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
->ctx_map
.map_size
; i
++)
60 if (hctx
->ctx_map
.map
[i
].word
)
66 static inline struct blk_align_bitmap
*get_bm(struct blk_mq_hw_ctx
*hctx
,
67 struct blk_mq_ctx
*ctx
)
69 return &hctx
->ctx_map
.map
[ctx
->index_hw
/ hctx
->ctx_map
.bits_per_word
];
72 #define CTX_TO_BIT(hctx, ctx) \
73 ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
76 * Mark this ctx as having pending work in this hardware queue
78 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
79 struct blk_mq_ctx
*ctx
)
81 struct blk_align_bitmap
*bm
= get_bm(hctx
, ctx
);
83 if (!test_bit(CTX_TO_BIT(hctx
, ctx
), &bm
->word
))
84 set_bit(CTX_TO_BIT(hctx
, ctx
), &bm
->word
);
87 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
88 struct blk_mq_ctx
*ctx
)
90 struct blk_align_bitmap
*bm
= get_bm(hctx
, ctx
);
92 clear_bit(CTX_TO_BIT(hctx
, ctx
), &bm
->word
);
95 static struct request
*__blk_mq_alloc_request(struct blk_mq_hw_ctx
*hctx
,
96 struct blk_mq_ctx
*ctx
,
97 gfp_t gfp
, bool reserved
)
102 tag
= blk_mq_get_tag(hctx
, &ctx
->last_tag
, gfp
, reserved
);
103 if (tag
!= BLK_MQ_TAG_FAIL
) {
104 rq
= hctx
->tags
->rqs
[tag
];
107 if (blk_mq_tag_busy(hctx
)) {
108 rq
->cmd_flags
= REQ_MQ_INFLIGHT
;
109 atomic_inc(&hctx
->nr_active
);
119 static int blk_mq_queue_enter(struct request_queue
*q
)
123 __percpu_counter_add(&q
->mq_usage_counter
, 1, 1000000);
125 /* we have problems to freeze the queue if it's initializing */
126 if (!blk_queue_bypass(q
) || !blk_queue_init_done(q
))
129 __percpu_counter_add(&q
->mq_usage_counter
, -1, 1000000);
131 spin_lock_irq(q
->queue_lock
);
132 ret
= wait_event_interruptible_lock_irq(q
->mq_freeze_wq
,
133 !blk_queue_bypass(q
) || blk_queue_dying(q
),
135 /* inc usage with lock hold to avoid freeze_queue runs here */
136 if (!ret
&& !blk_queue_dying(q
))
137 __percpu_counter_add(&q
->mq_usage_counter
, 1, 1000000);
138 else if (blk_queue_dying(q
))
140 spin_unlock_irq(q
->queue_lock
);
145 static void blk_mq_queue_exit(struct request_queue
*q
)
147 __percpu_counter_add(&q
->mq_usage_counter
, -1, 1000000);
150 static void __blk_mq_drain_queue(struct request_queue
*q
)
155 spin_lock_irq(q
->queue_lock
);
156 count
= percpu_counter_sum(&q
->mq_usage_counter
);
157 spin_unlock_irq(q
->queue_lock
);
161 blk_mq_run_queues(q
, false);
167 * Guarantee no request is in use, so we can change any data structure of
168 * the queue afterward.
170 static void blk_mq_freeze_queue(struct request_queue
*q
)
174 spin_lock_irq(q
->queue_lock
);
175 drain
= !q
->bypass_depth
++;
176 queue_flag_set(QUEUE_FLAG_BYPASS
, q
);
177 spin_unlock_irq(q
->queue_lock
);
180 __blk_mq_drain_queue(q
);
183 void blk_mq_drain_queue(struct request_queue
*q
)
185 __blk_mq_drain_queue(q
);
188 static void blk_mq_unfreeze_queue(struct request_queue
*q
)
192 spin_lock_irq(q
->queue_lock
);
193 if (!--q
->bypass_depth
) {
194 queue_flag_clear(QUEUE_FLAG_BYPASS
, q
);
197 WARN_ON_ONCE(q
->bypass_depth
< 0);
198 spin_unlock_irq(q
->queue_lock
);
200 wake_up_all(&q
->mq_freeze_wq
);
203 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
205 return blk_mq_has_free_tags(hctx
->tags
);
207 EXPORT_SYMBOL(blk_mq_can_queue
);
209 static void blk_mq_rq_ctx_init(struct request_queue
*q
, struct blk_mq_ctx
*ctx
,
210 struct request
*rq
, unsigned int rw_flags
)
212 if (blk_queue_io_stat(q
))
213 rw_flags
|= REQ_IO_STAT
;
215 INIT_LIST_HEAD(&rq
->queuelist
);
216 /* csd/requeue_work/fifo_time is initialized before use */
219 rq
->cmd_flags
|= rw_flags
;
221 /* do not touch atomic flags, it needs atomic ops against the timer */
224 rq
->__sector
= (sector_t
) -1;
227 INIT_HLIST_NODE(&rq
->hash
);
228 RB_CLEAR_NODE(&rq
->rb_node
);
229 memset(&rq
->flush
, 0, max(sizeof(rq
->flush
), sizeof(rq
->elv
)));
232 rq
->start_time
= jiffies
;
233 #ifdef CONFIG_BLK_CGROUP
235 set_start_time_ns(rq
);
236 rq
->io_start_time_ns
= 0;
238 rq
->nr_phys_segments
= 0;
239 #if defined(CONFIG_BLK_DEV_INTEGRITY)
240 rq
->nr_integrity_segments
= 0;
244 /* tag was already set */
246 memset(rq
->__cmd
, 0, sizeof(rq
->__cmd
));
248 rq
->cmd_len
= BLK_MAX_CDB
;
256 INIT_LIST_HEAD(&rq
->timeout_list
);
260 rq
->end_io_data
= NULL
;
263 ctx
->rq_dispatched
[rw_is_sync(rw_flags
)]++;
266 static struct request
*blk_mq_alloc_request_pinned(struct request_queue
*q
,
273 struct blk_mq_ctx
*ctx
= blk_mq_get_ctx(q
);
274 struct blk_mq_hw_ctx
*hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
276 rq
= __blk_mq_alloc_request(hctx
, ctx
, gfp
& ~__GFP_WAIT
,
279 blk_mq_rq_ctx_init(q
, ctx
, rq
, rw
);
283 if (gfp
& __GFP_WAIT
) {
284 __blk_mq_run_hw_queue(hctx
);
291 blk_mq_wait_for_tags(hctx
, reserved
);
297 struct request
