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 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
418 return tags
->rqs
[tag
];
420 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
422 struct blk_mq_timeout_data
{
423 struct blk_mq_hw_ctx
*hctx
;
425 unsigned int *next_set
;
428 static void blk_mq_timeout_check(void *__data
, unsigned long *free_tags
)
430 struct blk_mq_timeout_data
*data
= __data
;
431 struct blk_mq_hw_ctx
*hctx
= data
->hctx
;
434 /* It may not be in flight yet (this is where
435 * the REQ_ATOMIC_STARTED flag comes in). The requests are
436 * statically allocated, so we know it's always safe to access the
437 * memory associated with a bit offset into ->rqs[].
443 tag
= find_next_zero_bit(free_tags
, hctx
->tags
->nr_tags
, tag
);
444 if (tag
>= hctx
->tags
->nr_tags
)
447 rq
= blk_mq_tag_to_rq(hctx
->tags
, tag
++);
448 if (rq
->q
!= hctx
->queue
)
450 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
453 blk_rq_check_expired(rq
, data
->next
, data
->next_set
);
457 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx
*hctx
,
459 unsigned int *next_set
)
461 struct blk_mq_timeout_data data
= {
464 .next_set
= next_set
,
468 * Ask the tagging code to iterate busy requests, so we can
469 * check them for timeout.
471 blk_mq_tag_busy_iter(hctx
->tags
, blk_mq_timeout_check
, &data
);
474 static void blk_mq_rq_timer(unsigned long data
)
476 struct request_queue
*q
= (struct request_queue
*) data
;
477 struct blk_mq_hw_ctx
*hctx
;
478 unsigned long next
= 0;
481 queue_for_each_hw_ctx(q
, hctx
, i
)
482 blk_mq_hw_ctx_check_timeout(hctx
, &next
, &next_set
);
485 mod_timer(&q
->timeout
, round_jiffies_up(next
));
489 * Reverse check our software queue for entries that we could potentially
490 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
491 * too much time checking for merges.
493 static bool blk_mq_attempt_merge(struct request_queue
*q
,
494 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
499 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
505 if (!blk_rq_merge_ok(rq
, bio
))
508 el_ret
= blk_try_merge(rq
, bio
);
509 if (el_ret
== ELEVATOR_BACK_MERGE
) {
510 if (bio_attempt_back_merge(q
, rq
, bio
)) {
515 } else if (el_ret
== ELEVATOR_FRONT_MERGE
) {
516 if (bio_attempt_front_merge(q
, rq
, bio
)) {
527 void blk_mq_add_timer(struct request
*rq
)
529 __blk_add_timer(rq
, NULL
);
533 * Run this hardware queue, pulling any software queues mapped to it in.
534 * Note that this function currently has various problems around ordering
535 * of IO. In particular, we'd like FIFO behaviour on handling existing
536 * items on the hctx->dispatch list. Ignore that for now.
538 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
540 struct request_queue
*q
= hctx
->queue
;
541 struct blk_mq_ctx
*ctx
;
546 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
));
548 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)))
554 * Touch any software queue that has pending entries.
556 for_each_set_bit(bit
, hctx
->ctx_map
, hctx
->nr_ctx
) {
557 clear_bit(bit
, hctx
->ctx_map
);
558 ctx
= hctx
->ctxs
[bit
];
559 BUG_ON(bit
!= ctx
->index_hw
);
561 spin_lock(&ctx
->lock
);
562 list_splice_tail_init(&ctx
->rq_list
, &rq_list
);
563 spin_unlock(&ctx
->lock
);
567 * If we have previous entries on our dispatch list, grab them
568 * and stuff them at the front for more fair dispatch.
570 if (!list_empty_careful(&hctx
->dispatch
)) {
571 spin_lock(&hctx
->lock
);
572 if (!list_empty(&hctx
->dispatch
))
573 list_splice_init(&hctx
->dispatch
, &rq_list
);
574 spin_unlock(&hctx
->lock
);
578 * Delete and return all entries from our dispatch list
583 * Now process all the entries, sending them to the driver.
