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>
12 #include <linux/kmemleak.h>
14 #include <linux/init.h>
15 #include <linux/slab.h>
16 #include <linux/workqueue.h>
17 #include <linux/smp.h>
18 #include <linux/llist.h>
19 #include <linux/list_sort.h>
20 #include <linux/cpu.h>
21 #include <linux/cache.h>
22 #include <linux/sched/sysctl.h>
23 #include <linux/delay.h>
24 #include <linux/crash_dump.h>
25 #include <linux/prefetch.h>
27 #include <trace/events/block.h>
29 #include <linux/blk-mq.h>
32 #include "blk-mq-tag.h"
34 static DEFINE_MUTEX(all_q_mutex
);
35 static LIST_HEAD(all_q_list
);
38 * Check if any of the ctx's have pending work in this hardware queue
40 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
42 return sbitmap_any_bit_set(&hctx
->ctx_map
);
46 * Mark this ctx as having pending work in this hardware queue
48 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
49 struct blk_mq_ctx
*ctx
)
51 if (!sbitmap_test_bit(&hctx
->ctx_map
, ctx
->index_hw
))
52 sbitmap_set_bit(&hctx
->ctx_map
, ctx
->index_hw
);
55 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
56 struct blk_mq_ctx
*ctx
)
58 sbitmap_clear_bit(&hctx
->ctx_map
, ctx
->index_hw
);
61 void blk_mq_freeze_queue_start(struct request_queue
*q
)
65 freeze_depth
= atomic_inc_return(&q
->mq_freeze_depth
);
66 if (freeze_depth
== 1) {
67 percpu_ref_kill(&q
->q_usage_counter
);
68 blk_mq_run_hw_queues(q
, false);
71 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start
);
73 static void blk_mq_freeze_queue_wait(struct request_queue
*q
)
75 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
79 * Guarantee no request is in use, so we can change any data structure of
80 * the queue afterward.
82 void blk_freeze_queue(struct request_queue
*q
)
85 * In the !blk_mq case we are only calling this to kill the
86 * q_usage_counter, otherwise this increases the freeze depth
87 * and waits for it to return to zero. For this reason there is
88 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
89 * exported to drivers as the only user for unfreeze is blk_mq.
91 blk_mq_freeze_queue_start(q
);
92 blk_mq_freeze_queue_wait(q
);
95 void blk_mq_freeze_queue(struct request_queue
*q
)
98 * ...just an alias to keep freeze and unfreeze actions balanced
99 * in the blk_mq_* namespace
103 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
105 void blk_mq_unfreeze_queue(struct request_queue
*q
)
109 freeze_depth
= atomic_dec_return(&q
->mq_freeze_depth
);
110 WARN_ON_ONCE(freeze_depth
< 0);
112 percpu_ref_reinit(&q
->q_usage_counter
);
113 wake_up_all(&q
->mq_freeze_wq
);
116 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
118 void blk_mq_wake_waiters(struct request_queue
*q
)
120 struct blk_mq_hw_ctx
*hctx
;
123 queue_for_each_hw_ctx(q
, hctx
, i
)
124 if (blk_mq_hw_queue_mapped(hctx
))
125 blk_mq_tag_wakeup_all(hctx
->tags
, true);
128 * If we are called because the queue has now been marked as
129 * dying, we need to ensure that processes currently waiting on
130 * the queue are notified as well.
132 wake_up_all(&q
->mq_freeze_wq
);
135 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
137 return blk_mq_has_free_tags(hctx
->tags
);
139 EXPORT_SYMBOL(blk_mq_can_queue
);
141 static void blk_mq_rq_ctx_init(struct request_queue
*q
, struct blk_mq_ctx
*ctx
,
142 struct request
*rq
, unsigned int op
)
144 INIT_LIST_HEAD(&rq
->queuelist
);
145 /* csd/requeue_work/fifo_time is initialized before use */
149 if (blk_queue_io_stat(q
))
150 rq
->rq_flags
|= RQF_IO_STAT
;
151 /* do not touch atomic flags, it needs atomic ops against the timer */
153 INIT_HLIST_NODE(&rq
->hash
);
154 RB_CLEAR_NODE(&rq
->rb_node
);
157 rq
->start_time
= jiffies
;
158 #ifdef CONFIG_BLK_CGROUP
160 set_start_time_ns(rq
);
161 rq
->io_start_time_ns
= 0;
163 rq
->nr_phys_segments
= 0;
164 #if defined(CONFIG_BLK_DEV_INTEGRITY)
165 rq
->nr_integrity_segments
= 0;
168 /* tag was already set */
178 INIT_LIST_HEAD(&rq
->timeout_list
);
182 rq
->end_io_data
= NULL
;
185 ctx
->rq_dispatched
[op_is_sync(op
)]++;
188 static struct request
*
189 __blk_mq_alloc_request(struct blk_mq_alloc_data
*data
, unsigned int op
)
194 tag
= blk_mq_get_tag(data
);
195 if (tag
!= BLK_MQ_TAG_FAIL
) {
196 rq
= data
->hctx
->tags
->rqs
[tag
];
198 if (blk_mq_tag_busy(data
->hctx
)) {
199 rq
->rq_flags
= RQF_MQ_INFLIGHT
;
200 atomic_inc(&data
->hctx
->nr_active
);
204 blk_mq_rq_ctx_init(data
->q
, data
->ctx
, rq
, op
);
211 struct request
*blk_mq_alloc_request(struct request_queue
*q
, int rw
,
214 struct blk_mq_ctx
*ctx
;
215 struct blk_mq_hw_ctx
*hctx
;
217 struct blk_mq_alloc_data alloc_data
;
220 ret
= blk_queue_enter(q
, flags
& BLK_MQ_REQ_NOWAIT
);
224 ctx
= blk_mq_get_ctx(q
);
225 hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
226 blk_mq_set_alloc_data(&alloc_data
, q
, flags
, ctx
, hctx
);
227 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
232 return ERR_PTR(-EWOULDBLOCK
);
236 rq
->__sector
= (sector_t
) -1;
237 rq
->bio
= rq
->biotail
= NULL
;
240 EXPORT_SYMBOL(blk_mq_alloc_request
);
242 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
, int rw
,
243 unsigned int flags
, unsigned int hctx_idx
)
245 struct blk_mq_hw_ctx
*hctx
;
246 struct blk_mq_ctx
*ctx
;
248 struct blk_mq_alloc_data alloc_data
;
252 * If the tag allocator sleeps we could get an allocation for a
253 * different hardware context. No need to complicate the low level
254 * allocator for this for the rare use case of a command tied to
257 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
258 return ERR_PTR(-EINVAL
);
260 if (hctx_idx
>= q
->nr_hw_queues
)
261 return ERR_PTR(-EIO
);
263 ret
= blk_queue_enter(q
, true);
268 * Check if the hardware context is actually mapped to anything.
269 * If not tell the caller that it should skip this queue.