*blk_mq_alloc_request(struct request_queue
*q
, int rw
, gfp_t gfp
)
301 if (blk_mq_queue_enter(q
))
304 rq
= blk_mq_alloc_request_pinned(q
, rw
, gfp
, false);
306 blk_mq_put_ctx(rq
->mq_ctx
);
309 EXPORT_SYMBOL(blk_mq_alloc_request
);
311 struct request
*blk_mq_alloc_reserved_request(struct request_queue
*q
, int rw
,
316 if (blk_mq_queue_enter(q
))
319 rq
= blk_mq_alloc_request_pinned(q
, rw
, gfp
, true);
321 blk_mq_put_ctx(rq
->mq_ctx
);
324 EXPORT_SYMBOL(blk_mq_alloc_reserved_request
);
326 static void __blk_mq_free_request(struct blk_mq_hw_ctx
*hctx
,
327 struct blk_mq_ctx
*ctx
, struct request
*rq
)
329 const int tag
= rq
->tag
;
330 struct request_queue
*q
= rq
->q
;
332 if (rq
->cmd_flags
& REQ_MQ_INFLIGHT
)
333 atomic_dec(&hctx
->nr_active
);
335 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
336 blk_mq_put_tag(hctx
, tag
, &ctx
->last_tag
);
337 blk_mq_queue_exit(q
);
340 void blk_mq_free_request(struct request
*rq
)
342 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
343 struct blk_mq_hw_ctx
*hctx
;
344 struct request_queue
*q
= rq
->q
;
346 ctx
->rq_completed
[rq_is_sync(rq
)]++;
348 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
349 __blk_mq_free_request(hctx
, ctx
, rq
);
353 * Clone all relevant state from a request that has been put on hold in
354 * the flush state machine into the preallocated flush request that hangs
355 * off the request queue.
357 * For a driver the flush request should be invisible, that's why we are
358 * impersonating the original request here.
360 void blk_mq_clone_flush_request(struct request
*flush_rq
,
361 struct request
*orig_rq
)
363 struct blk_mq_hw_ctx
*hctx
=
364 orig_rq
->q
->mq_ops
->map_queue(orig_rq
->q
, orig_rq
->mq_ctx
->cpu
);
366 flush_rq
->mq_ctx
= orig_rq
->mq_ctx
;
367 flush_rq
->tag
= orig_rq
->tag
;
368 memcpy(blk_mq_rq_to_pdu(flush_rq
), blk_mq_rq_to_pdu(orig_rq
),
372 inline void __blk_mq_end_io(struct request
*rq
, int error
)
374 blk_account_io_done(rq
);
377 rq
->end_io(rq
, error
);
379 if (unlikely(blk_bidi_rq(rq
)))
380 blk_mq_free_request(rq
->next_rq
);
381 blk_mq_free_request(rq
);
384 EXPORT_SYMBOL(__blk_mq_end_io
);
386 void blk_mq_end_io(struct request
*rq
, int error
)
388 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
390 __blk_mq_end_io(rq
, error
);
392 EXPORT_SYMBOL(blk_mq_end_io
);
394 static void __blk_mq_complete_request_remote(void *data
)
396 struct request
*rq
= data
;
398 rq
->q
->softirq_done_fn(rq
);
401 void __blk_mq_complete_request(struct request
*rq
)
403 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
407 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
408 rq
->q
->softirq_done_fn(rq
);
413 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
414 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
416 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
417 rq
->csd
.func
= __blk_mq_complete_request_remote
;
420 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
422 rq
->q
->softirq_done_fn(rq
);
428 * blk_mq_complete_request - end I/O on a request
429 * @rq: the request being processed
432 * Ends all I/O on a request. It does not handle partial completions.
433 * The actual completion happens out-of-order, through a IPI handler.
435 void blk_mq_complete_request(struct request
*rq
)
437 if (unlikely(blk_should_fake_timeout(rq
->q
)))
439 if (!blk_mark_rq_complete(rq
))
440 __blk_mq_complete_request(rq
);
442 EXPORT_SYMBOL(blk_mq_complete_request
);
444 static void blk_mq_start_request(struct request
*rq
, bool last
)
446 struct request_queue
*q
= rq
->q
;
448 trace_block_rq_issue(q
, rq
);
450 rq
->resid_len
= blk_rq_bytes(rq
);
451 if (unlikely(blk_bidi_rq(rq
)))
452 rq
->next_rq
->resid_len
= blk_rq_bytes(rq
->next_rq
);
455 * Just mark start time and set the started bit. Due to memory
456 * ordering, we know we'll see the correct deadline as long as
457 * REQ_ATOMIC_STARTED is seen.
459 rq
->deadline
= jiffies
+ q
->rq_timeout
;
462 * Mark us as started and clear complete. Complete might have been
463 * set if requeue raced with timeout, which then marked it as
464 * complete. So be sure to clear complete again when we start
465 * the request, otherwise we'll ignore the completion event.
467 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
468 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
470 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
472 * Make sure space for the drain appears. We know we can do
473 * this because max_hw_segments has been adjusted to be one
474 * fewer than the device can handle.
476 rq
->nr_phys_segments
++;
480 * Flag the last request in the series so that drivers know when IO
481 * should be kicked off, if they don't do it on a per-request basis.
