2 * Block multiqueue core code
4 * Copyright (C) 2013-2014 Jens Axboe
5 * Copyright (C) 2013-2014 Christoph Hellwig
7 #include <linux/kernel.h>
8 #include <linux/module.h>
9 #include <linux/backing-dev.h>
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
13 #include <linux/init.h>
14 #include <linux/slab.h>
15 #include <linux/workqueue.h>
16 #include <linux/smp.h>
17 #include <linux/llist.h>
18 #include <linux/list_sort.h>
19 #include <linux/cpu.h>
20 #include <linux/cache.h>
21 #include <linux/sched/sysctl.h>
22 #include <linux/delay.h>
24 #include <trace/events/block.h>
26 #include <linux/blk-mq.h>
29 #include "blk-mq-tag.h"
31 static DEFINE_MUTEX(all_q_mutex
);
32 static LIST_HEAD(all_q_list
);
34 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
);
37 * Check if any of the ctx's have pending work in this hardware queue
39 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
43 for (i
= 0; i
< hctx
->ctx_map
.map_size
; i
++)
44 if (hctx
->ctx_map
.map
[i
].word
)
50 static inline struct blk_align_bitmap
*get_bm(struct blk_mq_hw_ctx
*hctx
,
51 struct blk_mq_ctx
*ctx
)
53 return &hctx
->ctx_map
.map
[ctx
->index_hw
/ hctx
->ctx_map
.bits_per_word
];
56 #define CTX_TO_BIT(hctx, ctx) \
57 ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
60 * Mark this ctx as having pending work in this hardware queue
62 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
63 struct blk_mq_ctx
*ctx
)
65 struct blk_align_bitmap
*bm
= get_bm(hctx
, ctx
);
67 if (!test_bit(CTX_TO_BIT(hctx
, ctx
), &bm
->word
))
68 set_bit(CTX_TO_BIT(hctx
, ctx
), &bm
->word
);
71 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
72 struct blk_mq_ctx
*ctx
)
74 struct blk_align_bitmap
*bm
= get_bm(hctx
, ctx
);
76 clear_bit(CTX_TO_BIT(hctx
, ctx
), &bm
->word
);
79 static int blk_mq_queue_enter(struct request_queue
*q
)
84 if (percpu_ref_tryget_live(&q
->mq_usage_counter
))
87 ret
= wait_event_interruptible(q
->mq_freeze_wq
,
88 !q
->mq_freeze_depth
|| blk_queue_dying(q
));
89 if (blk_queue_dying(q
))
96 static void blk_mq_queue_exit(struct request_queue
*q
)
98 percpu_ref_put(&q
->mq_usage_counter
);
101 static void blk_mq_usage_counter_release(struct percpu_ref
*ref
)
103 struct request_queue
*q
=
104 container_of(ref
, struct request_queue
, mq_usage_counter
);
106 wake_up_all(&q
->mq_freeze_wq
);
110 * Guarantee no request is in use, so we can change any data structure of
111 * the queue afterward.
113 void blk_mq_freeze_queue(struct request_queue
*q
)
117 spin_lock_irq(q
->queue_lock
);
118 freeze
= !q
->mq_freeze_depth
++;
119 spin_unlock_irq(q
->queue_lock
);
122 percpu_ref_kill(&q
->mq_usage_counter
);
123 blk_mq_run_queues(q
, false);
125 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->mq_usage_counter
));
128 static void blk_mq_unfreeze_queue(struct request_queue
*q
)
132 spin_lock_irq(q
->queue_lock
);
133 wake
= !--q
->mq_freeze_depth
;
134 WARN_ON_ONCE(q
->mq_freeze_depth
< 0);
135 spin_unlock_irq(q
->queue_lock
);
137 percpu_ref_reinit(&q
->mq_usage_counter
);
138 wake_up_all(&q
->mq_freeze_wq
);
142 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
144 return blk_mq_has_free_tags(hctx
->tags
);
146 EXPORT_SYMBOL(blk_mq_can_queue
);
148 static void blk_mq_rq_ctx_init(struct request_queue
*q
, struct blk_mq_ctx
*ctx
,
149 struct request
*rq
, unsigned int rw_flags
)
151 if (blk_queue_io_stat(q
))
152 rw_flags
|= REQ_IO_STAT
;
154 INIT_LIST_HEAD(&rq
->queuelist
);
155 /* csd/requeue_work/fifo_time is initialized before use */
158 rq
->cmd_flags
|= rw_flags
;
159 /* do not touch atomic flags, it needs atomic ops against the timer */
161 INIT_HLIST_NODE(&rq
->hash
);
162 RB_CLEAR_NODE(&rq
->rb_node
);
165 rq
->start_time
= jiffies
;
166 #ifdef CONFIG_BLK_CGROUP
168 set_start_time_ns(rq
);
169 rq
->io_start_time_ns
= 0;
171 rq
->nr_phys_segments
= 0;
172 #if defined(CONFIG_BLK_DEV_INTEGRITY)
173 rq
->nr_integrity_segments
= 0;
176 /* tag was already set */
186 INIT_LIST_HEAD(&rq
->timeout_list
);
190 rq
->end_io_data
= NULL
;
193 ctx
->rq_dispatched
[rw_is_sync(rw_flags
)]++;
196 static struct request
*
197 __blk_mq_alloc_request(struct blk_mq_alloc_data
*data
, int rw
)
202 tag
= blk_mq_get_tag(data
);
203 if (tag
!= BLK_MQ_TAG_FAIL
) {
204 rq
= data
->hctx
->tags
->rqs
[tag
];
206 if (blk_mq_tag_busy(data
->hctx
)) {
207 rq
->cmd_flags
= REQ_MQ_INFLIGHT
;
208 atomic_inc(&data
->hctx
->nr_active
);
212 blk_mq_rq_ctx_init(data
->q
, data
->ctx
, rq
, rw
);
219 struct request
*blk_mq_alloc_request(struct request_queue
*q
, int rw
, gfp_t gfp
,
222 struct blk_mq_ctx
*ctx
;
223 struct blk_mq_hw_ctx
*hctx
;
225 struct blk_mq_alloc_data alloc_data
;
228 ret
= blk_mq_queue_enter(q
);
232 ctx
= blk_mq_get_ctx(q
);
233 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
234 blk_mq_set_alloc_data(&alloc_data
, q
, gfp
& ~__GFP_WAIT
,
235 reserved
, ctx
, hctx
);
237 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
238 if (!rq
&& (gfp
& __GFP_WAIT
)) {
239 __blk_mq_run_hw_queue(hctx
);
242 ctx
= blk_mq_get_ctx(q
);
243 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
244 blk_mq_set_alloc_data(&alloc_data
, q
, gfp
, reserved
, ctx
,
246 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
247 ctx
= alloc_data
.ctx
;
251 return ERR_PTR(-EWOULDBLOCK
);
254 EXPORT_SYMBOL(blk_mq_alloc_request
);
256 static void __blk_mq_free_request(struct blk_mq_hw_ctx
*hctx
,
257 struct blk_mq_ctx
*ctx
, struct request
*rq
)
259 const int tag
= rq
->tag
;
260 struct request_queue
*q
= rq
->q
;
262 if (rq
->cmd_flags
& REQ_MQ_INFLIGHT
)
263 atomic_dec(&hctx
->nr_active
);
266 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
267 blk_mq_put_tag(hctx
, tag
, &ctx
->last_tag
);
268 blk_mq_queue_exit(q
);
271 void blk_mq_free_request(struct request
*rq
)
273 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
274 struct blk_mq_hw_ctx
*hctx
;
275 struct request_queue
*q
= rq
->q
;
277 ctx
->rq_completed
[rq_is_sync(rq
)]++;
279 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
280 __blk_mq_free_request(hctx
, ctx
, rq
);
284 * Clone all relevant state from a request that has been put on hold in
285 * the flush state machine into the preallocated flush request that hangs
286 * off the request queue.
