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>
13 #include <linux/init.h>
14 #include <linux/slab.h>
15 #include <linux/workqueue.h>
16 #include <linux/smp.h>
17 #include <linux/llist.h>
18 #include <linux/list_sort.h>
19 #include <linux/cpu.h>
20 #include <linux/cache.h>
21 #include <linux/sched/sysctl.h>
22 #include <linux/delay.h>
23 #include <linux/crash_dump.h>
25 #include <trace/events/block.h>
27 #include <linux/blk-mq.h>
30 #include "blk-mq-tag.h"
32 static DEFINE_MUTEX(all_q_mutex
);
33 static LIST_HEAD(all_q_list
);
35 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
);
38 * Check if any of the ctx's have pending work in this hardware queue
40 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
44 for (i
= 0; i
< hctx
->ctx_map
.map_size
; i
++)
45 if (hctx
->ctx_map
.map
[i
].word
)
51 static inline struct blk_align_bitmap
*get_bm(struct blk_mq_hw_ctx
*hctx
,
52 struct blk_mq_ctx
*ctx
)
54 return &hctx
->ctx_map
.map
[ctx
->index_hw
/ hctx
->ctx_map
.bits_per_word
];
57 #define CTX_TO_BIT(hctx, ctx) \
58 ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
61 * Mark this ctx as having pending work in this hardware queue
63 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
64 struct blk_mq_ctx
*ctx
)
66 struct blk_align_bitmap
*bm
= get_bm(hctx
, ctx
);
68 if (!test_bit(CTX_TO_BIT(hctx
, ctx
), &bm
->word
))
69 set_bit(CTX_TO_BIT(hctx
, ctx
), &bm
->word
);
72 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
73 struct blk_mq_ctx
*ctx
)
75 struct blk_align_bitmap
*bm
= get_bm(hctx
, ctx
);
77 clear_bit(CTX_TO_BIT(hctx
, ctx
), &bm
->word
);
80 static int blk_mq_queue_enter(struct request_queue
*q
)
85 if (percpu_ref_tryget_live(&q
->mq_usage_counter
))
88 ret
= wait_event_interruptible(q
->mq_freeze_wq
,
89 !q
->mq_freeze_depth
|| blk_queue_dying(q
));
90 if (blk_queue_dying(q
))
97 static void blk_mq_queue_exit(struct request_queue
*q
)
99 percpu_ref_put(&q
->mq_usage_counter
);
102 static void blk_mq_usage_counter_release(struct percpu_ref
*ref
)
104 struct request_queue
*q
=
105 container_of(ref
, struct request_queue
, mq_usage_counter
);
107 wake_up_all(&q
->mq_freeze_wq
);
111 * Guarantee no request is in use, so we can change any data structure of
112 * the queue afterward.
114 void blk_mq_freeze_queue(struct request_queue
*q
)
118 spin_lock_irq(q
->queue_lock
);
119 freeze
= !q
->mq_freeze_depth
++;
120 spin_unlock_irq(q
->queue_lock
);
123 percpu_ref_kill(&q
->mq_usage_counter
);
124 blk_mq_run_queues(q
, false);
126 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->mq_usage_counter
));
129 static void blk_mq_unfreeze_queue(struct request_queue
*q
)
133 spin_lock_irq(q
->queue_lock
);
134 wake
= !--q
->mq_freeze_depth
;
135 WARN_ON_ONCE(q
->mq_freeze_depth
< 0);
136 spin_unlock_irq(q
->queue_lock
);
138 percpu_ref_reinit(&q
->mq_usage_counter
);
139 wake_up_all(&q
->mq_freeze_wq
);
143 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
145 return blk_mq_has_free_tags(hctx
->tags
);
147 EXPORT_SYMBOL(blk_mq_can_queue
);
149 static void blk_mq_rq_ctx_init(struct request_queue
*q
, struct blk_mq_ctx
*ctx
,
150 struct request
*rq
, unsigned int rw_flags
)
152 if (blk_queue_io_stat(q
))
153 rw_flags
|= REQ_IO_STAT
;
155 INIT_LIST_HEAD(&rq
->queuelist
);
156 /* csd/requeue_work/fifo_time is initialized before use */
159 rq
->cmd_flags
|= rw_flags
;
160 /* do not touch atomic flags, it needs atomic ops against the timer */
162 INIT_HLIST_NODE(&rq
->hash
);
163 RB_CLEAR_NODE(&rq
->rb_node
);
166 rq
->start_time
= jiffies
;
167 #ifdef CONFIG_BLK_CGROUP
169 set_start_time_ns(rq
);
170 rq
->io_start_time_ns
= 0;
172 rq
->nr_phys_segments
= 0;
173 #if defined(CONFIG_BLK_DEV_INTEGRITY)
174 rq
->nr_integrity_segments
= 0;
177 /* tag was already set */
187 INIT_LIST_HEAD(&rq
->timeout_list
);
191 rq
->end_io_data
= NULL
;
194 ctx
->rq_dispatched
[rw_is_sync(rw_flags
)]++;
197 static struct request
*
198 __blk_mq_alloc_request(struct blk_mq_alloc_data
*data
, int rw
)
203 tag
= blk_mq_get_tag(data
);
204 if (tag
!= BLK_MQ_TAG_FAIL
) {
205 rq
= data
->hctx
->tags
->rqs
[tag
];
207 if (blk_mq_tag_busy(data
->hctx
)) {
208 rq
->cmd_flags
= REQ_MQ_INFLIGHT
;
209 atomic_inc(&data
->hctx
->nr_active
);
213 blk_mq_rq_ctx_init(data
->q
, data
->ctx
, rq
, rw
);
220 struct request
*blk_mq_alloc_request(struct request_queue
*q
, int rw
, gfp_t gfp
,
223 struct blk_mq_ctx
*ctx
;
224 struct blk_mq_hw_ctx
*hctx
;
226 struct blk_mq_alloc_data alloc_data
;
229 ret
= blk_mq_queue_enter(q
);
233 ctx
= blk_mq_get_ctx(q
);
234 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
235 blk_mq_set_alloc_data(&alloc_data
, q
, gfp
& ~__GFP_WAIT
,
236 reserved
, ctx
, hctx
);
238 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
239 if (!rq
&& (gfp
& __GFP_WAIT
)) {
240 __blk_mq_run_hw_queue(hctx
);
243 ctx
= blk_mq_get_ctx(q
);
244 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
245 blk_mq_set_alloc_data(&alloc_data
, q
, gfp
, reserved
, ctx
,
247 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
248 ctx
= alloc_data
.ctx
;
252 return ERR_PTR(-EWOULDBLOCK
);
255 EXPORT_SYMBOL(blk_mq_alloc_request
);
257 static void __blk_mq_free_request(struct blk_mq_hw_ctx
*hctx
,
258 struct blk_mq_ctx
*ctx
, struct request
*rq
)
260 const int tag
= rq
->tag
;
261 struct request_queue
*q
= rq
->q
;
263 if (rq
->cmd_flags
& REQ_MQ_INFLIGHT
)
264 atomic_dec(&hctx
->nr_active
);
267 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
268 blk_mq_put_tag(hctx
, tag
, &ctx
->last_tag
);
269 blk_mq_queue_exit(q
);
272 void blk_mq_free_request(struct request
*rq
)
274 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
275 struct blk_mq_hw_ctx
*hctx
;
276 struct request_queue
*q
= rq
->q
;
278 ctx
->rq_completed
[rq_is_sync(rq
)]++;
280 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
281 __blk_mq_free_request(hctx
, ctx
, rq
);
283 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
285 inline void __blk_mq_end_request(struct request
*rq
, int error
)
287 blk_account_io_done(rq
);
290 rq
->end_io(rq
, error
);
292 if (unlikely(blk_bidi_rq(rq
