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
)
115 spin_lock_irq(q
->queue_lock
);
116 q
->mq_freeze_depth
++;
117 spin_unlock_irq(q
->queue_lock
);
119 percpu_ref_kill(&q
->mq_usage_counter
);
120 blk_mq_run_queues(q
, false);
121 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->mq_usage_counter
));
124 static void blk_mq_unfreeze_queue(struct request_queue
*q
)
128 spin_lock_irq(q
->queue_lock
);
129 wake
= !--q
->mq_freeze_depth
;
130 WARN_ON_ONCE(q
->mq_freeze_depth
< 0);
131 spin_unlock_irq(q
->queue_lock
);
133 percpu_ref_reinit(&q
->mq_usage_counter
);
134 wake_up_all(&q
->mq_freeze_wq
);
138 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
140 return blk_mq_has_free_tags(hctx
->tags
);
142 EXPORT_SYMBOL(blk_mq_can_queue
);
144 static void blk_mq_rq_ctx_init(struct request_queue
*q
, struct blk_mq_ctx
*ctx
,
145 struct request
*rq
, unsigned int rw_flags
)
147 if (blk_queue_io_stat(q
))
148 rw_flags
|= REQ_IO_STAT
;
150 INIT_LIST_HEAD(&rq
->queuelist
);
151 /* csd/requeue_work/fifo_time is initialized before use */
154 rq
->cmd_flags
|= rw_flags
;
155 /* do not touch atomic flags, it needs atomic ops against the timer */
157 INIT_HLIST_NODE(&rq
->hash
);
158 RB_CLEAR_NODE(&rq
->rb_node
);
161 rq
->start_time
= jiffies
;
162 #ifdef CONFIG_BLK_CGROUP
164 set_start_time_ns(rq
);
165 rq
->io_start_time_ns
= 0;
167 rq
->nr_phys_segments
= 0;
168 #if defined(CONFIG_BLK_DEV_INTEGRITY)
169 rq
->nr_integrity_segments
= 0;
172 /* tag was already set */
180 INIT_LIST_HEAD(&rq
->timeout_list
);
184 rq
->end_io_data
= NULL
;
187 ctx
->rq_dispatched
[rw_is_sync(rw_flags
)]++;
190 static struct request
*
191 __blk_mq_alloc_request(struct blk_mq_alloc_data
*data
, int rw
)
196 tag
= blk_mq_get_tag(data
);
197 if (tag
!= BLK_MQ_TAG_FAIL
) {
198 rq
= data
->hctx
->tags
->rqs
[tag
];
201 if (blk_mq_tag_busy(data
->hctx
)) {
202 rq
->cmd_flags
= REQ_MQ_INFLIGHT
;
203 atomic_inc(&data
->hctx
->nr_active
);
207 blk_mq_rq_ctx_init(data
->q
, data
->ctx
, rq
, rw
);
214 struct request
*blk_mq_alloc_request(struct request_queue
*q
, int rw
, gfp_t gfp
,
217 struct blk_mq_ctx
*ctx
;
218 struct blk_mq_hw_ctx
*hctx
;
220 struct blk_mq_alloc_data alloc_data
;
222 if (blk_mq_queue_enter(q
))
225 ctx
= blk_mq_get_ctx(q
);
226 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
227 blk_mq_set_alloc_data(&alloc_data
, q
, gfp
& ~__GFP_WAIT
,
228 reserved
, ctx
, hctx
);
230 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
231 if (!rq
&& (gfp
& __GFP_WAIT
)) {
232 __blk_mq_run_hw_queue(hctx
);
235 ctx
= blk_mq_get_ctx(q
);
236 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
237 blk_mq_set_alloc_data(&alloc_data
, q
, gfp
, reserved
, ctx
,
239 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
240 ctx
= alloc_data
.ctx
;
245 EXPORT_SYMBOL(blk_mq_alloc_request
);
247 static void __blk_mq_free_request(struct blk_mq_hw_ctx
*hctx
,
248 struct blk_mq_ctx
*ctx
, struct request
*rq
)
250 const int tag
= rq
->tag
;
251 struct request_queue
*q
= rq
->q
;
253 if (rq
->cmd_flags
& REQ_MQ_INFLIGHT
)
254 atomic_dec(&hctx
->nr_active
);
256 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
257 blk_mq_put_tag(hctx
, tag
, &ctx
->last_tag
);
258 blk_mq_queue_exit(q
);
261 void blk_mq_free_request(struct request
*rq
)
263 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
264 struct blk_mq_hw_ctx
*hctx
;
265 struct request_queue
*q
= rq
->q
;
267 ctx
->rq_completed
[rq_is_sync(rq
)]++;
269 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
270 __blk_mq_free_request(hctx
, ctx
, rq
);
274 * Clone all relevant state from a request that has been put on hold in
275 * the flush state machine into the preallocated flush request that hangs
276 * off the request queue.
