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>
24 #include <trace/events/block.h>
26 #include <linux/blk-mq.h>
29 #include "blk-mq-tag.h"
31 static DEFINE_MUTEX(all_q_mutex
);
32 static LIST_HEAD(all_q_list
);
34 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
);
37 * Check if any of the ctx's have pending work in this hardware queue
39 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
43 for (i
= 0; i
< hctx
->ctx_map
.map_size
; i
++)
44 if (hctx
->ctx_map
.map
[i
].word
)
50 static inline struct blk_align_bitmap
*get_bm(struct blk_mq_hw_ctx
*hctx
,
51 struct blk_mq_ctx
*ctx
)
53 return &hctx
->ctx_map
.map
[ctx
->index_hw
/ hctx
->ctx_map
.bits_per_word
];
56 #define CTX_TO_BIT(hctx, ctx) \
57 ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
60 * Mark this ctx as having pending work in this hardware queue
62 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
63 struct blk_mq_ctx
*ctx
)
65 struct blk_align_bitmap
*bm
= get_bm(hctx
, ctx
);
67 if (!test_bit(CTX_TO_BIT(hctx
, ctx
), &bm
->word
))
68 set_bit(CTX_TO_BIT(hctx
, ctx
), &bm
->word
);
71 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
72 struct blk_mq_ctx
*ctx
)
74 struct blk_align_bitmap
*bm
= get_bm(hctx
, ctx
);
76 clear_bit(CTX_TO_BIT(hctx
, ctx
), &bm
->word
);
79 static int blk_mq_queue_enter(struct request_queue
*q
)
84 if (percpu_ref_tryget_live(&q
->mq_usage_counter
))
87 ret
= wait_event_interruptible(q
->mq_freeze_wq
,
88 !q
->mq_freeze_depth
|| blk_queue_dying(q
));
89 if (blk_queue_dying(q
))
96 static void blk_mq_queue_exit(struct request_queue
*q
)
98 percpu_ref_put(&q
->mq_usage_counter
);
101 static void blk_mq_usage_counter_release(struct percpu_ref
*ref
)
103 struct request_queue
*q
=
104 container_of(ref
, struct request_queue
, mq_usage_counter
);
106 wake_up_all(&q
->mq_freeze_wq
);
110 * Guarantee no request is in use, so we can change any data structure of
111 * the queue afterward.
113 void blk_mq_freeze_queue(struct request_queue
*q
)
117 spin_lock_irq(q
->queue_lock
);
118 freeze
= !q
->mq_freeze_depth
++;
119 spin_unlock_irq(q
->queue_lock
);
122 percpu_ref_kill(&q
->mq_usage_counter
);
123 blk_mq_run_queues(q
, false);
125 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->mq_usage_counter
));
128 static void blk_mq_unfreeze_queue(struct request_queue
*q
)
132 spin_lock_irq(q
->queue_lock
);
133 wake
= !--q
->mq_freeze_depth
;
134 WARN_ON_ONCE(q
->mq_freeze_depth
< 0);
135 spin_unlock_irq(q
->queue_lock
);
137 percpu_ref_reinit(&q
->mq_usage_counter
);
138 wake_up_all(&q
->mq_freeze_wq
);
142 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
144 return blk_mq_has_free_tags(hctx
->tags
);
146 EXPORT_SYMBOL(blk_mq_can_queue
);
148 static void blk_mq_rq_ctx_init(struct request_queue
*q
, struct blk_mq_ctx
*ctx
,
149 struct request
*rq
, unsigned int rw_flags
)
151 if (blk_queue_io_stat(q
))
152 rw_flags
|= REQ_IO_STAT
;
154 INIT_LIST_HEAD(&rq
->queuelist
);
155 /* csd/requeue_work/fifo_time is initialized before use */
158 rq
->cmd_flags
|= rw_flags
;
159 /* do not touch atomic flags, it needs atomic ops against the timer */
161 INIT_HLIST_NODE(&rq
->hash
);
162 RB_CLEAR_NODE(&rq
->rb_node
);
165 rq
->start_time
= jiffies
;
166 #ifdef CONFIG_BLK_CGROUP
168 set_start_time_ns(rq
);
169 rq
->io_start_time_ns
= 0;
171 rq
->nr_phys_segments
= 0;
172 #if defined(CONFIG_BLK_DEV_INTEGRITY)
173 rq
->nr_integrity_segments
= 0;
176 /* tag was already set */
184 INIT_LIST_HEAD(&rq
->timeout_list
);
188 rq
->end_io_data
= NULL
;
191 ctx
->rq_dispatched
[rw_is_sync(rw_flags
)]++;
194 static struct request
*
195 __blk_mq_alloc_request(struct blk_mq_alloc_data
*data
, int rw
)
200 tag
= blk_mq_get_tag(data
);
201 if (tag
!= BLK_MQ_TAG_FAIL
) {
202 rq
= data
->hctx
->tags
->rqs
[tag
];
205 if (blk_mq_tag_busy(data
->hctx
)) {
206 rq
->cmd_flags
= REQ_MQ_INFLIGHT
;
207 atomic_inc(&data
->hctx
->nr_active
);
211 blk_mq_rq_ctx_init(data
->q
, data
->ctx
, rq
, rw
);
218 struct request
*blk_mq_alloc_request(struct request_queue
*q
, int rw
, gfp_t gfp
,
221 struct blk_mq_ctx
*ctx
;
222 struct blk_mq_hw_ctx
*hctx
;
224 struct blk_mq_alloc_data alloc_data
;
226 if (blk_mq_queue_enter(q
))
229 ctx
= blk_mq_get_ctx(q
);
230 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
231 blk_mq_set_alloc_data(&alloc_data
, q
, gfp
& ~__GFP_WAIT
,
232 reserved
, ctx
, hctx
);
234 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
235 if (!rq
&& (gfp
& __GFP_WAIT
)) {
236 __blk_mq_run_hw_queue(hctx
);
239 ctx
= blk_mq_get_ctx(q
);
240 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
241 blk_mq_set_alloc_data(&alloc_data
, q
, gfp
, reserved
, ctx
,
243 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
244 ctx
= alloc_data
.ctx
;
249 EXPORT_SYMBOL(blk_mq_alloc_request
);
251 static void __blk_mq_free_request(struct blk_mq_hw_ctx
*hctx
,
252 struct blk_mq_ctx
*ctx
, struct request
*rq
)
254 const int tag
= rq
->tag
;
255 struct request_queue
*q
= rq
->q
;
257 if (rq
->cmd_flags
& REQ_MQ_INFLIGHT
)
258 atomic_dec(&hctx
->nr_active
);
260 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
261 blk_mq_put_tag(hctx
, tag
, &ctx
->last_tag
);
262 blk_mq_queue_exit(q
);
265 void blk_mq_free_request(struct request
*rq
)
267 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
268 struct blk_mq_hw_ctx
*hctx
;
269 struct request_queue
*q
= rq
->q
;
271 ctx
->rq_completed
[rq_is_sync(rq
)]++;
273 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
274 __blk_mq_free_request(hctx
, ctx
, rq
);
278 * Clone all relevant state from a request that has been put on hold in
279 * the flush state machine into the preallocated flush request that hangs
280 * off the request queue.