271 hctx
= q
->queue_hw_ctx
[hctx_idx
];
272 if (!blk_mq_hw_queue_mapped(hctx
)) {
276 ctx
= __blk_mq_get_ctx(q
, cpumask_first(hctx
->cpumask
));
278 blk_mq_set_alloc_data(&alloc_data
, q
, flags
, ctx
, hctx
);
279 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
291 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
293 static void __blk_mq_free_request(struct blk_mq_hw_ctx
*hctx
,
294 struct blk_mq_ctx
*ctx
, struct request
*rq
)
296 const int tag
= rq
->tag
;
297 struct request_queue
*q
= rq
->q
;
299 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
300 atomic_dec(&hctx
->nr_active
);
303 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
304 blk_mq_put_tag(hctx
, ctx
, tag
);
308 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
310 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
312 ctx
->rq_completed
[rq_is_sync(rq
)]++;
313 __blk_mq_free_request(hctx
, ctx
, rq
);
316 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request
);
318 void blk_mq_free_request(struct request
*rq
)
320 blk_mq_free_hctx_request(blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
), rq
);
322 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
324 inline void __blk_mq_end_request(struct request
*rq
, int error
)
326 blk_account_io_done(rq
);
329 rq
->end_io(rq
, error
);
331 if (unlikely(blk_bidi_rq(rq
)))
332 blk_mq_free_request(rq
->next_rq
);
333 blk_mq_free_request(rq
);
336 EXPORT_SYMBOL(__blk_mq_end_request
);
338 void blk_mq_end_request(struct request
*rq
, int error
)
340 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
342 __blk_mq_end_request(rq
, error
);
344 EXPORT_SYMBOL(blk_mq_end_request
);
346 static void __blk_mq_complete_request_remote(void *data
)
348 struct request
*rq
= data
;
350 rq
->q
->softirq_done_fn(rq
);
353 static void blk_mq_ipi_complete_request(struct request
*rq
)
355 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
359 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
360 rq
->q
->softirq_done_fn(rq
);
365 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
366 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
368 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
369 rq
->csd
.func
= __blk_mq_complete_request_remote
;
372 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
374 rq
->q
->softirq_done_fn(rq
);
379 static void __blk_mq_complete_request(struct request
*rq
)
381 struct request_queue
*q
= rq
->q
;
383 if (!q
->softirq_done_fn
)
384 blk_mq_end_request(rq
, rq
->errors
);
386 blk_mq_ipi_complete_request(rq
);
390 * blk_mq_complete_request - end I/O on a request
391 * @rq: the request being processed
394 * Ends all I/O on a request. It does not handle partial completions.
395 * The actual completion happens out-of-order, through a IPI handler.
397 void blk_mq_complete_request(struct request
*rq
, int error
)
399 struct request_queue
*q
= rq
->q
;
401 if (unlikely(blk_should_fake_timeout(q
)))
403 if (!blk_mark_rq_complete(rq
)) {
405 __blk_mq_complete_request(rq
);
408 EXPORT_SYMBOL(blk_mq_complete_request
);
410 int blk_mq_request_started(struct request
*rq
)
412 return test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
414 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
416 void blk_mq_start_request(struct request
*rq
)
418 struct request_queue
*q
= rq
->q
;
420 trace_block_rq_issue(q
, rq
);
422 rq
->resid_len
= blk_rq_bytes(rq
);
423 if (unlikely(blk_bidi_rq(rq
)))
424 rq
->next_rq
->resid_len
= blk_rq_bytes(rq
->next_rq
);
429 * Ensure that ->deadline is visible before set the started
430 * flag and clear the completed flag.
432 smp_mb__before_atomic();
435 * Mark us as started and clear complete. Complete might have been
436 * set if requeue raced with timeout, which then marked it as
437 * complete. So be sure to clear complete again when we start
438 * the request, otherwise we'll ignore the completion event.
440 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
441 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
442 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
443 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
445 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
447 * Make sure space for the drain appears. We know we can do
448 * this because max_hw_segments has been adjusted to be one
449 * fewer than the device can handle.
451 rq
->nr_phys_segments
++;
454 EXPORT_SYMBOL(blk_mq_start_request
);
456 static void __blk_mq_requeue_request(struct request
*rq
)
458 struct request_queue
*q
= rq
->q
;
460 trace_block_rq_requeue(q
, rq
);
462 if (test_and_clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
463 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
464 rq
->nr_phys_segments
--;
468 void blk_mq_requeue_request(struct request
*rq
)
470 __blk_mq_requeue_request(rq
);
472 BUG_ON(blk_queued_rq(rq
));
473 blk_mq_add_to_requeue_list(rq
, true);
475 EXPORT_SYMBOL(blk_mq_requeue_request
);
477 static void blk_mq_requeue_work(struct work_struct
*work
)
479 struct request_queue
*q
=
480 container_of(work
, struct request_queue
, requeue_work
.work
);
482 struct request
*rq
, *next
;
485 spin_lock_irqsave(&q
->requeue_lock
, flags
);
486 list_splice_init(&q
->requeue_list
, &rq_list
);
487 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
489 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
490 if (!(rq
->rq_flags
& RQF_SOFTBARRIER
))
493 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
494 list_del_init(&rq
->queuelist
);
495 blk_mq_insert_request(rq
, true, false, false);
498 while (!list_empty(&rq_list
)) {
499 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
500 list_del_init(&rq
->queuelist
);
501 blk_mq_insert_request(rq
, false, false, false);
504 blk_mq_run_hw_queues(q
, false);
507 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
)
509 struct request_queue
*q
= rq
->q
;
513 * We abuse this flag that is otherwise used by the I/O scheduler to
514 * request head insertation from the workqueue.
516 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
518 spin_lock_irqsave(&q
->requeue_lock
, flags
);
520 rq
->rq_flags
|= RQF_SOFTBARRIER
;
521 list_add(&rq
->queuelist
, &q
->requeue_list
);
523 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
525 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
527 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
529 void blk_mq_cancel_requeue_work(struct request_queue
*q
)
531 cancel_delayed_work_sync(&q
->requeue_work
);
533 EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work
);
535 void blk_mq_kick_requeue_list(struct request_queue
*q
)
537 kblockd_schedule_delayed_work(&q
->requeue_work
, 0);
539 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
541 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
544 kblockd_schedule_delayed_work(&q
->requeue_work
,
545 msecs_to_jiffies(msecs
));
547 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
549 void blk_mq_abort_requeue_list(struct request_queue
*q
)
554 spin_lock_irqsave(&q
->requeue_lock
, flags
);
555 list_splice_init(&q
->requeue_list
, &rq_list
);
556 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
558 while (!list_empty(&rq_list
)) {
561 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
562 list_del_init(&rq
->queuelist
);
564 blk_mq_end_request(rq
, rq
->errors
);
567 EXPORT_SYMBOL(blk_mq_abort_requeue_list
);
569 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
571 if (tag
< tags
->nr_tags
) {
572 prefetch(tags
->rqs
[tag
]);
573 return tags
->rqs
[tag
];
578 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
580 struct blk_mq_timeout_data
{
582 unsigned int next_set
;
585 void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
587 struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
588 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
591 * We know that complete is set at this point. If STARTED isn't set
592 * anymore, then the request isn't active and the "timeout" should
593 * just be ignored. This can happen due to the bitflag ordering.
