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 static 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);
124 static void blk_mq_freeze_queue_wait(struct request_queue
*q
)
126 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->mq_usage_counter
));
130 * Guarantee no request is in use, so we can change any data structure of
131 * the queue afterward.
133 void blk_mq_freeze_queue(struct request_queue
*q
)
135 blk_mq_freeze_queue_start(q
);
136 blk_mq_freeze_queue_wait(q
);
139 static void blk_mq_unfreeze_queue(struct request_queue
*q
)
143 spin_lock_irq(q
->queue_lock
);
144 wake
= !--q
->mq_freeze_depth
;
145 WARN_ON_ONCE(q
->mq_freeze_depth
< 0);
146 spin_unlock_irq(q
->queue_lock
);
148 percpu_ref_reinit(&q
->mq_usage_counter
);
149 wake_up_all(&q
->mq_freeze_wq
);
153 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
155 return blk_mq_has_free_tags(hctx
->tags
);
157 EXPORT_SYMBOL(blk_mq_can_queue
);
159 static void blk_mq_rq_ctx_init(struct request_queue
*q
, struct blk_mq_ctx
*ctx
,
160 struct request
*rq
, unsigned int rw_flags
)
162 if (blk_queue_io_stat(q
))
163 rw_flags
|= REQ_IO_STAT
;
165 INIT_LIST_HEAD(&rq
->queuelist
);
166 /* csd/requeue_work/fifo_time is initialized before use */
169 rq
->cmd_flags
|= rw_flags
;
170 /* do not touch atomic flags, it needs atomic ops against the timer */
172 INIT_HLIST_NODE(&rq
->hash
);
173 RB_CLEAR_NODE(&rq
->rb_node
);
176 rq
->start_time
= jiffies
;
177 #ifdef CONFIG_BLK_CGROUP
179 set_start_time_ns(rq
);
180 rq
->io_start_time_ns
= 0;
182 rq
->nr_phys_segments
= 0;
183 #if defined(CONFIG_BLK_DEV_INTEGRITY)
184 rq
->nr_integrity_segments
= 0;
187 /* tag was already set */
197 INIT_LIST_HEAD(&rq
->timeout_list
);
201 rq
->end_io_data
= NULL
;
204 ctx
->rq_dispatched
[rw_is_sync(rw_flags
)]++;
207 static struct request
*
208 __blk_mq_alloc_request(struct blk_mq_alloc_data
*data
, int rw
)
213 tag
= blk_mq_get_tag(data
);
214 if (tag
!= BLK_MQ_TAG_FAIL
) {
215 rq
= data
->hctx
->tags
->rqs
[tag
];
217 if (blk_mq_tag_busy(data
->hctx
)) {
218 rq
->cmd_flags
= REQ_MQ_INFLIGHT
;
219 atomic_inc(&data
->hctx
->nr_active
);
223 blk_mq_rq_ctx_init(data
->q
, data
->ctx
, rq
, rw
);
230 struct request
*blk_mq_alloc_request(struct request_queue
*q
, int rw
, gfp_t gfp
,
233 struct blk_mq_ctx
*ctx
;
234 struct blk_mq_hw_ctx
*hctx
;
236 struct blk_mq_alloc_data alloc_data
;
239 ret
= blk_mq_queue_enter(q
);
243 ctx
= blk_mq_get_ctx(q
);
244 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
245 blk_mq_set_alloc_data(&alloc_data
, q
, gfp
& ~__GFP_WAIT
,
246 reserved
, ctx
, hctx
);
248 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
249 if (!rq
&& (gfp
& __GFP_WAIT
)) {
250 __blk_mq_run_hw_queue(hctx
);
253 ctx
= blk_mq_get_ctx(q
);
254 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
255 blk_mq_set_alloc_data(&alloc_data
, q
, gfp
, reserved
, ctx
,
257 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
258 ctx
= alloc_data
.ctx
;
262 return ERR_PTR(-EWOULDBLOCK
);
265 EXPORT_SYMBOL(blk_mq_alloc_request
);
267 static void __blk_mq_free_request(struct blk_mq_hw_ctx
*hctx
,
268 struct blk_mq_ctx
*ctx
, struct request
*rq
)
270 const int tag
= rq
->tag
;
271 struct request_queue
*q
= rq
->q
;
273 if (rq
->cmd_flags
& REQ_MQ_INFLIGHT
)
274 atomic_dec(&hctx
->nr_active
);
277 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
278 blk_mq_put_tag(hctx
, tag
, &ctx
->last_tag
);
279 blk_mq_queue_exit(q
);
282 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
284 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
286 ctx
->rq_completed
[rq_is_sync(rq
)]++;
287 __blk_mq_free_request(hctx
, ctx
, rq
);
290 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request
);
292 void blk_mq_free_request(struct request
*rq
)
294 struct blk_mq_hw_ctx
*hctx
;
295 struct request_queue
*q
= rq
->q
;
297 hctx
= q
->mq_ops
->map_queue(q
, rq
->mq_ctx
->cpu
);
298 blk_mq_free_hctx_request(hctx
, rq
);
300 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
302 inline void __blk_mq_end_request(struct request
*rq
, int error
)
304 blk_account_io_done(rq
);
307 rq
->end_io(rq
, error
);
309 if (unlikely(blk_bidi_rq(rq
)))
310 blk_mq_free_request(rq
->next_rq
);
311 blk_mq_free_request(rq
);
314 EXPORT_SYMBOL(__blk_mq_end_request
);
316 void blk_mq_end_request(struct request
*rq
, int error
)
318 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
320 __blk_mq_end_request(rq
, error
);
322 EXPORT_SYMBOL(blk_mq_end_request
);
324 static void __blk_mq_complete_request_remote(void *data
)
326 struct request
*rq
= data
;
328 rq
->q
->softirq_done_fn(rq
);
331 static void blk_mq_ipi_complete_request(struct request
*rq
)
333 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
337 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
338 rq
->q
->softirq_done_fn(rq
);
343 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
344 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
346 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
347 rq
->csd
.func
= __blk_mq_complete_request_remote
;
350 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
352 rq
->q
->softirq_done_fn(rq
);
357 void __blk_mq_complete_request(struct request
*rq
)
359 struct request_queue
*q
= rq
->q
;
361 if (!q
->softirq_done_fn
)
362 blk_mq_end_request(rq
, rq
->errors
);
364 blk_mq_ipi_complete_request(rq
);
368 * blk_mq_complete_request - end I/O on a request
369 * @rq: the request being processed
372 * Ends all I/O on a request. It does not handle partial completions.
