2 * Block multiqueue core code
4 * Copyright (C) 2013-2014 Jens Axboe
5 * Copyright (C) 2013-2014 Christoph Hellwig
7 #include <linux/kernel.h>
8 #include <linux/module.h>
9 #include <linux/backing-dev.h>
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
13 #include <linux/init.h>
14 #include <linux/slab.h>
15 #include <linux/workqueue.h>
16 #include <linux/smp.h>
17 #include <linux/llist.h>
18 #include <linux/list_sort.h>
19 #include <linux/cpu.h>
20 #include <linux/cache.h>
21 #include <linux/sched/sysctl.h>
22 #include <linux/delay.h>
23 #include <linux/crash_dump.h>
25 #include <trace/events/block.h>
27 #include <linux/blk-mq.h>
30 #include "blk-mq-tag.h"
32 static DEFINE_MUTEX(all_q_mutex
);
33 static LIST_HEAD(all_q_list
);
35 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
);
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
)
44 for (i
= 0; i
< hctx
->ctx_map
.map_size
; i
++)
45 if (hctx
->ctx_map
.map
[i
].word
)
51 static inline struct blk_align_bitmap
*get_bm(struct blk_mq_hw_ctx
*hctx
,
52 struct blk_mq_ctx
*ctx
)
54 return &hctx
->ctx_map
.map
[ctx
->index_hw
/ hctx
->ctx_map
.bits_per_word
];
57 #define CTX_TO_BIT(hctx, ctx) \
58 ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
61 * Mark this ctx as having pending work in this hardware queue
63 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
64 struct blk_mq_ctx
*ctx
)
66 struct blk_align_bitmap
*bm
= get_bm(hctx
, ctx
);
68 if (!test_bit(CTX_TO_BIT(hctx
, ctx
), &bm
->word
))
69 set_bit(CTX_TO_BIT(hctx
, ctx
), &bm
->word
);
72 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
73 struct blk_mq_ctx
*ctx
)
75 struct blk_align_bitmap
*bm
= get_bm(hctx
, ctx
);
77 clear_bit(CTX_TO_BIT(hctx
, ctx
), &bm
->word
);
80 static int blk_mq_queue_enter(struct request_queue
*q
)
85 if (percpu_ref_tryget_live(&q
->mq_usage_counter
))
88 ret
= wait_event_interruptible(q
->mq_freeze_wq
,
89 !q
->mq_freeze_depth
|| blk_queue_dying(q
));
90 if (blk_queue_dying(q
))
97 static void blk_mq_queue_exit(struct request_queue
*q
)
99 percpu_ref_put(&q
->mq_usage_counter
);
102 static void blk_mq_usage_counter_release(struct percpu_ref
*ref
)
104 struct request_queue
*q
=
105 container_of(ref
, struct request_queue
, mq_usage_counter
);
107 wake_up_all(&q
->mq_freeze_wq
);
110 void blk_mq_freeze_queue_start(struct request_queue
*q
)
114 spin_lock_irq(q
->queue_lock
);
115 freeze
= !q
->mq_freeze_depth
++;
116 spin_unlock_irq(q
->queue_lock
);
119 percpu_ref_kill(&q
->mq_usage_counter
);
120 blk_mq_run_queues(q
, false);
123 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start
);
125 static void blk_mq_freeze_queue_wait(struct request_queue
*q
)
127 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->mq_usage_counter
));
131 * Guarantee no request is in use, so we can change any data structure of
132 * the queue afterward.
134 void blk_mq_freeze_queue(struct request_queue
*q
)
136 blk_mq_freeze_queue_start(q
);
137 blk_mq_freeze_queue_wait(q
);
140 void blk_mq_unfreeze_queue(struct request_queue
*q
)
144 spin_lock_irq(q
->queue_lock
);
145 wake
= !--q
->mq_freeze_depth
;
146 WARN_ON_ONCE(q
->mq_freeze_depth
< 0);
147 spin_unlock_irq(q
->queue_lock
);
149 percpu_ref_reinit(&q
->mq_usage_counter
);
150 wake_up_all(&q
->mq_freeze_wq
);
153 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
155 void blk_mq_wake_waiters(struct request_queue
*q
)
157 struct blk_mq_hw_ctx
*hctx
;
160 queue_for_each_hw_ctx(q
, hctx
, i
)
161 if (blk_mq_hw_queue_mapped(hctx
))
162 blk_mq_tag_wakeup_all(hctx
->tags
, true);
165 * If we are called because the queue has now been marked as
166 * dying, we need to ensure that processes currently waiting on
167 * the queue are notified as well.
169 wake_up_all(&q
->mq_freeze_wq
);
172 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
174 return blk_mq_has_free_tags(hctx
->tags
);
176 EXPORT_SYMBOL(blk_mq_can_queue
);
178 static void blk_mq_rq_ctx_init(struct request_queue
*q
, struct blk_mq_ctx
*ctx
,
179 struct request
*rq
, unsigned int rw_flags
)
181 if (blk_queue_io_stat(q
))
182 rw_flags
|= REQ_IO_STAT
;
184 INIT_LIST_HEAD(&rq
->queuelist
);
185 /* csd/requeue_work/fifo_time is initialized before use */
188 rq
->cmd_flags
|= rw_flags
;
189 /* do not touch atomic flags, it needs atomic ops against the timer */
191 INIT_HLIST_NODE(&rq
->hash
);
192 RB_CLEAR_NODE(&rq
->rb_node
);
195 rq
->start_time
= jiffies
;
196 #ifdef CONFIG_BLK_CGROUP
198 set_start_time_ns(rq
);
199 rq
->io_start_time_ns
= 0;
201 rq
->nr_phys_segments
= 0;
202 #if defined(CONFIG_BLK_DEV_INTEGRITY)
203 rq
->nr_integrity_segments
= 0;
206 /* tag was already set */
216 INIT_LIST_HEAD(&rq
->timeout_list
);
220 rq
->end_io_data
= NULL
;
223 ctx
->rq_dispatched
[rw_is_sync(rw_flags
)]++;
226 static struct request
*
227 __blk_mq_alloc_request(struct blk_mq_alloc_data
*data
, int rw
)
232 tag
= blk_mq_get_tag(data
);
233 if (tag
!= BLK_MQ_TAG_FAIL
) {
234 rq
= data
->hctx
->tags
->rqs
[tag
];
236 if (blk_mq_tag_busy(data
->hctx
)) {
237 rq
->cmd_flags
= REQ_MQ_INFLIGHT
;
238 atomic_inc(&data
->hctx
->nr_active
);
242 blk_mq_rq_ctx_init(data
->q
, data
->ctx
, rq
, rw
);
249 struct request
*blk_mq_alloc_request(struct request_queue
*q
, int rw
, gfp_t gfp
,
252 struct blk_mq_ctx
*ctx
;
253 struct blk_mq_hw_ctx
*hctx
;
255 struct blk_mq_alloc_data alloc_data
;
258 ret
= blk_mq_queue_enter(q
);
262 ctx
= blk_mq_get_ctx(q
);
263 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
264 blk_mq_set_alloc_data(&alloc_data
, q
, gfp
& ~__GFP_WAIT
,
265 reserved
, ctx
, hctx
);
267 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
268 if (!rq
&& (gfp
& __GFP_WAIT
)) {
269 __blk_mq_run_hw_queue(hctx
);
272 ctx
= blk_mq_get_ctx(q
);
273 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
274 blk_mq_set_alloc_data(&alloc_data
, q
, gfp
, reserved
, ctx
,
276 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
277 ctx
= alloc_data
.ctx
;