594 * Timeout first checks if STARTED is set, and if it is, assumes
595 * the request is active. But if we race with completion, then
596 * we both flags will get cleared. So check here again, and ignore
597 * a timeout event with a request that isn't active.
599 if (!test_bit(REQ_ATOM_STARTED
, &req
->atomic_flags
))
603 ret
= ops
->timeout(req
, reserved
);
607 __blk_mq_complete_request(req
);
609 case BLK_EH_RESET_TIMER
:
611 blk_clear_rq_complete(req
);
613 case BLK_EH_NOT_HANDLED
:
616 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
621 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
622 struct request
*rq
, void *priv
, bool reserved
)
624 struct blk_mq_timeout_data
*data
= priv
;
626 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
628 * If a request wasn't started before the queue was
629 * marked dying, kill it here or it'll go unnoticed.
631 if (unlikely(blk_queue_dying(rq
->q
))) {
633 blk_mq_end_request(rq
, rq
->errors
);
638 if (time_after_eq(jiffies
, rq
->deadline
)) {
639 if (!blk_mark_rq_complete(rq
))
640 blk_mq_rq_timed_out(rq
, reserved
);
641 } else if (!data
->next_set
|| time_after(data
->next
, rq
->deadline
)) {
642 data
->next
= rq
->deadline
;
647 static void blk_mq_timeout_work(struct work_struct
*work
)
649 struct request_queue
*q
=
650 container_of(work
, struct request_queue
, timeout_work
);
651 struct blk_mq_timeout_data data
= {
657 /* A deadlock might occur if a request is stuck requiring a
658 * timeout at the same time a queue freeze is waiting
659 * completion, since the timeout code would not be able to
660 * acquire the queue reference here.
662 * That's why we don't use blk_queue_enter here; instead, we use
663 * percpu_ref_tryget directly, because we need to be able to
664 * obtain a reference even in the short window between the queue
665 * starting to freeze, by dropping the first reference in
666 * blk_mq_freeze_queue_start, and the moment the last request is
667 * consumed, marked by the instant q_usage_counter reaches
670 if (!percpu_ref_tryget(&q
->q_usage_counter
))
673 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &data
);
676 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
677 mod_timer(&q
->timeout
, data
.next
);
679 struct blk_mq_hw_ctx
*hctx
;
681 queue_for_each_hw_ctx(q
, hctx
, i
) {
682 /* the hctx may be unmapped, so check it here */
683 if (blk_mq_hw_queue_mapped(hctx
))
684 blk_mq_tag_idle(hctx
);
691 * Reverse check our software queue for entries that we could potentially
692 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
693 * too much time checking for merges.
695 static bool blk_mq_attempt_merge(struct request_queue
*q
,
696 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
701 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
707 if (!blk_rq_merge_ok(rq
, bio
))
710 el_ret
= blk_try_merge(rq
, bio
);
711 if (el_ret
== ELEVATOR_BACK_MERGE
) {
712 if (bio_attempt_back_merge(q
, rq
, bio
)) {
717 } else if (el_ret
== ELEVATOR_FRONT_MERGE
) {
718 if (bio_attempt_front_merge(q
, rq
, bio
)) {
729 struct flush_busy_ctx_data
{
730 struct blk_mq_hw_ctx
*hctx
;
731 struct list_head
*list
;
734 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
736 struct flush_busy_ctx_data
*flush_data
= data
;
737 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
738 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
740 sbitmap_clear_bit(sb
, bitnr
);
741 spin_lock(&ctx
->lock
);
742 list_splice_tail_init(&ctx
->rq_list
, flush_data
->list
);
743 spin_unlock(&ctx
->lock
);
748 * Process software queues that have been marked busy, splicing them
749 * to the for-dispatch
751 static void flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
753 struct flush_busy_ctx_data data
= {
758 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
761 static inline unsigned int queued_to_index(unsigned int queued
)
766 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
770 * Run this hardware queue, pulling any software queues mapped to it in.
771 * Note that this function currently has various problems around ordering
772 * of IO. In particular, we'd like FIFO behaviour on handling existing
773 * items on the hctx->dispatch list. Ignore that for now.
775 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
777 struct request_queue
*q
= hctx
->queue
;
780 LIST_HEAD(driver_list
);
781 struct list_head
*dptr
;
784 if (unlikely(blk_mq_hctx_stopped(hctx
)))
787 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
788 cpu_online(hctx
->next_cpu
));
793 * Touch any software queue that has pending entries.
795 flush_busy_ctxs(hctx
, &rq_list
);
798 * If we have previous entries on our dispatch list, grab them
799 * and stuff them at the front for more fair dispatch.
801 if (!list_empty_careful(&hctx
->dispatch
)) {
802 spin_lock(&hctx
->lock
);
803 if (!list_empty(&hctx
->dispatch
))
804 list_splice_init(&hctx
->dispatch
, &rq_list
);
805 spin_unlock(&hctx
->lock
);
809 * Start off with dptr being NULL, so we start the first request
810 * immediately, even if we have more pending.
815 * Now process all the entries, sending them to the driver.
818 while (!list_empty(&rq_list
)) {
819 struct blk_mq_queue_data bd
;
822 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
823 list_del_init(&rq
->queuelist
);
827 bd
.last
= list_empty(&rq_list
);
829 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
831 case BLK_MQ_RQ_QUEUE_OK
:
834 case BLK_MQ_RQ_QUEUE_BUSY
:
835 list_add(&rq
->queuelist
, &rq_list
);
836 __blk_mq_requeue_request(rq
);
839 pr_err("blk-mq: bad return on queue: %d\n", ret
);
840 case BLK_MQ_RQ_QUEUE_ERROR
:
842 blk_mq_end_request(rq
, rq
->errors
);
846 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
850 * We've done the first request. If we have more than 1
851 * left in the list, set dptr to defer issue.
853 if (!dptr
&& rq_list
.next
!= rq_list
.prev
)
857 hctx
->dispatched
[queued_to_index(queued
)]++;
860 * Any items that need requeuing? Stuff them into hctx->dispatch,
861 * that is where we will continue on next queue run.
863 if (!list_empty(&rq_list
)) {
864 spin_lock(&hctx
->lock
);
865 list_splice(&rq_list
, &hctx
->dispatch
);
866 spin_unlock(&hctx
->lock
);
868 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
869 * it's possible the queue is stopped and restarted again
870 * before this. Queue restart will dispatch requests. And since
871 * requests in rq_list aren't added into hctx->dispatch yet,
872 * the requests in rq_list might get lost.
874 * blk_mq_run_hw_queue() already checks the STOPPED bit
876 blk_mq_run_hw_queue(hctx
, true);
881 * It'd be great if the workqueue API had a way to pass
882 * in a mask and had some smarts for more clever placement.