282 * For a driver the flush request should be invisible, that's why we are
283 * impersonating the original request here.
285 void blk_mq_clone_flush_request(struct request
*flush_rq
,
286 struct request
*orig_rq
)
288 struct blk_mq_hw_ctx
*hctx
=
289 orig_rq
->q
->mq_ops
->map_queue(orig_rq
->q
, orig_rq
->mq_ctx
->cpu
);
291 flush_rq
->mq_ctx
= orig_rq
->mq_ctx
;
292 flush_rq
->tag
= orig_rq
->tag
;
293 memcpy(blk_mq_rq_to_pdu(flush_rq
), blk_mq_rq_to_pdu(orig_rq
),
297 inline void __blk_mq_end_io(struct request
*rq
, int error
)
299 blk_account_io_done(rq
);
302 rq
->end_io(rq
, error
);
304 if (unlikely(blk_bidi_rq(rq
)))
305 blk_mq_free_request(rq
->next_rq
);
306 blk_mq_free_request(rq
);
309 EXPORT_SYMBOL(__blk_mq_end_io
);
311 void blk_mq_end_io(struct request
*rq
, int error
)
313 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
315 __blk_mq_end_io(rq
, error
);
317 EXPORT_SYMBOL(blk_mq_end_io
);
319 static void __blk_mq_complete_request_remote(void *data
)
321 struct request
*rq
= data
;
323 rq
->q
->softirq_done_fn(rq
);
326 static void blk_mq_ipi_complete_request(struct request
*rq
)
328 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
332 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
333 rq
->q
->softirq_done_fn(rq
);
338 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
339 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
341 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
342 rq
->csd
.func
= __blk_mq_complete_request_remote
;
345 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
347 rq
->q
->softirq_done_fn(rq
);
352 void __blk_mq_complete_request(struct request
*rq
)
354 struct request_queue
*q
= rq
->q
;
356 if (!q
->softirq_done_fn
)
357 blk_mq_end_io(rq
, rq
->errors
);
359 blk_mq_ipi_complete_request(rq
);
363 * blk_mq_complete_request - end I/O on a request
364 * @rq: the request being processed
367 * Ends all I/O on a request. It does not handle partial completions.
368 * The actual completion happens out-of-order, through a IPI handler.
370 void blk_mq_complete_request(struct request
*rq
)
372 struct request_queue
*q
= rq
->q
;
374 if (unlikely(blk_should_fake_timeout(q
)))
376 if (!blk_mark_rq_complete(rq
))
377 __blk_mq_complete_request(rq
);
379 EXPORT_SYMBOL(blk_mq_complete_request
);
381 static void blk_mq_start_request(struct request
*rq
, bool last
)
383 struct request_queue
*q
= rq
->q
;
385 trace_block_rq_issue(q
, rq
);
387 rq
->resid_len
= blk_rq_bytes(rq
);
388 if (unlikely(blk_bidi_rq(rq
)))
389 rq
->next_rq
->resid_len
= blk_rq_bytes(rq
->next_rq
);
394 * Mark us as started and clear complete. Complete might have been
395 * set if requeue raced with timeout, which then marked it as
396 * complete. So be sure to clear complete again when we start
397 * the request, otherwise we'll ignore the completion event.
399 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
400 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
401 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
402 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
404 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
406 * Make sure space for the drain appears. We know we can do
407 * this because max_hw_segments has been adjusted to be one
408 * fewer than the device can handle.
410 rq
->nr_phys_segments
++;
414 * Flag the last request in the series so that drivers know when IO
415 * should be kicked off, if they don't do it on a per-request basis.
417 * Note: the flag isn't the only condition drivers should do kick off.
418 * If drive is busy, the last request might not have the bit set.
421 rq
->cmd_flags
|= REQ_END
;
424 static void __blk_mq_requeue_request(struct request
*rq
)
426 struct request_queue
*q
= rq
->q
;
428 trace_block_rq_requeue(q
, rq
);
429 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
431 rq
->cmd_flags
&= ~REQ_END
;
433 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
434 rq
->nr_phys_segments
--;
437 void blk_mq_requeue_request(struct request
*rq
)
439 __blk_mq_requeue_request(rq
);
440 blk_clear_rq_complete(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);
474 blk_mq_run_queues(q
, false);
477 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
)
479 struct request_queue
*q
= rq
->q
;
483 * We abuse this flag that is otherwise used by the I/O scheduler to
484 * request head insertation from the workqueue.
486 BUG_ON(rq
->cmd_flags
& REQ_SOFTBARRIER
);
488 spin_lock_irqsave(&q
->requeue_lock
, flags
);
490 rq
->cmd_flags
|= REQ_SOFTBARRIER
;
491 list_add(&rq
->queuelist
, &q
->requeue_list
);
493 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
495 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
497 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
499 void blk_mq_kick_requeue_list(struct request_queue
*q
)
501 kblockd_schedule_work(&q
->requeue_work
);
503 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
505 static inline bool is_flush_request(struct request
*rq
, unsigned int tag
)
507 return ((rq
->cmd_flags
& REQ_FLUSH_SEQ
) &&
508 rq
->q
->flush_rq
->tag
== tag
);
511 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
513 struct request
*rq
= tags
->rqs
[tag
];
515 if (!is_flush_request(rq
, tag
))
518 return rq
->q
->flush_rq
;
520 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
522 struct blk_mq_timeout_data
{
523 struct blk_mq_hw_ctx
*hctx
;
525 unsigned int *next_set
;
528 static void blk_mq_timeout_check(void *__data
, unsigned long *free_tags
)
530 struct blk_mq_timeout_data
*data
= __data
;
531 struct blk_mq_hw_ctx
*hctx
= data
->hctx
;
534 /* It may not be in flight yet (this is where
535 * the REQ_ATOMIC_STARTED flag comes in). The requests are
536 * statically allocated, so we know it's always safe to access the
537 * memory associated with a bit offset into ->rqs[].
