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. Use the default queue timeout,
458 * unless one has been set in the request.
461 rq
->deadline
= jiffies
+ q
->rq_timeout
;
463 rq
->deadline
= jiffies
+ rq
->timeout
;
466 * Mark us as started and clear complete. Complete might have been
467 * set if requeue raced with timeout, which then marked it as
468 * complete. So be sure to clear complete again when we start
469 * the request, otherwise we'll ignore the completion event.
471 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
472 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
474 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
476 * Make sure space for the drain appears. We know we can do
477 * this because max_hw_segments has been adjusted to be one
478 * fewer than the device can handle.
480 rq
->nr_phys_segments
++;
484 * Flag the last request in the series so that drivers know when IO
485 * should be kicked off, if they don't do it on a per-request basis.
487 * Note: the flag isn't the only condition drivers should do kick off.
488 * If drive is busy, the last request might not have the bit set.
491 rq
->cmd_flags
|= REQ_END
;
494 static void __blk_mq_requeue_request(struct request
*rq
)
496 struct request_queue
*q
= rq
->q
;
498 trace_block_rq_requeue(q
, rq
);
499 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
501 rq
->cmd_flags
&= ~REQ_END
;
503 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
504 rq
->nr_phys_segments
--;
507 void blk_mq_requeue_request(struct request
*rq
)
509 __blk_mq_requeue_request(rq
);
510 blk_clear_rq_complete(rq
);
512 BUG_ON(blk_queued_rq(rq
));
513 blk_mq_insert_request(rq
, true, true, false);
515 EXPORT_SYMBOL(blk_mq_requeue_request
);
517 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
519 return tags
->rqs
[tag
];
521 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
523 struct blk_mq_timeout_data
{
524 struct blk_mq_hw_ctx
*hctx
;
526 unsigned int *next_set
;
529 static void blk_mq_timeout_check(void *__data
, unsigned long *free_tags
)
531 struct blk_mq_timeout_data
*data
= __data
;
532 struct blk_mq_hw_ctx
*hctx
= data
->hctx
;
535 /* It may not be in flight yet (this is where
536 * the REQ_ATOMIC_STARTED flag comes in). The requests are
537 * statically allocated, so we know it's always safe to access the
538 * memory associated with a bit offset into ->rqs[].
544 tag
= find_next_zero_bit(free_tags
, hctx
->tags
->nr_tags
, tag
);
545 if (tag
>= hctx
->tags
->nr_tags
)
548 rq
= blk_mq_tag_to_rq(hctx
->tags
, tag
++);
549 if (rq
->q
!= hctx
->queue
)
551 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
554 blk_rq_check_expired(rq
, data
->next
, data
->next_set
);
558 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx
*hctx
,
560 unsigned int *next_set
)
562 struct blk_mq_timeout_data data
= {
565 .next_set
= next_set
,
569 * Ask the tagging code to iterate busy requests, so we can
570 * check them for timeout.
572 blk_mq_tag_busy_iter(hctx
->tags
, blk_mq_timeout_check
, &data
);
575 static enum blk_eh_timer_return
blk_mq_rq_timed_out(struct request
*rq
)
577 struct request_queue
*q
= rq
->q
;
580 * We know that complete is set at this point. If STARTED isn't set
581 * anymore, then the request isn't active and the "timeout" should
582 * just be ignored. This can happen due to the bitflag ordering.
583 * Timeout first checks if STARTED is set, and if it is, assumes
584 * the request is active. But if we race with completion, then
585 * we both flags will get cleared. So check here again, and ignore
586 * a timeout event with a request that isn't active.
588 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
589 return BLK_EH_NOT_HANDLED
;
591 if (!q
->mq_ops
->timeout
)
592 return BLK_EH_RESET_TIMER
;
594 return q
->mq_ops
->timeout(rq
);
597 static void blk_mq_rq_timer(unsigned long data
)
599 struct request_queue
*q
= (struct request_queue
*) data
;
600 struct blk_mq_hw_ctx
*hctx
;
601 unsigned long next
= 0;
604 queue_for_each_hw_ctx(q
, hctx
, i
) {
606 * If not software queues are currently mapped to this
607 * hardware queue, there's nothing to check
609 if (!hctx
->nr_ctx
|| !hctx
->tags
)
612 blk_mq_hw_ctx_check_timeout(hctx
, &next
, &next_set
);
616 next
= blk_rq_timeout(round_jiffies_up(next
));
617 mod_timer(&q
->timeout
, next
);
619 queue_for_each_hw_ctx(q
, hctx
, i
)
620 blk_mq_tag_idle(hctx
);
625 * Reverse check our software queue for entries that we could potentially
626 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
627 * too much time checking for merges.