281 blk_mq_queue_exit(q
);
282 return ERR_PTR(-EWOULDBLOCK
);
286 EXPORT_SYMBOL(blk_mq_alloc_request
);
288 static void __blk_mq_free_request(struct blk_mq_hw_ctx
*hctx
,
289 struct blk_mq_ctx
*ctx
, struct request
*rq
)
291 const int tag
= rq
->tag
;
292 struct request_queue
*q
= rq
->q
;
294 if (rq
->cmd_flags
& REQ_MQ_INFLIGHT
)
295 atomic_dec(&hctx
->nr_active
);
298 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
299 blk_mq_put_tag(hctx
, tag
, &ctx
->last_tag
);
300 blk_mq_queue_exit(q
);
303 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
305 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
307 ctx
->rq_completed
[rq_is_sync(rq
)]++;
308 __blk_mq_free_request(hctx
, ctx
, rq
);
311 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request
);
313 void blk_mq_free_request(struct request
*rq
)
315 struct blk_mq_hw_ctx
*hctx
;
316 struct request_queue
*q
= rq
->q
;
318 hctx
= q
->mq_ops
->map_queue(q
, rq
->mq_ctx
->cpu
);
319 blk_mq_free_hctx_request(hctx
, rq
);
321 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
323 inline void __blk_mq_end_request(struct request
*rq
, int error
)
325 blk_account_io_done(rq
);
328 rq
->end_io(rq
, error
);
330 if (unlikely(blk_bidi_rq(rq
)))
331 blk_mq_free_request(rq
->next_rq
);
332 blk_mq_free_request(rq
);
335 EXPORT_SYMBOL(__blk_mq_end_request
);
337 void blk_mq_end_request(struct request
*rq
, int error
)
339 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
341 __blk_mq_end_request(rq
, error
);
343 EXPORT_SYMBOL(blk_mq_end_request
);
345 static void __blk_mq_complete_request_remote(void *data
)
347 struct request
*rq
= data
;
349 rq
->q
->softirq_done_fn(rq
);
352 static void blk_mq_ipi_complete_request(struct request
*rq
)
354 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
358 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
359 rq
->q
->softirq_done_fn(rq
);
364 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
365 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
367 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
368 rq
->csd
.func
= __blk_mq_complete_request_remote
;
371 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
373 rq
->q
->softirq_done_fn(rq
);
378 void __blk_mq_complete_request(struct request
*rq
)
380 struct request_queue
*q
= rq
->q
;
382 if (!q
->softirq_done_fn
)
383 blk_mq_end_request(rq
, rq
->errors
);
385 blk_mq_ipi_complete_request(rq
);
389 * blk_mq_complete_request - end I/O on a request
390 * @rq: the request being processed
393 * Ends all I/O on a request. It does not handle partial completions.
394 * The actual completion happens out-of-order, through a IPI handler.
396 void blk_mq_complete_request(struct request
*rq
)
398 struct request_queue
*q
= rq
->q
;
400 if (unlikely(blk_should_fake_timeout(q
)))
402 if (!blk_mark_rq_complete(rq
))
403 __blk_mq_complete_request(rq
);
405 EXPORT_SYMBOL(blk_mq_complete_request
);
407 int blk_mq_request_started(struct request
*rq
)
409 return test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
411 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
413 void blk_mq_start_request(struct request
*rq
)
415 struct request_queue
*q
= rq
->q
;
417 trace_block_rq_issue(q
, rq
);
419 rq
->resid_len
= blk_rq_bytes(rq
);
420 if (unlikely(blk_bidi_rq(rq
)))
421 rq
->next_rq
->resid_len
= blk_rq_bytes(rq
->next_rq
);
426 * Ensure that ->deadline is visible before set the started
427 * flag and clear the completed flag.
429 smp_mb__before_atomic();
432 * Mark us as started and clear complete. Complete might have been
433 * set if requeue raced with timeout, which then marked it as
434 * complete. So be sure to clear complete again when we start
435 * the request, otherwise we'll ignore the completion event.
437 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
438 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
439 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
440 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
442 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
444 * Make sure space for the drain appears. We know we can do
445 * this because max_hw_segments has been adjusted to be one
446 * fewer than the device can handle.
448 rq
->nr_phys_segments
++;
451 EXPORT_SYMBOL(blk_mq_start_request
);
453 static void __blk_mq_requeue_request(struct request
*rq
)
455 struct request_queue
*q
= rq
->q
;
457 trace_block_rq_requeue(q
, rq
);
459 if (test_and_clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
460 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
461 rq
->nr_phys_segments
--;
465 void blk_mq_requeue_request(struct request
*rq
)
467 __blk_mq_requeue_request(rq
);
469 BUG_ON(blk_queued_rq(rq
));
470 blk_mq_add_to_requeue_list(rq
, true);
472 EXPORT_SYMBOL(blk_mq_requeue_request
);
474 static void blk_mq_requeue_work(struct work_struct
*work
)
476 struct request_queue
*q
=
477 container_of(work
, struct request_queue
, requeue_work
);
479 struct request
*rq
, *next
;
482 spin_lock_irqsave(&q
->requeue_lock
, flags
);
483 list_splice_init(&q
->requeue_list
, &rq_list
);
484 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
486 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
487 if (!(rq
->cmd_flags
& REQ_SOFTBARRIER
))
490 rq
->cmd_flags
&= ~REQ_SOFTBARRIER
;
491 list_del_init(&rq
->queuelist
);
492 blk_mq_insert_request(rq
, true, false, false);
495 while (!list_empty(&rq_list
)) {
496 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
497 list_del_init(&rq
->queuelist
);
498 blk_mq_insert_request(rq
, false, false, false);
502 * Use the start variant of queue running here, so that running
503 * the requeue work will kick stopped queues.
505 blk_mq_start_hw_queues(q
);
508 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
)
510 struct request_queue
*q
= rq
->q
;
514 * We abuse this flag that is otherwise used by the I/O scheduler to
515 * request head insertation from the workqueue.