543 tag
= find_next_zero_bit(free_tags
, hctx
->tags
->nr_tags
, tag
);
544 if (tag
>= hctx
->tags
->nr_tags
)
547 rq
= blk_mq_tag_to_rq(hctx
->tags
, tag
++);
548 if (rq
->q
!= hctx
->queue
)
550 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
553 blk_rq_check_expired(rq
, data
->next
, data
->next_set
);
557 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx
*hctx
,
559 unsigned int *next_set
)
561 struct blk_mq_timeout_data data
= {
564 .next_set
= next_set
,
568 * Ask the tagging code to iterate busy requests, so we can
569 * check them for timeout.
571 blk_mq_tag_busy_iter(hctx
->tags
, blk_mq_timeout_check
, &data
);
574 static enum blk_eh_timer_return
blk_mq_rq_timed_out(struct request
*rq
)
576 struct request_queue
*q
= rq
->q
;
579 * We know that complete is set at this point. If STARTED isn't set
580 * anymore, then the request isn't active and the "timeout" should
581 * just be ignored. This can happen due to the bitflag ordering.
582 * Timeout first checks if STARTED is set, and if it is, assumes
583 * the request is active. But if we race with completion, then
584 * we both flags will get cleared. So check here again, and ignore
585 * a timeout event with a request that isn't active.
587 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
588 return BLK_EH_NOT_HANDLED
;
590 if (!q
->mq_ops
->timeout
)
591 return BLK_EH_RESET_TIMER
;
593 return q
->mq_ops
->timeout(rq
);
596 static void blk_mq_rq_timer(unsigned long data
)
598 struct request_queue
*q
= (struct request_queue
*) data
;
599 struct blk_mq_hw_ctx
*hctx
;
600 unsigned long next
= 0;
603 queue_for_each_hw_ctx(q
, hctx
, i
) {
605 * If not software queues are currently mapped to this
606 * hardware queue, there's nothing to check
608 if (!hctx
->nr_ctx
|| !hctx
->tags
)
611 blk_mq_hw_ctx_check_timeout(hctx
, &next
, &next_set
);
615 next
= blk_rq_timeout(round_jiffies_up(next
));
616 mod_timer(&q
->timeout
, next
);
618 queue_for_each_hw_ctx(q
, hctx
, i
)
619 blk_mq_tag_idle(hctx
);
624 * Reverse check our software queue for entries that we could potentially
625 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
626 * too much time checking for merges.
628 static bool blk_mq_attempt_merge(struct request_queue
*q
,
629 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
634 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
640 if (!blk_rq_merge_ok(rq
, bio
))
643 el_ret
= blk_try_merge(rq
, bio
);
644 if (el_ret
== ELEVATOR_BACK_MERGE
) {
645 if (bio_attempt_back_merge(q
, rq
, bio
)) {
650 } else if (el_ret
== ELEVATOR_FRONT_MERGE
) {
651 if (bio_attempt_front_merge(q
, rq
, bio
)) {
663 * Process software queues that have been marked busy, splicing them
664 * to the for-dispatch
666 static void flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
668 struct blk_mq_ctx
*ctx
;
671 for (i
= 0; i
< hctx
->ctx_map
.map_size
; i
++) {
672 struct blk_align_bitmap
*bm
= &hctx
->ctx_map
.map
[i
];
673 unsigned int off
, bit
;
679 off
= i
* hctx
->ctx_map
.bits_per_word
;
681 bit
= find_next_bit(&bm
->word
, bm
->depth
, bit
);
682 if (bit
>= bm
->depth
)
685 ctx
= hctx
->ctxs
[bit
+ off
];
686 clear_bit(bit
, &bm
->word
);
687 spin_lock(&ctx
->lock
);
688 list_splice_tail_init(&ctx
->rq_list
, list
);
689 spin_unlock(&ctx
->lock
);
697 * Run this hardware queue, pulling any software queues mapped to it in.
698 * Note that this function currently has various problems around ordering
699 * of IO. In particular, we'd like FIFO behaviour on handling existing
700 * items on the hctx->dispatch list. Ignore that for now.
702 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
704 struct request_queue
*q
= hctx
->queue
;
709 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
));
711 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)))
717 * Touch any software queue that has pending entries.
719 flush_busy_ctxs(hctx
, &rq_list
);
722 * If we have previous entries on our dispatch list, grab them
723 * and stuff them at the front for more fair dispatch.
725 if (!list_empty_careful(&hctx
->dispatch
)) {
726 spin_lock(&hctx
->lock
);
727 if (!list_empty(&hctx
->dispatch
))
728 list_splice_init(&hctx
->dispatch
, &rq_list
);
729 spin_unlock(&hctx
->lock
);
733 * Now process all the entries, sending them to the driver.
736 while (!list_empty(&rq_list
)) {
739 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
740 list_del_init(&rq
->queuelist
);
742 blk_mq_start_request(rq
, list_empty(&rq_list
));
744 ret
= q
->mq_ops
->queue_rq(hctx
, rq
);
746 case BLK_MQ_RQ_QUEUE_OK
:
749 case BLK_MQ_RQ_QUEUE_BUSY
:
750 list_add(&rq
->queuelist
, &rq_list
);
751 __blk_mq_requeue_request(rq
);
754 pr_err("blk-mq: bad return on queue: %d\n", ret
);
755 case BLK_MQ_RQ_QUEUE_ERROR
:
757 blk_mq_end_io(rq
, rq
->errors
);
761 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
766 hctx
->dispatched
[0]++;
767 else if (queued
< (1 << (BLK_MQ_MAX_DISPATCH_ORDER
- 1)))
768 hctx
->dispatched
[ilog2(queued
) + 1]++;
771 * Any items that need requeuing? Stuff them into hctx->dispatch,
772 * that is where we will continue on next queue run.
