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_rq(struct blk_mq_hw_ctx
*hctx
, gfp_t gfp
,
82 tag
= blk_mq_get_tag(hctx
->tags
, gfp
, reserved
);
83 if (tag
!= BLK_MQ_TAG_FAIL
) {
93 static int blk_mq_queue_enter(struct request_queue
*q
)
97 __percpu_counter_add(&q
->mq_usage_counter
, 1, 1000000);
99 /* we have problems to freeze the queue if it's initializing */
100 if (!blk_queue_bypass(q
) || !blk_queue_init_done(q
))
103 __percpu_counter_add(&q
->mq_usage_counter
, -1, 1000000);
105 spin_lock_irq(q
->queue_lock
);
106 ret
= wait_event_interruptible_lock_irq(q
->mq_freeze_wq
,
107 !blk_queue_bypass(q
) || blk_queue_dying(q
),
109 /* inc usage with lock hold to avoid freeze_queue runs here */
110 if (!ret
&& !blk_queue_dying(q
))
111 __percpu_counter_add(&q
->mq_usage_counter
, 1, 1000000);
112 else if (blk_queue_dying(q
))
114 spin_unlock_irq(q
->queue_lock
);
119 static void blk_mq_queue_exit(struct request_queue
*q
)
121 __percpu_counter_add(&q
->mq_usage_counter
, -1, 1000000);
124 static void __blk_mq_drain_queue(struct request_queue
*q
)
129 spin_lock_irq(q
->queue_lock
);
130 count
= percpu_counter_sum(&q
->mq_usage_counter
);
131 spin_unlock_irq(q
->queue_lock
);
135 blk_mq_run_queues(q
, false);
141 * Guarantee no request is in use, so we can change any data structure of
142 * the queue afterward.
144 static void blk_mq_freeze_queue(struct request_queue
*q
)
148 spin_lock_irq(q
->queue_lock
);
149 drain
= !q
->bypass_depth
++;
150 queue_flag_set(QUEUE_FLAG_BYPASS
, q
);
151 spin_unlock_irq(q
->queue_lock
);
154 __blk_mq_drain_queue(q
);
157 void blk_mq_drain_queue(struct request_queue
*q
)
159 __blk_mq_drain_queue(q
);
162 static void blk_mq_unfreeze_queue(struct request_queue
*q
)
166 spin_lock_irq(q
->queue_lock
);
167 if (!--q
->bypass_depth
) {
168 queue_flag_clear(QUEUE_FLAG_BYPASS
, q
);
171 WARN_ON_ONCE(q
->bypass_depth
< 0);
172 spin_unlock_irq(q
->queue_lock
);
174 wake_up_all(&q
->mq_freeze_wq
);
177 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
179 return blk_mq_has_free_tags(hctx
->tags
);
181 EXPORT_SYMBOL(blk_mq_can_queue
);
183 static void blk_mq_rq_ctx_init(struct request_queue
*q
, struct blk_mq_ctx
*ctx
,
184 struct request
*rq
, unsigned int rw_flags
)
186 if (blk_queue_io_stat(q
))
187 rw_flags
|= REQ_IO_STAT
;
190 rq
->cmd_flags
= rw_flags
;
191 rq
->start_time
= jiffies
;
192 set_start_time_ns(rq
);
193 ctx
->rq_dispatched
[rw_is_sync(rw_flags
)]++;
196 static struct request
*__blk_mq_alloc_request(struct blk_mq_hw_ctx
*hctx
,
197 gfp_t gfp
, bool reserved
)
199 return blk_mq_alloc_rq(hctx
, gfp
, reserved
);
202 static struct request
*blk_mq_alloc_request_pinned(struct request_queue
*q
,
209 struct blk_mq_ctx
*ctx
= blk_mq_get_ctx(q
);
210 struct blk_mq_hw_ctx
*hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
212 rq
= __blk_mq_alloc_request(hctx
, gfp
& ~__GFP_WAIT
, reserved
);
214 blk_mq_rq_ctx_init(q
, ctx
, rq
, rw
);
219 if (!(gfp
& __GFP_WAIT
))
222 __blk_mq_run_hw_queue(hctx
);
223 blk_mq_wait_for_tags(hctx
->tags
);
229 struct request
*blk_mq_alloc_request(struct request_queue
*q
, int rw
,
230 gfp_t gfp
, bool reserved
)
234 if (blk_mq_queue_enter(q
))
237 rq
= blk_mq_alloc_request_pinned(q
, rw
, gfp
, reserved
);
239 blk_mq_put_ctx(rq
->mq_ctx
);
243 struct request
*blk_mq_alloc_reserved_request(struct request_queue
*q
, int rw
,
248 if (blk_mq_queue_enter(q
))
251 rq
= blk_mq_alloc_request_pinned(q
, rw
, gfp
, true);
253 blk_mq_put_ctx(rq
->mq_ctx
);
256 EXPORT_SYMBOL(blk_mq_alloc_reserved_request
);
259 * Re-init and set pdu, if we have it