585 while (!list_empty(&rq_list
)) {
588 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
589 list_del_init(&rq
->queuelist
);
591 blk_mq_start_request(rq
, list_empty(&rq_list
));
593 ret
= q
->mq_ops
->queue_rq(hctx
, rq
);
595 case BLK_MQ_RQ_QUEUE_OK
:
598 case BLK_MQ_RQ_QUEUE_BUSY
:
600 * FIXME: we should have a mechanism to stop the queue
601 * like blk_stop_queue, otherwise we will waste cpu
604 list_add(&rq
->queuelist
, &rq_list
);
605 blk_mq_requeue_request(rq
);
608 pr_err("blk-mq: bad return on queue: %d\n", ret
);
609 case BLK_MQ_RQ_QUEUE_ERROR
:
611 blk_mq_end_io(rq
, rq
->errors
);
615 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
620 hctx
->dispatched
[0]++;
621 else if (queued
< (1 << (BLK_MQ_MAX_DISPATCH_ORDER
- 1)))
622 hctx
->dispatched
[ilog2(queued
) + 1]++;
625 * Any items that need requeuing? Stuff them into hctx->dispatch,
626 * that is where we will continue on next queue run.
628 if (!list_empty(&rq_list
)) {
629 spin_lock(&hctx
->lock
);
630 list_splice(&rq_list
, &hctx
->dispatch
);
631 spin_unlock(&hctx
->lock
);
635 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
637 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)))
640 if (!async
&& cpumask_test_cpu(smp_processor_id(), hctx
->cpumask
))
641 __blk_mq_run_hw_queue(hctx
);
642 else if (hctx
->queue
->nr_hw_queues
== 1)
643 kblockd_schedule_delayed_work(&hctx
->run_work
, 0);
648 * It'd be great if the workqueue API had a way to pass
649 * in a mask and had some smarts for more clever placement
650 * than the first CPU. Or we could round-robin here. For now,
651 * just queue on the first CPU.
653 cpu
= cpumask_first(hctx
->cpumask
);
654 kblockd_schedule_delayed_work_on(cpu
, &hctx
->run_work
, 0);
658 void blk_mq_run_queues(struct request_queue
*q
, bool async
)
660 struct blk_mq_hw_ctx
*hctx
;
663 queue_for_each_hw_ctx(q
, hctx
, i
) {
664 if ((!blk_mq_hctx_has_pending(hctx
) &&
665 list_empty_careful(&hctx
->dispatch
)) ||
666 test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
670 blk_mq_run_hw_queue(hctx
, async
);
674 EXPORT_SYMBOL(blk_mq_run_queues
);
676 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
678 cancel_delayed_work(&hctx
->run_work
);
679 cancel_delayed_work(&hctx
->delay_work
);
680 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
682 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
684 void blk_mq_stop_hw_queues(struct request_queue
*q
)
686 struct blk_mq_hw_ctx
*hctx
;
689 queue_for_each_hw_ctx(q
, hctx
, i
)
690 blk_mq_stop_hw_queue(hctx
);
692 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
694 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
696 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
699 __blk_mq_run_hw_queue(hctx
);
702 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
704 void blk_mq_start_hw_queues(struct request_queue
*q
)
706 struct blk_mq_hw_ctx
*hctx
;
709 queue_for_each_hw_ctx(q
, hctx
, i
)
710 blk_mq_start_hw_queue(hctx
);
712 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
715 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
717 struct blk_mq_hw_ctx
*hctx
;
720 queue_for_each_hw_ctx(q
, hctx
, i
) {
721 if (!test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
724 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
726 blk_mq_run_hw_queue(hctx
, async
);
730 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
732 static void blk_mq_run_work_fn(struct work_struct
*work
)
734 struct blk_mq_hw_ctx
*hctx
;
736 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
738 __blk_mq_run_hw_queue(hctx
);
741 static void blk_mq_delay_work_fn(struct work_struct
*work
)
743 struct blk_mq_hw_ctx
*hctx
;
745 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
747 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
748 __blk_mq_run_hw_queue(hctx
);
751 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
753 unsigned long tmo
= msecs_to_jiffies(msecs
);
755 if (hctx
->queue
->nr_hw_queues
== 1)
756 kblockd_schedule_delayed_work(&hctx
->delay_work
, tmo
);
761 * It'd be great if the workqueue API had a way to pass
762 * in a mask and had some smarts for more clever placement
763 * than the first CPU. Or we could round-robin here. For now,
764 * just queue on the first CPU.