278 * For a driver the flush request should be invisible, that's why we are
279 * impersonating the original request here.
281 void blk_mq_clone_flush_request(struct request
*flush_rq
,
282 struct request
*orig_rq
)
284 struct blk_mq_hw_ctx
*hctx
=
285 orig_rq
->q
->mq_ops
->map_queue(orig_rq
->q
, orig_rq
->mq_ctx
->cpu
);
287 flush_rq
->mq_ctx
= orig_rq
->mq_ctx
;
288 flush_rq
->tag
= orig_rq
->tag
;
289 memcpy(blk_mq_rq_to_pdu(flush_rq
), blk_mq_rq_to_pdu(orig_rq
),
293 inline void __blk_mq_end_io(struct request
*rq
, int error
)
295 blk_account_io_done(rq
);
298 rq
->end_io(rq
, error
);
300 if (unlikely(blk_bidi_rq(rq
)))
301 blk_mq_free_request(rq
->next_rq
);
302 blk_mq_free_request(rq
);
305 EXPORT_SYMBOL(__blk_mq_end_io
);
307 void blk_mq_end_io(struct request
*rq
, int error
)
309 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
311 __blk_mq_end_io(rq
, error
);
313 EXPORT_SYMBOL(blk_mq_end_io
);
315 static void __blk_mq_complete_request_remote(void *data
)
317 struct request
*rq
= data
;
319 rq
->q
->softirq_done_fn(rq
);
322 static void blk_mq_ipi_complete_request(struct request
*rq
)
324 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
328 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
329 rq
->q
->softirq_done_fn(rq
);
334 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
335 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
337 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
338 rq
->csd
.func
= __blk_mq_complete_request_remote
;
341 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
343 rq
->q
->softirq_done_fn(rq
);
348 void __blk_mq_complete_request(struct request
*rq
)
350 struct request_queue
*q
= rq
->q
;
352 if (!q
->softirq_done_fn
)
353 blk_mq_end_io(rq
, rq
->errors
);
355 blk_mq_ipi_complete_request(rq
);
359 * blk_mq_complete_request - end I/O on a request
360 * @rq: the request being processed
363 * Ends all I/O on a request. It does not handle partial completions.
364 * The actual completion happens out-of-order, through a IPI handler.
366 void blk_mq_complete_request(struct request
*rq
)
368 struct request_queue
*q
= rq
->q
;
370 if (unlikely(blk_should_fake_timeout(q
)))
372 if (!blk_mark_rq_complete(rq
))
373 __blk_mq_complete_request(rq
);
375 EXPORT_SYMBOL(blk_mq_complete_request
);
377 static void blk_mq_start_request(struct request
*rq
, bool last
)
379 struct request_queue
*q
= rq
->q
;
381 trace_block_rq_issue(q
, rq
);
383 rq
->resid_len
= blk_rq_bytes(rq
);
384 if (unlikely(blk_bidi_rq(rq
)))
385 rq
->next_rq
->resid_len
= blk_rq_bytes(rq
->next_rq
);
390 * Mark us as started and clear complete. Complete might have been
391 * set if requeue raced with timeout, which then marked it as
392 * complete. So be sure to clear complete again when we start
393 * the request, otherwise we'll ignore the completion event.
395 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
396 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
397 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
398 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
400 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
402 * Make sure space for the drain appears. We know we can do
403 * this because max_hw_segments has been adjusted to be one
404 * fewer than the device can handle.
406 rq
->nr_phys_segments
++;
410 * Flag the last request in the series so that drivers know when IO
411 * should be kicked off, if they don't do it on a per-request basis.
413 * Note: the flag isn't the only condition drivers should do kick off.
