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
))
631 if (time_after_eq(jiffies
, rq
->deadline
)) {
632 if (!blk_mark_rq_complete(rq
))
633 blk_mq_rq_timed_out(rq
, reserved
);
634 } else if (!data
->next_set
|| time_after(data
->next
, rq
->deadline
)) {
635 data
->next
= rq
->deadline
;
640 static void blk_mq_rq_timer(unsigned long priv
)
642 struct request_queue
*q
= (struct request_queue
*)priv
;
643 struct blk_mq_timeout_data data
= {
647 struct blk_mq_hw_ctx
*hctx
;
650 queue_for_each_hw_ctx(q
, hctx
, i
) {
652 * If not software queues are currently mapped to this
653 * hardware queue, there's nothing to check
655 if (!blk_mq_hw_queue_mapped(hctx
))
658 blk_mq_tag_busy_iter(hctx
, blk_mq_check_expired
, &data
);
662 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
663 mod_timer(&q
->timeout
, data
.next
);
665 queue_for_each_hw_ctx(q
, hctx
, i
)
666 blk_mq_tag_idle(hctx
);
671 * Reverse check our software queue for entries that we could potentially
672 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
673 * too much time checking for merges.
675 static bool blk_mq_attempt_merge(struct request_queue
*q
,
676 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
681 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
687 if (!blk_rq_merge_ok(rq
, bio
))
690 el_ret
= blk_try_merge(rq
, bio
);
691 if (el_ret
== ELEVATOR_BACK_MERGE
) {
692 if (bio_attempt_back_merge(q
, rq
, bio
)) {
697 } else if (el_ret
== ELEVATOR_FRONT_MERGE
) {
698 if (bio_attempt_front_merge(q
, rq
, bio
)) {
710 * Process software queues that have been marked busy, splicing them
711 * to the for-dispatch
713 static void flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
715 struct blk_mq_ctx
*ctx
;
718 for (i
= 0; i
< hctx
->ctx_map
.map_size
; i
++) {
719 struct blk_align_bitmap
*bm
= &hctx
->ctx_map
.map
[i
];
720 unsigned int off
, bit
;
726 off
= i
* hctx
->ctx_map
.bits_per_word
;
728 bit
= find_next_bit(&bm
->word
, bm
->depth
, bit
);
729 if (bit
>= bm
->depth
)
732 ctx
= hctx
->ctxs
[bit
+ off
];
733 clear_bit(bit
, &bm
->word
);
734 spin_lock(&ctx
->lock
);
735 list_splice_tail_init(&ctx
->rq_list
, list
);
736 spin_unlock(&ctx
->lock
);
744 * Run this hardware queue, pulling any software queues mapped to it in.
745 * Note that this function currently has various problems around ordering
746 * of IO. In particular, we'd like FIFO behaviour on handling existing
747 * items on the hctx->dispatch list. Ignore that for now.
749 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
751 struct request_queue
*q
= hctx
->queue
;
754 LIST_HEAD(driver_list
);
755 struct list_head
*dptr
;
758 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
));
760 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)))
766 * Touch any software queue that has pending entries.
768 flush_busy_ctxs(hctx
, &rq_list
);
771 * If we have previous entries on our dispatch list, grab them
772 * and stuff them at the front for more fair dispatch.
774 if (!list_empty_careful(&hctx
->dispatch
)) {
775 spin_lock(&hctx
->lock
);
776 if (!list_empty(&hctx
->dispatch
))
777 list_splice_init(&hctx
->dispatch
, &rq_list
);
778 spin_unlock(&hctx
->lock
);
782 * Start off with dptr being NULL, so we start the first request
783 * immediately, even if we have more pending.
788 * Now process all the entries, sending them to the driver.
791 while (!list_empty(&rq_list
)) {
792 struct blk_mq_queue_data bd
;
795 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
796 list_del_init(&rq
->queuelist
);
800 bd
.last
= list_empty(&rq_list
);
802 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
804 case BLK_MQ_RQ_QUEUE_OK
:
807 case BLK_MQ_RQ_QUEUE_BUSY
:
808 list_add(&rq
->queuelist
, &rq_list
);
809 __blk_mq_requeue_request(rq
);
812 pr_err("blk-mq: bad return on queue: %d\n", ret
);
813 case BLK_MQ_RQ_QUEUE_ERROR
:
815 blk_mq_end_request(rq
, rq
->errors
);
819 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
823 * We've done the first request. If we have more than 1
824 * left in the list, set dptr to defer issue.
826 if (!dptr
&& rq_list
.next
!= rq_list
.prev
)
831 hctx
->dispatched
[0]++;
832 else if (queued
< (1 << (BLK_MQ_MAX_DISPATCH_ORDER
- 1)))
833 hctx
->dispatched
[ilog2(queued
) + 1]++;
836 * Any items that need requeuing? Stuff them into hctx->dispatch,
837 * that is where we will continue on next queue run.