373 * The actual completion happens out-of-order, through a IPI handler.
375 void blk_mq_complete_request(struct request
*rq
)
377 struct request_queue
*q
= rq
->q
;
379 if (unlikely(blk_should_fake_timeout(q
)))
381 if (!blk_mark_rq_complete(rq
))
382 __blk_mq_complete_request(rq
);
384 EXPORT_SYMBOL(blk_mq_complete_request
);
386 void blk_mq_start_request(struct request
*rq
)
388 struct request_queue
*q
= rq
->q
;
390 trace_block_rq_issue(q
, rq
);
392 rq
->resid_len
= blk_rq_bytes(rq
);
393 if (unlikely(blk_bidi_rq(rq
)))
394 rq
->next_rq
->resid_len
= blk_rq_bytes(rq
->next_rq
);
399 * Ensure that ->deadline is visible before set the started
400 * flag and clear the completed flag.
402 smp_mb__before_atomic();
405 * Mark us as started and clear complete. Complete might have been
406 * set if requeue raced with timeout, which then marked it as
407 * complete. So be sure to clear complete again when we start
408 * the request, otherwise we'll ignore the completion event.
410 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
411 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
412 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
413 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
415 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
417 * Make sure space for the drain appears. We know we can do
418 * this because max_hw_segments has been adjusted to be one
419 * fewer than the device can handle.
421 rq
->nr_phys_segments
++;
424 EXPORT_SYMBOL(blk_mq_start_request
);
426 static void __blk_mq_requeue_request(struct request
*rq
)
428 struct request_queue
*q
= rq
->q
;
430 trace_block_rq_requeue(q
, rq
);
432 if (test_and_clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
433 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
434 rq
->nr_phys_segments
--;
438 void blk_mq_requeue_request(struct request
*rq
)
440 __blk_mq_requeue_request(rq
);
442 BUG_ON(blk_queued_rq(rq
));
443 blk_mq_add_to_requeue_list(rq
, true);
445 EXPORT_SYMBOL(blk_mq_requeue_request
);
447 static void blk_mq_requeue_work(struct work_struct
*work
)
449 struct request_queue
*q
=
450 container_of(work
, struct request_queue
, requeue_work
);
452 struct request
*rq
, *next
;
455 spin_lock_irqsave(&q
->requeue_lock
, flags
);
456 list_splice_init(&q
->requeue_list
, &rq_list
);
457 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
459 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
460 if (!(rq
->cmd_flags
& REQ_SOFTBARRIER
))
463 rq
->cmd_flags
&= ~REQ_SOFTBARRIER
;
464 list_del_init(&rq
->queuelist
);
465 blk_mq_insert_request(rq
, true, false, false);
468 while (!list_empty(&rq_list
)) {
469 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
470 list_del_init(&rq
->queuelist
);
471 blk_mq_insert_request(rq
, false, false, false);
475 * Use the start variant of queue running here, so that running
476 * the requeue work will kick stopped queues.
478 blk_mq_start_hw_queues(q
);
481 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
)
483 struct request_queue
*q
= rq
->q
;
487 * We abuse this flag that is otherwise used by the I/O scheduler to
488 * request head insertation from the workqueue.
490 BUG_ON(rq
->cmd_flags
& REQ_SOFTBARRIER
);
492 spin_lock_irqsave(&q
->requeue_lock
, flags
);
494 rq
->cmd_flags
|= REQ_SOFTBARRIER
;
495 list_add(&rq
->queuelist
, &q
->requeue_list
);
497 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
499 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
501 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
503 void blk_mq_kick_requeue_list(struct request_queue
*q
)
505 kblockd_schedule_work(&q
->requeue_work
);
507 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
509 static inline bool is_flush_request(struct request
*rq
,
510 struct blk_flush_queue
*fq
, unsigned int tag
)
512 return ((rq
->cmd_flags
& REQ_FLUSH_SEQ
) &&
513 fq
->flush_rq
->tag
== tag
);
516 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
518 struct request
*rq
= tags
->rqs
[tag
];
519 /* mq_ctx of flush rq is always cloned from the corresponding req */
520 struct blk_flush_queue
*fq
= blk_get_flush_queue(rq
->q
, rq
->mq_ctx
);
522 if (!is_flush_request(rq
, fq
, tag
))
527 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
529 struct blk_mq_timeout_data
{
531 unsigned int next_set
;
534 void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
536 struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
537 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
540 * We know that complete is set at this point. If STARTED isn't set
541 * anymore, then the request isn't active and the "timeout" should
542 * just be ignored. This can happen due to the bitflag ordering.
543 * Timeout first checks if STARTED is set, and if it is, assumes
544 * the request is active. But if we race with completion, then
545 * we both flags will get cleared. So check here again, and ignore
546 * a timeout event with a request that isn't active.
548 if (!test_bit(REQ_ATOM_STARTED
, &req
->atomic_flags
))
552 ret
= ops
->timeout(req
, reserved
);
556 __blk_mq_complete_request(req
);
558 case BLK_EH_RESET_TIMER
:
560 blk_clear_rq_complete(req
);
562 case BLK_EH_NOT_HANDLED
:
565 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
570 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
571 struct request
*rq
, void *priv
, bool reserved
)
573 struct blk_mq_timeout_data
*data
= priv
;
575 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
578 if (time_after_eq(jiffies
, rq
->deadline
)) {
579 if (!blk_mark_rq_complete(rq
))
580 blk_mq_rq_timed_out(rq
, reserved
);
581 } else if (!data
->next_set
|| time_after(data
->next
, rq
->deadline
)) {
582 data
->next
= rq
->deadline
;
587 static void blk_mq_rq_timer(unsigned long priv
)
589 struct request_queue
*q
= (struct request_queue
*)priv
;
590 struct blk_mq_timeout_data data
= {
594 struct blk_mq_hw_ctx
*hctx
;
597 queue_for_each_hw_ctx(q
, hctx
, i
) {
599 * If not software queues are currently mapped to this
600 * hardware queue, there's nothing to check
602 if (!blk_mq_hw_queue_mapped(hctx
))
605 blk_mq_tag_busy_iter(hctx
, blk_mq_check_expired
, &data
);
609 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
610 mod_timer(&q
->timeout
, data
.next
);
612 queue_for_each_hw_ctx(q
, hctx
, i
)
613 blk_mq_tag_idle(hctx
);
618 * Reverse check our software queue for entries that we could potentially
619 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
620 * too much time checking for merges.