414 * If drive is busy, the last request might not have the bit set.
417 rq
->cmd_flags
|= REQ_END
;
420 static void __blk_mq_requeue_request(struct request
*rq
)
422 struct request_queue
*q
= rq
->q
;
424 trace_block_rq_requeue(q
, rq
);
425 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
427 rq
->cmd_flags
&= ~REQ_END
;
429 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
430 rq
->nr_phys_segments
--;
433 void blk_mq_requeue_request(struct request
*rq
)
435 __blk_mq_requeue_request(rq
);
436 blk_clear_rq_complete(rq
);
438 BUG_ON(blk_queued_rq(rq
));
439 blk_mq_add_to_requeue_list(rq
, true);
441 EXPORT_SYMBOL(blk_mq_requeue_request
);
443 static void blk_mq_requeue_work(struct work_struct
*work
)
445 struct request_queue
*q
=
446 container_of(work
, struct request_queue
, requeue_work
);
448 struct request
*rq
, *next
;
451 spin_lock_irqsave(&q
->requeue_lock
, flags
);
452 list_splice_init(&q
->requeue_list
, &rq_list
);
453 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
455 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
456 if (!(rq
->cmd_flags
& REQ_SOFTBARRIER
))
459 rq
->cmd_flags
&= ~REQ_SOFTBARRIER
;
460 list_del_init(&rq
->queuelist
);
461 blk_mq_insert_request(rq
, true, false, false);
464 while (!list_empty(&rq_list
)) {
465 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
466 list_del_init(&rq
->queuelist
);
467 blk_mq_insert_request(rq
, false, false, false);
470 blk_mq_run_queues(q
, false);
473 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
)
475 struct request_queue
*q
= rq
->q
;
479 * We abuse this flag that is otherwise used by the I/O scheduler to
480 * request head insertation from the workqueue.
482 BUG_ON(rq
->cmd_flags
& REQ_SOFTBARRIER
);
484 spin_lock_irqsave(&q
->requeue_lock
, flags
);
486 rq
->cmd_flags
|= REQ_SOFTBARRIER
;
487 list_add(&rq
->queuelist
, &q
->requeue_list
);
489 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
491 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
493 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
495 void blk_mq_kick_requeue_list(struct request_queue
*q
)
497 kblockd_schedule_work(&q
->requeue_work
);
499 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
501 static inline bool is_flush_request(struct request
*rq
, unsigned int tag
)
503 return ((rq
->cmd_flags
& REQ_FLUSH_SEQ
) &&
504 rq
->q
->flush_rq
->tag
== tag
);
507 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
509 struct request
*rq
= tags
->rqs
[tag
];
511 if (!is_flush_request(rq
, tag
))
514 return rq
->q
->flush_rq
;
516 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
518 struct blk_mq_timeout_data
{
519 struct blk_mq_hw_ctx
*hctx
;
521 unsigned int *next_set
;
524 static void blk_mq_timeout_check(void *__data
, unsigned long *free_tags
)
526 struct blk_mq_timeout_data
*data
= __data
;
527 struct blk_mq_hw_ctx
*hctx
= data
->hctx
;
530 /* It may not be in flight yet (this is where
531 * the REQ_ATOMIC_STARTED flag comes in). The requests are
532 * statically allocated, so we know it's always safe to access the
533 * memory associated with a bit offset into ->rqs[].
539 tag
= find_next_zero_bit(free_tags
, hctx
->tags
->nr_tags
, tag
);
540 if (tag
>= hctx
->tags
->nr_tags
)
543 rq
= blk_mq_tag_to_rq(hctx
->tags
, tag
++);
544 if (rq
->q
!= hctx
->queue
)
546 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
549 blk_rq_check_expired(rq
, data
->next
, data
->next_set
);
553 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx
*hctx
,
555 unsigned int *next_set
)
557 struct blk_mq_timeout_data data
= {
560 .next_set
= next_set
,
564 * Ask the tagging code to iterate busy requests, so we can
565 * check them for timeout.
567 blk_mq_tag_busy_iter(hctx
->tags
, blk_mq_timeout_check
, &data
);
570 static enum blk_eh_timer_return
blk_mq_rq_timed_out(struct request
*rq
)
572 struct request_queue
*q
= rq
->q
;
575 * We know that complete is set at this point. If STARTED isn't set
576 * anymore, then the request isn't active and the "timeout" should
577 * just be ignored. This can happen due to the bitflag ordering.