261 static void blk_mq_rq_init(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
263 blk_rq_init(hctx
->queue
, rq
);
266 rq
->special
= blk_mq_rq_to_pdu(rq
);
269 static void __blk_mq_free_request(struct blk_mq_hw_ctx
*hctx
,
270 struct blk_mq_ctx
*ctx
, struct request
*rq
)
272 const int tag
= rq
->tag
;
273 struct request_queue
*q
= rq
->q
;
275 blk_mq_rq_init(hctx
, rq
);
276 blk_mq_put_tag(hctx
->tags
, tag
);
278 blk_mq_queue_exit(q
);
281 void blk_mq_free_request(struct request
*rq
)
283 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
284 struct blk_mq_hw_ctx
*hctx
;
285 struct request_queue
*q
= rq
->q
;
287 ctx
->rq_completed
[rq_is_sync(rq
)]++;
289 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
290 __blk_mq_free_request(hctx
, ctx
, rq
);
293 static void blk_mq_bio_endio(struct request
*rq
, struct bio
*bio
, int error
)
296 clear_bit(BIO_UPTODATE
, &bio
->bi_flags
);
297 else if (!test_bit(BIO_UPTODATE
, &bio
->bi_flags
))
300 if (unlikely(rq
->cmd_flags
& REQ_QUIET
))
301 set_bit(BIO_QUIET
, &bio
->bi_flags
);
303 /* don't actually finish bio if it's part of flush sequence */
304 if (!(rq
->cmd_flags
& REQ_FLUSH_SEQ
))
305 bio_endio(bio
, error
);
308 void blk_mq_complete_request(struct request
*rq
, int error
)
310 struct bio
*bio
= rq
->bio
;
311 unsigned int bytes
= 0;
313 trace_block_rq_complete(rq
->q
, rq
);
316 struct bio
*next
= bio
->bi_next
;
319 bytes
+= bio
->bi_iter
.bi_size
;
320 blk_mq_bio_endio(rq
, bio
, error
);
324 blk_account_io_completion(rq
, bytes
);
326 blk_account_io_done(rq
);
329 rq
->end_io(rq
, error
);
331 blk_mq_free_request(rq
);
334 void __blk_mq_end_io(struct request
*rq
, int error
)
336 if (!blk_mark_rq_complete(rq
))
337 blk_mq_complete_request(rq
, error
);
340 static void blk_mq_end_io_remote(void *data
)
342 struct request
*rq
= data
;
344 __blk_mq_end_io(rq
, rq
->errors
);
348 * End IO on this request on a multiqueue enabled driver. We'll either do
349 * it directly inline, or punt to a local IPI handler on the matching
352 void blk_mq_end_io(struct request
*rq
, int error
)
354 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
357 if (!ctx
->ipi_redirect
)
358 return __blk_mq_end_io(rq
, error
);
361 if (cpu
!= ctx
->cpu
&& cpu_online(ctx
->cpu
)) {
363 rq
->csd
.func
= blk_mq_end_io_remote
;
366 __smp_call_function_single(ctx
->cpu
, &rq
->csd
, 0);
368 __blk_mq_end_io(rq
, error
);
372 EXPORT_SYMBOL(blk_mq_end_io
);
374 static void blk_mq_start_request(struct request
*rq
)
376 struct request_queue
*q
= rq
->q
;
378 trace_block_rq_issue(q
, rq
);
381 * Just mark start time and set the started bit. Due to memory
382 * ordering, we know we'll see the correct deadline as long as
383 * REQ_ATOMIC_STARTED is seen.
385 rq
->deadline
= jiffies
+ q
->rq_timeout
;
386 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
389 static void blk_mq_requeue_request(struct request
*rq
)
391 struct request_queue
*q
= rq
->q
;
393 trace_block_rq_requeue(q
, rq
);
394 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
397 struct blk_mq_timeout_data
{
398 struct blk_mq_hw_ctx
*hctx
;
400 unsigned int *next_set
;
403 static void blk_mq_timeout_check(void *__data
, unsigned long *free_tags
)
405 struct blk_mq_timeout_data
*data
= __data
;
406 struct blk_mq_hw_ctx
*hctx
= data
->hctx
;
409 /* It may not be in flight yet (this is where
410 * the REQ_ATOMIC_STARTED flag comes in). The requests are
411 * statically allocated, so we know it's always safe to access the
412 * memory associated with a bit offset into ->rqs[].