483 * Note: the flag isn't the only condition drivers should do kick off.
484 * If drive is busy, the last request might not have the bit set.
487 rq
->cmd_flags
|= REQ_END
;
490 static void __blk_mq_requeue_request(struct request
*rq
)
492 struct request_queue
*q
= rq
->q
;
494 trace_block_rq_requeue(q
, rq
);
495 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
497 rq
->cmd_flags
&= ~REQ_END
;
499 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
500 rq
->nr_phys_segments
--;
503 void blk_mq_requeue_request(struct request
*rq
)
505 __blk_mq_requeue_request(rq
);
506 blk_clear_rq_complete(rq
);
508 BUG_ON(blk_queued_rq(rq
));
509 blk_mq_insert_request(rq
, true, true, false);
511 EXPORT_SYMBOL(blk_mq_requeue_request
);
513 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
515 return tags
->rqs
[tag
];
517 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
519 struct blk_mq_timeout_data
{
520 struct blk_mq_hw_ctx
*hctx
;
522 unsigned int *next_set
;
525 static void blk_mq_timeout_check(void *__data
, unsigned long *free_tags
)
527 struct blk_mq_timeout_data
*data
= __data
;
528 struct blk_mq_hw_ctx
*hctx
= data
->hctx
;
531 /* It may not be in flight yet (this is where
532 * the REQ_ATOMIC_STARTED flag comes in). The requests are
533 * statically allocated, so we know it's always safe to access the
534 * memory associated with a bit offset into ->rqs[].
540 tag
= find_next_zero_bit(free_tags
, hctx
->tags
->nr_tags
, tag
);
541 if (tag
>= hctx
->tags
->nr_tags
)
544 rq
= blk_mq_tag_to_rq(hctx
->tags
, tag
++);
545 if (rq
->q
!= hctx
->queue
)
547 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
550 blk_rq_check_expired(rq
, data
->next
, data
->next_set
);
554 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx
*hctx
,
556 unsigned int *next_set
)
558 struct blk_mq_timeout_data data
= {
561 .next_set
= next_set
,
565 * Ask the tagging code to iterate busy requests, so we can
566 * check them for timeout.
568 blk_mq_tag_busy_iter(hctx
->tags
, blk_mq_timeout_check
, &data
);
571 static enum blk_eh_timer_return
blk_mq_rq_timed_out(struct request
*rq
)
573 struct request_queue
*q
= rq
->q
;
576 * We know that complete is set at this point. If STARTED isn't set
577 * anymore, then the request isn't active and the "timeout" should
578 * just be ignored. This can happen due to the bitflag ordering.
579 * Timeout first checks if STARTED is set, and if it is, assumes
580 * the request is active. But if we race with completion, then
581 * we both flags will get cleared. So check here again, and ignore
582 * a timeout event with a request that isn't active.
584 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
585 return BLK_EH_NOT_HANDLED
;
587 if (!q
->mq_ops
->timeout
)
588 return BLK_EH_RESET_TIMER
;
590 return q
->mq_ops
->timeout(rq
);
593 static void blk_mq_rq_timer(unsigned long data
)
595 struct request_queue
*q
= (struct request_queue
*) data
;
596 struct blk_mq_hw_ctx
*hctx
;
597 unsigned long next
= 0;
600 queue_for_each_hw_ctx(q
, hctx
, i
) {
602 * If not software queues are currently mapped to this
603 * hardware queue, there's nothing to check
605 if (!hctx
->nr_ctx
|| !hctx
->tags
)
608 blk_mq_hw_ctx_check_timeout(hctx
, &next
, &next_set
);
612 next
= blk_rq_timeout(round_jiffies_up(next
));
613 mod_timer(&q
->timeout
, next
);
615 queue_for_each_hw_ctx(q
, hctx
, i
)
616 blk_mq_tag_idle(hctx
);
621 * Reverse check our software queue for entries that we could potentially
622 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
623 * too much time checking for merges.
625 static bool blk_mq_attempt_merge(struct request_queue
*q
,
626 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
631 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
637 if (!blk_rq_merge_ok(rq
, bio
))
640 el_ret
= blk_try_merge(rq
, bio
);
641 if (el_ret
== ELEVATOR_BACK_MERGE
) {
642 if (bio_attempt_back_merge(q
, rq
, bio
)) {
647 } else if (el_ret
== ELEVATOR_FRONT_MERGE
) {
648 if (bio_attempt_front_merge(q
, rq
, bio
)) {
660 * Process software queues that have been marked busy, splicing them
661 * to the for-dispatch
663 static void flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
665 struct blk_mq_ctx
*ctx
;
668 for (i
= 0; i
< hctx
->ctx_map
.map_size
; i
++) {
669 struct blk_align_bitmap
*bm
= &hctx
->ctx_map
.map
[i
];
670 unsigned int off
, bit
;
676 off
= i
* hctx
->ctx_map
.bits_per_word
;
678 bit
= find_next_bit(&bm
->word
, bm
->depth
, bit
);
679 if (bit
>= bm
->depth
)
682 ctx
= hctx
->ctxs
[bit
+ off
];
683 clear_bit(bit
, &bm
->word
);
684 spin_lock(&ctx
->lock
);
685 list_splice_tail_init(&ctx
->rq_list
, list
);
686 spin_unlock(&ctx
->lock
);
694 * Run this hardware queue, pulling any software queues mapped to it in.
695 * Note that this function currently has various problems around ordering
696 * of IO. In particular, we'd like FIFO behaviour on handling existing
697 * items on the hctx->dispatch list. Ignore that for now.
699 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
701 struct request_queue
*q
= hctx
->queue
;
706 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
));
708 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)))
714 * Touch any software queue that has pending entries.
716 flush_busy_ctxs(hctx
, &rq_list
);
719 * If we have previous entries on our dispatch list, grab them
720 * and stuff them at the front for more fair dispatch.