578 * Timeout first checks if STARTED is set, and if it is, assumes
579 * the request is active. But if we race with completion, then
580 * we both flags will get cleared. So check here again, and ignore
581 * a timeout event with a request that isn't active.
583 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
584 return BLK_EH_NOT_HANDLED
;
586 if (!q
->mq_ops
->timeout
)
587 return BLK_EH_RESET_TIMER
;
589 return q
->mq_ops
->timeout(rq
);
592 static void blk_mq_rq_timer(unsigned long data
)
594 struct request_queue
*q
= (struct request_queue
*) data
;
595 struct blk_mq_hw_ctx
*hctx
;
596 unsigned long next
= 0;
599 queue_for_each_hw_ctx(q
, hctx
, i
) {
601 * If not software queues are currently mapped to this
602 * hardware queue, there's nothing to check
604 if (!hctx
->nr_ctx
|| !hctx
->tags
)
607 blk_mq_hw_ctx_check_timeout(hctx
, &next
, &next_set
);
611 next
= blk_rq_timeout(round_jiffies_up(next
));
612 mod_timer(&q
->timeout
, next
);
614 queue_for_each_hw_ctx(q
, hctx
, i
)
615 blk_mq_tag_idle(hctx
);
620 * Reverse check our software queue for entries that we could potentially
621 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
622 * too much time checking for merges.
624 static bool blk_mq_attempt_merge(struct request_queue
*q
,
625 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
630 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
636 if (!blk_rq_merge_ok(rq
, bio
))
639 el_ret
= blk_try_merge(rq
, bio
);
640 if (el_ret
== ELEVATOR_BACK_MERGE
) {
641 if (bio_attempt_back_merge(q
, rq
, bio
)) {
646 } else if (el_ret
== ELEVATOR_FRONT_MERGE
) {
647 if (bio_attempt_front_merge(q
, rq
, bio
)) {
659 * Process software queues that have been marked busy, splicing them
660 * to the for-dispatch
662 static void flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
664 struct blk_mq_ctx
*ctx
;
667 for (i
= 0; i
< hctx
->ctx_map
.map_size
; i
++) {
668 struct blk_align_bitmap
*bm
= &hctx
->ctx_map
.map
[i
];
669 unsigned int off
, bit
;
675 off
= i
* hctx
->ctx_map
.bits_per_word
;
677 bit
= find_next_bit(&bm
->word
, bm
->depth
, bit
);
678 if (bit
>= bm
->depth
)
681 ctx
= hctx
->ctxs
[bit
+ off
];
682 clear_bit(bit
, &bm
->word
);
683 spin_lock(&ctx
->lock
);
684 list_splice_tail_init(&ctx
->rq_list
, list
);
685 spin_unlock(&ctx
->lock
);
693 * Run this hardware queue, pulling any software queues mapped to it in.
694 * Note that this function currently has various problems around ordering
695 * of IO. In particular, we'd like FIFO behaviour on handling existing
696 * items on the hctx->dispatch list. Ignore that for now.
698 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
700 struct request_queue
*q
= hctx
->queue
;
705 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
));
707 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)))
713 * Touch any software queue that has pending entries.
715 flush_busy_ctxs(hctx
, &rq_list
);
718 * If we have previous entries on our dispatch list, grab them
719 * and stuff them at the front for more fair dispatch.
721 if (!list_empty_careful(&hctx
->dispatch
)) {
722 spin_lock(&hctx
->lock
);
723 if (!list_empty(&hctx
->dispatch
))
724 list_splice_init(&hctx
->dispatch
, &rq_list
);
725 spin_unlock(&hctx
->lock
);
729 * Now process all the entries, sending them to the driver.
