1 #include <linux/kernel.h>
2 #include <linux/module.h>
3 #include <linux/backing-dev.h>
5 #include <linux/blkdev.h>
7 #include <linux/init.h>
8 #include <linux/slab.h>
9 #include <linux/workqueue.h>
10 #include <linux/smp.h>
11 #include <linux/llist.h>
12 #include <linux/list_sort.h>
13 #include <linux/cpu.h>
14 #include <linux/cache.h>
15 #include <linux/sched/sysctl.h>
16 #include <linux/delay.h>
18 #include <trace/events/block.h>
20 #include <linux/blk-mq.h>
23 #include "blk-mq-tag.h"
25 static DEFINE_MUTEX(all_q_mutex
);
26 static LIST_HEAD(all_q_list
);
28 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
);
30 static struct blk_mq_ctx
*__blk_mq_get_ctx(struct request_queue
*q
,
33 return per_cpu_ptr(q
->queue_ctx
, cpu
);
37 * This assumes per-cpu software queueing queues. They could be per-node
38 * as well, for instance. For now this is hardcoded as-is. Note that we don't
39 * care about preemption, since we know the ctx's are persistent. This does
40 * mean that we can't rely on ctx always matching the currently running CPU.
42 static struct blk_mq_ctx
*blk_mq_get_ctx(struct request_queue
*q
)
44 return __blk_mq_get_ctx(q
, get_cpu());
47 static void blk_mq_put_ctx(struct blk_mq_ctx
*ctx
)
53 * Check if any of the ctx's have pending work in this hardware queue
55 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
59 for (i
= 0; i
< hctx
->nr_ctx_map
; i
++)
67 * Mark this ctx as having pending work in this hardware queue
69 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
70 struct blk_mq_ctx
*ctx
)
72 if (!test_bit(ctx
->index_hw
, hctx
->ctx_map
))
73 set_bit(ctx
->index_hw
, hctx
->ctx_map
);
76 static struct request
*__blk_mq_alloc_request(struct blk_mq_hw_ctx
*hctx
,
77 gfp_t gfp
, bool reserved
)
82 tag
= blk_mq_get_tag(hctx
->tags
, gfp
, reserved
);
83 if (tag
!= BLK_MQ_TAG_FAIL
) {
84 rq
= hctx
->tags
->rqs
[tag
];
85 blk_rq_init(hctx
->queue
, rq
);
94 static int blk_mq_queue_enter(struct request_queue
*q
)
98 __percpu_counter_add(&q
->mq_usage_counter
, 1, 1000000);
100 /* we have problems to freeze the queue if it's initializing */
101 if (!blk_queue_bypass(q
) || !blk_queue_init_done(q
))
104 __percpu_counter_add(&q
->mq_usage_counter
, -1, 1000000);
106 spin_lock_irq(q
->queue_lock
);
107 ret
= wait_event_interruptible_lock_irq(q
->mq_freeze_wq
,
108 !blk_queue_bypass(q
) || blk_queue_dying(q
),
110 /* inc usage with lock hold to avoid freeze_queue runs here */
111 if (!ret
&& !blk_queue_dying(q
))
112 __percpu_counter_add(&q
->mq_usage_counter
, 1, 1000000);
113 else if (blk_queue_dying(q
))
115 spin_unlock_irq(q
->queue_lock
);
120 static void blk_mq_queue_exit(struct request_queue
*q
)
122 __percpu_counter_add(&q
->mq_usage_counter
, -1, 1000000);
125 static void __blk_mq_drain_queue(struct request_queue
*q
)
130 spin_lock_irq(q
->queue_lock
);
131 count
= percpu_counter_sum(&q
->mq_usage_counter
);
132 spin_unlock_irq(q
->queue_lock
);
136 blk_mq_run_queues(q
, false);
142 * Guarantee no request is in use, so we can change any data structure of
143 * the queue afterward.