622 static bool blk_mq_attempt_merge(struct request_queue
*q
,
623 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
628 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
634 if (!blk_rq_merge_ok(rq
, bio
))
637 el_ret
= blk_try_merge(rq
, bio
);
638 if (el_ret
== ELEVATOR_BACK_MERGE
) {
639 if (bio_attempt_back_merge(q
, rq
, bio
)) {
644 } else if (el_ret
== ELEVATOR_FRONT_MERGE
) {
645 if (bio_attempt_front_merge(q
, rq
, bio
)) {
657 * Process software queues that have been marked busy, splicing them
658 * to the for-dispatch
660 static void flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
662 struct blk_mq_ctx
*ctx
;
665 for (i
= 0; i
< hctx
->ctx_map
.map_size
; i
++) {
666 struct blk_align_bitmap
*bm
= &hctx
->ctx_map
.map
[i
];
667 unsigned int off
, bit
;
673 off
= i
* hctx
->ctx_map
.bits_per_word
;
675 bit
= find_next_bit(&bm
->word
, bm
->depth
, bit
);
676 if (bit
>= bm
->depth
)
679 ctx
= hctx
->ctxs
[bit
+ off
];
680 clear_bit(bit
, &bm
->word
);
681 spin_lock(&ctx
->lock
);
682 list_splice_tail_init(&ctx
->rq_list
, list
);
683 spin_unlock(&ctx
->lock
);
691 * Run this hardware queue, pulling any software queues mapped to it in.
692 * Note that this function currently has various problems around ordering
693 * of IO. In particular, we'd like FIFO behaviour on handling existing
694 * items on the hctx->dispatch list. Ignore that for now.
696 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
698 struct request_queue
*q
= hctx
->queue
;
701 LIST_HEAD(driver_list
);
702 struct list_head
*dptr
;
705 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
));
707 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)))
713 * Touch any software queue that has pending entries.
715 flush_busy_ctxs(hctx
, &rq_list
);
718 * If we have previous entries on our dispatch list, grab them
719 * and stuff them at the front for more fair dispatch.
721 if (!list_empty_careful(&hctx
->dispatch
)) {
722 spin_lock(&hctx
->lock
);
723 if (!list_empty(&hctx
->dispatch
))
724 list_splice_init(&hctx
->dispatch
, &rq_list
);
725 spin_unlock(&hctx
->lock
);
729 * Start off with dptr being NULL, so we start the first request
730 * immediately, even if we have more pending.
735 * Now process all the entries, sending them to the driver.
738 while (!list_empty(&rq_list
)) {
739 struct blk_mq_queue_data bd
;
742 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
743 list_del_init(&rq
->queuelist
);
747 bd
.last
= list_empty(&rq_list
);
749 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
751 case BLK_MQ_RQ_QUEUE_OK
:
754 case BLK_MQ_RQ_QUEUE_BUSY
:
755 list_add(&rq
->queuelist
, &rq_list
);
756 __blk_mq_requeue_request(rq
);
759 pr_err("blk-mq: bad return on queue: %d\n", ret
);
760 case BLK_MQ_RQ_QUEUE_ERROR
:
762 blk_mq_end_request(rq
, rq
->errors
);
766 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
770 * We've done the first request. If we have more than 1
771 * left in the list, set dptr to defer issue.
773 if (!dptr
&& rq_list
.next
!= rq_list
.prev
)
778 hctx
->dispatched
[0]++;
779 else if (queued
< (1 << (BLK_MQ_MAX_DISPATCH_ORDER
- 1)))
780 hctx
->dispatched
[ilog2(queued
) + 1]++;
783 * Any items that need requeuing? Stuff them into hctx->dispatch,
784 * that is where we will continue on next queue run.
786 if (!list_empty(&rq_list
)) {
787 spin_lock(&hctx
->lock
);
788 list_splice(&rq_list
, &hctx
->dispatch
);
789 spin_unlock(&hctx
->lock
);
794 * It'd be great if the workqueue API had a way to pass
795 * in a mask and had some smarts for more clever placement.
796 * For now we just round-robin here, switching for every
797 * BLK_MQ_CPU_WORK_BATCH queued items.