883 * For now we just round-robin here, switching for every
884 * BLK_MQ_CPU_WORK_BATCH queued items.
886 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
888 if (hctx
->queue
->nr_hw_queues
== 1)
889 return WORK_CPU_UNBOUND
;
891 if (--hctx
->next_cpu_batch
<= 0) {
892 int cpu
= hctx
->next_cpu
, next_cpu
;
894 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
895 if (next_cpu
>= nr_cpu_ids
)
896 next_cpu
= cpumask_first(hctx
->cpumask
);
898 hctx
->next_cpu
= next_cpu
;
899 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
904 return hctx
->next_cpu
;
907 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
909 if (unlikely(blk_mq_hctx_stopped(hctx
) ||
910 !blk_mq_hw_queue_mapped(hctx
)))
913 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
915 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
916 __blk_mq_run_hw_queue(hctx
);
924 kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
);
927 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
929 struct blk_mq_hw_ctx
*hctx
;
932 queue_for_each_hw_ctx(q
, hctx
, i
) {
933 if ((!blk_mq_hctx_has_pending(hctx
) &&
934 list_empty_careful(&hctx
->dispatch
)) ||
935 blk_mq_hctx_stopped(hctx
))
938 blk_mq_run_hw_queue(hctx
, async
);
941 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
944 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
947 * The caller is responsible for serializing this function against
948 * blk_mq_{start,stop}_hw_queue().
950 bool blk_mq_queue_stopped(struct request_queue
*q
)
952 struct blk_mq_hw_ctx
*hctx
;
955 queue_for_each_hw_ctx(q
, hctx
, i
)
956 if (blk_mq_hctx_stopped(hctx
))
961 EXPORT_SYMBOL(blk_mq_queue_stopped
);
963 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
965 cancel_work(&hctx
->run_work
);
966 cancel_delayed_work(&hctx
->delay_work
);
967 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
969 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
971 void blk_mq_stop_hw_queues(struct request_queue
*q
)
973 struct blk_mq_hw_ctx
*hctx
;
976 queue_for_each_hw_ctx(q
, hctx
, i
)
977 blk_mq_stop_hw_queue(hctx
);
979 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
981 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
983 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
985 blk_mq_run_hw_queue(hctx
, false);
987 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
989 void blk_mq_start_hw_queues(struct request_queue
*q
)
991 struct blk_mq_hw_ctx
*hctx
;
994 queue_for_each_hw_ctx(q
, hctx
, i
)
995 blk_mq_start_hw_queue(hctx
);
997 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
999 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1001 struct blk_mq_hw_ctx
*hctx
;
1004 queue_for_each_hw_ctx(q
, hctx
, i
) {
1005 if (!blk_mq_hctx_stopped(hctx
))
1008 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1009 blk_mq_run_hw_queue(hctx
, async
);
1012 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1014 static void blk_mq_run_work_fn(struct work_struct
*work
)
1016 struct blk_mq_hw_ctx
*hctx
;
1018 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
);
1020 __blk_mq_run_hw_queue(hctx
);
1023 static void blk_mq_delay_work_fn(struct work_struct
*work
)
1025 struct blk_mq_hw_ctx
*hctx
;
1027 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
1029 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
1030 __blk_mq_run_hw_queue(hctx
);
1033 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1035 if (unlikely(!blk_mq_hw_queue_mapped(hctx
)))
1038 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1039 &hctx
->delay_work
, msecs_to_jiffies(msecs
));
1041 EXPORT_SYMBOL(blk_mq_delay_queue
);
1043 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1047 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1049 trace_block_rq_insert(hctx
->queue
, rq
);
1052 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1054 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1057 static void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
,
1058 struct request
*rq
, bool at_head
)
1060 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1062 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1063 blk_mq_hctx_mark_pending(hctx
, ctx
);
1066 void blk_mq_insert_request(struct request
*rq
, bool at_head
, bool run_queue
,
1069 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1070 struct request_queue
*q
= rq
->q
;
1071 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
1073 spin_lock(&ctx
->lock
);
1074 __blk_mq_insert_request(hctx
, rq
, at_head
);
1075 spin_unlock(&ctx
->lock
);
1078 blk_mq_run_hw_queue(hctx
, async
);
1081 static void blk_mq_insert_requests(struct request_queue
*q
,
1082 struct blk_mq_ctx
*ctx
,
1083 struct list_head
*list
,
1088 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
1090 trace_block_unplug(q
, depth
, !from_schedule
);
1093 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1096 spin_lock(&ctx
->lock
);
1097 while (!list_empty(list
)) {
1100 rq
= list_first_entry(list
, struct request
, queuelist
);
1101 BUG_ON(rq
->mq_ctx
!= ctx
);
1102 list_del_init(&rq
->queuelist
);
1103 __blk_mq_insert_req_list(hctx
, rq
, false);
1105 blk_mq_hctx_mark_pending(hctx
, ctx
);
1106 spin_unlock(&ctx
->lock
);
1108 blk_mq_run_hw_queue(hctx
, from_schedule
);
1111 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1113 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1114 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1116 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1117 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1118 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1121 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1123 struct blk_mq_ctx
*this_ctx
;
1124 struct request_queue
*this_q
;
1127 LIST_HEAD(ctx_list
);
1130 list_splice_init(&plug
->mq_list
, &list
);
1132 list_sort(NULL
, &list
, plug_ctx_cmp
);
1138 while (!list_empty(&list
)) {
1139 rq
= list_entry_rq(list
.next
);
1140 list_del_init(&rq
->queuelist
);
1142 if (rq
->mq_ctx
!= this_ctx
) {
1144 blk_mq_insert_requests(this_q
, this_ctx
,
1149 this_ctx
= rq
->mq_ctx
;
1155 list_add_tail(&rq
->queuelist
, &ctx_list
);
1159 * If 'this_ctx' is set, we know we have entries to complete
1160 * on 'ctx_list'. Do those.
