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
);
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
;
331 if (!ctx
->ipi_redirect
) {
332 rq
->q
->softirq_done_fn(rq
);
337 if (cpu
!= ctx
->cpu
&& cpu_online(ctx
->cpu
)) {
338 rq
->csd
.func
= __blk_mq_complete_request_remote
;
341 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
343 rq
->q
->softirq_done_fn(rq
);
349 * blk_mq_complete_request - end I/O on a request
350 * @rq: the request being processed
353 * Ends all I/O on a request. It does not handle partial completions.
354 * The actual completion happens out-of-order, through a IPI handler.
356 void blk_mq_complete_request(struct request
*rq
)
358 if (unlikely(blk_should_fake_timeout(rq
->q
)))
360 if (!blk_mark_rq_complete(rq
))
361 __blk_mq_complete_request(rq
);
363 EXPORT_SYMBOL(blk_mq_complete_request
);
365 static void blk_mq_start_request(struct request
*rq
, bool last
)
367 struct request_queue
*q
= rq
->q
;
369 trace_block_rq_issue(q
, rq
);
371 rq
->resid_len
= blk_rq_bytes(rq
);
372 if (unlikely(blk_bidi_rq(rq
)))
373 rq
->next_rq
->resid_len
= blk_rq_bytes(rq
->next_rq
);
376 * Just mark start time and set the started bit. Due to memory
377 * ordering, we know we'll see the correct deadline as long as
378 * REQ_ATOMIC_STARTED is seen.
380 rq
->deadline
= jiffies
+ q
->rq_timeout
;
381 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
383 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
385 * Make sure space for the drain appears. We know we can do
386 * this because max_hw_segments has been adjusted to be one
387 * fewer than the device can handle.
389 rq
->nr_phys_segments
++;
393 * Flag the last request in the series so that drivers know when IO
394 * should be kicked off, if they don't do it on a per-request basis.
396 * Note: the flag isn't the only condition drivers should do kick off.
397 * If drive is busy, the last request might not have the bit set.
400 rq
->cmd_flags
|= REQ_END
;
403 static void __blk_mq_requeue_request(struct request
*rq
)
405 struct request_queue
*q
= rq
->q
;
407 trace_block_rq_requeue(q
, rq
);
408 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
410 rq
->cmd_flags
&= ~REQ_END
;
412 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
413 rq
->nr_phys_segments
--;
416 void blk_mq_requeue_request(struct request
*rq
)
418 struct request_queue
*q
= rq
->q
;
420 __blk_mq_requeue_request(rq
);
421 blk_clear_rq_complete(rq
);
423 trace_block_rq_requeue(q
, rq
);
425 BUG_ON(blk_queued_rq(rq
));
426 blk_mq_insert_request(rq
, true, true, false);
428 EXPORT_SYMBOL(blk_mq_requeue_request
);
430 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
432 return tags
->rqs
[tag
];
434 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
436 struct blk_mq_timeout_data
{
437 struct blk_mq_hw_ctx
*hctx
;
439 unsigned int *next_set
;
442 static void blk_mq_timeout_check(void *__data
, unsigned long *free_tags
)
444 struct blk_mq_timeout_data
*data
= __data
;
445 struct blk_mq_hw_ctx
*hctx
= data
->hctx
;
448 /* It may not be in flight yet (this is where
449 * the REQ_ATOMIC_STARTED flag comes in). The requests are
450 * statically allocated, so we know it's always safe to access the
451 * memory associated with a bit offset into ->rqs[].
457 tag
= find_next_zero_bit(free_tags
, hctx
->tags
->nr_tags
, tag
);
458 if (tag
>= hctx
->tags
->nr_tags
)
461 rq
= blk_mq_tag_to_rq(hctx
->tags
, tag
++);
462 if (rq
->q
!= hctx
->queue
)
464 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
467 blk_rq_check_expired(rq
, data
->next
, data
->next_set
);
471 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx
*hctx
,
473 unsigned int *next_set
)
475 struct blk_mq_timeout_data data
= {
478 .next_set
= next_set
,
482 * Ask the tagging code to iterate busy requests, so we can
483 * check them for timeout.