629 static bool blk_mq_attempt_merge(struct request_queue
*q
,
630 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
635 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
641 if (!blk_rq_merge_ok(rq
, bio
))
644 el_ret
= blk_try_merge(rq
, bio
);
645 if (el_ret
== ELEVATOR_BACK_MERGE
) {
646 if (bio_attempt_back_merge(q
, rq
, bio
)) {
651 } else if (el_ret
== ELEVATOR_FRONT_MERGE
) {
652 if (bio_attempt_front_merge(q
, rq
, bio
)) {
664 * Process software queues that have been marked busy, splicing them
665 * to the for-dispatch
667 static void flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
669 struct blk_mq_ctx
*ctx
;
672 for (i
= 0; i
< hctx
->ctx_map
.map_size
; i
++) {
673 struct blk_align_bitmap
*bm
= &hctx
->ctx_map
.map
[i
];
674 unsigned int off
, bit
;
680 off
= i
* hctx
->ctx_map
.bits_per_word
;
682 bit
= find_next_bit(&bm
->word
, bm
->depth
, bit
);
683 if (bit
>= bm
->depth
)
686 ctx
= hctx
->ctxs
[bit
+ off
];
687 clear_bit(bit
, &bm
->word
);
688 spin_lock(&ctx
->lock
);
689 list_splice_tail_init(&ctx
->rq_list
, list
);
690 spin_unlock(&ctx
->lock
);
698 * Run this hardware queue, pulling any software queues mapped to it in.
699 * Note that this function currently has various problems around ordering
700 * of IO. In particular, we'd like FIFO behaviour on handling existing
701 * items on the hctx->dispatch list. Ignore that for now.
703 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
705 struct request_queue
*q
= hctx
->queue
;
710 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
));
712 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)))
718 * Touch any software queue that has pending entries.
720 flush_busy_ctxs(hctx
, &rq_list
);
723 * If we have previous entries on our dispatch list, grab them
724 * and stuff them at the front for more fair dispatch.
726 if (!list_empty_careful(&hctx
->dispatch
)) {
727 spin_lock(&hctx
->lock
);
728 if (!list_empty(&hctx
->dispatch
))
729 list_splice_init(&hctx
->dispatch
, &rq_list
);
730 spin_unlock(&hctx
->lock
);
734 * Now process all the entries, sending them to the driver.
737 while (!list_empty(&rq_list
)) {
740 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
741 list_del_init(&rq
->queuelist
);
743 blk_mq_start_request(rq
, list_empty(&rq_list
));
745 ret
= q
->mq_ops
->queue_rq(hctx
, rq
);
747 case BLK_MQ_RQ_QUEUE_OK
:
750 case BLK_MQ_RQ_QUEUE_BUSY
:
751 list_add(&rq
->queuelist
, &rq_list
);
752 __blk_mq_requeue_request(rq
);
755 pr_err("blk-mq: bad return on queue: %d\n", ret
);
756 case BLK_MQ_RQ_QUEUE_ERROR
:
758 blk_mq_end_io(rq
, rq
->errors
);
762 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
767 hctx
->dispatched
[0]++;
768 else if (queued
< (1 << (BLK_MQ_MAX_DISPATCH_ORDER
- 1)))
769 hctx
->dispatched
[ilog2(queued
) + 1]++;
772 * Any items that need requeuing? Stuff them into hctx->dispatch,
773 * that is where we will continue on next queue run.
775 if (!list_empty(&rq_list
)) {
776 spin_lock(&hctx
->lock
);
777 list_splice(&rq_list
, &hctx
->dispatch
);
778 spin_unlock(&hctx
->lock
);
783 * It'd be great if the workqueue API had a way to pass
784 * in a mask and had some smarts for more clever placement.
785 * For now we just round-robin here, switching for every
786 * BLK_MQ_CPU_WORK_BATCH queued items.