517 BUG_ON(rq
->cmd_flags
& REQ_SOFTBARRIER
);
519 spin_lock_irqsave(&q
->requeue_lock
, flags
);
521 rq
->cmd_flags
|= REQ_SOFTBARRIER
;
522 list_add(&rq
->queuelist
, &q
->requeue_list
);
524 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
526 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
528 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
530 void blk_mq_cancel_requeue_work(struct request_queue
*q
)
532 cancel_work_sync(&q
->requeue_work
);
534 EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work
);
536 void blk_mq_kick_requeue_list(struct request_queue
*q
)
538 kblockd_schedule_work(&q
->requeue_work
);
540 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
542 void blk_mq_abort_requeue_list(struct request_queue
*q
)
547 spin_lock_irqsave(&q
->requeue_lock
, flags
);
548 list_splice_init(&q
->requeue_list
, &rq_list
);
549 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
551 while (!list_empty(&rq_list
)) {
554 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
555 list_del_init(&rq
->queuelist
);
557 blk_mq_end_request(rq
, rq
->errors
);
560 EXPORT_SYMBOL(blk_mq_abort_requeue_list
);
562 static inline bool is_flush_request(struct request
*rq
,
563 struct blk_flush_queue
*fq
, unsigned int tag
)
565 return ((rq
->cmd_flags
& REQ_FLUSH_SEQ
) &&
566 fq
->flush_rq
->tag
== tag
);
569 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
571 struct request
*rq
= tags
->rqs
[tag
];
572 /* mq_ctx of flush rq is always cloned from the corresponding req */
573 struct blk_flush_queue
*fq
= blk_get_flush_queue(rq
->q
, rq
->mq_ctx
);
575 if (!is_flush_request(rq
, fq
, tag
))
580 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
582 struct blk_mq_timeout_data
{
584 unsigned int next_set
;
587 void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
589 struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
590 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
593 * We know that complete is set at this point. If STARTED isn't set
594 * anymore, then the request isn't active and the "timeout" should
595 * just be ignored. This can happen due to the bitflag ordering.
596 * Timeout first checks if STARTED is set, and if it is, assumes
597 * the request is active. But if we race with completion, then
598 * we both flags will get cleared. So check here again, and ignore
599 * a timeout event with a request that isn't active.
601 if (!test_bit(REQ_ATOM_STARTED
, &req
->atomic_flags
))
605 ret
= ops
->timeout(req
, reserved
);
609 __blk_mq_complete_request(req
);
611 case BLK_EH_RESET_TIMER
:
613 blk_clear_rq_complete(req
);
615 case BLK_EH_NOT_HANDLED
:
618 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
623 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
624 struct request
*rq
, void *priv
, bool reserved
)
626 struct blk_mq_timeout_data
*data
= priv
;
628 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
630 if (rq
->cmd_flags
& REQ_NO_TIMEOUT
)
633 if (time_after_eq(jiffies
, rq
->deadline
)) {
634 if (!blk_mark_rq_complete(rq
))
635 blk_mq_rq_timed_out(rq
, reserved
);
636 } else if (!data
->next_set
|| time_after(data
->next
, rq
->deadline
)) {
637 data
->next
= rq
->deadline
;
642 static void blk_mq_rq_timer(unsigned long priv
)
644 struct request_queue
*q
= (struct request_queue
*)priv
;
645 struct blk_mq_timeout_data data
= {
649 struct blk_mq_hw_ctx
*hctx
;
652 queue_for_each_hw_ctx(q
, hctx
, i
) {
654 * If not software queues are currently mapped to this
655 * hardware queue, there's nothing to check
657 if (!blk_mq_hw_queue_mapped(hctx
))
660 blk_mq_tag_busy_iter(hctx
, blk_mq_check_expired
, &data
);
664 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
665 mod_timer(&q
->timeout
, data
.next
);
667 queue_for_each_hw_ctx(q
, hctx
, i
)
668 blk_mq_tag_idle(hctx
);
673 * Reverse check our software queue for entries that we could potentially
674 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
675 * too much time checking for merges.
677 static bool blk_mq_attempt_merge(struct request_queue
*q
,
678 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
683 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
689 if (!blk_rq_merge_ok(rq
, bio
))
692 el_ret
= blk_try_merge(rq
, bio
);
693 if (el_ret
== ELEVATOR_BACK_MERGE
) {
694 if (bio_attempt_back_merge(q
, rq
, bio
)) {
699 } else if (el_ret
== ELEVATOR_FRONT_MERGE
) {
700 if (bio_attempt_front_merge(q
, rq
, bio
)) {
712 * Process software queues that have been marked busy, splicing them
713 * to the for-dispatch
715 static void flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
717 struct blk_mq_ctx
*ctx
;
720 for (i
= 0; i
< hctx
->ctx_map
.map_size
; i
++) {
721 struct blk_align_bitmap
*bm
= &hctx
->ctx_map
.map
[i
];
722 unsigned int off
, bit
;
728 off
= i
* hctx
->ctx_map
.bits_per_word
;
730 bit
= find_next_bit(&bm
->word
, bm
->depth
, bit
);
731 if (bit
>= bm
->depth
)
734 ctx
= hctx
->ctxs
[bit
+ off
];
735 clear_bit(bit
, &bm
->word
);
736 spin_lock(&ctx
->lock
);
737 list_splice_tail_init(&ctx
->rq_list
, list
);
738 spin_unlock(&ctx
->lock
);
746 * Run this hardware queue, pulling any software queues mapped to it in.
747 * Note that this function currently has various problems around ordering
748 * of IO. In particular, we'd like FIFO behaviour on handling existing
749 * items on the hctx->dispatch list. Ignore that for now.
751 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
753 struct request_queue
*q
= hctx
->queue
;
756 LIST_HEAD(driver_list
);
757 struct list_head
*dptr
;
760 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
));
762 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)))
768 * Touch any software queue that has pending entries.
770 flush_busy_ctxs(hctx
, &rq_list
);
773 * If we have previous entries on our dispatch list, grab them
774 * and stuff them at the front for more fair dispatch.
776 if (!list_empty_careful(&hctx
->dispatch
)) {
777 spin_lock(&hctx
->lock
);
778 if (!list_empty(&hctx
->dispatch
))
779 list_splice_init(&hctx
->dispatch
, &rq_list
);
780 spin_unlock(&hctx
->lock
);
784 * Start off with dptr being NULL, so we start the first request
785 * immediately, even if we have more pending.
790 * Now process all the entries, sending them to the driver.