485 blk_mq_tag_busy_iter(hctx
->tags
, blk_mq_timeout_check
, &data
);
488 static void blk_mq_rq_timer(unsigned long data
)
490 struct request_queue
*q
= (struct request_queue
*) data
;
491 struct blk_mq_hw_ctx
*hctx
;
492 unsigned long next
= 0;
495 queue_for_each_hw_ctx(q
, hctx
, i
)
496 blk_mq_hw_ctx_check_timeout(hctx
, &next
, &next_set
);
499 mod_timer(&q
->timeout
, round_jiffies_up(next
));
503 * Reverse check our software queue for entries that we could potentially
504 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
505 * too much time checking for merges.
507 static bool blk_mq_attempt_merge(struct request_queue
*q
,
508 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
513 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
519 if (!blk_rq_merge_ok(rq
, bio
))
522 el_ret
= blk_try_merge(rq
, bio
);
523 if (el_ret
== ELEVATOR_BACK_MERGE
) {
524 if (bio_attempt_back_merge(q
, rq
, bio
)) {
529 } else if (el_ret
== ELEVATOR_FRONT_MERGE
) {
530 if (bio_attempt_front_merge(q
, rq
, bio
)) {
541 void blk_mq_add_timer(struct request
*rq
)
543 __blk_add_timer(rq
, NULL
);
547 * Run this hardware queue, pulling any software queues mapped to it in.
548 * Note that this function currently has various problems around ordering
549 * of IO. In particular, we'd like FIFO behaviour on handling existing
550 * items on the hctx->dispatch list. Ignore that for now.
552 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
554 struct request_queue
*q
= hctx
->queue
;
555 struct blk_mq_ctx
*ctx
;
560 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
));
562 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)))
568 * Touch any software queue that has pending entries.
570 for_each_set_bit(bit
, hctx
->ctx_map
, hctx
->nr_ctx
) {
571 clear_bit(bit
, hctx
->ctx_map
);
572 ctx
= hctx
->ctxs
[bit
];
573 BUG_ON(bit
!= ctx
->index_hw
);
575 spin_lock(&ctx
->lock
);
576 list_splice_tail_init(&ctx
->rq_list
, &rq_list
);
577 spin_unlock(&ctx
->lock
);
581 * If we have previous entries on our dispatch list, grab them
582 * and stuff them at the front for more fair dispatch.
584 if (!list_empty_careful(&hctx
->dispatch
)) {
585 spin_lock(&hctx
->lock
);
586 if (!list_empty(&hctx
->dispatch
))
587 list_splice_init(&hctx
->dispatch
, &rq_list
);
588 spin_unlock(&hctx
->lock
);
592 * Delete and return all entries from our dispatch list
597 * Now process all the entries, sending them to the driver.
599 while (!list_empty(&rq_list
)) {
602 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
603 list_del_init(&rq
->queuelist
);
605 blk_mq_start_request(rq
, list_empty(&rq_list
));
607 ret
= q
->mq_ops
->queue_rq(hctx
, rq
);
609 case BLK_MQ_RQ_QUEUE_OK
:
612 case BLK_MQ_RQ_QUEUE_BUSY
:
614 * FIXME: we should have a mechanism to stop the queue
615 * like blk_stop_queue, otherwise we will waste cpu
618 list_add(&rq
->queuelist
, &rq_list
);
619 __blk_mq_requeue_request(rq
);
622 pr_err("blk-mq: bad return on queue: %d\n", ret
);
623 case BLK_MQ_RQ_QUEUE_ERROR
:
625 blk_mq_end_io(rq
, rq
->errors
);
629 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
634 hctx
->dispatched
[0]++;
635 else if (queued
< (1 << (BLK_MQ_MAX_DISPATCH_ORDER
- 1)))
636 hctx
->dispatched
[ilog2(queued
) + 1]++;
639 * Any items that need requeuing? Stuff them into hctx->dispatch,
640 * that is where we will continue on next queue run.