766 cpu
= cpumask_first(hctx
->cpumask
);
767 kblockd_schedule_delayed_work_on(cpu
, &hctx
->delay_work
, tmo
);
770 EXPORT_SYMBOL(blk_mq_delay_queue
);
772 static void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
,
773 struct request
*rq
, bool at_head
)
775 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
777 trace_block_rq_insert(hctx
->queue
, rq
);
780 list_add(&rq
->queuelist
, &ctx
->rq_list
);
782 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
783 blk_mq_hctx_mark_pending(hctx
, ctx
);
786 * We do this early, to ensure we are on the right CPU.
788 blk_mq_add_timer(rq
);
791 void blk_mq_insert_request(struct request
*rq
, bool at_head
, bool run_queue
,
794 struct request_queue
*q
= rq
->q
;
795 struct blk_mq_hw_ctx
*hctx
;
796 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
, *current_ctx
;
798 current_ctx
= blk_mq_get_ctx(q
);
799 if (!cpu_online(ctx
->cpu
))
800 rq
->mq_ctx
= ctx
= current_ctx
;
802 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
804 if (rq
->cmd_flags
& (REQ_FLUSH
| REQ_FUA
) &&
805 !(rq
->cmd_flags
& (REQ_FLUSH_SEQ
))) {
806 blk_insert_flush(rq
);
808 spin_lock(&ctx
->lock
);
809 __blk_mq_insert_request(hctx
, rq
, at_head
);
810 spin_unlock(&ctx
->lock
);
814 blk_mq_run_hw_queue(hctx
, async
);
816 blk_mq_put_ctx(current_ctx
);
819 static void blk_mq_insert_requests(struct request_queue
*q
,
820 struct blk_mq_ctx
*ctx
,
821 struct list_head
*list
,
826 struct blk_mq_hw_ctx
*hctx
;
827 struct blk_mq_ctx
*current_ctx
;
829 trace_block_unplug(q
, depth
, !from_schedule
);
831 current_ctx
= blk_mq_get_ctx(q
);
833 if (!cpu_online(ctx
->cpu
))
835 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
838 * preemption doesn't flush plug list, so it's possible ctx->cpu is
841 spin_lock(&ctx
->lock
);
842 while (!list_empty(list
)) {
845 rq
= list_first_entry(list
, struct request
, queuelist
);
846 list_del_init(&rq
->queuelist
);
848 __blk_mq_insert_request(hctx
, rq
, false);
850 spin_unlock(&ctx
->lock
);
852 blk_mq_run_hw_queue(hctx
, from_schedule
);
853 blk_mq_put_ctx(current_ctx
);
856 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
858 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
859 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
861 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
862 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
863 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
866 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
868 struct blk_mq_ctx
*this_ctx
;
869 struct request_queue
*this_q
;
875 list_splice_init(&plug
->mq_list
, &list
);
877 list_sort(NULL
, &list
, plug_ctx_cmp
);
883 while (!list_empty(&list
)) {
884 rq
= list_entry_rq(list
.next
);
885 list_del_init(&rq
->queuelist
);
887 if (rq
->mq_ctx
!= this_ctx
) {
889 blk_mq_insert_requests(this_q
, this_ctx
,
894 this_ctx
= rq
->mq_ctx
;
900 list_add_tail(&rq
->queuelist
, &ctx_list
);
904 * If 'this_ctx' is set, we know we have entries to complete
905 * on 'ctx_list'. Do those.
908 blk_mq_insert_requests(this_q
, this_ctx
, &ctx_list
, depth
,
913 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
915 init_request_from_bio(rq
, bio
);
916 blk_account_io_start(rq
, 1);
919 static void blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
921 struct blk_mq_hw_ctx
*hctx
;
922 struct blk_mq_ctx
*ctx
;
923 const int is_sync
= rw_is_sync(bio
->bi_rw
);
924 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
925 int rw
= bio_data_dir(bio
);
927 unsigned int use_plug
, request_count
= 0;
930 * If we have multiple hardware queues, just go directly to
931 * one of those for sync IO.
