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
) {
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_pinned(struct request_queue
*q
,
203 struct blk_mq_ctx
*ctx
= blk_mq_get_ctx(q
);
204 struct blk_mq_hw_ctx
*hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
206 rq
= __blk_mq_alloc_request(hctx
, gfp
& ~__GFP_WAIT
, reserved
);
208 blk_mq_rq_ctx_init(q
, ctx
, rq
, rw
);
212 if (gfp
& __GFP_WAIT
) {
213 __blk_mq_run_hw_queue(hctx
);
220 blk_mq_wait_for_tags(hctx
->tags
);
226 struct request
*blk_mq_alloc_request(struct request_queue
*q
, int rw
, gfp_t gfp
)
230 if (blk_mq_queue_enter(q
))
233 rq
= blk_mq_alloc_request_pinned(q
, rw
, gfp
, false);
235 blk_mq_put_ctx(rq
->mq_ctx
);
239 struct request
*blk_mq_alloc_reserved_request(struct request_queue
*q
, int rw
,
244 if (blk_mq_queue_enter(q
))
247 rq
= blk_mq_alloc_request_pinned(q
, rw
, gfp
, true);
249 blk_mq_put_ctx(rq
->mq_ctx
);
252 EXPORT_SYMBOL(blk_mq_alloc_reserved_request
);
255 * Re-init and set pdu, if we have it
257 void blk_mq_rq_init(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
259 blk_rq_init(hctx
->queue
, rq
);
262 rq
->special
= blk_mq_rq_to_pdu(rq
);
265 static void __blk_mq_free_request(struct blk_mq_hw_ctx
*hctx
,
266 struct blk_mq_ctx
*ctx
, struct request
*rq
)
268 const int tag
= rq
->tag
;
269 struct request_queue
*q
= rq
->q
;
271 blk_mq_rq_init(hctx
, rq
);
272 blk_mq_put_tag(hctx
->tags
, tag
);
274 blk_mq_queue_exit(q
);
277 void blk_mq_free_request(struct request
*rq
)
279 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
280 struct blk_mq_hw_ctx
*hctx
;
281 struct request_queue
*q
= rq
->q
;
283 ctx
->rq_completed
[rq_is_sync(rq
)]++;
285 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
286 __blk_mq_free_request(hctx
, ctx
, rq
);
289 bool blk_mq_end_io_partial(struct request
*rq
, int error
, unsigned int nr_bytes
)
291 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
294 blk_account_io_done(rq
);
297 rq
->end_io(rq
, error
);
299 blk_mq_free_request(rq
);
302 EXPORT_SYMBOL(blk_mq_end_io_partial
);
304 static void __blk_mq_complete_request_remote(void *data
)
306 struct request
*rq
= data
;
308 rq
->q
->softirq_done_fn(rq
);
311 void __blk_mq_complete_request(struct request
*rq
)
313 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
316 if (!ctx
->ipi_redirect
) {
317 rq
->q
->softirq_done_fn(rq
);
322 if (cpu
!= ctx
->cpu
&& cpu_online(ctx
->cpu
)) {
323 rq
->csd
.func
= __blk_mq_complete_request_remote
;
326 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
328 rq
->q
->softirq_done_fn(rq
);
334 * blk_mq_complete_request - end I/O on a request
335 * @rq: the request being processed
338 * Ends all I/O on a request. It does not handle partial completions.
339 * The actual completion happens out-of-order, through a IPI handler.
341 void blk_mq_complete_request(struct request
*rq
)
343 if (unlikely(blk_should_fake_timeout(rq
->q
)))
345 if (!blk_mark_rq_complete(rq
))
346 __blk_mq_complete_request(rq
);
348 EXPORT_SYMBOL(blk_mq_complete_request
);
350 static void blk_mq_start_request(struct request
*rq
, bool last
)
352 struct request_queue
*q
= rq
->q
;
354 trace_block_rq_issue(q
, rq
);
357 * Just mark start time and set the started bit. Due to memory
358 * ordering, we know we'll see the correct deadline as long as
359 * REQ_ATOMIC_STARTED is seen.