642 if (!list_empty(&rq_list
)) {
643 spin_lock(&hctx
->lock
);
644 list_splice(&rq_list
, &hctx
->dispatch
);
645 spin_unlock(&hctx
->lock
);
649 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
651 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)))
654 if (!async
&& cpumask_test_cpu(smp_processor_id(), hctx
->cpumask
))
655 __blk_mq_run_hw_queue(hctx
);
656 else if (hctx
->queue
->nr_hw_queues
== 1)
657 kblockd_schedule_delayed_work(&hctx
->run_work
, 0);
662 * It'd be great if the workqueue API had a way to pass
663 * in a mask and had some smarts for more clever placement
664 * than the first CPU. Or we could round-robin here. For now,
665 * just queue on the first CPU.
667 cpu
= cpumask_first(hctx
->cpumask
);
668 kblockd_schedule_delayed_work_on(cpu
, &hctx
->run_work
, 0);
672 void blk_mq_run_queues(struct request_queue
*q
, bool async
)
674 struct blk_mq_hw_ctx
*hctx
;
677 queue_for_each_hw_ctx(q
, hctx
, i
) {
678 if ((!blk_mq_hctx_has_pending(hctx
) &&
679 list_empty_careful(&hctx
->dispatch
)) ||
680 test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
684 blk_mq_run_hw_queue(hctx
, async
);
688 EXPORT_SYMBOL(blk_mq_run_queues
);
690 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
692 cancel_delayed_work(&hctx
->run_work
);
693 cancel_delayed_work(&hctx
->delay_work
);
694 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
696 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
698 void blk_mq_stop_hw_queues(struct request_queue
*q
)
700 struct blk_mq_hw_ctx
*hctx
;
703 queue_for_each_hw_ctx(q
, hctx
, i
)
704 blk_mq_stop_hw_queue(hctx
);
706 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
708 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
710 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
713 __blk_mq_run_hw_queue(hctx
);
716 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
718 void blk_mq_start_hw_queues(struct request_queue
*q
)
720 struct blk_mq_hw_ctx
*hctx
;
723 queue_for_each_hw_ctx(q
, hctx
, i
)
724 blk_mq_start_hw_queue(hctx
);
726 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
729 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
731 struct blk_mq_hw_ctx
*hctx
;
734 queue_for_each_hw_ctx(q
, hctx
, i
) {
735 if (!test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
738 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
740 blk_mq_run_hw_queue(hctx
, async
);
744 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
746 static void blk_mq_run_work_fn(struct work_struct
*work
)
748 struct blk_mq_hw_ctx
*hctx
;
750 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
752 __blk_mq_run_hw_queue(hctx
);
755 static void blk_mq_delay_work_fn(struct work_struct
*work
)
757 struct blk_mq_hw_ctx
*hctx
;
759 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
761 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
762 __blk_mq_run_hw_queue(hctx
);
765 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
767 unsigned long tmo
= msecs_to_jiffies(msecs
);
769 if (hctx
->queue
->nr_hw_queues
== 1)
770 kblockd_schedule_delayed_work(&hctx
->delay_work
, tmo
);
775 * It'd be great if the workqueue API had a way to pass
776 * in a mask and had some smarts for more clever placement
777 * than the first CPU. Or we could round-robin here. For now,
778 * just queue on the first CPU.
780 cpu
= cpumask_first(hctx
->cpumask
);
781 kblockd_schedule_delayed_work_on(cpu
, &hctx
->delay_work
, tmo
);
784 EXPORT_SYMBOL(blk_mq_delay_queue
);
786 static void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
,
787 struct request
*rq
, bool at_head
)
789 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
791 trace_block_rq_insert(hctx
->queue
, rq
);
794 list_add(&rq
->queuelist
, &ctx
->rq_list
);
796 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
797 blk_mq_hctx_mark_pending(hctx
, ctx
);
800 * We do this early, to ensure we are on the right CPU.
