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
);
606 if (rq
->cmd_flags
& REQ_NO_TIMEOUT
)
609 if (time_after_eq(jiffies
, rq
->deadline
)) {
610 if (!blk_mark_rq_complete(rq
))
611 blk_mq_rq_timed_out(rq
, reserved
);
612 } else if (!data
->next_set
|| time_after(data
->next
, rq
->deadline
)) {
613 data
->next
= rq
->deadline
;
618 static void blk_mq_rq_timer(unsigned long priv
)
620 struct request_queue
*q
= (struct request_queue
*)priv
;
621 struct blk_mq_timeout_data data
= {
627 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &data
);
630 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
631 mod_timer(&q
->timeout
, data
.next
);
633 struct blk_mq_hw_ctx
*hctx
;
635 queue_for_each_hw_ctx(q
, hctx
, i
) {
636 /* the hctx may be unmapped, so check it here */
637 if (blk_mq_hw_queue_mapped(hctx
))
638 blk_mq_tag_idle(hctx
);
644 * Reverse check our software queue for entries that we could potentially
645 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
646 * too much time checking for merges.
648 static bool blk_mq_attempt_merge(struct request_queue
*q
,
649 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
654 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
660 if (!blk_rq_merge_ok(rq
, bio
))
663 el_ret
= blk_try_merge(rq
, bio
);
664 if (el_ret
== ELEVATOR_BACK_MERGE
) {
665 if (bio_attempt_back_merge(q
, rq
, bio
)) {
670 } else if (el_ret
== ELEVATOR_FRONT_MERGE
) {
671 if (bio_attempt_front_merge(q
, rq
, bio
)) {
683 * Process software queues that have been marked busy, splicing them
684 * to the for-dispatch
686 static void flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
688 struct blk_mq_ctx
*ctx
;
691 for (i
= 0; i
< hctx
->ctx_map
.size
; i
++) {
692 struct blk_align_bitmap
*bm
= &hctx
->ctx_map
.map
[i
];
693 unsigned int off
, bit
;
699 off
= i
* hctx
->ctx_map
.bits_per_word
;
701 bit
= find_next_bit(&bm
->word
, bm
->depth
, bit
);
702 if (bit
>= bm
->depth
)
705 ctx
= hctx
->ctxs
[bit
+ off
];
706 clear_bit(bit
, &bm
->word
);
707 spin_lock(&ctx
->lock
);
708 list_splice_tail_init(&ctx
->rq_list
, list
);
709 spin_unlock(&ctx
->lock
);
717 * Run this hardware queue, pulling any software queues mapped to it in.
718 * Note that this function currently has various problems around ordering
719 * of IO. In particular, we'd like FIFO behaviour on handling existing
720 * items on the hctx->dispatch list. Ignore that for now.
722 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
724 struct request_queue
*q
= hctx
->queue
;
727 LIST_HEAD(driver_list
);
728 struct list_head
*dptr
;
731 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
));
733 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)))
739 * Touch any software queue that has pending entries.
741 flush_busy_ctxs(hctx
, &rq_list
);
744 * If we have previous entries on our dispatch list, grab them
745 * and stuff them at the front for more fair dispatch.
747 if (!list_empty_careful(&hctx
->dispatch
)) {
748 spin_lock(&hctx
->lock
);
749 if (!list_empty(&hctx
->dispatch
))
750 list_splice_init(&hctx
->dispatch
, &rq_list
);
751 spin_unlock(&hctx
->lock
);
755 * Start off with dptr being NULL, so we start the first request
756 * immediately, even if we have more pending.
761 * Now process all the entries, sending them to the driver.
764 while (!list_empty(&rq_list
)) {
765 struct blk_mq_queue_data bd
;
768 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
769 list_del_init(&rq
->queuelist
);
773 bd
.last
= list_empty(&rq_list
);
775 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
777 case BLK_MQ_RQ_QUEUE_OK
:
780 case BLK_MQ_RQ_QUEUE_BUSY
:
781 list_add(&rq
->queuelist
, &rq_list
);
782 __blk_mq_requeue_request(rq
);
785 pr_err("blk-mq: bad return on queue: %d\n", ret
);
786 case BLK_MQ_RQ_QUEUE_ERROR
:
788 blk_mq_end_request(rq
, rq
->errors
);
792 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
796 * We've done the first request. If we have more than 1
797 * left in the list, set dptr to defer issue.
799 if (!dptr
&& rq_list
.next
!= rq_list
.prev
)
804 hctx
->dispatched
[0]++;
805 else if (queued
< (1 << (BLK_MQ_MAX_DISPATCH_ORDER
- 1)))
806 hctx
->dispatched
[ilog2(queued
) + 1]++;
809 * Any items that need requeuing? Stuff them into hctx->dispatch,
810 * that is where we will continue on next queue run.
812 if (!list_empty(&rq_list
)) {
813 spin_lock(&hctx
->lock
);
814 list_splice(&rq_list
, &hctx
->dispatch
);
815 spin_unlock(&hctx
->lock
);
817 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
818 * it's possible the queue is stopped and restarted again
819 * before this. Queue restart will dispatch requests. And since
820 * requests in rq_list aren't added into hctx->dispatch yet,
821 * the requests in rq_list might get lost.