145 static void blk_mq_freeze_queue(struct request_queue
*q
)
149 spin_lock_irq(q
->queue_lock
);
150 drain
= !q
->bypass_depth
++;
151 queue_flag_set(QUEUE_FLAG_BYPASS
, q
);
152 spin_unlock_irq(q
->queue_lock
);
155 __blk_mq_drain_queue(q
);
158 void blk_mq_drain_queue(struct request_queue
*q
)
160 __blk_mq_drain_queue(q
);
163 static void blk_mq_unfreeze_queue(struct request_queue
*q
)
167 spin_lock_irq(q
->queue_lock
);
168 if (!--q
->bypass_depth
) {
169 queue_flag_clear(QUEUE_FLAG_BYPASS
, q
);
172 WARN_ON_ONCE(q
->bypass_depth
< 0);
173 spin_unlock_irq(q
->queue_lock
);
175 wake_up_all(&q
->mq_freeze_wq
);
178 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
180 return blk_mq_has_free_tags(hctx
->tags
);
182 EXPORT_SYMBOL(blk_mq_can_queue
);
184 static void blk_mq_rq_ctx_init(struct request_queue
*q
, struct blk_mq_ctx
*ctx
,
185 struct request
*rq
, unsigned int rw_flags
)
187 if (blk_queue_io_stat(q
))
188 rw_flags
|= REQ_IO_STAT
;
191 rq
->cmd_flags
= rw_flags
;
192 rq
->start_time
= jiffies
;
193 set_start_time_ns(rq
);
194 ctx
->rq_dispatched
[rw_is_sync(rw_flags
)]++;
197 static struct request
*blk_mq_alloc_request_pinned(struct request_queue
*q
,
204 struct blk_mq_ctx
*ctx
= blk_mq_get_ctx(q
);
205 struct blk_mq_hw_ctx
*hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
207 rq
= __blk_mq_alloc_request(hctx
, gfp
& ~__GFP_WAIT
, reserved
);
209 blk_mq_rq_ctx_init(q
, ctx
, rq
, rw
);
213 if (gfp
& __GFP_WAIT
) {
214 __blk_mq_run_hw_queue(hctx
);
221 blk_mq_wait_for_tags(hctx
->tags
, reserved
);
227 struct request
*blk_mq_alloc_request(struct request_queue
*q
, int rw
, gfp_t gfp
)
231 if (blk_mq_queue_enter(q
))
234 rq
= blk_mq_alloc_request_pinned(q
, rw
, gfp
, false);
236 blk_mq_put_ctx(rq
->mq_ctx
);
240 struct request
*blk_mq_alloc_reserved_request(struct request_queue
*q
, int rw
,
245 if (blk_mq_queue_enter(q
))
248 rq
= blk_mq_alloc_request_pinned(q
, rw
, gfp
, true);
250 blk_mq_put_ctx(rq
->mq_ctx
);
253 EXPORT_SYMBOL(blk_mq_alloc_reserved_request
);
255 static void __blk_mq_free_request(struct blk_mq_hw_ctx
*hctx
,
256 struct blk_mq_ctx
*ctx
, struct request
*rq
)
258 const int tag
= rq
->tag
;
259 struct request_queue
*q
= rq
->q
;
261 blk_mq_put_tag(hctx
->tags
, tag
);
262 blk_mq_queue_exit(q
);
265 void blk_mq_free_request(struct request
*rq
)
267 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
268 struct blk_mq_hw_ctx
*hctx
;
269 struct request_queue
*q
= rq
->q
;
271 ctx
->rq_completed
[rq_is_sync(rq
)]++;
273 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
274 __blk_mq_free_request(hctx
, ctx
, rq
);
278 * Clone all relevant state from a request that has been put on hold in
279 * the flush state machine into the preallocated flush request that hangs
280 * off the request queue.
282 * For a driver the flush request should be invisible, that's why we are
283 * impersonating the original request here.
285 void blk_mq_clone_flush_request(struct request
*flush_rq
,
286 struct request
*orig_rq
)
288 struct blk_mq_hw_ctx
*hctx
=
289 orig_rq
->q
->mq_ops
->map_queue(orig_rq
->q
, orig_rq
->mq_ctx
->cpu
);
291 flush_rq
->mq_ctx
= orig_rq
->mq_ctx
;
292 flush_rq
->tag
= orig_rq
->tag
;
293 memcpy(blk_mq_rq_to_pdu(flush_rq
), blk_mq_rq_to_pdu(orig_rq
),
297 inline void __blk_mq_end_io(struct request
*rq
, int error
)
299 blk_account_io_done(rq
);
302 rq
->end_io(rq
, error
);
304 if (unlikely(blk_bidi_rq(rq
)))
305 blk_mq_free_request(rq
->next_rq
);
306 blk_mq_free_request(rq
);
309 EXPORT_SYMBOL(__blk_mq_end_io
);
311 void blk_mq_end_io(struct request
*rq
, int error
)
313 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
315 __blk_mq_end_io(rq
, error
);
317 EXPORT_SYMBOL(blk_mq_end_io
);
319 static void __blk_mq_complete_request_remote(void *data
)
321 struct request
*rq
= data
;
323 rq
->q
->softirq_done_fn(rq
);
326 void __blk_mq_complete_request(struct request
*rq
)
328 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
332 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
333 rq
->q
->softirq_done_fn(rq
);
338 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
339 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
341 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
342 rq
->csd
.func
= __blk_mq_complete_request_remote
;
345 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
347 rq
->q
->softirq_done_fn(rq
);
353 * blk_mq_complete_request - end I/O on a request
354 * @rq: the request being processed
357 * Ends all I/O on a request. It does not handle partial completions.
358 * The actual completion happens out-of-order, through a IPI handler.
360 void blk_mq_complete_request(struct request
*rq
)
362 if (unlikely(blk_should_fake_timeout(rq
->q
)))
364 if (!blk_mark_rq_complete(rq
))
365 __blk_mq_complete_request(rq
);
367 EXPORT_SYMBOL(blk_mq_complete_request
);
369 static void blk_mq_start_request(struct request
*rq
, bool last
)
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
);
380 * Just mark start time and set the started bit. Due to memory
381 * ordering, we know we'll see the correct deadline as long as
382 * REQ_ATOMIC_STARTED is seen.
384 rq
->deadline
= jiffies
+ q
->rq_timeout
;
387 * Mark us as started and clear complete. Complete might have been
388 * set if requeue raced with timeout, which then marked it as
389 * complete. So be sure to clear complete again when we start
390 * the request, otherwise we'll ignore the completion event.