774 if (!list_empty(&rq_list
)) {
775 spin_lock(&hctx
->lock
);
776 list_splice(&rq_list
, &hctx
->dispatch
);
777 spin_unlock(&hctx
->lock
);
782 * It'd be great if the workqueue API had a way to pass
783 * in a mask and had some smarts for more clever placement.
784 * For now we just round-robin here, switching for every
785 * BLK_MQ_CPU_WORK_BATCH queued items.
787 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
789 int cpu
= hctx
->next_cpu
;
791 if (--hctx
->next_cpu_batch
<= 0) {
794 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
795 if (next_cpu
>= nr_cpu_ids
)
796 next_cpu
= cpumask_first(hctx
->cpumask
);
798 hctx
->next_cpu
= next_cpu
;
799 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
805 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
807 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)))
810 if (!async
&& cpumask_test_cpu(smp_processor_id(), hctx
->cpumask
))
811 __blk_mq_run_hw_queue(hctx
);
812 else if (hctx
->queue
->nr_hw_queues
== 1)
813 kblockd_schedule_delayed_work(&hctx
->run_work
, 0);
817 cpu
= blk_mq_hctx_next_cpu(hctx
);
818 kblockd_schedule_delayed_work_on(cpu
, &hctx
->run_work
, 0);
822 void blk_mq_run_queues(struct request_queue
*q
, bool async
)
824 struct blk_mq_hw_ctx
*hctx
;
827 queue_for_each_hw_ctx(q
, hctx
, i
) {
828 if ((!blk_mq_hctx_has_pending(hctx
) &&
829 list_empty_careful(&hctx
->dispatch
)) ||
830 test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
834 blk_mq_run_hw_queue(hctx
, async
);
838 EXPORT_SYMBOL(blk_mq_run_queues
);
840 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
842 cancel_delayed_work(&hctx
->run_work
);
843 cancel_delayed_work(&hctx
->delay_work
);
844 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
846 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
848 void blk_mq_stop_hw_queues(struct request_queue
*q
)
850 struct blk_mq_hw_ctx
*hctx
;
853 queue_for_each_hw_ctx(q
, hctx
, i
)
854 blk_mq_stop_hw_queue(hctx
);
856 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
858 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
860 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
863 blk_mq_run_hw_queue(hctx
, false);
866 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
868 void blk_mq_start_hw_queues(struct request_queue
*q
)
870 struct blk_mq_hw_ctx
*hctx
;
873 queue_for_each_hw_ctx(q
, hctx
, i
)
874 blk_mq_start_hw_queue(hctx
);
876 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
879 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
881 struct blk_mq_hw_ctx
*hctx
;
884 queue_for_each_hw_ctx(q
, hctx
, i
) {
885 if (!test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
888 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
890 blk_mq_run_hw_queue(hctx
, async
);
894 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
896 static void blk_mq_run_work_fn(struct work_struct
*work
)
898 struct blk_mq_hw_ctx
*hctx
;
900 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
902 __blk_mq_run_hw_queue(hctx
);
905 static void blk_mq_delay_work_fn(struct work_struct
*work
)
907 struct blk_mq_hw_ctx
*hctx
;
909 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
911 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
912 __blk_mq_run_hw_queue(hctx
);
915 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
917 unsigned long tmo
= msecs_to_jiffies(msecs
);
919 if (hctx
->queue
->nr_hw_queues
== 1)
920 kblockd_schedule_delayed_work(&hctx
->delay_work
, tmo
);
924 cpu
= blk_mq_hctx_next_cpu(hctx
);
925 kblockd_schedule_delayed_work_on(cpu
, &hctx
->delay_work
, tmo
);
928 EXPORT_SYMBOL(blk_mq_delay_queue
);
930 static void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
,
931 struct request
*rq
, bool at_head
)
933 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
935 trace_block_rq_insert(hctx
->queue
, rq
);
938 list_add(&rq
->queuelist
, &ctx
->rq_list
);
940 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
942 blk_mq_hctx_mark_pending(hctx
, ctx
);
945 void blk_mq_insert_request(struct request
*rq
, bool at_head
, bool run_queue
,
948 struct request_queue
*q
= rq
->q
;
949 struct blk_mq_hw_ctx
*hctx
;
950 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
, *current_ctx
;
952 current_ctx
= blk_mq_get_ctx(q
);
953 if (!cpu_online(ctx
->cpu
))
954 rq
->mq_ctx
= ctx
= current_ctx
;
956 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
958 if (rq
->cmd_flags
& (REQ_FLUSH
| REQ_FUA
) &&
959 !(rq
->cmd_flags
& (REQ_FLUSH_SEQ
))) {
960 blk_insert_flush(rq
);
962 spin_lock(&ctx
->lock
);
963 __blk_mq_insert_request(hctx
, rq
, at_head
);
964 spin_unlock(&ctx
->lock
);
968 blk_mq_run_hw_queue(hctx
, async
);
970 blk_mq_put_ctx(current_ctx
);
973 static void blk_mq_insert_requests(struct request_queue
*q
,
974 struct blk_mq_ctx
*ctx
,
975 struct list_head
*list
,
980 struct blk_mq_hw_ctx
*hctx
;
981 struct blk_mq_ctx
*current_ctx
;
983 trace_block_unplug(q
, depth
, !from_schedule
);
985 current_ctx
= blk_mq_get_ctx(q
);
987 if (!cpu_online(ctx
->cpu
))
989 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
992 * preemption doesn't flush plug list, so it's possible ctx->cpu is
995 spin_lock(&ctx
->lock
);
996 while (!list_empty(list
)) {
999 rq
= list_first_entry(list
, struct request
, queuelist
);
1000 list_del_init(&rq
->queuelist
);
1002 __blk_mq_insert_request(hctx
, rq
, false);
1004 spin_unlock(&ctx
->lock
);
1006 blk_mq_run_hw_queue(hctx
, from_schedule
);
1007 blk_mq_put_ctx(current_ctx
);
1010 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1012 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1013 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1015 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1016 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1017 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1020 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1022 struct blk_mq_ctx
*this_ctx
;
1023 struct request_queue
*this_q
;
1026 LIST_HEAD(ctx_list
);
1029 list_splice_init(&plug
->mq_list
, &list
);
1031 list_sort(NULL
, &list
, plug_ctx_cmp
);
1037 while (!list_empty(&list
)) {
1038 rq
= list_entry_rq(list
.next
);
1039 list_del_init(&rq
->queuelist
);
1041 if (rq
->mq_ctx
!= this_ctx
) {
1043 blk_mq_insert_requests(this_q
, this_ctx
,
1048 this_ctx
= rq
->mq_ctx
;
1054 list_add_tail(&rq
->queuelist
, &ctx_list
);
1058 * If 'this_ctx' is set, we know we have entries to complete
1059 * on 'ctx_list'. Do those.