288 * For a driver the flush request should be invisible, that's why we are
289 * impersonating the original request here.
291 void blk_mq_clone_flush_request(struct request
*flush_rq
,
292 struct request
*orig_rq
)
294 struct blk_mq_hw_ctx
*hctx
=
295 orig_rq
->q
->mq_ops
->map_queue(orig_rq
->q
, orig_rq
->mq_ctx
->cpu
);
297 flush_rq
->mq_ctx
= orig_rq
->mq_ctx
;
298 flush_rq
->tag
= orig_rq
->tag
;
299 memcpy(blk_mq_rq_to_pdu(flush_rq
), blk_mq_rq_to_pdu(orig_rq
),
303 inline void __blk_mq_end_io(struct request
*rq
, int error
)
305 blk_account_io_done(rq
);
308 rq
->end_io(rq
, error
);
310 if (unlikely(blk_bidi_rq(rq
)))
311 blk_mq_free_request(rq
->next_rq
);
312 blk_mq_free_request(rq
);
315 EXPORT_SYMBOL(__blk_mq_end_io
);
317 void blk_mq_end_io(struct request
*rq
, int error
)
319 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
321 __blk_mq_end_io(rq
, error
);
323 EXPORT_SYMBOL(blk_mq_end_io
);
325 static void __blk_mq_complete_request_remote(void *data
)
327 struct request
*rq
= data
;
329 rq
->q
->softirq_done_fn(rq
);
332 static void blk_mq_ipi_complete_request(struct request
*rq
)
334 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
338 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
339 rq
->q
->softirq_done_fn(rq
);
344 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
345 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
347 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
348 rq
->csd
.func
= __blk_mq_complete_request_remote
;
351 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
353 rq
->q
->softirq_done_fn(rq
);
358 void __blk_mq_complete_request(struct request
*rq
)
360 struct request_queue
*q
= rq
->q
;
362 if (!q
->softirq_done_fn
)
363 blk_mq_end_io(rq
, rq
->errors
);
365 blk_mq_ipi_complete_request(rq
);
369 * blk_mq_complete_request - end I/O on a request
370 * @rq: the request being processed
373 * Ends all I/O on a request. It does not handle partial completions.
374 * The actual completion happens out-of-order, through a IPI handler.
376 void blk_mq_complete_request(struct request
*rq
)
378 struct request_queue
*q
= rq
->q
;
380 if (unlikely(blk_should_fake_timeout(q
)))
382 if (!blk_mark_rq_complete(rq
))
383 __blk_mq_complete_request(rq
);
385 EXPORT_SYMBOL(blk_mq_complete_request
);
387 static void blk_mq_start_request(struct request
*rq
, bool last
)
389 struct request_queue
*q
= rq
->q
;
391 trace_block_rq_issue(q
, rq
);
393 rq
->resid_len
= blk_rq_bytes(rq
);
394 if (unlikely(blk_bidi_rq(rq
)))
395 rq
->next_rq
->resid_len
= blk_rq_bytes(rq
->next_rq
);
400 * Ensure that ->deadline is visible before set the started
401 * flag and clear the completed flag.
403 smp_mb__before_atomic();
406 * Mark us as started and clear complete. Complete might have been
407 * set if requeue raced with timeout, which then marked it as
408 * complete. So be sure to clear complete again when we start
409 * the request, otherwise we'll ignore the completion event.
411 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
412 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
413 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
414 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
416 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
418 * Make sure space for the drain appears. We know we can do
419 * this because max_hw_segments has been adjusted to be one
420 * fewer than the device can handle.
422 rq
->nr_phys_segments
++;
426 * Flag the last request in the series so that drivers know when IO
427 * should be kicked off, if they don't do it on a per-request basis.
429 * Note: the flag isn't the only condition drivers should do kick off.
430 * If drive is busy, the last request might not have the bit set.
433 rq
->cmd_flags
|= REQ_END
;
436 static void __blk_mq_requeue_request(struct request
*rq
)
438 struct request_queue
*q
= rq
->q
;
440 trace_block_rq_requeue(q
, rq
);
441 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
443 rq
->cmd_flags
&= ~REQ_END
;
445 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
446 rq
->nr_phys_segments
--;
449 void blk_mq_requeue_request(struct request
*rq
)
451 __blk_mq_requeue_request(rq
);
452 blk_clear_rq_complete(rq
);
454 BUG_ON(blk_queued_rq(rq
));
455 blk_mq_add_to_requeue_list(rq
, true);
457 EXPORT_SYMBOL(blk_mq_requeue_request
);
459 static void blk_mq_requeue_work(struct work_struct
*work
)
461 struct request_queue
*q
=
462 container_of(work
, struct request_queue
, requeue_work
);
464 struct request
*rq
, *next
;
467 spin_lock_irqsave(&q
->requeue_lock
, flags
);
468 list_splice_init(&q
->requeue_list
, &rq_list
);
469 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
471 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
472 if (!(rq
->cmd_flags
& REQ_SOFTBARRIER
))
475 rq
->cmd_flags
&= ~REQ_SOFTBARRIER
;
476 list_del_init(&rq
->queuelist
);
477 blk_mq_insert_request(rq
, true, false, false);
480 while (!list_empty(&rq_list
)) {
481 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
482 list_del_init(&rq
->queuelist
);
483 blk_mq_insert_request(rq
, false, false, false);
487 * Use the start variant of queue running here, so that running
488 * the requeue work will kick stopped queues.