392 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
393 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
395 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
397 * Make sure space for the drain appears. We know we can do
398 * this because max_hw_segments has been adjusted to be one
399 * fewer than the device can handle.
401 rq
->nr_phys_segments
++;
405 * Flag the last request in the series so that drivers know when IO
406 * should be kicked off, if they don't do it on a per-request basis.
408 * Note: the flag isn't the only condition drivers should do kick off.
409 * If drive is busy, the last request might not have the bit set.
412 rq
->cmd_flags
|= REQ_END
;
415 static void __blk_mq_requeue_request(struct request
*rq
)
417 struct request_queue
*q
= rq
->q
;
419 trace_block_rq_requeue(q
, rq
);
420 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
422 rq
->cmd_flags
&= ~REQ_END
;
424 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
425 rq
->nr_phys_segments
--;
428 void blk_mq_requeue_request(struct request
*rq
)
430 __blk_mq_requeue_request(rq
);
431 blk_clear_rq_complete(rq
);
433 BUG_ON(blk_queued_rq(rq
));
434 blk_mq_insert_request(rq
, true, true, false);
436 EXPORT_SYMBOL(blk_mq_requeue_request
);
438 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
440 return tags
->rqs
[tag
];
442 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
444 struct blk_mq_timeout_data
{
445 struct blk_mq_hw_ctx
*hctx
;
447 unsigned int *next_set
;
450 static void blk_mq_timeout_check(void *__data
, unsigned long *free_tags
)
452 struct blk_mq_timeout_data
*data
= __data
;
453 struct blk_mq_hw_ctx
*hctx
= data
->hctx
;
456 /* It may not be in flight yet (this is where
457 * the REQ_ATOMIC_STARTED flag comes in). The requests are
458 * statically allocated, so we know it's always safe to access the
459 * memory associated with a bit offset into ->rqs[].
465 tag
= find_next_zero_bit(free_tags
, hctx
->tags
->nr_tags
, tag
);
466 if (tag
>= hctx
->tags
->nr_tags
)
469 rq
= blk_mq_tag_to_rq(hctx
->tags
, tag
++);
470 if (rq
->q
!= hctx
->queue
)
472 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
475 blk_rq_check_expired(rq
, data
->next
, data
->next_set
);
479 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx
*hctx
,
481 unsigned int *next_set
)
483 struct blk_mq_timeout_data data
= {
486 .next_set
= next_set
,
490 * Ask the tagging code to iterate busy requests, so we can
491 * check them for timeout.
493 blk_mq_tag_busy_iter(hctx
->tags
, blk_mq_timeout_check
, &data
);
496 static enum blk_eh_timer_return
blk_mq_rq_timed_out(struct request
*rq
)
498 struct request_queue
*q
= rq
->q
;
501 * We know that complete is set at this point. If STARTED isn't set
502 * anymore, then the request isn't active and the "timeout" should
503 * just be ignored. This can happen due to the bitflag ordering.
504 * Timeout first checks if STARTED is set, and if it is, assumes
505 * the request is active. But if we race with completion, then
506 * we both flags will get cleared. So check here again, and ignore
507 * a timeout event with a request that isn't active.
509 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
510 return BLK_EH_NOT_HANDLED
;
512 if (!q
->mq_ops
->timeout
)
513 return BLK_EH_RESET_TIMER
;
515 return q
->mq_ops
->timeout(rq
);
518 static void blk_mq_rq_timer(unsigned long data
)
520 struct request_queue
*q
= (struct request_queue
*) data
;
521 struct blk_mq_hw_ctx
*hctx
;
522 unsigned long next
= 0;
525 queue_for_each_hw_ctx(q
, hctx
, i
)
526 blk_mq_hw_ctx_check_timeout(hctx
, &next
, &next_set
);
529 mod_timer(&q
->timeout
, round_jiffies_up(next
));
533 * Reverse check our software queue for entries that we could potentially
534 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
535 * too much time checking for merges.
537 static bool blk_mq_attempt_merge(struct request_queue
*q
,
538 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
543 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
549 if (!blk_rq_merge_ok(rq
, bio
))
552 el_ret
= blk_try_merge(rq
, bio
);
553 if (el_ret
== ELEVATOR_BACK_MERGE
) {
554 if (bio_attempt_back_merge(q
, rq
, bio
)) {
559 } else if (el_ret
== ELEVATOR_FRONT_MERGE
) {
560 if (bio_attempt_front_merge(q
, rq
, bio
)) {
572 * Run this hardware queue, pulling any software queues mapped to it in.
573 * Note that this function currently has various problems around ordering
574 * of IO. In particular, we'd like FIFO behaviour on handling existing
575 * items on the hctx->dispatch list. Ignore that for now.
577 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
579 struct request_queue
*q
= hctx
->queue
;
580 struct blk_mq_ctx
*ctx
;
585 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
));
587 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)))
593 * Touch any software queue that has pending entries.
595 for_each_set_bit(bit
, hctx
->ctx_map
, hctx
->nr_ctx
) {
596 clear_bit(bit
, hctx
->ctx_map
);
597 ctx
= hctx
->ctxs
[bit
];
599 spin_lock(&ctx
->lock
);
600 list_splice_tail_init(&ctx
->rq_list
, &rq_list
);
601 spin_unlock(&ctx
->lock
);
605 * If we have previous entries on our dispatch list, grab them
606 * and stuff them at the front for more fair dispatch.