788 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
790 int cpu
= hctx
->next_cpu
;
792 if (--hctx
->next_cpu_batch
<= 0) {
795 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
796 if (next_cpu
>= nr_cpu_ids
)
797 next_cpu
= cpumask_first(hctx
->cpumask
);
799 hctx
->next_cpu
= next_cpu
;
800 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
806 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
808 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)))
811 if (!async
&& cpumask_test_cpu(smp_processor_id(), hctx
->cpumask
))
812 __blk_mq_run_hw_queue(hctx
);
813 else if (hctx
->queue
->nr_hw_queues
== 1)
814 kblockd_schedule_delayed_work(&hctx
->run_work
, 0);
818 cpu
= blk_mq_hctx_next_cpu(hctx
);
819 kblockd_schedule_delayed_work_on(cpu
, &hctx
->run_work
, 0);
823 void blk_mq_run_queues(struct request_queue
*q
, bool async
)
825 struct blk_mq_hw_ctx
*hctx
;
828 queue_for_each_hw_ctx(q
, hctx
, i
) {
829 if ((!blk_mq_hctx_has_pending(hctx
) &&
830 list_empty_careful(&hctx
->dispatch
)) ||
831 test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
835 blk_mq_run_hw_queue(hctx
, async
);
839 EXPORT_SYMBOL(blk_mq_run_queues
);
841 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
843 cancel_delayed_work(&hctx
->run_work
);
844 cancel_delayed_work(&hctx
->delay_work
);
845 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
847 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
849 void blk_mq_stop_hw_queues(struct request_queue
*q
)
851 struct blk_mq_hw_ctx
*hctx
;
854 queue_for_each_hw_ctx(q
, hctx
, i
)
855 blk_mq_stop_hw_queue(hctx
);
857 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
859 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
861 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
864 __blk_mq_run_hw_queue(hctx
);
867 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
869 void blk_mq_start_hw_queues(struct request_queue
*q
)
871 struct blk_mq_hw_ctx
*hctx
;
874 queue_for_each_hw_ctx(q
, hctx
, i
)
875 blk_mq_start_hw_queue(hctx
);
877 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
880 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
882 struct blk_mq_hw_ctx
*hctx
;
885 queue_for_each_hw_ctx(q
, hctx
, i
) {
886 if (!test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
889 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
891 blk_mq_run_hw_queue(hctx
, async
);
895 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
897 static void blk_mq_run_work_fn(struct work_struct
*work
)
899 struct blk_mq_hw_ctx
*hctx
;
901 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
903 __blk_mq_run_hw_queue(hctx
);
906 static void blk_mq_delay_work_fn(struct work_struct
*work
)
908 struct blk_mq_hw_ctx
*hctx
;
910 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
912 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
913 __blk_mq_run_hw_queue(hctx
);
916 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
918 unsigned long tmo
= msecs_to_jiffies(msecs
);
920 if (hctx
->queue
->nr_hw_queues
== 1)
921 kblockd_schedule_delayed_work(&hctx
->delay_work
, tmo
);
925 cpu
= blk_mq_hctx_next_cpu(hctx
);
926 kblockd_schedule_delayed_work_on(cpu
, &hctx
->delay_work
, tmo
);
929 EXPORT_SYMBOL(blk_mq_delay_queue
);
931 static void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
,
932 struct request
*rq
, bool at_head
)
934 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
936 trace_block_rq_insert(hctx
->queue
, rq
);
939 list_add(&rq
->queuelist
, &ctx
->rq_list
);
941 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
943 blk_mq_hctx_mark_pending(hctx
, ctx
);
946 * We do this early, to ensure we are on the right CPU.
951 void blk_mq_insert_request(struct request
*rq
, bool at_head
, bool run_queue
,
954 struct request_queue
*q
= rq
->q
;
955 struct blk_mq_hw_ctx
*hctx
;
956 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
, *current_ctx
;
958 current_ctx
= blk_mq_get_ctx(q
);
959 if (!cpu_online(ctx
->cpu
))
960 rq
->mq_ctx
= ctx
= current_ctx
;
962 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
964 if (rq
->cmd_flags
& (REQ_FLUSH
| REQ_FUA
) &&
965 !(rq
->cmd_flags
& (REQ_FLUSH_SEQ
))) {
966 blk_insert_flush(rq
);
968 spin_lock(&ctx
->lock
);
969 __blk_mq_insert_request(hctx
, rq
, at_head
);
970 spin_unlock(&ctx
->lock
);
974 blk_mq_run_hw_queue(hctx
, async
);
976 blk_mq_put_ctx(current_ctx
);
979 static void blk_mq_insert_requests(struct request_queue
*q
,
980 struct blk_mq_ctx
*ctx
,
981 struct list_head
*list
,
986 struct blk_mq_hw_ctx
*hctx
;
987 struct blk_mq_ctx
*current_ctx
;
989 trace_block_unplug(q
, depth
, !from_schedule
);
991 current_ctx
= blk_mq_get_ctx(q
);
993 if (!cpu_online(ctx
->cpu
))
995 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
998 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1001 spin_lock(&ctx
->lock
);
1002 while (!list_empty(list
)) {
1005 rq
= list_first_entry(list
, struct request
, queuelist
);
1006 list_del_init(&rq
->queuelist
);
1008 __blk_mq_insert_request(hctx
, rq
, false);
1010 spin_unlock(&ctx
->lock
);
1012 blk_mq_run_hw_queue(hctx
, from_schedule
);
1013 blk_mq_put_ctx(current_ctx
);
1016 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1018 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1019 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1021 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1022 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1023 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1026 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1028 struct blk_mq_ctx
*this_ctx
;
1029 struct request_queue
*this_q
;
1032 LIST_HEAD(ctx_list
);
1035 list_splice_init(&plug
->mq_list
, &list
);
1037 list_sort(NULL
, &list
, plug_ctx_cmp
);
1043 while (!list_empty(&list
)) {
1044 rq
= list_entry_rq(list
.next
);
1045 list_del_init(&rq
->queuelist
);
1047 if (rq
->mq_ctx
!= this_ctx
) {
1049 blk_mq_insert_requests(this_q
, this_ctx
,
1054 this_ctx
= rq
->mq_ctx
;
1060 list_add_tail(&rq
->queuelist
, &ctx_list
);
1064 * If 'this_ctx' is set, we know we have entries to complete
1065 * on 'ctx_list'. Do those.