823 * blk_mq_run_hw_queue() already checks the STOPPED bit
825 blk_mq_run_hw_queue(hctx
, true);
830 * It'd be great if the workqueue API had a way to pass
831 * in a mask and had some smarts for more clever placement.
832 * For now we just round-robin here, switching for every
833 * BLK_MQ_CPU_WORK_BATCH queued items.
835 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
837 if (hctx
->queue
->nr_hw_queues
== 1)
838 return WORK_CPU_UNBOUND
;
840 if (--hctx
->next_cpu_batch
<= 0) {
841 int cpu
= hctx
->next_cpu
, next_cpu
;
843 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
844 if (next_cpu
>= nr_cpu_ids
)
845 next_cpu
= cpumask_first(hctx
->cpumask
);
847 hctx
->next_cpu
= next_cpu
;
848 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
853 return hctx
->next_cpu
;
856 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
858 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
) ||
859 !blk_mq_hw_queue_mapped(hctx
)))
864 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
865 __blk_mq_run_hw_queue(hctx
);
873 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
877 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
879 struct blk_mq_hw_ctx
*hctx
;
882 queue_for_each_hw_ctx(q
, hctx
, i
) {
883 if ((!blk_mq_hctx_has_pending(hctx
) &&
884 list_empty_careful(&hctx
->dispatch
)) ||
885 test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
888 blk_mq_run_hw_queue(hctx
, async
);
891 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
893 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
895 cancel_delayed_work(&hctx
->run_work
);
896 cancel_delayed_work(&hctx
->delay_work
);
897 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
899 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
901 void blk_mq_stop_hw_queues(struct request_queue
*q
)
903 struct blk_mq_hw_ctx
*hctx
;
906 queue_for_each_hw_ctx(q
, hctx
, i
)
907 blk_mq_stop_hw_queue(hctx
);
909 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
911 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
913 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
915 blk_mq_run_hw_queue(hctx
, false);
917 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
919 void blk_mq_start_hw_queues(struct request_queue
*q
)
921 struct blk_mq_hw_ctx
*hctx
;
924 queue_for_each_hw_ctx(q
, hctx
, i
)
925 blk_mq_start_hw_queue(hctx
);
927 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
929 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
931 struct blk_mq_hw_ctx
*hctx
;
934 queue_for_each_hw_ctx(q
, hctx
, i
) {
935 if (!test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
938 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
939 blk_mq_run_hw_queue(hctx
, async
);
942 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
944 static void blk_mq_run_work_fn(struct work_struct
*work
)
946 struct blk_mq_hw_ctx
*hctx
;
948 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
950 __blk_mq_run_hw_queue(hctx
);
953 static void blk_mq_delay_work_fn(struct work_struct
*work
)
955 struct blk_mq_hw_ctx
*hctx
;
957 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
959 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
960 __blk_mq_run_hw_queue(hctx
);
963 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
965 if (unlikely(!blk_mq_hw_queue_mapped(hctx
)))
968 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
969 &hctx
->delay_work
, msecs_to_jiffies(msecs
));
971 EXPORT_SYMBOL(blk_mq_delay_queue
);
973 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
974 struct blk_mq_ctx
*ctx
,
978 trace_block_rq_insert(hctx
->queue
, rq
);
981 list_add(&rq
->queuelist
, &ctx
->rq_list
);
983 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
986 static void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
,
987 struct request
*rq
, bool at_head
)
989 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
991 __blk_mq_insert_req_list(hctx
, ctx
, rq
, at_head
);
992 blk_mq_hctx_mark_pending(hctx
, ctx
);
995 void blk_mq_insert_request(struct request
*rq
, bool at_head
, bool run_queue
,
998 struct request_queue
*q
= rq
->q
;
999 struct blk_mq_hw_ctx
*hctx
;
1000 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
, *current_ctx
;
1002 current_ctx
= blk_mq_get_ctx(q
);
1003 if (!cpu_online(ctx
->cpu
))
1004 rq
->mq_ctx
= ctx
= current_ctx
;
1006 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1008 spin_lock(&ctx
->lock
);
1009 __blk_mq_insert_request(hctx
, rq
, at_head
);
1010 spin_unlock(&ctx
->lock
);
1013 blk_mq_run_hw_queue(hctx
, async
);
1015 blk_mq_put_ctx(current_ctx
);
1018 static void blk_mq_insert_requests(struct request_queue
*q
,
1019 struct blk_mq_ctx
*ctx
,
1020 struct list_head
*list
,
1025 struct blk_mq_hw_ctx
*hctx
;
1026 struct blk_mq_ctx
*current_ctx
;
1028 trace_block_unplug(q
, depth
, !from_schedule
);
1030 current_ctx
= blk_mq_get_ctx(q
);
1032 if (!cpu_online(ctx
->cpu
))
1034 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1037 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1040 spin_lock(&ctx
->lock
);
1041 while (!list_empty(list
)) {
1044 rq
= list_first_entry(list
, struct request
, queuelist
);
1045 list_del_init(&rq
->queuelist
);
1047 __blk_mq_insert_req_list(hctx
, ctx
, rq
, false);
1049 blk_mq_hctx_mark_pending(hctx
, ctx
);
1050 spin_unlock(&ctx
->lock
);
1052 blk_mq_run_hw_queue(hctx
, from_schedule
);
1053 blk_mq_put_ctx(current_ctx
);
1056 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1058 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1059 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1061 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1062 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1063 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1066 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1068 struct blk_mq_ctx
*this_ctx
;
1069 struct request_queue
*this_q
;
1072 LIST_HEAD(ctx_list
);
1075 list_splice_init(&plug
->mq_list
, &list
);
1077 list_sort(NULL
, &list
, plug_ctx_cmp
);
1083 while (!list_empty(&list
)) {
1084 rq
= list_entry_rq(list
.next
);
1085 list_del_init(&rq
->queuelist
);
1087 if (rq
->mq_ctx
!= this_ctx
) {
1089 blk_mq_insert_requests(this_q
, this_ctx
,
1094 this_ctx
= rq
->mq_ctx
;
1100 list_add_tail(&rq
->queuelist
, &ctx_list
);
1104 * If 'this_ctx' is set, we know we have entries to complete
1105 * on 'ctx_list'. Do those.