722 if (!list_empty_careful(&hctx
->dispatch
)) {
723 spin_lock(&hctx
->lock
);
724 if (!list_empty(&hctx
->dispatch
))
725 list_splice_init(&hctx
->dispatch
, &rq_list
);
726 spin_unlock(&hctx
->lock
);
730 * Now process all the entries, sending them to the driver.
733 while (!list_empty(&rq_list
)) {
736 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
737 list_del_init(&rq
->queuelist
);
739 blk_mq_start_request(rq
, list_empty(&rq_list
));
741 ret
= q
->mq_ops
->queue_rq(hctx
, rq
);
743 case BLK_MQ_RQ_QUEUE_OK
:
746 case BLK_MQ_RQ_QUEUE_BUSY
:
747 list_add(&rq
->queuelist
, &rq_list
);
748 __blk_mq_requeue_request(rq
);
751 pr_err("blk-mq: bad return on queue: %d\n", ret
);
752 case BLK_MQ_RQ_QUEUE_ERROR
:
754 blk_mq_end_io(rq
, rq
->errors
);
758 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
763 hctx
->dispatched
[0]++;
764 else if (queued
< (1 << (BLK_MQ_MAX_DISPATCH_ORDER
- 1)))
765 hctx
->dispatched
[ilog2(queued
) + 1]++;
768 * Any items that need requeuing? Stuff them into hctx->dispatch,
769 * that is where we will continue on next queue run.
771 if (!list_empty(&rq_list
)) {
772 spin_lock(&hctx
->lock
);
773 list_splice(&rq_list
, &hctx
->dispatch
);
774 spin_unlock(&hctx
->lock
);
779 * It'd be great if the workqueue API had a way to pass
780 * in a mask and had some smarts for more clever placement.
781 * For now we just round-robin here, switching for every
782 * BLK_MQ_CPU_WORK_BATCH queued items.
784 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
786 int cpu
= hctx
->next_cpu
;
788 if (--hctx
->next_cpu_batch
<= 0) {
791 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
792 if (next_cpu
>= nr_cpu_ids
)
793 next_cpu
= cpumask_first(hctx
->cpumask
);
795 hctx
->next_cpu
= next_cpu
;
796 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
802 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
804 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)))
807 if (!async
&& cpumask_test_cpu(smp_processor_id(), hctx
->cpumask
))
808 __blk_mq_run_hw_queue(hctx
);
809 else if (hctx
->queue
->nr_hw_queues
== 1)
810 kblockd_schedule_delayed_work(&hctx
->run_work
, 0);
814 cpu
= blk_mq_hctx_next_cpu(hctx
);
815 kblockd_schedule_delayed_work_on(cpu
, &hctx
->run_work
, 0);
819 void blk_mq_run_queues(struct request_queue
*q
, bool async
)
821 struct blk_mq_hw_ctx
*hctx
;
824 queue_for_each_hw_ctx(q
, hctx
, i
) {
825 if ((!blk_mq_hctx_has_pending(hctx
) &&
826 list_empty_careful(&hctx
->dispatch
)) ||
827 test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
831 blk_mq_run_hw_queue(hctx
, async
);
835 EXPORT_SYMBOL(blk_mq_run_queues
);
837 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
839 cancel_delayed_work(&hctx
->run_work
);
840 cancel_delayed_work(&hctx
->delay_work
);
841 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
843 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
845 void blk_mq_stop_hw_queues(struct request_queue
*q
)
847 struct blk_mq_hw_ctx
*hctx
;
850 queue_for_each_hw_ctx(q
, hctx
, i
)
851 blk_mq_stop_hw_queue(hctx
);
853 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
855 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
857 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
860 __blk_mq_run_hw_queue(hctx
);
863 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
865 void blk_mq_start_hw_queues(struct request_queue
*q
)
867 struct blk_mq_hw_ctx
*hctx
;
870 queue_for_each_hw_ctx(q
, hctx
, i
)
871 blk_mq_start_hw_queue(hctx
);
873 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
876 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
878 struct blk_mq_hw_ctx
*hctx
;
881 queue_for_each_hw_ctx(q
, hctx
, i
) {
882 if (!test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
885 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
887 blk_mq_run_hw_queue(hctx
, async
);
891 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
893 static void blk_mq_run_work_fn(struct work_struct
*work
)
895 struct blk_mq_hw_ctx
*hctx
;
897 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
899 __blk_mq_run_hw_queue(hctx
);
902 static void blk_mq_delay_work_fn(struct work_struct
*work
)
904 struct blk_mq_hw_ctx
*hctx
;
906 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
908 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
909 __blk_mq_run_hw_queue(hctx
);
912 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
914 unsigned long tmo
= msecs_to_jiffies(msecs
);
916 if (hctx
->queue
->nr_hw_queues
== 1)
917 kblockd_schedule_delayed_work(&hctx
->delay_work
, tmo
);
921 cpu
= blk_mq_hctx_next_cpu(hctx
);
922 kblockd_schedule_delayed_work_on(cpu
, &hctx
->delay_work
, tmo
);
925 EXPORT_SYMBOL(blk_mq_delay_queue
);
927 static void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
,
928 struct request
*rq
, bool at_head
)
930 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
932 trace_block_rq_insert(hctx
->queue
, rq
);
935 list_add(&rq
->queuelist
, &ctx
->rq_list
);
937 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
939 blk_mq_hctx_mark_pending(hctx
, ctx
);
942 * We do this early, to ensure we are on the right CPU.
