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 void blk_mq_start_request(struct request
*rq
)
409 struct request_queue
*q
= rq
->q
;
411 trace_block_rq_issue(q
, rq
);
413 rq
->resid_len
= blk_rq_bytes(rq
);
414 if (unlikely(blk_bidi_rq(rq
)))
415 rq
->next_rq
->resid_len
= blk_rq_bytes(rq
->next_rq
);
420 * Ensure that ->deadline is visible before set the started
421 * flag and clear the completed flag.
423 smp_mb__before_atomic();
426 * Mark us as started and clear complete. Complete might have been
427 * set if requeue raced with timeout, which then marked it as
428 * complete. So be sure to clear complete again when we start
429 * the request, otherwise we'll ignore the completion event.
431 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
432 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
433 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
434 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
436 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
438 * Make sure space for the drain appears. We know we can do
439 * this because max_hw_segments has been adjusted to be one
440 * fewer than the device can handle.
442 rq
->nr_phys_segments
++;
445 EXPORT_SYMBOL(blk_mq_start_request
);
447 static void __blk_mq_requeue_request(struct request
*rq
)
449 struct request_queue
*q
= rq
->q
;
451 trace_block_rq_requeue(q
, rq
);
453 if (test_and_clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
454 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
455 rq
->nr_phys_segments
--;
459 void blk_mq_requeue_request(struct request
*rq
)
461 __blk_mq_requeue_request(rq
);
463 BUG_ON(blk_queued_rq(rq
));
464 blk_mq_add_to_requeue_list(rq
, true);
466 EXPORT_SYMBOL(blk_mq_requeue_request
);
468 static void blk_mq_requeue_work(struct work_struct
*work
)
470 struct request_queue
*q
=
471 container_of(work
, struct request_queue
, requeue_work
);
473 struct request
*rq
, *next
;
476 spin_lock_irqsave(&q
->requeue_lock
, flags
);
477 list_splice_init(&q
->requeue_list
, &rq_list
);
478 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
480 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
481 if (!(rq
->cmd_flags
& REQ_SOFTBARRIER
))
484 rq
->cmd_flags
&= ~REQ_SOFTBARRIER
;
485 list_del_init(&rq
->queuelist
);
486 blk_mq_insert_request(rq
, true, false, false);
489 while (!list_empty(&rq_list
)) {
490 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
491 list_del_init(&rq
->queuelist
);
492 blk_mq_insert_request(rq
, false, false, false);
496 * Use the start variant of queue running here, so that running
497 * the requeue work will kick stopped queues.
499 blk_mq_start_hw_queues(q
);
502 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
)
504 struct request_queue
*q
= rq
->q
;
508 * We abuse this flag that is otherwise used by the I/O scheduler to
509 * request head insertation from the workqueue.
511 BUG_ON(rq
->cmd_flags
& REQ_SOFTBARRIER
);
513 spin_lock_irqsave(&q
->requeue_lock
, flags
);
515 rq
->cmd_flags
|= REQ_SOFTBARRIER
;
516 list_add(&rq
->queuelist
, &q
->requeue_list
);
518 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
520 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
522 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
524 void blk_mq_kick_requeue_list(struct request_queue
*q
)
526 kblockd_schedule_work(&q
->requeue_work
);
528 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
530 static inline bool is_flush_request(struct request
*rq
,
531 struct blk_flush_queue
*fq
, unsigned int tag
)
533 return ((rq
->cmd_flags
& REQ_FLUSH_SEQ
) &&
534 fq
->flush_rq
->tag
== tag
);
537 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
539 struct request
*rq
= tags
->rqs
[tag
];
540 /* mq_ctx of flush rq is always cloned from the corresponding req */
541 struct blk_flush_queue
*fq
= blk_get_flush_queue(rq
->q
, rq
->mq_ctx
);
543 if (!is_flush_request(rq
, fq
, tag
))
548 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
550 struct blk_mq_timeout_data
{
552 unsigned int next_set
;
555 void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
557 struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
558 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
561 * We know that complete is set at this point. If STARTED isn't set
562 * anymore, then the request isn't active and the "timeout" should
563 * just be ignored. This can happen due to the bitflag ordering.
564 * Timeout first checks if STARTED is set, and if it is, assumes
565 * the request is active. But if we race with completion, then
566 * we both flags will get cleared. So check here again, and ignore
567 * a timeout event with a request that isn't active.
569 if (!test_bit(REQ_ATOM_STARTED
, &req
->atomic_flags
))
573 ret
= ops
->timeout(req
, reserved
);
577 __blk_mq_complete_request(req
);
579 case BLK_EH_RESET_TIMER
:
581 blk_clear_rq_complete(req
);
583 case BLK_EH_NOT_HANDLED
:
586 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
591 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
592 struct request
*rq
, void *priv
, bool reserved
)
594 struct blk_mq_timeout_data
*data
= priv
;
596 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
599 if (time_after_eq(jiffies
, rq
->deadline
)) {
600 if (!blk_mark_rq_complete(rq
))
601 blk_mq_rq_timed_out(rq
, reserved
);
602 } else if (!data
->next_set
|| time_after(data
->next
, rq
->deadline
)) {
603 data
->next
= rq
->deadline
;
608 static void blk_mq_rq_timer(unsigned long priv
)
610 struct request_queue
*q
= (struct request_queue
*)priv
;
611 struct blk_mq_timeout_data data
= {
615 struct blk_mq_hw_ctx
*hctx
;
618 queue_for_each_hw_ctx(q
, hctx
, i
) {
620 * If not software queues are currently mapped to this
621 * hardware queue, there's nothing to check
623 if (!blk_mq_hw_queue_mapped(hctx
))
626 blk_mq_tag_busy_iter(hctx
, blk_mq_check_expired
, &data
);
630 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
631 mod_timer(&q
->timeout
, data
.next
);
633 queue_for_each_hw_ctx(q
, hctx
, i
)
634 blk_mq_tag_idle(hctx
);
639 * Reverse check our software queue for entries that we could potentially
640 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
641 * too much time checking for merges.
