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
);
139 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
141 void blk_mq_unfreeze_queue(struct request_queue
*q
)
145 spin_lock_irq(q
->queue_lock
);
146 wake
= !--q
->mq_freeze_depth
;
147 WARN_ON_ONCE(q
->mq_freeze_depth
< 0);
148 spin_unlock_irq(q
->queue_lock
);
150 percpu_ref_reinit(&q
->mq_usage_counter
);
151 wake_up_all(&q
->mq_freeze_wq
);
154 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
156 void blk_mq_wake_waiters(struct request_queue
*q
)
158 struct blk_mq_hw_ctx
*hctx
;
161 queue_for_each_hw_ctx(q
, hctx
, i
)
162 if (blk_mq_hw_queue_mapped(hctx
))
163 blk_mq_tag_wakeup_all(hctx
->tags
, true);
166 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
168 return blk_mq_has_free_tags(hctx
->tags
);
170 EXPORT_SYMBOL(blk_mq_can_queue
);
172 static void blk_mq_rq_ctx_init(struct request_queue
*q
, struct blk_mq_ctx
*ctx
,
173 struct request
*rq
, unsigned int rw_flags
)
175 if (blk_queue_io_stat(q
))
176 rw_flags
|= REQ_IO_STAT
;
178 INIT_LIST_HEAD(&rq
->queuelist
);
179 /* csd/requeue_work/fifo_time is initialized before use */
182 rq
->cmd_flags
|= rw_flags
;
183 /* do not touch atomic flags, it needs atomic ops against the timer */
185 INIT_HLIST_NODE(&rq
->hash
);
186 RB_CLEAR_NODE(&rq
->rb_node
);
189 rq
->start_time
= jiffies
;
190 #ifdef CONFIG_BLK_CGROUP
192 set_start_time_ns(rq
);
193 rq
->io_start_time_ns
= 0;
195 rq
->nr_phys_segments
= 0;
196 #if defined(CONFIG_BLK_DEV_INTEGRITY)
197 rq
->nr_integrity_segments
= 0;
200 /* tag was already set */
210 INIT_LIST_HEAD(&rq
->timeout_list
);
214 rq
->end_io_data
= NULL
;
217 ctx
->rq_dispatched
[rw_is_sync(rw_flags
)]++;
220 static struct request
*
221 __blk_mq_alloc_request(struct blk_mq_alloc_data
*data
, int rw
)
226 tag
= blk_mq_get_tag(data
);
227 if (tag
!= BLK_MQ_TAG_FAIL
) {
228 rq
= data
->hctx
->tags
->rqs
[tag
];
230 if (blk_mq_tag_busy(data
->hctx
)) {
231 rq
->cmd_flags
= REQ_MQ_INFLIGHT
;
232 atomic_inc(&data
->hctx
->nr_active
);
236 blk_mq_rq_ctx_init(data
->q
, data
->ctx
, rq
, rw
);
243 struct request
*blk_mq_alloc_request(struct request_queue
*q
, int rw
, gfp_t gfp
,
246 struct blk_mq_ctx
*ctx
;
247 struct blk_mq_hw_ctx
*hctx
;
249 struct blk_mq_alloc_data alloc_data
;
252 ret
= blk_mq_queue_enter(q
);
256 ctx
= blk_mq_get_ctx(q
);
257 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
258 blk_mq_set_alloc_data(&alloc_data
, q
, gfp
& ~__GFP_WAIT
,
259 reserved
, ctx
, hctx
);
261 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
262 if (!rq
&& (gfp
& __GFP_WAIT
)) {
263 __blk_mq_run_hw_queue(hctx
);
266 ctx
= blk_mq_get_ctx(q
);
267 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
268 blk_mq_set_alloc_data(&alloc_data
, q
, gfp
, reserved
, ctx
,
270 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
271 ctx
= alloc_data
.ctx
;
275 blk_mq_queue_exit(q
);
276 return ERR_PTR(-EWOULDBLOCK
);
280 EXPORT_SYMBOL(blk_mq_alloc_request
);
282 static void __blk_mq_free_request(struct blk_mq_hw_ctx
*hctx
,
283 struct blk_mq_ctx
*ctx
, struct request
*rq
)
285 const int tag
= rq
->tag
;
286 struct request_queue
*q
= rq
->q
;
288 if (rq
->cmd_flags
& REQ_MQ_INFLIGHT
)
289 atomic_dec(&hctx
->nr_active
);
292 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
293 blk_mq_put_tag(hctx
, tag
, &ctx
->last_tag
);
294 blk_mq_queue_exit(q
);
297 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
299 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
301 ctx
->rq_completed
[rq_is_sync(rq
)]++;
302 __blk_mq_free_request(hctx
, ctx
, rq
);
305 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request
);
307 void blk_mq_free_request(struct request
*rq
)
309 struct blk_mq_hw_ctx
*hctx
;
310 struct request_queue
*q
= rq
->q
;
312 hctx
= q
->mq_ops
->map_queue(q
, rq
->mq_ctx
->cpu
);
313 blk_mq_free_hctx_request(hctx
, rq
);
315 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
317 inline void __blk_mq_end_request(struct request
*rq
, int error
)
319 blk_account_io_done(rq
);
322 rq
->end_io(rq
, error
);
324 if (unlikely(blk_bidi_rq(rq
)))
325 blk_mq_free_request(rq
->next_rq
);
326 blk_mq_free_request(rq
);
329 EXPORT_SYMBOL(__blk_mq_end_request
);
331 void blk_mq_end_request(struct request
*rq
, int error
)
333 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
335 __blk_mq_end_request(rq
, error
);
337 EXPORT_SYMBOL(blk_mq_end_request
);
339 static void __blk_mq_complete_request_remote(void *data
)
341 struct request
*rq
= data
;
343 rq
->q
->softirq_done_fn(rq
);
346 static void blk_mq_ipi_complete_request(struct request
*rq
)
348 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
352 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
353 rq
->q
->softirq_done_fn(rq
);
358 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
359 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
361 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
362 rq
->csd
.func
= __blk_mq_complete_request_remote
;
365 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
367 rq
->q
->softirq_done_fn(rq
);
372 void __blk_mq_complete_request(struct request
*rq
)
374 struct request_queue
*q
= rq
->q
;
376 if (!q
->softirq_done_fn
)
377 blk_mq_end_request(rq
, rq
->errors
);
379 blk_mq_ipi_complete_request(rq
);
383 * blk_mq_complete_request - end I/O on a request
384 * @rq: the request being processed
387 * Ends all I/O on a request. It does not handle partial completions.