490 blk_mq_start_hw_queues(q
);
493 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
)
495 struct request_queue
*q
= rq
->q
;
499 * We abuse this flag that is otherwise used by the I/O scheduler to
500 * request head insertation from the workqueue.
502 BUG_ON(rq
->cmd_flags
& REQ_SOFTBARRIER
);
504 spin_lock_irqsave(&q
->requeue_lock
, flags
);
506 rq
->cmd_flags
|= REQ_SOFTBARRIER
;
507 list_add(&rq
->queuelist
, &q
->requeue_list
);
509 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
511 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
513 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
515 void blk_mq_kick_requeue_list(struct request_queue
*q
)
517 kblockd_schedule_work(&q
->requeue_work
);
519 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
521 static inline bool is_flush_request(struct request
*rq
, unsigned int tag
)
523 return ((rq
->cmd_flags
& REQ_FLUSH_SEQ
) &&
524 rq
->q
->flush_rq
->tag
== tag
);
527 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
529 struct request
*rq
= tags
->rqs
[tag
];
531 if (!is_flush_request(rq
, tag
))
534 return rq
->q
->flush_rq
;
536 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
538 struct blk_mq_timeout_data
{
539 struct blk_mq_hw_ctx
*hctx
;
541 unsigned int *next_set
;
544 static void blk_mq_timeout_check(void *__data
, unsigned long *free_tags
)
546 struct blk_mq_timeout_data
*data
= __data
;
547 struct blk_mq_hw_ctx
*hctx
= data
->hctx
;
550 /* It may not be in flight yet (this is where
551 * the REQ_ATOMIC_STARTED flag comes in). The requests are
552 * statically allocated, so we know it's always safe to access the
553 * memory associated with a bit offset into ->rqs[].
559 tag
= find_next_zero_bit(free_tags
, hctx
->tags
->nr_tags
, tag
);
560 if (tag
>= hctx
->tags
->nr_tags
)
563 rq
= blk_mq_tag_to_rq(hctx
->tags
, tag
++);
564 if (rq
->q
!= hctx
->queue
)
566 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
569 blk_rq_check_expired(rq
, data
->next
, data
->next_set
);
573 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx
*hctx
,
575 unsigned int *next_set
)
577 struct blk_mq_timeout_data data
= {
580 .next_set
= next_set
,
584 * Ask the tagging code to iterate busy requests, so we can
585 * check them for timeout.
587 blk_mq_tag_busy_iter(hctx
->tags
, blk_mq_timeout_check
, &data
);
590 static enum blk_eh_timer_return
blk_mq_rq_timed_out(struct request
*rq
)
592 struct request_queue
*q
= rq
->q
;
595 * We know that complete is set at this point. If STARTED isn't set
596 * anymore, then the request isn't active and the "timeout" should
597 * just be ignored. This can happen due to the bitflag ordering.
598 * Timeout first checks if STARTED is set, and if it is, assumes
599 * the request is active. But if we race with completion, then
600 * we both flags will get cleared. So check here again, and ignore
601 * a timeout event with a request that isn't active.
603 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
604 return BLK_EH_NOT_HANDLED
;
606 if (!q
->mq_ops
->timeout
)
607 return BLK_EH_RESET_TIMER
;
609 return q
->mq_ops
->timeout(rq
);
612 static void blk_mq_rq_timer(unsigned long data
)
614 struct request_queue
*q
= (struct request_queue
*) data
;
615 struct blk_mq_hw_ctx
*hctx
;
616 unsigned long next
= 0;
619 queue_for_each_hw_ctx(q
, hctx
, i
) {
621 * If not software queues are currently mapped to this
622 * hardware queue, there's nothing to check
624 if (!hctx
->nr_ctx
|| !hctx
->tags
)
627 blk_mq_hw_ctx_check_timeout(hctx
, &next
, &next_set
);
631 next
= blk_rq_timeout(round_jiffies_up(next
));
632 mod_timer(&q
->timeout
, next
);
634 queue_for_each_hw_ctx(q
, hctx
, i
)
635 blk_mq_tag_idle(hctx
);
640 * Reverse check our software queue for entries that we could potentially
641 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
642 * too much time checking for merges.
644 static bool blk_mq_attempt_merge(struct request_queue
*q
,
645 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
650 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
656 if (!blk_rq_merge_ok(rq
, bio
))
659 el_ret
= blk_try_merge(rq
, bio
);
660 if (el_ret
== ELEVATOR_BACK_MERGE
) {
661 if (bio_attempt_back_merge(q
, rq
, bio
)) {
666 } else if (el_ret
== ELEVATOR_FRONT_MERGE
) {
667 if (bio_attempt_front_merge(q
, rq
, bio
)) {
679 * Process software queues that have been marked busy, splicing them
680 * to the for-dispatch
682 static void flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
684 struct blk_mq_ctx
*ctx
;
687 for (i
= 0; i
< hctx
->ctx_map
.map_size
; i
++) {
688 struct blk_align_bitmap
*bm
= &hctx
->ctx_map
.map
[i
];
689 unsigned int off
, bit
;
695 off
= i
* hctx
->ctx_map
.bits_per_word
;
697 bit
= find_next_bit(&bm
->word
, bm
->depth
, bit
);
698 if (bit
>= bm
->depth
)
701 ctx
= hctx
->ctxs
[bit
+ off
];
702 clear_bit(bit
, &bm
->word
);
703 spin_lock(&ctx
->lock
);
704 list_splice_tail_init(&ctx
->rq_list
, list
);
705 spin_unlock(&ctx
->lock
);
713 * Run this hardware queue, pulling any software queues mapped to it in.
714 * Note that this function currently has various problems around ordering
715 * of IO. In particular, we'd like FIFO behaviour on handling existing
716 * items on the hctx->dispatch list. Ignore that for now.
718 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
720 struct request_queue
*q
= hctx
->queue
;
725 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
));
727 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)))
733 * Touch any software queue that has pending entries.
735 flush_busy_ctxs(hctx
, &rq_list
);
738 * If we have previous entries on our dispatch list, grab them
739 * and stuff them at the front for more fair dispatch.
741 if (!list_empty_careful(&hctx
->dispatch
)) {
742 spin_lock(&hctx
->lock
);
743 if (!list_empty(&hctx
->dispatch
))
744 list_splice_init(&hctx
->dispatch
, &rq_list
);
745 spin_unlock(&hctx
->lock
);
749 * Now process all the entries, sending them to the driver.