643 static bool blk_mq_attempt_merge(struct request_queue
*q
,
644 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
649 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
655 if (!blk_rq_merge_ok(rq
, bio
))
658 el_ret
= blk_try_merge(rq
, bio
);
659 if (el_ret
== ELEVATOR_BACK_MERGE
) {
660 if (bio_attempt_back_merge(q
, rq
, bio
)) {
665 } else if (el_ret
== ELEVATOR_FRONT_MERGE
) {
666 if (bio_attempt_front_merge(q
, rq
, bio
)) {
678 * Process software queues that have been marked busy, splicing them
679 * to the for-dispatch
681 static void flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
683 struct blk_mq_ctx
*ctx
;
686 for (i
= 0; i
< hctx
->ctx_map
.map_size
; i
++) {
687 struct blk_align_bitmap
*bm
= &hctx
->ctx_map
.map
[i
];
688 unsigned int off
, bit
;
694 off
= i
* hctx
->ctx_map
.bits_per_word
;
696 bit
= find_next_bit(&bm
->word
, bm
->depth
, bit
);
697 if (bit
>= bm
->depth
)
700 ctx
= hctx
->ctxs
[bit
+ off
];
701 clear_bit(bit
, &bm
->word
);
702 spin_lock(&ctx
->lock
);
703 list_splice_tail_init(&ctx
->rq_list
, list
);
704 spin_unlock(&ctx
->lock
);
712 * Run this hardware queue, pulling any software queues mapped to it in.
713 * Note that this function currently has various problems around ordering
714 * of IO. In particular, we'd like FIFO behaviour on handling existing
715 * items on the hctx->dispatch list. Ignore that for now.
717 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
719 struct request_queue
*q
= hctx
->queue
;
722 LIST_HEAD(driver_list
);
723 struct list_head
*dptr
;
726 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
));
728 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)))
734 * Touch any software queue that has pending entries.
736 flush_busy_ctxs(hctx
, &rq_list
);
739 * If we have previous entries on our dispatch list, grab them
740 * and stuff them at the front for more fair dispatch.
742 if (!list_empty_careful(&hctx
->dispatch
)) {
743 spin_lock(&hctx
->lock
);
744 if (!list_empty(&hctx
->dispatch
))
745 list_splice_init(&hctx
->dispatch
, &rq_list
);
746 spin_unlock(&hctx
->lock
);
750 * Start off with dptr being NULL, so we start the first request
751 * immediately, even if we have more pending.
756 * Now process all the entries, sending them to the driver.
759 while (!list_empty(&rq_list
)) {
760 struct blk_mq_queue_data bd
;
763 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
764 list_del_init(&rq
->queuelist
);
768 bd
.last
= list_empty(&rq_list
);
770 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
772 case BLK_MQ_RQ_QUEUE_OK
:
775 case BLK_MQ_RQ_QUEUE_BUSY
:
776 list_add(&rq
->queuelist
, &rq_list
);
777 __blk_mq_requeue_request(rq
);
780 pr_err("blk-mq: bad return on queue: %d\n", ret
);
781 case BLK_MQ_RQ_QUEUE_ERROR
:
783 blk_mq_end_request(rq
, rq
->errors
);
787 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
791 * We've done the first request. If we have more than 1
792 * left in the list, set dptr to defer issue.
794 if (!dptr
&& rq_list
.next
!= rq_list
.prev
)
799 hctx
->dispatched
[0]++;
800 else if (queued
< (1 << (BLK_MQ_MAX_DISPATCH_ORDER
- 1)))
801 hctx
->dispatched
[ilog2(queued
) + 1]++;
804 * Any items that need requeuing? Stuff them into hctx->dispatch,
805 * that is where we will continue on next queue run.
807 if (!list_empty(&rq_list
)) {
808 spin_lock(&hctx
->lock
);
809 list_splice(&rq_list
, &hctx
->dispatch
);
810 spin_unlock(&hctx
->lock
);
815 * It'd be great if the workqueue API had a way to pass
816 * in a mask and had some smarts for more clever placement.
817 * For now we just round-robin here, switching for every
818 * BLK_MQ_CPU_WORK_BATCH queued items.