388 * The actual completion happens out-of-order, through a IPI handler.
390 void blk_mq_complete_request(struct request
*rq
)
392 struct request_queue
*q
= rq
->q
;
394 if (unlikely(blk_should_fake_timeout(q
)))
396 if (!blk_mark_rq_complete(rq
))
397 __blk_mq_complete_request(rq
);
399 EXPORT_SYMBOL(blk_mq_complete_request
);
401 void blk_mq_start_request(struct request
*rq
)
403 struct request_queue
*q
= rq
->q
;
405 trace_block_rq_issue(q
, rq
);
407 rq
->resid_len
= blk_rq_bytes(rq
);
408 if (unlikely(blk_bidi_rq(rq
)))
409 rq
->next_rq
->resid_len
= blk_rq_bytes(rq
->next_rq
);
414 * Ensure that ->deadline is visible before set the started
415 * flag and clear the completed flag.
417 smp_mb__before_atomic();
420 * Mark us as started and clear complete. Complete might have been
421 * set if requeue raced with timeout, which then marked it as
422 * complete. So be sure to clear complete again when we start
423 * the request, otherwise we'll ignore the completion event.
425 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
426 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
427 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
428 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
430 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
432 * Make sure space for the drain appears. We know we can do
433 * this because max_hw_segments has been adjusted to be one
434 * fewer than the device can handle.
436 rq
->nr_phys_segments
++;
439 EXPORT_SYMBOL(blk_mq_start_request
);
441 static void __blk_mq_requeue_request(struct request
*rq
)
443 struct request_queue
*q
= rq
->q
;
445 trace_block_rq_requeue(q
, rq
);
447 if (test_and_clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
448 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
449 rq
->nr_phys_segments
--;
453 void blk_mq_requeue_request(struct request
*rq
)
455 __blk_mq_requeue_request(rq
);
457 BUG_ON(blk_queued_rq(rq
));
458 blk_mq_add_to_requeue_list(rq
, true);
460 EXPORT_SYMBOL(blk_mq_requeue_request
);
462 static void blk_mq_requeue_work(struct work_struct
*work
)
464 struct request_queue
*q
=
465 container_of(work
, struct request_queue
, requeue_work
);
467 struct request
*rq
, *next
;
470 spin_lock_irqsave(&q
->requeue_lock
, flags
);
471 list_splice_init(&q
->requeue_list
, &rq_list
);
472 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
474 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
475 if (!(rq
->cmd_flags
& REQ_SOFTBARRIER
))
478 rq
->cmd_flags
&= ~REQ_SOFTBARRIER
;
479 list_del_init(&rq
->queuelist
);
480 blk_mq_insert_request(rq
, true, false, false);
483 while (!list_empty(&rq_list
)) {
484 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
485 list_del_init(&rq
->queuelist
);
486 blk_mq_insert_request(rq
, false, false, false);
490 * Use the start variant of queue running here, so that running
491 * the requeue work will kick stopped queues.
493 blk_mq_start_hw_queues(q
);
496 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
)
498 struct request_queue
*q
= rq
->q
;
502 * We abuse this flag that is otherwise used by the I/O scheduler to
503 * request head insertation from the workqueue.
505 BUG_ON(rq
->cmd_flags
& REQ_SOFTBARRIER
);
507 spin_lock_irqsave(&q
->requeue_lock
, flags
);
509 rq
->cmd_flags
|= REQ_SOFTBARRIER
;
510 list_add(&rq
->queuelist
, &q
->requeue_list
);
512 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
514 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
516 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
518 void blk_mq_kick_requeue_list(struct request_queue
*q
)
520 kblockd_schedule_work(&q
->requeue_work
);
522 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
524 static inline bool is_flush_request(struct request
*rq
,
525 struct blk_flush_queue
*fq
, unsigned int tag
)
527 return ((rq
->cmd_flags
& REQ_FLUSH_SEQ
) &&
528 fq
->flush_rq
->tag
== tag
);
531 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
533 struct request
*rq
= tags
->rqs
[tag
];
534 /* mq_ctx of flush rq is always cloned from the corresponding req */
535 struct blk_flush_queue
*fq
= blk_get_flush_queue(rq
->q
, rq
->mq_ctx
);
537 if (!is_flush_request(rq
, fq
, tag
))
542 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
544 struct blk_mq_timeout_data
{
546 unsigned int next_set
;
549 void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
551 struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
552 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
555 * We know that complete is set at this point. If STARTED isn't set
556 * anymore, then the request isn't active and the "timeout" should
557 * just be ignored. This can happen due to the bitflag ordering.
558 * Timeout first checks if STARTED is set, and if it is, assumes
559 * the request is active. But if we race with completion, then
560 * we both flags will get cleared. So check here again, and ignore
561 * a timeout event with a request that isn't active.
563 if (!test_bit(REQ_ATOM_STARTED
, &req
->atomic_flags
))
567 ret
= ops
->timeout(req
, reserved
);
571 __blk_mq_complete_request(req
);
573 case BLK_EH_RESET_TIMER
:
575 blk_clear_rq_complete(req
);
577 case BLK_EH_NOT_HANDLED
:
580 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
585 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
586 struct request
*rq
, void *priv
, bool reserved
)
588 struct blk_mq_timeout_data
*data
= priv
;
590 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
593 if (time_after_eq(jiffies
, rq
->deadline
)) {
594 if (!blk_mark_rq_complete(rq
))
595 blk_mq_rq_timed_out(rq
, reserved
);
596 } else if (!data
->next_set
|| time_after(data
->next
, rq
->deadline
)) {
597 data
->next
= rq
->deadline
;
602 static void blk_mq_rq_timer(unsigned long priv
)
604 struct request_queue
*q
= (struct request_queue
*)priv
;
605 struct blk_mq_timeout_data data
= {
609 struct blk_mq_hw_ctx
*hctx
;
612 queue_for_each_hw_ctx(q
, hctx
, i
) {
614 * If not software queues are currently mapped to this
615 * hardware queue, there's nothing to check
617 if (!blk_mq_hw_queue_mapped(hctx
))
620 blk_mq_tag_busy_iter(hctx
, blk_mq_check_expired
, &data
);
624 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
625 mod_timer(&q
->timeout
, data
.next
);
627 queue_for_each_hw_ctx(q
, hctx
, i
)
628 blk_mq_tag_idle(hctx
);
633 * Reverse check our software queue for entries that we could potentially
634 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
635 * too much time checking for merges.
