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
);
110 void blk_mq_freeze_queue_start(struct request_queue
*q
)
114 spin_lock_irq(q
->queue_lock
);
115 freeze
= !q
->mq_freeze_depth
++;
116 spin_unlock_irq(q
->queue_lock
);
119 percpu_ref_kill(&q
->mq_usage_counter
);
120 blk_mq_run_queues(q
, false);
123 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start
);
125 static void blk_mq_freeze_queue_wait(struct request_queue
*q
)
127 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->mq_usage_counter
));
131 * Guarantee no request is in use, so we can change any data structure of
132 * the queue afterward.
134 void blk_mq_freeze_queue(struct request_queue
*q
)
136 blk_mq_freeze_queue_start(q
);
137 blk_mq_freeze_queue_wait(q
);
140 void blk_mq_unfreeze_queue(struct request_queue
*q
)
144 spin_lock_irq(q
->queue_lock
);
145 wake
= !--q
->mq_freeze_depth
;
146 WARN_ON_ONCE(q
->mq_freeze_depth
< 0);
147 spin_unlock_irq(q
->queue_lock
);
149 percpu_ref_reinit(&q
->mq_usage_counter
);
150 wake_up_all(&q
->mq_freeze_wq
);
153 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
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
, gfp_t gfp
,
235 struct blk_mq_ctx
*ctx
;
236 struct blk_mq_hw_ctx
*hctx
;
238 struct blk_mq_alloc_data alloc_data
;
241 ret
= blk_mq_queue_enter(q
);
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
, gfp
& ~__GFP_WAIT
,
248 reserved
, ctx
, hctx
);
250 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
251 if (!rq
&& (gfp
& __GFP_WAIT
)) {
252 __blk_mq_run_hw_queue(hctx
);
255 ctx
= blk_mq_get_ctx(q
);
256 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
257 blk_mq_set_alloc_data(&alloc_data
, q
, gfp
, reserved
, ctx
,
259 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
260 ctx
= alloc_data
.ctx
;
264 blk_mq_queue_exit(q
);
265 return ERR_PTR(-EWOULDBLOCK
);
269 EXPORT_SYMBOL(blk_mq_alloc_request
);
271 static void __blk_mq_free_request(struct blk_mq_hw_ctx
*hctx
,
272 struct blk_mq_ctx
*ctx
, struct request
*rq
)
274 const int tag
= rq
->tag
;
275 struct request_queue
*q
= rq
->q
;
277 if (rq
->cmd_flags
& REQ_MQ_INFLIGHT
)
278 atomic_dec(&hctx
->nr_active
);
281 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
282 blk_mq_put_tag(hctx
, tag
, &ctx
->last_tag
);
283 blk_mq_queue_exit(q
);
286 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
288 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
290 ctx
->rq_completed
[rq_is_sync(rq
)]++;
291 __blk_mq_free_request(hctx
, ctx
, rq
);
294 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request
);
296 void blk_mq_free_request(struct request
*rq
)
298 struct blk_mq_hw_ctx
*hctx
;
299 struct request_queue
*q
= rq
->q
;
301 hctx
= q
->mq_ops
->map_queue(q
, rq
->mq_ctx
->cpu
);
302 blk_mq_free_hctx_request(hctx
, rq
);
304 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
306 inline void __blk_mq_end_request(struct request
*rq
, int error
)
308 blk_account_io_done(rq
);
311 rq
->end_io(rq
, error
);
313 if (unlikely(blk_bidi_rq(rq
)))
314 blk_mq_free_request(rq
->next_rq
);
315 blk_mq_free_request(rq
);
318 EXPORT_SYMBOL(__blk_mq_end_request
);
320 void blk_mq_end_request(struct request
*rq
, int error
)
322 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
324 __blk_mq_end_request(rq
, error
);
326 EXPORT_SYMBOL(blk_mq_end_request
);
328 static void __blk_mq_complete_request_remote(void *data
)
330 struct request
*rq
= data
;
332 rq
->q
->softirq_done_fn(rq
);
335 static void blk_mq_ipi_complete_request(struct request
*rq
)
337 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
341 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
342 rq
->q
->softirq_done_fn(rq
);
347 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
348 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
350 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
351 rq
->csd
.func
= __blk_mq_complete_request_remote
;
354 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
356 rq
->q
->softirq_done_fn(rq
);
361 void __blk_mq_complete_request(struct request
*rq
)
363 struct request_queue
*q
= rq
->q
;
365 if (!q
->softirq_done_fn
)
366 blk_mq_end_request(rq
, rq
->errors
);
368 blk_mq_ipi_complete_request(rq
);
372 * blk_mq_complete_request - end I/O on a request
373 * @rq: the request being processed
376 * Ends all I/O on a request. It does not handle partial completions.
377 * The actual completion happens out-of-order, through a IPI handler.
379 void blk_mq_complete_request(struct request
*rq
)
381 struct request_queue
*q
= rq
->q
;
383 if (unlikely(blk_should_fake_timeout(q
)))
385 if (!blk_mark_rq_complete(rq
))
386 __blk_mq_complete_request(rq
);
388 EXPORT_SYMBOL(blk_mq_complete_request
);
390 void blk_mq_start_request(struct request
*rq
)
392 struct request_queue
*q
= rq
->q
;
394 trace_block_rq_issue(q
, rq
);
396 rq
->resid_len
= blk_rq_bytes(rq
);
397 if (unlikely(blk_bidi_rq(rq
)))
398 rq
->next_rq
->resid_len
= blk_rq_bytes(rq
->next_rq
);
403 * Ensure that ->deadline is visible before set the started
404 * flag and clear the completed flag.
406 smp_mb__before_atomic();
409 * Mark us as started and clear complete. Complete might have been
410 * set if requeue raced with timeout, which then marked it as
411 * complete. So be sure to clear complete again when we start
412 * the request, otherwise we'll ignore the completion event.
414 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
415 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
416 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
417 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
419 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
421 * Make sure space for the drain appears. We know we can do
422 * this because max_hw_segments has been adjusted to be one
423 * fewer than the device can handle.