839 if (!list_empty(&rq_list
)) {
840 spin_lock(&hctx
->lock
);
841 list_splice(&rq_list
, &hctx
->dispatch
);
842 spin_unlock(&hctx
->lock
);
847 * It'd be great if the workqueue API had a way to pass
848 * in a mask and had some smarts for more clever placement.
849 * For now we just round-robin here, switching for every
850 * BLK_MQ_CPU_WORK_BATCH queued items.
852 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
854 if (hctx
->queue
->nr_hw_queues
== 1)
855 return WORK_CPU_UNBOUND
;
857 if (--hctx
->next_cpu_batch
<= 0) {
858 int cpu
= hctx
->next_cpu
, next_cpu
;
860 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
861 if (next_cpu
>= nr_cpu_ids
)
862 next_cpu
= cpumask_first(hctx
->cpumask
);
864 hctx
->next_cpu
= next_cpu
;
865 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
870 return hctx
->next_cpu
;
873 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
875 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
) ||
876 !blk_mq_hw_queue_mapped(hctx
)))
881 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
882 __blk_mq_run_hw_queue(hctx
);
890 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
894 void blk_mq_run_queues(struct request_queue
*q
, bool async
)
896 struct blk_mq_hw_ctx
*hctx
;
899 queue_for_each_hw_ctx(q
, hctx
, i
) {
900 if ((!blk_mq_hctx_has_pending(hctx
) &&
901 list_empty_careful(&hctx
->dispatch
)) ||
902 test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
905 blk_mq_run_hw_queue(hctx
, async
);
908 EXPORT_SYMBOL(blk_mq_run_queues
);
910 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
912 cancel_delayed_work(&hctx
->run_work
);
913 cancel_delayed_work(&hctx
->delay_work
);
914 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
916 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
918 void blk_mq_stop_hw_queues(struct request_queue
*q
)
920 struct blk_mq_hw_ctx
*hctx
;
923 queue_for_each_hw_ctx(q
, hctx
, i
)
924 blk_mq_stop_hw_queue(hctx
);
926 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
928 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
930 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
932 blk_mq_run_hw_queue(hctx
, false);
934 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
936 void blk_mq_start_hw_queues(struct request_queue
*q
)
938 struct blk_mq_hw_ctx
*hctx
;
941 queue_for_each_hw_ctx(q
, hctx
, i
)
942 blk_mq_start_hw_queue(hctx
);
944 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
947 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
949 struct blk_mq_hw_ctx
*hctx
;
952 queue_for_each_hw_ctx(q
, hctx
, i
) {
953 if (!test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
956 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
957 blk_mq_run_hw_queue(hctx
, async
);
960 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
962 static void blk_mq_run_work_fn(struct work_struct
*work
)
964 struct blk_mq_hw_ctx
*hctx
;
966 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
968 __blk_mq_run_hw_queue(hctx
);
971 static void blk_mq_delay_work_fn(struct work_struct
*work
)
973 struct blk_mq_hw_ctx
*hctx
;
975 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
977 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
978 __blk_mq_run_hw_queue(hctx
);
981 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
983 if (unlikely(!blk_mq_hw_queue_mapped(hctx
)))
986 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
987 &hctx
->delay_work
, msecs_to_jiffies(msecs
));
989 EXPORT_SYMBOL(blk_mq_delay_queue
);
991 static void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
,
992 struct request
*rq
, bool at_head
)
994 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
996 trace_block_rq_insert(hctx
->queue
, rq
);
999 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1001 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1003 blk_mq_hctx_mark_pending(hctx
, ctx
);
1006 void blk_mq_insert_request(struct request
*rq
, bool at_head
, bool run_queue
,
1009 struct request_queue
*q
= rq
->q
;
1010 struct blk_mq_hw_ctx
*hctx
;
1011 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
, *current_ctx
;
1013 current_ctx
= blk_mq_get_ctx(q
);
1014 if (!cpu_online(ctx
->cpu
))
1015 rq
->mq_ctx
= ctx
= current_ctx
;