)))
293 blk_mq_free_request(rq
->next_rq
);
294 blk_mq_free_request(rq
);
297 EXPORT_SYMBOL(__blk_mq_end_request
);
299 void blk_mq_end_request(struct request
*rq
, int error
)
301 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
303 __blk_mq_end_request(rq
, error
);
305 EXPORT_SYMBOL(blk_mq_end_request
);
307 static void __blk_mq_complete_request_remote(void *data
)
309 struct request
*rq
= data
;
311 rq
->q
->softirq_done_fn(rq
);
314 static void blk_mq_ipi_complete_request(struct request
*rq
)
316 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
320 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
321 rq
->q
->softirq_done_fn(rq
);
326 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
327 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
329 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
330 rq
->csd
.func
= __blk_mq_complete_request_remote
;
333 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
335 rq
->q
->softirq_done_fn(rq
);
340 void __blk_mq_complete_request(struct request
*rq
)
342 struct request_queue
*q
= rq
->q
;
344 if (!q
->softirq_done_fn
)
345 blk_mq_end_request(rq
, rq
->errors
);
347 blk_mq_ipi_complete_request(rq
);
351 * blk_mq_complete_request - end I/O on a request
352 * @rq: the request being processed
355 * Ends all I/O on a request. It does not handle partial completions.
356 * The actual completion happens out-of-order, through a IPI handler.
358 void blk_mq_complete_request(struct request
*rq
)
360 struct request_queue
*q
= rq
->q
;
362 if (unlikely(blk_should_fake_timeout(q
)))
364 if (!blk_mark_rq_complete(rq
))
365 __blk_mq_complete_request(rq
);
367 EXPORT_SYMBOL(blk_mq_complete_request
);
369 void blk_mq_start_request(struct request
*rq
)
371 struct request_queue
*q
= rq
->q
;
373 trace_block_rq_issue(q
, rq
);
375 rq
->resid_len
= blk_rq_bytes(rq
);
376 if (unlikely(blk_bidi_rq(rq
)))
377 rq
->next_rq
->resid_len
= blk_rq_bytes(rq
->next_rq
);
382 * Ensure that ->deadline is visible before set the started
383 * flag and clear the completed flag.
385 smp_mb__before_atomic();
388 * Mark us as started and clear complete. Complete might have been
389 * set if requeue raced with timeout, which then marked it as
390 * complete. So be sure to clear complete again when we start
391 * the request, otherwise we'll ignore the completion event.
393 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
394 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
395 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
396 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
398 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
400 * Make sure space for the drain appears. We know we can do
401 * this because max_hw_segments has been adjusted to be one
402 * fewer than the device can handle.
404 rq
->nr_phys_segments
++;
407 EXPORT_SYMBOL(blk_mq_start_request
);
409 static void __blk_mq_requeue_request(struct request
*rq
)
411 struct request_queue
*q
= rq
->q
;
413 trace_block_rq_requeue(q
, rq
);
415 if (test_and_clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
416 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
417 rq
->nr_phys_segments
--;
421 void blk_mq_requeue_request(struct request
*rq
)
423 __blk_mq_requeue_request(rq
);
425 BUG_ON(blk_queued_rq(rq
));
426 blk_mq_add_to_requeue_list(rq
, true);
428 EXPORT_SYMBOL(blk_mq_requeue_request
);
430 static void blk_mq_requeue_work(struct work_struct
*work
)
432 struct request_queue
*q
=
433 container_of(work
, struct request_queue
, requeue_work
);
435 struct request
*rq
, *next
;
438 spin_lock_irqsave(&q
->requeue_lock
, flags
);
439 list_splice_init(&q
->requeue_list
, &rq_list
);
440 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
442 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
443 if (!(rq
->cmd_flags
& REQ_SOFTBARRIER
))
446 rq
->cmd_flags
&= ~REQ_SOFTBARRIER
;
447 list_del_init(&rq
->queuelist
);
448 blk_mq_insert_request(rq
, true, false, false);
451 while (!list_empty(&rq_list
)) {
452 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
453 list_del_init(&rq
->queuelist
);
454 blk_mq_insert_request(rq
, false, false, false);
458 * Use the start variant of queue running here, so that running
459 * the requeue work will kick stopped queues.
461 blk_mq_start_hw_queues(q
);
464 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
)
466 struct request_queue
*q
= rq
->q
;
470 * We abuse this flag that is otherwise used by the I/O scheduler to
471 * request head insertation from the workqueue.
473 BUG_ON(rq
->cmd_flags
& REQ_SOFTBARRIER
);
475 spin_lock_irqsave(&q
->requeue_lock
, flags
);
477 rq
->cmd_flags
|= REQ_SOFTBARRIER
;
478 list_add(&rq
->queuelist
, &q
->requeue_list
);
480 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
482 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
484 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
486 void blk_mq_kick_requeue_list(struct request_queue
*q
)
488 kblockd_schedule_work(&q
->requeue_work
);
490 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
492 static inline bool is_flush_request(struct request
*rq
,
493 struct blk_flush_queue
*fq
, unsigned int tag
)
495 return ((rq
->cmd_flags
& REQ_FLUSH_SEQ
) &&
496 fq
->flush_rq
->tag
== tag
);
499 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
501 struct request
*rq
= tags
->rqs
[tag
];
502 /* mq_ctx of flush rq is always cloned from the corresponding req */
503 struct blk_flush_queue
*fq
= blk_get_flush_queue(rq
->q
, rq
->mq_ctx
);
505 if (!is_flush_request(rq
, fq
, tag
))
510 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
512 struct blk_mq_timeout_data
{
514 unsigned int next_set
;
517 void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
519 struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
520 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
523 * We know that complete is set at this point. If STARTED isn't set
524 * anymore, then the request isn't active and the "timeout" should
525 * just be ignored. This can happen due to the bitflag ordering.