425 rq
->nr_phys_segments
++;
428 EXPORT_SYMBOL(blk_mq_start_request
);
430 static void __blk_mq_requeue_request(struct request
*rq
)
432 struct request_queue
*q
= rq
->q
;
434 trace_block_rq_requeue(q
, rq
);
436 if (test_and_clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
437 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
438 rq
->nr_phys_segments
--;
442 void blk_mq_requeue_request(struct request
*rq
)
444 __blk_mq_requeue_request(rq
);
446 BUG_ON(blk_queued_rq(rq
));
447 blk_mq_add_to_requeue_list(rq
, true);
449 EXPORT_SYMBOL(blk_mq_requeue_request
);
451 static void blk_mq_requeue_work(struct work_struct
*work
)
453 struct request_queue
*q
=
454 container_of(work
, struct request_queue
, requeue_work
);
456 struct request
*rq
, *next
;
459 spin_lock_irqsave(&q
->requeue_lock
, flags
);
460 list_splice_init(&q
->requeue_list
, &rq_list
);
461 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
463 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
464 if (!(rq
->cmd_flags
& REQ_SOFTBARRIER
))
467 rq
->cmd_flags
&= ~REQ_SOFTBARRIER
;
468 list_del_init(&rq
->queuelist
);
469 blk_mq_insert_request(rq
, true, false, false);
472 while (!list_empty(&rq_list
)) {
473 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
474 list_del_init(&rq
->queuelist
);
475 blk_mq_insert_request(rq
, false, false, false);
479 * Use the start variant of queue running here, so that running
480 * the requeue work will kick stopped queues.
482 blk_mq_start_hw_queues(q
);
485 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
)
487 struct request_queue
*q
= rq
->q
;
491 * We abuse this flag that is otherwise used by the I/O scheduler to
492 * request head insertation from the workqueue.
494 BUG_ON(rq
->cmd_flags
& REQ_SOFTBARRIER
);
496 spin_lock_irqsave(&q
->requeue_lock
, flags
);
498 rq
->cmd_flags
|= REQ_SOFTBARRIER
;
499 list_add(&rq
->queuelist
, &q
->requeue_list
);
501 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
503 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
505 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
507 void blk_mq_kick_requeue_list(struct request_queue
*q
)
509 kblockd_schedule_work(&q
->requeue_work
);
511 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
513 static inline bool is_flush_request(struct request
*rq
,
514 struct blk_flush_queue
*fq
, unsigned int tag
)
516 return ((rq
->cmd_flags
& REQ_FLUSH_SEQ
) &&
517 fq
->flush_rq
->tag
== tag
);
520 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
522 struct request
*rq
= tags
->rqs
[tag
];
523 /* mq_ctx of flush rq is always cloned from the corresponding req */
524 struct blk_flush_queue
*fq
= blk_get_flush_queue(rq
->q
, rq
->mq_ctx
);
526 if (!is_flush_request(rq
, fq
, tag
))
531 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
533 struct blk_mq_timeout_data
{
535 unsigned int next_set
;
538 void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
540 struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
541 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
544 * We know that complete is set at this point. If STARTED isn't set
545 * anymore, then the request isn't active and the "timeout" should
546 * just be ignored. This can happen due to the bitflag ordering.
547 * Timeout first checks if STARTED is set, and if it is, assumes
548 * the request is active. But if we race with completion, then
549 * we both flags will get cleared. So check here again, and ignore
550 * a timeout event with a request that isn't active.
552 if (!test_bit(REQ_ATOM_STARTED
, &req
->atomic_flags
))
556 ret
= ops
->timeout(req
, reserved
);
560 __blk_mq_complete_request(req
);
562 case BLK_EH_RESET_TIMER
:
564 blk_clear_rq_complete(req
);
566 case BLK_EH_NOT_HANDLED
:
569 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
574 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
575 struct request
*rq
, void *priv
, bool reserved
)
577 struct blk_mq_timeout_data
*data
= priv
;
579 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
582 if (time_after_eq(jiffies
, rq
->deadline
)) {
583 if (!blk_mark_rq_complete(rq
))
584 blk_mq_rq_timed_out(rq
, reserved
);
585 } else if (!data
->next_set
|| time_after(data
->next
, rq
->deadline
)) {
586 data
->next
= rq
->deadline
;
591 static void blk_mq_rq_timer(unsigned long priv
)
593 struct request_queue
*q
= (struct request_queue
*)priv
;
594 struct blk_mq_timeout_data data
= {
598 struct blk_mq_hw_ctx
*hctx
;
601 queue_for_each_hw_ctx(q
, hctx
, i
) {
603 * If not software queues are currently mapped to this
604 * hardware queue, there's nothing to check
606 if (!blk_mq_hw_queue_mapped(hctx
))
609 blk_mq_tag_busy_iter(hctx
, blk_mq_check_expired
, &data
);
613 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
614 mod_timer(&q
->timeout
, data
.next
);
616 queue_for_each_hw_ctx(q
, hctx
, i
)
617 blk_mq_tag_idle(hctx
);
622 * Reverse check our software queue for entries that we could potentially
623 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
624 * too much time checking for merges.
626 static bool blk_mq_attempt_merge(struct request_queue
*q
,
627 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
632 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
638 if (!blk_rq_merge_ok(rq
, bio
))
641 el_ret
= blk_try_merge(rq
, bio
);
642 if (el_ret
== ELEVATOR_BACK_MERGE
) {
643 if (bio_attempt_back_merge(q
, rq
, bio
)) {
648 } else if (el_ret
== ELEVATOR_FRONT_MERGE
) {
649 if (bio_attempt_front_merge(q
, rq
, bio
)) {
661 * Process software queues that have been marked busy, splicing them
662 * to the for-dispatch
664 static void flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
666 struct blk_mq_ctx
*ctx
;
669 for (i
= 0; i
< hctx
->ctx_map
.map_size
; i
++) {
670 struct blk_align_bitmap
*bm
= &hctx
->ctx_map
.map
[i
];
671 unsigned int off
, bit
;
677 off
= i
* hctx
->ctx_map
.bits_per_word
;
679 bit
= find_next_bit(&bm
->word
, bm
->depth
, bit
);
680 if (bit
>= bm
->depth
)
683 ctx
= hctx
->ctxs
[bit
+ off
];
684 clear_bit(bit
, &bm
->word
);
685 spin_lock(&ctx
->lock
);
686 list_splice_tail_init(&ctx
->rq_list
, list
);
687 spin_unlock(&ctx
->lock
);
695 * Run this hardware queue, pulling any software queues mapped to it in.