1108 blk_mq_insert_requests(this_q
, this_ctx
, &ctx_list
, depth
,
1113 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1115 init_request_from_bio(rq
, bio
);
1117 if (blk_do_io_stat(rq
))
1118 blk_account_io_start(rq
, 1);
1121 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1123 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1124 !blk_queue_nomerges(hctx
->queue
);
1127 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx
*hctx
,
1128 struct blk_mq_ctx
*ctx
,
1129 struct request
*rq
, struct bio
*bio
)
1131 if (!hctx_allow_merges(hctx
) || !bio_mergeable(bio
)) {
1132 blk_mq_bio_to_request(rq
, bio
);
1133 spin_lock(&ctx
->lock
);
1135 __blk_mq_insert_request(hctx
, rq
, false);
1136 spin_unlock(&ctx
->lock
);
1139 struct request_queue
*q
= hctx
->queue
;
1141 spin_lock(&ctx
->lock
);
1142 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1143 blk_mq_bio_to_request(rq
, bio
);
1147 spin_unlock(&ctx
->lock
);
1148 __blk_mq_free_request(hctx
, ctx
, rq
);
1153 struct blk_map_ctx
{
1154 struct blk_mq_hw_ctx
*hctx
;
1155 struct blk_mq_ctx
*ctx
;
1158 static struct request
*blk_mq_map_request(struct request_queue
*q
,
1160 struct blk_map_ctx
*data
)
1162 struct blk_mq_hw_ctx
*hctx
;
1163 struct blk_mq_ctx
*ctx
;
1165 int rw
= bio_data_dir(bio
);
1166 struct blk_mq_alloc_data alloc_data
;
1168 blk_queue_enter_live(q
);
1169 ctx
= blk_mq_get_ctx(q
);
1170 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1172 if (rw_is_sync(bio
->bi_rw
))
1175 trace_block_getrq(q
, bio
, rw
);
1176 blk_mq_set_alloc_data(&alloc_data
, q
, BLK_MQ_REQ_NOWAIT
, ctx
, hctx
);
1177 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
1178 if (unlikely(!rq
)) {
1179 __blk_mq_run_hw_queue(hctx
);
1180 blk_mq_put_ctx(ctx
);
1181 trace_block_sleeprq(q
, bio
, rw
);
1183 ctx
= blk_mq_get_ctx(q
);
1184 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1185 blk_mq_set_alloc_data(&alloc_data
, q
, 0, ctx
, hctx
);
1186 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
1187 ctx
= alloc_data
.ctx
;
1188 hctx
= alloc_data
.hctx
;
1197 static int blk_mq_direct_issue_request(struct request
*rq
, blk_qc_t
*cookie
)
1200 struct request_queue
*q
= rq
->q
;
1201 struct blk_mq_hw_ctx
*hctx
= q
->mq_ops
->map_queue(q
,
1203 struct blk_mq_queue_data bd
= {
1208 blk_qc_t new_cookie
= blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
);
1211 * For OK queue, we are done. For error, kill it. Any other
1212 * error (busy), just add it to our list as we previously
1215 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1216 if (ret
== BLK_MQ_RQ_QUEUE_OK
) {
1217 *cookie
= new_cookie
;
1221 __blk_mq_requeue_request(rq
);
1223 if (ret
== BLK_MQ_RQ_QUEUE_ERROR
) {
1224 *cookie
= BLK_QC_T_NONE
;
1226 blk_mq_end_request(rq
, rq
->errors
);
1234 * Multiple hardware queue variant. This will not use per-process plugs,
1235 * but will attempt to bypass the hctx queueing if we can go straight to
1236 * hardware for SYNC IO.
1238 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1240 const int is_sync
= rw_is_sync(bio
->bi_rw
);
1241 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
1242 struct blk_map_ctx data
;
1244 unsigned int request_count
= 0;
1245 struct blk_plug
*plug
;
1246 struct request
*same_queue_rq
= NULL
;
1249 blk_queue_bounce(q
, &bio
);
1251 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1253 return BLK_QC_T_NONE
;
1256 blk_queue_split(q
, &bio
, q
->bio_split
);
1258 if (!is_flush_fua
&& !blk_queue_nomerges(q
)) {
1259 if (blk_attempt_plug_merge(q
, bio
, &request_count
,
1261 return BLK_QC_T_NONE
;
1263 request_count
= blk_plug_queued_count(q
);
1265 rq
= blk_mq_map_request(q
, bio
, &data
);
1267 return BLK_QC_T_NONE
;
1269 cookie
= blk_tag_to_qc_t(rq
->tag
, data
.hctx
->queue_num
);
1271 if (unlikely(is_flush_fua
)) {
1272 blk_mq_bio_to_request(rq
, bio
);
1273 blk_insert_flush(rq
);
1277 plug
= current
->plug
;
1279 * If the driver supports defer issued based on 'last', then
1280 * queue it up like normal since we can potentially save some
1283 if (((plug
&& !blk_queue_nomerges(q
)) || is_sync
) &&
1284 !(data
.hctx
->flags
& BLK_MQ_F_DEFER_ISSUE
)) {
1285 struct request
*old_rq
= NULL
;
1287 blk_mq_bio_to_request(rq
, bio
);
1290 * We do limited pluging. If the bio can be merged, do that.