608 if (!list_empty_careful(&hctx
->dispatch
)) {
609 spin_lock(&hctx
->lock
);
610 if (!list_empty(&hctx
->dispatch
))
611 list_splice_init(&hctx
->dispatch
, &rq_list
);
612 spin_unlock(&hctx
->lock
);
616 * Delete and return all entries from our dispatch list
621 * Now process all the entries, sending them to the driver.
623 while (!list_empty(&rq_list
)) {
626 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
627 list_del_init(&rq
->queuelist
);
629 blk_mq_start_request(rq
, list_empty(&rq_list
));
631 ret
= q
->mq_ops
->queue_rq(hctx
, rq
);
633 case BLK_MQ_RQ_QUEUE_OK
:
636 case BLK_MQ_RQ_QUEUE_BUSY
:
638 * FIXME: we should have a mechanism to stop the queue
639 * like blk_stop_queue, otherwise we will waste cpu
642 list_add(&rq
->queuelist
, &rq_list
);
643 __blk_mq_requeue_request(rq
);
646 pr_err("blk-mq: bad return on queue: %d\n", ret
);
647 case BLK_MQ_RQ_QUEUE_ERROR
:
649 blk_mq_end_io(rq
, rq
->errors
);
653 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
658 hctx
->dispatched
[0]++;
659 else if (queued
< (1 << (BLK_MQ_MAX_DISPATCH_ORDER
- 1)))
660 hctx
->dispatched
[ilog2(queued
) + 1]++;
663 * Any items that need requeuing? Stuff them into hctx->dispatch,
664 * that is where we will continue on next queue run.
666 if (!list_empty(&rq_list
)) {
667 spin_lock(&hctx
->lock
);
668 list_splice(&rq_list
, &hctx
->dispatch
);
669 spin_unlock(&hctx
->lock
);
674 * It'd be great if the workqueue API had a way to pass
675 * in a mask and had some smarts for more clever placement.
676 * For now we just round-robin here, switching for every
677 * BLK_MQ_CPU_WORK_BATCH queued items.
679 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
681 int cpu
= hctx
->next_cpu
;
683 if (--hctx
->next_cpu_batch
<= 0) {
686 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
687 if (next_cpu
>= nr_cpu_ids
)
688 next_cpu
= cpumask_first(hctx
->cpumask
);
690 hctx
->next_cpu
= next_cpu
;
691 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
697 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
699 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)))
702 if (!async
&& cpumask_test_cpu(smp_processor_id(), hctx
->cpumask
))
703 __blk_mq_run_hw_queue(hctx
);
704 else if (hctx
->queue
->nr_hw_queues
== 1)
705 kblockd_schedule_delayed_work(&hctx
->run_work
, 0);
709 cpu
= blk_mq_hctx_next_cpu(hctx
);
710 kblockd_schedule_delayed_work_on(cpu
, &hctx
->run_work
, 0);
714 void blk_mq_run_queues(struct request_queue
*q
, bool async
)
716 struct blk_mq_hw_ctx
*hctx
;
719 queue_for_each_hw_ctx(q
, hctx
, i
) {
720 if ((!blk_mq_hctx_has_pending(hctx
) &&
721 list_empty_careful(&hctx
->dispatch
)) ||
722 test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
726 blk_mq_run_hw_queue(hctx
, async
);
730 EXPORT_SYMBOL(blk_mq_run_queues
);
732 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
734 cancel_delayed_work(&hctx
->run_work
);
735 cancel_delayed_work(&hctx
->delay_work
);
736 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
738 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
740 void blk_mq_stop_hw_queues(struct request_queue
*q
)
742 struct blk_mq_hw_ctx
*hctx
;
745 queue_for_each_hw_ctx(q
, hctx
, i
)
746 blk_mq_stop_hw_queue(hctx
);
748 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
750 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
752 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
755 __blk_mq_run_hw_queue(hctx
);
758 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
760 void blk_mq_start_hw_queues(struct request_queue
*q
)
762 struct blk_mq_hw_ctx
*hctx
;
765 queue_for_each_hw_ctx(q
, hctx
, i
)
766 blk_mq_start_hw_queue(hctx
);
768 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
771 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
773 struct blk_mq_hw_ctx
*hctx
;
776 queue_for_each_hw_ctx(q
, hctx
, i
) {
777 if (!test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
780 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
782 blk_mq_run_hw_queue(hctx
, async
);
786 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
788 static void blk_mq_run_work_fn(struct work_struct
*work
)
790 struct blk_mq_hw_ctx
*hctx
;
792 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
794 __blk_mq_run_hw_queue(hctx
);
797 static void blk_mq_delay_work_fn(struct work_struct
*work
)
799 struct blk_mq_hw_ctx
*hctx
;
801 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
803 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
804 __blk_mq_run_hw_queue(hctx
);
807 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
809 unsigned long tmo
= msecs_to_jiffies(msecs
);
811 if (hctx
->queue
->nr_hw_queues
== 1)
812 kblockd_schedule_delayed_work(&hctx
->delay_work
, tmo
);
816 cpu
= blk_mq_hctx_next_cpu(hctx
);
817 kblockd_schedule_delayed_work_on(cpu
, &hctx
->delay_work
, tmo
);
820 EXPORT_SYMBOL(blk_mq_delay_queue
);
822 static void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
,
823 struct request
*rq
, bool at_head
)
825 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
827 trace_block_rq_insert(hctx
->queue
, rq
);
830 list_add(&rq
->queuelist
, &ctx
->rq_list
);
832 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
833 blk_mq_hctx_mark_pending(hctx
, ctx
);
836 * We do this early, to ensure we are on the right CPU.