1017 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1019 spin_lock(&ctx
->lock
);
1020 __blk_mq_insert_request(hctx
, rq
, at_head
);
1021 spin_unlock(&ctx
->lock
);
1024 blk_mq_run_hw_queue(hctx
, async
);
1026 blk_mq_put_ctx(current_ctx
);
1029 static void blk_mq_insert_requests(struct request_queue
*q
,
1030 struct blk_mq_ctx
*ctx
,
1031 struct list_head
*list
,
1036 struct blk_mq_hw_ctx
*hctx
;
1037 struct blk_mq_ctx
*current_ctx
;
1039 trace_block_unplug(q
, depth
, !from_schedule
);
1041 current_ctx
= blk_mq_get_ctx(q
);
1043 if (!cpu_online(ctx
->cpu
))
1045 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1048 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1051 spin_lock(&ctx
->lock
);
1052 while (!list_empty(list
)) {
1055 rq
= list_first_entry(list
, struct request
, queuelist
);
1056 list_del_init(&rq
->queuelist
);
1058 __blk_mq_insert_request(hctx
, rq
, false);
1060 spin_unlock(&ctx
->lock
);
1062 blk_mq_run_hw_queue(hctx
, from_schedule
);
1063 blk_mq_put_ctx(current_ctx
);
1066 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1068 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1069 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1071 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1072 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1073 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1076 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1078 struct blk_mq_ctx
*this_ctx
;
1079 struct request_queue
*this_q
;
1082 LIST_HEAD(ctx_list
);
1085 list_splice_init(&plug
->mq_list
, &list
);
1087 list_sort(NULL
, &list
, plug_ctx_cmp
);
1093 while (!list_empty(&list
)) {
1094 rq
= list_entry_rq(list
.next
);
1095 list_del_init(&rq
->queuelist
);
1097 if (rq
->mq_ctx
!= this_ctx
) {
1099 blk_mq_insert_requests(this_q
, this_ctx
,
1104 this_ctx
= rq
->mq_ctx
;
1110 list_add_tail(&rq
->queuelist
, &ctx_list
);
1114 * If 'this_ctx' is set, we know we have entries to complete
1115 * on 'ctx_list'. Do those.
1118 blk_mq_insert_requests(this_q
, this_ctx
, &ctx_list
, depth
,
1123 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1125 init_request_from_bio(rq
, bio
);
1127 if (blk_do_io_stat(rq
))
1128 blk_account_io_start(rq
, 1);
1131 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1133 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1134 !blk_queue_nomerges(hctx
->queue
);
1137 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx
*hctx
,
1138 struct blk_mq_ctx
*ctx
,
1139 struct request
*rq
, struct bio
*bio
)
1141 if (!hctx_allow_merges(hctx
)) {
1142 blk_mq_bio_to_request(rq
, bio
);
1143 spin_lock(&ctx
->lock
);
1145 __blk_mq_insert_request(hctx
, rq
, false);
1146 spin_unlock(&ctx
->lock
);
1149 struct request_queue
*q
= hctx
->queue
;
1151 spin_lock(&ctx
->lock
);
1152 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1153 blk_mq_bio_to_request(rq
, bio
);
1157 spin_unlock(&ctx
->lock
);
1158 __blk_mq_free_request(hctx
, ctx
, rq
);
1163 struct blk_map_ctx
{
1164 struct blk_mq_hw_ctx
*hctx
;
1165 struct blk_mq_ctx
*ctx
;
1168 static struct request
*blk_mq_map_request(struct request_queue
*q
,
1170 struct blk_map_ctx
*data
)
1172 struct blk_mq_hw_ctx
*hctx
;
1173 struct blk_mq_ctx
*ctx
;
1175 int rw
= bio_data_dir(bio
);
1176 struct blk_mq_alloc_data alloc_data
;
1178 if (unlikely(blk_mq_queue_enter(q
))) {
1179 bio_endio(bio
, -EIO
);
1183 ctx
= blk_mq_get_ctx(q
);
1184 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1186 if (rw_is_sync(bio
->bi_rw
))
1189 trace_block_getrq(q
, bio
, rw
);
1190 blk_mq_set_alloc_data(&alloc_data
, q
, GFP_ATOMIC
, false, ctx
,
1192 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
1193 if (unlikely(!rq
)) {
1194 __blk_mq_run_hw_queue(hctx
);
1195 blk_mq_put_ctx(ctx
);
1196 trace_block_sleeprq(q
, bio
, rw
);
1198 ctx
= blk_mq_get_ctx(q
);
1199 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1200 blk_mq_set_alloc_data(&alloc_data
, q
,
1201 __GFP_WAIT
|GFP_ATOMIC
, false, ctx
, hctx
);
1202 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
1203 ctx
= alloc_data
.ctx
;
1204 hctx
= alloc_data
.hctx
;
1214 * Multiple hardware queue variant. This will not use per-process plugs,
1215 * but will attempt to bypass the hctx queueing if we can go straight to
1216 * hardware for SYNC IO.