799 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
801 if (hctx
->queue
->nr_hw_queues
== 1)
802 return WORK_CPU_UNBOUND
;
804 if (--hctx
->next_cpu_batch
<= 0) {
805 int cpu
= hctx
->next_cpu
, next_cpu
;
807 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
808 if (next_cpu
>= nr_cpu_ids
)
809 next_cpu
= cpumask_first(hctx
->cpumask
);
811 hctx
->next_cpu
= next_cpu
;
812 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
817 return hctx
->next_cpu
;
820 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
822 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
) ||
823 !blk_mq_hw_queue_mapped(hctx
)))
828 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
829 __blk_mq_run_hw_queue(hctx
);
837 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
841 void blk_mq_run_queues(struct request_queue
*q
, bool async
)
843 struct blk_mq_hw_ctx
*hctx
;
846 queue_for_each_hw_ctx(q
, hctx
, i
) {
847 if ((!blk_mq_hctx_has_pending(hctx
) &&
848 list_empty_careful(&hctx
->dispatch
)) ||
849 test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
852 blk_mq_run_hw_queue(hctx
, async
);
855 EXPORT_SYMBOL(blk_mq_run_queues
);
857 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
859 cancel_delayed_work(&hctx
->run_work
);
860 cancel_delayed_work(&hctx
->delay_work
);
861 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
863 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
865 void blk_mq_stop_hw_queues(struct request_queue
*q
)
867 struct blk_mq_hw_ctx
*hctx
;
870 queue_for_each_hw_ctx(q
, hctx
, i
)
871 blk_mq_stop_hw_queue(hctx
);
873 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
875 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
877 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
879 blk_mq_run_hw_queue(hctx
, false);
881 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
883 void blk_mq_start_hw_queues(struct request_queue
*q
)
885 struct blk_mq_hw_ctx
*hctx
;
888 queue_for_each_hw_ctx(q
, hctx
, i
)
889 blk_mq_start_hw_queue(hctx
);
891 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
894 void blk_mq_start_stopped_hw_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 (!test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
903 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
904 blk_mq_run_hw_queue(hctx
, async
);
907 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
909 static void blk_mq_run_work_fn(struct work_struct
*work
)
911 struct blk_mq_hw_ctx
*hctx
;
913 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
915 __blk_mq_run_hw_queue(hctx
);
918 static void blk_mq_delay_work_fn(struct work_struct
*work
)
920 struct blk_mq_hw_ctx
*hctx
;
922 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
924 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
925 __blk_mq_run_hw_queue(hctx
);
928 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
930 if (unlikely(!blk_mq_hw_queue_mapped(hctx
)))
933 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
934 &hctx
->delay_work
, msecs_to_jiffies(msecs
));
936 EXPORT_SYMBOL(blk_mq_delay_queue
);
938 static void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
,
939 struct request
*rq
, bool at_head
)
941 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
943 trace_block_rq_insert(hctx
->queue
, rq
);
946 list_add(&rq
->queuelist
, &ctx
->rq_list
);
948 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
950 blk_mq_hctx_mark_pending(hctx
, ctx
);
953 void blk_mq_insert_request(struct request
*rq
, bool at_head
, bool run_queue
,
956 struct request_queue
*q
= rq
->q
;
957 struct blk_mq_hw_ctx
*hctx
;
958 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
, *current_ctx
;
960 current_ctx
= blk_mq_get_ctx(q
);
961 if (!cpu_online(ctx
->cpu
))
962 rq
->mq_ctx
= ctx
= current_ctx
;
964 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
966 spin_lock(&ctx
->lock
);
967 __blk_mq_insert_request(hctx
, rq
, at_head
);
968 spin_unlock(&ctx
->lock
);
971 blk_mq_run_hw_queue(hctx
, async
);
973 blk_mq_put_ctx(current_ctx
);
976 static void blk_mq_insert_requests(struct request_queue
*q
,
977 struct blk_mq_ctx
*ctx
,
978 struct list_head
*list
,
983 struct blk_mq_hw_ctx
*hctx
;
984 struct blk_mq_ctx
*current_ctx
;
986 trace_block_unplug(q
, depth
, !from_schedule
);
988 current_ctx
= blk_mq_get_ctx(q
);
990 if (!cpu_online(ctx
->cpu
))
992 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
995 * preemption doesn't flush plug list, so it's possible ctx->cpu is
998 spin_lock(&ctx
->lock
);
999 while (!list_empty(list
)) {
1002 rq
= list_first_entry(list
, struct request
, queuelist
);
1003 list_del_init(&rq
->queuelist
);
1005 __blk_mq_insert_request(hctx
, rq
, false);
1007 spin_unlock(&ctx
->lock
);
1009 blk_mq_run_hw_queue(hctx
, from_schedule
);
1010 blk_mq_put_ctx(current_ctx
);
1013 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1015 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1016 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1018 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1019 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1020 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1023 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1025 struct blk_mq_ctx
*this_ctx
;
1026 struct request_queue
*this_q
;
1029 LIST_HEAD(ctx_list
);
1032 list_splice_init(&plug
->mq_list
, &list
);
1034 list_sort(NULL
, &list
, plug_ctx_cmp
);
1040 while (!list_empty(&list
)) {
1041 rq
= list_entry_rq(list
.next
);
1042 list_del_init(&rq
->queuelist
);
1044 if (rq
->mq_ctx
!= this_ctx
) {
1046 blk_mq_insert_requests(this_q
, this_ctx
,
1051 this_ctx
= rq
->mq_ctx
;
1057 list_add_tail(&rq
->queuelist
, &ctx_list
);
1061 * If 'this_ctx' is set, we know we have entries to complete
1062 * on 'ctx_list'. Do those.
1065 blk_mq_insert_requests(this_q
, this_ctx
, &ctx_list
, depth
,
1070 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1072 init_request_from_bio(rq
, bio
);
1074 if (blk_do_io_stat(rq
))
1075 blk_account_io_start(rq
, 1);
1078 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1080 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1081 !blk_queue_nomerges(hctx
->queue
);
1084 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx
*hctx
,
1085 struct blk_mq_ctx
*ctx
,
1086 struct request
*rq
, struct bio
*bio
)
1088 if (!hctx_allow_merges(hctx
)) {
1089 blk_mq_bio_to_request(rq
, bio
);
1090 spin_lock(&ctx
->lock
);
1092 __blk_mq_insert_request(hctx
, rq
, false);
1093 spin_unlock(&ctx
->lock
);
1096 struct request_queue
*q
= hctx
->queue
;
1098 spin_lock(&ctx
->lock
);
1099 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1100 blk_mq_bio_to_request(rq
, bio
);
1104 spin_unlock(&ctx
->lock
);
1105 __blk_mq_free_request(hctx
, ctx
, rq
);
1110 struct blk_map_ctx
{
1111 struct blk_mq_hw_ctx
*hctx
;
1112 struct blk_mq_ctx
*ctx
;
1115 static struct request
*blk_mq_map_request(struct request_queue
*q
,
1117 struct blk_map_ctx
*data
)
1119 struct blk_mq_hw_ctx
*hctx
;
1120 struct blk_mq_ctx
*ctx
;
1122 int rw
= bio_data_dir(bio
);
1123 struct blk_mq_alloc_data alloc_data
;
1125 if (unlikely(blk_mq_queue_enter(q
))) {
1126 bio_endio(bio
, -EIO
);
1130 ctx
= blk_mq_get_ctx(q
);
1131 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1133 if (rw_is_sync(bio
->bi_rw
))
1136 trace_block_getrq(q
, bio
, rw
);
1137 blk_mq_set_alloc_data(&alloc_data
, q
, GFP_ATOMIC
, false, ctx
,
1139 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
1140 if (unlikely(!rq
)) {
1141 __blk_mq_run_hw_queue(hctx
);
1142 blk_mq_put_ctx(ctx
);
1143 trace_block_sleeprq(q
, bio
, rw
);
1145 ctx
= blk_mq_get_ctx(q
);
1146 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1147 blk_mq_set_alloc_data(&alloc_data
, q
,
1148 __GFP_WAIT
|GFP_ATOMIC
, false, ctx
, hctx
);
1149 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
1150 ctx
= alloc_data
.ctx
;
1151 hctx
= alloc_data
.hctx
;
1161 * Multiple hardware queue variant. This will not use per-process plugs,
1162 * but will attempt to bypass the hctx queueing if we can go straight to
1163 * hardware for SYNC IO.