696 * Note that this function currently has various problems around ordering
697 * of IO. In particular, we'd like FIFO behaviour on handling existing
698 * items on the hctx->dispatch list. Ignore that for now.
700 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
702 struct request_queue
*q
= hctx
->queue
;
705 LIST_HEAD(driver_list
);
706 struct list_head
*dptr
;
709 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
));
711 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)))
717 * Touch any software queue that has pending entries.
719 flush_busy_ctxs(hctx
, &rq_list
);
722 * If we have previous entries on our dispatch list, grab them
723 * and stuff them at the front for more fair dispatch.
725 if (!list_empty_careful(&hctx
->dispatch
)) {
726 spin_lock(&hctx
->lock
);
727 if (!list_empty(&hctx
->dispatch
))
728 list_splice_init(&hctx
->dispatch
, &rq_list
);
729 spin_unlock(&hctx
->lock
);
733 * Start off with dptr being NULL, so we start the first request
734 * immediately, even if we have more pending.
739 * Now process all the entries, sending them to the driver.
742 while (!list_empty(&rq_list
)) {
743 struct blk_mq_queue_data bd
;
746 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
747 list_del_init(&rq
->queuelist
);
751 bd
.last
= list_empty(&rq_list
);
753 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
755 case BLK_MQ_RQ_QUEUE_OK
:
758 case BLK_MQ_RQ_QUEUE_BUSY
:
759 list_add(&rq
->queuelist
, &rq_list
);
760 __blk_mq_requeue_request(rq
);
763 pr_err("blk-mq: bad return on queue: %d\n", ret
);
764 case BLK_MQ_RQ_QUEUE_ERROR
:
766 blk_mq_end_request(rq
, rq
->errors
);
770 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
774 * We've done the first request. If we have more than 1
775 * left in the list, set dptr to defer issue.
777 if (!dptr
&& rq_list
.next
!= rq_list
.prev
)
782 hctx
->dispatched
[0]++;
783 else if (queued
< (1 << (BLK_MQ_MAX_DISPATCH_ORDER
- 1)))
784 hctx
->dispatched
[ilog2(queued
) + 1]++;
787 * Any items that need requeuing? Stuff them into hctx->dispatch,
788 * that is where we will continue on next queue run.
790 if (!list_empty(&rq_list
)) {
791 spin_lock(&hctx
->lock
);
792 list_splice(&rq_list
, &hctx
->dispatch
);
793 spin_unlock(&hctx
->lock
);
798 * It'd be great if the workqueue API had a way to pass
799 * in a mask and had some smarts for more clever placement.
800 * For now we just round-robin here, switching for every
801 * BLK_MQ_CPU_WORK_BATCH queued items.
803 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
805 if (hctx
->queue
->nr_hw_queues
== 1)
806 return WORK_CPU_UNBOUND
;
808 if (--hctx
->next_cpu_batch
<= 0) {
809 int cpu
= hctx
->next_cpu
, next_cpu
;
811 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
812 if (next_cpu
>= nr_cpu_ids
)
813 next_cpu
= cpumask_first(hctx
->cpumask
);
815 hctx
->next_cpu
= next_cpu
;
816 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
821 return hctx
->next_cpu
;
824 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
826 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
) ||
827 !blk_mq_hw_queue_mapped(hctx
)))
832 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
833 __blk_mq_run_hw_queue(hctx
);
841 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
845 void blk_mq_run_queues(struct request_queue
*q
, bool async
)
847 struct blk_mq_hw_ctx
*hctx
;
850 queue_for_each_hw_ctx(q
, hctx
, i
) {
851 if ((!blk_mq_hctx_has_pending(hctx
) &&
852 list_empty_careful(&hctx
->dispatch
)) ||
853 test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
856 blk_mq_run_hw_queue(hctx
, async
);
859 EXPORT_SYMBOL(blk_mq_run_queues
);
861 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
863 cancel_delayed_work(&hctx
->run_work
);
864 cancel_delayed_work(&hctx
->delay_work
);
865 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
867 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
869 void blk_mq_stop_hw_queues(struct request_queue
*q
)
871 struct blk_mq_hw_ctx
*hctx
;
874 queue_for_each_hw_ctx(q
, hctx
, i
)
875 blk_mq_stop_hw_queue(hctx
);
877 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
879 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
881 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
883 blk_mq_run_hw_queue(hctx
, false);
885 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
887 void blk_mq_start_hw_queues(struct request_queue
*q
)
889 struct blk_mq_hw_ctx
*hctx
;
892 queue_for_each_hw_ctx(q
, hctx
, i
)
893 blk_mq_start_hw_queue(hctx
);
895 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
898 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
900 struct blk_mq_hw_ctx
*hctx
;
903 queue_for_each_hw_ctx(q
, hctx
, i
) {
904 if (!test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
907 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
908 blk_mq_run_hw_queue(hctx
, async
);
911 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
913 static void blk_mq_run_work_fn(struct work_struct
*work
)
915 struct blk_mq_hw_ctx
*hctx
;
917 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
919 __blk_mq_run_hw_queue(hctx
);
922 static void blk_mq_delay_work_fn(struct work_struct
*work
)
924 struct blk_mq_hw_ctx
*hctx
;
926 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
928 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