752 while (!list_empty(&rq_list
)) {
755 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
756 list_del_init(&rq
->queuelist
);
758 blk_mq_start_request(rq
, list_empty(&rq_list
));
760 ret
= q
->mq_ops
->queue_rq(hctx
, rq
);
762 case BLK_MQ_RQ_QUEUE_OK
:
765 case BLK_MQ_RQ_QUEUE_BUSY
:
766 list_add(&rq
->queuelist
, &rq_list
);
767 __blk_mq_requeue_request(rq
);
770 pr_err("blk-mq: bad return on queue: %d\n", ret
);
771 case BLK_MQ_RQ_QUEUE_ERROR
:
773 blk_mq_end_io(rq
, rq
->errors
);
777 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
782 hctx
->dispatched
[0]++;
783 else if (queued
< (1 << (BLK_MQ_MAX_DISPATCH_ORDER
- 1)))
784 hctx
->dispatched
[ilog2(queued
) + 1]++;
787 * Any items that need requeuing? Stuff them into hctx->dispatch,
788 * that is where we will continue on next queue run.
790 if (!list_empty(&rq_list
)) {
791 spin_lock(&hctx
->lock
);
792 list_splice(&rq_list
, &hctx
->dispatch
);
793 spin_unlock(&hctx
->lock
);
798 * It'd be great if the workqueue API had a way to pass
799 * in a mask and had some smarts for more clever placement.
800 * For now we just round-robin here, switching for every
801 * BLK_MQ_CPU_WORK_BATCH queued items.
803 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
805 int cpu
= hctx
->next_cpu
;
807 if (--hctx
->next_cpu_batch
<= 0) {
810 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
811 if (next_cpu
>= nr_cpu_ids
)
812 next_cpu
= cpumask_first(hctx
->cpumask
);
814 hctx
->next_cpu
= next_cpu
;
815 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
821 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
823 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)))
826 if (!async
&& cpumask_test_cpu(smp_processor_id(), hctx
->cpumask
))
827 __blk_mq_run_hw_queue(hctx
);
828 else if (hctx
->queue
->nr_hw_queues
== 1)
829 kblockd_schedule_delayed_work(&hctx
->run_work
, 0);
833 cpu
= blk_mq_hctx_next_cpu(hctx
);
834 kblockd_schedule_delayed_work_on(cpu
, &hctx
->run_work
, 0);
838 void blk_mq_run_queues(struct request_queue
*q
, bool async
)
840 struct blk_mq_hw_ctx
*hctx
;
843 queue_for_each_hw_ctx(q
, hctx
, i
) {
844 if ((!blk_mq_hctx_has_pending(hctx
) &&
845 list_empty_careful(&hctx
->dispatch
)) ||
846 test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
850 blk_mq_run_hw_queue(hctx
, async
);
854 EXPORT_SYMBOL(blk_mq_run_queues
);
856 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
858 cancel_delayed_work(&hctx
->run_work
);
859 cancel_delayed_work(&hctx
->delay_work
);
860 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
862 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
864 void blk_mq_stop_hw_queues(struct request_queue
*q
)
866 struct blk_mq_hw_ctx
*hctx
;
869 queue_for_each_hw_ctx(q
, hctx
, i
)
870 blk_mq_stop_hw_queue(hctx
);
872 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
874 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
876 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
879 blk_mq_run_hw_queue(hctx
, false);
882 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
884 void blk_mq_start_hw_queues(struct request_queue
*q
)
886 struct blk_mq_hw_ctx
*hctx
;
889 queue_for_each_hw_ctx(q
, hctx
, i
)
890 blk_mq_start_hw_queue(hctx
);
892 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
895 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
897 struct blk_mq_hw_ctx
*hctx
;
900 queue_for_each_hw_ctx(q
, hctx
, i
) {
901 if (!test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
904 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
906 blk_mq_run_hw_queue(hctx
, async
);
910 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
912 static void blk_mq_run_work_fn(struct work_struct
*work
)
914 struct blk_mq_hw_ctx
*hctx
;
916 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
918 __blk_mq_run_hw_queue(hctx
);
921 static void blk_mq_delay_work_fn(struct work_struct
*work
)
923 struct blk_mq_hw_ctx
*hctx
;
925 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
927 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
928 __blk_mq_run_hw_queue(hctx
);
931 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
933 unsigned long tmo
= msecs_to_jiffies(msecs
);
935 if (hctx
->queue
->nr_hw_queues
== 1)
936 kblockd_schedule_delayed_work(&hctx
->delay_work
, tmo
);
940 cpu
= blk_mq_hctx_next_cpu(hctx
);
941 kblockd_schedule_delayed_work_on(cpu
, &hctx
->delay_work
, tmo
);
944 EXPORT_SYMBOL(blk_mq_delay_queue
);
946 static void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
,
947 struct request
*rq
, bool at_head
)
949 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
951 trace_block_rq_insert(hctx
->queue
, rq
);
954 list_add(&rq
->queuelist
, &ctx
->rq_list
);
956 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
958 blk_mq_hctx_mark_pending(hctx
, ctx
);
961 void blk_mq_insert_request(struct request
*rq
, bool at_head
, bool run_queue
,
964 struct request_queue
*q
= rq
->q
;
965 struct blk_mq_hw_ctx
*hctx
;
966 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
, *current_ctx
;
968 current_ctx
= blk_mq_get_ctx(q
);
969 if (!cpu_online(ctx
->cpu
))
970 rq
->mq_ctx
= ctx
= current_ctx
;
972 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
974 spin_lock(&ctx
->lock
);
975 __blk_mq_insert_request(hctx
, rq
, at_head
);
976 spin_unlock(&ctx
->lock
);
979 blk_mq_run_hw_queue(hctx
, async
);
981 blk_mq_put_ctx(current_ctx
);
984 static void blk_mq_insert_requests(struct request_queue
*q
,
985 struct blk_mq_ctx
*ctx
,
986 struct list_head
*list
,
991 struct blk_mq_hw_ctx
*hctx
;
992 struct blk_mq_ctx
*current_ctx
;
994 trace_block_unplug(q
, depth
, !from_schedule
);
996 current_ctx
= blk_mq_get_ctx(q
);
998 if (!cpu_online(ctx
->cpu
))
1000 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1003 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1006 spin_lock(&ctx
->lock
);
1007 while (!list_empty(list
)) {
1010 rq
= list_first_entry(list
, struct request
, queuelist
);
1011 list_del_init(&rq
->queuelist
);
1013 __blk_mq_insert_request(hctx
, rq
, false);
1015 spin_unlock(&ctx
->lock
);
1017 blk_mq_run_hw_queue(hctx
, from_schedule
);
1018 blk_mq_put_ctx(current_ctx
);
1021 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1023 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1024 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1026 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1027 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1028 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1031 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1033 struct blk_mq_ctx
*this_ctx
;
1034 struct request_queue
*this_q
;
1037 LIST_HEAD(ctx_list
);
1040 list_splice_init(&plug
->mq_list
, &list
);
1042 list_sort(NULL
, &list
, plug_ctx_cmp
);
1048 while (!list_empty(&list
)) {
1049 rq
= list_entry_rq(list
.next
);
1050 list_del_init(&rq
->queuelist
);
1052 if (rq
->mq_ctx
!= this_ctx
) {
1054 blk_mq_insert_requests(this_q
, this_ctx
,
1059 this_ctx
= rq
->mq_ctx
;
1065 list_add_tail(&rq
->queuelist
, &ctx_list
);
1069 * If 'this_ctx' is set, we know we have entries to complete
1070 * on 'ctx_list'. Do those.