793 while (!list_empty(&rq_list
)) {
794 struct blk_mq_queue_data bd
;
797 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
798 list_del_init(&rq
->queuelist
);
802 bd
.last
= list_empty(&rq_list
);
804 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
806 case BLK_MQ_RQ_QUEUE_OK
:
809 case BLK_MQ_RQ_QUEUE_BUSY
:
810 list_add(&rq
->queuelist
, &rq_list
);
811 __blk_mq_requeue_request(rq
);
814 pr_err("blk-mq: bad return on queue: %d\n", ret
);
815 case BLK_MQ_RQ_QUEUE_ERROR
:
817 blk_mq_end_request(rq
, rq
->errors
);
821 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
825 * We've done the first request. If we have more than 1
826 * left in the list, set dptr to defer issue.
828 if (!dptr
&& rq_list
.next
!= rq_list
.prev
)
833 hctx
->dispatched
[0]++;
834 else if (queued
< (1 << (BLK_MQ_MAX_DISPATCH_ORDER
- 1)))
835 hctx
->dispatched
[ilog2(queued
) + 1]++;
838 * Any items that need requeuing? Stuff them into hctx->dispatch,
839 * that is where we will continue on next queue run.
841 if (!list_empty(&rq_list
)) {
842 spin_lock(&hctx
->lock
);
843 list_splice(&rq_list
, &hctx
->dispatch
);
844 spin_unlock(&hctx
->lock
);
849 * It'd be great if the workqueue API had a way to pass
850 * in a mask and had some smarts for more clever placement.
851 * For now we just round-robin here, switching for every
852 * BLK_MQ_CPU_WORK_BATCH queued items.
854 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
856 if (hctx
->queue
->nr_hw_queues
== 1)
857 return WORK_CPU_UNBOUND
;
859 if (--hctx
->next_cpu_batch
<= 0) {
860 int cpu
= hctx
->next_cpu
, next_cpu
;
862 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
863 if (next_cpu
>= nr_cpu_ids
)
864 next_cpu
= cpumask_first(hctx
->cpumask
);
866 hctx
->next_cpu
= next_cpu
;
867 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
872 return hctx
->next_cpu
;
875 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
877 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
) ||
878 !blk_mq_hw_queue_mapped(hctx
)))
883 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
884 __blk_mq_run_hw_queue(hctx
);
892 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
896 void blk_mq_run_queues(struct request_queue
*q
, bool async
)
898 struct blk_mq_hw_ctx
*hctx
;
901 queue_for_each_hw_ctx(q
, hctx
, i
) {
902 if ((!blk_mq_hctx_has_pending(hctx
) &&
903 list_empty_careful(&hctx
->dispatch
)) ||
904 test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
907 blk_mq_run_hw_queue(hctx
, async
);
910 EXPORT_SYMBOL(blk_mq_run_queues
);
912 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
914 cancel_delayed_work(&hctx
->run_work
);
915 cancel_delayed_work(&hctx
->delay_work
);
916 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
918 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
920 void blk_mq_stop_hw_queues(struct request_queue
*q
)
922 struct blk_mq_hw_ctx
*hctx
;
925 queue_for_each_hw_ctx(q
, hctx
, i
)
926 blk_mq_stop_hw_queue(hctx
);
928 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
930 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
932 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
934 blk_mq_run_hw_queue(hctx
, false);
936 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
938 void blk_mq_start_hw_queues(struct request_queue
*q
)
940 struct blk_mq_hw_ctx
*hctx
;
943 queue_for_each_hw_ctx(q
, hctx
, i
)
944 blk_mq_start_hw_queue(hctx
);
946 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
949 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
951 struct blk_mq_hw_ctx
*hctx
;
954 queue_for_each_hw_ctx(q
, hctx
, i
) {
955 if (!test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
958 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
959 blk_mq_run_hw_queue(hctx
, async
);
962 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
964 static void blk_mq_run_work_fn(struct work_struct
*work
)
966 struct blk_mq_hw_ctx
*hctx
;
968 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
970 __blk_mq_run_hw_queue(hctx
);
973 static void blk_mq_delay_work_fn(struct work_struct
*work
)
975 struct blk_mq_hw_ctx
*hctx
;
977 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
979 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
980 __blk_mq_run_hw_queue(hctx
);
983 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
985 if (unlikely(!blk_mq_hw_queue_mapped(hctx
)))
988 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
989 &hctx
->delay_work
, msecs_to_jiffies(msecs
));
991 EXPORT_SYMBOL(blk_mq_delay_queue
);
993 static void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
,
994 struct request
*rq
, bool at_head
)
996 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
998 trace_block_rq_insert(hctx
->queue
, rq
);
1001 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1003 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1005 blk_mq_hctx_mark_pending(hctx
, ctx
);
1008 void blk_mq_insert_request(struct request
*rq
, bool at_head
, bool run_queue
,
1011 struct request_queue
*q
= rq
->q
;
1012 struct blk_mq_hw_ctx
*hctx
;
1013 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
, *current_ctx
;
1015 current_ctx
= blk_mq_get_ctx(q
);
1016 if (!cpu_online(ctx
->cpu
))
1017 rq
->mq_ctx
= ctx
= current_ctx
;
1019 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1021 spin_lock(&ctx
->lock
);
1022 __blk_mq_insert_request(hctx
, rq
, at_head
);
1023 spin_unlock(&ctx
->lock
);
1026 blk_mq_run_hw_queue(hctx
, async
);
1028 blk_mq_put_ctx(current_ctx
);
1031 static void blk_mq_insert_requests(struct request_queue
*q
,
1032 struct blk_mq_ctx
*ctx
,
1033 struct list_head
*list
,
1038 struct blk_mq_hw_ctx
*hctx
;
1039 struct blk_mq_ctx
*current_ctx
;
1041 trace_block_unplug(q
, depth
, !from_schedule
);
1043 current_ctx
= blk_mq_get_ctx(q
);
1045 if (!cpu_online(ctx
->cpu
))
1047 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1050 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1053 spin_lock(&ctx
->lock
);
1054 while (!list_empty(list
)) {
1057 rq
= list_first_entry(list
, struct request
, queuelist
);
1058 list_del_init(&rq
->queuelist
);
1060 __blk_mq_insert_request(hctx
, rq
, false);
1062 spin_unlock(&ctx
->lock
);
1064 blk_mq_run_hw_queue(hctx
, from_schedule
);
1065 blk_mq_put_ctx(current_ctx
);
1068 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1070 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1071 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1073 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1074 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1075 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1078 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1080 struct blk_mq_ctx
*this_ctx
;
1081 struct request_queue
*this_q
;
1084 LIST_HEAD(ctx_list
);
1087 list_splice_init(&plug
->mq_list
, &list
);
1089 list_sort(NULL
, &list
, plug_ctx_cmp
);
1095 while (!list_empty(&list
)) {
1096 rq
= list_entry_rq(list
.next
);
1097 list_del_init(&rq
->queuelist
);
1099 if (rq
->mq_ctx
!= this_ctx
) {
1101 blk_mq_insert_requests(this_q
, this_ctx
,
1106 this_ctx
= rq
->mq_ctx
;
1112 list_add_tail(&rq
->queuelist
, &ctx_list
);
1116 * If 'this_ctx' is set, we know we have entries to complete
1117 * on 'ctx_list'. Do those.