732 while (!list_empty(&rq_list
)) {
735 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
736 list_del_init(&rq
->queuelist
);
738 blk_mq_start_request(rq
, list_empty(&rq_list
));
740 ret
= q
->mq_ops
->queue_rq(hctx
, rq
);
742 case BLK_MQ_RQ_QUEUE_OK
:
745 case BLK_MQ_RQ_QUEUE_BUSY
:
746 list_add(&rq
->queuelist
, &rq_list
);
747 __blk_mq_requeue_request(rq
);
750 pr_err("blk-mq: bad return on queue: %d\n", ret
);
751 case BLK_MQ_RQ_QUEUE_ERROR
:
753 blk_mq_end_io(rq
, rq
->errors
);
757 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
762 hctx
->dispatched
[0]++;
763 else if (queued
< (1 << (BLK_MQ_MAX_DISPATCH_ORDER
- 1)))
764 hctx
->dispatched
[ilog2(queued
) + 1]++;
767 * Any items that need requeuing? Stuff them into hctx->dispatch,
768 * that is where we will continue on next queue run.
770 if (!list_empty(&rq_list
)) {
771 spin_lock(&hctx
->lock
);
772 list_splice(&rq_list
, &hctx
->dispatch
);
773 spin_unlock(&hctx
->lock
);
778 * It'd be great if the workqueue API had a way to pass
779 * in a mask and had some smarts for more clever placement.
780 * For now we just round-robin here, switching for every
781 * BLK_MQ_CPU_WORK_BATCH queued items.
783 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
785 int cpu
= hctx
->next_cpu
;
787 if (--hctx
->next_cpu_batch
<= 0) {
790 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
791 if (next_cpu
>= nr_cpu_ids
)
792 next_cpu
= cpumask_first(hctx
->cpumask
);
794 hctx
->next_cpu
= next_cpu
;
795 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
801 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
803 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)))
806 if (!async
&& cpumask_test_cpu(smp_processor_id(), hctx
->cpumask
))
807 __blk_mq_run_hw_queue(hctx
);
808 else if (hctx
->queue
->nr_hw_queues
== 1)
809 kblockd_schedule_delayed_work(&hctx
->run_work
, 0);
813 cpu
= blk_mq_hctx_next_cpu(hctx
);
814 kblockd_schedule_delayed_work_on(cpu
, &hctx
->run_work
, 0);
818 void blk_mq_run_queues(struct request_queue
*q
, bool async
)
820 struct blk_mq_hw_ctx
*hctx
;
823 queue_for_each_hw_ctx(q
, hctx
, i
) {
824 if ((!blk_mq_hctx_has_pending(hctx
) &&
825 list_empty_careful(&hctx
->dispatch
)) ||
826 test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
830 blk_mq_run_hw_queue(hctx
, async
);
834 EXPORT_SYMBOL(blk_mq_run_queues
);
836 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
838 cancel_delayed_work(&hctx
->run_work
);
839 cancel_delayed_work(&hctx
->delay_work
);
840 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
842 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
844 void blk_mq_stop_hw_queues(struct request_queue
*q
)
846 struct blk_mq_hw_ctx
*hctx
;
849 queue_for_each_hw_ctx(q
, hctx
, i
)
850 blk_mq_stop_hw_queue(hctx
);
852 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
854 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
856 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
859 blk_mq_run_hw_queue(hctx
, false);
862 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
864 void blk_mq_start_hw_queues(struct request_queue
*q
)
866 struct blk_mq_hw_ctx
*hctx
;
869 queue_for_each_hw_ctx(q
, hctx
, i
)
870 blk_mq_start_hw_queue(hctx
);
872 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
875 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
877 struct blk_mq_hw_ctx
*hctx
;
880 queue_for_each_hw_ctx(q
, hctx
, i
) {
881 if (!test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
884 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
886 blk_mq_run_hw_queue(hctx
, async
);
890 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