929 __blk_mq_run_hw_queue(hctx
);
932 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
934 if (unlikely(!blk_mq_hw_queue_mapped(hctx
)))
937 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
938 &hctx
->delay_work
, msecs_to_jiffies(msecs
));
940 EXPORT_SYMBOL(blk_mq_delay_queue
);
942 static void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
,
943 struct request
*rq
, bool at_head
)
945 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
947 trace_block_rq_insert(hctx
->queue
, rq
);
950 list_add(&rq
->queuelist
, &ctx
->rq_list
);
952 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
954 blk_mq_hctx_mark_pending(hctx
, ctx
);
957 void blk_mq_insert_request(struct request
*rq
, bool at_head
, bool run_queue
,
960 struct request_queue
*q
= rq
->q
;
961 struct blk_mq_hw_ctx
*hctx
;
962 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
, *current_ctx
;
964 current_ctx
= blk_mq_get_ctx(q
);
965 if (!cpu_online(ctx
->cpu
))
966 rq
->mq_ctx
= ctx
= current_ctx
;
968 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
970 spin_lock(&ctx
->lock
);
971 __blk_mq_insert_request(hctx
, rq
, at_head
);
972 spin_unlock(&ctx
->lock
);
975 blk_mq_run_hw_queue(hctx
, async
);
977 blk_mq_put_ctx(current_ctx
);
980 static void blk_mq_insert_requests(struct request_queue
*q
,
981 struct blk_mq_ctx
*ctx
,
982 struct list_head
*list
,
987 struct blk_mq_hw_ctx
*hctx
;
988 struct blk_mq_ctx
*current_ctx
;
990 trace_block_unplug(q
, depth
, !from_schedule
);
992 current_ctx
= blk_mq_get_ctx(q
);
994 if (!cpu_online(ctx
->cpu
))
996 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
999 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1002 spin_lock(&ctx
->lock
);
1003 while (!list_empty(list
)) {
1006 rq
= list_first_entry(list
, struct request
, queuelist
);
1007 list_del_init(&rq
->queuelist
);
1009 __blk_mq_insert_request(hctx
, rq
, false);
1011 spin_unlock(&ctx
->lock
);
1013 blk_mq_run_hw_queue(hctx
, from_schedule
);
1014 blk_mq_put_ctx(current_ctx
);
1017 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1019 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1020 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1022 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1023 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1024 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1027 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1029 struct blk_mq_ctx
*this_ctx
;
1030 struct request_queue
*this_q
;
1033 LIST_HEAD(ctx_list
);
1036 list_splice_init(&plug
->mq_list
, &list
);
1038 list_sort(NULL
, &list
, plug_ctx_cmp
);
1044 while (!list_empty(&list
)) {
1045 rq
= list_entry_rq(list
.next
);
1046 list_del_init(&rq
->queuelist
);
1048 if (rq
->mq_ctx
!= this_ctx
) {
1050 blk_mq_insert_requests(this_q
, this_ctx
,
1055 this_ctx
= rq
->mq_ctx
;
1061 list_add_tail(&rq
->queuelist
, &ctx_list
);
1065 * If 'this_ctx' is set, we know we have entries to complete
1066 * on 'ctx_list'. Do those.
1069 blk_mq_insert_requests(this_q
, this_ctx
, &ctx_list
, depth
,
1074 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1076 init_request_from_bio(rq
, bio
);
1078 if (blk_do_io_stat(rq
))
1079 blk_account_io_start(rq
, 1);
1082 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1084 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1085 !blk_queue_nomerges(hctx
->queue
);
1088 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx
*hctx
,
1089 struct blk_mq_ctx
*ctx
,
1090 struct request
*rq
, struct bio
*bio
)
1092 if (!hctx_allow_merges(hctx
)) {
1093 blk_mq_bio_to_request(rq
, bio
);
1094 spin_lock(&ctx
->lock
);
1096 __blk_mq_insert_request(hctx
, rq
, false);
1097 spin_unlock(&ctx
->lock
);
1100 struct request_queue
*q
= hctx
->queue
;
1102 spin_lock(&ctx
->lock
);
1103 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1104 blk_mq_bio_to_request(rq
, bio
);
1108 spin_unlock(&ctx
->lock
);
1109 __blk_mq_free_request(hctx
, ctx
, rq
);
1114 struct blk_map_ctx
{
1115 struct blk_mq_hw_ctx
*hctx
;
1116 struct blk_mq_ctx
*ctx
;
1119 static struct request
*blk_mq_map_request(struct request_queue
*q
,
1121 struct blk_map_ctx
*data
)
1123 struct blk_mq_hw_ctx
*hctx
;
1124 struct blk_mq_ctx
*ctx
;
1126 int rw
= bio_data_dir(bio
);
1127 struct blk_mq_alloc_data alloc_data
;
1129 if (unlikely(blk_mq_queue_enter(q
))) {
1130 bio_endio(bio
, -EIO
);
1134 ctx
= blk_mq_get_ctx(q
);
1135 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1137 if (rw_is_sync(bio
->bi_rw
))
1140 trace_block_getrq(q
, bio
, rw
);
1141 blk_mq_set_alloc_data(&alloc_data
, q
, GFP_ATOMIC
, false, ctx
,
1143 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
1144 if (unlikely(!rq
)) {
1145 __blk_mq_run_hw_queue(hctx
);
1146 blk_mq_put_ctx(ctx
);
1147 trace_block_sleeprq(q
, bio
, rw
);
1149 ctx
= blk_mq_get_ctx(q
);
1150 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1151 blk_mq_set_alloc_data(&alloc_data
, q
,
1152 __GFP_WAIT
|GFP_ATOMIC
, false, ctx
, hctx
);
1153 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
1154 ctx
= alloc_data
.ctx
;
1155 hctx
= alloc_data
.hctx
;
1165 * Multiple hardware queue variant. This will not use per-process plugs,
1166 * but will attempt to bypass the hctx queueing if we can go straight to
1167 * hardware for SYNC IO.