418 tag
= find_next_zero_bit(free_tags
, hctx
->queue_depth
, tag
);
419 if (tag
>= hctx
->queue_depth
)
422 rq
= hctx
->rqs
[tag
++];
424 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
427 blk_rq_check_expired(rq
, data
->next
, data
->next_set
);
431 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx
*hctx
,
433 unsigned int *next_set
)
435 struct blk_mq_timeout_data data
= {
438 .next_set
= next_set
,
442 * Ask the tagging code to iterate busy requests, so we can
443 * check them for timeout.
445 blk_mq_tag_busy_iter(hctx
->tags
, blk_mq_timeout_check
, &data
);
448 static void blk_mq_rq_timer(unsigned long data
)
450 struct request_queue
*q
= (struct request_queue
*) data
;
451 struct blk_mq_hw_ctx
*hctx
;
452 unsigned long next
= 0;
455 queue_for_each_hw_ctx(q
, hctx
, i
)
456 blk_mq_hw_ctx_check_timeout(hctx
, &next
, &next_set
);
459 mod_timer(&q
->timeout
, round_jiffies_up(next
));
463 * Reverse check our software queue for entries that we could potentially
464 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
465 * too much time checking for merges.
467 static bool blk_mq_attempt_merge(struct request_queue
*q
,
468 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
473 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
479 if (!blk_rq_merge_ok(rq
, bio
))
482 el_ret
= blk_try_merge(rq
, bio
);
483 if (el_ret
== ELEVATOR_BACK_MERGE
) {
484 if (bio_attempt_back_merge(q
, rq
, bio
)) {
489 } else if (el_ret
== ELEVATOR_FRONT_MERGE
) {
490 if (bio_attempt_front_merge(q
, rq
, bio
)) {
501 void blk_mq_add_timer(struct request
*rq
)
503 __blk_add_timer(rq
, NULL
);
507 * Run this hardware queue, pulling any software queues mapped to it in.
508 * Note that this function currently has various problems around ordering
509 * of IO. In particular, we'd like FIFO behaviour on handling existing
510 * items on the hctx->dispatch list. Ignore that for now.
512 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
514 struct request_queue
*q
= hctx
->queue
;
515 struct blk_mq_ctx
*ctx
;
520 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->flags
)))
526 * Touch any software queue that has pending entries.
528 for_each_set_bit(bit
, hctx
->ctx_map
, hctx
->nr_ctx
) {
529 clear_bit(bit
, hctx
->ctx_map
);
530 ctx
= hctx
->ctxs
[bit
];
531 BUG_ON(bit
!= ctx
->index_hw
);
533 spin_lock(&ctx
->lock
);
534 list_splice_tail_init(&ctx
->rq_list
, &rq_list
);
535 spin_unlock(&ctx
->lock
);
539 * If we have previous entries on our dispatch list, grab them
540 * and stuff them at the front for more fair dispatch.
542 if (!list_empty_careful(&hctx
->dispatch
)) {
543 spin_lock(&hctx
->lock
);
544 if (!list_empty(&hctx
->dispatch
))
545 list_splice_init(&hctx
->dispatch
, &rq_list
);
546 spin_unlock(&hctx
->lock
);
550 * Delete and return all entries from our dispatch list
555 * Now process all the entries, sending them to the driver.
557 while (!list_empty(&rq_list
)) {
560 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
561 list_del_init(&rq
->queuelist
);
562 blk_mq_start_request(rq
);
565 * Last request in the series. Flag it as such, this
566 * enables drivers to know when IO should be kicked off,
567 * if they don't do it on a per-request basis.
569 * Note: the flag isn't the only condition drivers
570 * should do kick off. If drive is busy, the last
571 * request might not have the bit set.
573 if (list_empty(&rq_list
))
574 rq
->cmd_flags
|= REQ_END
;
576 ret
= q
->mq_ops
->queue_rq(hctx
, rq
);
578 case BLK_MQ_RQ_QUEUE_OK
:
581 case BLK_MQ_RQ_QUEUE_BUSY
:
583 * FIXME: we should have a mechanism to stop the queue
584 * like blk_stop_queue, otherwise we will waste cpu
587 list_add(&rq
->queuelist
, &rq_list
);
588 blk_mq_requeue_request(rq
);
591 pr_err("blk-mq: bad return on queue: %d\n", ret
);
593 case BLK_MQ_RQ_QUEUE_ERROR
:
594 blk_mq_end_io(rq
, rq
->errors
);
598 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
603 hctx
->dispatched
[0]++;
604 else if (queued
< (1 << (BLK_MQ_MAX_DISPATCH_ORDER
- 1)))
605 hctx
->dispatched
[ilog2(queued
) + 1]++;
608 * Any items that need requeuing? Stuff them into hctx->dispatch,
609 * that is where we will continue on next queue run.