892 static void blk_mq_run_work_fn(struct work_struct
*work
)
894 struct blk_mq_hw_ctx
*hctx
;
896 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
898 __blk_mq_run_hw_queue(hctx
);
901 static void blk_mq_delay_work_fn(struct work_struct
*work
)
903 struct blk_mq_hw_ctx
*hctx
;
905 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
907 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
908 __blk_mq_run_hw_queue(hctx
);
911 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
913 unsigned long tmo
= msecs_to_jiffies(msecs
);
915 if (hctx
->queue
->nr_hw_queues
== 1)
916 kblockd_schedule_delayed_work(&hctx
->delay_work
, tmo
);
920 cpu
= blk_mq_hctx_next_cpu(hctx
);
921 kblockd_schedule_delayed_work_on(cpu
, &hctx
->delay_work
, tmo
);
924 EXPORT_SYMBOL(blk_mq_delay_queue
);
926 static void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
,
927 struct request
*rq
, bool at_head
)
929 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
931 trace_block_rq_insert(hctx
->queue
, rq
);
934 list_add(&rq
->queuelist
, &ctx
->rq_list
);
936 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
938 blk_mq_hctx_mark_pending(hctx
, ctx
);
941 void blk_mq_insert_request(struct request
*rq
, bool at_head
, bool run_queue
,
944 struct request_queue
*q
= rq
->q
;
945 struct blk_mq_hw_ctx
*hctx
;
946 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
, *current_ctx
;
948 current_ctx
= blk_mq_get_ctx(q
);
949 if (!cpu_online(ctx
->cpu
))
950 rq
->mq_ctx
= ctx
= current_ctx
;
952 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
954 if (rq
->cmd_flags
& (REQ_FLUSH
| REQ_FUA
) &&
955 !(rq
->cmd_flags
& (REQ_FLUSH_SEQ
))) {
956 blk_insert_flush(rq
);
958 spin_lock(&ctx
->lock
);
959 __blk_mq_insert_request(hctx
, rq
, at_head
);
960 spin_unlock(&ctx
->lock
);
964 blk_mq_run_hw_queue(hctx
, async
);
966 blk_mq_put_ctx(current_ctx
);
969 static void blk_mq_insert_requests(struct request_queue
*q
,
970 struct blk_mq_ctx
*ctx
,
971 struct list_head
*list
,
976 struct blk_mq_hw_ctx
*hctx
;
977 struct blk_mq_ctx
*current_ctx
;
979 trace_block_unplug(q
, depth
, !from_schedule
);
981 current_ctx
= blk_mq_get_ctx(q
);
983 if (!cpu_online(ctx
->cpu
))
985 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
988 * preemption doesn't flush plug list, so it's possible ctx->cpu is
991 spin_lock(&ctx
->lock
);
992 while (!list_empty(list
)) {
995 rq
= list_first_entry(list
, struct request
, queuelist
);
996 list_del_init(&rq
->queuelist
);
998 __blk_mq_insert_request(hctx
, rq
, false);
1000 spin_unlock(&ctx
->lock
);
1002 blk_mq_run_hw_queue(hctx
, from_schedule
);
1003 blk_mq_put_ctx(current_ctx
);
1006 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1008 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1009 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1011 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1012 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1013 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1016 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1018 struct blk_mq_ctx
*this_ctx
;
1019 struct request_queue
*this_q
;
1022 LIST_HEAD(ctx_list
);
1025 list_splice_init(&plug
->mq_list
, &list
);
1027 list_sort(NULL
, &list
, plug_ctx_cmp
);
1033 while (!list_empty(&list
)) {
1034 rq
= list_entry_rq(list
.next
);
1035 list_del_init(&rq
->queuelist
);
1037 if (rq
->mq_ctx
!= this_ctx
) {
1039 blk_mq_insert_requests(this_q
, this_ctx
,
1044 this_ctx
= rq
->mq_ctx
;
1050 list_add_tail(&rq
->queuelist
, &ctx_list
);
1054 * If 'this_ctx' is set, we know we have entries to complete
1055 * on 'ctx_list'. Do those.