1291 * Otherwise the existing request in the plug list will be
1292 * issued. So the plug list will have one request at most
1296 * The plug list might get flushed before this. If that
1297 * happens, same_queue_rq is invalid and plug list is
1300 if (same_queue_rq
&& !list_empty(&plug
->mq_list
)) {
1301 old_rq
= same_queue_rq
;
1302 list_del_init(&old_rq
->queuelist
);
1304 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1305 } else /* is_sync */
1307 blk_mq_put_ctx(data
.ctx
);
1310 if (!blk_mq_direct_issue_request(old_rq
, &cookie
))
1312 blk_mq_insert_request(old_rq
, false, true, true);
1316 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1318 * For a SYNC request, send it to the hardware immediately. For
1319 * an ASYNC request, just ensure that we run it later on. The
1320 * latter allows for merging opportunities and more efficient
1324 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1326 blk_mq_put_ctx(data
.ctx
);
1332 * Single hardware queue variant. This will attempt to use any per-process
1333 * plug for merging and IO deferral.
1335 static blk_qc_t
blk_sq_make_request(struct request_queue
*q
, struct bio
*bio
)
1337 const int is_sync
= rw_is_sync(bio
->bi_rw
);
1338 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
1339 struct blk_plug
*plug
;
1340 unsigned int request_count
= 0;
1341 struct blk_map_ctx data
;
1345 blk_queue_bounce(q
, &bio
);
1347 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1349 return BLK_QC_T_NONE
;
1352 blk_queue_split(q
, &bio
, q
->bio_split
);
1354 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1355 blk_attempt_plug_merge(q
, bio
, &request_count
, NULL
))
1356 return BLK_QC_T_NONE
;
1358 rq
= blk_mq_map_request(q
, bio
, &data
);
1360 return BLK_QC_T_NONE
;
1362 cookie
= blk_tag_to_qc_t(rq
->tag
, data
.hctx
->queue_num
);
1364 if (unlikely(is_flush_fua
)) {
1365 blk_mq_bio_to_request(rq
, bio
);
1366 blk_insert_flush(rq
);
1371 * A task plug currently exists. Since this is completely lockless,
1372 * utilize that to temporarily store requests until the task is
1373 * either done or scheduled away.
1375 plug
= current
->plug
;
1377 blk_mq_bio_to_request(rq
, bio
);
1379 trace_block_plug(q
);
1381 blk_mq_put_ctx(data
.ctx
);
1383 if (request_count
>= BLK_MAX_REQUEST_COUNT
) {
1384 blk_flush_plug_list(plug
, false);
1385 trace_block_plug(q
);
1388 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1392 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1394 * For a SYNC request, send it to the hardware immediately. For
1395 * an ASYNC request, just ensure that we run it later on. The
1396 * latter allows for merging opportunities and more efficient
1400 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1403 blk_mq_put_ctx(data
.ctx
);
1408 * Default mapping to a software queue, since we use one per CPU.
1410 struct blk_mq_hw_ctx
*blk_mq_map_queue(struct request_queue
*q
, const int cpu
)
1412 return q
->queue_hw_ctx
[q
->mq_map
[cpu
]];
1414 EXPORT_SYMBOL(blk_mq_map_queue
);
1416 static void blk_mq_free_rq_map(struct blk_mq_tag_set
*set
,
1417 struct blk_mq_tags
*tags
, unsigned int hctx_idx
)
1421 if (tags
->rqs
&& set
->ops
->exit_request
) {
1424 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1427 set
->ops
->exit_request(set
->driver_data
, tags
->rqs
[i
],
1429 tags
->rqs
[i
] = NULL
;
1433 while (!list_empty(&tags
->page_list
)) {
1434 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1435 list_del_init(&page
->lru
);
1437 * Remove kmemleak object previously allocated in
1438 * blk_mq_init_rq_map().