1218 static void blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1220 const int is_sync
= rw_is_sync(bio
->bi_rw
);
1221 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
1222 struct blk_map_ctx data
;
1225 blk_queue_bounce(q
, &bio
);
1227 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1228 bio_endio(bio
, -EIO
);
1232 rq
= blk_mq_map_request(q
, bio
, &data
);
1236 if (unlikely(is_flush_fua
)) {
1237 blk_mq_bio_to_request(rq
, bio
);
1238 blk_insert_flush(rq
);
1243 * If the driver supports defer issued based on 'last', then
1244 * queue it up like normal since we can potentially save some
1247 if (is_sync
&& !(data
.hctx
->flags
& BLK_MQ_F_DEFER_ISSUE
)) {
1248 struct blk_mq_queue_data bd
= {
1255 blk_mq_bio_to_request(rq
, bio
);
1258 * For OK queue, we are done. For error, kill it. Any other
1259 * error (busy), just add it to our list as we previously
1262 ret
= q
->mq_ops
->queue_rq(data
.hctx
, &bd
);
1263 if (ret
== BLK_MQ_RQ_QUEUE_OK
)
1266 __blk_mq_requeue_request(rq
);
1268 if (ret
== BLK_MQ_RQ_QUEUE_ERROR
) {
1270 blk_mq_end_request(rq
, rq
->errors
);
1276 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1278 * For a SYNC request, send it to the hardware immediately. For
1279 * an ASYNC request, just ensure that we run it later on. The
1280 * latter allows for merging opportunities and more efficient
1284 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1287 blk_mq_put_ctx(data
.ctx
);
1291 * Single hardware queue variant. This will attempt to use any per-process
1292 * plug for merging and IO deferral.
1294 static void blk_sq_make_request(struct request_queue
*q
, struct bio
*bio
)
1296 const int is_sync
= rw_is_sync(bio
->bi_rw
);
1297 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
1298 unsigned int use_plug
, request_count
= 0;
1299 struct blk_map_ctx data
;
1303 * If we have multiple hardware queues, just go directly to
1304 * one of those for sync IO.
1306 use_plug
= !is_flush_fua
&& !is_sync
;
1308 blk_queue_bounce(q
, &bio
);
1310 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1311 bio_endio(bio
, -EIO
);
1315 if (use_plug
&& !blk_queue_nomerges(q
) &&
1316 blk_attempt_plug_merge(q
, bio
, &request_count
))
1319 rq
= blk_mq_map_request(q
, bio
, &data
);
1323 if (unlikely(is_flush_fua
)) {
1324 blk_mq_bio_to_request(rq
, bio
);
1325 blk_insert_flush(rq
);
1330 * A task plug currently exists. Since this is completely lockless,
1331 * utilize that to temporarily store requests until the task is
1332 * either done or scheduled away.
1335 struct blk_plug
*plug
= current
->plug
;
1338 blk_mq_bio_to_request(rq
, bio
);
1339 if (list_empty(&plug
->mq_list
))
1340 trace_block_plug(q
);
1341 else if (request_count
>= BLK_MAX_REQUEST_COUNT
) {
1342 blk_flush_plug_list(plug
, false);
1343 trace_block_plug(q
);
1345 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1346 blk_mq_put_ctx(data
.ctx
);
1351 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1353 * For a SYNC request, send it to the hardware immediately. For
1354 * an ASYNC request, just ensure that we run it later on. The
1355 * latter allows for merging opportunities and more efficient
1359 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1362 blk_mq_put_ctx(data
.ctx
);
1366 * Default mapping to a software queue, since we use one per CPU.
1368 struct blk_mq_hw_ctx
*blk_mq_map_queue(struct request_queue
*q
, const int cpu
)
1370 return q
->queue_hw_ctx
[q
->mq_map
[cpu
]];
1372 EXPORT_SYMBOL(blk_mq_map_queue
);
1374 static void blk_mq_free_rq_map(struct blk_mq_tag_set
*set
,
1375 struct blk_mq_tags
*tags
, unsigned int hctx_idx
)
1379 if (tags
->rqs
&& set
->ops
->exit_request
) {
1382 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1385 set
->ops
->exit_request(set
->driver_data
, tags
->rqs
[i
],
1387 tags
->rqs
[i
] = NULL
;
1391 while (!list_empty(&tags
->page_list
)) {
1392 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1393 list_del_init(&page
->lru
);
1394 __free_pages(page
, page
->private);
1399 blk_mq_free_tags(tags
);
1402 static size_t order_to_size(unsigned int order