611 if (!list_empty(&rq_list
)) {
612 spin_lock(&hctx
->lock
);
613 list_splice(&rq_list
, &hctx
->dispatch
);
614 spin_unlock(&hctx
->lock
);
618 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
620 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->flags
)))
624 __blk_mq_run_hw_queue(hctx
);
626 struct request_queue
*q
= hctx
->queue
;
628 kblockd_schedule_delayed_work(q
, &hctx
->delayed_work
, 0);
632 void blk_mq_run_queues(struct request_queue
*q
, bool async
)
634 struct blk_mq_hw_ctx
*hctx
;
637 queue_for_each_hw_ctx(q
, hctx
, i
) {
638 if ((!blk_mq_hctx_has_pending(hctx
) &&
639 list_empty_careful(&hctx
->dispatch
)) ||
640 test_bit(BLK_MQ_S_STOPPED
, &hctx
->flags
))
643 blk_mq_run_hw_queue(hctx
, async
);
646 EXPORT_SYMBOL(blk_mq_run_queues
);
648 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
650 cancel_delayed_work(&hctx
->delayed_work
);
651 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
653 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
655 void blk_mq_stop_hw_queues(struct request_queue
*q
)
657 struct blk_mq_hw_ctx
*hctx
;
660 queue_for_each_hw_ctx(q
, hctx
, i
)
661 blk_mq_stop_hw_queue(hctx
);
663 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
665 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
667 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
668 __blk_mq_run_hw_queue(hctx
);
670 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
672 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
)
674 struct blk_mq_hw_ctx
*hctx
;
677 queue_for_each_hw_ctx(q
, hctx
, i
) {
678 if (!test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
681 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
682 blk_mq_run_hw_queue(hctx
, true);
685 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
687 static void blk_mq_work_fn(struct work_struct
*work
)
689 struct blk_mq_hw_ctx
*hctx
;
691 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delayed_work
.work
);
692 __blk_mq_run_hw_queue(hctx
);
695 static void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
,
698 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
700 trace_block_rq_insert(hctx
->queue
, rq
);
702 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
703 blk_mq_hctx_mark_pending(hctx
, ctx
);
706 * We do this early, to ensure we are on the right CPU.
708 blk_mq_add_timer(rq
);
711 void blk_mq_insert_request(struct request_queue
*q
, struct request
*rq
,
714 struct blk_mq_hw_ctx
*hctx
;
715 struct blk_mq_ctx
*ctx
, *current_ctx
;
718 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
720 if (rq
->cmd_flags
& (REQ_FLUSH
| REQ_FUA
)) {
721 blk_insert_flush(rq
);
723 current_ctx
= blk_mq_get_ctx(q
);
725 if (!cpu_online(ctx
->cpu
)) {
727 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
730 spin_lock(&ctx
->lock
);
731 __blk_mq_insert_request(hctx
, rq
);
732 spin_unlock(&ctx
->lock
);
734 blk_mq_put_ctx(current_ctx
);
738 __blk_mq_run_hw_queue(hctx
);
740 EXPORT_SYMBOL(blk_mq_insert_request
);
743 * This is a special version of blk_mq_insert_request to bypass FLUSH request
744 * check. Should only be used internally.
746 void blk_mq_run_request(struct request
*rq
, bool run_queue
, bool async
)
748 struct request_queue
*q
= rq
->q
;
749 struct blk_mq_hw_ctx
*hctx
;
750 struct blk_mq_ctx
*ctx
, *current_ctx
;
752 current_ctx
= blk_mq_get_ctx(q
);
755 if (!cpu_online(ctx
->cpu
)) {
759 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
761 /* ctx->cpu might be offline */
762 spin_lock(&ctx
->lock
);
763 __blk_mq_insert_request(hctx
, rq
);
764 spin_unlock(&ctx
->lock
);
766 blk_mq_put_ctx(current_ctx
);
769 blk_mq_run_hw_queue(hctx
, async
);
772 static void blk_mq_insert_requests(struct request_queue
*q
,
773 struct blk_mq_ctx
*ctx
,
774 struct list_head
*list
,
779 struct blk_mq_hw_ctx
*hctx
;
780 struct blk_mq_ctx
*current_ctx
;
782 trace_block_unplug(q
, depth
, !from_schedule
);
784 current_ctx
= blk_mq_get_ctx(q
);
786 if (!cpu_online(ctx
->cpu
))
788 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
791 * preemption doesn't flush plug list, so it's possible ctx->cpu is
794 spin_lock(&ctx
->lock
);
795 while (!list_empty(list
)) {
798 rq
= list_first_entry(list
, struct request
, queuelist
);
799 list_del_init(&rq
->queuelist
);
801 __blk_mq_insert_request(hctx
, rq
);
803 spin_unlock(&ctx
->lock
);
805 blk_mq_put_ctx(current_ctx
);
807 blk_mq_run_hw_queue(hctx
, from_schedule
);
810 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
812 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
813 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
815 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
816 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
817 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
820 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
822 struct blk_mq_ctx
*this_ctx
;
823 struct request_queue
*this_q
;
829 list_splice_init(&plug
->mq_list
, &list
);
831 list_sort(NULL
, &list
, plug_ctx_cmp
);
837 while (!list_empty(&list
)) {
838 rq
= list_entry_rq(list
.next
);
839 list_del_init(&rq
->queuelist
);
841 if (rq
->mq_ctx
!= this_ctx
) {
843 blk_mq_insert_requests(this_q
, this_ctx
,
848 this_ctx
= rq
->mq_ctx
;
854 list_add_tail(&rq
->queuelist
, &ctx_list
);
858 * If 'this_ctx' is set, we know we have entries to complete
859 * on 'ctx_list'. Do those.