526 * Timeout first checks if STARTED is set, and if it is, assumes
527 * the request is active. But if we race with completion, then
528 * we both flags will get cleared. So check here again, and ignore
529 * a timeout event with a request that isn't active.
531 if (!test_bit(REQ_ATOM_STARTED
, &req
->atomic_flags
))
535 ret
= ops
->timeout(req
, reserved
);
539 __blk_mq_complete_request(req
);
541 case BLK_EH_RESET_TIMER
:
543 blk_clear_rq_complete(req
);
545 case BLK_EH_NOT_HANDLED
:
548 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
553 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
554 struct request
*rq
, void *priv
, bool reserved
)
556 struct blk_mq_timeout_data
*data
= priv
;
558 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
561 if (time_after_eq(jiffies
, rq
->deadline
)) {
562 if (!blk_mark_rq_complete(rq
))
563 blk_mq_rq_timed_out(rq
, reserved
);
564 } else if (!data
->next_set
|| time_after(data
->next
, rq
->deadline
)) {
565 data
->next
= rq
->deadline
;
570 static void blk_mq_rq_timer(unsigned long priv
)
572 struct request_queue
*q
= (struct request_queue
*)priv
;
573 struct blk_mq_timeout_data data
= {
577 struct blk_mq_hw_ctx
*hctx
;
580 queue_for_each_hw_ctx(q
, hctx
, i
) {
582 * If not software queues are currently mapped to this
583 * hardware queue, there's nothing to check
585 if (!hctx
->nr_ctx
|| !hctx
->tags
)
588 blk_mq_tag_busy_iter(hctx
, blk_mq_check_expired
, &data
);
592 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
593 mod_timer(&q
->timeout
, data
.next
);
595 queue_for_each_hw_ctx(q
, hctx
, i
)
596 blk_mq_tag_idle(hctx
);
601 * Reverse check our software queue for entries that we could potentially
602 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
603 * too much time checking for merges.
605 static bool blk_mq_attempt_merge(struct request_queue
*q
,
606 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
611 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
617 if (!blk_rq_merge_ok(rq
, bio
))
620 el_ret
= blk_try_merge(rq
, bio
);
621 if (el_ret
== ELEVATOR_BACK_MERGE
) {
622 if (bio_attempt_back_merge(q
, rq
, bio
)) {
627 } else if (el_ret
== ELEVATOR_FRONT_MERGE
) {
628 if (bio_attempt_front_merge(q
, rq
, bio
)) {
640 * Process software queues that have been marked busy, splicing them
641 * to the for-dispatch
643 static void flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
645 struct blk_mq_ctx
*ctx
;
648 for (i
= 0; i
< hctx
->ctx_map
.map_size
; i
++) {
649 struct blk_align_bitmap
*bm
= &hctx
->ctx_map
.map
[i
];
650 unsigned int off
, bit
;
656 off
= i
* hctx
->ctx_map
.bits_per_word
;
658 bit
= find_next_bit(&bm
->word
, bm
->depth
, bit
);
659 if (bit
>= bm
->depth
)
662 ctx
= hctx
->ctxs
[bit
+ off
];
663 clear_bit(bit
, &bm
->word
);
664 spin_lock(&ctx
->lock
);
665 list_splice_tail_init(&ctx
->rq_list
, list
);
666 spin_unlock(&ctx
->lock
);
674 * Run this hardware queue, pulling any software queues mapped to it in.
675 * Note that this function currently has various problems around ordering
676 * of IO. In particular, we'd like FIFO behaviour on handling existing
677 * items on the hctx->dispatch list. Ignore that for now.
679 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
681 struct request_queue
*q
= hctx
->queue
;
684 LIST_HEAD(driver_list
);
685 struct list_head
*dptr
;
688 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
));
690 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)))
696 * Touch any software queue that has pending entries.
698 flush_busy_ctxs(hctx
, &rq_list
);
701 * If we have previous entries on our dispatch list, grab them
702 * and stuff them at the front for more fair dispatch.
704 if (!list_empty_careful(&hctx
->dispatch
)) {
705 spin_lock(&hctx
->lock
);
706 if (!list_empty(&hctx
->dispatch
))
707 list_splice_init(&hctx
->dispatch
, &rq_list
);
708 spin_unlock(&hctx
->lock
);
712 * Start off with dptr being NULL, so we start the first request
713 * immediately, even if we have more pending.
718 * Now process all the entries, sending them to the driver.
721 while (!list_empty(&rq_list
)) {
722 struct blk_mq_queue_data bd
;
725 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
726 list_del_init(&rq
->queuelist
);
730 bd
.last
= list_empty(&rq_list
);
732 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
734 case BLK_MQ_RQ_QUEUE_OK
:
737 case BLK_MQ_RQ_QUEUE_BUSY
:
738 list_add(&rq
->queuelist
, &rq_list
);
739 __blk_mq_requeue_request(rq
);
742 pr_err("blk-mq: bad return on queue: %d\n", ret
);
743 case BLK_MQ_RQ_QUEUE_ERROR
:
745 blk_mq_end_request(rq
, rq
->errors
);
749 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
753 * We've done the first request. If we have more than 1
754 * left in the list, set dptr to defer issue.
756 if (!dptr
&& rq_list
.next
!= rq_list
.prev
)
761 hctx
->dispatched
[0]++;
762 else if (queued
< (1 << (BLK_MQ_MAX_DISPATCH_ORDER
- 1)))
763 hctx
->dispatched
[ilog2(queued
) + 1]++;
766 * Any items that need requeuing? Stuff them into hctx->dispatch,
767 * that is where we will continue on next queue run.
769 if (!list_empty(&rq_list
)) {
770 spin_lock(&hctx
->lock
);
771 list_splice(&rq_list
, &hctx
->dispatch
);
772 spin_unlock(&hctx
->lock
);
777 * It'd be great if the workqueue API had a way to pass
778 * in a mask and had some smarts for more clever placement.
779 * For now we just round-robin here, switching for every
780 * BLK_MQ_CPU_WORK_BATCH queued items.