1440 kmemleak_free(page_address(page
));
1441 __free_pages(page
, page
->private);
1446 blk_mq_free_tags(tags
);
1449 static size_t order_to_size(unsigned int order
)
1451 return (size_t)PAGE_SIZE
<< order
;
1454 static struct blk_mq_tags
*blk_mq_init_rq_map(struct blk_mq_tag_set
*set
,
1455 unsigned int hctx_idx
)
1457 struct blk_mq_tags
*tags
;
1458 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1459 size_t rq_size
, left
;
1461 tags
= blk_mq_init_tags(set
->queue_depth
, set
->reserved_tags
,
1463 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1467 INIT_LIST_HEAD(&tags
->page_list
);
1469 tags
->rqs
= kzalloc_node(set
->queue_depth
* sizeof(struct request
*),
1470 GFP_KERNEL
| __GFP_NOWARN
| __GFP_NORETRY
,
1473 blk_mq_free_tags(tags
);
1478 * rq_size is the size of the request plus driver payload, rounded
1479 * to the cacheline size
1481 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1483 left
= rq_size
* set
->queue_depth
;
1485 for (i
= 0; i
< set
->queue_depth
; ) {
1486 int this_order
= max_order
;
1491 while (left
< order_to_size(this_order
- 1) && this_order
)
1495 page
= alloc_pages_node(set
->numa_node
,
1496 GFP_KERNEL
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
1502 if (order_to_size(this_order
) < rq_size
)
1509 page
->private = this_order
;
1510 list_add_tail(&page
->lru
, &tags
->page_list
);
1512 p
= page_address(page
);
1514 * Allow kmemleak to scan these pages as they contain pointers
1515 * to additional allocations like via ops->init_request().
1517 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_KERNEL
);
1518 entries_per_page
= order_to_size(this_order
) / rq_size
;
1519 to_do
= min(entries_per_page
, set
->queue_depth
- i
);
1520 left
-= to_do
* rq_size
;
1521 for (j
= 0; j
< to_do
; j
++) {
1523 if (set
->ops
->init_request
) {
1524 if (set
->ops
->init_request(set
->driver_data
,
1525 tags
->rqs
[i
], hctx_idx
, i
,
1527 tags
->rqs
[i
] = NULL
;
1539 blk_mq_free_rq_map(set
, tags
, hctx_idx
);
1543 static void blk_mq_free_bitmap(struct blk_mq_ctxmap
*bitmap
)
1548 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap
*bitmap
, int node
)
1550 unsigned int bpw
= 8, total
, num_maps
, i
;
1552 bitmap
->bits_per_word
= bpw
;
1554 num_maps
= ALIGN(nr_cpu_ids
, bpw
) / bpw
;
1555 bitmap
->map
= kzalloc_node(num_maps
* sizeof(struct blk_align_bitmap
),
1561 for (i
= 0; i
< num_maps
; i
++) {
1562 bitmap
->map
[i
].depth
= min(total
, bitmap
->bits_per_word
);
1563 total
-= bitmap
->map
[i
].depth
;
1569 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx
*hctx
, int cpu
)
1571 struct request_queue
*q
= hctx
->queue
;
1572 struct blk_mq_ctx
*ctx
;
1576 * Move ctx entries to new CPU, if this one is going away.
1578 ctx
= __blk_mq_get_ctx(q
, cpu
);
1580 spin_lock(&ctx
->lock
);
1581 if (!list_empty(&ctx
->rq_list
)) {
1582 list_splice_init(&ctx
->rq_list
, &tmp
);
1583 blk_mq_hctx_clear_pending(hctx
, ctx
);
1585 spin_unlock(&ctx
->lock
);
1587 if (list_empty(&tmp
))
1590 ctx
= blk_mq_get_ctx(q
);
1591 spin_lock(&ctx
->lock
);
1593 while (!list_empty(&tmp
)) {
1596 rq
= list_first_entry(&tmp
, struct request
, queuelist
);
1598 list_move_tail(&rq
->queuelist
, &ctx
->rq_list
);
1601 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1602 blk_mq_hctx_mark_pending(hctx
, ctx
);
1604 spin_unlock(&ctx
->lock
);
1606 blk_mq_run_hw_queue(hctx
, true);
1607 blk_mq_put_ctx(ctx
);
1611 static int blk_mq_hctx_notify(void *data
, unsigned long action
,
1614 struct blk_mq_hw_ctx
*hctx
= data
;
1616 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
)
1617 return blk_mq_hctx_cpu_offline(hctx
, cpu
);
1620 * In case of CPU online, tags may be reallocated
1621 * in blk_mq_map_swqueue() after mapping is updated.