862 blk_mq_insert_requests(this_q
, this_ctx
, &ctx_list
, depth
,
867 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
869 init_request_from_bio(rq
, bio
);
870 blk_account_io_start(rq
, 1);
873 static void blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
875 struct blk_mq_hw_ctx
*hctx
;
876 struct blk_mq_ctx
*ctx
;
877 const int is_sync
= rw_is_sync(bio
->bi_rw
);
878 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
879 int rw
= bio_data_dir(bio
);
881 unsigned int use_plug
, request_count
= 0;
884 * If we have multiple hardware queues, just go directly to
885 * one of those for sync IO.
887 use_plug
= !is_flush_fua
&& ((q
->nr_hw_queues
== 1) || !is_sync
);
889 blk_queue_bounce(q
, &bio
);
891 if (use_plug
&& blk_attempt_plug_merge(q
, bio
, &request_count
))
894 if (blk_mq_queue_enter(q
)) {
895 bio_endio(bio
, -EIO
);
899 ctx
= blk_mq_get_ctx(q
);
900 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
902 trace_block_getrq(q
, bio
, rw
);
903 rq
= __blk_mq_alloc_request(hctx
, GFP_ATOMIC
, false);
905 blk_mq_rq_ctx_init(q
, ctx
, rq
, rw
);
908 trace_block_sleeprq(q
, bio
, rw
);
909 rq
= blk_mq_alloc_request_pinned(q
, rw
, __GFP_WAIT
|GFP_ATOMIC
,
912 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
917 if (unlikely(is_flush_fua
)) {
918 blk_mq_bio_to_request(rq
, bio
);
920 blk_insert_flush(rq
);
925 * A task plug currently exists. Since this is completely lockless,
926 * utilize that to temporarily store requests until the task is
927 * either done or scheduled away.
930 struct blk_plug
*plug
= current
->plug
;
933 blk_mq_bio_to_request(rq
, bio
);
934 if (list_empty(&plug
->mq_list
))
936 else if (request_count
>= BLK_MAX_REQUEST_COUNT
) {
937 blk_flush_plug_list(plug
, false);
940 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
946 spin_lock(&ctx
->lock
);
948 if ((hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
949 blk_mq_attempt_merge(q
, ctx
, bio
))
950 __blk_mq_free_request(hctx
, ctx
, rq
);
952 blk_mq_bio_to_request(rq
, bio
);
953 __blk_mq_insert_request(hctx
, rq
);
956 spin_unlock(&ctx
->lock
);
960 * For a SYNC request, send it to the hardware immediately. For an
961 * ASYNC request, just ensure that we run it later on. The latter
962 * allows for merging opportunities and more efficient dispatching.
965 blk_mq_run_hw_queue(hctx
, !is_sync
|| is_flush_fua
);
969 * Default mapping to a software queue, since we use one per CPU.
971 struct blk_mq_hw_ctx
*blk_mq_map_queue(struct request_queue
*q
, const int cpu
)
973 return q
->queue_hw_ctx
[q
->mq_map
[cpu
]];
975 EXPORT_SYMBOL(blk_mq_map_queue
);
977 struct blk_mq_hw_ctx
*blk_mq_alloc_single_hw_queue(struct blk_mq_reg
*reg
,
978 unsigned int hctx_index
)
980 return kmalloc_node(sizeof(struct blk_mq_hw_ctx
),
981 GFP_KERNEL
| __GFP_ZERO
, reg
->numa_node
);
983 EXPORT_SYMBOL(blk_mq_alloc_single_hw_queue
);
985 void blk_mq_free_single_hw_queue(struct blk_mq_hw_ctx
*hctx
,
986 unsigned int hctx_index
)
990 EXPORT_SYMBOL(blk_mq_free_single_hw_queue
);
992 static void blk_mq_hctx_notify(void *data
, unsigned long action
,
995 struct blk_mq_hw_ctx
*hctx
= data
;
996 struct blk_mq_ctx
*ctx
;
999 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
1003 * Move ctx entries to new CPU, if this one is going away.