1165 static void blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1167 const int is_sync
= rw_is_sync(bio
->bi_rw
);
1168 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
1169 struct blk_map_ctx data
;
1172 blk_queue_bounce(q
, &bio
);
1174 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1175 bio_endio(bio
, -EIO
);
1179 rq
= blk_mq_map_request(q
, bio
, &data
);
1183 if (unlikely(is_flush_fua
)) {
1184 blk_mq_bio_to_request(rq
, bio
);
1185 blk_insert_flush(rq
);
1190 * If the driver supports defer issued based on 'last', then
1191 * queue it up like normal since we can potentially save some
1194 if (is_sync
&& !(data
.hctx
->flags
& BLK_MQ_F_DEFER_ISSUE
)) {
1195 struct blk_mq_queue_data bd
= {
1202 blk_mq_bio_to_request(rq
, bio
);
1205 * For OK queue, we are done. For error, kill it. Any other
1206 * error (busy), just add it to our list as we previously
1209 ret
= q
->mq_ops
->queue_rq(data
.hctx
, &bd
);
1210 if (ret
== BLK_MQ_RQ_QUEUE_OK
)
1213 __blk_mq_requeue_request(rq
);
1215 if (ret
== BLK_MQ_RQ_QUEUE_ERROR
) {
1217 blk_mq_end_request(rq
, rq
->errors
);
1223 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1225 * For a SYNC request, send it to the hardware immediately. For
1226 * an ASYNC request, just ensure that we run it later on. The
1227 * latter allows for merging opportunities and more efficient
1231 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1234 blk_mq_put_ctx(data
.ctx
);
1238 * Single hardware queue variant. This will attempt to use any per-process
1239 * plug for merging and IO deferral.
1241 static void blk_sq_make_request(struct request_queue
*q
, struct bio
*bio
)
1243 const int is_sync
= rw_is_sync(bio
->bi_rw
);
1244 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
1245 unsigned int use_plug
, request_count
= 0;
1246 struct blk_map_ctx data
;
1250 * If we have multiple hardware queues, just go directly to
1251 * one of those for sync IO.
1253 use_plug
= !is_flush_fua
&& !is_sync
;
1255 blk_queue_bounce(q
, &bio
);
1257 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1258 bio_endio(bio
, -EIO
);
1262 if (use_plug
&& !blk_queue_nomerges(q
) &&
1263 blk_attempt_plug_merge(q
, bio
, &request_count
))
1266 rq
= blk_mq_map_request(q
, bio
, &data
);
1270 if (unlikely(is_flush_fua
)) {
1271 blk_mq_bio_to_request(rq
, bio
);
1272 blk_insert_flush(rq
);
1277 * A task plug currently exists. Since this is completely lockless,
1278 * utilize that to temporarily store requests until the task is
1279 * either done or scheduled away.
1282 struct blk_plug
*plug
= current
->plug
;
1285 blk_mq_bio_to_request(rq
, bio
);
1286 if (list_empty(&plug
->mq_list
))
1287 trace_block_plug(q
);
1288 else if (request_count
>= BLK_MAX_REQUEST_COUNT
) {
1289 blk_flush_plug_list(plug
, false);
1290 trace_block_plug(q
);
1292 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1293 blk_mq_put_ctx(data
.ctx
);
1298 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1300 * For a SYNC request, send it to the hardware immediately. For
1301 * an ASYNC request, just ensure that we run it later on. The
1302 * latter allows for merging opportunities and more efficient
1306 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1309 blk_mq_put_ctx(data
.ctx
);
1313 * Default mapping to a software queue, since we use one per CPU.
1315 struct blk_mq_hw_ctx
*blk_mq_map_queue(struct request_queue
*q
, const int cpu
)
1317 return q
->queue_hw_ctx
[q
->mq_map
[cpu
]];
1319 EXPORT_SYMBOL(blk_mq_map_queue
);
1321 static void blk_mq_free_rq_map(struct blk_mq_tag_set
*set
,
1322 struct blk_mq_tags
*tags
, unsigned int hctx_idx
)
1326 if (tags
->rqs
&& set
->ops
->exit_request
) {
1329 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1332 set
->ops
->exit_request(set
->driver_data
, tags
->rqs
[i
],
1334 tags
->rqs
[i
] = NULL
;
1338 while (!list_empty(&tags
->page_list
)) {
1339 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1340 list_del_init(&page
->lru
);
1341 __free_pages(page
, page
->private);
1346 blk_mq_free_tags(tags
);
1349 static size_t order_to_size(unsigned int order
)
1351 return (size_t)PAGE_SIZE
<< order
;
1354 static struct blk_mq_tags
*blk_mq_init_rq_map(struct blk_mq_tag_set
*set
,
1355 unsigned int hctx_idx
)
1357 struct blk_mq_tags
*tags
;
1358 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1359 size_t rq_size
, left
;
1361 tags
= blk_mq_init_tags(set
->queue_depth
, set
->reserved_tags
,
1366 INIT_LIST_HEAD(&tags
->page_list
);
1368 tags
->rqs
= kzalloc_node(set
->queue_depth
* sizeof(struct request
*),
1369 GFP_KERNEL
| __GFP_NOWARN
| __GFP_NORETRY
,
1372 blk_mq_free_tags(tags
);
1377 * rq_size is the size of the request plus driver payload, rounded
1378 * to the cacheline size
1380 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1382 left
= rq_size
* set
->queue_depth
;
1384 for (i
= 0; i
< set
->queue_depth
; ) {
1385 int this_order
= max_order
;
1390 while (left
< order_to_size(this_order
- 1) && this_order
)
1394 page
= alloc_pages_node(set
->numa_node
,
1395 GFP_KERNEL
| __GFP_NOWARN
| __GFP_NORETRY
,
1401 if (order_to_size(this_order
) < rq_size
)
1408 page
->private = this_order