637 static bool blk_mq_attempt_merge(struct request_queue
*q
,
638 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
643 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
649 if (!blk_rq_merge_ok(rq
, bio
))
652 el_ret
= blk_try_merge(rq
, bio
);
653 if (el_ret
== ELEVATOR_BACK_MERGE
) {
654 if (bio_attempt_back_merge(q
, rq
, bio
)) {
659 } else if (el_ret
== ELEVATOR_FRONT_MERGE
) {
660 if (bio_attempt_front_merge(q
, rq
, bio
)) {
672 * Process software queues that have been marked busy, splicing them
673 * to the for-dispatch
675 static void flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
677 struct blk_mq_ctx
*ctx
;
680 for (i
= 0; i
< hctx
->ctx_map
.map_size
; i
++) {
681 struct blk_align_bitmap
*bm
= &hctx
->ctx_map
.map
[i
];
682 unsigned int off
, bit
;
688 off
= i
* hctx
->ctx_map
.bits_per_word
;
690 bit
= find_next_bit(&bm
->word
, bm
->depth
, bit
);
691 if (bit
>= bm
->depth
)
694 ctx
= hctx
->ctxs
[bit
+ off
];
695 clear_bit(bit
, &bm
->word
);
696 spin_lock(&ctx
->lock
);
697 list_splice_tail_init(&ctx
->rq_list
, list
);
698 spin_unlock(&ctx
->lock
);
706 * Run this hardware queue, pulling any software queues mapped to it in.
707 * Note that this function currently has various problems around ordering
708 * of IO. In particular, we'd like FIFO behaviour on handling existing
709 * items on the hctx->dispatch list. Ignore that for now.
711 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
713 struct request_queue
*q
= hctx
->queue
;
716 LIST_HEAD(driver_list
);
717 struct list_head
*dptr
;
720 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
));
722 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)))
728 * Touch any software queue that has pending entries.
730 flush_busy_ctxs(hctx
, &rq_list
);
733 * If we have previous entries on our dispatch list, grab them
734 * and stuff them at the front for more fair dispatch.
736 if (!list_empty_careful(&hctx
->dispatch
)) {
737 spin_lock(&hctx
->lock
);
738 if (!list_empty(&hctx
->dispatch
))
739 list_splice_init(&hctx
->dispatch
, &rq_list
);
740 spin_unlock(&hctx
->lock
);
744 * Start off with dptr being NULL, so we start the first request
745 * immediately, even if we have more pending.
750 * Now process all the entries, sending them to the driver.
753 while (!list_empty(&rq_list
)) {
754 struct blk_mq_queue_data bd
;
757 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
758 list_del_init(&rq
->queuelist
);
762 bd
.last
= list_empty(&rq_list
);
764 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
766 case BLK_MQ_RQ_QUEUE_OK
:
769 case BLK_MQ_RQ_QUEUE_BUSY
:
770 list_add(&rq
->queuelist
, &rq_list
);
771 __blk_mq_requeue_request(rq
);
774 pr_err("blk-mq: bad return on queue: %d\n", ret
);
775 case BLK_MQ_RQ_QUEUE_ERROR
:
777 blk_mq_end_request(rq
, rq
->errors
);
781 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
785 * We've done the first request. If we have more than 1
786 * left in the list, set dptr to defer issue.
788 if (!dptr
&& rq_list
.next
!= rq_list
.prev
)
793 hctx
->dispatched
[0]++;
794 else if (queued
< (1 << (BLK_MQ_MAX_DISPATCH_ORDER
- 1)))
795 hctx
->dispatched
[ilog2(queued
) + 1]++;
798 * Any items that need requeuing? Stuff them into hctx->dispatch,
799 * that is where we will continue on next queue run.
801 if (!list_empty(&rq_list
)) {
802 spin_lock(&hctx
->lock
);
803 list_splice(&rq_list
, &hctx
->dispatch
);
804 spin_unlock(&hctx
->lock
);
809 * It'd be great if the workqueue API had a way to pass
810 * in a mask and had some smarts for more clever placement.
811 * For now we just round-robin here, switching for every
812 * BLK_MQ_CPU_WORK_BATCH queued items.