820 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
822 if (hctx
->queue
->nr_hw_queues
== 1)
823 return WORK_CPU_UNBOUND
;
825 if (--hctx
->next_cpu_batch
<= 0) {
826 int cpu
= hctx
->next_cpu
, next_cpu
;
828 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
829 if (next_cpu
>= nr_cpu_ids
)
830 next_cpu
= cpumask_first(hctx
->cpumask
);
832 hctx
->next_cpu
= next_cpu
;
833 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
838 return hctx
->next_cpu
;
841 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
843 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
) ||
844 !blk_mq_hw_queue_mapped(hctx
)))
849 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
850 __blk_mq_run_hw_queue(hctx
);
858 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
862 void blk_mq_run_queues(struct request_queue
*q
, bool async
)
864 struct blk_mq_hw_ctx
*hctx
;
867 queue_for_each_hw_ctx(q
, hctx
, i
) {
868 if ((!blk_mq_hctx_has_pending(hctx
) &&
869 list_empty_careful(&hctx
->dispatch
)) ||
870 test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
873 blk_mq_run_hw_queue(hctx
, async
);
876 EXPORT_SYMBOL(blk_mq_run_queues
);
878 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
880 cancel_delayed_work(&hctx
->run_work
);
881 cancel_delayed_work(&hctx
->delay_work
);
882 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
884 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
886 void blk_mq_stop_hw_queues(struct request_queue
*q
)
888 struct blk_mq_hw_ctx
*hctx
;
891 queue_for_each_hw_ctx(q
, hctx
, i
)
892 blk_mq_stop_hw_queue(hctx
);
894 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
896 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
898 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
900 blk_mq_run_hw_queue(hctx
, false);
902 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
904 void blk_mq_start_hw_queues(struct request_queue
*q
)
906 struct blk_mq_hw_ctx
*hctx
;
909 queue_for_each_hw_ctx(q
, hctx
, i
)
910 blk_mq_start_hw_queue(hctx
);
912 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
915 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
917 struct blk_mq_hw_ctx
*hctx
;
920 queue_for_each_hw_ctx(q
, hctx
, i
) {
921 if (!test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
924 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
925 blk_mq_run_hw_queue(hctx
, async
);
928 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
930 static void blk_mq_run_work_fn(struct work_struct
*work
)
932 struct blk_mq_hw_ctx
*hctx
;
934 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
936 __blk_mq_run_hw_queue(hctx
);
939 static void blk_mq_delay_work_fn(struct work_struct
*work
)
941 struct blk_mq_hw_ctx
*hctx
;
943 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
945 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
946 __blk_mq_run_hw_queue(hctx
);
949 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
951 if (unlikely(!blk_mq_hw_queue_mapped(hctx
)))
954 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
955 &hctx
->delay_work
, msecs_to_jiffies(msecs
));
957 EXPORT_SYMBOL(blk_mq_delay_queue
);
959 static void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
,
960 struct request
*rq
, bool at_head
)
962 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
964 trace_block_rq_insert(hctx
->queue
, rq
);
967 list_add(&rq
->queuelist
, &ctx
->rq_list
);
969 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
971 blk_mq_hctx_mark_pending(hctx
, ctx
);
974 void blk_mq_insert_request(struct request
*rq
, bool at_head
, bool run_queue
,
977 struct request_queue
*q
= rq
->q
;
978 struct blk_mq_hw_ctx
*hctx
;
979 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
, *current_ctx
;
981 current_ctx
= blk_mq_get_ctx(q
);
982 if (!cpu_online(ctx
->cpu
))
983 rq
->mq_ctx
= ctx
= current_ctx
;
985 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
987 spin_lock(&ctx
->lock
);
988 __blk_mq_insert_request(hctx
, rq
, at_head
);
989 spin_unlock(&ctx
->lock
);
992 blk_mq_run_hw_queue(hctx
, async
);
994 blk_mq_put_ctx(current_ctx
);
997 static void blk_mq_insert_requests(struct request_queue
*q
,
998 struct blk_mq_ctx
*ctx
,
999 struct list_head
*list
,
1004 struct blk_mq_hw_ctx
*hctx
;
1005 struct blk_mq_ctx
*current_ctx
;
1007 trace_block_unplug(q
, depth
, !from_schedule
);
1009 current_ctx
= blk_mq_get_ctx(q
);
1011 if (!cpu_online(ctx
->cpu
))
1013 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1016 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1019 spin_lock(&ctx
->lock
);
1020 while (!list_empty(list
)) {
1023 rq
= list_first_entry(list
, struct request
, queuelist
);
1024 list_del_init(&rq
->queuelist
);
1026 __blk_mq_insert_request(hctx
, rq
, false);
1028 spin_unlock(&ctx
->lock
);
1030 blk_mq_run_hw_queue(hctx
, from_schedule
);
1031 blk_mq_put_ctx(current_ctx
);
1034 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1036 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1037 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1039 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1040 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1041 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1044 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1046 struct blk_mq_ctx
*this_ctx
;
1047 struct request_queue
*this_q
;
1050 LIST_HEAD(ctx_list
);
1053 list_splice_init(&plug
->mq_list
, &list
);
1055 list_sort(NULL
, &list
, plug_ctx_cmp
);
1061 while (!list_empty(&list
)) {
1062 rq
= list_entry_rq(list
.next
);
1063 list_del_init(&rq
->queuelist
);
1065 if (rq
->mq_ctx
!= this_ctx
) {
1067 blk_mq_insert_requests(this_q
, this_ctx
,
1072 this_ctx
= rq
->mq_ctx
;
1078 list_add_tail(&rq
->queuelist
, &ctx_list
);
1082 * If 'this_ctx' is set, we know we have entries to complete
1083 * on 'ctx_list'. Do those.