)
1404 return (size_t)PAGE_SIZE
<< order
;
1407 static struct blk_mq_tags
*blk_mq_init_rq_map(struct blk_mq_tag_set
*set
,
1408 unsigned int hctx_idx
)
1410 struct blk_mq_tags
*tags
;
1411 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1412 size_t rq_size
, left
;
1414 tags
= blk_mq_init_tags(set
->queue_depth
, set
->reserved_tags
,
1419 INIT_LIST_HEAD(&tags
->page_list
);
1421 tags
->rqs
= kzalloc_node(set
->queue_depth
* sizeof(struct request
*),
1422 GFP_KERNEL
| __GFP_NOWARN
| __GFP_NORETRY
,
1425 blk_mq_free_tags(tags
);
1430 * rq_size is the size of the request plus driver payload, rounded
1431 * to the cacheline size
1433 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1435 left
= rq_size
* set
->queue_depth
;
1437 for (i
= 0; i
< set
->queue_depth
; ) {
1438 int this_order
= max_order
;
1443 while (left
< order_to_size(this_order
- 1) && this_order
)
1447 page
= alloc_pages_node(set
->numa_node
,
1448 GFP_KERNEL
| __GFP_NOWARN
| __GFP_NORETRY
,
1454 if (order_to_size(this_order
) < rq_size
)
1461 page
->private = this_order
;
1462 list_add_tail(&page
->lru
, &tags
->page_list
);
1464 p
= page_address(page
);
1465 entries_per_page
= order_to_size(this_order
) / rq_size
;
1466 to_do
= min(entries_per_page
, set
->queue_depth
- i
);
1467 left
-= to_do
* rq_size
;
1468 for (j
= 0; j
< to_do
; j
++) {
1470 tags
->rqs
[i
]->atomic_flags
= 0;
1471 tags
->rqs
[i
]->cmd_flags
= 0;
1472 if (set
->ops
->init_request
) {
1473 if (set
->ops
->init_request(set
->driver_data
,
1474 tags
->rqs
[i
], hctx_idx
, i
,
1476 tags
->rqs
[i
] = NULL
;
1489 blk_mq_free_rq_map(set
, tags
, hctx_idx
);
1493 static void blk_mq_free_bitmap(struct blk_mq_ctxmap
*bitmap
)
1498 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap
*bitmap
, int node
)
1500 unsigned int bpw
= 8, total
, num_maps
, i
;
1502 bitmap
->bits_per_word
= bpw
;
1504 num_maps
= ALIGN(nr_cpu_ids
, bpw
) / bpw
;
1505 bitmap
->map
= kzalloc_node(num_maps
* sizeof(struct blk_align_bitmap
),
1510 bitmap
->map_size
= num_maps
;
1513 for (i
= 0; i
< num_maps
; i
++) {
1514 bitmap
->map
[i
].depth
= min(total
, bitmap
->bits_per_word
);
1515 total
-= bitmap
->map
[i
].depth
;
1521 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx
*hctx
, int cpu
)
1523 struct request_queue
*q
= hctx
->queue
;
1524 struct blk_mq_ctx
*ctx
;
1528 * Move ctx entries to new CPU, if this one is going away.
1530 ctx
= __blk_mq_get_ctx(q
, cpu
);
1532 spin_lock(&ctx
->lock
);
1533 if (!list_empty(&ctx
->rq_list
)) {
1534 list_splice_init(&ctx
->rq_list
, &tmp
);
1535 blk_mq_hctx_clear_pending(hctx
, ctx
);
1537 spin_unlock(&ctx
->lock
);
1539 if (list_empty(&tmp
))
1542 ctx
= blk_mq_get_ctx(q
);
1543 spin_lock(&ctx
->lock
);
1545 while (!list_empty(&tmp
)) {
1548 rq
= list_first_entry(&tmp
, struct request
, queuelist
);
1550 list_move_tail(&rq
->queuelist
, &ctx
->rq_list
);
1553 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1554 blk_mq_hctx_mark_pending(hctx
, ctx
);
1556 spin_unlock(&ctx
->lock
);
1558 blk_mq_run_hw_queue(hctx
, true);
1559 blk_mq_put_ctx(ctx
);
1563 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx
*hctx
, int cpu
)
1565 struct request_queue
*q
= hctx
->queue
;
1566 struct blk_mq_tag_set
*set
= q
->tag_set
;
1568 if (set
->tags
[hctx
->queue_num
])
1571 set
->tags
[hctx
->queue_num
] = blk_mq_init_rq_map(set
, hctx
->queue_num
);
1572 if (!set
->tags
[hctx
->queue_num
])
1575 hctx
->tags
= set
->tags
[hctx
->queue_num
];
1579 static int blk_mq_hctx_notify(void *data
, unsigned long action
,
1582 struct blk_mq_hw_ctx
*hctx
= data
;
1584 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
)
1585 return blk_mq_hctx_cpu_offline(hctx
, cpu
);
1586 else if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
)
1587 return blk_mq_hctx_cpu_online(hctx
, cpu
);
1592 static void blk_mq_exit_hctx(struct request_queue
*q
,
1593 struct blk_mq_tag_set
*set
,
1594 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
1596 unsigned flush_start_tag
= set
->queue_depth
;
1598 blk_mq_tag_idle(hctx
);
1600 if (set
->ops