814 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
816 if (hctx
->queue
->nr_hw_queues
== 1)
817 return WORK_CPU_UNBOUND
;
819 if (--hctx
->next_cpu_batch
<= 0) {
820 int cpu
= hctx
->next_cpu
, next_cpu
;
822 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
823 if (next_cpu
>= nr_cpu_ids
)
824 next_cpu
= cpumask_first(hctx
->cpumask
);
826 hctx
->next_cpu
= next_cpu
;
827 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
832 return hctx
->next_cpu
;
835 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
837 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
) ||
838 !blk_mq_hw_queue_mapped(hctx
)))
843 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
844 __blk_mq_run_hw_queue(hctx
);
852 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
856 void blk_mq_run_queues(struct request_queue
*q
, bool async
)
858 struct blk_mq_hw_ctx
*hctx
;
861 queue_for_each_hw_ctx(q
, hctx
, i
) {
862 if ((!blk_mq_hctx_has_pending(hctx
) &&
863 list_empty_careful(&hctx
->dispatch
)) ||
864 test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
867 blk_mq_run_hw_queue(hctx
, async
);
870 EXPORT_SYMBOL(blk_mq_run_queues
);
872 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
874 cancel_delayed_work(&hctx
->run_work
);
875 cancel_delayed_work(&hctx
->delay_work
);
876 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
878 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
880 void blk_mq_stop_hw_queues(struct request_queue
*q
)
882 struct blk_mq_hw_ctx
*hctx
;
885 queue_for_each_hw_ctx(q
, hctx
, i
)
886 blk_mq_stop_hw_queue(hctx
);
888 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
890 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
892 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
894 blk_mq_run_hw_queue(hctx
, false);
896 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
898 void blk_mq_start_hw_queues(struct request_queue
*q
)
900 struct blk_mq_hw_ctx
*hctx
;
903 queue_for_each_hw_ctx(q
, hctx
, i
)
904 blk_mq_start_hw_queue(hctx
);
906 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
909 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
911 struct blk_mq_hw_ctx
*hctx
;
914 queue_for_each_hw_ctx(q
, hctx
, i
) {
915 if (!test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
918 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
919 blk_mq_run_hw_queue(hctx
, async
);
922 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
924 static void blk_mq_run_work_fn(struct work_struct
*work
)
926 struct blk_mq_hw_ctx
*hctx
;
928 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
930 __blk_mq_run_hw_queue(hctx
);
933 static void blk_mq_delay_work_fn(struct work_struct
*work
)
935 struct blk_mq_hw_ctx
*hctx
;
937 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
939 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
940 __blk_mq_run_hw_queue(hctx
);
943 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
945 if (unlikely(!blk_mq_hw_queue_mapped(hctx
)))
948 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
949 &hctx
->delay_work
, msecs_to_jiffies(msecs
));
951 EXPORT_SYMBOL(blk_mq_delay_queue
);
953 static void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
,
954 struct request
*rq
, bool at_head
)
956 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
958 trace_block_rq_insert(hctx
->queue
, rq
);
961 list_add(&rq
->queuelist
, &ctx
->rq_list
);
963 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
965 blk_mq_hctx_mark_pending(hctx
, ctx
);
968 void blk_mq_insert_request(struct request
*rq
, bool at_head
, bool run_queue
,
971 struct request_queue
*q
= rq
->q
;
972 struct blk_mq_hw_ctx
*hctx
;
973 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
, *current_ctx
;
975 current_ctx
= blk_mq_get_ctx(q
);
976 if (!cpu_online(ctx
->cpu
))
977 rq
->mq_ctx
= ctx
= current_ctx
;
979 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
981 spin_lock(&ctx
->lock
);
982 __blk_mq_insert_request(hctx
, rq
, at_head
);
983 spin_unlock(&ctx
->lock
);
986 blk_mq_run_hw_queue(hctx
, async
);
988 blk_mq_put_ctx(current_ctx
);
991 static void blk_mq_insert_requests(struct request_queue
*q
,
992 struct blk_mq_ctx
*ctx
,
993 struct list_head
*list
,
998 struct blk_mq_hw_ctx
*hctx
;
999 struct blk_mq_ctx
*current_ctx
;
1001 trace_block_unplug(q
, depth
, !from_schedule
);
1003 current_ctx
= blk_mq_get_ctx(q
);
1005 if (!cpu_online(ctx
->cpu
))
1007 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1010 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1013 spin_lock(&ctx
->lock
);
1014 while (!list_empty(list
)) {
1017 rq
= list_first_entry(list
, struct request
, queuelist
);
1018 list_del_init(&rq
->queuelist
);
1020 __blk_mq_insert_request(hctx
, rq
, false);
1022 spin_unlock(&ctx
->lock
);
1024 blk_mq_run_hw_queue(hctx
, from_schedule
);
1025 blk_mq_put_ctx(current_ctx
);
1028 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1030 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1031 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1033 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1034 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1035 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1038 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1040 struct blk_mq_ctx
*this_ctx
;
1041 struct request_queue
*this_q
;
1044 LIST_HEAD(ctx_list
);
1047 list_splice_init(&plug
->mq_list
, &list
);
1049 list_sort(NULL
, &list
, plug_ctx_cmp
);
1055 while (!list_empty(&list
)) {
1056 rq
= list_entry_rq(list
.next
);
1057 list_del_init(&rq
->queuelist
);
1059 if (rq
->mq_ctx
!= this_ctx
) {
1061 blk_mq_insert_requests(this_q
, this_ctx
,
1066 this_ctx
= rq
->mq_ctx
;
1072 list_add_tail(&rq
->queuelist
, &ctx_list
);
1076 * If 'this_ctx' is set, we know we have entries to complete
1077 * on 'ctx_list'. Do those.
1080 blk_mq_insert_requests(this_q
, this_ctx
, &ctx_list
, depth
,
1085 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1087 init_request_from_bio(rq
, bio
);
1089 if (blk_do_io_stat(rq
))
1090 blk_account_io_start(rq
, 1);
1093 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1095 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1096 !blk_queue_nomerges(hctx
->queue
);
1099 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx
*hctx
,
1100 struct blk_mq_ctx
*ctx
,
1101 struct request
*rq
, struct bio
*bio
)
1103 if (!hctx_allow_merges(hctx
)) {
1104 blk_mq_bio_to_request(rq
, bio
);
1105 spin_lock(&ctx
->lock
);
1107 __blk_mq_insert_request(hctx
, rq
, false);
1108 spin_unlock(&ctx
->lock
);
1111 struct request_queue
*q
= hctx
->queue
;
1113 spin_lock(&ctx
->lock
);
1114 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1115 blk_mq_bio_to_request(rq
, bio
);
1119 spin_unlock(&ctx
->lock
);
1120 __blk_mq_free_request(hctx
, ctx
, rq
);
1125 struct blk_map_ctx
{
1126 struct blk_mq_hw_ctx
*hctx
;
1127 struct blk_mq_ctx
*ctx
;
1130 static struct request
*blk_mq_map_request(struct request_queue
*q
,
1132 struct blk_map_ctx
*data
)
1134 struct blk_mq_hw_ctx
*hctx
;
1135 struct blk_mq_ctx
*ctx
;
1137 int rw
= bio_data_dir(bio
);
1138 struct blk_mq_alloc_data alloc_data
;
1140 if (unlikely(blk_mq_queue_enter(q
))) {
1141 bio_endio(bio
, -EIO
);
1145 ctx
= blk_mq_get_ctx(q
);
1146 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1148 if (rw_is_sync(bio
->bi_rw
))
1151 trace_block_getrq(q
, bio
, rw
);
1152 blk_mq_set_alloc_data(&alloc_data
, q
, GFP_ATOMIC
, false, ctx
,
1154 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
1155 if (unlikely(!rq
)) {
1156 __blk_mq_run_hw_queue(hctx
);
1157 blk_mq_put_ctx(ctx
);
1158 trace_block_sleeprq(q
, bio
, rw
);
1160 ctx
= blk_mq_get_ctx(q
);
1161 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1162 blk_mq_set_alloc_data(&alloc_data
, q
,
1163 __GFP_WAIT
|GFP_ATOMIC
, false, ctx
, hctx
);
1164 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
1165 ctx
= alloc_data
.ctx
;
1166 hctx
= alloc_data
.hctx
;
1176 * Multiple hardware queue variant. This will not use per-process plugs,
1177 * but will attempt to bypass the hctx queueing if we can go straight to
1178 * hardware for SYNC IO.