1120 blk_mq_insert_requests(this_q
, this_ctx
, &ctx_list
, depth
,
1125 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1127 init_request_from_bio(rq
, bio
);
1129 if (blk_do_io_stat(rq
))
1130 blk_account_io_start(rq
, 1);
1133 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1135 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1136 !blk_queue_nomerges(hctx
->queue
);
1139 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx
*hctx
,
1140 struct blk_mq_ctx
*ctx
,
1141 struct request
*rq
, struct bio
*bio
)
1143 if (!hctx_allow_merges(hctx
)) {
1144 blk_mq_bio_to_request(rq
, bio
);
1145 spin_lock(&ctx
->lock
);
1147 __blk_mq_insert_request(hctx
, rq
, false);
1148 spin_unlock(&ctx
->lock
);
1151 struct request_queue
*q
= hctx
->queue
;
1153 spin_lock(&ctx
->lock
);
1154 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1155 blk_mq_bio_to_request(rq
, bio
);
1159 spin_unlock(&ctx
->lock
);
1160 __blk_mq_free_request(hctx
, ctx
, rq
);
1165 struct blk_map_ctx
{
1166 struct blk_mq_hw_ctx
*hctx
;
1167 struct blk_mq_ctx
*ctx
;
1170 static struct request
*blk_mq_map_request(struct request_queue
*q
,
1172 struct blk_map_ctx
*data
)
1174 struct blk_mq_hw_ctx
*hctx
;
1175 struct blk_mq_ctx
*ctx
;
1177 int rw
= bio_data_dir(bio
);
1178 struct blk_mq_alloc_data alloc_data
;
1180 if (unlikely(blk_mq_queue_enter(q
))) {
1181 bio_endio(bio
, -EIO
);
1185 ctx
= blk_mq_get_ctx(q
);
1186 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1188 if (rw_is_sync(bio
->bi_rw
))
1191 trace_block_getrq(q
, bio
, rw
);
1192 blk_mq_set_alloc_data(&alloc_data
, q
, GFP_ATOMIC
, false, ctx
,
1194 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
1195 if (unlikely(!rq
)) {
1196 __blk_mq_run_hw_queue(hctx
);
1197 blk_mq_put_ctx(ctx
);
1198 trace_block_sleeprq(q
, bio
, rw
);
1200 ctx
= blk_mq_get_ctx(q
);
1201 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1202 blk_mq_set_alloc_data(&alloc_data
, q
,
1203 __GFP_WAIT
|GFP_ATOMIC
, false, ctx
, hctx
);
1204 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
1205 ctx
= alloc_data
.ctx
;
1206 hctx
= alloc_data
.hctx
;
1216 * Multiple hardware queue variant. This will not use per-process plugs,
1217 * but will attempt to bypass the hctx queueing if we can go straight to
1218 * hardware for SYNC IO.
1220 static void blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1222 const int is_sync
= rw_is_sync(bio
->bi_rw
);
1223 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
1224 struct blk_map_ctx data
;
1227 blk_queue_bounce(q
, &bio
);
1229 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1230 bio_endio(bio
, -EIO
);
1234 rq
= blk_mq_map_request(q
, bio
, &data
);
1238 if (unlikely(is_flush_fua
)) {
1239 blk_mq_bio_to_request(rq
, bio
);
1240 blk_insert_flush(rq
);
1245 * If the driver supports defer issued based on 'last', then
1246 * queue it up like normal since we can potentially save some
1249 if (is_sync
&& !(data
.hctx
->flags
& BLK_MQ_F_DEFER_ISSUE
)) {
1250 struct blk_mq_queue_data bd
= {
1257 blk_mq_bio_to_request(rq
, bio
);
1260 * For OK queue, we are done. For error, kill it. Any other
1261 * error (busy), just add it to our list as we previously
1264 ret
= q
->mq_ops
->queue_rq(data
.hctx
, &bd
);
1265 if (ret
== BLK_MQ_RQ_QUEUE_OK
)
1268 __blk_mq_requeue_request(rq
);
1270 if (ret
== BLK_MQ_RQ_QUEUE_ERROR
) {
1272 blk_mq_end_request(rq
, rq
->errors
);
1278 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1280 * For a SYNC request, send it to the hardware immediately. For
1281 * an ASYNC request, just ensure that we run it later on. The
1282 * latter allows for merging opportunities and more efficient
1286 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1289 blk_mq_put_ctx(data
.ctx
);
1293 * Single hardware queue variant. This will attempt to use any per-process
1294 * plug for merging and IO deferral.
1296 static void blk_sq_make_request(struct request_queue
*q
, struct bio
*bio
)
1298 const int is_sync
= rw_is_sync(bio
->bi_rw
);
1299 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
1300 unsigned int use_plug
, request_count
= 0;
1301 struct blk_map_ctx data
;
1305 * If we have multiple hardware queues, just go directly to
1306 * one of those for sync IO.
1308 use_plug
= !is_flush_fua
&& !is_sync
;
1310 blk_queue_bounce(q
, &bio
);
1312 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1313 bio_endio(bio
, -EIO
);
1317 if (use_plug
&& !blk_queue_nomerges(q
) &&
1318 blk_attempt_plug_merge(q
, bio
, &request_count
))
1321 rq
= blk_mq_map_request(q
, bio
, &data
);
1325 if (unlikely(is_flush_fua
)) {
1326 blk_mq_bio_to_request(rq
, bio
);
1327 blk_insert_flush(rq
);
1332 * A task plug currently exists. Since this is completely lockless,
1333 * utilize that to temporarily store requests until the task is
1334 * either done or scheduled away.
1337 struct blk_plug
*plug
= current
->plug
;
1340 blk_mq_bio_to_request(rq
, bio
);
1341 if (list_empty(&plug
->mq_list
))
1342 trace_block_plug(q
);
1343 else if (request_count
>= BLK_MAX_REQUEST_COUNT
) {
1344 blk_flush_plug_list(plug
, false);
1345 trace_block_plug(q
);
1347 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1348 blk_mq_put_ctx(data
.ctx
);
1353 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1355 * For a SYNC request, send it to the hardware immediately. For
1356 * an ASYNC request, just ensure that we run it later on. The
1357 * latter allows for merging opportunities and more efficient
1361 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1364 blk_mq_put_ctx(data
.ctx
);
1368 * Default mapping to a software queue, since we use one per CPU.