782 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
784 int cpu
= hctx
->next_cpu
;
786 if (--hctx
->next_cpu_batch
<= 0) {
789 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
790 if (next_cpu
>= nr_cpu_ids
)
791 next_cpu
= cpumask_first(hctx
->cpumask
);
793 hctx
->next_cpu
= next_cpu
;
794 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
800 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
802 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)))
807 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
808 __blk_mq_run_hw_queue(hctx
);
816 if (hctx
->queue
->nr_hw_queues
== 1)
817 kblockd_schedule_delayed_work(&hctx
->run_work
, 0);
821 cpu
= blk_mq_hctx_next_cpu(hctx
);
822 kblockd_schedule_delayed_work_on(cpu
, &hctx
->run_work
, 0);
826 void blk_mq_run_queues(struct request_queue
*q
, bool async
)
828 struct blk_mq_hw_ctx
*hctx
;
831 queue_for_each_hw_ctx(q
, hctx
, i
) {
832 if ((!blk_mq_hctx_has_pending(hctx
) &&
833 list_empty_careful(&hctx
->dispatch
)) ||
834 test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
837 blk_mq_run_hw_queue(hctx
, async
);
840 EXPORT_SYMBOL(blk_mq_run_queues
);
842 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
844 cancel_delayed_work(&hctx
->run_work
);
845 cancel_delayed_work(&hctx
->delay_work
);
846 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
848 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
850 void blk_mq_stop_hw_queues(struct request_queue
*q
)
852 struct blk_mq_hw_ctx
*hctx
;
855 queue_for_each_hw_ctx(q
, hctx
, i
)
856 blk_mq_stop_hw_queue(hctx
);
858 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
860 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
862 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
864 blk_mq_run_hw_queue(hctx
, false);
866 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
868 void blk_mq_start_hw_queues(struct request_queue
*q
)
870 struct blk_mq_hw_ctx
*hctx
;
873 queue_for_each_hw_ctx(q
, hctx
, i
)
874 blk_mq_start_hw_queue(hctx
);
876 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
879 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
881 struct blk_mq_hw_ctx
*hctx
;
884 queue_for_each_hw_ctx(q
, hctx
, i
) {
885 if (!test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
888 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
889 blk_mq_run_hw_queue(hctx
, async
);
892 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
894 static void blk_mq_run_work_fn(struct work_struct
*work
)
896 struct blk_mq_hw_ctx
*hctx
;
898 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
900 __blk_mq_run_hw_queue(hctx
);
903 static void blk_mq_delay_work_fn(struct work_struct
*work
)
905 struct blk_mq_hw_ctx
*hctx
;
907 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
909 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
910 __blk_mq_run_hw_queue(hctx
);
913 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
915 unsigned long tmo
= msecs_to_jiffies(msecs
);
917 if (hctx
->queue
->nr_hw_queues
== 1)
918 kblockd_schedule_delayed_work(&hctx
->delay_work
, tmo
);
922 cpu
= blk_mq_hctx_next_cpu(hctx
);
923 kblockd_schedule_delayed_work_on(cpu
, &hctx
->delay_work
, tmo
);
926 EXPORT_SYMBOL(blk_mq_delay_queue
);
928 static void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
,
929 struct request
*rq
, bool at_head
)
931 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
933 trace_block_rq_insert(hctx
->queue
, rq
);
936 list_add(&rq
->queuelist
, &ctx
->rq_list
);
938 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
940 blk_mq_hctx_mark_pending(hctx
, ctx
);
943 void blk_mq_insert_request(struct request
*rq
, bool at_head
, bool run_queue
,
946 struct request_queue
*q
= rq
->q
;
947 struct blk_mq_hw_ctx
*hctx
;
948 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
, *current_ctx
;
950 current_ctx
= blk_mq_get_ctx(q
);
951 if (!cpu_online(ctx
->cpu
))
952 rq
->mq_ctx
= ctx
= current_ctx
;
954 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
956 spin_lock(&ctx
->lock
);
957 __blk_mq_insert_request(hctx
, rq
, at_head
);
958 spin_unlock(&ctx
->lock
);
961 blk_mq_run_hw_queue(hctx
, async
);
963 blk_mq_put_ctx(current_ctx
);
966 static void blk_mq_insert_requests(struct request_queue
*q
,
967 struct blk_mq_ctx
*ctx
,
968 struct list_head
*list
,
973 struct blk_mq_hw_ctx
*hctx
;
974 struct blk_mq_ctx
*current_ctx
;
976 trace_block_unplug(q
, depth
, !from_schedule
);
978 current_ctx
= blk_mq_get_ctx(q
);
980 if (!cpu_online(ctx
->cpu
))
982 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
985 * preemption doesn't flush plug list, so it's possible ctx->cpu is
988 spin_lock(&ctx
->lock
);
989 while (!list_empty(list
)) {
992 rq
= list_first_entry(list
, struct request
, queuelist
);
993 list_del_init(&rq
->queuelist
);
995 __blk_mq_insert_request(hctx
, rq
, false);
997 spin_unlock(&ctx
->lock
);
999 blk_mq_run_hw_queue(hctx
, from_schedule
);
1000 blk_mq_put_ctx(current_ctx
);
1003 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1005 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1006 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1008 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1009 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1010 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1013 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1015 struct blk_mq_ctx
*this_ctx
;
1016 struct request_queue
*this_q
;
1019 LIST_HEAD(ctx_list
);
1022 list_splice_init(&plug
->mq_list
, &list
);
1024 list_sort(NULL
, &list
, plug_ctx_cmp
);
1030 while (!list_empty(&list
)) {
1031 rq
= list_entry_rq(list
.next
);
1032 list_del_init(&rq
->queuelist
);
1034 if (rq
->mq_ctx
!= this_ctx
) {
1036 blk_mq_insert_requests(this_q
, this_ctx
,
1041 this_ctx
= rq
->mq_ctx
;
1047 list_add_tail(&rq
->queuelist
, &ctx_list
);
1051 * If 'this_ctx' is set, we know we have entries to complete
1052 * on 'ctx_list'. Do those.