->exit_request
)
1601 set
->ops
->exit_request(set
->driver_data
,
1602 hctx
->fq
->flush_rq
, hctx_idx
,
1603 flush_start_tag
+ hctx_idx
);
1605 if (set
->ops
->exit_hctx
)
1606 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1608 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1609 blk_free_flush_queue(hctx
->fq
);
1611 blk_mq_free_bitmap(&hctx
->ctx_map
);
1614 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1615 struct blk_mq_tag_set
*set
, int nr_queue
)
1617 struct blk_mq_hw_ctx
*hctx
;
1620 queue_for_each_hw_ctx(q
, hctx
, i
) {
1623 blk_mq_exit_hctx(q
, set
, hctx
, i
);
1627 static void blk_mq_free_hw_queues(struct request_queue
*q
,
1628 struct blk_mq_tag_set
*set
)
1630 struct blk_mq_hw_ctx
*hctx
;
1633 queue_for_each_hw_ctx(q
, hctx
, i
) {
1634 free_cpumask_var(hctx
->cpumask
);
1639 static int blk_mq_init_hctx(struct request_queue
*q
,
1640 struct blk_mq_tag_set
*set
,
1641 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
1644 unsigned flush_start_tag
= set
->queue_depth
;
1646 node
= hctx
->numa_node
;
1647 if (node
== NUMA_NO_NODE
)
1648 node
= hctx
->numa_node
= set
->numa_node
;
1650 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1651 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1652 spin_lock_init(&hctx
->lock
);
1653 INIT_LIST_HEAD(&hctx
->dispatch
);
1655 hctx
->queue_num
= hctx_idx
;
1656 hctx
->flags
= set
->flags
;
1658 blk_mq_init_cpu_notifier(&hctx
->cpu_notifier
,
1659 blk_mq_hctx_notify
, hctx
);
1660 blk_mq_register_cpu_notifier(&hctx
->cpu_notifier
);
1662 hctx
->tags
= set
->tags
[hctx_idx
];
1665 * Allocate space for all possible cpus to avoid allocation at
1668 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1671 goto unregister_cpu_notifier
;
1673 if (blk_mq_alloc_bitmap(&hctx
->ctx_map
, node
))
1678 if (set
->ops
->init_hctx
&&
1679 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
1682 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
1686 if (set
->ops
->init_request
&&
1687 set
->ops
->init_request(set
->driver_data
,
1688 hctx
->fq
->flush_rq
, hctx_idx
,
1689 flush_start_tag
+ hctx_idx
, node
))
1697 if (set
->ops
->exit_hctx
)
1698 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1700 blk_mq_free_bitmap(&hctx
->ctx_map
);
1703 unregister_cpu_notifier
:
1704 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1709 static int blk_mq_init_hw_queues(struct request_queue
*q
,
1710 struct blk_mq_tag_set
*set
)
1712 struct blk_mq_hw_ctx
*hctx
;
1716 * Initialize hardware queues
1718 queue_for_each_hw_ctx(q
, hctx
, i
) {
1719 if (blk_mq_init_hctx(q
, set
, hctx
, i
))
1723 if (i
== q
->nr_hw_queues
)
1729 blk_mq_exit_hw_queues(q
, set
, i
);
1734 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1735 unsigned int nr_hw_queues
)
1739 for_each_possible_cpu(i
) {
1740 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1741 struct blk_mq_hw_ctx
*hctx
;
1743 memset(__ctx
, 0, sizeof(*__ctx
));
1745 spin_lock_init(&__ctx
->lock
);
1746 INIT_LIST_HEAD(&__ctx
->rq_list
);
1749 /* If the cpu isn't online, the cpu is mapped to first hctx */
1753 hctx
= q
->mq_ops
->map_queue(q
, i
);
1754 cpumask_set_cpu(i
, hctx
->cpumask
);
1758 * Set local node, IFF we have more than one hw queue. If
1759 * not, we remain on the home node of the device
1761 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1762 hctx
->numa_node
= cpu_to_node(i
);
1766 static void blk_mq_map_swqueue(struct request_queue
*q
)
1769 struct blk_mq_hw_ctx
*hctx
;
1770 struct blk_mq_ctx
*ctx
;
1772 queue_for_each_hw_ctx(q
, hctx
, i
) {
1773 cpumask_clear(hctx
->cpumask
);
1778 * Map software to hardware queues
1780 queue_for_each_ctx(q
, ctx
, i
) {
1781 /* If the cpu isn't online, the cpu is mapped to first hctx */
1785 hctx
= q
->mq_ops
->map_queue(q
, i
);
1786 cpumask_set_cpu(i
, hctx
->cpumask
);
1787 ctx
->index_hw
= hctx
->nr_ctx
;
1788 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
1791 queue_for_each_hw_ctx(q
, hctx
, i
) {
1793 * If no software queues are mapped to this hardware queue,
1794 * disable it and free the request entries.