802 blk_mq_add_timer(rq
);
805 void blk_mq_insert_request(struct request
*rq
, bool at_head
, bool run_queue
,
808 struct request_queue
*q
= rq
->q
;
809 struct blk_mq_hw_ctx
*hctx
;
810 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
, *current_ctx
;
812 current_ctx
= blk_mq_get_ctx(q
);
813 if (!cpu_online(ctx
->cpu
))
814 rq
->mq_ctx
= ctx
= current_ctx
;
816 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
818 if (rq
->cmd_flags
& (REQ_FLUSH
| REQ_FUA
) &&
819 !(rq
->cmd_flags
& (REQ_FLUSH_SEQ
))) {
820 blk_insert_flush(rq
);
822 spin_lock(&ctx
->lock
);
823 __blk_mq_insert_request(hctx
, rq
, at_head
);
824 spin_unlock(&ctx
->lock
);
828 blk_mq_run_hw_queue(hctx
, async
);
830 blk_mq_put_ctx(current_ctx
);
833 static void blk_mq_insert_requests(struct request_queue
*q
,
834 struct blk_mq_ctx
*ctx
,
835 struct list_head
*list
,
840 struct blk_mq_hw_ctx
*hctx
;
841 struct blk_mq_ctx
*current_ctx
;
843 trace_block_unplug(q
, depth
, !from_schedule
);
845 current_ctx
= blk_mq_get_ctx(q
);
847 if (!cpu_online(ctx
->cpu
))
849 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
852 * preemption doesn't flush plug list, so it's possible ctx->cpu is
855 spin_lock(&ctx
->lock
);
856 while (!list_empty(list
)) {
859 rq
= list_first_entry(list
, struct request
, queuelist
);
860 list_del_init(&rq
->queuelist
);
862 __blk_mq_insert_request(hctx
, rq
, false);
864 spin_unlock(&ctx
->lock
);
866 blk_mq_run_hw_queue(hctx
, from_schedule
);
867 blk_mq_put_ctx(current_ctx
);
870 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
872 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
873 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
875 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
876 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
877 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
880 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
882 struct blk_mq_ctx
*this_ctx
;
883 struct request_queue
*this_q
;
889 list_splice_init(&plug
->mq_list
, &list
);
891 list_sort(NULL
, &list
, plug_ctx_cmp
);
897 while (!list_empty(&list
)) {
898 rq
= list_entry_rq(list
.next
);
899 list_del_init(&rq
->queuelist
);
901 if (rq
->mq_ctx
!= this_ctx
) {
903 blk_mq_insert_requests(this_q
, this_ctx
,
908 this_ctx
= rq
->mq_ctx
;
914 list_add_tail(&rq
->queuelist
, &ctx_list
);
918 * If 'this_ctx' is set, we know we have entries to complete
919 * on 'ctx_list'. Do those.
922 blk_mq_insert_requests(this_q
, this_ctx
, &ctx_list
, depth
,
927 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
929 init_request_from_bio(rq
, bio
);
930 blk_account_io_start(rq
, 1);
933 static void blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
935 struct blk_mq_hw_ctx
*hctx
;
936 struct blk_mq_ctx
*ctx
;
937 const int is_sync
= rw_is_sync(bio
->bi_rw
);
938 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
939 int rw
= bio_data_dir(bio
);
941 unsigned int use_plug
, request_count
= 0;
944 * If we have multiple hardware queues, just go directly to
945 * one of those for sync IO.
947 use_plug
= !is_flush_fua
&& ((q
->nr_hw_queues
== 1) || !is_sync
);
949 blk_queue_bounce(q
, &bio
);
951 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
952 bio_endio(bio
, -EIO
);
956 if (use_plug
&& blk_attempt_plug_merge(q
, bio
, &request_count
))
959 if (blk_mq_queue_enter(q
)) {
960 bio_endio(bio
, -EIO
);
964 ctx
= blk_mq_get_ctx(q
);
965 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
969 trace_block_getrq(q
, bio
, rw
);
970 rq
= __blk_mq_alloc_request(hctx
, GFP_ATOMIC
, false);
972 blk_mq_rq_ctx_init(q
, ctx
, rq
, rw
);
975 trace_block_sleeprq(q
, bio
, rw
);
976 rq
= blk_mq_alloc_request_pinned(q
, rw
, __GFP_WAIT
|GFP_ATOMIC
,
979 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
984 if (unlikely(is_flush_fua
)) {
985 blk_mq_bio_to_request(rq
, bio
);
986 blk_insert_flush(rq
);
991 * A task plug currently exists. Since this is completely lockless,
992 * utilize that to temporarily store requests until the task is
993 * either done or scheduled away.