1163 blk_mq_insert_requests(this_q
, this_ctx
, &ctx_list
, depth
,
1168 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1170 init_request_from_bio(rq
, bio
);
1172 blk_account_io_start(rq
, 1);
1175 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1177 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1178 !blk_queue_nomerges(hctx
->queue
);
1181 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx
*hctx
,
1182 struct blk_mq_ctx
*ctx
,
1183 struct request
*rq
, struct bio
*bio
)
1185 if (!hctx_allow_merges(hctx
) || !bio_mergeable(bio
)) {
1186 blk_mq_bio_to_request(rq
, bio
);
1187 spin_lock(&ctx
->lock
);
1189 __blk_mq_insert_request(hctx
, rq
, false);
1190 spin_unlock(&ctx
->lock
);
1193 struct request_queue
*q
= hctx
->queue
;
1195 spin_lock(&ctx
->lock
);
1196 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1197 blk_mq_bio_to_request(rq
, bio
);
1201 spin_unlock(&ctx
->lock
);
1202 __blk_mq_free_request(hctx
, ctx
, rq
);
1207 static struct request
*blk_mq_map_request(struct request_queue
*q
,
1209 struct blk_mq_alloc_data
*data
)
1211 struct blk_mq_hw_ctx
*hctx
;
1212 struct blk_mq_ctx
*ctx
;
1215 blk_queue_enter_live(q
);
1216 ctx
= blk_mq_get_ctx(q
);
1217 hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
1219 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1220 blk_mq_set_alloc_data(data
, q
, 0, ctx
, hctx
);
1221 rq
= __blk_mq_alloc_request(data
, bio
->bi_opf
);
1223 data
->hctx
->queued
++;
1227 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1228 struct request
*rq
, blk_qc_t
*cookie
)
1231 struct request_queue
*q
= rq
->q
;
1232 struct blk_mq_queue_data bd
= {
1237 blk_qc_t new_cookie
= blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
);
1239 if (blk_mq_hctx_stopped(hctx
))
1243 * For OK queue, we are done. For error, kill it. Any other
1244 * error (busy), just add it to our list as we previously
1247 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1248 if (ret
== BLK_MQ_RQ_QUEUE_OK
) {
1249 *cookie
= new_cookie
;
1253 __blk_mq_requeue_request(rq
);
1255 if (ret
== BLK_MQ_RQ_QUEUE_ERROR
) {
1256 *cookie
= BLK_QC_T_NONE
;
1258 blk_mq_end_request(rq
, rq
->errors
);
1263 blk_mq_insert_request(rq
, false, true, true);
1267 * Multiple hardware queue variant. This will not use per-process plugs,
1268 * but will attempt to bypass the hctx queueing if we can go straight to
1269 * hardware for SYNC IO.
1271 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1273 const int is_sync
= op_is_sync(bio
->bi_opf
);
1274 const int is_flush_fua
= bio
->bi_opf
& (REQ_PREFLUSH
| REQ_FUA
);
1275 struct blk_mq_alloc_data data
;
1277 unsigned int request_count
= 0;
1278 struct blk_plug
*plug
;
1279 struct request
*same_queue_rq
= NULL
;
1282 blk_queue_bounce(q
, &bio
);
1284 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1286 return BLK_QC_T_NONE
;
1289 blk_queue_split(q
, &bio
, q
->bio_split
);
1291 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1292 blk_attempt_plug_merge(q
, bio
, &request_count
, &same_queue_rq
))
1293 return BLK_QC_T_NONE
;
1295 rq
= blk_mq_map_request(q
, bio
, &data
);
1297 return BLK_QC_T_NONE
;
1299 cookie
= blk_tag_to_qc_t(rq
->tag
, data
.hctx
->queue_num
);
1301 if (unlikely(is_flush_fua
)) {
1302 blk_mq_bio_to_request(rq
, bio
);
1303 blk_insert_flush(rq
);
1307 plug
= current
->plug
;
1309 * If the driver supports defer issued based on 'last', then
1310 * queue it up like normal since we can potentially save some
1313 if (((plug
&& !blk_queue_nomerges(q
)) || is_sync
) &&
1314 !(data
.hctx
->flags
& BLK_MQ_F_DEFER_ISSUE
)) {
1315 struct request
*old_rq
= NULL
;
1317 blk_mq_bio_to_request(rq
, bio
);
1320 * We do limited pluging. If the bio can be merged, do that.
1321 * Otherwise the existing request in the plug list will be
1322 * issued. So the plug list will have one request at most
1326 * The plug list might get flushed before this. If that
1327 * happens, same_queue_rq is invalid and plug list is
1330 if (same_queue_rq
&& !list_empty(&plug
->mq_list
)) {
1331 old_rq
= same_queue_rq
;
1332 list_del_init(&old_rq
->queuelist
);
1334 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1335 } else /* is_sync */
1337 blk_mq_put_ctx(data
.ctx
);
1340 blk_mq_try_issue_directly(data
.hctx
, old_rq
, &cookie
);
1344 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1346 * For a SYNC request, send it to the hardware immediately. For
1347 * an ASYNC request, just ensure that we run it later on. The
1348 * latter allows for merging opportunities and more efficient
1352 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1354 blk_mq_put_ctx(data
.ctx
);
1360 * Single hardware queue variant. This will attempt to use any per-process
1361 * plug for merging and IO deferral.
1363 static blk_qc_t
blk_sq_make_request(struct request_queue
*q
, struct bio
*bio
)
1365 const int is_sync
= op_is_sync(bio
->bi_opf
);
1366 const int is_flush_fua
= bio
->bi_opf
& (REQ_PREFLUSH
| REQ_FUA
);
1367 struct blk_plug
*plug
;
1368 unsigned int request_count
= 0;
1369 struct blk_mq_alloc_data data
;
1373 blk_queue_bounce(q
, &bio
);
1375 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1377 return BLK_QC_T_NONE
;
1380 blk_queue_split(q
, &bio
, q
->bio_split
);
1382 if (!is_flush_fua
&& !blk_queue_nomerges(q
)) {
1383 if (blk_attempt_plug_merge(q
, bio
, &request_count
, NULL
))
1384 return BLK_QC_T_NONE
;
1386 request_count
= blk_plug_queued_count(q
);
1388 rq
= blk_mq_map_request(q
, bio
, &data
);
1390 return BLK_QC_T_NONE
;
1392 cookie
= blk_tag_to_qc_t(rq
->tag
, data
.hctx
->queue_num
);
1394 if (unlikely(is_flush_fua
)) {
1395 blk_mq_bio_to_request(rq
, bio
);
1396 blk_insert_flush(rq
);
1401 * A task plug currently exists. Since this is completely lockless,
1402 * utilize that to temporarily store requests until the task is
1403 * either done or scheduled away.
1405 plug
= current
->plug
;
1407 blk_mq_bio_to_request(rq
, bio
);
1409 trace_block_plug(q
);
1411 blk_mq_put_ctx(data
.ctx
);
1413 if (request_count
>= BLK_MAX_REQUEST_COUNT
) {
1414 blk_flush_plug_list(plug
, false);
1415 trace_block_plug(q
);
1418 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1422 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1424 * For a SYNC request, send it to the hardware immediately. For
1425 * an ASYNC request, just ensure that we run it later on. The
1426 * latter allows for merging opportunities and more efficient
1430 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1433 blk_mq_put_ctx(data
.ctx
);
1437 static void blk_mq_free_rq_map(struct blk_mq_tag_set
*set
,
1438 struct blk_mq_tags
*tags
, unsigned int hctx_idx
)
1442 if (tags
->rqs
&& set
->ops
->exit_request
) {
1445 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1448 set
->ops
->exit_request(set
->driver_data
, tags
->rqs
[i
],
1450 tags
->rqs
[i
] = NULL
;
1454 while (!list_empty(&tags
->page_list
)) {
1455 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1456 list_del_init(&page
->lru
);
1458 * Remove kmemleak object previously allocated in
1459 * blk_mq_init_rq_map().