1169 static void blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1171 const int is_sync
= rw_is_sync(bio
->bi_rw
);
1172 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
1173 struct blk_map_ctx data
;
1176 blk_queue_bounce(q
, &bio
);
1178 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1179 bio_endio(bio
, -EIO
);
1183 rq
= blk_mq_map_request(q
, bio
, &data
);
1187 if (unlikely(is_flush_fua
)) {
1188 blk_mq_bio_to_request(rq
, bio
);
1189 blk_insert_flush(rq
);
1194 * If the driver supports defer issued based on 'last', then
1195 * queue it up like normal since we can potentially save some
1198 if (is_sync
&& !(data
.hctx
->flags
& BLK_MQ_F_DEFER_ISSUE
)) {
1199 struct blk_mq_queue_data bd
= {
1206 blk_mq_bio_to_request(rq
, bio
);
1209 * For OK queue, we are done. For error, kill it. Any other
1210 * error (busy), just add it to our list as we previously
1213 ret
= q
->mq_ops
->queue_rq(data
.hctx
, &bd
);
1214 if (ret
== BLK_MQ_RQ_QUEUE_OK
)
1217 __blk_mq_requeue_request(rq
);
1219 if (ret
== BLK_MQ_RQ_QUEUE_ERROR
) {
1221 blk_mq_end_request(rq
, rq
->errors
);
1227 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1229 * For a SYNC request, send it to the hardware immediately. For
1230 * an ASYNC request, just ensure that we run it later on. The
1231 * latter allows for merging opportunities and more efficient
1235 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1238 blk_mq_put_ctx(data
.ctx
);
1242 * Single hardware queue variant. This will attempt to use any per-process
1243 * plug for merging and IO deferral.
1245 static void blk_sq_make_request(struct request_queue
*q
, struct bio
*bio
)
1247 const int is_sync
= rw_is_sync(bio
->bi_rw
);
1248 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
1249 unsigned int use_plug
, request_count
= 0;
1250 struct blk_map_ctx data
;
1254 * If we have multiple hardware queues, just go directly to
1255 * one of those for sync IO.
1257 use_plug
= !is_flush_fua
&& !is_sync
;
1259 blk_queue_bounce(q
, &bio
);
1261 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1262 bio_endio(bio
, -EIO
);
1266 if (use_plug
&& !blk_queue_nomerges(q
) &&
1267 blk_attempt_plug_merge(q
, bio
, &request_count
))
1270 rq
= blk_mq_map_request(q
, bio
, &data
);
1274 if (unlikely(is_flush_fua
)) {
1275 blk_mq_bio_to_request(rq
, bio
);
1276 blk_insert_flush(rq
);
1281 * A task plug currently exists. Since this is completely lockless,
1282 * utilize that to temporarily store requests until the task is
1283 * either done or scheduled away.
1286 struct blk_plug
*plug
= current
->plug
;
1289 blk_mq_bio_to_request(rq
, bio
);
1290 if (list_empty(&plug
->mq_list
))
1291 trace_block_plug(q
);
1292 else if (request_count
>= BLK_MAX_REQUEST_COUNT
) {
1293 blk_flush_plug_list(plug
, false);
1294 trace_block_plug(q
);
1296 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1297 blk_mq_put_ctx(data
.ctx
);
1302 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1304 * For a SYNC request, send it to the hardware immediately. For
1305 * an ASYNC request, just ensure that we run it later on. The
1306 * latter allows for merging opportunities and more efficient
1310 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1313 blk_mq_put_ctx(data
.ctx
);
1317 * Default mapping to a software queue, since we use one per CPU.
1319 struct blk_mq_hw_ctx
*blk_mq_map_queue(struct request_queue
*q
, const int cpu
)
1321 return q
->queue_hw_ctx
[q
->mq_map
[cpu
]];
1323 EXPORT_SYMBOL(blk_mq_map_queue
);
1325 static void blk_mq_free_rq_map(struct blk_mq_tag_set
*set
,
1326 struct blk_mq_tags
*tags
, unsigned int hctx_idx
)
1330 if (tags
->rqs
&& set
->ops
->exit_request
) {
1333 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1336 set
->ops
->exit_request(set
->driver_data
, tags
->rqs
[i
],
1338 tags
->rqs
[i
] = NULL
;
1342 while (!list_empty(&tags
->page_list
)) {
1343 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1344 list_del_init(&page
->lru
);
1345 __free_pages(page
, page
->private);
1350 blk_mq_free_tags(tags
);
1353 static size_t order_to_size(unsigned int order
)
1355 return (size_t)PAGE_SIZE
<< order
;
1358 static struct blk_mq_tags
*blk_mq_init_rq_map(struct blk_mq_tag_set
*set
,
1359 unsigned int hctx_idx
)
1361 struct blk_mq_tags
*tags
;
1362 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1363 size_t rq_size
, left
;
1365 tags
= blk_mq_init_tags(set
->queue_depth
, set
->reserved_tags
,
1370 INIT_LIST_HEAD(&tags
->page_list
);
1372 tags
->rqs
= kzalloc_node(set
->queue_depth
* sizeof(struct request
*),
1373 GFP_KERNEL
| __GFP_NOWARN
| __GFP_NORETRY
,
1376 blk_mq_free_tags(tags
);
1381 * rq_size is the size of the request plus driver payload, rounded
1382 * to the cacheline size
1384 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1386 left
= rq_size
* set
->queue_depth
;
1388 for (i
= 0; i
< set
->queue_depth
; ) {
1389 int this_order
= max_order
;
1394 while (left
< order_to_size(this_order
- 1) && this_order
)
1398 page
= alloc_pages_node(set
->numa_node
,
1399 GFP_KERNEL
| __GFP_NOWARN
| __GFP_NORETRY
,
1405 if (order_to_size(this_order
) < rq_size
)
1412 page
->private = this_order
;
1413 list_add_tail(&page
->lru
, &tags
->page_list
);
1415 p
= page_address(page
);
1416 entries_per_page
= order_to_size(this_order
) / rq_size
;
1417 to_do
= min(entries_per_page
, set
->queue_depth
- i
);
1418 left
-= to_do
* rq_size
;
1419 for (j
= 0; j
< to_do
; j
++) {
1421 tags
->rqs
[i
]->atomic_flags
= 0;
1422 tags
->rqs
[i
]->cmd_flags
= 0;
1423 if (set
->ops
->init_request
) {
1424 if (set
->ops
->init_request(set
->driver_data
,
1425 tags
->rqs
[i
], hctx_idx
, i
,
1427 tags
->rqs
[i
] = NULL
;
1440 blk_mq_free_rq_map(set
, tags
, hctx_idx
);
1444 static void blk_mq_free_bitmap(struct blk_mq_ctxmap
*bitmap
)
1449 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap
*bitmap
, int node
)
1451 unsigned int bpw
= 8, total
, num_maps
, i
;
1453 bitmap
->bits_per_word
= bpw
;
1455 num_maps
= ALIGN(nr_cpu_ids
, bpw
) / bpw
;
1456 bitmap
->map
= kzalloc_node(num_maps
* sizeof(struct blk_align_bitmap
),
1461 bitmap
->map_size
= num_maps
;
1464 for (i
= 0; i
< num_maps
; i
++) {
1465 bitmap
->map
[i
].depth
= min(total
, bitmap
->bits_per_word
);
1466 total
-= bitmap
->map
[i
].depth
;
1472 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx
*hctx
, int cpu
)
1474 struct request_queue
*q
= hctx
->queue
;
1475 struct blk_mq_ctx
*ctx
;
1479 * Move ctx entries to new CPU, if this one is going away.