947 void blk_mq_insert_request(struct request
*rq
, bool at_head
, bool run_queue
,
950 struct request_queue
*q
= rq
->q
;
951 struct blk_mq_hw_ctx
*hctx
;
952 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
, *current_ctx
;
954 current_ctx
= blk_mq_get_ctx(q
);
955 if (!cpu_online(ctx
->cpu
))
956 rq
->mq_ctx
= ctx
= current_ctx
;
958 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
960 if (rq
->cmd_flags
& (REQ_FLUSH
| REQ_FUA
) &&
961 !(rq
->cmd_flags
& (REQ_FLUSH_SEQ
))) {
962 blk_insert_flush(rq
);
964 spin_lock(&ctx
->lock
);
965 __blk_mq_insert_request(hctx
, rq
, at_head
);
966 spin_unlock(&ctx
->lock
);
970 blk_mq_run_hw_queue(hctx
, async
);
972 blk_mq_put_ctx(current_ctx
);
975 static void blk_mq_insert_requests(struct request_queue
*q
,
976 struct blk_mq_ctx
*ctx
,
977 struct list_head
*list
,
982 struct blk_mq_hw_ctx
*hctx
;
983 struct blk_mq_ctx
*current_ctx
;
985 trace_block_unplug(q
, depth
, !from_schedule
);
987 current_ctx
= blk_mq_get_ctx(q
);
989 if (!cpu_online(ctx
->cpu
))
991 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
994 * preemption doesn't flush plug list, so it's possible ctx->cpu is
997 spin_lock(&ctx
->lock
);
998 while (!list_empty(list
)) {
1001 rq
= list_first_entry(list
, struct request
, queuelist
);
1002 list_del_init(&rq
->queuelist
);
1004 __blk_mq_insert_request(hctx
, rq
, false);
1006 spin_unlock(&ctx
->lock
);
1008 blk_mq_run_hw_queue(hctx
, from_schedule
);
1009 blk_mq_put_ctx(current_ctx
);
1012 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1014 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1015 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1017 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1018 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1019 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1022 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1024 struct blk_mq_ctx
*this_ctx
;
1025 struct request_queue
*this_q
;
1028 LIST_HEAD(ctx_list
);
1031 list_splice_init(&plug
->mq_list
, &list
);
1033 list_sort(NULL
, &list
, plug_ctx_cmp
);
1039 while (!list_empty(&list
)) {
1040 rq
= list_entry_rq(list
.next
);
1041 list_del_init(&rq
->queuelist
);
1043 if (rq
->mq_ctx
!= this_ctx
) {
1045 blk_mq_insert_requests(this_q
, this_ctx
,
1050 this_ctx
= rq
->mq_ctx
;
1056 list_add_tail(&rq
->queuelist
, &ctx_list
);
1060 * If 'this_ctx' is set, we know we have entries to complete
1061 * on 'ctx_list'. Do those.
1064 blk_mq_insert_requests(this_q
, this_ctx
, &ctx_list
, depth
,
1069 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1071 init_request_from_bio(rq
, bio
);
1072 blk_account_io_start(rq
, 1);
1075 static void blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1077 struct blk_mq_hw_ctx
*hctx
;
1078 struct blk_mq_ctx
*ctx
;
1079 const int is_sync
= rw_is_sync(bio
->bi_rw
);
1080 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
1081 int rw
= bio_data_dir(bio
);
1083 unsigned int use_plug
, request_count
= 0;
1086 * If we have multiple hardware queues, just go directly to
1087 * one of those for sync IO.
1089 use_plug
= !is_flush_fua
&& ((q
->nr_hw_queues
== 1) || !is_sync
);
1091 blk_queue_bounce(q
, &bio
);
1093 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1094 bio_endio(bio
, -EIO
);
1098 if (use_plug
&& !blk_queue_nomerges(q
) &&
1099 blk_attempt_plug_merge(q
, bio
, &request_count
))
1102 if (blk_mq_queue_enter(q
)) {
1103 bio_endio(bio
, -EIO
);
1107 ctx
= blk_mq_get_ctx(q
);
1108 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1112 trace_block_getrq(q
, bio
, rw
);
1113 rq
= __blk_mq_alloc_request(hctx
, ctx
, GFP_ATOMIC
, false);
1115 blk_mq_rq_ctx_init(q
, ctx
, rq
, rw
);
1117 blk_mq_put_ctx(ctx
);
1118 trace_block_sleeprq(q
, bio
, rw
);
1119 rq
= blk_mq_alloc_request_pinned(q
, rw
, __GFP_WAIT
|GFP_ATOMIC
,
1122 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1127 if (unlikely(is_flush_fua
)) {
1128 blk_mq_bio_to_request(rq
, bio
);
1129 blk_insert_flush(rq
);
1134 * A task plug currently exists. Since this is completely lockless,
1135 * utilize that to temporarily store requests until the task is
1136 * either done or scheduled away.
1139 struct blk_plug
*plug
= current
->plug
;
1142 blk_mq_bio_to_request(rq
, bio
);
1143 if (list_empty(&plug
->mq_list
))
1144 trace_block_plug(q
);
1145 else if (request_count
>= BLK_MAX_REQUEST_COUNT
) {
1146 blk_flush_plug_list(plug
, false);
1147 trace_block_plug(q
);
1149 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1150 blk_mq_put_ctx(ctx
);
1155 if (!(hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
)) {
1156 blk_mq_bio_to_request(rq
, bio
);
1157 spin_lock(&ctx
->lock
);
1159 __blk_mq_insert_request(hctx
, rq
, false);
1160 spin_unlock(&ctx
->lock
);
1162 spin_lock(&ctx
->lock
);
1163 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1164 blk_mq_bio_to_request(rq
, bio
);
1168 spin_unlock(&ctx
->lock
);
1169 __blk_mq_free_request(hctx
, ctx
, rq
);
1174 * For a SYNC request, send it to the hardware immediately. For an
1175 * ASYNC request, just ensure that we run it later on. The latter
1176 * allows for merging opportunities and more efficient dispatching.
1179 blk_mq_run_hw_queue(hctx
, !is_sync
|| is_flush_fua
);
1180 blk_mq_put_ctx(ctx
);
1184 * Default mapping to a software queue, since we use one per CPU.