;
1409 list_add_tail(&page
->lru
, &tags
->page_list
);
1411 p
= page_address(page
);
1412 entries_per_page
= order_to_size(this_order
) / rq_size
;
1413 to_do
= min(entries_per_page
, set
->queue_depth
- i
);
1414 left
-= to_do
* rq_size
;
1415 for (j
= 0; j
< to_do
; j
++) {
1417 tags
->rqs
[i
]->atomic_flags
= 0;
1418 tags
->rqs
[i
]->cmd_flags
= 0;
1419 if (set
->ops
->init_request
) {
1420 if (set
->ops
->init_request(set
->driver_data
,
1421 tags
->rqs
[i
], hctx_idx
, i
,
1423 tags
->rqs
[i
] = NULL
;
1436 blk_mq_free_rq_map(set
, tags
, hctx_idx
);
1440 static void blk_mq_free_bitmap(struct blk_mq_ctxmap
*bitmap
)
1445 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap
*bitmap
, int node
)
1447 unsigned int bpw
= 8, total
, num_maps
, i
;
1449 bitmap
->bits_per_word
= bpw
;
1451 num_maps
= ALIGN(nr_cpu_ids
, bpw
) / bpw
;
1452 bitmap
->map
= kzalloc_node(num_maps
* sizeof(struct blk_align_bitmap
),
1457 bitmap
->map_size
= num_maps
;
1460 for (i
= 0; i
< num_maps
; i
++) {
1461 bitmap
->map
[i
].depth
= min(total
, bitmap
->bits_per_word
);
1462 total
-= bitmap
->map
[i
].depth
;
1468 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx
*hctx
, int cpu
)
1470 struct request_queue
*q
= hctx
->queue
;
1471 struct blk_mq_ctx
*ctx
;
1475 * Move ctx entries to new CPU, if this one is going away.
1477 ctx
= __blk_mq_get_ctx(q
, cpu
);
1479 spin_lock(&ctx
->lock
);
1480 if (!list_empty(&ctx
->rq_list
)) {
1481 list_splice_init(&ctx
->rq_list
, &tmp
);
1482 blk_mq_hctx_clear_pending(hctx
, ctx
);
1484 spin_unlock(&ctx
->lock
);
1486 if (list_empty(&tmp
))
1489 ctx
= blk_mq_get_ctx(q
);
1490 spin_lock(&ctx
->lock
);
1492 while (!list_empty(&tmp
)) {
1495 rq
= list_first_entry(&tmp
, struct request
, queuelist
);
1497 list_move_tail(&rq
->queuelist
, &ctx
->rq_list
);
1500 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1501 blk_mq_hctx_mark_pending(hctx
, ctx
);
1503 spin_unlock(&ctx
->lock
);
1505 blk_mq_run_hw_queue(hctx
, true);
1506 blk_mq_put_ctx(ctx
);
1510 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx
*hctx
, int cpu
)
1512 struct request_queue
*q
= hctx
->queue
;
1513 struct blk_mq_tag_set
*set
= q
->tag_set
;
1515 if (set
->tags
[hctx
->queue_num
])
1518 set
->tags
[hctx
->queue_num
] = blk_mq_init_rq_map(set
, hctx
->queue_num
);
1519 if (!set
->tags
[hctx
->queue_num
])
1522 hctx
->tags
= set
->tags
[hctx
->queue_num
];
1526 static int blk_mq_hctx_notify(void *data
, unsigned long action
,
1529 struct blk_mq_hw_ctx
*hctx
= data
;
1531 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
)
1532 return blk_mq_hctx_cpu_offline(hctx
, cpu
);
1533 else if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
)
1534 return blk_mq_hctx_cpu_online(hctx
, cpu
);
1539 static void blk_mq_exit_hctx(struct request_queue
*q
,
1540 struct blk_mq_tag_set
*set
,
1541 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
1543 unsigned flush_start_tag
= set
->queue_depth
;
1545 blk_mq_tag_idle(hctx
);
1547 if (set
->ops
->exit_request
)
1548 set
->ops
->exit_request(set
->driver_data
,
1549 hctx
->fq
->flush_rq
, hctx_idx
,
1550 flush_start_tag
+ hctx_idx
);
1552 if (set
->ops
->exit_hctx
)
1553 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1555 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1556 blk_free_flush_queue(hctx
->fq
);
1558 blk_mq_free_bitmap(&hctx
->ctx_map
);
1561 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1562 struct blk_mq_tag_set
*set
, int nr_queue
)
1564 struct blk_mq_hw_ctx
*hctx
;
1567 queue_for_each_hw_ctx(q
, hctx
, i
) {
1570 blk_mq_exit_hctx(q
, set
, hctx
, i
);
1574 static void blk_mq_free_hw_queues(struct request_queue
*q
,
1575 struct blk_mq_tag_set
*set
)
1577 struct blk_mq_hw_ctx
*hctx
;
1580 queue_for_each_hw_ctx(q
, hctx
, i
) {
1581 free_cpumask_var(hctx
->cpumask
);
1586 static int blk_mq_init_hctx(struct request_queue
*q
,
1587 struct blk_mq_tag_set
*set
,
1588 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
1591 unsigned flush_start_tag
= set
->queue_depth
;
1593 node
= hctx
->numa_node
;
1594 if (node
== NUMA_NO_NODE
)
1595 node
= hctx
->numa_node
= set
->numa_node
;
1597 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1598 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1599 spin_lock_init(&hctx
->lock
);
1600 INIT_LIST_HEAD(&hctx
->dispatch
);
1602 hctx
->queue_num
= hctx_idx
;
1603 hctx
->flags
= set
->flags
;
1604 hctx
->cmd_size
= set
->cmd_size
;
1606 blk_mq_init_cpu_notifier(&hctx
->cpu_notifier
,
1607 blk_mq_hctx_notify
, hctx
);
1608 blk_mq_register_cpu_notifier(&hctx
->cpu_notifier
);
1610 hctx
->tags
= set
->tags
[hctx_idx
];
1613 * Allocate space for all possible cpus to avoid allocation at
1616 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1619 goto unregister_cpu_notifier
;
1621 if (blk_mq_alloc_bitmap(&hctx
->ctx_map
, node
))
1626 if (set
->ops
->init_hctx
&&
1627 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
1630 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
1634 if (set
->ops
->init_request
&&
1635 set
->ops
->init_request(set
->driver_data
,
1636 hctx
->fq
->flush_rq
, hctx_idx
,
1637 flush_start_tag
+ hctx_idx
, node