1068 blk_mq_insert_requests(this_q
, this_ctx
, &ctx_list
, depth
,
1073 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1075 init_request_from_bio(rq
, bio
);
1076 blk_account_io_start(rq
, 1);
1079 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx
*hctx
,
1080 struct blk_mq_ctx
*ctx
,
1081 struct request
*rq
, struct bio
*bio
)
1083 struct request_queue
*q
= hctx
->queue
;
1085 if (!(hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
)) {
1086 blk_mq_bio_to_request(rq
, bio
);
1087 spin_lock(&ctx
->lock
);
1089 __blk_mq_insert_request(hctx
, rq
, false);
1090 spin_unlock(&ctx
->lock
);
1093 spin_lock(&ctx
->lock
);
1094 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1095 blk_mq_bio_to_request(rq
, bio
);
1099 spin_unlock(&ctx
->lock
);
1100 __blk_mq_free_request(hctx
, ctx
, rq
);
1105 struct blk_map_ctx
{
1106 struct blk_mq_hw_ctx
*hctx
;
1107 struct blk_mq_ctx
*ctx
;
1110 static struct request
*blk_mq_map_request(struct request_queue
*q
,
1112 struct blk_map_ctx
*data
)
1114 struct blk_mq_hw_ctx
*hctx
;
1115 struct blk_mq_ctx
*ctx
;
1117 int rw
= bio_data_dir(bio
);
1119 if (unlikely(blk_mq_queue_enter(q
))) {
1120 bio_endio(bio
, -EIO
);
1124 ctx
= blk_mq_get_ctx(q
);
1125 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1127 if (rw_is_sync(bio
->bi_rw
))
1130 trace_block_getrq(q
, bio
, rw
);
1131 rq
= __blk_mq_alloc_request(hctx
, ctx
, GFP_ATOMIC
, false);
1133 blk_mq_rq_ctx_init(q
, ctx
, rq
, rw
);
1135 blk_mq_put_ctx(ctx
);
1136 trace_block_sleeprq(q
, bio
, rw
);
1137 rq
= blk_mq_alloc_request_pinned(q
, rw
, __GFP_WAIT
|GFP_ATOMIC
,
1140 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1150 * Multiple hardware queue variant. This will not use per-process plugs,
1151 * but will attempt to bypass the hctx queueing if we can go straight to
1152 * hardware for SYNC IO.
1154 static void blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1156 const int is_sync
= rw_is_sync(bio
->bi_rw
);
1157 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
1158 struct blk_map_ctx data
;
1161 blk_queue_bounce(q
, &bio
);
1163 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1164 bio_endio(bio
, -EIO
);
1168 rq
= blk_mq_map_request(q
, bio
, &data
);
1172 if (unlikely(is_flush_fua
)) {
1173 blk_mq_bio_to_request(rq
, bio
);
1174 blk_insert_flush(rq
);
1181 blk_mq_bio_to_request(rq
, bio
);
1182 blk_mq_start_request(rq
, true);
1185 * For OK queue, we are done. For error, kill it. Any other
1186 * error (busy), just add it to our list as we previously
1189 ret
= q
->mq_ops
->queue_rq(data
.hctx
, rq
);
1190 if (ret
== BLK_MQ_RQ_QUEUE_OK
)
1193 __blk_mq_requeue_request(rq
);
1195 if (ret
== BLK_MQ_RQ_QUEUE_ERROR
) {
1197 blk_mq_end_io(rq
, rq
->errors
);
1203 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1205 * For a SYNC request, send it to the hardware immediately. For
1206 * an ASYNC request, just ensure that we run it later on. The
1207 * latter allows for merging opportunities and more efficient
1211 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1214 blk_mq_put_ctx(data
.ctx
);
1218 * Single hardware queue variant. This will attempt to use any per-process
1219 * plug for merging and IO deferral.
1221 static void blk_sq_make_request(struct request_queue
*q
, struct bio
*bio
)
1223 const int is_sync
= rw_is_sync(bio
->bi_rw
);
1224 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
1225 unsigned int use_plug
, request_count
= 0;
1226 struct blk_map_ctx data
;
1230 * If we have multiple hardware queues, just go directly to
1231 * one of those for sync IO.