1055 blk_mq_insert_requests(this_q
, this_ctx
, &ctx_list
, depth
,
1060 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1062 init_request_from_bio(rq
, bio
);
1064 if (blk_do_io_stat(rq
))
1065 blk_account_io_start(rq
, 1);
1068 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1070 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1071 !blk_queue_nomerges(hctx
->queue
);
1074 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx
*hctx
,
1075 struct blk_mq_ctx
*ctx
,
1076 struct request
*rq
, struct bio
*bio
)
1078 if (!hctx_allow_merges(hctx
)) {
1079 blk_mq_bio_to_request(rq
, bio
);
1080 spin_lock(&ctx
->lock
);
1082 __blk_mq_insert_request(hctx
, rq
, false);
1083 spin_unlock(&ctx
->lock
);
1086 struct request_queue
*q
= hctx
->queue
;
1088 spin_lock(&ctx
->lock
);
1089 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1090 blk_mq_bio_to_request(rq
, bio
);
1094 spin_unlock(&ctx
->lock
);
1095 __blk_mq_free_request(hctx
, ctx
, rq
);
1100 struct blk_map_ctx
{
1101 struct blk_mq_hw_ctx
*hctx
;
1102 struct blk_mq_ctx
*ctx
;
1105 static struct request
*blk_mq_map_request(struct request_queue
*q
,
1107 struct blk_map_ctx
*data
)
1109 struct blk_mq_hw_ctx
*hctx
;
1110 struct blk_mq_ctx
*ctx
;
1112 int rw
= bio_data_dir(bio
);
1113 struct blk_mq_alloc_data alloc_data
;
1115 if (unlikely(blk_mq_queue_enter(q
))) {
1116 bio_endio(bio
, -EIO
);
1120 ctx
= blk_mq_get_ctx(q
);
1121 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1123 if (rw_is_sync(bio
->bi_rw
))
1126 trace_block_getrq(q
, bio
, rw
);
1127 blk_mq_set_alloc_data(&alloc_data
, q
, GFP_ATOMIC
, false, ctx
,
1129 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
1130 if (unlikely(!rq
)) {
1131 __blk_mq_run_hw_queue(hctx
);
1132 blk_mq_put_ctx(ctx
);
1133 trace_block_sleeprq(q
, bio
, rw
);
1135 ctx
= blk_mq_get_ctx(q
);
1136 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1137 blk_mq_set_alloc_data(&alloc_data
, q
,
1138 __GFP_WAIT
|GFP_ATOMIC
, false, ctx
, hctx
);
1139 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
1140 ctx
= alloc_data
.ctx
;
1141 hctx
= alloc_data
.hctx
;
1151 * Multiple hardware queue variant. This will not use per-process plugs,
1152 * but will attempt to bypass the hctx queueing if we can go straight to
1153 * hardware for SYNC IO.
1155 static void blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1157 const int is_sync
= rw_is_sync(bio
->bi_rw
);
1158 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
1159 struct blk_map_ctx data
;
1162 blk_queue_bounce(q
, &bio
);
1164 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1165 bio_endio(bio
, -EIO
);
1169 rq
= blk_mq_map_request(q
, bio
, &data
);
1173 if (unlikely(is_flush_fua
)) {
1174 blk_mq_bio_to_request(rq
, bio
);
1175 blk_insert_flush(rq
);
1180 * If the driver supports defer issued based on 'last', then
1181 * queue it up like normal since we can potentially save some
1184 if (is_sync
&& !(data
.hctx
->flags
& BLK_MQ_F_DEFER_ISSUE
)) {
1185 struct blk_mq_queue_data bd
= {
1192 blk_mq_bio_to_request(rq
, bio
);
1195 * For OK queue, we are done. For error, kill it. Any other
1196 * error (busy), just add it to our list as we previously
1199 ret
= q
->mq_ops
->queue_rq(data
.hctx
, &bd
);
1200 if (ret
== BLK_MQ_RQ_QUEUE_OK
)
1203 __blk_mq_requeue_request(rq
);
1205 if (ret
== BLK_MQ_RQ_QUEUE_ERROR
) {
1207 blk_mq_end_request(rq
, rq
->errors
);
1213 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1215 * For a SYNC request, send it to the hardware immediately. For
1216 * an ASYNC request, just ensure that we run it later on. The
1217 * latter allows for merging opportunities and more efficient
1221 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1224 blk_mq_put_ctx(data
.ctx
);
1228 * Single hardware queue variant. This will attempt to use any per-process
1229 * plug for merging and IO deferral.
1231 static void blk_sq_make_request(struct request_queue
*q
, struct bio
*bio
)
1233 const int is_sync
= rw_is_sync(bio
->bi_rw
);
1234 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
1235 unsigned int use_plug
, request_count
= 0;
1236 struct blk_map_ctx data
;
1240 * If we have multiple hardware queues, just go directly to
1241 * one of those for sync IO.
1243 use_plug
= !is_flush_fua
&& !is_sync
;
1245 blk_queue_bounce(q
, &bio
);
1247 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1248 bio_endio(bio
, -EIO
);
1252 if (use_plug
&& !blk_queue_nomerges(q
) &&
1253 blk_attempt_plug_merge(q
, bio
, &request_count
))
1256 rq
= blk_mq_map_request(q
, bio
, &data
);
1260 if (unlikely(is_flush_fua
)) {
1261 blk_mq_bio_to_request(rq
, bio
);
1262 blk_insert_flush(rq
);
1267 * A task plug currently exists. Since this is completely lockless,
1268 * utilize that to temporarily store requests until the task is
1269 * either done or scheduled away.
1272 struct blk_plug
*plug
= current
->plug
;
1275 blk_mq_bio_to_request(rq
, bio
);
1276 if (list_empty(&plug
->mq_list
))
1277 trace_block_plug(q
);
1278 else if (request_count
>= BLK_MAX_REQUEST_COUNT
) {
1279 blk_flush_plug_list(plug
, false);
1280 trace_block_plug(q
);
1282 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1283 blk_mq_put_ctx(data
.ctx
);
1288 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1290 * For a SYNC request, send it to the hardware immediately. For
1291 * an ASYNC request, just ensure that we run it later on. The
1292 * latter allows for merging opportunities and more efficient
1296 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1299 blk_mq_put_ctx(data
.ctx
);
1303 * Default mapping to a software queue, since we use one per CPU.