361 rq
->deadline
= jiffies
+ q
->rq_timeout
;
362 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
364 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
366 * Make sure space for the drain appears. We know we can do
367 * this because max_hw_segments has been adjusted to be one
368 * fewer than the device can handle.
370 rq
->nr_phys_segments
++;
374 * Flag the last request in the series so that drivers know when IO
375 * should be kicked off, if they don't do it on a per-request basis.
377 * Note: the flag isn't the only condition drivers should do kick off.
378 * If drive is busy, the last request might not have the bit set.
381 rq
->cmd_flags
|= REQ_END
;
384 static void blk_mq_requeue_request(struct request
*rq
)
386 struct request_queue
*q
= rq
->q
;
388 trace_block_rq_requeue(q
, rq
);
389 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
391 rq
->cmd_flags
&= ~REQ_END
;
393 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
394 rq
->nr_phys_segments
--;
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 WARN_ON(!preempt_count());
522 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)))
528 * Touch any software queue that has pending entries.
530 for_each_set_bit(bit
, hctx
->ctx_map
, hctx
->nr_ctx
) {
531 clear_bit(bit
, hctx
->ctx_map
);
532 ctx
= hctx
->ctxs
[bit
];
533 BUG_ON(bit
!= ctx
->index_hw
);
535 spin_lock(&ctx
->lock
);
536 list_splice_tail_init(&ctx
->rq_list
, &rq_list
);
537 spin_unlock(&ctx
->lock
);
541 * If we have previous entries on our dispatch list, grab them
542 * and stuff them at the front for more fair dispatch.
544 if (!list_empty_careful(&hctx
->dispatch
)) {
545 spin_lock(&hctx
->lock
);
546 if (!list_empty(&hctx
->dispatch
))
547 list_splice_init(&hctx
->dispatch
, &rq_list
);
548 spin_unlock(&hctx
->lock
);
552 * Delete and return all entries from our dispatch list
557 * Now process all the entries, sending them to the driver.
559 while (!list_empty(&rq_list
)) {
562 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
563 list_del_init(&rq
->queuelist
);
565 blk_mq_start_request(rq
, list_empty(&rq_list
));
567 ret
= q
->mq_ops
->queue_rq(hctx
, rq
);
569 case BLK_MQ_RQ_QUEUE_OK
:
572 case BLK_MQ_RQ_QUEUE_BUSY
:
574 * FIXME: we should have a mechanism to stop the queue
575 * like blk_stop_queue, otherwise we will waste cpu
578 list_add(&rq
->queuelist
, &rq_list
);
579 blk_mq_requeue_request(rq
);
582 pr_err("blk-mq: bad return on queue: %d\n", ret
);
583 case BLK_MQ_RQ_QUEUE_ERROR
:
585 blk_mq_end_io(rq
, rq
->errors
);
589 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
594 hctx
->dispatched
[0]++;
595 else if (queued
< (1 << (BLK_MQ_MAX_DISPATCH_ORDER
- 1)))
596 hctx
->dispatched
[ilog2(queued
) + 1]++;
599 * Any items that need requeuing? Stuff them into hctx->dispatch,
600 * that is where we will continue on next queue run.
602 if (!list_empty(&rq_list
)) {
603 spin_lock(&hctx
->lock
);
604 list_splice(&rq_list
, &hctx
->dispatch
);
605 spin_unlock(&hctx
->lock
);
609 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
611 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)))
614 if (!async
&& cpumask_test_cpu(smp_processor_id(), hctx
->cpumask
))
615 __blk_mq_run_hw_queue(hctx
);
616 else if (hctx
->queue
->nr_hw_queues
== 1)
617 kblockd_schedule_delayed_work(&hctx
->delayed_work
, 0);
622 * It'd be great if the workqueue API had a way to pass
623 * in a mask and had some smarts for more clever placement
624 * than the first CPU. Or we could round-robin here. For now,
625 * just queue on the first CPU.