1186 struct blk_mq_hw_ctx
*blk_mq_map_queue(struct request_queue
*q
, const int cpu
)
1188 return q
->queue_hw_ctx
[q
->mq_map
[cpu
]];
1190 EXPORT_SYMBOL(blk_mq_map_queue
);
1192 struct blk_mq_hw_ctx
*blk_mq_alloc_single_hw_queue(struct blk_mq_tag_set
*set
,
1193 unsigned int hctx_index
)
1195 return kzalloc_node(sizeof(struct blk_mq_hw_ctx
), GFP_KERNEL
,
1198 EXPORT_SYMBOL(blk_mq_alloc_single_hw_queue
);
1200 void blk_mq_free_single_hw_queue(struct blk_mq_hw_ctx
*hctx
,
1201 unsigned int hctx_index
)
1205 EXPORT_SYMBOL(blk_mq_free_single_hw_queue
);
1207 static void blk_mq_free_rq_map(struct blk_mq_tag_set
*set
,
1208 struct blk_mq_tags
*tags
, unsigned int hctx_idx
)
1212 if (tags
->rqs
&& set
->ops
->exit_request
) {
1215 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1218 set
->ops
->exit_request(set
->driver_data
, tags
->rqs
[i
],
1223 while (!list_empty(&tags
->page_list
)) {
1224 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1225 list_del_init(&page
->lru
);
1226 __free_pages(page
, page
->private);
1231 blk_mq_free_tags(tags
);
1234 static size_t order_to_size(unsigned int order
)
1236 return (size_t)PAGE_SIZE
<< order
;
1239 static struct blk_mq_tags
*blk_mq_init_rq_map(struct blk_mq_tag_set
*set
,
1240 unsigned int hctx_idx
)
1242 struct blk_mq_tags
*tags
;
1243 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1244 size_t rq_size
, left
;
1246 tags
= blk_mq_init_tags(set
->queue_depth
, set
->reserved_tags
,
1251 INIT_LIST_HEAD(&tags
->page_list
);
1253 tags
->rqs
= kmalloc_node(set
->queue_depth
* sizeof(struct request
*),
1254 GFP_KERNEL
, set
->numa_node
);
1256 blk_mq_free_tags(tags
);
1261 * rq_size is the size of the request plus driver payload, rounded
1262 * to the cacheline size
1264 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1266 left
= rq_size
* set
->queue_depth
;
1268 for (i
= 0; i
< set
->queue_depth
; ) {
1269 int this_order
= max_order
;
1274 while (left
< order_to_size(this_order
- 1) && this_order
)
1278 page
= alloc_pages_node(set
->numa_node
, GFP_KERNEL
,
1284 if (order_to_size(this_order
) < rq_size
)
1291 page
->private = this_order
;
1292 list_add_tail(&page
->lru
, &tags
->page_list
);
1294 p
= page_address(page
);
1295 entries_per_page
= order_to_size(this_order
) / rq_size
;
1296 to_do
= min(entries_per_page
, set
->queue_depth
- i
);
1297 left
-= to_do
* rq_size
;
1298 for (j
= 0; j
< to_do
; j
++) {
1300 if (set
->ops
->init_request
) {
1301 if (set
->ops
->init_request(set
->driver_data
,
1302 tags
->rqs
[i
], hctx_idx
, i
,
1315 pr_warn("%s: failed to allocate requests\n", __func__
);
1316 blk_mq_free_rq_map(set
, tags
, hctx_idx
);
1320 static void blk_mq_free_bitmap(struct blk_mq_ctxmap
*bitmap
)
1325 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap
*bitmap
, int node
)
1327 unsigned int bpw
= 8, total
, num_maps
, i
;
1329 bitmap
->bits_per_word
= bpw
;
1331 num_maps
= ALIGN(nr_cpu_ids
, bpw
) / bpw
;
1332 bitmap
->map
= kzalloc_node(num_maps
* sizeof(struct blk_align_bitmap
),
1337 bitmap
->map_size
= num_maps
;
1340 for (i
= 0; i
< num_maps
; i
++) {
1341 bitmap
->map
[i
].depth
= min(total
, bitmap
->bits_per_word
);
1342 total
-= bitmap
->map
[i
].depth
;
1348 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx
*hctx
, int cpu
)
1350 struct request_queue
*q
= hctx
->queue
;
1351 struct blk_mq_ctx
*ctx
;
1355 * Move ctx entries to new CPU, if this one is going away.