1073 blk_mq_insert_requests(this_q
, this_ctx
, &ctx_list
, depth
,
1078 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1080 init_request_from_bio(rq
, bio
);
1082 if (blk_do_io_stat(rq
))
1083 blk_account_io_start(rq
, 1);
1086 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1088 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1089 !blk_queue_nomerges(hctx
->queue
);
1092 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx
*hctx
,
1093 struct blk_mq_ctx
*ctx
,
1094 struct request
*rq
, struct bio
*bio
)
1096 if (!hctx_allow_merges(hctx
)) {
1097 blk_mq_bio_to_request(rq
, bio
);
1098 spin_lock(&ctx
->lock
);
1100 __blk_mq_insert_request(hctx
, rq
, false);
1101 spin_unlock(&ctx
->lock
);
1104 struct request_queue
*q
= hctx
->queue
;
1106 spin_lock(&ctx
->lock
);
1107 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1108 blk_mq_bio_to_request(rq
, bio
);
1112 spin_unlock(&ctx
->lock
);
1113 __blk_mq_free_request(hctx
, ctx
, rq
);
1118 struct blk_map_ctx
{
1119 struct blk_mq_hw_ctx
*hctx
;
1120 struct blk_mq_ctx
*ctx
;
1123 static struct request
*blk_mq_map_request(struct request_queue
*q
,
1125 struct blk_map_ctx
*data
)
1127 struct blk_mq_hw_ctx
*hctx
;
1128 struct blk_mq_ctx
*ctx
;
1130 int rw
= bio_data_dir(bio
);
1131 struct blk_mq_alloc_data alloc_data
;
1133 if (unlikely(blk_mq_queue_enter(q
))) {
1134 bio_endio(bio
, -EIO
);
1138 ctx
= blk_mq_get_ctx(q
);
1139 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1141 if (rw_is_sync(bio
->bi_rw
))
1144 trace_block_getrq(q
, bio
, rw
);
1145 blk_mq_set_alloc_data(&alloc_data
, q
, GFP_ATOMIC
, false, ctx
,
1147 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
1148 if (unlikely(!rq
)) {
1149 __blk_mq_run_hw_queue(hctx
);
1150 blk_mq_put_ctx(ctx
);
1151 trace_block_sleeprq(q
, bio
, rw
);
1153 ctx
= blk_mq_get_ctx(q
);
1154 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1155 blk_mq_set_alloc_data(&alloc_data
, q
,
1156 __GFP_WAIT
|GFP_ATOMIC
, false, ctx
, hctx
);
1157 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
1158 ctx
= alloc_data
.ctx
;
1159 hctx
= alloc_data
.hctx
;
1169 * Multiple hardware queue variant. This will not use per-process plugs,
1170 * but will attempt to bypass the hctx queueing if we can go straight to
1171 * hardware for SYNC IO.
1173 static void blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1175 const int is_sync
= rw_is_sync(bio
->bi_rw
);
1176 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
1177 struct blk_map_ctx data
;
1180 blk_queue_bounce(q
, &bio
);
1182 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1183 bio_endio(bio
, -EIO
);
1187 rq
= blk_mq_map_request(q
, bio
, &data
);
1191 if (unlikely(is_flush_fua
)) {
1192 blk_mq_bio_to_request(rq
, bio
);
1193 blk_insert_flush(rq
);
1200 blk_mq_bio_to_request(rq
, bio
);
1201 blk_mq_start_request(rq
, true);
1204 * For OK queue, we are done. For error, kill it. Any other
1205 * error (busy), just add it to our list as we previously
1208 ret
= q
->mq_ops
->queue_rq(data
.hctx
, rq
);
1209 if (ret
== BLK_MQ_RQ_QUEUE_OK
)
1212 __blk_mq_requeue_request(rq
);
1214 if (ret
== BLK_MQ_RQ_QUEUE_ERROR
) {
1216 blk_mq_end_io(rq
, rq
->errors
);
1222 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1224 * For a SYNC request, send it to the hardware immediately. For
1225 * an ASYNC request, just ensure that we run it later on. The
1226 * latter allows for merging opportunities and more efficient
1230 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1233 blk_mq_put_ctx(data
.ctx
);
1237 * Single hardware queue variant. This will attempt to use any per-process
1238 * plug for merging and IO deferral.
1240 static void blk_sq_make_request(struct request_queue
*q
, struct bio
*bio
)
1242 const int is_sync
= rw_is_sync(bio
->bi_rw
);
1243 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
1244 unsigned int use_plug
, request_count
= 0;
1245 struct blk_map_ctx data
;
1249 * If we have multiple hardware queues, just go directly to
1250 * one of those for sync IO.
1252 use_plug
= !is_flush_fua
&& !is_sync
;
1254 blk_queue_bounce(q
, &bio
);
1256 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1257 bio_endio(bio
, -EIO
);
1261 if (use_plug
&& !blk_queue_nomerges(q
) &&
1262 blk_attempt_plug_merge(q
, bio
, &request_count
))
1265 rq
= blk_mq_map_request(q
, bio
, &data
);
1269 if (unlikely(is_flush_fua
)) {
1270 blk_mq_bio_to_request(rq
, bio
);
1271 blk_insert_flush(rq
);
1276 * A task plug currently exists. Since this is completely lockless,
1277 * utilize that to temporarily store requests until the task is
1278 * either done or scheduled away.