933 use_plug
= !is_flush_fua
&& ((q
->nr_hw_queues
== 1) || !is_sync
);
935 blk_queue_bounce(q
, &bio
);
937 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
938 bio_endio(bio
, -EIO
);
942 if (use_plug
&& blk_attempt_plug_merge(q
, bio
, &request_count
))
945 if (blk_mq_queue_enter(q
)) {
946 bio_endio(bio
, -EIO
);
950 ctx
= blk_mq_get_ctx(q
);
951 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
955 trace_block_getrq(q
, bio
, rw
);
956 rq
= __blk_mq_alloc_request(hctx
, GFP_ATOMIC
, false);
958 blk_mq_rq_ctx_init(q
, ctx
, rq
, rw
);
961 trace_block_sleeprq(q
, bio
, rw
);
962 rq
= blk_mq_alloc_request_pinned(q
, rw
, __GFP_WAIT
|GFP_ATOMIC
,
965 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
970 if (unlikely(is_flush_fua
)) {
971 blk_mq_bio_to_request(rq
, bio
);
972 blk_insert_flush(rq
);
977 * A task plug currently exists. Since this is completely lockless,
978 * utilize that to temporarily store requests until the task is
979 * either done or scheduled away.
982 struct blk_plug
*plug
= current
->plug
;
985 blk_mq_bio_to_request(rq
, bio
);
986 if (list_empty(&plug
->mq_list
))
988 else if (request_count
>= BLK_MAX_REQUEST_COUNT
) {
989 blk_flush_plug_list(plug
, false);
992 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
998 spin_lock(&ctx
->lock
);
1000 if ((hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1001 blk_mq_attempt_merge(q
, ctx
, bio
))
1002 __blk_mq_free_request(hctx
, ctx
, rq
);
1004 blk_mq_bio_to_request(rq
, bio
);
1005 __blk_mq_insert_request(hctx
, rq
, false);
1008 spin_unlock(&ctx
->lock
);
1011 * For a SYNC request, send it to the hardware immediately. For an
1012 * ASYNC request, just ensure that we run it later on. The latter
1013 * allows for merging opportunities and more efficient dispatching.
1016 blk_mq_run_hw_queue(hctx
, !is_sync
|| is_flush_fua
);
1017 blk_mq_put_ctx(ctx
);
1021 * Default mapping to a software queue, since we use one per CPU.
1023 struct blk_mq_hw_ctx
*blk_mq_map_queue(struct request_queue
*q
, const int cpu
)
1025 return q
->queue_hw_ctx
[q
->mq_map
[cpu
]];
1027 EXPORT_SYMBOL(blk_mq_map_queue
);
1029 struct blk_mq_hw_ctx
*blk_mq_alloc_single_hw_queue(struct blk_mq_tag_set
*set
,
1030 unsigned int hctx_index
)
1032 return kmalloc_node(sizeof(struct blk_mq_hw_ctx
),
1033 GFP_KERNEL
| __GFP_ZERO
, set
->numa_node
);
1035 EXPORT_SYMBOL(blk_mq_alloc_single_hw_queue
);
1037 void blk_mq_free_single_hw_queue(struct blk_mq_hw_ctx
*hctx
,
1038 unsigned int hctx_index
)
1042 EXPORT_SYMBOL(blk_mq_free_single_hw_queue
);
1044 static void blk_mq_hctx_notify(void *data
, unsigned long action
,
1047 struct blk_mq_hw_ctx
*hctx
= data
;
1048 struct request_queue
*q
= hctx
->queue
;
1049 struct blk_mq_ctx
*ctx
;
1052 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
1056 * Move ctx entries to new CPU, if this one is going away.