1062 blk_mq_insert_requests(this_q
, this_ctx
, &ctx_list
, depth
,
1067 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1069 init_request_from_bio(rq
, bio
);
1071 if (blk_do_io_stat(rq
))
1072 blk_account_io_start(rq
, 1);
1075 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1077 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1078 !blk_queue_nomerges(hctx
->queue
);
1081 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx
*hctx
,
1082 struct blk_mq_ctx
*ctx
,
1083 struct request
*rq
, struct bio
*bio
)
1085 if (!hctx_allow_merges(hctx
)) {
1086 blk_mq_bio_to_request(rq
, bio
);
1087 spin_lock(&ctx
->lock
);
1089 __blk_mq_insert_request(hctx
, rq
, false);
1090 spin_unlock(&ctx
->lock
);
1093 struct request_queue
*q
= hctx
->queue
;
1095 spin_lock(&ctx
->lock
);
1096 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1097 blk_mq_bio_to_request(rq
, bio
);
1101 spin_unlock(&ctx
->lock
);
1102 __blk_mq_free_request(hctx
, ctx
, rq
);
1107 struct blk_map_ctx
{
1108 struct blk_mq_hw_ctx
*hctx
;
1109 struct blk_mq_ctx
*ctx
;
1112 static struct request
*blk_mq_map_request(struct request_queue
*q
,
1114 struct blk_map_ctx
*data
)
1116 struct blk_mq_hw_ctx
*hctx
;
1117 struct blk_mq_ctx
*ctx
;
1119 int rw
= bio_data_dir(bio
);
1120 struct blk_mq_alloc_data alloc_data
;
1122 if (unlikely(blk_mq_queue_enter(q
))) {
1123 bio_endio(bio
, -EIO
);
1127 ctx
= blk_mq_get_ctx(q
);
1128 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1130 if (rw_is_sync(bio
->bi_rw
))
1133 trace_block_getrq(q
, bio
, rw
);
1134 blk_mq_set_alloc_data(&alloc_data
, q
, GFP_ATOMIC
, false, ctx
,
1136 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
1137 if (unlikely(!rq
)) {
1138 __blk_mq_run_hw_queue(hctx
);
1139 blk_mq_put_ctx(ctx
);
1140 trace_block_sleeprq(q
, bio
, rw
);
1142 ctx
= blk_mq_get_ctx(q
);
1143 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1144 blk_mq_set_alloc_data(&alloc_data
, q
,
1145 __GFP_WAIT
|GFP_ATOMIC
, false, ctx
, hctx
);
1146 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
1147 ctx
= alloc_data
.ctx
;
1148 hctx
= alloc_data
.hctx
;
1158 * Multiple hardware queue variant. This will not use per-process plugs,
1159 * but will attempt to bypass the hctx queueing if we can go straight to
1160 * hardware for SYNC IO.
1162 static void blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1164 const int is_sync
= rw_is_sync(bio
->bi_rw
);
1165 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
1166 struct blk_map_ctx data
;
1169 blk_queue_bounce(q
, &bio
);
1171 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1172 bio_endio(bio
, -EIO
);
1176 rq
= blk_mq_map_request(q
, bio
, &data
);
1180 if (unlikely(is_flush_fua
)) {
1181 blk_mq_bio_to_request(rq
, bio
);
1182 blk_insert_flush(rq
);
1189 blk_mq_bio_to_request(rq
, bio
);
1190 blk_mq_start_request(rq
, true);
1193 * For OK queue, we are done. For error, kill it. Any other
1194 * error (busy), just add it to our list as we previously
1197 ret
= q
->mq_ops
->queue_rq(data
.hctx
, rq
);
1198 if (ret
== BLK_MQ_RQ_QUEUE_OK
)
1201 __blk_mq_requeue_request(rq
);
1203 if (ret
== BLK_MQ_RQ_QUEUE_ERROR
) {
1205 blk_mq_end_io(rq
, rq
->errors
);
1211 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1213 * For a SYNC request, send it to the hardware immediately. For
1214 * an ASYNC request, just ensure that we run it later on. The
1215 * latter allows for merging opportunities and more efficient
1219 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1222 blk_mq_put_ctx(data
.ctx
);
1226 * Single hardware queue variant. This will attempt to use any per-process
1227 * plug for merging and IO deferral.
1229 static void blk_sq_make_request(struct request_queue
*q
, struct bio
*bio
)
1231 const int is_sync
= rw_is_sync(bio
->bi_rw
);
1232 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
1233 unsigned int use_plug
, request_count
= 0;
1234 struct blk_map_ctx data
;
1238 * If we have multiple hardware queues, just go directly to
1239 * one of those for sync IO.
1241 use_plug
= !is_flush_fua
&& !is_sync
;
1243 blk_queue_bounce(q
, &bio
);
1245 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1246 bio_endio(bio
, -EIO
);
1250 if (use_plug
&& !blk_queue_nomerges(q
) &&
1251 blk_attempt_plug_merge(q
, bio
, &request_count
))
1254 rq
= blk_mq_map_request(q
, bio
, &data
);
1258 if (unlikely(is_flush_fua
)) {
1259 blk_mq_bio_to_request(rq
, bio
);
1260 blk_insert_flush(rq
);
1265 * A task plug currently exists. Since this is completely lockless,
1266 * utilize that to temporarily store requests until the task is
1267 * either done or scheduled away.