1627 /* hctx->ctxs will be freed in queue's release handler */
1628 static void blk_mq_exit_hctx(struct request_queue
*q
,
1629 struct blk_mq_tag_set
*set
,
1630 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
1632 unsigned flush_start_tag
= set
->queue_depth
;
1634 blk_mq_tag_idle(hctx
);
1636 if (set
->ops
->exit_request
)
1637 set
->ops
->exit_request(set
->driver_data
,
1638 hctx
->fq
->flush_rq
, hctx_idx
,
1639 flush_start_tag
+ hctx_idx
);
1641 if (set
->ops
->exit_hctx
)
1642 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1644 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1645 blk_free_flush_queue(hctx
->fq
);
1646 blk_mq_free_bitmap(&hctx
->ctx_map
);
1649 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1650 struct blk_mq_tag_set
*set
, int nr_queue
)
1652 struct blk_mq_hw_ctx
*hctx
;
1655 queue_for_each_hw_ctx(q
, hctx
, i
) {
1658 blk_mq_exit_hctx(q
, set
, hctx
, i
);
1662 static void blk_mq_free_hw_queues(struct request_queue
*q
,
1663 struct blk_mq_tag_set
*set
)
1665 struct blk_mq_hw_ctx
*hctx
;
1668 queue_for_each_hw_ctx(q
, hctx
, i
)
1669 free_cpumask_var(hctx
->cpumask
);
1672 static int blk_mq_init_hctx(struct request_queue
*q
,
1673 struct blk_mq_tag_set
*set
,
1674 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
1677 unsigned flush_start_tag
= set
->queue_depth
;
1679 node
= hctx
->numa_node
;
1680 if (node
== NUMA_NO_NODE
)
1681 node
= hctx
->numa_node
= set
->numa_node
;
1683 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1684 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1685 spin_lock_init(&hctx
->lock
);
1686 INIT_LIST_HEAD(&hctx
->dispatch
);
1688 hctx
->queue_num
= hctx_idx
;
1689 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
1691 blk_mq_init_cpu_notifier(&hctx
->cpu_notifier
,
1692 blk_mq_hctx_notify
, hctx
);
1693 blk_mq_register_cpu_notifier(&hctx
->cpu_notifier
);
1695 hctx
->tags
= set
->tags
[hctx_idx
];
1698 * Allocate space for all possible cpus to avoid allocation at
1701 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1704 goto unregister_cpu_notifier
;
1706 if (blk_mq_alloc_bitmap(&hctx
->ctx_map
, node
))
1711 if (set
->ops
->init_hctx
&&
1712 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
1715 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
1719 if (set
->ops
->init_request
&&
1720 set
->ops
->init_request(set
->driver_data
,
1721 hctx
->fq
->flush_rq
, hctx_idx
,
1722 flush_start_tag
+ hctx_idx
, node
))
1730 if (set
->ops
->exit_hctx
)
1731 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1733 blk_mq_free_bitmap(&hctx
->ctx_map
);
1736 unregister_cpu_notifier
:
1737 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1742 static int blk_mq_init_hw_queues(struct request_queue
*q
,
1743 struct blk_mq_tag_set
*set
)
1745 struct blk_mq_hw_ctx
*hctx
;
1749 * Initialize hardware queues
1751 queue_for_each_hw_ctx(q
, hctx
, i
) {
1752 if (blk_mq_init_hctx(q
, set
, hctx
, i
))
1756 if (i
== q
->nr_hw_queues
)
1762 blk_mq_exit_hw_queues(q
, set
, i
);
1767 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1768 unsigned int nr_hw_queues
)
1772 for_each_possible_cpu(i
) {
1773 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1774 struct blk_mq_hw_ctx
*hctx
;
1776 memset(__ctx
, 0, sizeof(*__ctx
));
1778 spin_lock_init(&__ctx
->lock
);
1779 INIT_LIST_HEAD(&__ctx
->rq_list
);
1782 /* If the cpu isn't online, the cpu is mapped to first hctx */
1786 hctx
= q
->mq_ops
->map_queue(q
, i
);
1789 * Set local node, IFF we have more than one hw queue. If
1790 * not, we remain on the home node of the device
1792 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1793 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
1797 static void blk_mq_map_swqueue(struct request_queue
*q
,
1798 const struct cpumask
*online_mask
)
1801 struct blk_mq_hw_ctx
*hctx
;
1802 struct blk_mq_ctx
*ctx
;
1803 struct blk_mq_tag_set
*set
= q
->tag_set
;
1806 * Avoid others reading imcomplete hctx->cpumask through sysfs
1808 mutex_lock(&q
->sysfs_lock
);
1810 queue_for_each_hw_ctx(q
, hctx
, i
) {
1811 cpumask_clear(hctx
->cpumask
);
1816 * Map software to hardware queues
1818 queue_for_each_ctx(q
, ctx
, i
) {
1819 /* If the cpu isn't online, the cpu is mapped to first hctx */
1820 if (!cpumask_test_cpu(i
, online_mask
))
1823 hctx
= q
->mq_ops
->map_queue(q
, i
);
1824 cpumask_set_cpu(i
, hctx
->cpumask
);
1825 ctx
->index_hw
= hctx
->nr_ctx
;
1826 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
1829 mutex_unlock(&q
->sysfs_lock
);
1831 queue_for_each_hw_ctx(q
, hctx
, i
) {
1832 struct blk_mq_ctxmap
*map
= &hctx
->ctx_map
;
1835 * If no software queues are mapped to this hardware queue,
1836 * disable it and free the request entries.
1838 if (!hctx
->nr_ctx
) {
1840 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1841 set
->tags
[i
] = NULL
;
1847 /* unmapped hw queue can be remapped after CPU topo changed */
1849 set
->tags
[i
] = blk_mq_init_rq_map(set
, i
);
1850 hctx
->tags
= set
->tags
[i
];
1851 WARN_ON(!hctx
->tags
);
1853 cpumask_copy(hctx
->tags
->cpumask
, hctx
->cpumask
);
1855 * Set the map size to the number of mapped software queues.
1856 * This is more accurate and more efficient than looping
1857 * over all possibly mapped software queues.