1305 struct blk_mq_hw_ctx
*blk_mq_map_queue(struct request_queue
*q
, const int cpu
)
1307 return q
->queue_hw_ctx
[q
->mq_map
[cpu
]];
1309 EXPORT_SYMBOL(blk_mq_map_queue
);
1311 static void blk_mq_free_rq_map(struct blk_mq_tag_set
*set
,
1312 struct blk_mq_tags
*tags
, unsigned int hctx_idx
)
1316 if (tags
->rqs
&& set
->ops
->exit_request
) {
1319 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1322 set
->ops
->exit_request(set
->driver_data
, tags
->rqs
[i
],
1324 tags
->rqs
[i
] = NULL
;
1328 while (!list_empty(&tags
->page_list
)) {
1329 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1330 list_del_init(&page
->lru
);
1331 __free_pages(page
, page
->private);
1336 blk_mq_free_tags(tags
);
1339 static size_t order_to_size(unsigned int order
)
1341 return (size_t)PAGE_SIZE
<< order
;
1344 static struct blk_mq_tags
*blk_mq_init_rq_map(struct blk_mq_tag_set
*set
,
1345 unsigned int hctx_idx
)
1347 struct blk_mq_tags
*tags
;
1348 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1349 size_t rq_size
, left
;
1351 tags
= blk_mq_init_tags(set
->queue_depth
, set
->reserved_tags
,
1356 INIT_LIST_HEAD(&tags
->page_list
);
1358 tags
->rqs
= kzalloc_node(set
->queue_depth
* sizeof(struct request
*),
1359 GFP_KERNEL
| __GFP_NOWARN
| __GFP_NORETRY
,
1362 blk_mq_free_tags(tags
);
1367 * rq_size is the size of the request plus driver payload, rounded
1368 * to the cacheline size
1370 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1372 left
= rq_size
* set
->queue_depth
;
1374 for (i
= 0; i
< set
->queue_depth
; ) {
1375 int this_order
= max_order
;
1380 while (left
< order_to_size(this_order
- 1) && this_order
)
1384 page
= alloc_pages_node(set
->numa_node
,
1385 GFP_KERNEL
| __GFP_NOWARN
| __GFP_NORETRY
,
1391 if (order_to_size(this_order
) < rq_size
)
1398 page
->private = this_order
;
1399 list_add_tail(&page
->lru
, &tags
->page_list
);
1401 p
= page_address(page
);
1402 entries_per_page
= order_to_size(this_order
) / rq_size
;
1403 to_do
= min(entries_per_page
, set
->queue_depth
- i
);
1404 left
-= to_do
* rq_size
;
1405 for (j
= 0; j
< to_do
; j
++) {
1407 tags
->rqs
[i
]->atomic_flags
= 0;
1408 tags
->rqs
[i
]->cmd_flags
= 0;
1409 if (set
->ops
->init_request
) {
1410 if (set
->ops
->init_request(set
->driver_data
,
1411 tags
->rqs
[i
], hctx_idx
, i
,
1413 tags
->rqs
[i
] = NULL
;
1426 blk_mq_free_rq_map(set
, tags
, hctx_idx
);
1430 static void blk_mq_free_bitmap(struct blk_mq_ctxmap
*bitmap
)
1435 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap
*bitmap
, int node
)
1437 unsigned int bpw
= 8, total
, num_maps
, i
;
1439 bitmap
->bits_per_word
= bpw
;
1441 num_maps
= ALIGN(nr_cpu_ids
, bpw
) / bpw
;
1442 bitmap
->map
= kzalloc_node(num_maps
* sizeof(struct blk_align_bitmap
),
1447 bitmap
->map_size
= num_maps
;
1450 for (i
= 0; i
< num_maps
; i
++) {
1451 bitmap
->map
[i
].depth
= min(total
, bitmap
->bits_per_word
);
1452 total
-= bitmap
->map
[i
].depth
;
1458 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx
*hctx
, int cpu
)
1460 struct request_queue
*q
= hctx
->queue
;
1461 struct blk_mq_ctx
*ctx
;
1465 * Move ctx entries to new CPU, if this one is going away.
1467 ctx
= __blk_mq_get_ctx(q
, cpu
);
1469 spin_lock(&ctx
->lock
);
1470 if (!list_empty(&ctx
->rq_list
)) {
1471 list_splice_init(&ctx
->rq_list
, &tmp
);
1472 blk_mq_hctx_clear_pending(hctx
, ctx
);
1474 spin_unlock(&ctx
->lock
);
1476 if (list_empty(&tmp
))
1479 ctx
= blk_mq_get_ctx(q
);
1480 spin_lock(&ctx
->lock
);
1482 while (!list_empty(&tmp
)) {
1485 rq
= list_first_entry(&tmp
, struct request
, queuelist
);
1487 list_move_tail(&rq
->queuelist
, &ctx
->rq_list
);
1490 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1491 blk_mq_hctx_mark_pending(hctx
, ctx
);
1493 spin_unlock(&ctx
->lock
);
1495 blk_mq_run_hw_queue(hctx
, true);
1496 blk_mq_put_ctx(ctx
);
1500 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx
*hctx
, int cpu
)
1502 struct request_queue
*q
= hctx
->queue
;
1503 struct blk_mq_tag_set
*set
= q
->tag_set
;
1505 if (set
->tags
[hctx
->queue_num
])
1508 set
->tags
[hctx
->queue_num
] = blk_mq_init_rq_map(set
, hctx
->queue_num
);
1509 if (!set
->tags
[hctx
->queue_num
])
1512 hctx
->tags
= set
->tags
[hctx
->queue_num
];
1516 static int blk_mq_hctx_notify(void *data
, unsigned long action
,
1519 struct blk_mq_hw_ctx
*hctx
= data
;
1521 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
)
1522 return blk_mq_hctx_cpu_offline(hctx
, cpu
);
1523 else if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
)
1524 return blk_mq_hctx_cpu_online(hctx
, cpu
);
1529 static void blk_mq_exit_hctx(struct request_queue
*q
,
1530 struct blk_mq_tag_set
*set
,
1531 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
1533 unsigned flush_start_tag
= set
->queue_depth
;
1535 blk_mq_tag_idle(hctx
);
1537 if (set
->ops
->exit_request
)
1538 set
->ops
->exit_request(set
->driver_data
,
1539 hctx
->fq
->flush_rq
, hctx_idx
,
1540 flush_start_tag
+ hctx_idx
);
1542 if (set
->ops
->exit_hctx
)
1543 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1545 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1546 blk_free_flush_queue(hctx
->fq
);
1548 blk_mq_free_bitmap(&hctx
->ctx_map
);
1551 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1552 struct blk_mq_tag_set
*set
, int nr_queue
)
1554 struct blk_mq_hw_ctx
*hctx
;
1557 queue_for_each_hw_ctx(q
, hctx
, i
) {
1560 blk_mq_exit_hctx(q
, set
, hctx
, i
);
1564 static void blk_mq_free_hw_queues(struct request_queue
*q
,
1565 struct blk_mq_tag_set
*set
)
1567 struct blk_mq_hw_ctx
*hctx
;
1570 queue_for_each_hw_ctx(q
, hctx
, i
) {
1571 free_cpumask_var(hctx
->cpumask
);
1576 static int blk_mq_init_hctx(struct request_queue
*q
,
1577 struct blk_mq_tag_set