627 cpu
= cpumask_first(hctx
->cpumask
);
628 kblockd_schedule_delayed_work_on(cpu
, &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
->state
))
644 blk_mq_run_hw_queue(hctx
, async
);
648 EXPORT_SYMBOL(blk_mq_run_queues
);
650 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
652 cancel_delayed_work(&hctx
->delayed_work
);
653 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
655 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
657 void blk_mq_stop_hw_queues(struct request_queue
*q
)
659 struct blk_mq_hw_ctx
*hctx
;
662 queue_for_each_hw_ctx(q
, hctx
, i
)
663 blk_mq_stop_hw_queue(hctx
);
665 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
667 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
669 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
672 __blk_mq_run_hw_queue(hctx
);
675 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
677 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
)
679 struct blk_mq_hw_ctx
*hctx
;
682 queue_for_each_hw_ctx(q
, hctx
, i
) {
683 if (!test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
686 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
688 blk_mq_run_hw_queue(hctx
, true);
692 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
694 static void blk_mq_work_fn(struct work_struct
*work
)
696 struct blk_mq_hw_ctx
*hctx
;
698 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delayed_work
.work
);
701 __blk_mq_run_hw_queue(hctx
);
705 static void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
,
706 struct request
*rq
, bool at_head
)
708 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
710 trace_block_rq_insert(hctx
->queue
, rq
);
713 list_add(&rq
->queuelist
, &ctx
->rq_list
);
715 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
716 blk_mq_hctx_mark_pending(hctx
, ctx
);
719 * We do this early, to ensure we are on the right CPU.
721 blk_mq_add_timer(rq
);
724 void blk_mq_insert_request(struct request
*rq
, bool at_head
, bool run_queue
,
727 struct request_queue
*q
= rq
->q
;
728 struct blk_mq_hw_ctx
*hctx
;
729 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
, *current_ctx
;
731 current_ctx
= blk_mq_get_ctx(q
);
732 if (!cpu_online(ctx
->cpu
))
733 rq
->mq_ctx
= ctx
= current_ctx
;
735 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
737 if (rq
->cmd_flags
& (REQ_FLUSH
| REQ_FUA
) &&
738 !(rq
->cmd_flags
& (REQ_FLUSH_SEQ
))) {
739 blk_insert_flush(rq
);
741 spin_lock(&ctx
->lock
);
742 __blk_mq_insert_request(hctx
, rq
, at_head
);
743 spin_unlock(&ctx
->lock
);
747 blk_mq_run_hw_queue(hctx
, async
);
749 blk_mq_put_ctx(current_ctx
);
752 static void blk_mq_insert_requests(struct request_queue
*q
,
753 struct blk_mq_ctx
*ctx
,
754 struct list_head
*list
,
759 struct blk_mq_hw_ctx
*hctx
;
760 struct blk_mq_ctx
*current_ctx
;
762 trace_block_unplug(q
, depth
, !from_schedule
);
764 current_ctx
= blk_mq_get_ctx(q
);
766 if (!cpu_online(ctx
->cpu
))
768 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
771 * preemption doesn't flush plug list, so it's possible ctx->cpu is
774 spin_lock(&ctx
->lock
);
775 while (!list_empty(list
)) {
778 rq
= list_first_entry(list
, struct request
, queuelist
);
779 list_del_init(&rq
->queuelist
);
781 __blk_mq_insert_request(hctx
, rq
, false);
783 spin_unlock(&ctx
->lock
);
785 blk_mq_run_hw_queue(hctx
, from_schedule
);
786 blk_mq_put_ctx(current_ctx
);
789 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
791 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
792 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
794 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
795 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
796 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
799 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
801 struct blk_mq_ctx
*this_ctx
;
802 struct request_queue
*this_q
;
808 list_splice_init(&plug
->mq_list
, &list
);
810 list_sort(NULL
, &list
, plug_ctx_cmp
);
816 while (!list_empty(&list
)) {
817 rq
= list_entry_rq(list
.next
);
818 list_del_init(&rq
->queuelist
);
820 if (rq
->mq_ctx
!= this_ctx
) {
822 blk_mq_insert_requests(this_q
, this_ctx
,
827 this_ctx
= rq
->mq_ctx
;
833 list_add_tail(&rq
->queuelist
, &ctx_list
);
837 * If 'this_ctx' is set, we know we have entries to complete
838 * on 'ctx_list'. Do those.