1233 use_plug
= !is_flush_fua
&& !is_sync
;
1235 blk_queue_bounce(q
, &bio
);
1237 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1238 bio_endio(bio
, -EIO
);
1242 if (use_plug
&& !blk_queue_nomerges(q
) &&
1243 blk_attempt_plug_merge(q
, bio
, &request_count
))
1246 rq
= blk_mq_map_request(q
, bio
, &data
);
1248 if (unlikely(is_flush_fua
)) {
1249 blk_mq_bio_to_request(rq
, bio
);
1250 blk_insert_flush(rq
);
1255 * A task plug currently exists. Since this is completely lockless,
1256 * utilize that to temporarily store requests until the task is
1257 * either done or scheduled away.
1260 struct blk_plug
*plug
= current
->plug
;
1263 blk_mq_bio_to_request(rq
, bio
);
1264 if (list_empty(&plug
->mq_list
))
1265 trace_block_plug(q
);
1266 else if (request_count
>= BLK_MAX_REQUEST_COUNT
) {
1267 blk_flush_plug_list(plug
, false);
1268 trace_block_plug(q
);
1270 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1271 blk_mq_put_ctx(data
.ctx
);
1276 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1278 * For a SYNC request, send it to the hardware immediately. For
1279 * an ASYNC request, just ensure that we run it later on. The
1280 * latter allows for merging opportunities and more efficient
1284 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1287 blk_mq_put_ctx(data
.ctx
);
1291 * Default mapping to a software queue, since we use one per CPU.
1293 struct blk_mq_hw_ctx
*blk_mq_map_queue(struct request_queue
*q
, const int cpu
)
1295 return q
->queue_hw_ctx
[q
->mq_map
[cpu
]];
1297 EXPORT_SYMBOL(blk_mq_map_queue
);
1299 struct blk_mq_hw_ctx
*blk_mq_alloc_single_hw_queue(struct blk_mq_tag_set
*set
,
1300 unsigned int hctx_index
)
1302 return kzalloc_node(sizeof(struct blk_mq_hw_ctx
), GFP_KERNEL
,
1305 EXPORT_SYMBOL(blk_mq_alloc_single_hw_queue
);
1307 void blk_mq_free_single_hw_queue(struct blk_mq_hw_ctx
*hctx
,
1308 unsigned int hctx_index
)
1312 EXPORT_SYMBOL(blk_mq_free_single_hw_queue
);
1314 static void blk_mq_free_rq_map(struct blk_mq_tag_set
*set
,
1315 struct blk_mq_tags
*tags
, unsigned int hctx_idx
)
1319 if (tags
->rqs
&& set
->ops
->exit_request
) {
1322 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1325 set
->ops
->exit_request(set
->driver_data
, tags
->rqs
[i
],
1330 while (!list_empty(&tags
->page_list
)) {
1331 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1332 list_del_init(&page
->lru
);
1333 __free_pages(page
, page
->private);
1338 blk_mq_free_tags(tags
);
1341 static size_t order_to_size(unsigned int order
)
1343 return (size_t)PAGE_SIZE
<< order
;
1346 static struct blk_mq_tags
*blk_mq_init_rq_map(struct blk_mq_tag_set
*set
,
1347 unsigned int hctx_idx
)
1349 struct blk_mq_tags
*tags
;
1350 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1351 size_t rq_size
, left
;
1353 tags
= blk_mq_init_tags(set
->queue_depth
, set
->reserved_tags
,
1358 INIT_LIST_HEAD(&tags
->page_list
);
1360 tags
->rqs
= kmalloc_node(set
->queue_depth
* sizeof(struct request
*),
1361 GFP_KERNEL
, set
->numa_node
);
1363 blk_mq_free_tags(tags
);
1368 * rq_size is the size of the request plus driver payload, rounded
1369 * to the cacheline size
1371 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1373 left
= rq_size
* set
->queue_depth
;
1375 for (i
= 0; i
< set
->queue_depth
; ) {
1376 int this_order
= max_order
;
1381 while (left
< order_to_size(this_order
- 1) && this_order
)
1385 page
= alloc_pages_node(set
->numa_node
, GFP_KERNEL
,
1391 if (order_to_size(this_order
) < rq_size
)
1398 page
->private = this_order
;
1399 list_add_tail(&page
->lru
, &tags
->page_list
);
1401 p
= page_address(page
);
1402 entries_per_page
= order_to_size(this_order
) / rq_size
;
1403 to_do
= min(entries_per_page
, set
->queue_depth
- i
);
1404 left
-= to_do
* rq_size
;
1405 for (j
= 0; j
< to_do
; j
++) {
1407 if (set
->ops
->init_request
) {
1408 if (set
->ops
->init_request(set
->driver_data
,
1409 tags
->rqs
[i
], hctx_idx
, i
,
1422 pr_warn("%s: failed to allocate requests\n", __func__
);
1423 blk_mq_free_rq_map(set
, tags
, hctx_idx
);
1427 static void blk_mq_free_bitmap(struct blk_mq_ctxmap
*bitmap
)
1432 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap
*bitmap
, int node
)
1434 unsigned int bpw
= 8, total
, num_maps
, i
;
1436 bitmap
->bits_per_word
= bpw
;
1438 num_maps
= ALIGN(nr_cpu_ids
, bpw
) / bpw
;
1439 bitmap
->map
= kzalloc_node(num_maps
* sizeof(struct blk_align_bitmap
),
1444 bitmap
->map_size
= num_maps
;
1447 for (i
= 0; i
< num_maps
; i
++) {
1448 bitmap
->map
[i
].depth
= min(total
, bitmap
->bits_per_word
);
1449 total
-= bitmap
->map
[i
].depth
;
1455 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx
*hctx
, int cpu
)
1457 struct request_queue
*q
= hctx
->queue
;
1458 struct blk_mq_ctx
*ctx
;
1462 * Move ctx entries to new CPU, if this one is going away.