1370 struct blk_mq_hw_ctx
*blk_mq_map_queue(struct request_queue
*q
, const int cpu
)
1372 return q
->queue_hw_ctx
[q
->mq_map
[cpu
]];
1374 EXPORT_SYMBOL(blk_mq_map_queue
);
1376 static void blk_mq_free_rq_map(struct blk_mq_tag_set
*set
,
1377 struct blk_mq_tags
*tags
, unsigned int hctx_idx
)
1381 if (tags
->rqs
&& set
->ops
->exit_request
) {
1384 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1387 set
->ops
->exit_request(set
->driver_data
, tags
->rqs
[i
],
1389 tags
->rqs
[i
] = NULL
;
1393 while (!list_empty(&tags
->page_list
)) {
1394 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1395 list_del_init(&page
->lru
);
1396 __free_pages(page
, page
->private);
1401 blk_mq_free_tags(tags
);
1404 static size_t order_to_size(unsigned int order
)
1406 return (size_t)PAGE_SIZE
<< order
;
1409 static struct blk_mq_tags
*blk_mq_init_rq_map(struct blk_mq_tag_set
*set
,
1410 unsigned int hctx_idx
)
1412 struct blk_mq_tags
*tags
;
1413 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1414 size_t rq_size
, left
;
1416 tags
= blk_mq_init_tags(set
->queue_depth
, set
->reserved_tags
,
1421 INIT_LIST_HEAD(&tags
->page_list
);
1423 tags
->rqs
= kzalloc_node(set
->queue_depth
* sizeof(struct request
*),
1424 GFP_KERNEL
| __GFP_NOWARN
| __GFP_NORETRY
,
1427 blk_mq_free_tags(tags
);
1432 * rq_size is the size of the request plus driver payload, rounded
1433 * to the cacheline size
1435 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1437 left
= rq_size
* set
->queue_depth
;
1439 for (i
= 0; i
< set
->queue_depth
; ) {
1440 int this_order
= max_order
;
1445 while (left
< order_to_size(this_order
- 1) && this_order
)
1449 page
= alloc_pages_node(set
->numa_node
,
1450 GFP_KERNEL
| __GFP_NOWARN
| __GFP_NORETRY
,
1456 if (order_to_size(this_order
) < rq_size
)
1463 page
->private = this_order
;
1464 list_add_tail(&page
->lru
, &tags
->page_list
);
1466 p
= page_address(page
);
1467 entries_per_page
= order_to_size(this_order
) / rq_size
;
1468 to_do
= min(entries_per_page
, set
->queue_depth
- i
);
1469 left
-= to_do
* rq_size
;
1470 for (j
= 0; j
< to_do
; j
++) {
1472 tags
->rqs
[i
]->atomic_flags
= 0;
1473 tags
->rqs
[i
]->cmd_flags
= 0;
1474 if (set
->ops
->init_request
) {
1475 if (set
->ops
->init_request(set
->driver_data
,
1476 tags
->rqs
[i
], hctx_idx
, i
,
1478 tags
->rqs
[i
] = NULL
;
1491 blk_mq_free_rq_map(set
, tags
, hctx_idx
);
1495 static void blk_mq_free_bitmap(struct blk_mq_ctxmap
*bitmap
)
1500 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap
*bitmap
, int node
)
1502 unsigned int bpw
= 8, total
, num_maps
, i
;
1504 bitmap
->bits_per_word
= bpw
;
1506 num_maps
= ALIGN(nr_cpu_ids
, bpw
) / bpw
;
1507 bitmap
->map
= kzalloc_node(num_maps
* sizeof(struct blk_align_bitmap
),
1512 bitmap
->map_size
= num_maps
;
1515 for (i
= 0; i
< num_maps
; i
++) {
1516 bitmap
->map
[i
].depth
= min(total
, bitmap
->bits_per_word
);
1517 total
-= bitmap
->map
[i
].depth
;
1523 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx
*hctx
, int cpu
)
1525 struct request_queue
*q
= hctx
->queue
;
1526 struct blk_mq_ctx
*ctx
;
1530 * Move ctx entries to new CPU, if this one is going away.
1532 ctx
= __blk_mq_get_ctx(q
, cpu
);
1534 spin_lock(&ctx
->lock
);
1535 if (!list_empty(&ctx
->rq_list
)) {
1536 list_splice_init(&ctx
->rq_list
, &tmp
);
1537 blk_mq_hctx_clear_pending(hctx
, ctx
);
1539 spin_unlock(&ctx
->lock
);
1541 if (list_empty(&tmp
))
1544 ctx
= blk_mq_get_ctx(q
);
1545 spin_lock(&ctx
->lock
);
1547 while (!list_empty(&tmp
)) {
1550 rq
= list_first_entry(&tmp
, struct request
, queuelist
);
1552 list_move_tail(&rq
->queuelist
, &ctx
->rq_list
);
1555 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1556 blk_mq_hctx_mark_pending(hctx
, ctx
);
1558 spin_unlock(&ctx
->lock
);
1560 blk_mq_run_hw_queue(hctx
, true);
1561 blk_mq_put_ctx(ctx
);
1565 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx
*hctx
, int cpu
)
1567 struct request_queue
*q
= hctx
->queue
;
1568 struct blk_mq_tag_set
*set
= q
->tag_set
;
1570 if (set
->tags
[hctx
->queue_num
])
1573 set
->tags
[hctx
->queue_num
] = blk_mq_init_rq_map(set
, hctx
->queue_num
);
1574 if (!set
->tags
[hctx
->queue_num
])
1577 hctx
->tags
= set
->tags
[hctx
->queue_num
];
1581 static int blk_mq_hctx_notify(void *data
, unsigned long action
,
1584 struct blk_mq_hw_ctx
*hctx
= data
;
1586 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
)
1587 return blk_mq_hctx_cpu_offline(hctx
, cpu
);
1588 else if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
)
1589 return blk_mq_hctx_cpu_online(hctx
, cpu
);
1594 static void blk_mq_exit_hctx(struct request_queue
*q
,
1595 struct blk_mq_tag_set
*set
,
1596 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
1598 unsigned flush_start_tag
= set
->queue_depth
;
1600 blk_mq_tag_idle(hctx
);
1602 if (set
->ops
->exit_request
)
1603 set
->ops
->exit_request(set
->driver_data
,
1604 hctx
->fq
->flush_rq
, hctx_idx
,
1605 flush_start_tag
+ hctx_idx
);
1607 if (set
->ops
->exit_hctx
)
1608 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1610 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1611 blk_free_flush_queue(hctx
->fq
);
1613 blk_mq_free_bitmap(&hctx
->ctx_map
);
1616 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1617 struct blk_mq_tag_set
*set
, int nr_queue
)
1619 struct blk_mq_hw_ctx
*hctx
;
1622 queue_for_each_hw_ctx(q
, hctx
, i
) {
1625 blk_mq_exit_hctx(q
, set
, hctx
, i
);
1629 static void blk_mq_free_hw_queues(struct request_queue
*q
,
1630 struct blk_mq_tag_set
*set
)
1632 struct blk_mq_hw_ctx
*hctx