1086 blk_mq_insert_requests(this_q
, this_ctx
, &ctx_list
, depth
,
1091 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1093 init_request_from_bio(rq
, bio
);
1095 if (blk_do_io_stat(rq
))
1096 blk_account_io_start(rq
, 1);
1099 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1101 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1102 !blk_queue_nomerges(hctx
->queue
);
1105 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx
*hctx
,
1106 struct blk_mq_ctx
*ctx
,
1107 struct request
*rq
, struct bio
*bio
)
1109 if (!hctx_allow_merges(hctx
)) {
1110 blk_mq_bio_to_request(rq
, bio
);
1111 spin_lock(&ctx
->lock
);
1113 __blk_mq_insert_request(hctx
, rq
, false);
1114 spin_unlock(&ctx
->lock
);
1117 struct request_queue
*q
= hctx
->queue
;
1119 spin_lock(&ctx
->lock
);
1120 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1121 blk_mq_bio_to_request(rq
, bio
);
1125 spin_unlock(&ctx
->lock
);
1126 __blk_mq_free_request(hctx
, ctx
, rq
);
1131 struct blk_map_ctx
{
1132 struct blk_mq_hw_ctx
*hctx
;
1133 struct blk_mq_ctx
*ctx
;
1136 static struct request
*blk_mq_map_request(struct request_queue
*q
,
1138 struct blk_map_ctx
*data
)
1140 struct blk_mq_hw_ctx
*hctx
;
1141 struct blk_mq_ctx
*ctx
;
1143 int rw
= bio_data_dir(bio
);
1144 struct blk_mq_alloc_data alloc_data
;
1146 if (unlikely(blk_mq_queue_enter(q
))) {
1147 bio_endio(bio
, -EIO
);
1151 ctx
= blk_mq_get_ctx(q
);
1152 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1154 if (rw_is_sync(bio
->bi_rw
))
1157 trace_block_getrq(q
, bio
, rw
);
1158 blk_mq_set_alloc_data(&alloc_data
, q
, GFP_ATOMIC
, false, ctx
,
1160 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
1161 if (unlikely(!rq
)) {
1162 __blk_mq_run_hw_queue(hctx
);
1163 blk_mq_put_ctx(ctx
);
1164 trace_block_sleeprq(q
, bio
, rw
);
1166 ctx
= blk_mq_get_ctx(q
);
1167 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1168 blk_mq_set_alloc_data(&alloc_data
, q
,
1169 __GFP_WAIT
|GFP_ATOMIC
, false, ctx
, hctx
);
1170 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
1171 ctx
= alloc_data
.ctx
;
1172 hctx
= alloc_data
.hctx
;
1182 * Multiple hardware queue variant. This will not use per-process plugs,
1183 * but will attempt to bypass the hctx queueing if we can go straight to
1184 * hardware for SYNC IO.
1186 static void blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1188 const int is_sync
= rw_is_sync(bio
->bi_rw
);
1189 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
1190 struct blk_map_ctx data
;
1193 blk_queue_bounce(q
, &bio
);
1195 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1196 bio_endio(bio
, -EIO
);
1200 rq
= blk_mq_map_request(q
, bio
, &data
);
1204 if (unlikely(is_flush_fua
)) {
1205 blk_mq_bio_to_request(rq
, bio
);
1206 blk_insert_flush(rq
);
1211 * If the driver supports defer issued based on 'last', then
1212 * queue it up like normal since we can potentially save some
1215 if (is_sync
&& !(data
.hctx
->flags
& BLK_MQ_F_DEFER_ISSUE
)) {
1216 struct blk_mq_queue_data bd
= {
1223 blk_mq_bio_to_request(rq
, bio
);
1226 * For OK queue, we are done. For error, kill it. Any other
1227 * error (busy), just add it to our list as we previously
1230 ret
= q
->mq_ops
->queue_rq(data
.hctx
, &bd
);
1231 if (ret
== BLK_MQ_RQ_QUEUE_OK
)
1234 __blk_mq_requeue_request(rq
);
1236 if (ret
== BLK_MQ_RQ_QUEUE_ERROR
) {
1238 blk_mq_end_request(rq
, rq
->errors
);
1244 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1246 * For a SYNC request, send it to the hardware immediately. For
1247 * an ASYNC request, just ensure that we run it later on. The
1248 * latter allows for merging opportunities and more efficient
1252 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1255 blk_mq_put_ctx(data
.ctx
);
1259 * Single hardware queue variant. This will attempt to use any per-process
1260 * plug for merging and IO deferral.
1262 static void blk_sq_make_request(struct request_queue
*q
, struct bio
*bio
)
1264 const int is_sync
= rw_is_sync(bio
->bi_rw
);
1265 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
1266 unsigned int use_plug
, request_count
= 0;
1267 struct blk_map_ctx data
;
1271 * If we have multiple hardware queues, just go directly to
1272 * one of those for sync IO.
1274 use_plug
= !is_flush_fua
&& !is_sync
;
1276 blk_queue_bounce(q
, &bio
);
1278 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1279 bio_endio(bio
, -EIO
);
1283 if (use_plug
&& !blk_queue_nomerges(q
) &&
1284 blk_attempt_plug_merge(q
, bio
, &request_count
))
1287 rq
= blk_mq_map_request(q
, bio
, &data
);
1291 if (unlikely(is_flush_fua
)) {
1292 blk_mq_bio_to_request(rq
, bio
);
1293 blk_insert_flush(rq
);
1298 * A task plug currently exists. Since this is completely lockless,
1299 * utilize that to temporarily store requests until the task is
1300 * either done or scheduled away.