1180 static void blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1182 const int is_sync
= rw_is_sync(bio
->bi_rw
);
1183 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
1184 struct blk_map_ctx data
;
1187 blk_queue_bounce(q
, &bio
);
1189 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1190 bio_endio(bio
, -EIO
);
1194 rq
= blk_mq_map_request(q
, bio
, &data
);
1198 if (unlikely(is_flush_fua
)) {
1199 blk_mq_bio_to_request(rq
, bio
);
1200 blk_insert_flush(rq
);
1205 * If the driver supports defer issued based on 'last', then
1206 * queue it up like normal since we can potentially save some
1209 if (is_sync
&& !(data
.hctx
->flags
& BLK_MQ_F_DEFER_ISSUE
)) {
1210 struct blk_mq_queue_data bd
= {
1217 blk_mq_bio_to_request(rq
, bio
);
1220 * For OK queue, we are done. For error, kill it. Any other
1221 * error (busy), just add it to our list as we previously
1224 ret
= q
->mq_ops
->queue_rq(data
.hctx
, &bd
);
1225 if (ret
== BLK_MQ_RQ_QUEUE_OK
)
1228 __blk_mq_requeue_request(rq
);
1230 if (ret
== BLK_MQ_RQ_QUEUE_ERROR
) {
1232 blk_mq_end_request(rq
, rq
->errors
);
1238 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1240 * For a SYNC request, send it to the hardware immediately. For
1241 * an ASYNC request, just ensure that we run it later on. The
1242 * latter allows for merging opportunities and more efficient
1246 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1249 blk_mq_put_ctx(data
.ctx
);
1253 * Single hardware queue variant. This will attempt to use any per-process
1254 * plug for merging and IO deferral.
1256 static void blk_sq_make_request(struct request_queue
*q
, struct bio
*bio
)
1258 const int is_sync
= rw_is_sync(bio
->bi_rw
);
1259 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
1260 unsigned int use_plug
, request_count
= 0;
1261 struct blk_map_ctx data
;
1265 * If we have multiple hardware queues, just go directly to
1266 * one of those for sync IO.
1268 use_plug
= !is_flush_fua
&& !is_sync
;
1270 blk_queue_bounce(q
, &bio
);
1272 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1273 bio_endio(bio
, -EIO
);
1277 if (use_plug
&& !blk_queue_nomerges(q
) &&
1278 blk_attempt_plug_merge(q
, bio
, &request_count
))
1281 rq
= blk_mq_map_request(q
, bio
, &data
);
1285 if (unlikely(is_flush_fua
)) {
1286 blk_mq_bio_to_request(rq
, bio
);
1287 blk_insert_flush(rq
);
1292 * A task plug currently exists. Since this is completely lockless,
1293 * utilize that to temporarily store requests until the task is
1294 * either done or scheduled away.
1297 struct blk_plug
*plug
= current
->plug
;
1300 blk_mq_bio_to_request(rq
, bio
);
1301 if (list_empty(&plug
->mq_list
))
1302 trace_block_plug(q
);
1303 else if (request_count
>= BLK_MAX_REQUEST_COUNT
) {
1304 blk_flush_plug_list(plug
, false);
1305 trace_block_plug(q
);
1307 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1308 blk_mq_put_ctx(data
.ctx
);
1313 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1315 * For a SYNC request, send it to the hardware immediately. For
1316 * an ASYNC request, just ensure that we run it later on. The
1317 * latter allows for merging opportunities and more efficient
1321 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1324 blk_mq_put_ctx(data
.ctx
);
1328 * Default mapping to a software queue, since we use one per CPU.
1330 struct blk_mq_hw_ctx
*blk_mq_map_queue(struct request_queue
*q
, const int cpu
)
1332 return q
->queue_hw_ctx
[q
->mq_map
[cpu
]];
1334 EXPORT_SYMBOL(blk_mq_map_queue
);
1336 static void blk_mq_free_rq_map(struct blk_mq_tag_set
*set
,
1337 struct blk_mq_tags
*tags
, unsigned int hctx_idx
)
1341 if (tags
->rqs
&& set
->ops
->exit_request
) {
1344 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1347 set
->ops
->exit_request(set
->driver_data
, tags
->rqs
[i
],
1349 tags
->rqs
[i
] = NULL
;
1353 while (!list_empty(&tags
->page_list
)) {
1354 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1355 list_del_init(&page
->lru
);
1356 __free_pages(page
, page
->private);
1361 blk_mq_free_tags(tags
);
1364 static size_t order_to_size(unsigned int order
)
1366 return (size_t)PAGE_SIZE
<< order
;
1369 static struct blk_mq_tags
*blk_mq_init_rq_map(struct blk_mq_tag_set
*set
,
1370 unsigned int hctx_idx
)
1372 struct blk_mq_tags
*tags
;
1373 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1374 size_t rq_size
, left
;
1376 tags
= blk_mq_init_tags(set
->queue_depth
, set
->reserved_tags
,
1378 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1382 INIT_LIST_HEAD(&tags
->page_list
);
1384 tags
->rqs
= kzalloc_node(set
->queue_depth
* sizeof(struct request
*),
1385 GFP_KERNEL
| __GFP_NOWARN
| __GFP_NORETRY
,
1388 blk_mq_free_tags(tags
);
1393 * rq_size is the size of the request plus driver payload, rounded
1394 * to the cacheline size
1396 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1398 left
= rq_size
* set
->queue_depth
;
1400 for (i
= 0; i
< set
->queue_depth
; ) {
1401 int this_order
= max_order
;
1406 while (left
< order_to_size(this_order
- 1) && this_order
)
1410 page
= alloc_pages_node(set
->numa_node
,
1411 GFP_KERNEL
| __GFP_NOWARN
| __GFP_NORETRY
,
1417 if (order_to_size(this_order
) < rq_size