1481 ctx
= __blk_mq_get_ctx(q
, cpu
);
1483 spin_lock(&ctx
->lock
);
1484 if (!list_empty(&ctx
->rq_list
)) {
1485 list_splice_init(&ctx
->rq_list
, &tmp
);
1486 blk_mq_hctx_clear_pending(hctx
, ctx
);
1488 spin_unlock(&ctx
->lock
);
1490 if (list_empty(&tmp
))
1493 ctx
= blk_mq_get_ctx(q
);
1494 spin_lock(&ctx
->lock
);
1496 while (!list_empty(&tmp
)) {
1499 rq
= list_first_entry(&tmp
, struct request
, queuelist
);
1501 list_move_tail(&rq
->queuelist
, &ctx
->rq_list
);
1504 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1505 blk_mq_hctx_mark_pending(hctx
, ctx
);
1507 spin_unlock(&ctx
->lock
);
1509 blk_mq_run_hw_queue(hctx
, true);
1510 blk_mq_put_ctx(ctx
);
1514 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx
*hctx
, int cpu
)
1516 struct request_queue
*q
= hctx
->queue
;
1517 struct blk_mq_tag_set
*set
= q
->tag_set
;
1519 if (set
->tags
[hctx
->queue_num
])
1522 set
->tags
[hctx
->queue_num
] = blk_mq_init_rq_map(set
, hctx
->queue_num
);
1523 if (!set
->tags
[hctx
->queue_num
])
1526 hctx
->tags
= set
->tags
[hctx
->queue_num
];
1530 static int blk_mq_hctx_notify(void *data
, unsigned long action
,
1533 struct blk_mq_hw_ctx
*hctx
= data
;
1535 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
)
1536 return blk_mq_hctx_cpu_offline(hctx
, cpu
);
1537 else if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
)
1538 return blk_mq_hctx_cpu_online(hctx
, cpu
);
1543 static void blk_mq_exit_hctx(struct request_queue
*q
,
1544 struct blk_mq_tag_set
*set
,
1545 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
1547 unsigned flush_start_tag
= set
->queue_depth
;
1549 blk_mq_tag_idle(hctx
);
1551 if (set
->ops
->exit_request
)
1552 set
->ops
->exit_request(set
->driver_data
,
1553 hctx
->fq
->flush_rq
, hctx_idx
,
1554 flush_start_tag
+ hctx_idx
);
1556 if (set
->ops
->exit_hctx
)
1557 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1559 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1560 blk_free_flush_queue(hctx
->fq
);
1562 blk_mq_free_bitmap(&hctx
->ctx_map
);
1565 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1566 struct blk_mq_tag_set
*set
, int nr_queue
)
1568 struct blk_mq_hw_ctx
*hctx
;
1571 queue_for_each_hw_ctx(q
, hctx
, i
) {
1574 blk_mq_exit_hctx(q
, set
, hctx
, i
);
1578 static void blk_mq_free_hw_queues(struct request_queue
*q
,
1579 struct blk_mq_tag_set
*set
)
1581 struct blk_mq_hw_ctx
*hctx
;
1584 queue_for_each_hw_ctx(q
, hctx
, i
) {
1585 free_cpumask_var(hctx
->cpumask
);
1590 static int blk_mq_init_hctx(struct request_queue
*q
,
1591 struct blk_mq_tag_set
*set
,
1592 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
1595 unsigned flush_start_tag
= set
->queue_depth
;
1597 node
= hctx
->numa_node
;
1598 if (node
== NUMA_NO_NODE
)
1599 node
= hctx
->numa_node
= set
->numa_node
;
1601 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1602 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1603 spin_lock_init(&hctx
->lock
);
1604 INIT_LIST_HEAD(&hctx
->dispatch
);
1606 hctx
->queue_num
= hctx_idx
;
1607 hctx
->flags
= set
->flags
;
1608 hctx
->cmd_size
= set
->cmd_size
;
1610 blk_mq_init_cpu_notifier(&hctx
->cpu_notifier
,
1611 blk_mq_hctx_notify
, hctx
);
1612 blk_mq_register_cpu_notifier(&hctx
->cpu_notifier
);
1614 hctx
->tags
= set
->tags
[hctx_idx
];
1617 * Allocate space for all possible cpus to avoid allocation at
1620 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1623 goto unregister_cpu_notifier
;
1625 if (blk_mq_alloc_bitmap(&hctx
->ctx_map
, node
))
1630 if (set
->ops
->init_hctx
&&
1631 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
1634 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
1638 if (set
->ops
->init_request
&&
1639 set
->ops
->init_request(set
->driver_data
,
1640 hctx
->fq
->flush_rq
, hctx_idx
,
1641 flush_start_tag
+ hctx_idx
, node
))
1649 if (set
->ops
->exit_hctx
)
1650 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1652 blk_mq_free_bitmap(&hctx
->ctx_map
);
1655 unregister_cpu_notifier
:
1656 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1661 static int blk_mq_init_hw_queues(struct request_queue
*q
,
1662 struct blk_mq_tag_set
*set
)
1664 struct blk_mq_hw_ctx
*hctx
;
1668 * Initialize hardware queues
1670 queue_for_each_hw_ctx(q
, hctx
, i
) {
1671 if (blk_mq_init_hctx(q
, set
, hctx
, i
))
1675 if (i
== q
->nr_hw_queues
)
1681 blk_mq_exit_hw_queues(q
, set
, i
);
1686 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1687 unsigned int nr_hw_queues
)