1281 struct blk_plug
*plug
= current
->plug
;
1284 blk_mq_bio_to_request(rq
, bio
);
1285 if (list_empty(&plug
->mq_list
))
1286 trace_block_plug(q
);
1287 else if (request_count
>= BLK_MAX_REQUEST_COUNT
) {
1288 blk_flush_plug_list(plug
, false);
1289 trace_block_plug(q
);
1291 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1292 blk_mq_put_ctx(data
.ctx
);
1297 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1299 * For a SYNC request, send it to the hardware immediately. For
1300 * an ASYNC request, just ensure that we run it later on. The
1301 * latter allows for merging opportunities and more efficient
1305 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1308 blk_mq_put_ctx(data
.ctx
);
1312 * Default mapping to a software queue, since we use one per CPU.
1314 struct blk_mq_hw_ctx
*blk_mq_map_queue(struct request_queue
*q
, const int cpu
)
1316 return q
->queue_hw_ctx
[q
->mq_map
[cpu
]];
1318 EXPORT_SYMBOL(blk_mq_map_queue
);
1320 static void blk_mq_free_rq_map(struct blk_mq_tag_set
*set
,
1321 struct blk_mq_tags
*tags
, unsigned int hctx_idx
)
1325 if (tags
->rqs
&& set
->ops
->exit_request
) {
1328 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1331 set
->ops
->exit_request(set
->driver_data
, tags
->rqs
[i
],
1333 tags
->rqs
[i
] = NULL
;
1337 while (!list_empty(&tags
->page_list
)) {
1338 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1339 list_del_init(&page
->lru
);
1340 __free_pages(page
, page
->private);
1345 blk_mq_free_tags(tags
);
1348 static size_t order_to_size(unsigned int order
)
1350 return (size_t)PAGE_SIZE
<< order
;
1353 static struct blk_mq_tags
*blk_mq_init_rq_map(struct blk_mq_tag_set
*set
,
1354 unsigned int hctx_idx
)
1356 struct blk_mq_tags
*tags
;
1357 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1358 size_t rq_size
, left
;
1360 tags
= blk_mq_init_tags(set
->queue_depth
, set
->reserved_tags
,
1365 INIT_LIST_HEAD(&tags
->page_list
);
1367 tags
->rqs
= kzalloc_node(set
->queue_depth
* sizeof(struct request
*),
1368 GFP_KERNEL
| __GFP_NOWARN
| __GFP_NORETRY
,
1371 blk_mq_free_tags(tags
);
1376 * rq_size is the size of the request plus driver payload, rounded
1377 * to the cacheline size
1379 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1381 left
= rq_size
* set
->queue_depth
;
1383 for (i
= 0; i
< set
->queue_depth
; ) {
1384 int this_order
= max_order
;
1389 while (left
< order_to_size(this_order
- 1) && this_order
)
1393 page
= alloc_pages_node(set
->numa_node
,
1394 GFP_KERNEL
| __GFP_NOWARN
| __GFP_NORETRY
,
1400 if (order_to_size(this_order
) < rq_size
)
1407 page
->private = this_order
;
1408 list_add_tail(&page
->lru
, &tags
->page_list
);
1410 p
= page_address(page
);
1411 entries_per_page
= order_to_size(this_order
) / rq_size
;
1412 to_do
= min(entries_per_page
, set
->queue_depth
- i
);
1413 left
-= to_do
* rq_size
;
1414 for (j
= 0; j
< to_do
; j
++) {
1416 tags
->rqs
[i
]->atomic_flags
= 0;
1417 tags
->rqs
[i
]->cmd_flags
= 0;
1418 if (set
->ops
->init_request
) {
1419 if (set
->ops
->init_request(set
->driver_data
,
1420 tags
->rqs
[i
], hctx_idx
, i
,
1422 tags
->rqs
[i
] = NULL
;
1435 blk_mq_free_rq_map(set
, tags
, hctx_idx
);
1439 static void blk_mq_free_bitmap(struct blk_mq_ctxmap
*bitmap
)
1444 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap
*bitmap
, int node
)
1446 unsigned int bpw
= 8, total
, num_maps
, i
;
1448 bitmap
->bits_per_word
= bpw
;
1450 num_maps
= ALIGN(nr_cpu_ids
, bpw
) / bpw
;
1451 bitmap
->map
= kzalloc_node(num_maps
* sizeof(struct blk_align_bitmap
),
1456 bitmap
->map_size
= num_maps
;
1459 for (i
= 0; i
< num_maps
; i
++) {
1460 bitmap
->map
[i
].depth
= min(total
, bitmap
->bits_per_word
);
1461 total
-= bitmap
->map
[i
].depth
;
1467 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx
*hctx
, int cpu
)
1469 struct request_queue
*q
= hctx
->queue
;
1470 struct blk_mq_ctx
*ctx
;
1474 * Move ctx entries to new CPU, if this one is going away.