)
1424 page
->private = this_order
;
1425 list_add_tail(&page
->lru
, &tags
->page_list
);
1427 p
= page_address(page
);
1428 entries_per_page
= order_to_size(this_order
) / rq_size
;
1429 to_do
= min(entries_per_page
, set
->queue_depth
- i
);
1430 left
-= to_do
* rq_size
;
1431 for (j
= 0; j
< to_do
; j
++) {
1433 tags
->rqs
[i
]->atomic_flags
= 0;
1434 tags
->rqs
[i
]->cmd_flags
= 0;
1435 if (set
->ops
->init_request
) {
1436 if (set
->ops
->init_request(set
->driver_data
,
1437 tags
->rqs
[i
], hctx_idx
, i
,
1439 tags
->rqs
[i
] = NULL
;
1452 blk_mq_free_rq_map(set
, tags
, hctx_idx
);
1456 static void blk_mq_free_bitmap(struct blk_mq_ctxmap
*bitmap
)
1461 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap
*bitmap
, int node
)
1463 unsigned int bpw
= 8, total
, num_maps
, i
;
1465 bitmap
->bits_per_word
= bpw
;
1467 num_maps
= ALIGN(nr_cpu_ids
, bpw
) / bpw
;
1468 bitmap
->map
= kzalloc_node(num_maps
* sizeof(struct blk_align_bitmap
),
1473 bitmap
->map_size
= num_maps
;
1476 for (i
= 0; i
< num_maps
; i
++) {
1477 bitmap
->map
[i
].depth
= min(total
, bitmap
->bits_per_word
);
1478 total
-= bitmap
->map
[i
].depth
;
1484 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx
*hctx
, int cpu
)
1486 struct request_queue
*q
= hctx
->queue
;
1487 struct blk_mq_ctx
*ctx
;
1491 * Move ctx entries to new CPU, if this one is going away.
1493 ctx
= __blk_mq_get_ctx(q
, cpu
);
1495 spin_lock(&ctx
->lock
);
1496 if (!list_empty(&ctx
->rq_list
)) {
1497 list_splice_init(&ctx
->rq_list
, &tmp
);
1498 blk_mq_hctx_clear_pending(hctx
, ctx
);
1500 spin_unlock(&ctx
->lock
);
1502 if (list_empty(&tmp
))
1505 ctx
= blk_mq_get_ctx(q
);
1506 spin_lock(&ctx
->lock
);
1508 while (!list_empty(&tmp
)) {
1511 rq
= list_first_entry(&tmp
, struct request
, queuelist
);
1513 list_move_tail(&rq
->queuelist
, &ctx
->rq_list
);
1516 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1517 blk_mq_hctx_mark_pending(hctx
, ctx
);
1519 spin_unlock(&ctx
->lock
);
1521 blk_mq_run_hw_queue(hctx
, true);
1522 blk_mq_put_ctx(ctx
);
1526 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx
*hctx
, int cpu
)
1528 struct request_queue
*q
= hctx
->queue
;
1529 struct blk_mq_tag_set
*set
= q
->tag_set
;
1531 if (set
->tags
[hctx
->queue_num
])
1534 set
->tags
[hctx
->queue_num
] = blk_mq_init_rq_map(set
, hctx
->queue_num
);
1535 if (!set
->tags
[hctx
->queue_num
])
1538 hctx
->tags
= set
->tags
[hctx
->queue_num
];
1542 static int blk_mq_hctx_notify(void *data
, unsigned long action
,
1545 struct blk_mq_hw_ctx
*hctx
= data
;
1547 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
)
1548 return blk_mq_hctx_cpu_offline(hctx
, cpu
);
1549 else if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
)
1550 return blk_mq_hctx_cpu_online(hctx
, cpu
);
1555 static void blk_mq_exit_hctx(struct request_queue
*q
,
1556 struct blk_mq_tag_set
*set
,
1557 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
1559 unsigned flush_start_tag
= set
->queue_depth
;
1561 blk_mq_tag_idle(hctx
);
1563 if (set
->ops
->exit_request
)
1564 set
->ops
->exit_request(set
->driver_data
,
1565 hctx
->fq
->flush_rq
, hctx_idx
,
1566 flush_start_tag
+ hctx_idx
);
1568 if (set
->ops
->exit_hctx
)
1569 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1571 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1572 blk_free_flush_queue(hctx
->fq
);
1574 blk_mq_free_bitmap(&hctx
->ctx_map
);
1577 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1578 struct blk_mq_tag_set
*set
, int nr_queue
)
1580 struct blk_mq_hw_ctx
*hctx
;
1583 queue_for_each_hw_ctx(q
, hctx
, i
) {
1586 blk_mq_exit_hctx(q
, set
, hctx
, i
);
1590 static void blk_mq_free_hw_queues(struct request_queue
*q
,
1591 struct blk_mq_tag_set
*set
)
1593 struct blk_mq_hw_ctx
*hctx
;
1596 queue_for_each_hw_ctx(q
, hctx
, i
) {
1597 free_cpumask_var(hctx
->cpumask
);
1602 static int blk_mq_init_hctx(struct request_queue
*q
,
1603 struct blk_mq_tag_set
*set
,
1604 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
1607 unsigned flush_start_tag
= set
->queue_depth
;
1609 node
= hctx
->numa_node
;
1610 if (node
== NUMA_NO_NODE
)
1611 node
= hctx
->numa_node
= set
->numa_node
;
1613 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1614 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1615 spin_lock_init(&hctx
->lock
);
1616 INIT_LIST_HEAD(&hctx
->dispatch
);
1618 hctx
->queue_num
= hctx_idx
;
1619 hctx
->flags
= set
->flags
;
1620 hctx
->cmd_size
= set
->cmd_size
;
1622 blk_mq_init_cpu_notifier(&hctx
->cpu_notifier
,
1623 blk_mq_hctx_notify
, hctx
);
1624 blk_mq_register_cpu_notifier(&hctx
->cpu_notifier
);
1626 hctx
->tags
= set
->tags
[hctx_idx
];
1629 * Allocate space for all possible cpus to avoid allocation at
1632 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1635 goto unregister_cpu_notifier
;
1637 if (blk_mq_alloc_bitmap(&hctx
->ctx_map
, node
))
1642 if (set
->ops
->init_hctx
&&
1643 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
1646 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
1650 if (set
->ops
->init_request
&&
1651 set
->ops
->init_request(set
->driver_data
,
1652 hctx
->fq
->flush_rq
, hctx_idx
,
1653 flush_start_tag