996 struct blk_plug
*plug
= current
->plug
;
999 blk_mq_bio_to_request(rq
, bio
);
1000 if (list_empty(&plug
->mq_list
))
1001 trace_block_plug(q
);
1002 else if (request_count
>= BLK_MAX_REQUEST_COUNT
) {
1003 blk_flush_plug_list(plug
, false);
1004 trace_block_plug(q
);
1006 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1007 blk_mq_put_ctx(ctx
);
1012 spin_lock(&ctx
->lock
);
1014 if ((hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1015 blk_mq_attempt_merge(q
, ctx
, bio
))
1016 __blk_mq_free_request(hctx
, ctx
, rq
);
1018 blk_mq_bio_to_request(rq
, bio
);
1019 __blk_mq_insert_request(hctx
, rq
, false);
1022 spin_unlock(&ctx
->lock
);
1025 * For a SYNC request, send it to the hardware immediately. For an
1026 * ASYNC request, just ensure that we run it later on. The latter
1027 * allows for merging opportunities and more efficient dispatching.
1030 blk_mq_run_hw_queue(hctx
, !is_sync
|| is_flush_fua
);
1031 blk_mq_put_ctx(ctx
);
1035 * Default mapping to a software queue, since we use one per CPU.
1037 struct blk_mq_hw_ctx
*blk_mq_map_queue(struct request_queue
*q
, const int cpu
)
1039 return q
->queue_hw_ctx
[q
->mq_map
[cpu
]];
1041 EXPORT_SYMBOL(blk_mq_map_queue
);
1043 struct blk_mq_hw_ctx
*blk_mq_alloc_single_hw_queue(struct blk_mq_tag_set
*set
,
1044 unsigned int hctx_index
)
1046 return kmalloc_node(sizeof(struct blk_mq_hw_ctx
),
1047 GFP_KERNEL
| __GFP_ZERO
, set
->numa_node
);
1049 EXPORT_SYMBOL(blk_mq_alloc_single_hw_queue
);
1051 void blk_mq_free_single_hw_queue(struct blk_mq_hw_ctx
*hctx
,
1052 unsigned int hctx_index
)
1056 EXPORT_SYMBOL(blk_mq_free_single_hw_queue
);
1058 static void blk_mq_hctx_notify(void *data
, unsigned long action
,
1061 struct blk_mq_hw_ctx
*hctx
= data
;
1062 struct request_queue
*q
= hctx
->queue
;
1063 struct blk_mq_ctx
*ctx
;
1066 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
1070 * Move ctx entries to new CPU, if this one is going away.