1357 ctx
= __blk_mq_get_ctx(q
, cpu
);
1359 spin_lock(&ctx
->lock
);
1360 if (!list_empty(&ctx
->rq_list
)) {
1361 list_splice_init(&ctx
->rq_list
, &tmp
);
1362 blk_mq_hctx_clear_pending(hctx
, ctx
);
1364 spin_unlock(&ctx
->lock
);
1366 if (list_empty(&tmp
))
1369 ctx
= blk_mq_get_ctx(q
);
1370 spin_lock(&ctx
->lock
);
1372 while (!list_empty(&tmp
)) {
1375 rq
= list_first_entry(&tmp
, struct request
, queuelist
);
1377 list_move_tail(&rq
->queuelist
, &ctx
->rq_list
);
1380 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1381 blk_mq_hctx_mark_pending(hctx
, ctx
);
1383 spin_unlock(&ctx
->lock
);
1385 blk_mq_run_hw_queue(hctx
, true);
1386 blk_mq_put_ctx(ctx
);
1390 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx
*hctx
, int cpu
)
1392 struct request_queue
*q
= hctx
->queue
;
1393 struct blk_mq_tag_set
*set
= q
->tag_set
;
1395 if (set
->tags
[hctx
->queue_num
])
1398 set
->tags
[hctx
->queue_num
] = blk_mq_init_rq_map(set
, hctx
->queue_num
);
1399 if (!set
->tags
[hctx
->queue_num
])
1402 hctx
->tags
= set
->tags
[hctx
->queue_num
];
1406 static int blk_mq_hctx_notify(void *data
, unsigned long action
,
1409 struct blk_mq_hw_ctx
*hctx
= data
;
1411 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
)
1412 return blk_mq_hctx_cpu_offline(hctx
, cpu
);
1413 else if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
)
1414 return blk_mq_hctx_cpu_online(hctx
, cpu
);
1419 static int blk_mq_init_hw_queues(struct request_queue
*q
,
1420 struct blk_mq_tag_set
*set
)
1422 struct blk_mq_hw_ctx
*hctx
;
1426 * Initialize hardware queues
1428 queue_for_each_hw_ctx(q
, hctx
, i
) {
1431 node
= hctx
->numa_node
;
1432 if (node
== NUMA_NO_NODE
)
1433 node
= hctx
->numa_node
= set
->numa_node
;
1435 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1436 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1437 spin_lock_init(&hctx
->lock
);
1438 INIT_LIST_HEAD(&hctx
->dispatch
);
1440 hctx
->queue_num
= i
;
1441 hctx
->flags
= set
->flags
;
1442 hctx
->cmd_size
= set
->cmd_size
;
1444 blk_mq_init_cpu_notifier(&hctx
->cpu_notifier
,
1445 blk_mq_hctx_notify
, hctx
);
1446 blk_mq_register_cpu_notifier(&hctx
->cpu_notifier
);
1448 hctx
->tags
= set
->tags
[i
];
1451 * Allocate space for all possible cpus to avoid allocation in
1454 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1459 if (blk_mq_alloc_bitmap(&hctx
->ctx_map
, node
))
1464 if (set
->ops
->init_hctx
&&
1465 set
->ops
->init_hctx(hctx
, set
->driver_data
, i
))
1469 if (i
== q
->nr_hw_queues
)
1475 queue_for_each_hw_ctx(q
, hctx
, j
) {
1479 if (set
->ops
->exit_hctx
)
1480 set
->ops
->exit_hctx(hctx
, j
);
1482 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1484 blk_mq_free_bitmap(&hctx
->ctx_map
);
1490 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1491 unsigned int nr_hw_queues
)
1495 for_each_possible_cpu(i
) {
1496 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1497 struct blk_mq_hw_ctx
*hctx
;
1499 memset(__ctx
, 0, sizeof(*__ctx
));
1501 spin_lock_init(&__ctx
->lock
);
1502 INIT_LIST_HEAD(&__ctx
->rq_list
);
1505 /* If the cpu isn't online, the cpu is mapped to first hctx */
1509 hctx
= q
->mq_ops
->map_queue(q
, i
);
1510 cpumask_set_cpu(i
, hctx
->cpumask
);
1514 * Set local node, IFF we have more than one hw queue. If
1515 * not, we remain on the home node of the device
1517 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1518 hctx
->numa_node
= cpu_to_node(i
);
1522 static void blk_mq_map_swqueue(struct request_queue
*q
)
1525 struct blk_mq_hw_ctx
*hctx
;
1526 struct blk_mq_ctx
*ctx
;
1528 queue_for_each_hw_ctx(q
, hctx
, i
) {
1529 cpumask_clear(hctx
->cpumask
);
1534 * Map software to hardware queues
1536 queue_for_each_ctx(q
, ctx
, i
) {
1537 /* If the cpu isn't online, the cpu is mapped to first hctx */
1541 hctx
= q
->mq_ops
->map_queue(q
, i
);
1542 cpumask_set_cpu(i
, hctx
->cpumask
);
1543 ctx
->index_hw
= hctx
->nr_ctx
;
1544 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
1547 queue_for_each_hw_ctx(q
, hctx
, i
) {
1549 * If not software queues are mapped to this hardware queue,
1550 * disable it and free the request entries
1552 if (!hctx
->nr_ctx
) {
1553 struct blk_mq_tag_set
*set
= q
->tag_set
;
1556 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1557 set
->tags
[i
] = NULL
;
1564 * Initialize batch roundrobin counts
1566 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
1567 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1571 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
)
1573 struct blk_mq_hw_ctx
*hctx
;
1574 struct request_queue
*q
;
1578 if (set
->tag_list
.next
== set
->tag_list
.prev
)
1583 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
1584 blk_mq_freeze_queue(q
);
1586 queue_for_each_hw_ctx(q
, hctx
, i
) {
1588 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
1590 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
1592 blk_mq_unfreeze_queue(q
);
1596 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
1598 struct blk_mq_tag_set
*set
= q
->tag_set
;
1600 blk_mq_freeze_queue(q
);
1602 mutex_lock(&set
->tag_list_lock
);
1603 list_del_init(&q
->tag_set_list
);
1604 blk_mq_update_tag_set_depth(set
);
1605 mutex_unlock(&set
->tag_list_lock
);
1607 blk_mq_unfreeze_queue(q
);
1610 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
1611 struct request_queue
*q
)
1615 mutex_lock(&set
->tag_list_lock
);
1616 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
1617 blk_mq_update_tag_set_depth(set
);
1618 mutex_unlock(&set
->tag_list_lock
);
1621 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
1623 struct blk_mq_hw_ctx
**hctxs
;
1624 struct blk_mq_ctx
*ctx
;
1625 struct request_queue
*q
;
1628 ctx
= alloc_percpu(struct blk_mq_ctx
);
1630 return ERR_PTR(-ENOMEM
);
1632 hctxs
= kmalloc_node(set
->nr_hw_queues
* sizeof(*hctxs
), GFP_KERNEL
,
1638 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1639 hctxs
[i
] = set
->ops
->alloc_hctx(set
, i
);
1643 if (!zalloc_cpumask_var(&hctxs
[i
]->cpumask
, GFP_KERNEL
))
1646 atomic_set(&hctxs
[i
]->nr_active
, 0);
1647 hctxs
[i
]->numa_node
= NUMA_NO_NODE
;
1648 hctxs
[i
]->queue_num
= i
;
1651 q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
1655 q
->mq_map
= blk_mq_make_queue_map(set
);
1659 setup_timer(&q
->timeout
, blk_mq_rq_timer
, (unsigned long) q
);
1660 blk_queue_rq_timeout(q
, 30000);
1662 q
->nr_queues
= nr_cpu_ids
;
1663 q
->nr_hw_queues
= set
->nr_hw_queues
;
1666 q
->queue_hw_ctx
= hctxs
;
1668 q
->mq_ops
= set
->ops
;
1669 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
1671 q
->sg_reserved_size
= INT_MAX
;
1673 blk_queue_make_request(q
, blk_mq_make_request
);
1674 blk_queue_rq_timed_out(q
, blk_mq_rq_timed_out
);
1676 blk_queue_rq_timeout(q
, set
->timeout
);
1679 * Do this after blk_queue_make_request() overrides it...