+ hctx_idx
, node
))
1661 if (set
->ops
->exit_hctx
)
1662 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1664 blk_mq_free_bitmap(&hctx
->ctx_map
);
1667 unregister_cpu_notifier
:
1668 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1673 static int blk_mq_init_hw_queues(struct request_queue
*q
,
1674 struct blk_mq_tag_set
*set
)
1676 struct blk_mq_hw_ctx
*hctx
;
1680 * Initialize hardware queues
1682 queue_for_each_hw_ctx(q
, hctx
, i
) {
1683 if (blk_mq_init_hctx(q
, set
, hctx
, i
))
1687 if (i
== q
->nr_hw_queues
)
1693 blk_mq_exit_hw_queues(q
, set
, i
);
1698 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1699 unsigned int nr_hw_queues
)
1703 for_each_possible_cpu(i
) {
1704 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1705 struct blk_mq_hw_ctx
*hctx
;
1707 memset(__ctx
, 0, sizeof(*__ctx
));
1709 spin_lock_init(&__ctx
->lock
);
1710 INIT_LIST_HEAD(&__ctx
->rq_list
);
1713 /* If the cpu isn't online, the cpu is mapped to first hctx */
1717 hctx
= q
->mq_ops
->map_queue(q
, i
);
1718 cpumask_set_cpu(i
, hctx
->cpumask
);
1722 * Set local node, IFF we have more than one hw queue. If
1723 * not, we remain on the home node of the device
1725 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1726 hctx
->numa_node
= cpu_to_node(i
);
1730 static void blk_mq_map_swqueue(struct request_queue
*q
)
1733 struct blk_mq_hw_ctx
*hctx
;
1734 struct blk_mq_ctx
*ctx
;
1736 queue_for_each_hw_ctx(q
, hctx
, i
) {
1737 cpumask_clear(hctx
->cpumask
);
1742 * Map software to hardware queues
1744 queue_for_each_ctx(q
, ctx
, i
) {
1745 /* If the cpu isn't online, the cpu is mapped to first hctx */
1749 hctx
= q
->mq_ops
->map_queue(q
, i
);
1750 cpumask_set_cpu(i
, hctx
->cpumask
);
1751 ctx
->index_hw
= hctx
->nr_ctx
;
1752 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
1755 queue_for_each_hw_ctx(q
, hctx
, i
) {
1757 * If no software queues are mapped to this hardware queue,
1758 * disable it and free the request entries.
1760 if (!hctx
->nr_ctx
) {
1761 struct blk_mq_tag_set
*set
= q
->tag_set
;
1764 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1765 set
->tags
[i
] = NULL
;
1772 * Initialize batch roundrobin counts
1774 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
1775 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1779 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
)
1781 struct blk_mq_hw_ctx
*hctx
;
1782 struct request_queue
*q
;
1786 if (set
->tag_list
.next
== set
->tag_list
.prev
)
1791 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
1792 blk_mq_freeze_queue(q
);
1794 queue_for_each_hw_ctx(q
, hctx
, i
) {
1796 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
1798 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
1800 blk_mq_unfreeze_queue(q
);
1804 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
1806 struct blk_mq_tag_set
*set
= q
->tag_set
;
1808 mutex_lock(&set
->tag_list_lock
);
1809 list_del_init(&q
->tag_set_list
);
1810 blk_mq_update_tag_set_depth(set
);
1811 mutex_unlock(&set
->tag_list_lock
);
1814 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
1815 struct request_queue
*q
)
1819 mutex_lock(&set
->tag_list_lock
);
1820 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
1821 blk_mq_update_tag_set_depth(set
);
1822 mutex_unlock(&set
->tag_list_lock
);
1825 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
1827 struct blk_mq_hw_ctx
**hctxs
;
1828 struct blk_mq_ctx __percpu
*ctx
;
1829 struct request_queue
*q
;
1833 ctx
= alloc_percpu(struct blk_mq_ctx
);
1835 return ERR_PTR(-ENOMEM
);
1837 hctxs
= kmalloc_node(set
->nr_hw_queues
* sizeof(*hctxs
), GFP_KERNEL
,
1843 map
= blk_mq_make_queue_map(set
);
1847 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1848 int node
= blk_mq_hw_queue_to_node(map
, i
);
1850 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
1855 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
1859 atomic_set(&hctxs
[i
]->nr_active
, 0);
1860 hctxs
[i
]->numa_node
= node
;
1861 hctxs
[i
]->queue_num
= i
;
1864 q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
1869 * Init percpu_ref in atomic mode so that it's faster to shutdown.
1870 * See blk_register_queue() for details.
1872 if (percpu_ref_init(&q
->mq_usage_counter
, blk_mq_usage_counter_release
,
1873 PERCPU_REF_INIT_ATOMIC
, GFP_KERNEL
))
1876 setup_timer(&q
->timeout
, blk_mq_rq_timer
, (unsigned long) q
);
1877 blk_queue_rq_timeout(q
, 30000);
1879 q
->nr_queues
= nr_cpu_ids
;
1880 q
->nr_hw_queues
= set
->nr_hw_queues
;
1884 q
->queue_hw_ctx
= hctxs
;
1886 q
->mq_ops
= set
->ops
;
1887 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
1889 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
1890 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
1892 q
->sg_reserved_size
= INT_MAX
;
1894 INIT_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
1895 INIT_LIST_HEAD(&q
->requeue_list
);
1896 spin_lock_init(&q
->requeue_lock
);
1898 if (q
->nr_hw_queues
> 1)
1899 blk_queue_make_request(q
, blk_mq_make_request
);
1901 blk_queue_make_request(q
, blk_sq_make_request
);
1904 blk_queue_rq_timeout(q
, set
->timeout
);
1907 * Do this after blk_queue_make_request() overrides it...