1796 if (!hctx
->nr_ctx
) {
1797 struct blk_mq_tag_set
*set
= q
->tag_set
;
1800 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1801 set
->tags
[i
] = NULL
;
1808 * Initialize batch roundrobin counts
1810 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
1811 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1815 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
)
1817 struct blk_mq_hw_ctx
*hctx
;
1818 struct request_queue
*q
;
1822 if (set
->tag_list
.next
== set
->tag_list
.prev
)
1827 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
1828 blk_mq_freeze_queue(q
);
1830 queue_for_each_hw_ctx(q
, hctx
, i
) {
1832 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
1834 hctx
->flags
&= ~BLK_MQ_F_TAG_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 blk_mq_update_tag_set_depth(set
);
1847 mutex_unlock(&set
->tag_list_lock
);
1850 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
1851 struct request_queue
*q
)
1855 mutex_lock(&set
->tag_list_lock
);
1856 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
1857 blk_mq_update_tag_set_depth(set
);
1858 mutex_unlock(&set
->tag_list_lock
);
1861 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
1863 struct blk_mq_hw_ctx
**hctxs
;
1864 struct blk_mq_ctx __percpu
*ctx
;
1865 struct request_queue
*q
;
1869 ctx
= alloc_percpu(struct blk_mq_ctx
);
1871 return ERR_PTR(-ENOMEM
);
1873 hctxs
= kmalloc_node(set
->nr_hw_queues
* sizeof(*hctxs
), GFP_KERNEL
,
1879 map
= blk_mq_make_queue_map(set
);
1883 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1884 int node
= blk_mq_hw_queue_to_node(map
, i
);
1886 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
1891 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
1895 atomic_set(&hctxs
[i
]->nr_active
, 0);
1896 hctxs
[i
]->numa_node
= node
;
1897 hctxs
[i
]->queue_num
= i
;
1900 q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
1905 * Init percpu_ref in atomic mode so that it's faster to shutdown.
1906 * See blk_register_queue() for details.
1908 if (percpu_ref_init(&q
->mq_usage_counter
, blk_mq_usage_counter_release
,
1909 PERCPU_REF_INIT_ATOMIC
, GFP_KERNEL
))
1912 setup_timer(&q
->timeout
, blk_mq_rq_timer
, (unsigned long) q
);
1913 blk_queue_rq_timeout(q
, 30000);
1915 q
->nr_queues
= nr_cpu_ids
;
1916 q
->nr_hw_queues
= set
->nr_hw_queues
;
1920 q
->queue_hw_ctx
= hctxs
;
1922 q
->mq_ops
= set
->ops
;
1923 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
1925 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
1926 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
1928 q
->sg_reserved_size
= INT_MAX
;
1930 INIT_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
1931 INIT_LIST_HEAD(&q
->requeue_list
);
1932 spin_lock_init(&q
->requeue_lock
);
1934 if (q
->nr_hw_queues
> 1)
1935 blk_queue_make_request(q
, blk_mq_make_request
);
1937 blk_queue_make_request(q
, blk_sq_make_request
);
1940 blk_queue_rq_timeout(q
, set
->timeout
);
1943 * Do this after blk_queue_make_request() overrides it...
1945 q
->nr_requests
= set
->queue_depth
;
1947 if (set
->ops
->complete
)
1948 blk_queue_softirq_done(q
, set
->ops
->complete
);
1950 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
1952 if (blk_mq_init_hw_queues(q
, set
))
1955 mutex_lock(&all_q_mutex
);
1956 list_add_tail(&q
->all_q_node
, &all_q_list
);
1957 mutex_unlock(&all_q_mutex
);
1959 blk_mq_add_queue_tag_set(set
, q
);
1961 blk_mq_map_swqueue(q
);
1966 blk_cleanup_queue(q
);
1969 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1972 free_cpumask_var(hctxs
[i
]->cpumask
);
1979 return ERR_PTR(-ENOMEM
);
1981 EXPORT_SYMBOL(blk_mq_init_queue
);
1983 void blk_mq_free_queue(struct request_queue
*q
)
1985 struct blk_mq_tag_set
*set
= q
->tag_set
;
1987 blk_mq_del_queue_tag_set(q
);
1989 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
1990 blk_mq_free_hw_queues(q
, set
);
1992 percpu_ref_exit(&q
->mq_usage_counter
);
1994 free_percpu(q
->queue_ctx
);
1995 kfree(q
->queue_hw_ctx
);
1998 q
->queue_ctx
= NULL
;
1999 q
->queue_hw_ctx
= NULL
;
2002 mutex_lock(&all_q_mutex
);
2003 list_del_init(&q
->all_q_node
);
2004 mutex_unlock(&all_q_mutex
);
2007 /* Basically redo blk_mq_init_queue with queue frozen */
2008 static void blk_mq_queue_reinit(struct request_queue
*q
)
2010 WARN_ON_ONCE(!q
->mq_freeze_depth
);
2012 blk_mq_sysfs_unregister(q
);
2014 blk_mq_update_queue_map(q
->mq_map
, q
->nr_hw_queues
);
2017 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2018 * we should change hctx numa_node according to new topology (this
2019 * involves free and re-allocate memory, worthy doing?)