1464 ctx
= __blk_mq_get_ctx(q
, cpu
);
1466 spin_lock(&ctx
->lock
);
1467 if (!list_empty(&ctx
->rq_list
)) {
1468 list_splice_init(&ctx
->rq_list
, &tmp
);
1469 blk_mq_hctx_clear_pending(hctx
, ctx
);
1471 spin_unlock(&ctx
->lock
);
1473 if (list_empty(&tmp
))
1476 ctx
= blk_mq_get_ctx(q
);
1477 spin_lock(&ctx
->lock
);
1479 while (!list_empty(&tmp
)) {
1482 rq
= list_first_entry(&tmp
, struct request
, queuelist
);
1484 list_move_tail(&rq
->queuelist
, &ctx
->rq_list
);
1487 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1488 blk_mq_hctx_mark_pending(hctx
, ctx
);
1490 spin_unlock(&ctx
->lock
);
1492 blk_mq_run_hw_queue(hctx
, true);
1493 blk_mq_put_ctx(ctx
);
1497 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx
*hctx
, int cpu
)
1499 struct request_queue
*q
= hctx
->queue
;
1500 struct blk_mq_tag_set
*set
= q
->tag_set
;
1502 if (set
->tags
[hctx
->queue_num
])
1505 set
->tags
[hctx
->queue_num
] = blk_mq_init_rq_map(set
, hctx
->queue_num
);
1506 if (!set
->tags
[hctx
->queue_num
])
1509 hctx
->tags
= set
->tags
[hctx
->queue_num
];
1513 static int blk_mq_hctx_notify(void *data
, unsigned long action
,
1516 struct blk_mq_hw_ctx
*hctx
= data
;
1518 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
)
1519 return blk_mq_hctx_cpu_offline(hctx
, cpu
);
1520 else if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
)
1521 return blk_mq_hctx_cpu_online(hctx
, cpu
);
1526 static int blk_mq_init_hw_queues(struct request_queue
*q
,
1527 struct blk_mq_tag_set
*set
)
1529 struct blk_mq_hw_ctx
*hctx
;
1533 * Initialize hardware queues
1535 queue_for_each_hw_ctx(q
, hctx
, i
) {
1538 node
= hctx
->numa_node
;
1539 if (node
== NUMA_NO_NODE
)
1540 node
= hctx
->numa_node
= set
->numa_node
;
1542 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1543 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1544 spin_lock_init(&hctx
->lock
);
1545 INIT_LIST_HEAD(&hctx
->dispatch
);
1547 hctx
->queue_num
= i
;
1548 hctx
->flags
= set
->flags
;
1549 hctx
->cmd_size
= set
->cmd_size
;
1551 blk_mq_init_cpu_notifier(&hctx
->cpu_notifier
,
1552 blk_mq_hctx_notify
, hctx
);
1553 blk_mq_register_cpu_notifier(&hctx
->cpu_notifier
);
1555 hctx
->tags
= set
->tags
[i
];
1558 * Allocate space for all possible cpus to avoid allocation in
1561 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1566 if (blk_mq_alloc_bitmap(&hctx
->ctx_map
, node
))
1571 if (set
->ops
->init_hctx
&&
1572 set
->ops
->init_hctx(hctx
, set
->driver_data
, i
))
1576 if (i
== q
->nr_hw_queues
)
1582 queue_for_each_hw_ctx(q
, hctx
, j
) {
1586 if (set
->ops
->exit_hctx
)
1587 set
->ops
->exit_hctx(hctx
, j
);
1589 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1591 blk_mq_free_bitmap(&hctx
->ctx_map
);
1597 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1598 unsigned int nr_hw_queues
)
1602 for_each_possible_cpu(i
) {
1603 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1604 struct blk_mq_hw_ctx
*hctx
;
1606 memset(__ctx
, 0, sizeof(*__ctx
));
1608 spin_lock_init(&__ctx
->lock
);
1609 INIT_LIST_HEAD(&__ctx
->rq_list
);
1612 /* If the cpu isn't online, the cpu is mapped to first hctx */
1616 hctx
= q
->mq_ops
->map_queue(q
, i
);
1617 cpumask_set_cpu(i
, hctx
->cpumask
);
1621 * Set local node, IFF we have more than one hw queue. If
1622 * not, we remain on the home node of the device
1624 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1625 hctx
->numa_node
= cpu_to_node(i
);