1909 q
->nr_requests
= set
->queue_depth
;
1911 if (set
->ops
->complete
)
1912 blk_queue_softirq_done(q
, set
->ops
->complete
);
1914 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
1916 if (blk_mq_init_hw_queues(q
, set
))
1919 mutex_lock(&all_q_mutex
);
1920 list_add_tail(&q
->all_q_node
, &all_q_list
);
1921 mutex_unlock(&all_q_mutex
);
1923 blk_mq_add_queue_tag_set(set
, q
);
1925 blk_mq_map_swqueue(q
);
1930 blk_cleanup_queue(q
);
1933 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1936 free_cpumask_var(hctxs
[i
]->cpumask
);
1943 return ERR_PTR(-ENOMEM
);
1945 EXPORT_SYMBOL(blk_mq_init_queue
);
1947 void blk_mq_free_queue(struct request_queue
*q
)
1949 struct blk_mq_tag_set
*set
= q
->tag_set
;
1951 blk_mq_del_queue_tag_set(q
);
1953 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
1954 blk_mq_free_hw_queues(q
, set
);
1956 percpu_ref_exit(&q
->mq_usage_counter
);
1958 free_percpu(q
->queue_ctx
);
1959 kfree(q
->queue_hw_ctx
);
1962 q
->queue_ctx
= NULL
;
1963 q
->queue_hw_ctx
= NULL
;
1966 mutex_lock(&all_q_mutex
);
1967 list_del_init(&q
->all_q_node
);
1968 mutex_unlock(&all_q_mutex
);
1971 /* Basically redo blk_mq_init_queue with queue frozen */
1972 static void blk_mq_queue_reinit(struct request_queue
*q
)
1974 WARN_ON_ONCE(!q
->mq_freeze_depth
);
1976 blk_mq_sysfs_unregister(q
);
1978 blk_mq_update_queue_map(q
->mq_map
, q
->nr_hw_queues
);
1981 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1982 * we should change hctx numa_node according to new topology (this
1983 * involves free and re-allocate memory, worthy doing?)
1986 blk_mq_map_swqueue(q
);
1988 blk_mq_sysfs_register(q
);
1991 static int blk_mq_queue_reinit_notify(struct notifier_block
*nb
,
1992 unsigned long action
, void *hcpu
)
1994 struct request_queue
*q
;
1997 * Before new mappings are established, hotadded cpu might already
1998 * start handling requests. This doesn't break anything as we map
1999 * offline CPUs to first hardware queue. We will re-init the queue
2000 * below to get optimal settings.
2002 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
&&
2003 action
!= CPU_ONLINE
&& action
!= CPU_ONLINE_FROZEN
)
2006 mutex_lock(&all_q_mutex
);
2009 * We need to freeze and reinit all existing queues. Freezing
2010 * involves synchronous wait for an RCU grace period and doing it
2011 * one by one may take a long time. Start freezing all queues in
2012 * one swoop and then wait for the completions so that freezing can
2013 * take place in parallel.
2015 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2016 blk_mq_freeze_queue_start(q
);
2017 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2018 blk_mq_freeze_queue_wait(q
);
2020 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2021 blk_mq_queue_reinit(q
);
2023 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2024 blk_mq_unfreeze_queue(q
);
2026 mutex_unlock(&all_q_mutex
);
2030 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2034 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2035 set
->tags
[i
] = blk_mq_init_rq_map(set
, i
);
2044 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2050 * Allocate the request maps associated with this tag_set. Note that this
2051 * may reduce the depth asked for, if memory is tight. set->queue_depth
2052 * will be updated to reflect the allocated depth.
2054 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2059 depth
= set
->queue_depth
;
2061 err
= __blk_mq_alloc_rq_maps(set
);
2065 set
->queue_depth
>>= 1;
2066 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2070 } while (set
->queue_depth
);
2072 if (!set
->queue_depth
|| err
) {
2073 pr_err("blk-mq: failed to allocate request map\n");
2077 if (depth
!= set
->queue_depth
)
2078 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2079 depth
, set
->queue_depth
);
2085 * Alloc a tag set to be associated with one or more request queues.
2086 * May fail with EINVAL for various error conditions. May adjust the
2087 * requested depth down, if if it too large. In that case, the set
2088 * value will be stored in set->queue_depth.
2090 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2092 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2094 if (!set
->nr_hw_queues
)
2096 if (!set
->queue_depth
)
2098 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2101 if (!set
->nr_hw_queues
|| !set
->ops
->queue_rq
|| !set
->ops
->map_queue
)
2104 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2105 pr_info("blk-mq: reduced tag depth to %u\n",
2107 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2111 * If a crashdump is active, then we are potentially in a very
2112 * memory constrained environment. Limit us to 1 queue and
2113 * 64 tags to prevent using too much memory.
2115 if (is_kdump_kernel()) {
2116 set
->nr_hw_queues
= 1;
2117 set
->queue_depth
= min(64U, set
->queue_depth
);
2120 set
->tags
= kmalloc_node(set
->nr_hw_queues
*
2121 sizeof(struct blk_mq_tags
*),
2122 GFP_KERNEL
, set
->numa_node
);
2126 if (blk_mq_alloc_rq_maps(set
))
2129 mutex_init(&set
->tag_list_lock
);
2130 INIT_LIST_HEAD(&set
->tag_list
);
2138 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2140 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2144 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2146 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2152 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2154 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2156 struct blk_mq_tag_set
*set
= q
->tag_set
;
2157 struct blk_mq_hw_ctx
*hctx
;
2160 if (!set
|| nr
> set
->queue_depth
)
2164 queue_for_each_hw_ctx(q
, hctx
, i
) {
2165 ret
= blk_mq_tag_update_depth(hctx
->tags
, nr
);
2171 q
->nr_requests
= nr
;
2176 void blk_mq_disable_hotplug(void)
2178 mutex_lock(&all_q_mutex
);
2181 void blk_mq_enable_hotplug(void)
2183 mutex_unlock(&all_q_mutex
);
2186 static int __init
blk_mq_init(void)
2190 hotcpu_notifier(blk_mq_queue_reinit_notify
, 0);
2194 subsys_initcall(blk_mq_init
);