0, and ctx1 for CPU1).
2137 * Now CPU1 is just onlined and a request is inserted into
2138 * ctx1->rq_list and set bit0 in pending bitmap as ctx1->index_hw is
2141 * And then while running hw queue, flush_busy_ctxs() finds bit0 is
2142 * set in pending bitmap and tries to retrieve requests in
2143 * hctx->ctxs[0]->rq_list. But htx->ctxs[0] is a pointer to ctx0,
2144 * so the request in ctx1->rq_list is ignored.
2146 switch (action
& ~CPU_TASKS_FROZEN
) {
2148 case CPU_UP_CANCELED
:
2149 cpumask_copy(&online_new
, cpu_online_mask
);
2151 case CPU_UP_PREPARE
:
2152 cpumask_copy(&online_new
, cpu_online_mask
);
2153 cpumask_set_cpu(cpu
, &online_new
);
2159 mutex_lock(&all_q_mutex
);
2162 * We need to freeze and reinit all existing queues. Freezing
2163 * involves synchronous wait for an RCU grace period and doing it
2164 * one by one may take a long time. Start freezing all queues in
2165 * one swoop and then wait for the completions so that freezing can
2166 * take place in parallel.
2168 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2169 blk_mq_freeze_queue_start(q
);
2170 list_for_each_entry(q
, &all_q_list
, all_q_node
) {
2171 blk_mq_freeze_queue_wait(q
);
2174 * timeout handler can't touch hw queue during the
2177 del_timer_sync(&q
->timeout
);
2180 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2181 blk_mq_queue_reinit(q
, &online_new
);
2183 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2184 blk_mq_unfreeze_queue(q
);
2186 mutex_unlock(&all_q_mutex
);
2190 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2194 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2195 set
->tags
[i
] = blk_mq_init_rq_map(set
, i
);
2204 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2210 * Allocate the request maps associated with this tag_set. Note that this
2211 * may reduce the depth asked for, if memory is tight. set->queue_depth
2212 * will be updated to reflect the allocated depth.
2214 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2219 depth
= set
->queue_depth
;
2221 err
= __blk_mq_alloc_rq_maps(set
);
2225 set
->queue_depth
>>= 1;
2226 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2230 } while (set
->queue_depth
);
2232 if (!set
->queue_depth
|| err
) {
2233 pr_err("blk-mq: failed to allocate request map\n");
2237 if (depth
!= set
->queue_depth
)
2238 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2239 depth
, set
->queue_depth
);
2244 struct cpumask
*blk_mq_tags_cpumask(struct blk_mq_tags
*tags
)
2246 return tags
->cpumask
;
2248 EXPORT_SYMBOL_GPL(blk_mq_tags_cpumask
);
2251 * Alloc a tag set to be associated with one or more request queues.
2252 * May fail with EINVAL for various error conditions. May adjust the
2253 * requested depth down, if if it too large. In that case, the set
2254 * value will be stored in set->queue_depth.
2256 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2258 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2260 if (!set
->nr_hw_queues
)
2262 if (!set
->queue_depth
)
2264 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2267 if (!set
->ops
->queue_rq
|| !set
->ops
->map_queue
)
2270 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2271 pr_info("blk-mq: reduced tag depth to %u\n",
2273 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2277 * If a crashdump is active, then we are potentially in a very
2278 * memory constrained environment. Limit us to 1 queue and
2279 * 64 tags to prevent using too much memory.
2281 if (is_kdump_kernel()) {
2282 set
->nr_hw_queues
= 1;
2283 set
->queue_depth
= min(64U, set
->queue_depth
);
2286 set
->tags
= kmalloc_node(set
->nr_hw_queues
*
2287 sizeof(struct blk_mq_tags
*),
2288 GFP_KERNEL
, set
->numa_node
);
2292 if (blk_mq_alloc_rq_maps(set
))
2295 mutex_init(&set
->tag_list_lock
);
2296 INIT_LIST_HEAD(&set
->tag_list
);
2304 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2306 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2310 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2312 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2318 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2320 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2322 struct blk_mq_tag_set
*set
= q
->tag_set
;
2323 struct blk_mq_hw_ctx
*hctx
;
2326 if (!set
|| nr
> set
->queue_depth
)
2330 queue_for_each_hw_ctx(q
, hctx
, i
) {
2331 ret
= blk_mq_tag_update_depth(hctx
->tags
, nr
);
2337 q
->nr_requests
= nr
;
2342 void blk_mq_disable_hotplug(void)
2344 mutex_lock(&all_q_mutex
);
2347 void blk_mq_enable_hotplug(void)
2349 mutex_unlock(&all_q_mutex
);
2352 static int __init
blk_mq_init(void)
2356 hotcpu_notifier(blk_mq_queue_reinit_notify
, 0);
2360 subsys_initcall(blk_mq_init
);