841 void blk_mq_insert_request(struct request
*rq
, bool at_head
, bool run_queue
,
844 struct request_queue
*q
= rq
->q
;
845 struct blk_mq_hw_ctx
*hctx
;
846 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
, *current_ctx
;
848 current_ctx
= blk_mq_get_ctx(q
);
849 if (!cpu_online(ctx
->cpu
))
850 rq
->mq_ctx
= ctx
= current_ctx
;
852 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
854 if (rq
->cmd_flags
& (REQ_FLUSH
| REQ_FUA
) &&
855 !(rq
->cmd_flags
& (REQ_FLUSH_SEQ
))) {
856 blk_insert_flush(rq
);
858 spin_lock(&ctx
->lock
);
859 __blk_mq_insert_request(hctx
, rq
, at_head
);
860 spin_unlock(&ctx
->lock
);
864 blk_mq_run_hw_queue(hctx
, async
);
866 blk_mq_put_ctx(current_ctx
);
869 static void blk_mq_insert_requests(struct request_queue
*q
,
870 struct blk_mq_ctx
*ctx
,
871 struct list_head
*list
,
876 struct blk_mq_hw_ctx
*hctx
;
877 struct blk_mq_ctx
*current_ctx
;
879 trace_block_unplug(q
, depth
, !from_schedule
);
881 current_ctx
= blk_mq_get_ctx(q
);
883 if (!cpu_online(ctx
->cpu
))
885 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
888 * preemption doesn't flush plug list, so it's possible ctx->cpu is
891 spin_lock(&ctx
->lock
);
892 while (!list_empty(list
)) {
895 rq
= list_first_entry(list
, struct request
, queuelist
);
896 list_del_init(&rq
->queuelist
);
898 __blk_mq_insert_request(hctx
, rq
, false);
900 spin_unlock(&ctx
->lock
);
902 blk_mq_run_hw_queue(hctx
, from_schedule
);
903 blk_mq_put_ctx(current_ctx
);
906 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
908 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
909 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
911 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
912 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
913 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
916 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
918 struct blk_mq_ctx
*this_ctx
;
919 struct request_queue
*this_q
;
925 list_splice_init(&plug
->mq_list
, &list
);
927 list_sort(NULL
, &list
, plug_ctx_cmp
);
933 while (!list_empty(&list
)) {
934 rq
= list_entry_rq(list
.next
);
935 list_del_init(&rq
->queuelist
);
937 if (rq
->mq_ctx
!= this_ctx
) {
939 blk_mq_insert_requests(this_q
, this_ctx
,
944 this_ctx
= rq
->mq_ctx
;
950 list_add_tail(&rq
->queuelist
, &ctx_list
);
954 * If 'this_ctx' is set, we know we have entries to complete
955 * on 'ctx_list'. Do those.
958 blk_mq_insert_requests(this_q
, this_ctx
, &ctx_list
, depth
,
963 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
965 init_request_from_bio(rq
, bio
);
966 blk_account_io_start(rq
, 1);
969 static void blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
971 struct blk_mq_hw_ctx
*hctx
;
972 struct blk_mq_ctx
*ctx
;
973 const int is_sync
= rw_is_sync(bio
->bi_rw
);
974 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
975 int rw
= bio_data_dir(bio
);
977 unsigned int use_plug
, request_count
= 0;
980 * If we have multiple hardware queues, just go directly to
981 * one of those for sync IO.
983 use_plug
= !is_flush_fua
&& ((q
->nr_hw_queues
== 1) || !is_sync
);
985 blk_queue_bounce(q
, &bio
);
987 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
988 bio_endio(bio
, -EIO
);
992 if (use_plug
&& blk_attempt_plug_merge(q
, bio
, &request_count
))
995 if (blk_mq_queue_enter(q
)) {
996 bio_endio(bio
, -EIO
);
1000 ctx
= blk_mq_get_ctx(q
);
1001 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1005 trace_block_getrq(q
, bio
, rw
);
1006 rq
= __blk_mq_alloc_request(hctx
, GFP_ATOMIC
, false);
1008 blk_mq_rq_ctx_init(q
, ctx
, rq
, rw
);
1010 blk_mq_put_ctx(ctx
);
1011 trace_block_sleeprq(q
, bio
, rw
);
1012 rq
= blk_mq_alloc_request_pinned(q
, rw
, __GFP_WAIT
|GFP_ATOMIC
,
1015 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1020 if (unlikely(is_flush_fua
)) {
1021 blk_mq_bio_to_request(rq
, bio
);
1022 blk_insert_flush(rq
);
1027 * A task plug currently exists. Since this is completely lockless,
1028 * utilize that to temporarily store requests until the task is
1029 * either done or scheduled away.