;
1635 queue_for_each_hw_ctx(q
, hctx
, i
) {
1636 free_cpumask_var(hctx
->cpumask
);
1641 static int blk_mq_init_hctx(struct request_queue
*q
,
1642 struct blk_mq_tag_set
*set
,
1643 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
1646 unsigned flush_start_tag
= set
->queue_depth
;
1648 node
= hctx
->numa_node
;
1649 if (node
== NUMA_NO_NODE
)
1650 node
= hctx
->numa_node
= set
->numa_node
;
1652 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1653 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1654 spin_lock_init(&hctx
->lock
);
1655 INIT_LIST_HEAD(&hctx
->dispatch
);
1657 hctx
->queue_num
= hctx_idx
;
1658 hctx
->flags
= set
->flags
;
1660 blk_mq_init_cpu_notifier(&hctx
->cpu_notifier
,
1661 blk_mq_hctx_notify
, hctx
);
1662 blk_mq_register_cpu_notifier(&hctx
->cpu_notifier
);
1664 hctx
->tags
= set
->tags
[hctx_idx
];
1667 * Allocate space for all possible cpus to avoid allocation at
1670 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1673 goto unregister_cpu_notifier
;
1675 if (blk_mq_alloc_bitmap(&hctx
->ctx_map
, node
))
1680 if (set
->ops
->init_hctx
&&
1681 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
1684 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
1688 if (set
->ops
->init_request
&&
1689 set
->ops
->init_request(set
->driver_data
,
1690 hctx
->fq
->flush_rq
, hctx_idx
,
1691 flush_start_tag
+ hctx_idx
, node
))
1699 if (set
->ops
->exit_hctx
)
1700 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1702 blk_mq_free_bitmap(&hctx
->ctx_map
);
1705 unregister_cpu_notifier
:
1706 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1711 static int blk_mq_init_hw_queues(struct request_queue
*q
,
1712 struct blk_mq_tag_set
*set
)
1714 struct blk_mq_hw_ctx
*hctx
;
1718 * Initialize hardware queues
1720 queue_for_each_hw_ctx(q
, hctx
, i
) {
1721 if (blk_mq_init_hctx(q
, set
, hctx
, i
))
1725 if (i
== q
->nr_hw_queues
)
1731 blk_mq_exit_hw_queues(q
, set
, i
);
1736 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1737 unsigned int nr_hw_queues
)
1741 for_each_possible_cpu(i
) {
1742 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1743 struct blk_mq_hw_ctx
*hctx
;
1745 memset(__ctx
, 0, sizeof(*__ctx
));
1747 spin_lock_init(&__ctx
->lock
);
1748 INIT_LIST_HEAD(&__ctx
->rq_list
);
1751 /* If the cpu isn't online, the cpu is mapped to first hctx */
1755 hctx
= q
->mq_ops
->map_queue(q
, i
);
1756 cpumask_set_cpu(i
, hctx
->cpumask
);
1760 * Set local node, IFF we have more than one hw queue. If
1761 * not, we remain on the home node of the device
1763 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1764 hctx
->numa_node
= cpu_to_node(i
);
1768 static void blk_mq_map_swqueue(struct request_queue
*q
)
1771 struct blk_mq_hw_ctx
*hctx
;
1772 struct blk_mq_ctx
*ctx
;
1774 queue_for_each_hw_ctx(q
, hctx
, i
) {
1775 cpumask_clear(hctx
->cpumask
);
1780 * Map software to hardware queues
1782 queue_for_each_ctx(q
, ctx
, i
) {
1783 /* If the cpu isn't online, the cpu is mapped to first hctx */
1787 hctx
= q
->mq_ops
->map_queue(q
, i
);
1788 cpumask_set_cpu(i
, hctx
->cpumask
);
1789 ctx
->index_hw
= hctx
->nr_ctx
;
1790 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
1793 queue_for_each_hw_ctx(q
, hctx
, i
) {
1795 * If no software queues are mapped to this hardware queue,
1796 * disable it and free the request entries.
1798 if (!hctx
->nr_ctx
) {
1799 struct blk_mq_tag_set
*set
= q
->tag_set
;
1802 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1803 set
->tags
[i
] = NULL
;
1810 * Initialize batch roundrobin counts
1812 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
1813 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1817 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
)
1819 struct blk_mq_hw_ctx
*hctx
;
1820 struct request_queue
*q
;
1824 if (set
->tag_list
.next
== set
->tag_list
.prev
)
1829 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
1830 blk_mq_freeze_queue(q
);
1832 queue_for_each_hw_ctx(q
, hctx
, i
) {
1834 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
1836 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
1838 blk_mq_unfreeze_queue(q
);
1842 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
1844 struct blk_mq_tag_set
*set
= q
->tag_set
;
1846 mutex_lock(&set
->tag_list_lock
);
1847 list_del_init(&q
->tag_set_list
);
1848 blk_mq_update_tag_set_depth(set
);
1849 mutex_unlock(&set
->tag_list_lock
);
1852 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
1853 struct request_queue
*q
)
1857 mutex_lock(&set
->tag_list_lock
);
1858 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
1859 blk_mq_update_tag_set_depth(set
);
1860 mutex_unlock(&set
->tag_list_lock
);
1863 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
1865 struct blk_mq_hw_ctx
**hctxs
;
1866 struct blk_mq_ctx __percpu
*ctx
;
1867 struct request_queue
*q
;
1871 ctx
= alloc_percpu(struct blk_mq_ctx
);
1873 return ERR_PTR(-ENOMEM
);
1875 hctxs
= kmalloc_node(set
->nr_hw_queues
* sizeof(*hctxs
), GFP_KERNEL
,
1881 map
= blk_mq_make_queue_map(set
);
1885 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1886 int node
= blk_mq_hw_queue_to_node(map
, i
);
1888 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
1893 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
1897 atomic_set(&hctxs
[i
]->nr_active
, 0);
1898 hctxs
[i
]->numa_node
= node
;
1899 hctxs
[i
]->queue_num
= i
;
1902 q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
1907 * Init percpu_ref in atomic mode so that it's faster to shutdown.
1908 * See blk_register_queue() for details.