1072 ctx
= __blk_mq_get_ctx(q
, cpu
);
1074 spin_lock(&ctx
->lock
);
1075 if (!list_empty(&ctx
->rq_list
)) {
1076 list_splice_init(&ctx
->rq_list
, &tmp
);
1077 clear_bit(ctx
->index_hw
, hctx
->ctx_map
);
1079 spin_unlock(&ctx
->lock
);
1081 if (list_empty(&tmp
))
1084 ctx
= blk_mq_get_ctx(q
);
1085 spin_lock(&ctx
->lock
);
1087 while (!list_empty(&tmp
)) {
1090 rq
= list_first_entry(&tmp
, struct request
, queuelist
);
1092 list_move_tail(&rq
->queuelist
, &ctx
->rq_list
);
1095 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1096 blk_mq_hctx_mark_pending(hctx
, ctx
);
1098 spin_unlock(&ctx
->lock
);
1100 blk_mq_run_hw_queue(hctx
, true);
1101 blk_mq_put_ctx(ctx
);
1104 static void blk_mq_free_rq_map(struct blk_mq_tag_set
*set
,
1105 struct blk_mq_tags
*tags
, unsigned int hctx_idx
)
1109 if (tags
->rqs
&& set
->ops
->exit_request
) {
1112 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1115 set
->ops
->exit_request(set
->driver_data
, tags
->rqs
[i
],
1120 while (!list_empty(&tags
->page_list
)) {
1121 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1122 list_del_init(&page
->lru
);
1123 __free_pages(page
, page
->private);
1128 blk_mq_free_tags(tags
);
1131 static size_t order_to_size(unsigned int order
)
1133 return (size_t)PAGE_SIZE
<< order
;
1136 static struct blk_mq_tags
*blk_mq_init_rq_map(struct blk_mq_tag_set
*set
,
1137 unsigned int hctx_idx
)
1139 struct blk_mq_tags
*tags
;
1140 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1141 size_t rq_size
, left
;
1143 tags
= blk_mq_init_tags(set
->queue_depth
, set
->reserved_tags
,
1148 INIT_LIST_HEAD(&tags
->page_list
);
1150 tags
->rqs
= kmalloc_node(set
->queue_depth
* sizeof(struct request
*),
1151 GFP_KERNEL
, set
->numa_node
);
1153 blk_mq_free_tags(tags
);
1158 * rq_size is the size of the request plus driver payload, rounded
1159 * to the cacheline size
1161 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1163 left
= rq_size
* set
->queue_depth
;
1165 for (i
= 0; i
< set
->queue_depth
; ) {
1166 int this_order
= max_order
;
1171 while (left
< order_to_size(this_order
- 1) && this_order
)
1175 page
= alloc_pages_node(set
->numa_node
, GFP_KERNEL
,
1181 if (order_to_size(this_order
) < rq_size
)
1188 page
->private = this_order
;
1189 list_add_tail(&page
->lru
, &tags
->page_list
);
1191 p
= page_address(page
);
1192 entries_per_page
= order_to_size(this_order
) / rq_size
;
1193 to_do
= min(entries_per_page
, set
->queue_depth
- i
);
1194 left
-= to_do
* rq_size
;
1195 for (j
= 0; j
< to_do
; j
++) {
1197 if (set
->ops
->init_request
) {
1198 if (set
->ops
->init_request(set
->driver_data
,
1199 tags
->rqs
[i
], hctx_idx
, i
,
1212 pr_warn("%s: failed to allocate requests\n", __func__
);
1213 blk_mq_free_rq_map(set
, tags
, hctx_idx
);
1217 static int blk_mq_init_hw_queues(struct request_queue
*q
,
1218 struct blk_mq_tag_set
*set
)
1220 struct blk_mq_hw_ctx
*hctx
;
1224 * Initialize hardware queues
1226 queue_for_each_hw_ctx(q
, hctx
, i
) {
1227 unsigned int num_maps
;
1230 node
= hctx
->numa_node
;
1231 if (node
== NUMA_NO_NODE
)
1232 node
= hctx
->numa_node
= set
->numa_node
;
1234 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1235 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1236 spin_lock_init(&hctx
->lock
);
1237 INIT_LIST_HEAD(&hctx
->dispatch
);
1239 hctx
->queue_num
= i
;
1240 hctx
->flags
= set
->flags
;
1241 hctx
->cmd_size
= set
->cmd_size
;
1243 blk_mq_init_cpu_notifier(&hctx