*set
,
1578 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
1581 unsigned flush_start_tag
= set
->queue_depth
;
1583 node
= hctx
->numa_node
;
1584 if (node
== NUMA_NO_NODE
)
1585 node
= hctx
->numa_node
= set
->numa_node
;
1587 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1588 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1589 spin_lock_init(&hctx
->lock
);
1590 INIT_LIST_HEAD(&hctx
->dispatch
);
1592 hctx
->queue_num
= hctx_idx
;
1593 hctx
->flags
= set
->flags
;
1594 hctx
->cmd_size
= set
->cmd_size
;
1596 blk_mq_init_cpu_notifier(&hctx
->cpu_notifier
,
1597 blk_mq_hctx_notify
, hctx
);
1598 blk_mq_register_cpu_notifier(&hctx
->cpu_notifier
);
1600 hctx
->tags
= set
->tags
[hctx_idx
];
1603 * Allocate space for all possible cpus to avoid allocation at
1606 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1609 goto unregister_cpu_notifier
;
1611 if (blk_mq_alloc_bitmap(&hctx
->ctx_map
, node
))
1616 if (set
->ops
->init_hctx
&&
1617 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
1620 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
1624 if (set
->ops
->init_request
&&
1625 set
->ops
->init_request(set
->driver_data
,
1626 hctx
->fq
->flush_rq
, hctx_idx
,
1627 flush_start_tag
+ hctx_idx
, node
))
1635 if (set
->ops
->exit_hctx
)
1636 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1638 blk_mq_free_bitmap(&hctx
->ctx_map
);
1641 unregister_cpu_notifier
:
1642 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1647 static int blk_mq_init_hw_queues(struct request_queue
*q
,
1648 struct blk_mq_tag_set
*set
)
1650 struct blk_mq_hw_ctx
*hctx
;
1654 * Initialize hardware queues
1656 queue_for_each_hw_ctx(q
, hctx
, i
) {
1657 if (blk_mq_init_hctx(q
, set
, hctx
, i
))
1661 if (i
== q
->nr_hw_queues
)
1667 blk_mq_exit_hw_queues(q
, set
, i
);
1672 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1673 unsigned int nr_hw_queues
)
1677 for_each_possible_cpu(i
) {
1678 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1679 struct blk_mq_hw_ctx
*hctx
;
1681 memset(__ctx
, 0, sizeof(*__ctx
));
1683 spin_lock_init(&__ctx
->lock
);
1684 INIT_LIST_HEAD(&__ctx
->rq_list
);
1687 /* If the cpu isn't online, the cpu is mapped to first hctx */
1691 hctx
= q
->mq_ops
->map_queue(q
, i
);
1692 cpumask_set_cpu(i
, hctx
->cpumask
);
1696 * Set local node, IFF we have more than one hw queue. If
1697 * not, we remain on the home node of the device
1699 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1700 hctx
->numa_node
= cpu_to_node(i
);
1704 static void blk_mq_map_swqueue(struct request_queue
*q
)
1707 struct blk_mq_hw_ctx
*hctx
;
1708 struct blk_mq_ctx
*ctx
;
1710 queue_for_each_hw_ctx(q
, hctx
, i
) {
1711 cpumask_clear(hctx
->cpumask
);
1716 * Map software to hardware queues
1718 queue_for_each_ctx(q
, ctx
, i
) {
1719 /* If the cpu isn't online, the cpu is mapped to first hctx */
1723 hctx
= q
->mq_ops
->map_queue(q
, i
);
1724 cpumask_set_cpu(i
, hctx
->cpumask
);
1725 ctx
->index_hw
= hctx
->nr_ctx
;
1726 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
1729 queue_for_each_hw_ctx(q
, hctx
, i
) {
1731 * If no software queues are mapped to this hardware queue,
1732 * disable it and free the request entries.
1734 if (!hctx
->nr_ctx
) {
1735 struct blk_mq_tag_set
*set
= q
->tag_set
;
1738 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1739 set
->tags
[i
] = NULL
;
1746 * Initialize batch roundrobin counts
1748 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
1749 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1753 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
)
1755 struct blk_mq_hw_ctx
*hctx
;
1756 struct request_queue
*q
;
1760 if (set
->tag_list
.next
== set
->tag_list
.prev
)
1765 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
1766 blk_mq_freeze_queue(q
);
1768 queue_for_each_hw_ctx(q
, hctx
, i
) {
1770 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
1772 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
1774 blk_mq_unfreeze_queue(q
);
1778 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
1780 struct blk_mq_tag_set
*set
= q
->tag_set
;
1782 mutex_lock(&set
->tag_list_lock
);
1783 list_del_init(&q
->tag_set_list
);
1784 blk_mq_update_tag_set_depth(set
);
1785 mutex_unlock(&set
->tag_list_lock
);
1788 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
1789 struct request_queue
*q
)
1793 mutex_lock(&set
->tag_list_lock
);
1794 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
1795 blk_mq_update_tag_set_depth(set
);
1796 mutex_unlock(&set
->tag_list_lock
);
1799 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
1801 struct blk_mq_hw_ctx
**hctxs
;
1802 struct blk_mq_ctx __percpu
*ctx
;
1803 struct request_queue
*q
;
1807 ctx
= alloc_percpu(struct blk_mq_ctx
);
1809 return ERR_PTR(-ENOMEM
);
1812 * If a crashdump is active, then we are potentially in a very
1813 * memory constrained environment. Limit us to 1 queue and
1814 * 64 tags to prevent using too much memory.
1816 if (is_kdump_kernel()) {
1817 set
->nr_hw_queues
= 1;
1818 set
->queue_depth
= min(64U, set
->queue_depth
);
1821 hctxs
= kmalloc_node(set
->nr_hw_queues
* sizeof(*hctxs
), GFP_KERNEL
,
1827 map
= blk_mq_make_queue_map(set
);
1831 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1832 int node
= blk_mq_hw_queue_to_node(map
, i
);
1834 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
1839 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
1843 atomic_set(&hctxs
[i
]->nr_active
, 0);
1844 hctxs
[i
]->numa_node
= node
;
1845 hctxs
[i
]->queue_num
= i
;
1848 q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
1853 * Init percpu_ref in atomic mode so that it's faster to shutdown.
1854 * See blk_register_queue() for details.