1058 ctx
= __blk_mq_get_ctx(q
, cpu
);
1060 spin_lock(&ctx
->lock
);
1061 if (!list_empty(&ctx
->rq_list
)) {
1062 list_splice_init(&ctx
->rq_list
, &tmp
);
1063 clear_bit(ctx
->index_hw
, hctx
->ctx_map
);
1065 spin_unlock(&ctx
->lock
);
1067 if (list_empty(&tmp
))
1070 ctx
= blk_mq_get_ctx(q
);
1071 spin_lock(&ctx
->lock
);
1073 while (!list_empty(&tmp
)) {
1076 rq
= list_first_entry(&tmp
, struct request
, queuelist
);
1078 list_move_tail(&rq
->queuelist
, &ctx
->rq_list
);
1081 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1082 blk_mq_hctx_mark_pending(hctx
, ctx
);
1084 spin_unlock(&ctx
->lock
);
1086 blk_mq_run_hw_queue(hctx
, true);
1087 blk_mq_put_ctx(ctx
);
1090 static void blk_mq_free_rq_map(struct blk_mq_tag_set
*set
,
1091 struct blk_mq_tags
*tags
, unsigned int hctx_idx
)
1095 if (tags
->rqs
&& set
->ops
->exit_request
) {
1098 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1101 set
->ops
->exit_request(set
->driver_data
, tags
->rqs
[i
],
1106 while (!list_empty(&tags
->page_list
)) {
1107 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1108 list_del_init(&page
->lru
);
1109 __free_pages(page
, page
->private);
1114 blk_mq_free_tags(tags
);
1117 static size_t order_to_size(unsigned int order
)
1119 size_t ret
= PAGE_SIZE
;
1127 static struct blk_mq_tags
*blk_mq_init_rq_map(struct blk_mq_tag_set
*set
,
1128 unsigned int hctx_idx
)
1130 struct blk_mq_tags
*tags
;
1131 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1132 size_t rq_size
, left
;
1134 tags
= blk_mq_init_tags(set
->queue_depth
, set
->reserved_tags
,
1139 INIT_LIST_HEAD(&tags
->page_list
);
1141 tags
->rqs
= kmalloc_node(set
->queue_depth
* sizeof(struct request
*),
1142 GFP_KERNEL
, set
->numa_node
);
1144 blk_mq_free_tags(tags
);
1149 * rq_size is the size of the request plus driver payload, rounded
1150 * to the cacheline size
1152 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1154 left
= rq_size
* set
->queue_depth
;
1156 for (i
= 0; i
< set
->queue_depth
; ) {
1157 int this_order
= max_order
;
1162 while (left
< order_to_size(this_order
- 1) && this_order
)
1166 page
= alloc_pages_node(set
->numa_node
, GFP_KERNEL
,
1172 if (order_to_size(this_order
) < rq_size
)
1179 page
->private = this_order
;
1180 list_add_tail(&page
->lru
, &tags
->page_list
);
1182 p
= page_address(page
);
1183 entries_per_page
= order_to_size(this_order
) / rq_size
;
1184 to_do
= min(entries_per_page
, set
->queue_depth
- i
);
1185 left
-= to_do
* rq_size
;
1186 for (j
= 0; j
< to_do
; j
++) {
1188 if (set
->ops
->init_request
) {
1189 if (set
->ops
->init_request(set
->driver_data
,
1190 tags
->rqs
[i
], hctx_idx
, i
,
1203 pr_warn("%s: failed to allocate requests\n", __func__
);
1204 blk_mq_free_rq_map(set
, tags
, hctx_idx
);
1208 static int blk_mq_init_hw_queues(struct request_queue
*q
,
1209 struct blk_mq_tag_set
*set
)
1211 struct blk_mq_hw_ctx
*hctx
;
1215 * Initialize hardware queues
1217 queue_for_each_hw_ctx(q
, hctx
, i
) {
1218 unsigned int num_maps
;
1221 node
= hctx
->numa_node
;
1222 if (node
== NUMA_NO_NODE
)
1223 node
= hctx
->numa_node
= set
->numa_node
;
1225 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1226 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1227 spin_lock_init(&hctx
->lock
);
1228 INIT_LIST_HEAD(&hctx
->dispatch
);
1230 hctx
->queue_num
= i
;
1231 hctx
->flags
= set
->flags
;
1232 hctx
->cmd_size
= set
->cmd_size
;
1234 blk_mq_init_cpu_notifier(&hctx