1005 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
1007 spin_lock(&ctx
->lock
);
1008 if (!list_empty(&ctx
->rq_list
)) {
1009 list_splice_init(&ctx
->rq_list
, &tmp
);
1010 clear_bit(ctx
->index_hw
, hctx
->ctx_map
);
1012 spin_unlock(&ctx
->lock
);
1014 if (list_empty(&tmp
))
1017 ctx
= blk_mq_get_ctx(hctx
->queue
);
1018 spin_lock(&ctx
->lock
);
1020 while (!list_empty(&tmp
)) {
1023 rq
= list_first_entry(&tmp
, struct request
, queuelist
);
1025 list_move_tail(&rq
->queuelist
, &ctx
->rq_list
);
1028 blk_mq_hctx_mark_pending(hctx
, ctx
);
1030 spin_unlock(&ctx
->lock
);
1031 blk_mq_put_ctx(ctx
);
1034 static void blk_mq_init_hw_commands(struct blk_mq_hw_ctx
*hctx
,
1035 void (*init
)(void *, struct blk_mq_hw_ctx
*,
1036 struct request
*, unsigned int),
1041 for (i
= 0; i
< hctx
->queue_depth
; i
++) {
1042 struct request
*rq
= hctx
->rqs
[i
];
1044 init(data
, hctx
, rq
, i
);
1048 void blk_mq_init_commands(struct request_queue
*q
,
1049 void (*init
)(void *, struct blk_mq_hw_ctx
*,
1050 struct request
*, unsigned int),
1053 struct blk_mq_hw_ctx
*hctx
;
1056 queue_for_each_hw_ctx(q
, hctx
, i
)
1057 blk_mq_init_hw_commands(hctx
, init
, data
);
1059 EXPORT_SYMBOL(blk_mq_init_commands
);
1061 static void blk_mq_free_rq_map(struct blk_mq_hw_ctx
*hctx
)
1065 while (!list_empty(&hctx
->page_list
)) {
1066 page
= list_first_entry(&hctx
->page_list
, struct page
, lru
);
1067 list_del_init(&page
->lru
);
1068 __free_pages(page
, page
->private);
1074 blk_mq_free_tags(hctx
->tags
);
1077 static size_t order_to_size(unsigned int order
)
1079 size_t ret
= PAGE_SIZE
;
1087 static int blk_mq_init_rq_map(struct blk_mq_hw_ctx
*hctx
,
1088 unsigned int reserved_tags
, int node
)
1090 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1091 size_t rq_size
, left
;
1093 INIT_LIST_HEAD(&hctx
->page_list
);
1095 hctx
->rqs
= kmalloc_node(hctx
->queue_depth
* sizeof(struct request
*),
1101 * rq_size is the size of the request plus driver payload, rounded
1102 * to the cacheline size
1104 rq_size
= round_up(sizeof(struct request
) + hctx
->cmd_size
,
1106 left
= rq_size
* hctx
->queue_depth
;
1108 for (i
= 0; i
< hctx
->queue_depth
;) {
1109 int this_order
= max_order
;
1114 while (left
< order_to_size(this_order
- 1) && this_order
)
1118 page
= alloc_pages_node(node
, GFP_KERNEL
, this_order
);
1123 if (order_to_size(this_order
) < rq_size
)
1130 page
->private = this_order
;
1131 list_add_tail(&page
->lru
, &hctx
->page_list
);
1133 p
= page_address(page
);
1134 entries_per_page
= order_to_size(this_order
) / rq_size
;
1135 to_do
= min(entries_per_page
, hctx
->queue_depth
- i
);
1136 left
-= to_do
* rq_size
;
1137 for (j
= 0; j
< to_do
; j
++) {
1139 blk_mq_rq_init(hctx
, hctx
->rqs
[i
]);
1145 if (i
< (reserved_tags
+ BLK_MQ_TAG_MIN
))
1147 else if (i
!= hctx
->queue_depth
) {
1148 hctx
->queue_depth
= i
;
1149 pr_warn("%s: queue depth set to %u because of low memory\n",
1153 hctx
->tags
= blk_mq_init_tags(hctx
->queue_depth
, reserved_tags
, node
);
1156 blk_mq_free_rq_map(hctx
);
1163 static int blk_mq_init_hw_queues(struct request_queue
*q
,
1164 struct blk_mq_reg
*reg
, void *driver_data
)
1166 struct blk_mq_hw_ctx
*hctx
;