1910 if (percpu_ref_init(&q
->mq_usage_counter
, blk_mq_usage_counter_release
,
1911 PERCPU_REF_INIT_ATOMIC
, GFP_KERNEL
))
1914 setup_timer(&q
->timeout
, blk_mq_rq_timer
, (unsigned long) q
);
1915 blk_queue_rq_timeout(q
, 30000);
1917 q
->nr_queues
= nr_cpu_ids
;
1918 q
->nr_hw_queues
= set
->nr_hw_queues
;
1922 q
->queue_hw_ctx
= hctxs
;
1924 q
->mq_ops
= set
->ops
;
1925 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
1927 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
1928 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
1930 q
->sg_reserved_size
= INT_MAX
;
1932 INIT_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
1933 INIT_LIST_HEAD(&q
->requeue_list
);
1934 spin_lock_init(&q
->requeue_lock
);
1936 if (q
->nr_hw_queues
> 1)
1937 blk_queue_make_request(q
, blk_mq_make_request
);
1939 blk_queue_make_request(q
, blk_sq_make_request
);
1942 blk_queue_rq_timeout(q
, set
->timeout
);
1945 * Do this after blk_queue_make_request() overrides it...
1947 q
->nr_requests
= set
->queue_depth
;
1949 if (set
->ops
->complete
)
1950 blk_queue_softirq_done(q
, set
->ops
->complete
);
1952 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
1954 if (blk_mq_init_hw_queues(q
, set
))
1957 mutex_lock(&all_q_mutex
);
1958 list_add_tail(&q
->all_q_node
, &all_q_list
);
1959 mutex_unlock(&all_q_mutex
);
1961 blk_mq_add_queue_tag_set(set
, q
);
1963 blk_mq_map_swqueue(q
);
1968 blk_cleanup_queue(q
);
1971 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1974 free_cpumask_var(hctxs
[i
]->cpumask
);
1981 return ERR_PTR(-ENOMEM
);
1983 EXPORT_SYMBOL(blk_mq_init_queue
);
1985 void blk_mq_free_queue(struct request_queue
*q
)
1987 struct blk_mq_tag_set
*set
= q
->tag_set
;
1989 blk_mq_del_queue_tag_set(q
);
1991 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
1992 blk_mq_free_hw_queues(q
, set
);
1994 percpu_ref_exit(&q
->mq_usage_counter
);
1996 free_percpu(q
->queue_ctx
);
1997 kfree(q
->queue_hw_ctx
);
2000 q
->queue_ctx
= NULL
;
2001 q
->queue_hw_ctx
= NULL
;
2004 mutex_lock(&all_q_mutex
);
2005 list_del_init(&q
->all_q_node
);
2006 mutex_unlock(&all_q_mutex
);
2009 /* Basically redo blk_mq_init_queue with queue frozen */
2010 static void blk_mq_queue_reinit(struct request_queue
*q
)
2012 WARN_ON_ONCE(!q
->mq_freeze_depth
);
2014 blk_mq_sysfs_unregister(q
);
2016 blk_mq_update_queue_map(q
->mq_map
, q
->nr_hw_queues
);
2019 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2020 * we should change hctx numa_node according to new topology (this
2021 * involves free and re-allocate memory, worthy doing?)
2024 blk_mq_map_swqueue(q
);
2026 blk_mq_sysfs_register(q
);
2029 static int blk_mq_queue_reinit_notify(struct notifier_block
*nb
,
2030 unsigned long action
, void *hcpu
)
2032 struct request_queue
*q
;
2035 * Before new mappings are established, hotadded cpu might already
2036 * start handling requests. This doesn't break anything as we map
2037 * offline CPUs to first hardware queue. We will re-init the queue
2038 * below to get optimal settings.
2040 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
&&
2041 action
!= CPU_ONLINE
&& action
!= CPU_ONLINE_FROZEN
)
2044 mutex_lock(&all_q_mutex
);
2047 * We need to freeze and reinit all existing queues. Freezing
2048 * involves synchronous wait for an RCU grace period and doing it
2049 * one by one may take a long time. Start freezing all queues in
2050 * one swoop and then wait for the completions so that freezing can
2051 * take place in parallel.
2053 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2054 blk_mq_freeze_queue_start(q
);
2055 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2056 blk_mq_freeze_queue_wait(q
);
2058 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2059 blk_mq_queue_reinit(q
);
2061 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2062 blk_mq_unfreeze_queue(q
);
2064 mutex_unlock(&all_q_mutex
);
2068 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2072 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2073 set
->tags
[i
] = blk_mq_init_rq_map(set
, i
);
2082 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2088 * Allocate the request maps associated with this tag_set. Note that this
2089 * may reduce the depth asked for, if memory is tight. set->queue_depth
2090 * will be updated to reflect the allocated depth.
2092 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2097 depth
= set
->queue_depth
;
2099 err
= __blk_mq_alloc_rq_maps(set
);
2103 set
->queue_depth
>>= 1;
2104 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2108 } while (set
->queue_depth
);
2110 if (!set
->queue_depth
|| err
) {
2111 pr_err("blk-mq: failed to allocate request map\n");
2115 if (depth
!= set
->queue_depth
)
2116 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2117 depth
, set
->queue_depth
);
2123 * Alloc a tag set to be associated with one or more request queues.
2124 * May fail with EINVAL for various error conditions. May adjust the
2125 * requested depth down, if if it too large. In that case, the set
2126 * value will be stored in set->queue_depth.
2128 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2130 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2132 if (!set
->nr_hw_queues
)
2134 if (!set
->queue_depth
)
2136 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2139 if (!set
->nr_hw_queues
|| !set
->ops
->queue_rq
|| !set
->ops
->map_queue
)
2142 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2143 pr_info("blk-mq: reduced tag depth to %u\n",
2145 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2149 * If a crashdump is active, then we are potentially in a very
2150 * memory constrained environment. Limit us to 1 queue and
2151 * 64 tags to prevent using too much memory.
2153 if (is_kdump_kernel()) {
2154 set
->nr_hw_queues
= 1;
2155 set
->queue_depth
= min(64U, set
->queue_depth
);
2158 set
->tags
= kmalloc_node(set
->nr_hw_queues
*
2159 sizeof(struct blk_mq_tags
*),
2160 GFP_KERNEL
, set
->numa_node
);
2164 if (blk_mq_alloc_rq_maps(set
))
2167 mutex_init(&set
->tag_list_lock
);
2168 INIT_LIST_HEAD(&set
->tag_list
);
2176 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2178 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2182 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2184 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2190 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2192 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2194 struct blk_mq_tag_set
*set
= q
->tag_set
;
2195 struct blk_mq_hw_ctx
*hctx
;
2198 if (!set
|| nr
> set
->queue_depth
)
2202 queue_for_each_hw_ctx(q
, hctx
, i
) {
2203 ret
= blk_mq_tag_update_depth(hctx
->tags
, nr
);
2209 q
->nr_requests
= nr
;
2214 void blk_mq_disable_hotplug(void)
2216 mutex_lock(&all_q_mutex
);
2219 void blk_mq_enable_hotplug(void)
2221 mutex_unlock(&all_q_mutex
);
2224 static int __init
blk_mq_init(void)
2228 hotcpu_notifier(blk_mq_queue_reinit_notify
, 0);
2232 subsys_initcall(blk_mq_init
);