->cpu_notifier
,
1235 blk_mq_hctx_notify
, hctx
);
1236 blk_mq_register_cpu_notifier(&hctx
->cpu_notifier
);
1238 hctx
->tags
= set
->tags
[i
];
1241 * Allocate space for all possible cpus to avoid allocation in
1244 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1249 num_maps
= ALIGN(nr_cpu_ids
, BITS_PER_LONG
) / BITS_PER_LONG
;
1250 hctx
->ctx_map
= kzalloc_node(num_maps
* sizeof(unsigned long),
1255 hctx
->nr_ctx_map
= num_maps
;
1258 if (set
->ops
->init_hctx
&&
1259 set
->ops
->init_hctx(hctx
, set
->driver_data
, i
))
1263 if (i
== q
->nr_hw_queues
)
1269 queue_for_each_hw_ctx(q
, hctx
, j
) {
1273 if (set
->ops
->exit_hctx
)
1274 set
->ops
->exit_hctx(hctx
, j
);
1276 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1283 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1284 unsigned int nr_hw_queues
)
1288 for_each_possible_cpu(i
) {
1289 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1290 struct blk_mq_hw_ctx
*hctx
;
1292 memset(__ctx
, 0, sizeof(*__ctx
));
1294 spin_lock_init(&__ctx
->lock
);
1295 INIT_LIST_HEAD(&__ctx
->rq_list
);
1298 /* If the cpu isn't online, the cpu is mapped to first hctx */
1302 hctx
= q
->mq_ops
->map_queue(q
, i
);
1303 cpumask_set_cpu(i
, hctx
->cpumask
);
1307 * Set local node, IFF we have more than one hw queue. If
1308 * not, we remain on the home node of the device
1310 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1311 hctx
->numa_node
= cpu_to_node(i
);
1315 static void blk_mq_map_swqueue(struct request_queue
*q
)
1318 struct blk_mq_hw_ctx
*hctx
;
1319 struct blk_mq_ctx
*ctx
;
1321 queue_for_each_hw_ctx(q
, hctx
, i
) {
1322 cpumask_clear(hctx
->cpumask
);
1327 * Map software to hardware queues
1329 queue_for_each_ctx(q
, ctx
, i
) {
1330 /* If the cpu isn't online, the cpu is mapped to first hctx */
1334 hctx
= q
->mq_ops
->map_queue(q
, i
);
1335 cpumask_set_cpu(i
, hctx
->cpumask
);
1336 ctx
->index_hw
= hctx
->nr_ctx
;
1337 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
1341 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
1343 struct blk_mq_hw_ctx
**hctxs
;
1344 struct blk_mq_ctx
*ctx
;
1345 struct request_queue
*q
;
1348 ctx
= alloc_percpu(struct blk_mq_ctx
);
1350 return ERR_PTR(-ENOMEM
);
1352 hctxs
= kmalloc_node(set
->nr_hw_queues
* sizeof(*hctxs
), GFP_KERNEL
,
1358 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1359 hctxs
[i
] = set
->ops
->alloc_hctx(set
, i
);
1363 if (!zalloc_cpumask_var(&hctxs
[i
]->cpumask
, GFP_KERNEL
))
1366 hctxs
[i
]->numa_node
= NUMA_NO_NODE
;
1367 hctxs
[i
]->queue_num
= i
;
1370 q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
1374 q
->mq_map
= blk_mq_make_queue_map(set
);
1378 setup_timer(&q
->timeout
, blk_mq_rq_timer
, (unsigned long) q
);
1379 blk_queue_rq_timeout(q
, 30000);
1381 q
->nr_queues
= nr_cpu_ids
;
1382 q
->nr_hw_queues
= set
->nr_hw_queues
;
1385 q
->queue_hw_ctx
= hctxs
;
1387 q
->mq_ops
= set
->ops
;
1388 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
1390 q
->sg_reserved_size
= INT_MAX
;
1392 blk_queue_make_request(q
, blk_mq_make_request
);
1393 blk_queue_rq_timed_out(q
, set
->ops
->timeout
);
1395 blk_queue_rq_timeout(q
, set
->timeout
);
1397 if (set
->ops
->complete
)
1398 blk_queue_softirq_done(q
, set
->ops
->complete
);
1400 blk_mq_init_flush(q
);
1401 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
1403 q
->flush_rq
= kzalloc(round_up(sizeof(struct request
) +
1404 set
->cmd_size