1270 struct blk_plug
*plug
= current
->plug
;
1273 blk_mq_bio_to_request(rq
, bio
);
1274 if (list_empty(&plug
->mq_list
))
1275 trace_block_plug(q
);
1276 else if (request_count
>= BLK_MAX_REQUEST_COUNT
) {
1277 blk_flush_plug_list(plug
, false);
1278 trace_block_plug(q
);
1280 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1281 blk_mq_put_ctx(data
.ctx
);
1286 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1288 * For a SYNC request, send it to the hardware immediately. For
1289 * an ASYNC request, just ensure that we run it later on. The
1290 * latter allows for merging opportunities and more efficient
1294 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1297 blk_mq_put_ctx(data
.ctx
);
1301 * Default mapping to a software queue, since we use one per CPU.
1303 struct blk_mq_hw_ctx
*blk_mq_map_queue(struct request_queue
*q
, const int cpu
)
1305 return q
->queue_hw_ctx
[q
->mq_map
[cpu
]];
1307 EXPORT_SYMBOL(blk_mq_map_queue
);
1309 static void blk_mq_free_rq_map(struct blk_mq_tag_set
*set
,
1310 struct blk_mq_tags
*tags
, unsigned int hctx_idx
)
1314 if (tags
->rqs
&& set
->ops
->exit_request
) {
1317 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1320 set
->ops
->exit_request(set
->driver_data
, tags
->rqs
[i
],
1325 while (!list_empty(&tags
->page_list
)) {
1326 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1327 list_del_init(&page
->lru
);
1328 __free_pages(page
, page
->private);
1333 blk_mq_free_tags(tags
);
1336 static size_t order_to_size(unsigned int order
)
1338 return (size_t)PAGE_SIZE
<< order
;
1341 static struct blk_mq_tags
*blk_mq_init_rq_map(struct blk_mq_tag_set
*set
,
1342 unsigned int hctx_idx
)
1344 struct blk_mq_tags
*tags
;
1345 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1346 size_t rq_size
, left
;
1348 tags
= blk_mq_init_tags(set
->queue_depth
, set
->reserved_tags
,
1353 INIT_LIST_HEAD(&tags
->page_list
);
1355 tags
->rqs
= kmalloc_node(set
->queue_depth
* sizeof(struct request
*),
1356 GFP_KERNEL
, set
->numa_node
);
1358 blk_mq_free_tags(tags
);
1363 * rq_size is the size of the request plus driver payload, rounded
1364 * to the cacheline size
1366 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1368 left
= rq_size
* set
->queue_depth
;
1370 for (i
= 0; i
< set
->queue_depth
; ) {
1371 int this_order
= max_order
;
1376 while (left
< order_to_size(this_order
- 1) && this_order
)
1380 page
= alloc_pages_node(set
->numa_node
, GFP_KERNEL
,
1386 if (order_to_size(this_order
) < rq_size
)
1393 page
->private = this_order
;
1394 list_add_tail(&page
->lru
, &tags
->page_list
);
1396 p
= page_address(page
);
1397 entries_per_page
= order_to_size(this_order
) / rq_size
;
1398 to_do
= min(entries_per_page
, set
->queue_depth
- i
);
1399 left
-= to_do
* rq_size
;
1400 for (j
= 0; j
< to_do
; j
++) {
1402 if (set
->ops
->init_request
) {
1403 if (set
->ops
->init_request(set
->driver_data
,
1404 tags
->rqs
[i
], hctx_idx
, i
,
1417 pr_warn("%s: failed to allocate requests\n", __func__
);
1418 blk_mq_free_rq_map(set
, tags
, hctx_idx
);
1422 static void blk_mq_free_bitmap(struct blk_mq_ctxmap
*bitmap
)
1427 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap
*bitmap
, int node
)
1429 unsigned int bpw
= 8, total
, num_maps
, i
;
1431 bitmap
->bits_per_word
= bpw
;
1433 num_maps
= ALIGN(nr_cpu_ids
, bpw
) / bpw
;
1434 bitmap
->map
= kzalloc_node(num_maps
* sizeof(struct blk_align_bitmap
),
1439 bitmap
->map_size
= num_maps
;
1442 for (i
= 0; i
< num_maps
; i
++) {
1443 bitmap
->map
[i
].depth
= min(total
, bitmap
->bits_per_word
);
1444 total
-= bitmap
->map
[i
].depth
;
1450 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx
*hctx
, int cpu
)
1452 struct request_queue
*q
= hctx
->queue
;
1453 struct blk_mq_ctx
*ctx
;
1457 * Move ctx entries to new CPU, if this one is going away.