1303 struct blk_plug
*plug
= current
->plug
;
1306 blk_mq_bio_to_request(rq
, bio
);
1307 if (list_empty(&plug
->mq_list
))
1308 trace_block_plug(q
);
1309 else if (request_count
>= BLK_MAX_REQUEST_COUNT
) {
1310 blk_flush_plug_list(plug
, false);
1311 trace_block_plug(q
);
1313 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1314 blk_mq_put_ctx(data
.ctx
);
1319 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1321 * For a SYNC request, send it to the hardware immediately. For
1322 * an ASYNC request, just ensure that we run it later on. The
1323 * latter allows for merging opportunities and more efficient
1327 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1330 blk_mq_put_ctx(data
.ctx
);
1334 * Default mapping to a software queue, since we use one per CPU.
1336 struct blk_mq_hw_ctx
*blk_mq_map_queue(struct request_queue
*q
, const int cpu
)
1338 return q
->queue_hw_ctx
[q
->mq_map
[cpu
]];
1340 EXPORT_SYMBOL(blk_mq_map_queue
);
1342 static void blk_mq_free_rq_map(struct blk_mq_tag_set
*set
,
1343 struct blk_mq_tags
*tags
, unsigned int hctx_idx
)
1347 if (tags
->rqs
&& set
->ops
->exit_request
) {
1350 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1353 set
->ops
->exit_request(set
->driver_data
, tags
->rqs
[i
],
1355 tags
->rqs
[i
] = NULL
;
1359 while (!list_empty(&tags
->page_list
)) {
1360 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1361 list_del_init(&page
->lru
);
1362 __free_pages(page
, page
->private);
1367 blk_mq_free_tags(tags
);
1370 static size_t order_to_size(unsigned int order
)
1372 return (size_t)PAGE_SIZE
<< order
;
1375 static struct blk_mq_tags
*blk_mq_init_rq_map(struct blk_mq_tag_set
*set
,
1376 unsigned int hctx_idx
)
1378 struct blk_mq_tags
*tags
;
1379 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1380 size_t rq_size
, left
;
1382 tags
= blk_mq_init_tags(set
->queue_depth
, set
->reserved_tags
,
1387 INIT_LIST_HEAD(&tags
->page_list
);
1389 tags
->rqs
= kzalloc_node(set
->queue_depth
* sizeof(struct request
*),
1390 GFP_KERNEL
| __GFP_NOWARN
| __GFP_NORETRY
,
1393 blk_mq_free_tags(tags
);
1398 * rq_size is the size of the request plus driver payload, rounded
1399 * to the cacheline size
1401 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1403 left
= rq_size
* set
->queue_depth
;
1405 for (i
= 0; i
< set
->queue_depth
; ) {
1406 int this_order
= max_order
;
1411 while (left
< order_to_size(this_order
- 1) && this_order
)
1415 page
= alloc_pages_node(set
->numa_node
,
1416 GFP_KERNEL
| __GFP_NOWARN
| __GFP_NORETRY
,
1422 if (order_to_size(this_order
) < rq_size
)
1429 page
->private = this_order
;
1430 list_add_tail(&page
->lru
, &tags
->page_list
);
1432 p
= page_address(page
);
1433 entries_per_page
= order_to_size(this_order
) / rq_size
;
1434 to_do
= min(entries_per_page
, set
->queue_depth
- i
);
1435 left
-= to_do
* rq_size
;
1436 for (j
= 0; j
< to_do
; j
++) {
1438 tags
->rqs
[i
]->atomic_flags
= 0;
1439 tags
->rqs
[i
]->cmd_flags
= 0;
1440 if (set
->ops
->init_request
) {
1441 if (set
->ops
->init_request(set
->driver_data
,
1442 tags
->rqs
[i
], hctx_idx
, i
,
1444 tags
->rqs
[i
] = NULL
;
1457 blk_mq_free_rq_map(set
, tags
, hctx_idx
);
1461 static void blk_mq_free_bitmap(struct blk_mq_ctxmap
*bitmap
)
1466 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap
*bitmap
, int node
)
1468 unsigned int bpw
= 8, total
, num_maps
, i
;
1470 bitmap
->bits_per_word
= bpw
;
1472 num_maps
= ALIGN(nr_cpu_ids
, bpw
) / bpw
;
1473 bitmap
->map
= kzalloc_node(num_maps
* sizeof(struct blk_align_bitmap
),
1478 bitmap
->map_size
= num_maps
;
1481 for (i
= 0; i
< num_maps
; i
++) {
1482 bitmap
->map
[i
].depth
= min(total
, bitmap
->bits_per_word
);
1483 total
-= bitmap
->map
[i
].depth
;
1489 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx
*hctx
, int cpu
)
1491 struct request_queue
*q
= hctx
->queue
;
1492 struct blk_mq_ctx
*ctx
;
1496 * Move ctx entries to new CPU, if this one is going away.