841 blk_mq_insert_requests(this_q
, this_ctx
, &ctx_list
, depth
,
846 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
848 init_request_from_bio(rq
, bio
);
849 blk_account_io_start(rq
, 1);
852 static void blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
854 struct blk_mq_hw_ctx
*hctx
;
855 struct blk_mq_ctx
*ctx
;
856 const int is_sync
= rw_is_sync(bio
->bi_rw
);
857 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
858 int rw
= bio_data_dir(bio
);
860 unsigned int use_plug
, request_count
= 0;
863 * If we have multiple hardware queues, just go directly to
864 * one of those for sync IO.
866 use_plug
= !is_flush_fua
&& ((q
->nr_hw_queues
== 1) || !is_sync
);
868 blk_queue_bounce(q
, &bio
);
870 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
871 bio_endio(bio
, -EIO
);
875 if (use_plug
&& blk_attempt_plug_merge(q
, bio
, &request_count
))
878 if (blk_mq_queue_enter(q
)) {
879 bio_endio(bio
, -EIO
);
883 ctx
= blk_mq_get_ctx(q
);
884 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
888 trace_block_getrq(q
, bio
, rw
);
889 rq
= __blk_mq_alloc_request(hctx
, GFP_ATOMIC
, false);
891 blk_mq_rq_ctx_init(q
, ctx
, rq
, rw
);
894 trace_block_sleeprq(q
, bio
, rw
);
895 rq
= blk_mq_alloc_request_pinned(q
, rw
, __GFP_WAIT
|GFP_ATOMIC
,
898 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
903 if (unlikely(is_flush_fua
)) {
904 blk_mq_bio_to_request(rq
, bio
);
905 blk_insert_flush(rq
);
910 * A task plug currently exists. Since this is completely lockless,
911 * utilize that to temporarily store requests until the task is
912 * either done or scheduled away.
915 struct blk_plug
*plug
= current
->plug
;
918 blk_mq_bio_to_request(rq
, bio
);
919 if (list_empty(&plug
->mq_list
))
921 else if (request_count
>= BLK_MAX_REQUEST_COUNT
) {
922 blk_flush_plug_list(plug
, false);
925 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
931 spin_lock(&ctx
->lock
);
933 if ((hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
934 blk_mq_attempt_merge(q
, ctx
, bio
))
935 __blk_mq_free_request(hctx
, ctx
, rq
);
937 blk_mq_bio_to_request(rq
, bio
);
938 __blk_mq_insert_request(hctx
, rq
, false);
941 spin_unlock(&ctx
->lock
);
944 * For a SYNC request, send it to the hardware immediately. For an
945 * ASYNC request, just ensure that we run it later on. The latter
946 * allows for merging opportunities and more efficient dispatching.
949 blk_mq_run_hw_queue(hctx
, !is_sync
|| is_flush_fua
);
954 * Default mapping to a software queue, since we use one per CPU.
956 struct blk_mq_hw_ctx
*blk_mq_map_queue(struct request_queue
*q
, const int cpu
)
958 return q
->queue_hw_ctx
[q
->mq_map
[cpu
]];
960 EXPORT_SYMBOL(blk_mq_map_queue
);
962 struct blk_mq_hw_ctx
*blk_mq_alloc_single_hw_queue(struct blk_mq_reg
*reg
,
963 unsigned int hctx_index
)
965 return kmalloc_node(sizeof(struct blk_mq_hw_ctx
),
966 GFP_KERNEL
| __GFP_ZERO
, reg
->numa_node
);
968 EXPORT_SYMBOL(blk_mq_alloc_single_hw_queue
);
970 void blk_mq_free_single_hw_queue(struct blk_mq_hw_ctx
*hctx
,
971 unsigned int hctx_index
)
975 EXPORT_SYMBOL(blk_mq_free_single_hw_queue
);
977 static void blk_mq_hctx_notify(void *data
, unsigned long action
,
980 struct blk_mq_hw_ctx
*hctx
= data
;
981 struct request_queue
*q
= hctx
->queue
;
982 struct blk_mq_ctx
*ctx
;
985 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
989 * Move ctx entries to new CPU, if this one is going away.