1461 kmemleak_free(page_address(page
));
1462 __free_pages(page
, page
->private);
1467 blk_mq_free_tags(tags
);
1470 static size_t order_to_size(unsigned int order
)
1472 return (size_t)PAGE_SIZE
<< order
;
1475 static struct blk_mq_tags
*blk_mq_init_rq_map(struct blk_mq_tag_set
*set
,
1476 unsigned int hctx_idx
)
1478 struct blk_mq_tags
*tags
;
1479 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1480 size_t rq_size
, left
;
1482 tags
= blk_mq_init_tags(set
->queue_depth
, set
->reserved_tags
,
1484 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1488 INIT_LIST_HEAD(&tags
->page_list
);
1490 tags
->rqs
= kzalloc_node(set
->queue_depth
* sizeof(struct request
*),
1491 GFP_KERNEL
| __GFP_NOWARN
| __GFP_NORETRY
,
1494 blk_mq_free_tags(tags
);
1499 * rq_size is the size of the request plus driver payload, rounded
1500 * to the cacheline size
1502 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1504 left
= rq_size
* set
->queue_depth
;
1506 for (i
= 0; i
< set
->queue_depth
; ) {
1507 int this_order
= max_order
;
1512 while (this_order
&& left
< order_to_size(this_order
- 1))
1516 page
= alloc_pages_node(set
->numa_node
,
1517 GFP_KERNEL
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
1523 if (order_to_size(this_order
) < rq_size
)
1530 page
->private = this_order
;
1531 list_add_tail(&page
->lru
, &tags
->page_list
);
1533 p
= page_address(page
);
1535 * Allow kmemleak to scan these pages as they contain pointers
1536 * to additional allocations like via ops->init_request().
1538 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_KERNEL
);
1539 entries_per_page
= order_to_size(this_order
) / rq_size
;
1540 to_do
= min(entries_per_page
, set
->queue_depth
- i
);
1541 left
-= to_do
* rq_size
;
1542 for (j
= 0; j
< to_do
; j
++) {
1544 if (set
->ops
->init_request
) {
1545 if (set
->ops
->init_request(set
->driver_data
,
1546 tags
->rqs
[i
], hctx_idx
, i
,
1548 tags
->rqs
[i
] = NULL
;
1560 blk_mq_free_rq_map(set
, tags
, hctx_idx
);
1565 * 'cpu' is going away. splice any existing rq_list entries from this
1566 * software queue to the hw queue dispatch list, and ensure that it
1569 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
1571 struct blk_mq_hw_ctx
*hctx
;
1572 struct blk_mq_ctx
*ctx
;
1575 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
1576 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
1578 spin_lock(&ctx
->lock
);
1579 if (!list_empty(&ctx
->rq_list
)) {
1580 list_splice_init(&ctx
->rq_list
, &tmp
);
1581 blk_mq_hctx_clear_pending(hctx
, ctx
);
1583 spin_unlock(&ctx
->lock
);
1585 if (list_empty(&tmp
))
1588 spin_lock(&hctx
->lock
);
1589 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
1590 spin_unlock(&hctx
->lock
);
1592 blk_mq_run_hw_queue(hctx
, true);
1596 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
1598 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
1602 /* hctx->ctxs will be freed in queue's release handler */
1603 static void blk_mq_exit_hctx(struct request_queue
*q
,
1604 struct blk_mq_tag_set
*set
,
1605 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
1607 unsigned flush_start_tag
= set
->queue_depth
;
1609 blk_mq_tag_idle(hctx
);
1611 if (set
->ops
->exit_request
)
1612 set
->ops
->exit_request(set
->driver_data
,
1613 hctx
->fq
->flush_rq
, hctx_idx
,
1614 flush_start_tag
+ hctx_idx
);
1616 if (set
->ops
->exit_hctx
)
1617 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1619 blk_mq_remove_cpuhp(hctx
);
1620 blk_free_flush_queue(hctx
->fq
);
1621 sbitmap_free(&hctx
->ctx_map
);
1624 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1625 struct blk_mq_tag_set
*set
, int nr_queue
)
1627 struct blk_mq_hw_ctx
*hctx
;
1630 queue_for_each_hw_ctx(q
, hctx
, i
) {
1633 blk_mq_exit_hctx(q
, set
, hctx
, i
);
1637 static void blk_mq_free_hw_queues(struct request_queue
*q
,
1638 struct blk_mq_tag_set
*set
)
1640 struct blk_mq_hw_ctx
*hctx
;
1643 queue_for_each_hw_ctx(q
, hctx
, i
)
1644 free_cpumask_var(hctx
->cpumask
);
1647 static int blk_mq_init_hctx(struct request_queue
*q
,
1648 struct blk_mq_tag_set
*set
,
1649 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
1652 unsigned flush_start_tag
= set
->queue_depth
;
1654 node
= hctx
->numa_node
;
1655 if (node
== NUMA_NO_NODE
)
1656 node
= hctx
->numa_node
= set
->numa_node
;
1658 INIT_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1659 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1660 spin_lock_init(&hctx
->lock
);
1661 INIT_LIST_HEAD(&hctx
->dispatch
);
1663 hctx
->queue_num
= hctx_idx
;
1664 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
1666 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
1668 hctx
->tags
= set
->tags
[hctx_idx
];
1671 * Allocate space for all possible cpus to avoid allocation at
1674 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1677 goto unregister_cpu_notifier
;
1679 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8), GFP_KERNEL
,
1685 if (set
->ops
->init_hctx
&&
1686 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
1689 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
1693 if (set
->ops
->init_request
&&
1694 set
->ops
->init_request(set
->driver_data
,
1695 hctx
->fq
->flush_rq
, hctx_idx
,
1696 flush_start_tag
+ hctx_idx
, node
))
1704 if (set
->ops
->exit_hctx
)
1705 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1707 sbitmap_free(&hctx
->ctx_map
);
1710 unregister_cpu_notifier
:
1711 blk_mq_remove_cpuhp(hctx
);
1715 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1716 unsigned int nr_hw_queues
)
1720 for_each_possible_cpu(i
) {
1721 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1722 struct blk_mq_hw_ctx
*hctx
;
1724 memset(__ctx
, 0, sizeof(*__ctx
));
1726 spin_lock_init(&__ctx
->lock
);
1727 INIT_LIST_HEAD(&__ctx
->rq_list
);
1730 /* If the cpu isn't online, the cpu is mapped to first hctx */
1734 hctx
= blk_mq_map_queue(q
, i
);
1737 * Set local node, IFF we have more than one hw queue. If
1738 * not, we remain on the home node of the device
1740 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1741 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
1745 static void blk_mq_map_swqueue(struct request_queue
*q
,
1746 const struct cpumask
*online_mask
)
1749 struct blk_mq_hw_ctx
*hctx
;
1750 struct blk_mq_ctx
*ctx
;
1751 struct blk_mq_tag_set
*set
= q
->tag_set
;
1754 * Avoid others reading imcomplete hctx->cpumask through sysfs
1756 mutex_lock(&q
->sysfs_lock
);
1758 queue_for_each_hw_ctx(q
, hctx
, i
) {
1759 cpumask_clear(hctx
->cpumask
);
1764 * Map software to hardware queues
1766 for_each_possible_cpu(i
) {
1767 /* If the cpu isn't online, the cpu is mapped to first hctx */
1768 if (!cpumask_test_cpu(i
, online_mask
))
1771 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1772 hctx
= blk_mq_map_queue(q
, i
);
1774 cpumask_set_cpu(i
, hctx
->cpumask
);
1775 ctx
->index_hw
= hctx
->nr_ctx
;
1776 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
1779 mutex_unlock(&q
->sysfs_lock
);
1781 queue_for_each_hw_ctx(q
, hctx
, i
) {
1783 * If no software queues are mapped to this hardware queue,
1784 * disable it and free the request entries.