1629 static void blk_mq_map_swqueue(struct request_queue
*q
)
1632 struct blk_mq_hw_ctx
*hctx
;
1633 struct blk_mq_ctx
*ctx
;
1635 queue_for_each_hw_ctx(q
, hctx
, i
) {
1636 cpumask_clear(hctx
->cpumask
);
1641 * Map software to hardware queues
1643 queue_for_each_ctx(q
, ctx
, i
) {
1644 /* If the cpu isn't online, the cpu is mapped to first hctx */
1648 hctx
= q
->mq_ops
->map_queue(q
, i
);
1649 cpumask_set_cpu(i
, hctx
->cpumask
);
1650 ctx
->index_hw
= hctx
->nr_ctx
;
1651 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
1654 queue_for_each_hw_ctx(q
, hctx
, i
) {
1656 * If not software queues are mapped to this hardware queue,
1657 * disable it and free the request entries
1659 if (!hctx
->nr_ctx
) {
1660 struct blk_mq_tag_set
*set
= q
->tag_set
;
1663 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1664 set
->tags
[i
] = NULL
;
1671 * Initialize batch roundrobin counts
1673 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
1674 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1678 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
)
1680 struct blk_mq_hw_ctx
*hctx
;
1681 struct request_queue
*q
;
1685 if (set
->tag_list
.next
== set
->tag_list
.prev
)
1690 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
1691 blk_mq_freeze_queue(q
);
1693 queue_for_each_hw_ctx(q
, hctx
, i
) {
1695 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
1697 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
1699 blk_mq_unfreeze_queue(q
);
1703 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
1705 struct blk_mq_tag_set
*set
= q
->tag_set
;
1707 blk_mq_freeze_queue(q
);
1709 mutex_lock(&set
->tag_list_lock
);
1710 list_del_init(&q
->tag_set_list
);
1711 blk_mq_update_tag_set_depth(set
);
1712 mutex_unlock(&set
->tag_list_lock
);
1714 blk_mq_unfreeze_queue(q
);
1717 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
1718 struct request_queue
*q
)
1722 mutex_lock(&set
->tag_list_lock
);
1723 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
1724 blk_mq_update_tag_set_depth(set
);
1725 mutex_unlock(&set
->tag_list_lock
);
1728 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
1730 struct blk_mq_hw_ctx
**hctxs
;
1731 struct blk_mq_ctx
*ctx
;
1732 struct request_queue
*q
;
1735 ctx
= alloc_percpu(struct blk_mq_ctx
);
1737 return ERR_PTR(-ENOMEM
);
1739 hctxs
= kmalloc_node(set
->nr_hw_queues
* sizeof(*hctxs
), GFP_KERNEL
,
1745 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1746 hctxs
[i
] = set
->ops
->alloc_hctx(set
, i
);
1750 if (!zalloc_cpumask_var(&hctxs
[i
]->cpumask
, GFP_KERNEL
))
1753 atomic_set(&hctxs
[i
]->nr_active
, 0);
1754 hctxs
[i
]->numa_node
= NUMA_NO_NODE
;
1755 hctxs
[i
]->queue_num
= i
;
1758 q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
1762 q
->mq_map
= blk_mq_make_queue_map(set
);
1766 setup_timer(&q
->timeout
, blk_mq_rq_timer
, (unsigned long) q
);
1767 blk_queue_rq_timeout(q
, 30000);
1769 q
->nr_queues
= nr_cpu_ids
;
1770 q
->nr_hw_queues
= set
->nr_hw_queues
;
1773 q
->queue_hw_ctx
= hctxs
;
1775 q
->mq_ops
= set
->ops
;
1776 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
1778 q
->sg_reserved_size
= INT_MAX
;
1780 if (q
->nr_hw_queues
> 1)
1781 blk_queue_make_request(q
, blk_mq_make_request
);
1783 blk_queue_make_request(q
, blk_sq_make_request
);
1785 blk_queue_rq_timed_out(q
, blk_mq_rq_timed_out
);
1787 blk_queue_rq_timeout(q
, set
->timeout
);
1790 * Do this after blk_queue_make_request() overrides it...