991 ctx
= __blk_mq_get_ctx(q
, cpu
);
993 spin_lock(&ctx
->lock
);
994 if (!list_empty(&ctx
->rq_list
)) {
995 list_splice_init(&ctx
->rq_list
, &tmp
);
996 clear_bit(ctx
->index_hw
, hctx
->ctx_map
);
998 spin_unlock(&ctx
->lock
);
1000 if (list_empty(&tmp
))
1003 ctx
= blk_mq_get_ctx(q
);
1004 spin_lock(&ctx
->lock
);
1006 while (!list_empty(&tmp
)) {
1009 rq
= list_first_entry(&tmp
, struct request
, queuelist
);
1011 list_move_tail(&rq
->queuelist
, &ctx
->rq_list
);
1014 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1015 blk_mq_hctx_mark_pending(hctx
, ctx
);
1017 spin_unlock(&ctx
->lock
);
1019 blk_mq_run_hw_queue(hctx
, true);
1020 blk_mq_put_ctx(ctx
);
1023 static int blk_mq_init_hw_commands(struct blk_mq_hw_ctx
*hctx
,
1024 int (*init
)(void *, struct blk_mq_hw_ctx
*,
1025 struct request
*, unsigned int),
1031 for (i
= 0; i
< hctx
->queue_depth
; i
++) {
1032 struct request
*rq
= hctx
->rqs
[i
];
1034 ret
= init(data
, hctx
, rq
, i
);
1042 int blk_mq_init_commands(struct request_queue
*q
,
1043 int (*init
)(void *, struct blk_mq_hw_ctx
*,
1044 struct request
*, unsigned int),
1047 struct blk_mq_hw_ctx
*hctx
;
1051 queue_for_each_hw_ctx(q
, hctx
, i
) {
1052 ret
= blk_mq_init_hw_commands(hctx
, init
, data
);
1059 EXPORT_SYMBOL(blk_mq_init_commands
);
1061 static void blk_mq_free_hw_commands(struct blk_mq_hw_ctx
*hctx
,
1062 void (*free
)(void *, struct blk_mq_hw_ctx
*,
1063 struct request
*, unsigned int),
1068 for (i
= 0; i
< hctx
->queue_depth
; i
++) {
1069 struct request
*rq
= hctx
->rqs
[i
];
1071 free(data
, hctx
, rq
, i
);
1075 void blk_mq_free_commands(struct request_queue
*q
,
1076 void (*free
)(void *, struct blk_mq_hw_ctx
*,
1077 struct request
*, unsigned int),
1080 struct blk_mq_hw_ctx
*hctx
;
1083 queue_for_each_hw_ctx(q
, hctx
, i
)
1084 blk_mq_free_hw_commands(hctx
, free
, data
);
1086 EXPORT_SYMBOL(blk_mq_free_commands
);
1088 static void blk_mq_free_rq_map(struct blk_mq_hw_ctx
*hctx
)
1092 while (!list_empty(&hctx
->page_list
)) {
1093 page
= list_first_entry(&hctx
->page_list
, struct page
, lru
);
1094 list_del_init(&page
->lru
);
1095 __free_pages(page
, page
->private);
1101 blk_mq_free_tags(hctx
->tags
);
1104 static size_t order_to_size(unsigned int order
)
1106 size_t ret
= PAGE_SIZE
;
1114 static int blk_mq_init_rq_map(struct blk_mq_hw_ctx
*hctx
,
1115 unsigned int reserved_tags
, int node
)
1117 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1118 size_t rq_size
, left
;
1120 INIT_LIST_HEAD(&hctx
->page_list
);
1122 hctx
->rqs
= kmalloc_node(hctx
->queue_depth
* sizeof(struct request
*),
1128 * rq_size is the size of the request plus driver payload, rounded
1129 * to the cacheline size
1131 rq_size
= round_up(sizeof(struct request
) + hctx
->cmd_size
,
1133 left
= rq_size
* hctx
->queue_depth
;
1135 for (i
= 0; i
< hctx
->queue_depth
;) {
1136 int this_order
= max_order
;
1141 while (left
< order_to_size(this_order
- 1) && this_order
)
1145 page
= alloc_pages_node(node
, GFP_KERNEL
, this_order
);
1150 if (order_to_size(this_order