1856 if (percpu_ref_init(&q
->mq_usage_counter
, blk_mq_usage_counter_release
,
1857 PERCPU_REF_INIT_ATOMIC
, GFP_KERNEL
))
1860 setup_timer(&q
->timeout
, blk_mq_rq_timer
, (unsigned long) q
);
1861 blk_queue_rq_timeout(q
, 30000);
1863 q
->nr_queues
= nr_cpu_ids
;
1864 q
->nr_hw_queues
= set
->nr_hw_queues
;
1868 q
->queue_hw_ctx
= hctxs
;
1870 q
->mq_ops
= set
->ops
;
1871 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
1873 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
1874 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
1876 q
->sg_reserved_size
= INT_MAX
;
1878 INIT_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
1879 INIT_LIST_HEAD(&q
->requeue_list
);
1880 spin_lock_init(&q
->requeue_lock
);
1882 if (q
->nr_hw_queues
> 1)
1883 blk_queue_make_request(q
, blk_mq_make_request
);
1885 blk_queue_make_request(q
, blk_sq_make_request
);
1888 blk_queue_rq_timeout(q
, set
->timeout
);
1891 * Do this after blk_queue_make_request() overrides it...
1893 q
->nr_requests
= set
->queue_depth
;
1895 if (set
->ops
->complete
)
1896 blk_queue_softirq_done(q
, set
->ops
->complete
);
1898 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
1900 if (blk_mq_init_hw_queues(q
, set
))
1903 mutex_lock(&all_q_mutex
);
1904 list_add_tail(&q
->all_q_node
, &all_q_list
);
1905 mutex_unlock(&all_q_mutex
);
1907 blk_mq_add_queue_tag_set(set
, q
);
1909 blk_mq_map_swqueue(q
);
1914 blk_cleanup_queue(q
);
1917 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1920 free_cpumask_var(hctxs
[i
]->cpumask
);
1927 return ERR_PTR(-ENOMEM
);
1929 EXPORT_SYMBOL(blk_mq_init_queue
);
1931 void blk_mq_free_queue(struct request_queue
*q
)
1933 struct blk_mq_tag_set
*set
= q
->tag_set
;
1935 blk_mq_del_queue_tag_set(q
);
1937 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
1938 blk_mq_free_hw_queues(q
, set
);
1940 percpu_ref_exit(&q
->mq_usage_counter
);
1942 free_percpu(q
->queue_ctx
);
1943 kfree(q
->queue_hw_ctx
);
1946 q
->queue_ctx
= NULL
;
1947 q
->queue_hw_ctx
= NULL
;
1950 mutex_lock(&all_q_mutex
);
1951 list_del_init(&q
->all_q_node
);
1952 mutex_unlock(&all_q_mutex
);
1955 /* Basically redo blk_mq_init_queue with queue frozen */
1956 static void blk_mq_queue_reinit(struct request_queue
*q
)
1958 blk_mq_freeze_queue(q
);
1960 blk_mq_sysfs_unregister(q
);
1962 blk_mq_update_queue_map(q
->mq_map
, q
->nr_hw_queues
);
1965 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1966 * we should change hctx numa_node according to new topology (this
1967 * involves free and re-allocate memory, worthy doing?)
1970 blk_mq_map_swqueue(q
);
1972 blk_mq_sysfs_register(q
);
1974 blk_mq_unfreeze_queue(q
);
1977 static int blk_mq_queue_reinit_notify(struct notifier_block
*nb
,
1978 unsigned long action
, void *hcpu
)
1980 struct request_queue
*q
;
1983 * Before new mappings are established, hotadded cpu might already
1984 * start handling requests. This doesn't break anything as we map
1985 * offline CPUs to first hardware queue. We will re-init the queue
1986 * below to get optimal settings.
1988 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
&&
1989 action
!= CPU_ONLINE
&& action
!= CPU_ONLINE_FROZEN
)
1992 mutex_lock(&all_q_mutex
);
1993 list_for_each_entry(q
, &all_q_list
, all_q_node
)
1994 blk_mq_queue_reinit(q
);
1995 mutex_unlock(&all_q_mutex
);
1999 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2003 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2004 set
->tags
[i
] = blk_mq_init_rq_map(set
, i
);
2013 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2019 * Allocate the request maps associated with this tag_set. Note that this
2020 * may reduce the depth asked for, if memory is tight. set->queue_depth
2021 * will be updated to reflect the allocated depth.
2023 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2028 depth
= set
->queue_depth
;
2030 err
= __blk_mq_alloc_rq_maps(set
);
2034 set
->queue_depth
>>= 1;
2035 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2039 } while (set
->queue_depth
);
2041 if (!set
->queue_depth
|| err
) {
2042 pr_err("blk-mq: failed to allocate request map\n");
2046 if (depth
!= set
->queue_depth
)
2047 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2048 depth
, set
->queue_depth
);
2054 * Alloc a tag set to be associated with one or more request queues.
2055 * May fail with EINVAL for various error conditions. May adjust the
2056 * requested depth down, if if it too large. In that case, the set
2057 * value will be stored in set->queue_depth.
2059 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2061 if (!set
->nr_hw_queues
)
2063 if (!set
->queue_depth
)
2065 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2068 if (!set
->nr_hw_queues
|| !set
->ops
->queue_rq
|| !set
->ops
->map_queue
)
2071 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2072 pr_info("blk-mq: reduced tag depth to %u\n",
2074 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2077 set
->tags
= kmalloc_node(set
->nr_hw_queues
*
2078 sizeof(struct blk_mq_tags
*),
2079 GFP_KERNEL
, set
->numa_node
);
2083 if (blk_mq_alloc_rq_maps(set
))
2086 mutex_init(&set
->tag_list_lock
);
2087 INIT_LIST_HEAD(&set
->tag_list
);
2095 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2097 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2101 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2103 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2109 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2111 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2113 struct blk_mq_tag_set
*set
= q
->tag_set
;
2114 struct blk_mq_hw_ctx
*hctx
;
2117 if (!set
|| nr
> set
->queue_depth
)
2121 queue_for_each_hw_ctx(q
, hctx
, i
) {
2122 ret
= blk_mq_tag_update_depth(hctx
->tags
, nr
);
2128 q
->nr_requests
= nr
;
2133 void blk_mq_disable_hotplug(void)
2135 mutex_lock(&all_q_mutex
);
2138 void blk_mq_enable_hotplug(void)
2140 mutex_unlock(&all_q_mutex
);
2143 static int __init
blk_mq_init(void)
2147 hotcpu_notifier(blk_mq_queue_reinit_notify
, 0);
2151 subsys_initcall(blk_mq_init
);