1859 map
->size
= DIV_ROUND_UP(hctx
->nr_ctx
, map
->bits_per_word
);
1862 * Initialize batch roundrobin counts
1864 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
1865 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1869 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
1871 struct blk_mq_hw_ctx
*hctx
;
1874 queue_for_each_hw_ctx(q
, hctx
, i
) {
1876 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
1878 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
1882 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
, bool shared
)
1884 struct request_queue
*q
;
1886 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
1887 blk_mq_freeze_queue(q
);
1888 queue_set_hctx_shared(q
, shared
);
1889 blk_mq_unfreeze_queue(q
);
1893 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
1895 struct blk_mq_tag_set
*set
= q
->tag_set
;
1897 mutex_lock(&set
->tag_list_lock
);
1898 list_del_init(&q
->tag_set_list
);
1899 if (list_is_singular(&set
->tag_list
)) {
1900 /* just transitioned to unshared */
1901 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
1902 /* update existing queue */
1903 blk_mq_update_tag_set_depth(set
, false);
1905 mutex_unlock(&set
->tag_list_lock
);
1908 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
1909 struct request_queue
*q
)
1913 mutex_lock(&set
->tag_list_lock
);
1915 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
1916 if (!list_empty(&set
->tag_list
) && !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
1917 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
1918 /* update existing queue */
1919 blk_mq_update_tag_set_depth(set
, true);
1921 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
1922 queue_set_hctx_shared(q
, true);
1923 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
1925 mutex_unlock(&set
->tag_list_lock
);
1929 * It is the actual release handler for mq, but we do it from
1930 * request queue's release handler for avoiding use-after-free
1931 * and headache because q->mq_kobj shouldn't have been introduced,
1932 * but we can't group ctx/kctx kobj without it.
1934 void blk_mq_release(struct request_queue
*q
)
1936 struct blk_mq_hw_ctx
*hctx
;
1939 /* hctx kobj stays in hctx */
1940 queue_for_each_hw_ctx(q
, hctx
, i
) {
1950 kfree(q
->queue_hw_ctx
);
1952 /* ctx kobj stays in queue_ctx */
1953 free_percpu(q
->queue_ctx
);
1956 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
1958 struct request_queue
*uninit_q
, *q
;
1960 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
1962 return ERR_PTR(-ENOMEM
);
1964 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
1966 blk_cleanup_queue(uninit_q
);
1970 EXPORT_SYMBOL(blk_mq_init_queue
);
1972 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
1973 struct request_queue
*q
)
1975 struct blk_mq_hw_ctx
**hctxs
;
1976 struct blk_mq_ctx __percpu
*ctx
;
1980 ctx
= alloc_percpu(struct blk_mq_ctx
);
1982 return ERR_PTR(-ENOMEM
);
1984 hctxs
= kmalloc_node(set
->nr_hw_queues
* sizeof(*hctxs
), GFP_KERNEL
,
1990 map
= blk_mq_make_queue_map(set
);
1994 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1995 int node
= blk_mq_hw_queue_to_node(map
, i
);
1997 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
2002 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
2006 atomic_set(&hctxs
[i
]->nr_active
, 0);
2007 hctxs
[i
]->numa_node
= node
;
2008 hctxs
[i
]->queue_num
= i
;
2011 setup_timer(&q
->timeout
, blk_mq_rq_timer
, (unsigned long) q
);
2012 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2014 q
->nr_queues
= nr_cpu_ids
;
2015 q
->nr_hw_queues
= set
->nr_hw_queues
;
2019 q
->queue_hw_ctx
= hctxs
;
2021 q
->mq_ops
= set
->ops
;
2022 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2024 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2025 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
2027 q
->sg_reserved_size
= INT_MAX
;
2029 INIT_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2030 INIT_LIST_HEAD(&q
->requeue_list
);
2031 spin_lock_init(&q
->requeue_lock
);
2033 if (q
->nr_hw_queues
> 1)
2034 blk_queue_make_request(q
, blk_mq_make_request
);
2036 blk_queue_make_request(q
, blk_sq_make_request
);
2039 * Do this after blk_queue_make_request() overrides it...
2041 q
->nr_requests
= set
->queue_depth
;
2043 if (set
->ops
->complete
)
2044 blk_queue_softirq_done(q
, set
->ops
->complete
);
2046 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2048 if (blk_mq_init_hw_queues(q
, set
))
2052 mutex_lock(&all_q_mutex
);
2054 list_add_tail(&q
->all_q_node
, &all_q_list
);
2055 blk_mq_add_queue_tag_set(set
, q
);
2056 blk_mq_map_swqueue(q
, cpu_online_mask
);
2058 mutex_unlock(&all_q_mutex
);
2065 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2068 free_cpumask_var(hctxs
[i
]->cpumask
);
2075 return ERR_PTR(-ENOMEM
);
2077 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2079 void blk_mq_free_queue(struct request_queue
*q
)
2081 struct blk_mq_tag_set
*set
= q
->tag_set
;
2083 mutex_lock(&all_q_mutex
);
2084 list_del_init(&q
->all_q_node
);
2085 mutex_unlock(&all_q_mutex
);
2087 blk_mq_del_queue_tag_set(q
);
2089 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2090 blk_mq_free_hw_queues(q
, set
);
2093 /* Basically redo blk_mq_init_queue with queue frozen */
2094 static void blk_mq_queue_reinit(struct request_queue
*q
,
2095 const struct cpumask
*online_mask
)
2097 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2099 blk_mq_sysfs_unregister(q
);
2101 blk_mq_update_queue_map(q
->mq_map
, q
->nr_hw_queues
, online_mask
);
2104 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2105 * we should change hctx numa_node according to new topology (this
2106 * involves free and re-allocate memory, worthy doing?)
2109 blk_mq_map_swqueue(q
, online_mask
);
2111 blk_mq_sysfs_register(q
);
2114 static int blk_mq_queue_reinit_notify(struct notifier_block
*nb
,
2115 unsigned long action
, void *hcpu
)
2117 struct request_queue
*q
;
2118 int cpu
= (unsigned long)hcpu
;
2120 * New online cpumask which is going to be set in this hotplug event.