) < rq_size
)
1157 page
->private = this_order
;
1158 list_add_tail(&page
->lru
, &hctx
->page_list
);
1160 p
= page_address(page
);
1161 entries_per_page
= order_to_size(this_order
) / rq_size
;
1162 to_do
= min(entries_per_page
, hctx
->queue_depth
- i
);
1163 left
-= to_do
* rq_size
;
1164 for (j
= 0; j
< to_do
; j
++) {
1166 blk_mq_rq_init(hctx
, hctx
->rqs
[i
]);
1172 if (i
< (reserved_tags
+ BLK_MQ_TAG_MIN
))
1174 else if (i
!= hctx
->queue_depth
) {
1175 hctx
->queue_depth
= i
;
1176 pr_warn("%s: queue depth set to %u because of low memory\n",
1180 hctx
->tags
= blk_mq_init_tags(hctx
->queue_depth
, reserved_tags
, node
);
1183 blk_mq_free_rq_map(hctx
);
1190 static int blk_mq_init_hw_queues(struct request_queue
*q
,
1191 struct blk_mq_reg
*reg
, void *driver_data
)
1193 struct blk_mq_hw_ctx
*hctx
;
1197 * Initialize hardware queues
1199 queue_for_each_hw_ctx(q
, hctx
, i
) {
1200 unsigned int num_maps
;
1203 node
= hctx
->numa_node
;
1204 if (node
== NUMA_NO_NODE
)
1205 node
= hctx
->numa_node
= reg
->numa_node
;
1207 INIT_DELAYED_WORK(&hctx
->delayed_work
, blk_mq_work_fn
);
1208 spin_lock_init(&hctx
->lock
);
1209 INIT_LIST_HEAD(&hctx
->dispatch
);
1211 hctx
->queue_num
= i
;
1212 hctx
->flags
= reg
->flags
;
1213 hctx
->queue_depth
= reg
->queue_depth
;
1214 hctx
->cmd_size
= reg
->cmd_size
;
1216 blk_mq_init_cpu_notifier(&hctx
->cpu_notifier
,
1217 blk_mq_hctx_notify
, hctx
);
1218 blk_mq_register_cpu_notifier(&hctx
->cpu_notifier
);
1220 if (blk_mq_init_rq_map(hctx
, reg
->reserved_tags
, node
))
1224 * Allocate space for all possible cpus to avoid allocation in
1227 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1232 num_maps
= ALIGN(nr_cpu_ids
, BITS_PER_LONG
) / BITS_PER_LONG
;
1233 hctx
->ctx_map
= kzalloc_node(num_maps
* sizeof(unsigned long),
1238 hctx
->nr_ctx_map
= num_maps
;
1241 if (reg
->ops
->init_hctx
&&
1242 reg
->ops
->init_hctx(hctx
, driver_data
, i
))
1246 if (i
== q
->nr_hw_queues
)
1252 queue_for_each_hw_ctx(q
, hctx
, j
) {
1256 if (reg
->ops
->exit_hctx
)
1257 reg
->ops
->exit_hctx(hctx
, j
);
1259 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1260 blk_mq_free_rq_map(hctx
);
1267 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1268 unsigned int nr_hw_queues
)
1272 for_each_possible_cpu(i
) {
1273 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1274 struct blk_mq_hw_ctx
*hctx
;
1276 memset(__ctx
, 0, sizeof(*__ctx
));
1278 spin_lock_init(&__ctx
->lock
);
1279 INIT_LIST_HEAD(&__ctx
->rq_list
);
1282 /* If the cpu isn't online, the cpu is mapped to first hctx */
1286 hctx
= q
->mq_ops
->map_queue(q
, i
);
1287 cpumask_set_cpu(i
, hctx
->cpumask
);
1291 * Set local node, IFF we have more than one hw queue. If
1292 * not, we remain on the home node of the device
1294 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1295 hctx
->numa_node
= cpu_to_node(i
);
1299 static void blk_mq_map_swqueue(struct request_queue
*q
)
1302 struct blk_mq_hw_ctx
*hctx
;
1303 struct blk_mq_ctx
*ctx
;
1305 queue_for_each_hw_ctx(q