2022 blk_mq_map_swqueue(q
);
2024 blk_mq_sysfs_register(q
);
2027 static int blk_mq_queue_reinit_notify(struct notifier_block
*nb
,
2028 unsigned long action
, void *hcpu
)
2030 struct request_queue
*q
;
2033 * Before new mappings are established, hotadded cpu might already
2034 * start handling requests. This doesn't break anything as we map
2035 * offline CPUs to first hardware queue. We will re-init the queue
2036 * below to get optimal settings.
2038 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
&&
2039 action
!= CPU_ONLINE
&& action
!= CPU_ONLINE_FROZEN
)
2042 mutex_lock(&all_q_mutex
);
2045 * We need to freeze and reinit all existing queues. Freezing
2046 * involves synchronous wait for an RCU grace period and doing it
2047 * one by one may take a long time. Start freezing all queues in
2048 * one swoop and then wait for the completions so that freezing can
2049 * take place in parallel.
2051 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2052 blk_mq_freeze_queue_start(q
);
2053 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2054 blk_mq_freeze_queue_wait(q
);
2056 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2057 blk_mq_queue_reinit(q
);
2059 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2060 blk_mq_unfreeze_queue(q
);
2062 mutex_unlock(&all_q_mutex
);
2066 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2070 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2071 set
->tags
[i
] = blk_mq_init_rq_map(set
, i
);
2080 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2086 * Allocate the request maps associated with this tag_set. Note that this
2087 * may reduce the depth asked for, if memory is tight. set->queue_depth
2088 * will be updated to reflect the allocated depth.
2090 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2095 depth
= set
->queue_depth
;
2097 err
= __blk_mq_alloc_rq_maps(set
);
2101 set
->queue_depth
>>= 1;
2102 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2106 } while (set
->queue_depth
);
2108 if (!set
->queue_depth
|| err
) {
2109 pr_err("blk-mq: failed to allocate request map\n");
2113 if (depth
!= set
->queue_depth
)
2114 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2115 depth
, set
->queue_depth
);
2121 * Alloc a tag set to be associated with one or more request queues.
2122 * May fail with EINVAL for various error conditions. May adjust the
2123 * requested depth down, if if it too large. In that case, the set
2124 * value will be stored in set->queue_depth.
2126 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2128 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2130 if (!set
->nr_hw_queues
)
2132 if (!set
->queue_depth
)
2134 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2137 if (!set
->nr_hw_queues
|| !set
->ops
->queue_rq
|| !set
->ops
->map_queue
)
2140 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2141 pr_info("blk-mq: reduced tag depth to %u\n",
2143 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2147 * If a crashdump is active, then we are potentially in a very
2148 * memory constrained environment. Limit us to 1 queue and
2149 * 64 tags to prevent using too much memory.
2151 if (is_kdump_kernel()) {
2152 set
->nr_hw_queues
= 1;
2153 set
->queue_depth
= min(64U, set
->queue_depth
);
2156 set
->tags
= kmalloc_node(set
->nr_hw_queues
*
2157 sizeof(struct blk_mq_tags
*),
2158 GFP_KERNEL
, set
->numa_node
);
2162 if (blk_mq_alloc_rq_maps(set
))
2165 mutex_init(&set
->tag_list_lock
);
2166 INIT_LIST_HEAD(&set
->tag_list
);
2174 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2176 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2180 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2182 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2188 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2190 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2192 struct blk_mq_tag_set
*set
= q
->tag_set
;
2193 struct blk_mq_hw_ctx
*hctx
;
2196 if (!set
|| nr
> set
->queue_depth
)
2200 queue_for_each_hw_ctx(q
, hctx
, i
) {
2201 ret
= blk_mq_tag_update_depth(hctx
->tags
, nr
);
2207 q
->nr_requests
= nr
;
2212 void blk_mq_disable_hotplug(void)
2214 mutex_lock(&all_q_mutex
);
2217 void blk_mq_enable_hotplug(void)
2219 mutex_unlock(&all_q_mutex
);
2222 static int __init
blk_mq_init(void)
2226 hotcpu_notifier(blk_mq_queue_reinit_notify
, 0);
2230 subsys_initcall(blk_mq_init
);