1498 ctx
= __blk_mq_get_ctx(q
, cpu
);
1500 spin_lock(&ctx
->lock
);
1501 if (!list_empty(&ctx
->rq_list
)) {
1502 list_splice_init(&ctx
->rq_list
, &tmp
);
1503 blk_mq_hctx_clear_pending(hctx
, ctx
);
1505 spin_unlock(&ctx
->lock
);
1507 if (list_empty(&tmp
))
1510 ctx
= blk_mq_get_ctx(q
);
1511 spin_lock(&ctx
->lock
);
1513 while (!list_empty(&tmp
)) {
1516 rq
= list_first_entry(&tmp
, struct request
, queuelist
);
1518 list_move_tail(&rq
->queuelist
, &ctx
->rq_list
);
1521 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1522 blk_mq_hctx_mark_pending(hctx
, ctx
);
1524 spin_unlock(&ctx
->lock
);
1526 blk_mq_run_hw_queue(hctx
, true);
1527 blk_mq_put_ctx(ctx
);
1531 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx
*hctx
, int cpu
)
1533 struct request_queue
*q
= hctx
->queue
;
1534 struct blk_mq_tag_set
*set
= q
->tag_set
;
1536 if (set
->tags
[hctx
->queue_num
])
1539 set
->tags
[hctx
->queue_num
] = blk_mq_init_rq_map(set
, hctx
->queue_num
);
1540 if (!set
->tags
[hctx
->queue_num
])
1543 hctx
->tags
= set
->tags
[hctx
->queue_num
];
1547 static int blk_mq_hctx_notify(void *data
, unsigned long action
,
1550 struct blk_mq_hw_ctx
*hctx
= data
;
1552 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
)
1553 return blk_mq_hctx_cpu_offline(hctx
, cpu
);
1554 else if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
)
1555 return blk_mq_hctx_cpu_online(hctx
, cpu
);
1560 static void blk_mq_exit_hctx(struct request_queue
*q
,
1561 struct blk_mq_tag_set
*set
,
1562 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
1564 unsigned flush_start_tag
= set
->queue_depth
;
1566 blk_mq_tag_idle(hctx
);
1568 if (set
->ops
->exit_request
)
1569 set
->ops
->exit_request(set
->driver_data
,
1570 hctx
->fq
->flush_rq
, hctx_idx
,
1571 flush_start_tag
+ hctx_idx
);
1573 if (set
->ops
->exit_hctx
)
1574 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1576 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1577 blk_free_flush_queue(hctx
->fq
);
1579 blk_mq_free_bitmap(&hctx
->ctx_map
);
1582 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1583 struct blk_mq_tag_set
*set
, int nr_queue
)
1585 struct blk_mq_hw_ctx
*hctx
;
1588 queue_for_each_hw_ctx(q
, hctx
, i
) {
1591 blk_mq_exit_hctx(q
, set
, hctx
, i
);
1595 static void blk_mq_free_hw_queues(struct request_queue
*q
,
1596 struct blk_mq_tag_set
*set
)
1598 struct blk_mq_hw_ctx
*hctx
;
1601 queue_for_each_hw_ctx(q
, hctx
, i
) {
1602 free_cpumask_var(hctx
->cpumask
);
1607 static int blk_mq_init_hctx(struct request_queue
*q
,
1608 struct blk_mq_tag_set
*set
,
1609 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
1612 unsigned flush_start_tag
= set
->queue_depth
;
1614 node
= hctx
->numa_node
;
1615 if (node
== NUMA_NO_NODE
)
1616 node
= hctx
->numa_node
= set
->numa_node
;
1618 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1619 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1620 spin_lock_init(&hctx
->lock
);
1621 INIT_LIST_HEAD(&hctx
->dispatch
);
1623 hctx
->queue_num
= hctx_idx
;
1624 hctx
->flags
= set
->flags
;
1626 blk_mq_init_cpu_notifier(&hctx
->cpu_notifier
,
1627 blk_mq_hctx_notify
, hctx
);
1628 blk_mq_register_cpu_notifier(&hctx
->cpu_notifier
);
1630 hctx
->tags
= set
->tags
[hctx_idx
];
1633 * Allocate space for all possible cpus to avoid allocation at
1636 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1639 goto unregister_cpu_notifier
;
1641 if (blk_mq_alloc_bitmap(&hctx
->ctx_map
, node
))
1646 if (set
->ops
->init_hctx
&&
1647 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
1650 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
1654 if (set
->ops
->init_request
&&
1655 set
->ops
->init_request(set
->driver_data
,
1656 hctx
->fq
->flush_rq
, hctx_idx
,
1657 flush_start_tag
+ hctx_idx
, node
))
1665 if (set
->ops
->exit_hctx
)
1666 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1668 blk_mq_free_bitmap(&hctx
->ctx_map
);
1671 unregister_cpu_notifier
:
1672 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1677 static int blk_mq_init_hw_queues(struct request_queue
*q
,
1678 struct blk_mq_tag_set
*set
)
1680 struct blk_mq_hw_ctx
*hctx
;
1684 * Initialize hardware queues
1686 queue_for_each_hw_ctx(q
, hctx
, i
) {
1687 if (blk_mq_init_hctx(q
, set
, hctx
, i
))
1691 if (i
== q
->nr_hw_queues
)
1697 blk_mq_exit_hw_queues(q
, set
, i
);
1702 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1703 unsigned int nr_hw_queues
)
1707 for_each_possible_cpu(i
) {
1708 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1709 struct blk_mq_hw_ctx
*hctx
;
1711 memset(__ctx
, 0, sizeof(*__ctx
));
1713 spin_lock_init(&__ctx
->lock
);
1714 INIT_LIST_HEAD(&__ctx
->rq_list
);
1717 /* If the cpu isn't online, the cpu is mapped to first hctx */
1721 hctx
= q
->mq_ops
->map_queue(q
, i
);
1722 cpumask_set_cpu(i
, hctx
->cpumask
);
1726 * Set local node, IFF we have more than one hw queue. If
1727 * not, we remain on the home node of the device
1729 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1730 hctx
->numa_node
= cpu_to_node(i
);
1734 static void blk_mq_map_swqueue(struct request_queue
*q
)
1737 struct blk_mq_hw_ctx
*hctx
;
1738 struct blk_mq_ctx
*ctx
;
1740 queue_for_each_hw_ctx(q
, hctx
, i
) {
1741 cpumask_clear(hctx
->cpumask
);
1746 * Map software to hardware queues
1748 queue_for_each_ctx(q
, ctx
, i
) {
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
);
1755 ctx
->index_hw
= hctx
->nr_ctx
;
1756 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
1759 queue_for_each_hw_ctx(q
, hctx
, i
) {
1761 * If no software queues are mapped to this hardware queue,
1762 * disable it and free the request entries.