, hctx
, i
) {
1306 cpumask_clear(hctx
->cpumask
);
1311 * Map software to hardware queues
1313 queue_for_each_ctx(q
, ctx
, i
) {
1314 /* If the cpu isn't online, the cpu is mapped to first hctx */
1318 hctx
= q
->mq_ops
->map_queue(q
, i
);
1319 cpumask_set_cpu(i
, hctx
->cpumask
);
1320 ctx
->index_hw
= hctx
->nr_ctx
;
1321 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
1325 struct request_queue
*blk_mq_init_queue(struct blk_mq_reg
*reg
,
1328 struct blk_mq_hw_ctx
**hctxs
;
1329 struct blk_mq_ctx
*ctx
;
1330 struct request_queue
*q
;
1333 if (!reg
->nr_hw_queues
||
1334 !reg
->ops
->queue_rq
|| !reg
->ops
->map_queue
||
1335 !reg
->ops
->alloc_hctx
|| !reg
->ops
->free_hctx
)
1336 return ERR_PTR(-EINVAL
);
1338 if (!reg
->queue_depth
)
1339 reg
->queue_depth
= BLK_MQ_MAX_DEPTH
;
1340 else if (reg
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
1341 pr_err("blk-mq: queuedepth too large (%u)\n", reg
->queue_depth
);
1342 reg
->queue_depth
= BLK_MQ_MAX_DEPTH
;
1345 if (reg
->queue_depth
< (reg
->reserved_tags
+ BLK_MQ_TAG_MIN
))
1346 return ERR_PTR(-EINVAL
);
1348 ctx
= alloc_percpu(struct blk_mq_ctx
);
1350 return ERR_PTR(-ENOMEM
);
1352 hctxs
= kmalloc_node(reg
->nr_hw_queues
* sizeof(*hctxs
), GFP_KERNEL
,
1358 for (i
= 0; i
< reg
->nr_hw_queues
; i
++) {
1359 hctxs
[i
] = reg
->ops
->alloc_hctx(reg
, 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
, reg
->numa_node
);
1374 q
->mq_map
= blk_mq_make_queue_map(reg
);
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
= reg
->nr_hw_queues
;
1385 q
->queue_hw_ctx
= hctxs
;
1387 q
->mq_ops
= reg
->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
, reg
->ops
->timeout
);
1395 blk_queue_rq_timeout(q
, reg
->timeout
);
1397 if (reg
->ops
->complete
)
1398 blk_queue_softirq_done(q
, reg
->ops
->complete
);
1400 blk_mq_init_flush(q
);
1401 blk_mq_init_cpu_queues(q
, reg
->nr_hw_queues
);
1403 q
->flush_rq
= kzalloc(round_up(sizeof(struct request
) + reg
->cmd_size
,
1404 cache_line_size()), GFP_KERNEL
);
1408 if (blk_mq_init_hw_queues(q
, reg
, driver_data
))
1411 blk_mq_map_swqueue(q
);
1413 mutex_lock(&all_q_mutex
);
1414 list_add_tail(&q
->all_q_node
, &all_q_list
);
1415 mutex_unlock(&all_q_mutex
);
1424 blk_cleanup_queue(q
);
1426 for (i
= 0; i
< reg
->nr_hw_queues
; i
++) {
1429 free_cpumask_var(hctxs
[i
]->cpumask
);
1430 reg
->ops
->free_hctx(hctxs
[i
], i
);
1435 return ERR_PTR(-ENOMEM
);
1437 EXPORT_SYMBOL(blk_mq_init_queue
);
1439 void blk_mq_free_queue(struct request_queue
*q
)
1441 struct blk_mq_hw_ctx
*hctx
;
1444 queue_for_each_hw_ctx(q
, hctx
, i
) {
1445 kfree(hctx
->ctx_map
);
1447 blk_mq_free_rq_map(hctx
);
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 void blk_mq_disable_hotplug(void)
1510 mutex_lock(&all_q_mutex
);
1513 void blk_mq_enable_hotplug(void)
1515 mutex_unlock(&all_q_mutex
);
1518 static int __init
blk_mq_init(void)
1522 /* Must be called after percpu_counter_hotcpu_callback() */
1523 hotcpu_notifier(blk_mq_queue_reinit_notify
, -10);
1527 subsys_initcall(blk_mq_init
);