1476 ctx
= __blk_mq_get_ctx(q
, cpu
);
1478 spin_lock(&ctx
->lock
);
1479 if (!list_empty(&ctx
->rq_list
)) {
1480 list_splice_init(&ctx
->rq_list
, &tmp
);
1481 blk_mq_hctx_clear_pending(hctx
, ctx
);
1483 spin_unlock(&ctx
->lock
);
1485 if (list_empty(&tmp
))
1488 ctx
= blk_mq_get_ctx(q
);
1489 spin_lock(&ctx
->lock
);
1491 while (!list_empty(&tmp
)) {
1494 rq
= list_first_entry(&tmp
, struct request
, queuelist
);
1496 list_move_tail(&rq
->queuelist
, &ctx
->rq_list
);
1499 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1500 blk_mq_hctx_mark_pending(hctx
, ctx
);
1502 spin_unlock(&ctx
->lock
);
1504 blk_mq_run_hw_queue(hctx
, true);
1505 blk_mq_put_ctx(ctx
);
1509 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx
*hctx
, int cpu
)
1511 struct request_queue
*q
= hctx
->queue
;
1512 struct blk_mq_tag_set
*set
= q
->tag_set
;
1514 if (set
->tags
[hctx
->queue_num
])
1517 set
->tags
[hctx
->queue_num
] = blk_mq_init_rq_map(set
, hctx
->queue_num
);
1518 if (!set
->tags
[hctx
->queue_num
])
1521 hctx
->tags
= set
->tags
[hctx
->queue_num
];
1525 static int blk_mq_hctx_notify(void *data
, unsigned long action
,
1528 struct blk_mq_hw_ctx
*hctx
= data
;
1530 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
)
1531 return blk_mq_hctx_cpu_offline(hctx
, cpu
);
1532 else if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
)
1533 return blk_mq_hctx_cpu_online(hctx
, cpu
);
1538 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1539 struct blk_mq_tag_set
*set
, int nr_queue
)
1541 struct blk_mq_hw_ctx
*hctx
;
1544 queue_for_each_hw_ctx(q
, hctx
, i
) {
1548 blk_mq_tag_idle(hctx
);
1550 if (set
->ops
->exit_hctx
)
1551 set
->ops
->exit_hctx(hctx
, i
);
1553 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1555 blk_mq_free_bitmap(&hctx
->ctx_map
);
1560 static void blk_mq_free_hw_queues(struct request_queue
*q
,
1561 struct blk_mq_tag_set
*set
)
1563 struct blk_mq_hw_ctx
*hctx
;
1566 queue_for_each_hw_ctx(q
, hctx
, i
) {
1567 free_cpumask_var(hctx
->cpumask
);
1572 static int blk_mq_init_hw_queues(struct request_queue
*q
,
1573 struct blk_mq_tag_set
*set
)
1575 struct blk_mq_hw_ctx
*hctx
;
1579 * Initialize hardware queues
1581 queue_for_each_hw_ctx(q
, hctx
, i
) {
1584 node
= hctx
->numa_node
;
1585 if (node
== NUMA_NO_NODE
)
1586 node
= hctx
->numa_node
= set
->numa_node
;
1588 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1589 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1590 spin_lock_init(&hctx
->lock
);
1591 INIT_LIST_HEAD(&hctx
->dispatch
);
1593 hctx
->queue_num
= i
;
1594 hctx
->flags
= set
->flags
;
1595 hctx
->cmd_size
= set
->cmd_size
;
1597 blk_mq_init_cpu_notifier(&hctx
->cpu_notifier
,
1598 blk_mq_hctx_notify
, hctx
);
1599 blk_mq_register_cpu_notifier(&hctx
->cpu_notifier
);
1601 hctx
->tags
= set
->tags
[i
];
1604 * Allocate space for all possible cpus to avoid allocation at
1607 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1612 if (blk_mq_alloc_bitmap(&hctx
->ctx_map
, node
))
1617 if (set
->ops
->init_hctx
&&
1618 set
->ops
->init_hctx(hctx
, set
->driver_data
, i
))
1622 if (i
== q
->nr_hw_queues
)
1628 blk_mq_exit_hw_queues(q
, set
, i
);
1633 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1634 unsigned int nr_hw_queues
)
1638 for_each_possible_cpu(i
) {
1639 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1640 struct blk_mq_hw_ctx
*hctx
;
1642 memset(__ctx
, 0, sizeof(*__ctx
));
1644 spin_lock_init(&__ctx
->lock
);
1645 INIT_LIST_HEAD(&__ctx
->rq_list
);
1648 /* If the cpu isn't online, the cpu is mapped to first hctx */
1652 hctx
= q
->mq_ops
->map_queue(q
, i
);
1653 cpumask_set_cpu(i
, hctx
->cpumask
);
1657 * Set local node, IFF we have more than one hw queue. If
1658 * not, we remain on the home node of the device
1660 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1661 hctx
->numa_node
= cpu_to_node(i
);
1665 static void blk_mq_map_swqueue(struct request_queue
*q
)
1668 struct blk_mq_hw_ctx
*hctx
;
1669 struct blk_mq_ctx
*ctx
;
1671 queue_for_each_hw_ctx(q
, hctx
, i
) {
1672 cpumask_clear(hctx
->cpumask
);
1677 * Map software to hardware queues
1679 queue_for_each_ctx(q
, ctx
, i
) {
1680 /* If the cpu isn't online, the cpu is mapped to first hctx */
1684 hctx
= q
->mq_ops
->map_queue(q
, i
);
1685 cpumask_set_cpu(i
, hctx
->cpumask
);
1686 ctx
->index_hw
= hctx
->nr_ctx
;
1687 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
1690 queue_for_each_hw_ctx(q
, hctx
, i
) {
1692 * If no software queues are mapped to this hardware queue,
1693 * disable it and free the request entries.
1695 if (!hctx
->nr_ctx
) {
1696 struct blk_mq_tag_set
*set
= q
->tag_set
;
1699 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1700 set
->tags
[i
] = NULL
;
1707 * Initialize batch roundrobin counts
1709 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
1710 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1714 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
)
1716 struct blk_mq_hw_ctx
*hctx
;
1717 struct request_queue
*q
;
1721 if (set
->tag_list
.next
== set
->tag_list
.prev
)
1726 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
1727 blk_mq_freeze_queue(q
);
1729 queue_for_each_hw_ctx(q
, hctx
, i
) {
1731 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
1733 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
1735 blk_mq_unfreeze_queue(q
);
1739 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
1741 struct blk_mq_tag_set
*set
= q
->tag_set
;
1743 mutex_lock(&set
->tag_list_lock
);
1744 list_del_init(&q
->tag_set_list
);
1745 blk_mq_update_tag_set_depth(set
);
1746 mutex_unlock(&set
->tag_list_lock
);
1749 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
1750 struct request_queue
*q
)
1754 mutex_lock(&set
->tag_list_lock
);
1755 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
1756 blk_mq_update_tag_set_depth(set
);
1757 mutex_unlock(&set
->tag_list_lock
);
1760 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
1762 struct blk_mq_hw_ctx
**hctxs
;
1763 struct blk_mq_ctx __percpu
*ctx
;
1764 struct request_queue
*q
;
1768 ctx
= alloc_percpu(struct blk_mq_ctx
);
1770 return ERR_PTR(-ENOMEM
);
1772 hctxs
= kmalloc_node(set
->nr_hw_queues
* sizeof(*hctxs
), GFP_KERNEL
,
1778 map
= blk_mq_make_queue_map(set
);
1782 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1783 int node
= blk_mq_hw_queue_to_node(map
, i
);
1785 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
1790 if (!zalloc_cpumask_var(&hctxs
[i
]->cpumask
, GFP_KERNEL
))
1793 atomic_set(&hctxs
[i
]->nr_active
, 0);
1794 hctxs
[i
]->numa_node
= node
;
1795 hctxs
[i
]->queue_num
= i
;
1798 q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
1802 if (percpu_ref_init(&q
->mq_usage_counter
, blk_mq_usage_counter_release
))
1805 setup_timer(&q
->timeout
, blk_mq_rq_timer
, (unsigned long) q
);
1806 blk_queue_rq_timeout(q
, 30000);
1808 q
->nr_queues
= nr_cpu_ids
;
1809 q
->nr_hw_queues
= set
->nr_hw_queues
;
1813 q
->queue_hw_ctx
= hctxs
;
1815 q
->mq_ops
= set
->ops
;
1816 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
1818 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
1819 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
1821 q
->sg_reserved_size
= INT_MAX
;
1823 INIT_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
1824 INIT_LIST_HEAD(&q
->requeue_list
);
1825 spin_lock_init(&q
->requeue_lock
);
1827 if (q
->nr_hw_queues
> 1)
1828 blk_queue_make_request(q
, blk_mq_make_request
);
1830 blk_queue_make_request(q
, blk_sq_make_request
);
1832 blk_queue_rq_timed_out(q
, blk_mq_rq_timed_out
);
1834 blk_queue_rq_timeout(q
, set
->timeout
);
1837 * Do this after blk_queue_make_request() overrides it...