1786 if (!hctx
->nr_ctx
) {
1788 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1789 set
->tags
[i
] = NULL
;
1795 /* unmapped hw queue can be remapped after CPU topo changed */
1797 set
->tags
[i
] = blk_mq_init_rq_map(set
, i
);
1798 hctx
->tags
= set
->tags
[i
];
1799 WARN_ON(!hctx
->tags
);
1802 * Set the map size to the number of mapped software queues.
1803 * This is more accurate and more efficient than looping
1804 * over all possibly mapped software queues.
1806 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
1809 * Initialize batch roundrobin counts
1811 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
1812 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1816 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
1818 struct blk_mq_hw_ctx
*hctx
;
1821 queue_for_each_hw_ctx(q
, hctx
, i
) {
1823 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
1825 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
1829 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
, bool shared
)
1831 struct request_queue
*q
;
1833 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
1834 blk_mq_freeze_queue(q
);
1835 queue_set_hctx_shared(q
, shared
);
1836 blk_mq_unfreeze_queue(q
);
1840 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
1842 struct blk_mq_tag_set
*set
= q
->tag_set
;
1844 mutex_lock(&set
->tag_list_lock
);
1845 list_del_init(&q
->tag_set_list
);
1846 if (list_is_singular(&set
->tag_list
)) {
1847 /* just transitioned to unshared */
1848 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
1849 /* update existing queue */
1850 blk_mq_update_tag_set_depth(set
, false);
1852 mutex_unlock(&set
->tag_list_lock
);
1855 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
1856 struct request_queue
*q
)
1860 mutex_lock(&set
->tag_list_lock
);
1862 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
1863 if (!list_empty(&set
->tag_list
) && !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
1864 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
1865 /* update existing queue */
1866 blk_mq_update_tag_set_depth(set
, true);
1868 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
1869 queue_set_hctx_shared(q
, true);
1870 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
1872 mutex_unlock(&set
->tag_list_lock
);
1876 * It is the actual release handler for mq, but we do it from
1877 * request queue's release handler for avoiding use-after-free
1878 * and headache because q->mq_kobj shouldn't have been introduced,
1879 * but we can't group ctx/kctx kobj without it.
1881 void blk_mq_release(struct request_queue
*q
)
1883 struct blk_mq_hw_ctx
*hctx
;
1886 /* hctx kobj stays in hctx */
1887 queue_for_each_hw_ctx(q
, hctx
, i
) {
1896 kfree(q
->queue_hw_ctx
);
1898 /* ctx kobj stays in queue_ctx */
1899 free_percpu(q
->queue_ctx
);
1902 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
1904 struct request_queue
*uninit_q
, *q
;
1906 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
1908 return ERR_PTR(-ENOMEM
);
1910 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
1912 blk_cleanup_queue(uninit_q
);
1916 EXPORT_SYMBOL(blk_mq_init_queue
);
1918 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
1919 struct request_queue
*q
)
1922 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
1924 blk_mq_sysfs_unregister(q
);
1925 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1931 node
= blk_mq_hw_queue_to_node(q
->mq_map
, i
);
1932 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
1937 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
1944 atomic_set(&hctxs
[i
]->nr_active
, 0);
1945 hctxs
[i
]->numa_node
= node
;
1946 hctxs
[i
]->queue_num
= i
;
1948 if (blk_mq_init_hctx(q
, set
, hctxs
[i
], i
)) {
1949 free_cpumask_var(hctxs
[i
]->cpumask
);
1954 blk_mq_hctx_kobj_init(hctxs
[i
]);
1956 for (j
= i
; j
< q
->nr_hw_queues
; j
++) {
1957 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
1961 blk_mq_free_rq_map(set
, hctx
->tags
, j
);
1962 set
->tags
[j
] = NULL
;
1964 blk_mq_exit_hctx(q
, set
, hctx
, j
);
1965 free_cpumask_var(hctx
->cpumask
);
1966 kobject_put(&hctx
->kobj
);
1973 q
->nr_hw_queues
= i
;
1974 blk_mq_sysfs_register(q
);
1977 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
1978 struct request_queue
*q
)
1980 /* mark the queue as mq asap */
1981 q
->mq_ops
= set
->ops
;
1983 q
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
1987 q
->queue_hw_ctx
= kzalloc_node(nr_cpu_ids
* sizeof(*(q
->queue_hw_ctx
)),
1988 GFP_KERNEL
, set
->numa_node
);
1989 if (!q
->queue_hw_ctx
)
1992 q
->mq_map
= set
->mq_map
;
1994 blk_mq_realloc_hw_ctxs(set
, q
);
1995 if (!q
->nr_hw_queues
)
1998 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
1999 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2001 q
->nr_queues
= nr_cpu_ids
;
2003 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2005 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2006 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
2008 q
->sg_reserved_size
= INT_MAX
;
2010 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2011 INIT_LIST_HEAD(&q
->requeue_list
);
2012 spin_lock_init(&q
->requeue_lock
);
2014 if (q
->nr_hw_queues
> 1)
2015 blk_queue_make_request(q
, blk_mq_make_request
);
2017 blk_queue_make_request(q
, blk_sq_make_request
);
2020 * Do this after blk_queue_make_request() overrides it...
2022 q
->nr_requests
= set
->queue_depth
;
2024 if (set
->ops
->complete
)
2025 blk_queue_softirq_done(q
, set
->ops
->complete
);
2027 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2030 mutex_lock(&all_q_mutex
);
2032 list_add_tail(&q
->all_q_node
, &all_q_list
);
2033 blk_mq_add_queue_tag_set(set
, q
);
2034 blk_mq_map_swqueue(q
, cpu_online_mask
);
2036 mutex_unlock(&all_q_mutex
);
2042 kfree(q
->queue_hw_ctx
);
2044 free_percpu(q
->queue_ctx
);
2047 return ERR_PTR(-ENOMEM
);
2049 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2051 void blk_mq_free_queue(struct request_queue
*q
)
2053 struct blk_mq_tag_set
*set
= q
->tag_set
;
2055 mutex_lock(&all_q_mutex
);
2056 list_del_init(&q
->all_q_node
);
2057 mutex_unlock(&all_q_mutex
);
2059 blk_mq_del_queue_tag_set(q
);
2061 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2062 blk_mq_free_hw_queues(q
, set
);
2065 /* Basically redo blk_mq_init_queue with queue frozen */
2066 static void blk_mq_queue_reinit(struct request_queue
*q
,
2067 const struct cpumask
*online_mask
)
2069 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2071 blk_mq_sysfs_unregister(q
);
2074 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2075 * we should change hctx numa_node according to new topology (this
2076 * involves free and re-allocate memory, worthy doing?)