1032 struct blk_plug
*plug
= current
->plug
;
1035 blk_mq_bio_to_request(rq
, bio
);
1036 if (list_empty(&plug
->mq_list
))
1037 trace_block_plug(q
);
1038 else if (request_count
>= BLK_MAX_REQUEST_COUNT
) {
1039 blk_flush_plug_list(plug
, false);
1040 trace_block_plug(q
);
1042 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1043 blk_mq_put_ctx(ctx
);
1048 if (!(hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
)) {
1049 init_request_from_bio(rq
, bio
);
1051 spin_lock(&ctx
->lock
);
1053 __blk_mq_insert_request(hctx
, rq
, false);
1054 spin_unlock(&ctx
->lock
);
1055 blk_account_io_start(rq
, 1);
1057 spin_lock(&ctx
->lock
);
1058 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1059 init_request_from_bio(rq
, bio
);
1063 spin_unlock(&ctx
->lock
);
1064 __blk_mq_free_request(hctx
, ctx
, rq
);
1069 * For a SYNC request, send it to the hardware immediately. For an
1070 * ASYNC request, just ensure that we run it later on. The latter
1071 * allows for merging opportunities and more efficient dispatching.
1074 blk_mq_run_hw_queue(hctx
, !is_sync
|| is_flush_fua
);
1075 blk_mq_put_ctx(ctx
);
1079 * Default mapping to a software queue, since we use one per CPU.
1081 struct blk_mq_hw_ctx
*blk_mq_map_queue(struct request_queue
*q
, const int cpu
)
1083 return q
->queue_hw_ctx
[q
->mq_map
[cpu
]];
1085 EXPORT_SYMBOL(blk_mq_map_queue
);
1087 struct blk_mq_hw_ctx
*blk_mq_alloc_single_hw_queue(struct blk_mq_tag_set
*set
,
1088 unsigned int hctx_index
)
1090 return kmalloc_node(sizeof(struct blk_mq_hw_ctx
),
1091 GFP_KERNEL
| __GFP_ZERO
, set
->numa_node
);
1093 EXPORT_SYMBOL(blk_mq_alloc_single_hw_queue
);
1095 void blk_mq_free_single_hw_queue(struct blk_mq_hw_ctx
*hctx
,
1096 unsigned int hctx_index
)
1100 EXPORT_SYMBOL(blk_mq_free_single_hw_queue
);
1102 static void blk_mq_hctx_notify(void *data
, unsigned long action
,
1105 struct blk_mq_hw_ctx
*hctx
= data
;
1106 struct request_queue
*q
= hctx
->queue
;
1107 struct blk_mq_ctx
*ctx
;
1110 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
1114 * Move ctx entries to new CPU, if this one is going away.
1116 ctx
= __blk_mq_get_ctx(q
, cpu
);
1118 spin_lock(&ctx
->lock
);
1119 if (!list_empty(&ctx
->rq_list
)) {
1120 list_splice_init(&ctx
->rq_list
, &tmp
);
1121 clear_bit(ctx
->index_hw
, hctx
->ctx_map
);
1123 spin_unlock(&ctx
->lock
);
1125 if (list_empty(&tmp
))
1128 ctx
= blk_mq_get_ctx(q
);
1129 spin_lock(&ctx
->lock
);
1131 while (!list_empty(&tmp
)) {
1134 rq
= list_first_entry(&tmp
, struct request
, queuelist
);
1136 list_move_tail(&rq
->queuelist
, &ctx
->rq_list
);
1139 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1140 blk_mq_hctx_mark_pending(hctx
, ctx
);
1142 spin_unlock(&ctx
->lock
);
1144 blk_mq_run_hw_queue(hctx
, true);
1145 blk_mq_put_ctx(ctx
);
1148 static void blk_mq_free_rq_map(struct blk_mq_tag_set
*set
,
1149 struct blk_mq_tags
*tags
, unsigned int hctx_idx
)
1153 if (tags
->rqs
&& set
->ops
->exit_request
) {
1156 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1159 set
->ops
->exit_request(set
->driver_data
, tags
->rqs
[i
],
1164 while (!list_empty(&tags
->page_list
)) {
1165 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1166 list_del_init(&page
->lru
);
1167 __free_pages(page
, page
->private);
1172 blk_mq_free_tags(tags
);
1175 static size_t order_to_size(unsigned int order
)
1177 return (size_t)PAGE_SIZE
<< order
;
1180 static struct blk_mq_tags
*blk_mq_init_rq_map(struct blk_mq_tag_set
*set
,
1181 unsigned int hctx_idx
)
1183 struct blk_mq_tags
*tags
;
1184 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1185 size_t rq_size
, left
;
1187 tags
= blk_mq_init_tags(set
->queue_depth
, set
->reserved_tags
,
1192 INIT_LIST_HEAD(&tags
->page_list
);
1194 tags
->rqs
= kmalloc_node(set
->queue_depth
* sizeof(struct request
*),
1195 GFP_KERNEL
, set
->numa_node
);
1197 blk_mq_free_tags(tags
);
1202 * rq_size is the size of the request plus driver payload, rounded
1203 * to the cacheline size
1205 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1207 left
= rq_size
* set
->queue_depth
;
1209 for (i
= 0; i
< set
->queue_depth
; ) {
1210 int this_order
= max_order
;
1215 while (left
< order_to_size(this_order