))
1645 if (set
->ops
->exit_hctx
)
1646 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1648 blk_mq_free_bitmap(&hctx
->ctx_map
);
1651 unregister_cpu_notifier
:
1652 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1657 static int blk_mq_init_hw_queues(struct request_queue
*q
,
1658 struct blk_mq_tag_set
*set
)
1660 struct blk_mq_hw_ctx
*hctx
;
1664 * Initialize hardware queues
1666 queue_for_each_hw_ctx(q
, hctx
, i
) {
1667 if (blk_mq_init_hctx(q
, set
, hctx
, i
))
1671 if (i
== q
->nr_hw_queues
)
1677 blk_mq_exit_hw_queues(q
, set
, i
);
1682 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1683 unsigned int nr_hw_queues
)
1687 for_each_possible_cpu(i
) {
1688 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1689 struct blk_mq_hw_ctx
*hctx
;
1691 memset(__ctx
, 0, sizeof(*__ctx
));
1693 spin_lock_init(&__ctx
->lock
);
1694 INIT_LIST_HEAD(&__ctx
->rq_list
);
1697 /* If the cpu isn't online, the cpu is mapped to first hctx */
1701 hctx
= q
->mq_ops
->map_queue(q
, i
);
1702 cpumask_set_cpu(i
, hctx
->cpumask
);
1706 * Set local node, IFF we have more than one hw queue. If
1707 * not, we remain on the home node of the device
1709 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1710 hctx
->numa_node
= cpu_to_node(i
);
1714 static void blk_mq_map_swqueue(struct request_queue
*q
)
1717 struct blk_mq_hw_ctx
*hctx
;
1718 struct blk_mq_ctx
*ctx
;
1720 queue_for_each_hw_ctx(q
, hctx
, i
) {
1721 cpumask_clear(hctx
->cpumask
);
1726 * Map software to hardware queues
1728 queue_for_each_ctx(q
, ctx
, i
) {
1729 /* If the cpu isn't online, the cpu is mapped to first hctx */
1733 hctx
= q
->mq_ops
->map_queue(q
, i
);
1734 cpumask_set_cpu(i
, hctx
->cpumask
);
1735 ctx
->index_hw
= hctx
->nr_ctx
;
1736 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
1739 queue_for_each_hw_ctx(q
, hctx
, i
) {
1741 * If no software queues are mapped to this hardware queue,
1742 * disable it and free the request entries.
1744 if (!hctx
->nr_ctx
) {
1745 struct blk_mq_tag_set
*set
= q
->tag_set
;
1748 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1749 set
->tags
[i
] = NULL
;
1756 * Initialize batch roundrobin counts
1758 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
1759 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1763 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
)
1765 struct blk_mq_hw_ctx
*hctx
;
1766 struct request_queue
*q
;
1770 if (set
->tag_list
.next
== set
->tag_list
.prev
)
1775 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
1776 blk_mq_freeze_queue(q
);
1778 queue_for_each_hw_ctx(q
, hctx
, i
) {
1780 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
1782 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
1784 blk_mq_unfreeze_queue(q
);
1788 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
1790 struct blk_mq_tag_set
*set
= q
->tag_set
;
1792 mutex_lock(&set
->tag_list_lock
);
1793 list_del_init(&q
->tag_set_list
);
1794 blk_mq_update_tag_set_depth(set
);
1795 mutex_unlock(&set
->tag_list_lock
);
1798 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
1799 struct request_queue
*q
)
1803 mutex_lock(&set
->tag_list_lock
);
1804 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
1805 blk_mq_update_tag_set_depth(set
);
1806 mutex_unlock(&set
->tag_list_lock
);
1809 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
1811 struct blk_mq_hw_ctx
**hctxs
;
1812 struct blk_mq_ctx __percpu
*ctx
;
1813 struct request_queue
*q
;
1817 ctx
= alloc_percpu(struct blk_mq_ctx
);
1819 return ERR_PTR(-ENOMEM
);
1821 hctxs
= kmalloc_node(set
->nr_hw_queues
* sizeof(*hctxs
), GFP_KERNEL
,
1827 map
= blk_mq_make_queue_map(set
);
1831 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1832 int node
= blk_mq_hw_queue_to_node(map
, i
);
1834 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
1839 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
1843 atomic_set(&hctxs
[i
]->nr_active
, 0);
1844 hctxs
[i
]->numa_node
= node
;
1845 hctxs
[i
]->queue_num
= i
;
1848 q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
1853 * Init percpu_ref in atomic mode so that it's faster to shutdown.
1854 * See blk_register_queue() for details.
1856 if (percpu_ref_init(&q
->mq_usage_counter
, blk_mq_usage_counter_release
,
1857 PERCPU_REF_INIT_ATOMIC
, GFP_KERNEL
))
1860 setup_timer(&q
->timeout
, blk_mq_rq_timer
, (unsigned long) q
);
1861 blk_queue_rq_timeout(q
, 30000);
1863 q
->nr_queues
= nr_cpu_ids
;
1864 q
->nr_hw_queues
= set
->nr_hw_queues
;
1868 q
->queue_hw_ctx
= hctxs
;
1870 q
->mq_ops
= set
->ops
;
1871 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
1873 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
1874 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
1876 q
->sg_reserved_size
= INT_MAX
;
1878 INIT_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
1879 INIT_LIST_HEAD(&q
->requeue_list
);
1880 spin_lock_init(&q
->requeue_lock
);
1882 if (q
->nr_hw_queues
> 1)
1883 blk_queue_make_request(q
, blk_mq_make_request
);
1885 blk_queue_make_request(q
, blk_sq_make_request
);
1888 blk_queue_rq_timeout(q
, set
->timeout
);
1891 * Do this after blk_queue_make_request() overrides it...