1170 * Initialize hardware queues
1172 queue_for_each_hw_ctx(q
, hctx
, i
) {
1173 unsigned int num_maps
;
1176 node
= hctx
->numa_node
;
1177 if (node
== NUMA_NO_NODE
)
1178 node
= hctx
->numa_node
= reg
->numa_node
;
1180 INIT_DELAYED_WORK(&hctx
->delayed_work
, blk_mq_work_fn
);
1181 spin_lock_init(&hctx
->lock
);
1182 INIT_LIST_HEAD(&hctx
->dispatch
);
1184 hctx
->queue_num
= i
;
1185 hctx
->flags
= reg
->flags
;
1186 hctx
->queue_depth
= reg
->queue_depth
;
1187 hctx
->cmd_size
= reg
->cmd_size
;
1189 blk_mq_init_cpu_notifier(&hctx
->cpu_notifier
,
1190 blk_mq_hctx_notify
, hctx
);
1191 blk_mq_register_cpu_notifier(&hctx
->cpu_notifier
);
1193 if (blk_mq_init_rq_map(hctx
, reg
->reserved_tags
, node
))
1197 * Allocate space for all possible cpus to avoid allocation in
1200 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1205 num_maps
= ALIGN(nr_cpu_ids
, BITS_PER_LONG
) / BITS_PER_LONG
;
1206 hctx
->ctx_map
= kzalloc_node(num_maps
* sizeof(unsigned long),
1211 hctx
->nr_ctx_map
= num_maps
;
1214 if (reg
->ops
->init_hctx
&&
1215 reg
->ops
->init_hctx(hctx
, driver_data
, i
))
1219 if (i
== q
->nr_hw_queues
)
1225 queue_for_each_hw_ctx(q
, hctx
, j
) {
1229 if (reg
->ops
->exit_hctx
)
1230 reg
->ops
->exit_hctx(hctx
, j
);
1232 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1233 blk_mq_free_rq_map(hctx
);
1240 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1241 unsigned int nr_hw_queues
)
1245 for_each_possible_cpu(i
) {
1246 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1247 struct blk_mq_hw_ctx
*hctx
;
1249 memset(__ctx
, 0, sizeof(*__ctx
));
1251 spin_lock_init(&__ctx
->lock
);
1252 INIT_LIST_HEAD(&__ctx
->rq_list
);
1255 /* If the cpu isn't online, the cpu is mapped to first hctx */
1256 hctx
= q
->mq_ops
->map_queue(q
, i
);
1263 * Set local node, IFF we have more than one hw queue. If
1264 * not, we remain on the home node of the device
1266 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1267 hctx
->numa_node
= cpu_to_node(i
);
1271 static void blk_mq_map_swqueue(struct request_queue
*q
)
1274 struct blk_mq_hw_ctx
*hctx
;
1275 struct blk_mq_ctx
*ctx
;
1277 queue_for_each_hw_ctx(q
, hctx
, i
) {
1282 * Map software to hardware queues
1284 queue_for_each_ctx(q
, ctx
, i
) {
1285 /* If the cpu isn't online, the cpu is mapped to first hctx */
1286 hctx
= q
->mq_ops
->map_queue(q
, i
);
1287 ctx
->index_hw
= hctx
->nr_ctx
;
1288 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
1292 struct request_queue
*blk_mq_init_queue(struct blk_mq_reg
*reg
,
1295 struct blk_mq_hw_ctx
**hctxs
;
1296 struct blk_mq_ctx
*ctx
;
1297 struct request_queue
*q
;
1300 if (!reg
->nr_hw_queues
||
1301 !reg
->ops
->queue_rq
|| !reg
->ops
->map_queue
||
1302 !reg
->ops
->alloc_hctx
|| !reg
->ops
->free_hctx
)
1303 return ERR_PTR(-EINVAL
);
1305 if (!reg
->queue_depth
)
1306 reg
->queue_depth
= BLK_MQ_MAX_DEPTH
;
1307 else if (reg
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
1308 pr_err("blk-mq: queuedepth too large (%u)\n", reg
->queue_depth
);
1309 reg
->queue_depth
= BLK_MQ_MAX_DEPTH
;
1313 * Set aside a tag for flush requests. It will only be used while
1314 * another flush request is in progress but outside the driver.