2079 blk_mq_map_swqueue(q
, online_mask
);
2081 blk_mq_sysfs_register(q
);
2085 * New online cpumask which is going to be set in this hotplug event.
2086 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2087 * one-by-one and dynamically allocating this could result in a failure.
2089 static struct cpumask cpuhp_online_new
;
2091 static void blk_mq_queue_reinit_work(void)
2093 struct request_queue
*q
;
2095 mutex_lock(&all_q_mutex
);
2097 * We need to freeze and reinit all existing queues. Freezing
2098 * involves synchronous wait for an RCU grace period and doing it
2099 * one by one may take a long time. Start freezing all queues in
2100 * one swoop and then wait for the completions so that freezing can
2101 * take place in parallel.
2103 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2104 blk_mq_freeze_queue_start(q
);
2105 list_for_each_entry(q
, &all_q_list
, all_q_node
) {
2106 blk_mq_freeze_queue_wait(q
);
2109 * timeout handler can't touch hw queue during the
2112 del_timer_sync(&q
->timeout
);
2115 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2116 blk_mq_queue_reinit(q
, &cpuhp_online_new
);
2118 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2119 blk_mq_unfreeze_queue(q
);
2121 mutex_unlock(&all_q_mutex
);
2124 static int blk_mq_queue_reinit_dead(unsigned int cpu
)
2126 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2127 blk_mq_queue_reinit_work();
2132 * Before hotadded cpu starts handling requests, new mappings must be
2133 * established. Otherwise, these requests in hw queue might never be
2136 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2137 * for CPU0, and ctx1 for CPU1).
2139 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2140 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2142 * And then while running hw queue, flush_busy_ctxs() finds bit0 is set in
2143 * pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2144 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list
2147 static int blk_mq_queue_reinit_prepare(unsigned int cpu
)
2149 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2150 cpumask_set_cpu(cpu
, &cpuhp_online_new
);
2151 blk_mq_queue_reinit_work();
2155 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2159 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2160 set
->tags
[i
] = blk_mq_init_rq_map(set
, i
);
2169 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2175 * Allocate the request maps associated with this tag_set. Note that this
2176 * may reduce the depth asked for, if memory is tight. set->queue_depth
2177 * will be updated to reflect the allocated depth.
2179 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2184 depth
= set
->queue_depth
;
2186 err
= __blk_mq_alloc_rq_maps(set
);
2190 set
->queue_depth
>>= 1;
2191 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2195 } while (set
->queue_depth
);
2197 if (!set
->queue_depth
|| err
) {
2198 pr_err("blk-mq: failed to allocate request map\n");
2202 if (depth
!= set
->queue_depth
)
2203 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2204 depth
, set
->queue_depth
);
2210 * Alloc a tag set to be associated with one or more request queues.
2211 * May fail with EINVAL for various error conditions. May adjust the
2212 * requested depth down, if if it too large. In that case, the set
2213 * value will be stored in set->queue_depth.
2215 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2219 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2221 if (!set
->nr_hw_queues
)
2223 if (!set
->queue_depth
)
2225 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2228 if (!set
->ops
->queue_rq
)
2231 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2232 pr_info("blk-mq: reduced tag depth to %u\n",
2234 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2238 * If a crashdump is active, then we are potentially in a very
2239 * memory constrained environment. Limit us to 1 queue and
2240 * 64 tags to prevent using too much memory.
2242 if (is_kdump_kernel()) {
2243 set
->nr_hw_queues
= 1;
2244 set
->queue_depth
= min(64U, set
->queue_depth
);
2247 * There is no use for more h/w queues than cpus.
2249 if (set
->nr_hw_queues
> nr_cpu_ids
)
2250 set
->nr_hw_queues
= nr_cpu_ids
;
2252 set
->tags
= kzalloc_node(nr_cpu_ids
* sizeof(struct blk_mq_tags
*),
2253 GFP_KERNEL
, set
->numa_node
);
2258 set
->mq_map
= kzalloc_node(sizeof(*set
->mq_map
) * nr_cpu_ids
,
2259 GFP_KERNEL
, set
->numa_node
);
2263 if (set
->ops
->map_queues
)
2264 ret
= set
->ops
->map_queues(set
);
2266 ret
= blk_mq_map_queues(set
);
2268 goto out_free_mq_map
;
2270 ret
= blk_mq_alloc_rq_maps(set
);
2272 goto out_free_mq_map
;
2274 mutex_init(&set
->tag_list_lock
);
2275 INIT_LIST_HEAD(&set
->tag_list
);
2287 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2289 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2293 for (i
= 0; i
< nr_cpu_ids
; i
++) {
2295 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2304 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2306 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2308 struct blk_mq_tag_set
*set
= q
->tag_set
;
2309 struct blk_mq_hw_ctx
*hctx
;
2312 if (!set
|| nr
> set
->queue_depth
)
2316 queue_for_each_hw_ctx(q
, hctx
, i
) {
2319 ret
= blk_mq_tag_update_depth(hctx
->tags
, nr
);
2325 q
->nr_requests
= nr
;
2330 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
2332 struct request_queue
*q
;
2334 if (nr_hw_queues
> nr_cpu_ids
)
2335 nr_hw_queues
= nr_cpu_ids
;
2336 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
2339 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2340 blk_mq_freeze_queue(q
);
2342 set
->nr_hw_queues
= nr_hw_queues
;
2343 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2344 blk_mq_realloc_hw_ctxs(set
, q
);
2346 if (q
->nr_hw_queues
> 1)
2347 blk_queue_make_request(q
, blk_mq_make_request
);
2349 blk_queue_make_request(q
, blk_sq_make_request
);
2351 blk_mq_queue_reinit(q
, cpu_online_mask
);
2354 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2355 blk_mq_unfreeze_queue(q
);
2357 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
2359 void blk_mq_disable_hotplug(void)
2361 mutex_lock(&all_q_mutex
);
2364 void blk_mq_enable_hotplug(void)
2366 mutex_unlock(&all_q_mutex
);
2369 static int __init
blk_mq_init(void)
2371 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
2372 blk_mq_hctx_notify_dead
);
2374 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE
, "block/mq:prepare",
2375 blk_mq_queue_reinit_prepare
,
2376 blk_mq_queue_reinit_dead
);
2379 subsys_initcall(blk_mq_init
);