1691 for_each_possible_cpu(i
) {
1692 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1693 struct blk_mq_hw_ctx
*hctx
;
1695 memset(__ctx
, 0, sizeof(*__ctx
));
1697 spin_lock_init(&__ctx
->lock
);
1698 INIT_LIST_HEAD(&__ctx
->rq_list
);
1701 /* If the cpu isn't online, the cpu is mapped to first hctx */
1705 hctx
= q
->mq_ops
->map_queue(q
, i
);
1706 cpumask_set_cpu(i
, hctx
->cpumask
);
1710 * Set local node, IFF we have more than one hw queue. If
1711 * not, we remain on the home node of the device
1713 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1714 hctx
->numa_node
= cpu_to_node(i
);
1718 static void blk_mq_map_swqueue(struct request_queue
*q
)
1721 struct blk_mq_hw_ctx
*hctx
;
1722 struct blk_mq_ctx
*ctx
;
1724 queue_for_each_hw_ctx(q
, hctx
, i
) {
1725 cpumask_clear(hctx
->cpumask
);
1730 * Map software to hardware queues
1732 queue_for_each_ctx(q
, ctx
, i
) {
1733 /* If the cpu isn't online, the cpu is mapped to first hctx */
1737 hctx
= q
->mq_ops
->map_queue(q
, i
);
1738 cpumask_set_cpu(i
, hctx
->cpumask
);
1739 ctx
->index_hw
= hctx
->nr_ctx
;
1740 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
1743 queue_for_each_hw_ctx(q
, hctx
, i
) {
1745 * If no software queues are mapped to this hardware queue,
1746 * disable it and free the request entries.
1748 if (!hctx
->nr_ctx
) {
1749 struct blk_mq_tag_set
*set
= q
->tag_set
;
1752 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1753 set
->tags
[i
] = NULL
;
1760 * Initialize batch roundrobin counts
1762 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
1763 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1767 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
)
1769 struct blk_mq_hw_ctx
*hctx
;
1770 struct request_queue
*q
;
1774 if (set
->tag_list
.next
== set
->tag_list
.prev
)
1779 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
1780 blk_mq_freeze_queue(q
);
1782 queue_for_each_hw_ctx(q
, hctx
, i
) {
1784 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
1786 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
1788 blk_mq_unfreeze_queue(q
);
1792 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
1794 struct blk_mq_tag_set
*set
= q
->tag_set
;
1796 mutex_lock(&set
->tag_list_lock
);
1797 list_del_init(&q
->tag_set_list
);
1798 blk_mq_update_tag_set_depth(set
);
1799 mutex_unlock(&set
->tag_list_lock
);
1802 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
1803 struct request_queue
*q
)
1807 mutex_lock(&set
->tag_list_lock
);
1808 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
1809 blk_mq_update_tag_set_depth(set
);
1810 mutex_unlock(&set
->tag_list_lock
);
1813 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
1815 struct blk_mq_hw_ctx
**hctxs
;
1816 struct blk_mq_ctx __percpu
*ctx
;
1817 struct request_queue
*q
;
1821 ctx
= alloc_percpu(struct blk_mq_ctx
);
1823 return ERR_PTR(-ENOMEM
);
1825 hctxs
= kmalloc_node(set
->nr_hw_queues
* sizeof(*hctxs
), GFP_KERNEL
,
1831 map
= blk_mq_make_queue_map(set
);
1835 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1836 int node
= blk_mq_hw_queue_to_node(map
, i
);
1838 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
1843 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
1847 atomic_set(&hctxs
[i
]->nr_active
, 0);
1848 hctxs
[i
]->numa_node
= node
;
1849 hctxs
[i
]->queue_num
= i
;
1852 q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
1857 * Init percpu_ref in atomic mode so that it's faster to shutdown.
1858 * See blk_register_queue() for details.
1860 if (percpu_ref_init(&q
->mq_usage_counter
, blk_mq_usage_counter_release
,
1861 PERCPU_REF_INIT_ATOMIC
, GFP_KERNEL
))
1864 setup_timer(&q
->timeout
, blk_mq_rq_timer
, (unsigned long) q
);
1865 blk_queue_rq_timeout(q
, 30000);
1867 q
->nr_queues
= nr_cpu_ids
;
1868 q
->nr_hw_queues
= set
->nr_hw_queues
;
1872 q
->queue_hw_ctx
= hctxs
;
1874 q
->mq_ops
= set
->ops
;
1875 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
1877 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
1878 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
1880 q
->sg_reserved_size
= INT_MAX
;
1882 INIT_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
1883 INIT_LIST_HEAD(&q
->requeue_list
);
1884 spin_lock_init(&q
->requeue_lock
);
1886 if (q
->nr_hw_queues
> 1)
1887 blk_queue_make_request(q
, blk_mq_make_request
);
1889 blk_queue_make_request(q
, blk_sq_make_request
);
1892 blk_queue_rq_timeout(q
, set
->timeout
);
1895 * Do this after blk_queue_make_request() overrides it...