->cpu_notifier
,
1244 blk_mq_hctx_notify
, hctx
);
1245 blk_mq_register_cpu_notifier(&hctx
->cpu_notifier
);
1247 hctx
->tags
= set
->tags
[i
];
1250 * Allocate space for all possible cpus to avoid allocation in
1253 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1258 num_maps
= ALIGN(nr_cpu_ids
, BITS_PER_LONG
) / BITS_PER_LONG
;
1259 hctx
->ctx_map
= kzalloc_node(num_maps
* sizeof(unsigned long),
1264 hctx
->nr_ctx_map
= num_maps
;
1267 if (set
->ops
->init_hctx
&&
1268 set
->ops
->init_hctx(hctx
, set
->driver_data
, i
))
1272 if (i
== q
->nr_hw_queues
)
1278 queue_for_each_hw_ctx(q
, hctx
, j
) {
1282 if (set
->ops
->exit_hctx
)
1283 set
->ops
->exit_hctx(hctx
, j
);
1285 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1287 kfree(hctx
->ctx_map
);
1293 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1294 unsigned int nr_hw_queues
)
1298 for_each_possible_cpu(i
) {
1299 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1300 struct blk_mq_hw_ctx
*hctx
;
1302 memset(__ctx
, 0, sizeof(*__ctx
));
1304 spin_lock_init(&__ctx
->lock
);
1305 INIT_LIST_HEAD(&__ctx
->rq_list
);
1308 /* If the cpu isn't online, the cpu is mapped to first hctx */
1312 hctx
= q
->mq_ops
->map_queue(q
, i
);
1313 cpumask_set_cpu(i
, hctx
->cpumask
);
1317 * Set local node, IFF we have more than one hw queue. If
1318 * not, we remain on the home node of the device
1320 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1321 hctx
->numa_node
= cpu_to_node(i
);
1325 static void blk_mq_map_swqueue(struct request_queue
*q
)
1328 struct blk_mq_hw_ctx
*hctx
;
1329 struct blk_mq_ctx
*ctx
;
1331 queue_for_each_hw_ctx(q
, hctx
, i
) {
1332 cpumask_clear(hctx
->cpumask
);
1337 * Map software to hardware queues
1339 queue_for_each_ctx(q
, ctx
, i
) {
1340 /* If the cpu isn't online, the cpu is mapped to first hctx */
1344 hctx
= q
->mq_ops
->map_queue(q
, i
);
1345 cpumask_set_cpu(i
, hctx
->cpumask
);
1346 ctx
->index_hw
= hctx
->nr_ctx
;
1347 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
1351 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
1353 struct blk_mq_hw_ctx
**hctxs
;
1354 struct blk_mq_ctx
*ctx
;
1355 struct request_queue
*q
;
1358 ctx
= alloc_percpu(struct blk_mq_ctx
);
1360 return ERR_PTR(-ENOMEM
);
1362 hctxs
= kmalloc_node(set
->nr_hw_queues
* sizeof(*hctxs
), GFP_KERNEL
,
1368 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1369 hctxs
[i
] = set
->ops
->alloc_hctx(set
, i
);
1373 if (!zalloc_cpumask_var(&hctxs
[i
]->cpumask
, GFP_KERNEL
))
1376 hctxs
[i
]->numa_node
= NUMA_NO_NODE
;
1377 hctxs
[i
]->queue_num
= i
;
1380 q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
1384 q
->mq_map
= blk_mq_make_queue_map(set
);
1388 setup_timer(&q
->timeout
, blk_mq_rq_timer
, (unsigned long) q
);
1389 blk_queue_rq_timeout(q
, 30000);
1391 q
->nr_queues
= nr_cpu_ids
;
1392 q
->nr_hw_queues
= set
->nr_hw_queues
;
1395 q
->queue_hw_ctx
= hctxs
;
1397 q
->mq_ops
= set
->ops
;
1398 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
1400 q
->sg_reserved_size
= INT_MAX
;
1402 blk_queue_make_request(q
, blk_mq_make_request
);
1403 blk_queue_rq_timed_out(q
, set
->ops
->timeout
);
1405 blk_queue_rq_timeout(q
, set
->timeout
);
1407 if (set
->ops
->complete
)
1408 blk_queue_softirq_done(q
, set
->ops
->complete
);
1410 blk_mq_init_flush(q
);
1411 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
1413 q
->flush_rq
= kzalloc(round_up(sizeof(struct request
) +
1414 set