2121 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2122 * one-by-one and dynamically allocating this could result in a failure.
2124 static struct cpumask online_new
;
2127 * Before hotadded cpu starts handling requests, new mappings must
2128 * be established. Otherwise, these requests in hw queue might
2129 * never be dispatched.
2131 * For example, there is a single hw queue (hctx) and two CPU queues
2132 * (ctx0 for CPU0, and ctx1 for CPU1).
2134 * Now CPU1 is just onlined and a request is inserted into
2135 * ctx1->rq_list and set bit0 in pending bitmap as ctx1->index_hw is
2138 * And then while running hw queue, flush_busy_ctxs() finds bit0 is
2139 * set in pending bitmap and tries to retrieve requests in
2140 * hctx->ctxs[0]->rq_list. But htx->ctxs[0] is a pointer to ctx0,
2141 * so the request in ctx1->rq_list is ignored.
2143 switch (action
& ~CPU_TASKS_FROZEN
) {
2145 case CPU_UP_CANCELED
:
2146 cpumask_copy(&online_new
, cpu_online_mask
);
2148 case CPU_UP_PREPARE
:
2149 cpumask_copy(&online_new
, cpu_online_mask
);
2150 cpumask_set_cpu(cpu
, &online_new
);
2156 mutex_lock(&all_q_mutex
);
2159 * We need to freeze and reinit all existing queues. Freezing
2160 * involves synchronous wait for an RCU grace period and doing it
2161 * one by one may take a long time. Start freezing all queues in
2162 * one swoop and then wait for the completions so that freezing can
2163 * take place in parallel.
2165 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2166 blk_mq_freeze_queue_start(q
);
2167 list_for_each_entry(q
, &all_q_list
, all_q_node
) {
2168 blk_mq_freeze_queue_wait(q
);
2171 * timeout handler can't touch hw queue during the
2174 del_timer_sync(&q
->timeout
);
2177 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2178 blk_mq_queue_reinit(q
, &online_new
);
2180 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2181 blk_mq_unfreeze_queue(q
);
2183 mutex_unlock(&all_q_mutex
);
2187 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2191 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2192 set
->tags
[i
] = blk_mq_init_rq_map(set
, i
);
2201 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2207 * Allocate the request maps associated with this tag_set. Note that this
2208 * may reduce the depth asked for, if memory is tight. set->queue_depth
2209 * will be updated to reflect the allocated depth.
2211 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2216 depth
= set
->queue_depth
;
2218 err
= __blk_mq_alloc_rq_maps(set
);
2222 set
->queue_depth
>>= 1;
2223 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2227 } while (set
->queue_depth
);
2229 if (!set
->queue_depth
|| err
) {
2230 pr_err("blk-mq: failed to allocate request map\n");
2234 if (depth
!= set
->queue_depth
)
2235 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2236 depth
, set
->queue_depth
);
2241 struct cpumask
*blk_mq_tags_cpumask(struct blk_mq_tags
*tags
)
2243 return tags
->cpumask
;
2245 EXPORT_SYMBOL_GPL(blk_mq_tags_cpumask
);
2248 * Alloc a tag set to be associated with one or more request queues.
2249 * May fail with EINVAL for various error conditions. May adjust the
2250 * requested depth down, if if it too large. In that case, the set
2251 * value will be stored in set->queue_depth.
2253 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2255 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2257 if (!set
->nr_hw_queues
)
2259 if (!set
->queue_depth
)
2261 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2264 if (!set
->ops
->queue_rq
|| !set
->ops
->map_queue
)
2267 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2268 pr_info("blk-mq: reduced tag depth to %u\n",
2270 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2274 * If a crashdump is active, then we are potentially in a very
2275 * memory constrained environment. Limit us to 1 queue and
2276 * 64 tags to prevent using too much memory.
2278 if (is_kdump_kernel()) {
2279 set
->nr_hw_queues
= 1;
2280 set
->queue_depth
= min(64U, set
->queue_depth
);
2283 set
->tags
= kmalloc_node(set
->nr_hw_queues
*
2284 sizeof(struct blk_mq_tags
*),
2285 GFP_KERNEL
, set
->numa_node
);
2289 if (blk_mq_alloc_rq_maps(set
))
2292 mutex_init(&set
->tag_list_lock
);
2293 INIT_LIST_HEAD(&set
->tag_list
);
2301 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2303 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2307 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2309 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2315 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2317 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2319 struct blk_mq_tag_set
*set
= q
->tag_set
;
2320 struct blk_mq_hw_ctx
*hctx
;
2323 if (!set
|| nr
> set
->queue_depth
)
2327 queue_for_each_hw_ctx(q
, hctx
, i
) {
2328 ret
= blk_mq_tag_update_depth(hctx
->tags
, nr
);
2334 q
->nr_requests
= nr
;
2339 void blk_mq_disable_hotplug(void)
2341 mutex_lock(&all_q_mutex
);
2344 void blk_mq_enable_hotplug(void)
2346 mutex_unlock(&all_q_mutex
);
2349 static int __init
blk_mq_init(void)
2353 hotcpu_notifier(blk_mq_queue_reinit_notify
, 0);
2357 subsys_initcall(blk_mq_init
);