1681 q
->nr_requests
= set
->queue_depth
;
1683 if (set
->ops
->complete
)
1684 blk_queue_softirq_done(q
, set
->ops
->complete
);
1686 blk_mq_init_flush(q
);
1687 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
1689 q
->flush_rq
= kzalloc(round_up(sizeof(struct request
) +
1690 set
->cmd_size
, cache_line_size()),
1695 if (blk_mq_init_hw_queues(q
, set
))
1698 mutex_lock(&all_q_mutex
);
1699 list_add_tail(&q
->all_q_node
, &all_q_list
);
1700 mutex_unlock(&all_q_mutex
);
1702 blk_mq_add_queue_tag_set(set
, q
);
1704 blk_mq_map_swqueue(q
);
1713 blk_cleanup_queue(q
);
1715 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1718 free_cpumask_var(hctxs
[i
]->cpumask
);
1719 set
->ops
->free_hctx(hctxs
[i
], i
);
1724 return ERR_PTR(-ENOMEM
);
1726 EXPORT_SYMBOL(blk_mq_init_queue
);
1728 void blk_mq_free_queue(struct request_queue
*q
)
1730 struct blk_mq_hw_ctx
*hctx
;
1733 blk_mq_del_queue_tag_set(q
);
1735 queue_for_each_hw_ctx(q
, hctx
, i
) {
1737 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1738 if (q
->mq_ops
->exit_hctx
)
1739 q
->mq_ops
->exit_hctx(hctx
, i
);
1740 free_cpumask_var(hctx
->cpumask
);
1741 q
->mq_ops
->free_hctx(hctx
, i
);
1744 free_percpu(q
->queue_ctx
);
1745 kfree(q
->queue_hw_ctx
);
1748 q
->queue_ctx
= NULL
;
1749 q
->queue_hw_ctx
= NULL
;
1752 mutex_lock(&all_q_mutex
);
1753 list_del_init(&q
->all_q_node
);
1754 mutex_unlock(&all_q_mutex
);
1757 /* Basically redo blk_mq_init_queue with queue frozen */
1758 static void blk_mq_queue_reinit(struct request_queue
*q
)
1760 blk_mq_freeze_queue(q
);
1762 blk_mq_update_queue_map(q
->mq_map
, q
->nr_hw_queues
);
1765 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1766 * we should change hctx numa_node according to new topology (this
1767 * involves free and re-allocate memory, worthy doing?)
1770 blk_mq_map_swqueue(q
);
1772 blk_mq_unfreeze_queue(q
);
1775 static int blk_mq_queue_reinit_notify(struct notifier_block
*nb
,
1776 unsigned long action
, void *hcpu
)
1778 struct request_queue
*q
;
1781 * Before new mappings are established, hotadded cpu might already
1782 * start handling requests. This doesn't break anything as we map
1783 * offline CPUs to first hardware queue. We will re-init the queue
1784 * below to get optimal settings.
1786 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
&&
1787 action
!= CPU_ONLINE
&& action
!= CPU_ONLINE_FROZEN
)
1790 mutex_lock(&all_q_mutex
);
1791 list_for_each_entry(q
, &all_q_list
, all_q_node
)
1792 blk_mq_queue_reinit(q
);
1793 mutex_unlock(&all_q_mutex
);
1797 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
1801 if (!set
->nr_hw_queues
)
1803 if (!set
->queue_depth
|| set
->queue_depth
> BLK_MQ_MAX_DEPTH
)
1805 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
1808 if (!set
->nr_hw_queues
||
1809 !set
->ops
->queue_rq
|| !set
->ops
->map_queue
||
1810 !set
->ops
->alloc_hctx
|| !set
->ops
->free_hctx
)
1814 set
->tags
= kmalloc_node(set
->nr_hw_queues
*
1815 sizeof(struct blk_mq_tags
*),
1816 GFP_KERNEL
, set
->numa_node
);
1820 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1821 set
->tags
[i
] = blk_mq_init_rq_map(set
, i
);
1826 mutex_init(&set
->tag_list_lock
);
1827 INIT_LIST_HEAD(&set
->tag_list
);
1833 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1837 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
1839 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
1843 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1845 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1850 EXPORT_SYMBOL(blk_mq_free_tag_set
);
1852 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
1854 struct blk_mq_tag_set
*set
= q
->tag_set
;
1855 struct blk_mq_hw_ctx
*hctx
;
1858 if (!set
|| nr
> set
->queue_depth
)
1862 queue_for_each_hw_ctx(q
, hctx
, i
) {
1863 ret
= blk_mq_tag_update_depth(hctx
->tags
, nr
);
1869 q
->nr_requests
= nr
;
1874 void blk_mq_disable_hotplug(void)
1876 mutex_lock(&all_q_mutex
);
1879 void blk_mq_enable_hotplug(void)
1881 mutex_unlock(&all_q_mutex
);
1884 static int __init
blk_mq_init(void)
1888 /* Must be called after percpu_counter_hotcpu_callback() */
1889 hotcpu_notifier(blk_mq_queue_reinit_notify
, -10);
1893 subsys_initcall(blk_mq_init
);