- 1) && this_order
)
1219 page
= alloc_pages_node(set
->numa_node
, GFP_KERNEL
,
1225 if (order_to_size(this_order
) < rq_size
)
1232 page
->private = this_order
;
1233 list_add_tail(&page
->lru
, &tags
->page_list
);
1235 p
= page_address(page
);
1236 entries_per_page
= order_to_size(this_order
) / rq_size
;
1237 to_do
= min(entries_per_page
, set
->queue_depth
- i
);
1238 left
-= to_do
* rq_size
;
1239 for (j
= 0; j
< to_do
; j
++) {
1241 if (set
->ops
->init_request
) {
1242 if (set
->ops
->init_request(set
->driver_data
,
1243 tags
->rqs
[i
], hctx_idx
, i
,
1256 pr_warn("%s: failed to allocate requests\n", __func__
);
1257 blk_mq_free_rq_map(set
, tags
, hctx_idx
);
1261 static int blk_mq_init_hw_queues(struct request_queue
*q
,
1262 struct blk_mq_tag_set
*set
)
1264 struct blk_mq_hw_ctx
*hctx
;
1268 * Initialize hardware queues
1270 queue_for_each_hw_ctx(q
, hctx
, i
) {
1271 unsigned int num_maps
;
1274 node
= hctx
->numa_node
;
1275 if (node
== NUMA_NO_NODE
)
1276 node
= hctx
->numa_node
= set
->numa_node
;
1278 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1279 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1280 spin_lock_init(&hctx
->lock
);
1281 INIT_LIST_HEAD(&hctx
->dispatch
);
1283 hctx
->queue_num
= i
;
1284 hctx
->flags
= set
->flags
;
1285 hctx
->cmd_size
= set
->cmd_size
;
1287 blk_mq_init_cpu_notifier(&hctx
->cpu_notifier
,
1288 blk_mq_hctx_notify
, hctx
);
1289 blk_mq_register_cpu_notifier(&hctx
->cpu_notifier
);
1291 hctx
->tags
= set
->tags
[i
];
1294 * Allocate space for all possible cpus to avoid allocation in
1297 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1302 num_maps
= ALIGN(nr_cpu_ids
, BITS_PER_LONG
) / BITS_PER_LONG
;
1303 hctx
->ctx_map
= kzalloc_node(num_maps
* sizeof(unsigned long),
1308 hctx
->nr_ctx_map
= num_maps
;
1311 if (set
->ops
->init_hctx
&&
1312 set
->ops
->init_hctx(hctx
, set
->driver_data
, i
))
1316 if (i
== q
->nr_hw_queues
)
1322 queue_for_each_hw_ctx(q
, hctx
, j
) {
1326 if (set
->ops
->exit_hctx
)
1327 set
->ops
->exit_hctx(hctx
, j
);
1329 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1331 kfree(hctx
->ctx_map
);
1337 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1338 unsigned int nr_hw_queues
)
1342 for_each_possible_cpu(i
) {
1343 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1344 struct blk_mq_hw_ctx
*hctx
;
1346 memset(__ctx
, 0, sizeof(*__ctx
));
1348 spin_lock_init(&__ctx
->lock
);
1349 INIT_LIST_HEAD(&__ctx
->rq_list
);
1352 /* If the cpu isn't online, the cpu is mapped to first hctx */
1356 hctx
= q
->mq_ops
->map_queue(q
, i
);
1357 cpumask_set_cpu(i
, hctx
->cpumask
);
1361 * Set local node, IFF we have more than one hw queue. If
1362 * not, we remain on the home node of the device
1364 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1365 hctx
->numa_node
= cpu_to_node(i
);
1369 static void blk_mq_map_swqueue(struct request_queue
*q
)
1372 struct blk_mq_hw_ctx
*hctx
;
1373 struct blk_mq_ctx
*ctx
;
1375 queue_for_each_hw_ctx(q
, hctx
, i
) {
1376 cpumask_clear(hctx
->cpumask
);
1381 * Map software to hardware queues
1383 queue_for_each_ctx(q
, ctx
, i
) {
1384 /* If the cpu isn't online, the cpu is mapped to first hctx */
1388 hctx
= q
->mq_ops
->map_queue(q
, i
);
1389 cpumask_set_cpu(i
, hctx
->cpumask
);
1390 ctx
->index_hw
= hctx
->nr_ctx
;
1391 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
1394 queue_for_each_hw_ctx(q
, hctx
, i
) {
1395 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
1396 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1400 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
1402 struct blk_mq_hw_ctx
**hctxs
;
1403 struct blk_mq_ctx
*ctx
;
1404 struct request_queue
*q
;
1407 ctx
= alloc_percpu(struct blk_mq_ctx
);
1409 return ERR_PTR(-ENOMEM
);
1411 hctxs
= kmalloc_node(set
->nr_hw_queues
* sizeof(*hctxs
), GFP_KERNEL
,
1417 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1418 hctxs
[i
] = set
->ops
->alloc_hctx(set
, i
);
1422 if (!zalloc_cpumask_var(&hctxs
[i
]->cpumask
, GFP_KERNEL
))
1425 hctxs
[i
]->numa_node
= NUMA_NO_NODE
;
1426 hctxs
[i
]->queue_num
= i
;
1429 q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
1433 q
->mq_map
= blk_mq_make_queue_map(set
);