, cache_line_size()),
1409 if (blk_mq_init_hw_queues(q
, set
))
1412 blk_mq_map_swqueue(q
);
1414 mutex_lock(&all_q_mutex
);
1415 list_add_tail(&q
->all_q_node
, &all_q_list
);
1416 mutex_unlock(&all_q_mutex
);
1425 blk_cleanup_queue(q
);
1427 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1430 free_cpumask_var(hctxs
[i
]->cpumask
);
1431 set
->ops
->free_hctx(hctxs
[i
], i
);
1436 return ERR_PTR(-ENOMEM
);
1438 EXPORT_SYMBOL(blk_mq_init_queue
);
1440 void blk_mq_free_queue(struct request_queue
*q
)
1442 struct blk_mq_hw_ctx
*hctx
;
1445 queue_for_each_hw_ctx(q
, hctx
, i
) {
1446 kfree(hctx
->ctx_map
);
1448 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1449 if (q
->mq_ops
->exit_hctx
)
1450 q
->mq_ops
->exit_hctx(hctx
, i
);
1451 free_cpumask_var(hctx
->cpumask
);
1452 q
->mq_ops
->free_hctx(hctx
, i
);
1455 free_percpu(q
->queue_ctx
);
1456 kfree(q
->queue_hw_ctx
);
1459 q
->queue_ctx
= NULL
;
1460 q
->queue_hw_ctx
= NULL
;
1463 mutex_lock(&all_q_mutex
);
1464 list_del_init(&q
->all_q_node
);
1465 mutex_unlock(&all_q_mutex
);
1468 /* Basically redo blk_mq_init_queue with queue frozen */
1469 static void blk_mq_queue_reinit(struct request_queue
*q
)
1471 blk_mq_freeze_queue(q
);
1473 blk_mq_update_queue_map(q
->mq_map
, q
->nr_hw_queues
);
1476 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1477 * we should change hctx numa_node according to new topology (this
1478 * involves free and re-allocate memory, worthy doing?)
1481 blk_mq_map_swqueue(q
);
1483 blk_mq_unfreeze_queue(q
);
1486 static int blk_mq_queue_reinit_notify(struct notifier_block
*nb
,
1487 unsigned long action
, void *hcpu
)
1489 struct request_queue
*q
;
1492 * Before new mapping is established, hotadded cpu might already start
1493 * handling requests. This doesn't break anything as we map offline
1494 * CPUs to first hardware queue. We will re-init queue below to get
1497 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
&&
1498 action
!= CPU_ONLINE
&& action
!= CPU_ONLINE_FROZEN
)
1501 mutex_lock(&all_q_mutex
);
1502 list_for_each_entry(q
, &all_q_list
, all_q_node
)
1503 blk_mq_queue_reinit(q
);
1504 mutex_unlock(&all_q_mutex
);
1508 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
1512 if (!set
->nr_hw_queues
)
1514 if (!set
->queue_depth
|| set
->queue_depth
> BLK_MQ_MAX_DEPTH
)
1516 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
1519 if (!set
->nr_hw_queues
||
1520 !set
->ops
->queue_rq
|| !set
->ops
->map_queue
||
1521 !set
->ops
->alloc_hctx
|| !set
->ops
->free_hctx
)
1525 set
->tags
= kmalloc_node(set
->nr_hw_queues
* sizeof(struct blk_mq_tags
),
1526 GFP_KERNEL
, set
->numa_node
);
1530 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1531 set
->tags
[i
] = blk_mq_init_rq_map(set
, i
);
1540 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1544 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
1546 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
1550 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
1551 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1553 EXPORT_SYMBOL(blk_mq_free_tag_set
);
1555 void blk_mq_disable_hotplug(void)
1557 mutex_lock(&all_q_mutex
);
1560 void blk_mq_enable_hotplug(void)
1562 mutex_unlock(&all_q_mutex
);
1565 static int __init
blk_mq_init(void)
1569 /* Must be called after percpu_counter_hotcpu_callback() */
1570 hotcpu_notifier(blk_mq_queue_reinit_notify
, -10);
1574 subsys_initcall(blk_mq_init
);