1764 if (!hctx
->nr_ctx
) {
1765 struct blk_mq_tag_set
*set
= q
->tag_set
;
1768 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1769 set
->tags
[i
] = NULL
;
1776 * Initialize batch roundrobin counts
1778 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
1779 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1783 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
)
1785 struct blk_mq_hw_ctx
*hctx
;
1786 struct request_queue
*q
;
1790 if (set
->tag_list
.next
== set
->tag_list
.prev
)
1795 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
1796 blk_mq_freeze_queue(q
);
1798 queue_for_each_hw_ctx(q
, hctx
, i
) {
1800 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
1802 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
1804 blk_mq_unfreeze_queue(q
);
1808 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
1810 struct blk_mq_tag_set
*set
= q
->tag_set
;
1812 mutex_lock(&set
->tag_list_lock
);
1813 list_del_init(&q
->tag_set_list
);
1814 blk_mq_update_tag_set_depth(set
);
1815 mutex_unlock(&set
->tag_list_lock
);
1818 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
1819 struct request_queue
*q
)
1823 mutex_lock(&set
->tag_list_lock
);
1824 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
1825 blk_mq_update_tag_set_depth(set
);
1826 mutex_unlock(&set
->tag_list_lock
);
1829 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
1831 struct blk_mq_hw_ctx
**hctxs
;
1832 struct blk_mq_ctx __percpu
*ctx
;
1833 struct request_queue
*q
;
1837 ctx
= alloc_percpu(struct blk_mq_ctx
);
1839 return ERR_PTR(-ENOMEM
);
1841 hctxs
= kmalloc_node(set
->nr_hw_queues
* sizeof(*hctxs
), GFP_KERNEL
,
1847 map
= blk_mq_make_queue_map(set
);
1851 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1852 int node
= blk_mq_hw_queue_to_node(map
, i
);
1854 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
1859 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
1863 atomic_set(&hctxs
[i
]->nr_active
, 0);
1864 hctxs
[i
]->numa_node
= node
;
1865 hctxs
[i
]->queue_num
= i
;
1868 q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
1873 * Init percpu_ref in atomic mode so that it's faster to shutdown.
1874 * See blk_register_queue() for details.
1876 if (percpu_ref_init(&q
->mq_usage_counter
, blk_mq_usage_counter_release
,
1877 PERCPU_REF_INIT_ATOMIC
, GFP_KERNEL
))
1880 setup_timer(&q
->timeout
, blk_mq_rq_timer
, (unsigned long) q
);
1881 blk_queue_rq_timeout(q
, 30000);
1883 q
->nr_queues
= nr_cpu_ids
;
1884 q
->nr_hw_queues
= set
->nr_hw_queues
;
1888 q
->queue_hw_ctx
= hctxs
;
1890 q
->mq_ops
= set
->ops
;
1891 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
1893 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
1894 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
1896 q
->sg_reserved_size
= INT_MAX
;
1898 INIT_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
1899 INIT_LIST_HEAD(&q
->requeue_list
);
1900 spin_lock_init(&q
->requeue_lock
);
1902 if (q
->nr_hw_queues
> 1)
1903 blk_queue_make_request(q
, blk_mq_make_request
);
1905 blk_queue_make_request(q
, blk_sq_make_request
);
1908 blk_queue_rq_timeout(q
, set
->timeout
);
1911 * Do this after blk_queue_make_request() overrides it...