1839 q
->nr_requests
= set
->queue_depth
;
1841 if (set
->ops
->complete
)
1842 blk_queue_softirq_done(q
, set
->ops
->complete
);
1844 blk_mq_init_flush(q
);
1845 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
1847 q
->flush_rq
= kzalloc(round_up(sizeof(struct request
) +
1848 set
->cmd_size
, cache_line_size()),
1853 if (blk_mq_init_hw_queues(q
, set
))
1856 mutex_lock(&all_q_mutex
);
1857 list_add_tail(&q
->all_q_node
, &all_q_list
);
1858 mutex_unlock(&all_q_mutex
);
1860 blk_mq_add_queue_tag_set(set
, q
);
1862 blk_mq_map_swqueue(q
);
1869 blk_cleanup_queue(q
);
1872 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1875 free_cpumask_var(hctxs
[i
]->cpumask
);
1882 return ERR_PTR(-ENOMEM
);
1884 EXPORT_SYMBOL(blk_mq_init_queue
);
1886 void blk_mq_free_queue(struct request_queue
*q
)
1888 struct blk_mq_tag_set
*set
= q
->tag_set
;
1890 blk_mq_del_queue_tag_set(q
);
1892 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
1893 blk_mq_free_hw_queues(q
, set
);
1895 percpu_ref_exit(&q
->mq_usage_counter
);
1897 free_percpu(q
->queue_ctx
);
1898 kfree(q
->queue_hw_ctx
);
1901 q
->queue_ctx
= NULL
;
1902 q
->queue_hw_ctx
= NULL
;
1905 mutex_lock(&all_q_mutex
);
1906 list_del_init(&q
->all_q_node
);
1907 mutex_unlock(&all_q_mutex
);
1910 /* Basically redo blk_mq_init_queue with queue frozen */
1911 static void blk_mq_queue_reinit(struct request_queue
*q
)
1913 blk_mq_freeze_queue(q
);
1915 blk_mq_sysfs_unregister(q
);
1917 blk_mq_update_queue_map(q
->mq_map
, q
->nr_hw_queues
);
1920 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1921 * we should change hctx numa_node according to new topology (this
1922 * involves free and re-allocate memory, worthy doing?)
1925 blk_mq_map_swqueue(q
);
1927 blk_mq_sysfs_register(q
);
1929 blk_mq_unfreeze_queue(q
);
1932 static int blk_mq_queue_reinit_notify(struct notifier_block
*nb
,
1933 unsigned long action
, void *hcpu
)
1935 struct request_queue
*q
;
1938 * Before new mappings are established, hotadded cpu might already
1939 * start handling requests. This doesn't break anything as we map
1940 * offline CPUs to first hardware queue. We will re-init the queue
1941 * below to get optimal settings.
1943 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
&&
1944 action
!= CPU_ONLINE
&& action
!= CPU_ONLINE_FROZEN
)
1947 mutex_lock(&all_q_mutex
);
1948 list_for_each_entry(q
, &all_q_list
, all_q_node
)
1949 blk_mq_queue_reinit(q
);
1950 mutex_unlock(&all_q_mutex
);
1954 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
1958 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1959 set
->tags
[i
] = blk_mq_init_rq_map(set
, i
);
1968 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1974 * Allocate the request maps associated with this tag_set. Note that this
1975 * may reduce the depth asked for, if memory is tight. set->queue_depth
1976 * will be updated to reflect the allocated depth.
1978 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
1983 depth
= set
->queue_depth
;
1985 err
= __blk_mq_alloc_rq_maps(set
);
1989 set
->queue_depth
>>= 1;
1990 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
1994 } while (set
->queue_depth
);
1996 if (!set
->queue_depth
|| err
) {
1997 pr_err("blk-mq: failed to allocate request map\n");
2001 if (depth
!= set
->queue_depth
)
2002 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2003 depth
, set
->queue_depth
);
2009 * Alloc a tag set to be associated with one or more request queues.
2010 * May fail with EINVAL for various error conditions. May adjust the
2011 * requested depth down, if if it too large. In that case, the set
2012 * value will be stored in set->queue_depth.
2014 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2016 if (!set
->nr_hw_queues
)
2018 if (!set
->queue_depth
)
2020 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2023 if (!set
->nr_hw_queues
|| !set
->ops
->queue_rq
|| !set
->ops
->map_queue
)
2026 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2027 pr_info("blk-mq: reduced tag depth to %u\n",
2029 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2032 set
->tags
= kmalloc_node(set
->nr_hw_queues
*
2033 sizeof(struct blk_mq_tags
*),
2034 GFP_KERNEL
, set
->numa_node
);
2038 if (blk_mq_alloc_rq_maps(set
))
2041 mutex_init(&set
->tag_list_lock
);
2042 INIT_LIST_HEAD(&set
->tag_list
);
2050 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2052 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2056 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2058 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2064 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2066 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2068 struct blk_mq_tag_set
*set
= q
->tag_set
;
2069 struct blk_mq_hw_ctx
*hctx
;
2072 if (!set
|| nr
> set
->queue_depth
)
2076 queue_for_each_hw_ctx(q
, hctx
, i
) {
2077 ret
= blk_mq_tag_update_depth(hctx
->tags
, nr
);
2083 q
->nr_requests
= nr
;
2088 void blk_mq_disable_hotplug(void)
2090 mutex_lock(&all_q_mutex
);
2093 void blk_mq_enable_hotplug(void)
2095 mutex_unlock(&all_q_mutex
);
2098 static int __init
blk_mq_init(void)
2102 hotcpu_notifier(blk_mq_queue_reinit_notify
, 0);
2106 subsys_initcall(blk_mq_init
);