1893 q
->nr_requests
= set
->queue_depth
;
1895 if (set
->ops
->complete
)
1896 blk_queue_softirq_done(q
, set
->ops
->complete
);
1898 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
1900 if (blk_mq_init_hw_queues(q
, set
))
1903 mutex_lock(&all_q_mutex
);
1904 list_add_tail(&q
->all_q_node
, &all_q_list
);
1905 mutex_unlock(&all_q_mutex
);
1907 blk_mq_add_queue_tag_set(set
, q
);
1909 blk_mq_map_swqueue(q
);
1914 blk_cleanup_queue(q
);
1917 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1920 free_cpumask_var(hctxs
[i
]->cpumask
);
1927 return ERR_PTR(-ENOMEM
);
1929 EXPORT_SYMBOL(blk_mq_init_queue
);
1931 void blk_mq_free_queue(struct request_queue
*q
)
1933 struct blk_mq_tag_set
*set
= q
->tag_set
;
1935 blk_mq_del_queue_tag_set(q
);
1937 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
1938 blk_mq_free_hw_queues(q
, set
);
1940 percpu_ref_exit(&q
->mq_usage_counter
);
1942 free_percpu(q
->queue_ctx
);
1943 kfree(q
->queue_hw_ctx
);
1946 q
->queue_ctx
= NULL
;
1947 q
->queue_hw_ctx
= NULL
;
1950 mutex_lock(&all_q_mutex
);
1951 list_del_init(&q
->all_q_node
);
1952 mutex_unlock(&all_q_mutex
);
1955 /* Basically redo blk_mq_init_queue with queue frozen */
1956 static void blk_mq_queue_reinit(struct request_queue
*q
)
1958 WARN_ON_ONCE(!q
->mq_freeze_depth
);
1960 blk_mq_sysfs_unregister(q
);
1962 blk_mq_update_queue_map(q
->mq_map
, q
->nr_hw_queues
);
1965 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1966 * we should change hctx numa_node according to new topology (this
1967 * involves free and re-allocate memory, worthy doing?)
1970 blk_mq_map_swqueue(q
);
1972 blk_mq_sysfs_register(q
);
1975 static int blk_mq_queue_reinit_notify(struct notifier_block
*nb
,
1976 unsigned long action
, void *hcpu
)
1978 struct request_queue
*q
;
1981 * Before new mappings are established, hotadded cpu might already
1982 * start handling requests. This doesn't break anything as we map
1983 * offline CPUs to first hardware queue. We will re-init the queue
1984 * below to get optimal settings.
1986 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
&&
1987 action
!= CPU_ONLINE
&& action
!= CPU_ONLINE_FROZEN
)
1990 mutex_lock(&all_q_mutex
);
1993 * We need to freeze and reinit all existing queues. Freezing
1994 * involves synchronous wait for an RCU grace period and doing it
1995 * one by one may take a long time. Start freezing all queues in
1996 * one swoop and then wait for the completions so that freezing can
1997 * take place in parallel.
1999 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2000 blk_mq_freeze_queue_start(q
);
2001 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2002 blk_mq_freeze_queue_wait(q
);
2004 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2005 blk_mq_queue_reinit(q
);
2007 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2008 blk_mq_unfreeze_queue(q
);
2010 mutex_unlock(&all_q_mutex
);
2014 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2018 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2019 set
->tags
[i
] = blk_mq_init_rq_map(set
, i
);
2028 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2034 * Allocate the request maps associated with this tag_set. Note that this
2035 * may reduce the depth asked for, if memory is tight. set->queue_depth
2036 * will be updated to reflect the allocated depth.
2038 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2043 depth
= set
->queue_depth
;
2045 err
= __blk_mq_alloc_rq_maps(set
);
2049 set
->queue_depth
>>= 1;
2050 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2054 } while (set
->queue_depth
);
2056 if (!set
->queue_depth
|| err
) {
2057 pr_err("blk-mq: failed to allocate request map\n");
2061 if (depth
!= set
->queue_depth
)
2062 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2063 depth
, set
->queue_depth
);
2069 * Alloc a tag set to be associated with one or more request queues.
2070 * May fail with EINVAL for various error conditions. May adjust the
2071 * requested depth down, if if it too large. In that case, the set
2072 * value will be stored in set->queue_depth.
2074 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2076 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2078 if (!set
->nr_hw_queues
)
2080 if (!set
->queue_depth
)
2082 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2085 if (!set
->nr_hw_queues
|| !set
->ops
->queue_rq
|| !set
->ops
->map_queue
)
2088 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2089 pr_info("blk-mq: reduced tag depth to %u\n",
2091 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2095 * If a crashdump is active, then we are potentially in a very
2096 * memory constrained environment. Limit us to 1 queue and
2097 * 64 tags to prevent using too much memory.
2099 if (is_kdump_kernel()) {
2100 set
->nr_hw_queues
= 1;
2101 set
->queue_depth
= min(64U, set
->queue_depth
);
2104 set
->tags
= kmalloc_node(set
->nr_hw_queues
*
2105 sizeof(struct blk_mq_tags
*),
2106 GFP_KERNEL
, set
->numa_node
);
2110 if (blk_mq_alloc_rq_maps(set
))
2113 mutex_init(&set
->tag_list_lock
);
2114 INIT_LIST_HEAD(&set
->tag_list
);
2122 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2124 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2128 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2130 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2136 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2138 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2140 struct blk_mq_tag_set
*set
= q
->tag_set
;
2141 struct blk_mq_hw_ctx
*hctx
;
2144 if (!set
|| nr
> set
->queue_depth
)
2148 queue_for_each_hw_ctx(q
, hctx
, i
) {
2149 ret
= blk_mq_tag_update_depth(hctx
->tags
, nr
);
2155 q
->nr_requests
= nr
;
2160 void blk_mq_disable_hotplug(void)
2162 mutex_lock(&all_q_mutex
);
2165 void blk_mq_enable_hotplug(void)
2167 mutex_unlock(&all_q_mutex
);
2170 static int __init
blk_mq_init(void)
2174 hotcpu_notifier(blk_mq_queue_reinit_notify
, 0);
2178 subsys_initcall(blk_mq_init
);