1459 ctx
= __blk_mq_get_ctx(q
, cpu
);
1461 spin_lock(&ctx
->lock
);
1462 if (!list_empty(&ctx
->rq_list
)) {
1463 list_splice_init(&ctx
->rq_list
, &tmp
);
1464 blk_mq_hctx_clear_pending(hctx
, ctx
);
1466 spin_unlock(&ctx
->lock
);
1468 if (list_empty(&tmp
))
1471 ctx
= blk_mq_get_ctx(q
);
1472 spin_lock(&ctx
->lock
);
1474 while (!list_empty(&tmp
)) {
1477 rq
= list_first_entry(&tmp
, struct request
, queuelist
);
1479 list_move_tail(&rq
->queuelist
, &ctx
->rq_list
);
1482 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1483 blk_mq_hctx_mark_pending(hctx
, ctx
);
1485 spin_unlock(&ctx
->lock
);
1487 blk_mq_run_hw_queue(hctx
, true);
1488 blk_mq_put_ctx(ctx
);
1492 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx
*hctx
, int cpu
)
1494 struct request_queue
*q
= hctx
->queue
;
1495 struct blk_mq_tag_set
*set
= q
->tag_set
;
1497 if (set
->tags
[hctx
->queue_num
])
1500 set
->tags
[hctx
->queue_num
] = blk_mq_init_rq_map(set
, hctx
->queue_num
);
1501 if (!set
->tags
[hctx
->queue_num
])
1504 hctx
->tags
= set
->tags
[hctx
->queue_num
];
1508 static int blk_mq_hctx_notify(void *data
, unsigned long action
,
1511 struct blk_mq_hw_ctx
*hctx
= data
;
1513 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
)
1514 return blk_mq_hctx_cpu_offline(hctx
, cpu
);
1515 else if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
)
1516 return blk_mq_hctx_cpu_online(hctx
, cpu
);
1521 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1522 struct blk_mq_tag_set
*set
, int nr_queue
)
1524 struct blk_mq_hw_ctx
*hctx
;
1527 queue_for_each_hw_ctx(q
, hctx
, i
) {
1531 blk_mq_tag_idle(hctx
);
1533 if (set
->ops
->exit_hctx
)
1534 set
->ops
->exit_hctx(hctx
, i
);
1536 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1538 blk_mq_free_bitmap(&hctx
->ctx_map
);
1543 static void blk_mq_free_hw_queues(struct request_queue
*q
,
1544 struct blk_mq_tag_set
*set
)
1546 struct blk_mq_hw_ctx
*hctx
;
1549 queue_for_each_hw_ctx(q
, hctx
, i
) {
1550 free_cpumask_var(hctx
->cpumask
);
1555 static int blk_mq_init_hw_queues(struct request_queue
*q
,
1556 struct blk_mq_tag_set
*set
)
1558 struct blk_mq_hw_ctx
*hctx
;
1562 * Initialize hardware queues
1564 queue_for_each_hw_ctx(q
, hctx
, i
) {
1567 node
= hctx
->numa_node
;
1568 if (node
== NUMA_NO_NODE
)
1569 node
= hctx
->numa_node
= set
->numa_node
;
1571 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1572 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1573 spin_lock_init(&hctx
->lock
);
1574 INIT_LIST_HEAD(&hctx
->dispatch
);
1576 hctx
->queue_num
= i
;
1577 hctx
->flags
= set
->flags
;
1578 hctx
->cmd_size
= set
->cmd_size
;
1580 blk_mq_init_cpu_notifier(&hctx
->cpu_notifier
,
1581 blk_mq_hctx_notify
, hctx
);
1582 blk_mq_register_cpu_notifier(&hctx
->cpu_notifier
);
1584 hctx
->tags
= set
->tags
[i
];
1587 * Allocate space for all possible cpus to avoid allocation at
1590 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1595 if (blk_mq_alloc_bitmap(&hctx
->ctx_map
, node
))
1600 if (set
->ops
->init_hctx
&&
1601 set
->ops
->init_hctx(hctx
, set
->driver_data
, i
))
1605 if (i
== q
->nr_hw_queues
)
1611 blk_mq_exit_hw_queues(q
, set
, i
);
1616 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1617 unsigned int nr_hw_queues
)
1621 for_each_possible_cpu(i
) {
1622 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1623 struct blk_mq_hw_ctx
*hctx
;
1625 memset(__ctx
, 0, sizeof(*__ctx
));
1627 spin_lock_init(&__ctx
->lock
);
1628 INIT_LIST_HEAD(&__ctx
->rq_list
);
1631 /* If the cpu isn't online, the cpu is mapped to first hctx */
1635 hctx
= q
->mq_ops
->map_queue(q
, i
);
1636 cpumask_set_cpu(i
, hctx
->cpumask
);
1640 * Set local node, IFF we have more than one hw queue. If
1641 * not, we remain on the home node of the device
1643 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1644 hctx
->numa_node
= cpu_to_node(i
);
1648 static void blk_mq_map_swqueue(struct request_queue
*q
)
1651 struct blk_mq_hw_ctx
*hctx
;
1652 struct blk_mq_ctx
*ctx
;
1654 queue_for_each_hw_ctx(q
, hctx
, i
) {
1655 cpumask_clear(hctx
->cpumask
);
1660 * Map software to hardware queues
1662 queue_for_each_ctx(q
, ctx
, i
) {
1663 /* If the cpu isn't online, the cpu is mapped to first hctx */
1667 hctx
= q
->mq_ops
->map_queue(q
, i
);
1668 cpumask_set_cpu(i
, hctx
->cpumask
);
1669 ctx
->index_hw
= hctx
->nr_ctx
;
1670 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
1673 queue_for_each_hw_ctx(q
, hctx
, i
) {
1675 * If no software queues are mapped to this hardware queue,
1676 * disable it and free the request entries.
1678 if (!hctx
->nr_ctx
) {
1679 struct blk_mq_tag_set
*set
= q
->tag_set
;
1682 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1683 set
->tags
[i
] = NULL
;
1690 * Initialize batch roundrobin counts
1692 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
1693 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1697 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
)
1699 struct blk_mq_hw_ctx
*hctx
;
1700 struct request_queue
*q
;
1704 if (set
->tag_list
.next
== set
->tag_list
.prev
)
1709 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
1710 blk_mq_freeze_queue(q
);
1712 queue_for_each_hw_ctx(q
, hctx
, i
) {
1714 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
1716 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
1718 blk_mq_unfreeze_queue(q
);
1722 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
1724 struct blk_mq_tag_set
*set
= q
->tag_set
;
1726 mutex_lock(&set
->tag_list_lock
);
1727 list_del_init(&q
->tag_set_list
);
1728 blk_mq_update_tag_set_depth(set
);
1729 mutex_unlock(&set
->tag_list_lock
);
1732 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
1733 struct request_queue
*q
)
1737 mutex_lock(&set
->tag_list_lock
);
1738 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
1739 blk_mq_update_tag_set_depth(set
);
1740 mutex_unlock(&set
->tag_list_lock
);
1743 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
1745 struct blk_mq_hw_ctx
**hctxs
;
1746 struct blk_mq_ctx __percpu
*ctx
;
1747 struct request_queue
*q
;
1751 ctx
= alloc_percpu(struct blk_mq_ctx
);
1753 return ERR_PTR(-ENOMEM
);
1755 hctxs
= kmalloc_node(set
->nr_hw_queues
* sizeof(*hctxs
), GFP_KERNEL
,
1761 map
= blk_mq_make_queue_map(set
);
1765 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1766 int node
= blk_mq_hw_queue_to_node(map
, i
);
1768 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
1773 if (!zalloc_cpumask_var(&hctxs
[i
]->cpumask
, GFP_KERNEL
))
1776 atomic_set(&hctxs
[i
]->nr_active
, 0);
1777 hctxs
[i
]->numa_node
= node
;
1778 hctxs
[i
]->queue_num
= i
;
1781 q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
1785 if (percpu_ref_init(&q
->mq_usage_counter
, blk_mq_usage_counter_release
))
1788 setup_timer(&q
->timeout
, blk_mq_rq_timer
, (unsigned long) q
);
1789 blk_queue_rq_timeout(q
, 30000);
1791 q
->nr_queues
= nr_cpu_ids
;
1792 q
->nr_hw_queues
= set
->nr_hw_queues
;
1796 q
->queue_hw_ctx
= hctxs
;
1798 q
->mq_ops
= set
->ops
;
1799 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
1801 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
1802 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
1804 q
->sg_reserved_size
= INT_MAX
;
1806 INIT_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
1807 INIT_LIST_HEAD(&q
->requeue_list
);
1808 spin_lock_init(&q
->requeue_lock
);
1810 if (q
->nr_hw_queues
> 1)
1811 blk_queue_make_request(q
, blk_mq_make_request
);
1813 blk_queue_make_request(q
, blk_sq_make_request
);
1815 blk_queue_rq_timed_out(q
, blk_mq_rq_timed_out
);
1817 blk_queue_rq_timeout(q
, set
->timeout
);
1820 * Do this after blk_queue_make_request() overrides it...