1316 * TODO: only allocate if flushes are supported
1319 reg
->reserved_tags
++;
1321 if (reg
->queue_depth
< (reg
->reserved_tags
+ BLK_MQ_TAG_MIN
))
1322 return ERR_PTR(-EINVAL
);
1324 ctx
= alloc_percpu(struct blk_mq_ctx
);
1326 return ERR_PTR(-ENOMEM
);
1328 hctxs
= kmalloc_node(reg
->nr_hw_queues
* sizeof(*hctxs
), GFP_KERNEL
,
1334 for (i
= 0; i
< reg
->nr_hw_queues
; i
++) {
1335 hctxs
[i
] = reg
->ops
->alloc_hctx(reg
, i
);
1339 hctxs
[i
]->numa_node
= NUMA_NO_NODE
;
1340 hctxs
[i
]->queue_num
= i
;
1343 q
= blk_alloc_queue_node(GFP_KERNEL
, reg
->numa_node
);
1347 q
->mq_map
= blk_mq_make_queue_map(reg
);
1351 setup_timer(&q
->timeout
, blk_mq_rq_timer
, (unsigned long) q
);
1352 blk_queue_rq_timeout(q
, 30000);
1354 q
->nr_queues
= nr_cpu_ids
;
1355 q
->nr_hw_queues
= reg
->nr_hw_queues
;
1358 q
->queue_hw_ctx
= hctxs
;
1360 q
->mq_ops
= reg
->ops
;
1361 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
1363 blk_queue_make_request(q
, blk_mq_make_request
);
1364 blk_queue_rq_timed_out(q
, reg
->ops
->timeout
);
1366 blk_queue_rq_timeout(q
, reg
->timeout
);
1368 blk_mq_init_flush(q
);
1369 blk_mq_init_cpu_queues(q
, reg
->nr_hw_queues
);
1371 if (blk_mq_init_hw_queues(q
, reg
, driver_data
))
1374 blk_mq_map_swqueue(q
);
1376 mutex_lock(&all_q_mutex
);
1377 list_add_tail(&q
->all_q_node
, &all_q_list
);
1378 mutex_unlock(&all_q_mutex
);
1384 blk_cleanup_queue(q
);
1386 for (i
= 0; i
< reg
->nr_hw_queues
; i
++) {
1389 reg
->ops
->free_hctx(hctxs
[i
], i
);
1394 return ERR_PTR(-ENOMEM
);
1396 EXPORT_SYMBOL(blk_mq_init_queue
);
1398 void blk_mq_free_queue(struct request_queue
*q
)
1400 struct blk_mq_hw_ctx
*hctx
;
1403 queue_for_each_hw_ctx(q
, hctx
, i
) {
1404 kfree(hctx
->ctx_map
);
1406 blk_mq_free_rq_map(hctx
);
1407 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1408 if (q
->mq_ops
->exit_hctx
)
1409 q
->mq_ops
->exit_hctx(hctx
, i
);
1410 q
->mq_ops
->free_hctx(hctx
, i
);
1413 free_percpu(q
->queue_ctx
);
1414 kfree(q
->queue_hw_ctx
);
1417 q
->queue_ctx
= NULL
;
1418 q
->queue_hw_ctx
= NULL
;
1421 mutex_lock(&all_q_mutex
);
1422 list_del_init(&q
->all_q_node
);
1423 mutex_unlock(&all_q_mutex
);
1426 /* Basically redo blk_mq_init_queue with queue frozen */
1427 static void blk_mq_queue_reinit(struct request_queue
*q
)
1429 blk_mq_freeze_queue(q
);
1431 blk_mq_update_queue_map(q
->mq_map
, q
->nr_hw_queues
);
1434 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1435 * we should change hctx numa_node according to new topology (this
1436 * involves free and re-allocate memory, worthy doing?)
1439 blk_mq_map_swqueue(q
);
1441 blk_mq_unfreeze_queue(q
);
1444 static int blk_mq_queue_reinit_notify(struct notifier_block
*nb
,
1445 unsigned long action
, void *hcpu
)
1447 struct request_queue
*q
;
1450 * Before new mapping is established, hotadded cpu might already start
1451 * handling requests. This doesn't break anything as we map offline
1452 * CPUs to first hardware queue. We will re-init queue below to get
1455 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
&&
1456 action
!= CPU_ONLINE
&& action
!= CPU_ONLINE_FROZEN
)
1459 mutex_lock(&all_q_mutex
);
1460 list_for_each_entry(q
, &all_q_list
, all_q_node
)
1461 blk_mq_queue_reinit(q
);
1462 mutex_unlock(&all_q_mutex
);
1466 static int __init
blk_mq_init(void)
1470 /* Must be called after percpu_counter_hotcpu_callback() */
1471 hotcpu_notifier(blk_mq_queue_reinit_notify
, -10);
1475 subsys_initcall(blk_mq_init
);