1792 q
->nr_requests
= set
->queue_depth
;
1794 if (set
->ops
->complete
)
1795 blk_queue_softirq_done(q
, set
->ops
->complete
);
1797 blk_mq_init_flush(q
);
1798 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
1800 q
->flush_rq
= kzalloc(round_up(sizeof(struct request
) +
1801 set
->cmd_size
, cache_line_size()),
1806 if (blk_mq_init_hw_queues(q
, set
))
1809 mutex_lock(&all_q_mutex
);
1810 list_add_tail(&q
->all_q_node
, &all_q_list
);
1811 mutex_unlock(&all_q_mutex
);
1813 blk_mq_add_queue_tag_set(set
, q
);
1815 blk_mq_map_swqueue(q
);
1824 blk_cleanup_queue(q
);
1826 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1829 free_cpumask_var(hctxs
[i
]->cpumask
);
1830 set
->ops
->free_hctx(hctxs
[i
], i
);
1835 return ERR_PTR(-ENOMEM
);
1837 EXPORT_SYMBOL(blk_mq_init_queue
);
1839 void blk_mq_free_queue(struct request_queue
*q
)
1841 struct blk_mq_hw_ctx
*hctx
;
1844 blk_mq_del_queue_tag_set(q
);
1846 queue_for_each_hw_ctx(q
, hctx
, i
) {
1847 blk_mq_tag_idle(hctx
);
1849 blk_mq_free_bitmap(&hctx
->ctx_map
);
1850 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1851 if (q
->mq_ops
->exit_hctx
)
1852 q
->mq_ops
->exit_hctx(hctx
, i
);
1853 free_cpumask_var(hctx
->cpumask
);
1854 q
->mq_ops
->free_hctx(hctx
, i
);
1857 free_percpu(q
->queue_ctx
);
1858 kfree(q
->queue_hw_ctx
);
1861 q
->queue_ctx
= NULL
;
1862 q
->queue_hw_ctx
= NULL
;
1865 mutex_lock(&all_q_mutex
);
1866 list_del_init(&q
->all_q_node
);
1867 mutex_unlock(&all_q_mutex
);
1870 /* Basically redo blk_mq_init_queue with queue frozen */
1871 static void blk_mq_queue_reinit(struct request_queue
*q
)
1873 blk_mq_freeze_queue(q
);
1875 blk_mq_update_queue_map(q
->mq_map
, q
->nr_hw_queues
);
1878 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1879 * we should change hctx numa_node according to new topology (this
1880 * involves free and re-allocate memory, worthy doing?)
1883 blk_mq_map_swqueue(q
);
1885 blk_mq_unfreeze_queue(q
);
1888 static int blk_mq_queue_reinit_notify(struct notifier_block
*nb
,
1889 unsigned long action
, void *hcpu
)
1891 struct request_queue
*q
;
1894 * Before new mappings are established, hotadded cpu might already
1895 * start handling requests. This doesn't break anything as we map
1896 * offline CPUs to first hardware queue. We will re-init the queue
1897 * below to get optimal settings.
1899 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
&&
1900 action
!= CPU_ONLINE
&& action
!= CPU_ONLINE_FROZEN
)
1903 mutex_lock(&all_q_mutex
);
1904 list_for_each_entry(q
, &all_q_list
, all_q_node
)
1905 blk_mq_queue_reinit(q
);
1906 mutex_unlock(&all_q_mutex
);
1910 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
1914 if (!set
->nr_hw_queues
)
1916 if (!set
->queue_depth
|| set
->queue_depth
> BLK_MQ_MAX_DEPTH
)
1918 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
1921 if (!set
->nr_hw_queues
||
1922 !set
->ops
->queue_rq
|| !set
->ops
->map_queue
||
1923 !set
->ops
->alloc_hctx
|| !set
->ops
->free_hctx
)
1927 set
->tags
= kmalloc_node(set
->nr_hw_queues
*
1928 sizeof(struct blk_mq_tags
*),
1929 GFP_KERNEL
, set
->numa_node
);
1933 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1934 set
->tags
[i
] = blk_mq_init_rq_map(set
, i
);
1939 mutex_init(&set
->tag_list_lock
);
1940 INIT_LIST_HEAD(&set
->tag_list
);
1946 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1950 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
1952 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
1956 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1958 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1963 EXPORT_SYMBOL(blk_mq_free_tag_set
);
1965 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
1967 struct blk_mq_tag_set
*set
= q
->tag_set
;
1968 struct blk_mq_hw_ctx
*hctx
;
1971 if (!set
|| nr
> set
->queue_depth
)
1975 queue_for_each_hw_ctx(q
, hctx
, i
) {
1976 ret
= blk_mq_tag_update_depth(hctx
->tags
, nr
);
1982 q
->nr_requests
= nr
;
1987 void blk_mq_disable_hotplug(void)
1989 mutex_lock(&all_q_mutex
);
1992 void blk_mq_enable_hotplug(void)
1994 mutex_unlock(&all_q_mutex
);
1997 static int __init
blk_mq_init(void)
2001 /* Must be called after percpu_counter_hotcpu_callback() */
2002 hotcpu_notifier(blk_mq_queue_reinit_notify
, -10);
2006 subsys_initcall(blk_mq_init
);