1822 q
->nr_requests
= set
->queue_depth
;
1824 if (set
->ops
->complete
)
1825 blk_queue_softirq_done(q
, set
->ops
->complete
);
1827 blk_mq_init_flush(q
);
1828 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
1830 q
->flush_rq
= kzalloc(round_up(sizeof(struct request
) +
1831 set
->cmd_size
, cache_line_size()),
1836 if (blk_mq_init_hw_queues(q
, set
))
1839 mutex_lock(&all_q_mutex
);
1840 list_add_tail(&q
->all_q_node
, &all_q_list
);
1841 mutex_unlock(&all_q_mutex
);
1843 blk_mq_add_queue_tag_set(set
, q
);
1845 blk_mq_map_swqueue(q
);
1852 blk_cleanup_queue(q
);
1855 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1858 free_cpumask_var(hctxs
[i
]->cpumask
);
1865 return ERR_PTR(-ENOMEM
);
1867 EXPORT_SYMBOL(blk_mq_init_queue
);
1869 void blk_mq_free_queue(struct request_queue
*q
)
1871 struct blk_mq_tag_set
*set
= q
->tag_set
;
1873 blk_mq_del_queue_tag_set(q
);
1875 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
1876 blk_mq_free_hw_queues(q
, set
);
1878 percpu_ref_exit(&q
->mq_usage_counter
);
1880 free_percpu(q
->queue_ctx
);
1881 kfree(q
->queue_hw_ctx
);
1884 q
->queue_ctx
= NULL
;
1885 q
->queue_hw_ctx
= NULL
;
1888 mutex_lock(&all_q_mutex
);
1889 list_del_init(&q
->all_q_node
);
1890 mutex_unlock(&all_q_mutex
);
1893 /* Basically redo blk_mq_init_queue with queue frozen */
1894 static void blk_mq_queue_reinit(struct request_queue
*q
)
1896 blk_mq_freeze_queue(q
);
1898 blk_mq_sysfs_unregister(q
);
1900 blk_mq_update_queue_map(q
->mq_map
, q
->nr_hw_queues
);
1903 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1904 * we should change hctx numa_node according to new topology (this
1905 * involves free and re-allocate memory, worthy doing?)
1908 blk_mq_map_swqueue(q
);
1910 blk_mq_sysfs_register(q
);
1912 blk_mq_unfreeze_queue(q
);
1915 static int blk_mq_queue_reinit_notify(struct notifier_block
*nb
,
1916 unsigned long action
, void *hcpu
)
1918 struct request_queue
*q
;
1921 * Before new mappings are established, hotadded cpu might already
1922 * start handling requests. This doesn't break anything as we map
1923 * offline CPUs to first hardware queue. We will re-init the queue
1924 * below to get optimal settings.
1926 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
&&
1927 action
!= CPU_ONLINE
&& action
!= CPU_ONLINE_FROZEN
)
1930 mutex_lock(&all_q_mutex
);
1931 list_for_each_entry(q
, &all_q_list
, all_q_node
)
1932 blk_mq_queue_reinit(q
);
1933 mutex_unlock(&all_q_mutex
);
1938 * Alloc a tag set to be associated with one or more request queues.
1939 * May fail with EINVAL for various error conditions. May adjust the
1940 * requested depth down, if if it too large. In that case, the set
1941 * value will be stored in set->queue_depth.
1943 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
1947 if (!set
->nr_hw_queues
)
1949 if (!set
->queue_depth
)
1951 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
1954 if (!set
->nr_hw_queues
|| !set
->ops
->queue_rq
|| !set
->ops
->map_queue
)
1957 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
1958 pr_info("blk-mq: reduced tag depth to %u\n",
1960 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
1963 set
->tags
= kmalloc_node(set
->nr_hw_queues
*
1964 sizeof(struct blk_mq_tags
*),
1965 GFP_KERNEL
, set
->numa_node
);
1969 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1970 set
->tags
[i
] = blk_mq_init_rq_map(set
, i
);
1975 mutex_init(&set
->tag_list_lock
);
1976 INIT_LIST_HEAD(&set
->tag_list
);
1982 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1986 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
1988 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
1992 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1994 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1999 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2001 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2003 struct blk_mq_tag_set
*set
= q
->tag_set
;
2004 struct blk_mq_hw_ctx
*hctx
;
2007 if (!set
|| nr
> set
->queue_depth
)
2011 queue_for_each_hw_ctx(q
, hctx
, i
) {
2012 ret
= blk_mq_tag_update_depth(hctx
->tags
, nr
);
2018 q
->nr_requests
= nr
;
2023 void blk_mq_disable_hotplug(void)
2025 mutex_lock(&all_q_mutex
);
2028 void blk_mq_enable_hotplug(void)
2030 mutex_unlock(&all_q_mutex
);
2033 static int __init
blk_mq_init(void)
2037 hotcpu_notifier(blk_mq_queue_reinit_notify
, 0);
2041 subsys_initcall(blk_mq_init
);