1913 q
->nr_requests
= set
->queue_depth
;
1915 if (set
->ops
->complete
)
1916 blk_queue_softirq_done(q
, set
->ops
->complete
);
1918 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
1920 if (blk_mq_init_hw_queues(q
, set
))
1923 mutex_lock(&all_q_mutex
);
1924 list_add_tail(&q
->all_q_node
, &all_q_list
);
1925 mutex_unlock(&all_q_mutex
);
1927 blk_mq_add_queue_tag_set(set
, q
);
1929 blk_mq_map_swqueue(q
);
1934 blk_cleanup_queue(q
);
1937 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1940 free_cpumask_var(hctxs
[i
]->cpumask
);
1947 return ERR_PTR(-ENOMEM
);
1949 EXPORT_SYMBOL(blk_mq_init_queue
);
1951 void blk_mq_free_queue(struct request_queue
*q
)
1953 struct blk_mq_tag_set
*set
= q
->tag_set
;
1955 blk_mq_del_queue_tag_set(q
);
1957 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
1958 blk_mq_free_hw_queues(q
, set
);
1960 percpu_ref_exit(&q
->mq_usage_counter
);
1962 free_percpu(q
->queue_ctx
);
1963 kfree(q
->queue_hw_ctx
);
1966 q
->queue_ctx
= NULL
;
1967 q
->queue_hw_ctx
= NULL
;
1970 mutex_lock(&all_q_mutex
);
1971 list_del_init(&q
->all_q_node
);
1972 mutex_unlock(&all_q_mutex
);
1975 /* Basically redo blk_mq_init_queue with queue frozen */
1976 static void blk_mq_queue_reinit(struct request_queue
*q
)
1978 WARN_ON_ONCE(!q
->mq_freeze_depth
);
1980 blk_mq_sysfs_unregister(q
);
1982 blk_mq_update_queue_map(q
->mq_map
, q
->nr_hw_queues
);
1985 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1986 * we should change hctx numa_node according to new topology (this
1987 * involves free and re-allocate memory, worthy doing?)
1990 blk_mq_map_swqueue(q
);
1992 blk_mq_sysfs_register(q
);
1995 static int blk_mq_queue_reinit_notify(struct notifier_block
*nb
,
1996 unsigned long action
, void *hcpu
)
1998 struct request_queue
*q
;
2001 * Before new mappings are established, hotadded cpu might already
2002 * start handling requests. This doesn't break anything as we map
2003 * offline CPUs to first hardware queue. We will re-init the queue
2004 * below to get optimal settings.
2006 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
&&
2007 action
!= CPU_ONLINE
&& action
!= CPU_ONLINE_FROZEN
)
2010 mutex_lock(&all_q_mutex
);
2013 * We need to freeze and reinit all existing queues. Freezing
2014 * involves synchronous wait for an RCU grace period and doing it
2015 * one by one may take a long time. Start freezing all queues in
2016 * one swoop and then wait for the completions so that freezing can
2017 * take place in parallel.
2019 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2020 blk_mq_freeze_queue_start(q
);
2021 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2022 blk_mq_freeze_queue_wait(q
);
2024 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2025 blk_mq_queue_reinit(q
);
2027 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2028 blk_mq_unfreeze_queue(q
);
2030 mutex_unlock(&all_q_mutex
);
2034 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2038 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2039 set
->tags
[i
] = blk_mq_init_rq_map(set
, i
);
2048 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2054 * Allocate the request maps associated with this tag_set. Note that this
2055 * may reduce the depth asked for, if memory is tight. set->queue_depth
2056 * will be updated to reflect the allocated depth.
2058 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2063 depth
= set
->queue_depth
;
2065 err
= __blk_mq_alloc_rq_maps(set
);
2069 set
->queue_depth
>>= 1;
2070 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2074 } while (set
->queue_depth
);
2076 if (!set
->queue_depth
|| err
) {
2077 pr_err("blk-mq: failed to allocate request map\n");
2081 if (depth
!= set
->queue_depth
)
2082 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2083 depth
, set
->queue_depth
);
2089 * Alloc a tag set to be associated with one or more request queues.
2090 * May fail with EINVAL for various error conditions. May adjust the
2091 * requested depth down, if if it too large. In that case, the set
2092 * value will be stored in set->queue_depth.
2094 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2096 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2098 if (!set
->nr_hw_queues
)
2100 if (!set
->queue_depth
)
2102 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2105 if (!set
->nr_hw_queues
|| !set
->ops
->queue_rq
|| !set
->ops
->map_queue
)
2108 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2109 pr_info("blk-mq: reduced tag depth to %u\n",
2111 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2115 * If a crashdump is active, then we are potentially in a very
2116 * memory constrained environment. Limit us to 1 queue and
2117 * 64 tags to prevent using too much memory.
2119 if (is_kdump_kernel()) {
2120 set
->nr_hw_queues
= 1;
2121 set
->queue_depth
= min(64U, set
->queue_depth
);
2124 set
->tags
= kmalloc_node(set
->nr_hw_queues
*
2125 sizeof(struct blk_mq_tags
*),
2126 GFP_KERNEL
, set
->numa_node
);
2130 if (blk_mq_alloc_rq_maps(set
))
2133 mutex_init(&set
->tag_list_lock
);
2134 INIT_LIST_HEAD(&set
->tag_list
);
2142 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2144 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2148 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2150 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2156 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2158 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2160 struct blk_mq_tag_set
*set
= q
->tag_set
;
2161 struct blk_mq_hw_ctx
*hctx
;
2164 if (!set
|| nr
> set
->queue_depth
)
2168 queue_for_each_hw_ctx(q
, hctx
, i
) {
2169 ret
= blk_mq_tag_update_depth(hctx
->tags
, nr
);
2175 q
->nr_requests
= nr
;
2180 void blk_mq_disable_hotplug(void)
2182 mutex_lock(&all_q_mutex
);
2185 void blk_mq_enable_hotplug(void)
2187 mutex_unlock(&all_q_mutex
);
2190 static int __init
blk_mq_init(void)
2194 hotcpu_notifier(blk_mq_queue_reinit_notify
, 0);
2198 subsys_initcall(blk_mq_init
);