->cmd_size
, cache_line_size()),
1419 if (blk_mq_init_hw_queues(q
, set
))
1422 blk_mq_map_swqueue(q
);
1424 mutex_lock(&all_q_mutex
);
1425 list_add_tail(&q
->all_q_node
, &all_q_list
);
1426 mutex_unlock(&all_q_mutex
);
1435 blk_cleanup_queue(q
);
1437 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1440 free_cpumask_var(hctxs
[i
]->cpumask
);
1441 set
->ops
->free_hctx(hctxs
[i
], i
);
1446 return ERR_PTR(-ENOMEM
);
1448 EXPORT_SYMBOL(blk_mq_init_queue
);
1450 void blk_mq_free_queue(struct request_queue
*q
)
1452 struct blk_mq_hw_ctx
*hctx
;
1455 queue_for_each_hw_ctx(q
, hctx
, i
) {
1456 kfree(hctx
->ctx_map
);
1458 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1459 if (q
->mq_ops
->exit_hctx
)
1460 q
->mq_ops
->exit_hctx(hctx
, i
);
1461 free_cpumask_var(hctx
->cpumask
);
1462 q
->mq_ops
->free_hctx(hctx
, i
);
1465 free_percpu(q
->queue_ctx
);
1466 kfree(q
->queue_hw_ctx
);
1469 q
->queue_ctx
= NULL
;
1470 q
->queue_hw_ctx
= NULL
;
1473 mutex_lock(&all_q_mutex
);
1474 list_del_init(&q
->all_q_node
);
1475 mutex_unlock(&all_q_mutex
);
1478 /* Basically redo blk_mq_init_queue with queue frozen */
1479 static void blk_mq_queue_reinit(struct request_queue
*q
)
1481 blk_mq_freeze_queue(q
);
1483 blk_mq_update_queue_map(q
->mq_map
, q
->nr_hw_queues
);
1486 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1487 * we should change hctx numa_node according to new topology (this
1488 * involves free and re-allocate memory, worthy doing?)
1491 blk_mq_map_swqueue(q
);
1493 blk_mq_unfreeze_queue(q
);
1496 static int blk_mq_queue_reinit_notify(struct notifier_block
*nb
,
1497 unsigned long action
, void *hcpu
)
1499 struct request_queue
*q
;
1502 * Before new mapping is established, hotadded cpu might already start
1503 * handling requests. This doesn't break anything as we map offline
1504 * CPUs to first hardware queue. We will re-init queue below to get
1507 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
&&
1508 action
!= CPU_ONLINE
&& action
!= CPU_ONLINE_FROZEN
)
1511 mutex_lock(&all_q_mutex
);
1512 list_for_each_entry(q
, &all_q_list
, all_q_node
)
1513 blk_mq_queue_reinit(q
);
1514 mutex_unlock(&all_q_mutex
);
1518 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
1522 if (!set
->nr_hw_queues
)
1524 if (!set
->queue_depth
|| set
->queue_depth
> BLK_MQ_MAX_DEPTH
)
1526 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
1529 if (!set
->nr_hw_queues
||
1530 !set
->ops
->queue_rq
|| !set
->ops
->map_queue
||
1531 !set
->ops
->alloc_hctx
|| !set
->ops
->free_hctx
)
1535 set
->tags
= kmalloc_node(set
->nr_hw_queues
*
1536 sizeof(struct blk_mq_tags
*),
1537 GFP_KERNEL
, set
->numa_node
);
1541 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1542 set
->tags
[i
] = blk_mq_init_rq_map(set
, i
);
1551 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1555 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
1557 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
1561 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
1562 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1565 EXPORT_SYMBOL(blk_mq_free_tag_set
);
1567 void blk_mq_disable_hotplug(void)
1569 mutex_lock(&all_q_mutex
);
1572 void blk_mq_enable_hotplug(void)
1574 mutex_unlock(&all_q_mutex
);
1577 static int __init
blk_mq_init(void)
1581 /* Must be called after percpu_counter_hotcpu_callback() */
1582 hotcpu_notifier(blk_mq_queue_reinit_notify
, -10);
1586 subsys_initcall(blk_mq_init
);