1897 q
->nr_requests
= set
->queue_depth
;
1899 if (set
->ops
->complete
)
1900 blk_queue_softirq_done(q
, set
->ops
->complete
);
1902 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
1904 if (blk_mq_init_hw_queues(q
, set
))
1907 mutex_lock(&all_q_mutex
);
1908 list_add_tail(&q
->all_q_node
, &all_q_list
);
1909 mutex_unlock(&all_q_mutex
);
1911 blk_mq_add_queue_tag_set(set
, q
);
1913 blk_mq_map_swqueue(q
);
1918 blk_cleanup_queue(q
);
1921 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1924 free_cpumask_var(hctxs
[i
]->cpumask
);
1931 return ERR_PTR(-ENOMEM
);
1933 EXPORT_SYMBOL(blk_mq_init_queue
);
1935 void blk_mq_free_queue(struct request_queue
*q
)
1937 struct blk_mq_tag_set
*set
= q
->tag_set
;
1939 blk_mq_del_queue_tag_set(q
);
1941 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
1942 blk_mq_free_hw_queues(q
, set
);
1944 percpu_ref_exit(&q
->mq_usage_counter
);
1946 free_percpu(q
->queue_ctx
);
1947 kfree(q
->queue_hw_ctx
);
1950 q
->queue_ctx
= NULL
;
1951 q
->queue_hw_ctx
= NULL
;
1954 mutex_lock(&all_q_mutex
);
1955 list_del_init(&q
->all_q_node
);
1956 mutex_unlock(&all_q_mutex
);
1959 /* Basically redo blk_mq_init_queue with queue frozen */
1960 static void blk_mq_queue_reinit(struct request_queue
*q
)
1962 WARN_ON_ONCE(!q
->mq_freeze_depth
);
1964 blk_mq_sysfs_unregister(q
);
1966 blk_mq_update_queue_map(q
->mq_map
, q
->nr_hw_queues
);
1969 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1970 * we should change hctx numa_node according to new topology (this
1971 * involves free and re-allocate memory, worthy doing?)
1974 blk_mq_map_swqueue(q
);
1976 blk_mq_sysfs_register(q
);
1979 static int blk_mq_queue_reinit_notify(struct notifier_block
*nb
,
1980 unsigned long action
, void *hcpu
)
1982 struct request_queue
*q
;
1985 * Before new mappings are established, hotadded cpu might already
1986 * start handling requests. This doesn't break anything as we map
1987 * offline CPUs to first hardware queue. We will re-init the queue
1988 * below to get optimal settings.
1990 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
&&
1991 action
!= CPU_ONLINE
&& action
!= CPU_ONLINE_FROZEN
)
1994 mutex_lock(&all_q_mutex
);
1997 * We need to freeze and reinit all existing queues. Freezing
1998 * involves synchronous wait for an RCU grace period and doing it
1999 * one by one may take a long time. Start freezing all queues in
2000 * one swoop and then wait for the completions so that freezing can
2001 * take place in parallel.
2003 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2004 blk_mq_freeze_queue_start(q
);
2005 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2006 blk_mq_freeze_queue_wait(q
);
2008 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2009 blk_mq_queue_reinit(q
);
2011 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2012 blk_mq_unfreeze_queue(q
);
2014 mutex_unlock(&all_q_mutex
);
2018 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2022 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2023 set
->tags
[i
] = blk_mq_init_rq_map(set
, i
);
2032 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2038 * Allocate the request maps associated with this tag_set. Note that this
2039 * may reduce the depth asked for, if memory is tight. set->queue_depth
2040 * will be updated to reflect the allocated depth.
2042 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2047 depth
= set
->queue_depth
;
2049 err
= __blk_mq_alloc_rq_maps(set
);
2053 set
->queue_depth
>>= 1;
2054 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2058 } while (set
->queue_depth
);
2060 if (!set
->queue_depth
|| err
) {
2061 pr_err("blk-mq: failed to allocate request map\n");
2065 if (depth
!= set
->queue_depth
)
2066 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2067 depth
, set
->queue_depth
);
2073 * Alloc a tag set to be associated with one or more request queues.
2074 * May fail with EINVAL for various error conditions. May adjust the
2075 * requested depth down, if if it too large. In that case, the set
2076 * value will be stored in set->queue_depth.
2078 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2080 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2082 if (!set
->nr_hw_queues
)
2084 if (!set
->queue_depth
)
2086 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2089 if (!set
->nr_hw_queues
|| !set
->ops
->queue_rq
|| !set
->ops
->map_queue
)
2092 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2093 pr_info("blk-mq: reduced tag depth to %u\n",
2095 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2099 * If a crashdump is active, then we are potentially in a very
2100 * memory constrained environment. Limit us to 1 queue and
2101 * 64 tags to prevent using too much memory.
2103 if (is_kdump_kernel()) {
2104 set
->nr_hw_queues
= 1;
2105 set
->queue_depth
= min(64U, set
->queue_depth
);
2108 set
->tags
= kmalloc_node(set
->nr_hw_queues
*
2109 sizeof(struct blk_mq_tags
*),
2110 GFP_KERNEL
, set
->numa_node
);
2114 if (blk_mq_alloc_rq_maps(set
))
2117 mutex_init(&set
->tag_list_lock
);
2118 INIT_LIST_HEAD(&set
->tag_list
);
2126 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2128 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2132 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2134 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2140 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2142 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2144 struct blk_mq_tag_set
*set
= q
->tag_set
;
2145 struct blk_mq_hw_ctx
*hctx
;
2148 if (!set
|| nr
> set
->queue_depth
)
2152 queue_for_each_hw_ctx(q
, hctx
, i
) {
2153 ret
= blk_mq_tag_update_depth(hctx
->tags
, nr
);
2159 q
->nr_requests
= nr
;
2164 void blk_mq_disable_hotplug(void)
2166 mutex_lock(&all_q_mutex
);
2169 void blk_mq_enable_hotplug(void)
2171 mutex_unlock(&all_q_mutex
);
2174 static int __init
blk_mq_init(void)
2178 hotcpu_notifier(blk_mq_queue_reinit_notify
, 0);
2182 subsys_initcall(blk_mq_init
);