1437 setup_timer(&q
->timeout
, blk_mq_rq_timer
, (unsigned long) q
);
1438 blk_queue_rq_timeout(q
, 30000);
1440 q
->nr_queues
= nr_cpu_ids
;
1441 q
->nr_hw_queues
= set
->nr_hw_queues
;
1444 q
->queue_hw_ctx
= hctxs
;
1446 q
->mq_ops
= set
->ops
;
1447 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
1449 q
->sg_reserved_size
= INT_MAX
;
1451 blk_queue_make_request(q
, blk_mq_make_request
);
1452 blk_queue_rq_timed_out(q
, blk_mq_rq_timed_out
);
1454 blk_queue_rq_timeout(q
, set
->timeout
);
1456 if (set
->ops
->complete
)
1457 blk_queue_softirq_done(q
, set
->ops
->complete
);
1459 blk_mq_init_flush(q
);
1460 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
1462 q
->flush_rq
= kzalloc(round_up(sizeof(struct request
) +
1463 set
->cmd_size
, cache_line_size()),
1468 if (blk_mq_init_hw_queues(q
, set
))
1471 blk_mq_map_swqueue(q
);
1473 mutex_lock(&all_q_mutex
);
1474 list_add_tail(&q
->all_q_node
, &all_q_list
);
1475 mutex_unlock(&all_q_mutex
);
1484 blk_cleanup_queue(q
);
1486 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1489 free_cpumask_var(hctxs
[i
]->cpumask
);
1490 set
->ops
->free_hctx(hctxs
[i
], i
);
1495 return ERR_PTR(-ENOMEM
);
1497 EXPORT_SYMBOL(blk_mq_init_queue
);
1499 void blk_mq_free_queue(struct request_queue
*q
)
1501 struct blk_mq_hw_ctx
*hctx
;
1504 queue_for_each_hw_ctx(q
, hctx
, i
) {
1505 kfree(hctx
->ctx_map
);
1507 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1508 if (q
->mq_ops
->exit_hctx
)
1509 q
->mq_ops
->exit_hctx(hctx
, i
);
1510 free_cpumask_var(hctx
->cpumask
);
1511 q
->mq_ops
->free_hctx(hctx
, i
);
1514 free_percpu(q
->queue_ctx
);
1515 kfree(q
->queue_hw_ctx
);
1518 q
->queue_ctx
= NULL
;
1519 q
->queue_hw_ctx
= NULL
;
1522 mutex_lock(&all_q_mutex
);
1523 list_del_init(&q
->all_q_node
);
1524 mutex_unlock(&all_q_mutex
);
1527 /* Basically redo blk_mq_init_queue with queue frozen */
1528 static void blk_mq_queue_reinit(struct request_queue
*q
)
1530 blk_mq_freeze_queue(q
);
1532 blk_mq_update_queue_map(q
->mq_map
, q
->nr_hw_queues
);
1535 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1536 * we should change hctx numa_node according to new topology (this
1537 * involves free and re-allocate memory, worthy doing?)
1540 blk_mq_map_swqueue(q
);
1542 blk_mq_unfreeze_queue(q
);
1545 static int blk_mq_queue_reinit_notify(struct notifier_block
*nb
,
1546 unsigned long action
, void *hcpu
)
1548 struct request_queue
*q
;
1551 * Before new mapping is established, hotadded cpu might already start
1552 * handling requests. This doesn't break anything as we map offline
1553 * CPUs to first hardware queue. We will re-init queue below to get
1556 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
&&
1557 action
!= CPU_ONLINE
&& action
!= CPU_ONLINE_FROZEN
)
1560 mutex_lock(&all_q_mutex
);
1561 list_for_each_entry(q
, &all_q_list
, all_q_node
)
1562 blk_mq_queue_reinit(q
);
1563 mutex_unlock(&all_q_mutex
);
1567 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
1571 if (!set
->nr_hw_queues
)
1573 if (!set
->queue_depth
|| set
->queue_depth
> BLK_MQ_MAX_DEPTH
)
1575 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
1578 if (!set
->nr_hw_queues
||
1579 !set
->ops
->queue_rq
|| !set
->ops
->map_queue
||
1580 !set
->ops
->alloc_hctx
|| !set
->ops
->free_hctx
)
1584 set
->tags
= kmalloc_node(set
->nr_hw_queues
*
1585 sizeof(struct blk_mq_tags
*),
1586 GFP_KERNEL
, set
->numa_node
);
1590 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1591 set
->tags
[i
] = blk_mq_init_rq_map(set
, i
);
1600 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1604 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
1606 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
1610 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
1611 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1614 EXPORT_SYMBOL(blk_mq_free_tag_set
);
1616 void blk_mq_disable_hotplug(void)
1618 mutex_lock(&all_q_mutex
);
1621 void blk_mq_enable_hotplug(void)
1623 mutex_unlock(&all_q_mutex
);
1626 static int __init
blk_mq_init(void)
1630 /* Must be called after percpu_counter_hotcpu_callback() */
1631 hotcpu_notifier(blk_mq_queue_reinit_notify
, -10);
1635 subsys_initcall(blk_mq_init
);