1058 blk_mq_insert_requests(this_q
, this_ctx
, &ctx_list
, depth
,
1063 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1065 init_request_from_bio(rq
, bio
);
1067 if (blk_do_io_stat(rq
))
1068 blk_account_io_start(rq
, 1);
1071 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx
*hctx
,
1072 struct blk_mq_ctx
*ctx
,
1073 struct request
*rq
, struct bio
*bio
)
1075 struct request_queue
*q
= hctx
->queue
;
1077 if (!(hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
)) {
1078 blk_mq_bio_to_request(rq
, bio
);
1079 spin_lock(&ctx
->lock
);
1081 __blk_mq_insert_request(hctx
, rq
, false);
1082 spin_unlock(&ctx
->lock
);
1085 spin_lock(&ctx
->lock
);
1086 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1087 blk_mq_bio_to_request(rq
, bio
);
1091 spin_unlock(&ctx
->lock
);
1092 __blk_mq_free_request(hctx
, ctx
, rq
);
1097 struct blk_map_ctx
{
1098 struct blk_mq_hw_ctx
*hctx
;
1099 struct blk_mq_ctx
*ctx
;
1102 static struct request
*blk_mq_map_request(struct request_queue
*q
,
1104 struct blk_map_ctx
*data
)
1106 struct blk_mq_hw_ctx
*hctx
;
1107 struct blk_mq_ctx
*ctx
;
1109 int rw
= bio_data_dir(bio
);
1110 struct blk_mq_alloc_data alloc_data
;
1112 if (unlikely(blk_mq_queue_enter(q
))) {
1113 bio_endio(bio
, -EIO
);
1117 ctx
= blk_mq_get_ctx(q
);
1118 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1120 if (rw_is_sync(bio
->bi_rw
))
1123 trace_block_getrq(q
, bio
, rw
);
1124 blk_mq_set_alloc_data(&alloc_data
, q
, GFP_ATOMIC
, false, ctx
,
1126 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
1127 if (unlikely(!rq
)) {
1128 __blk_mq_run_hw_queue(hctx
);
1129 blk_mq_put_ctx(ctx
);
1130 trace_block_sleeprq(q
, bio
, rw
);
1132 ctx
= blk_mq_get_ctx(q
);
1133 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1134 blk_mq_set_alloc_data(&alloc_data
, q
,
1135 __GFP_WAIT
|GFP_ATOMIC
, false, ctx
, hctx
);
1136 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
1137 ctx
= alloc_data
.ctx
;
1138 hctx
= alloc_data
.hctx
;
1148 * Multiple hardware queue variant. This will not use per-process plugs,
1149 * but will attempt to bypass the hctx queueing if we can go straight to
1150 * hardware for SYNC IO.
1152 static void blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1154 const int is_sync
= rw_is_sync(bio
->bi_rw
);
1155 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
1156 struct blk_map_ctx data
;
1159 blk_queue_bounce(q
, &bio
);
1161 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1162 bio_endio(bio
, -EIO
);
1166 rq
= blk_mq_map_request(q
, bio
, &data
);
1170 if (unlikely(is_flush_fua
)) {
1171 blk_mq_bio_to_request(rq
, bio
);
1172 blk_insert_flush(rq
);
1179 blk_mq_bio_to_request(rq
, bio
);
1180 blk_mq_start_request(rq
, true);
1183 * For OK queue, we are done. For error, kill it. Any other
1184 * error (busy), just add it to our list as we previously
1187 ret
= q
->mq_ops
->queue_rq(data
.hctx
, rq
);
1188 if (ret
== BLK_MQ_RQ_QUEUE_OK
)
1191 __blk_mq_requeue_request(rq
);
1193 if (ret
== BLK_MQ_RQ_QUEUE_ERROR
) {
1195 blk_mq_end_io(rq
, rq
->errors
);
1201 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1203 * For a SYNC request, send it to the hardware immediately. For
1204 * an ASYNC request, just ensure that we run it later on. The
1205 * latter allows for merging opportunities and more efficient
1209 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1212 blk_mq_put_ctx(data
.ctx
);
1216 * Single hardware queue variant. This will attempt to use any per-process
1217 * plug for merging and IO deferral.
1219 static void blk_sq_make_request(struct request_queue
*q
, struct bio
*bio
)
1221 const int is_sync
= rw_is_sync(bio
->bi_rw
);
1222 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
1223 unsigned int use_plug
, request_count
= 0;
1224 struct blk_map_ctx data
;
1228 * If we have multiple hardware queues, just go directly to
1229 * one of those for sync IO.
1231 use_plug
= !is_flush_fua
&& !is_sync
;
1233 blk_queue_bounce(q
, &bio
);
1235 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1236 bio_endio(bio
, -EIO
);
1240 if (use_plug
&& !blk_queue_nomerges(q
) &&
1241 blk_attempt_plug_merge(q
, bio
, &request_count
))
1244 rq
= blk_mq_map_request(q
, bio
, &data
);
1248 if (unlikely(is_flush_fua
)) {
1249 blk_mq_bio_to_request(rq
, bio
);
1250 blk_insert_flush(rq
);
1255 * A task plug currently exists. Since this is completely lockless,
1256 * utilize that to temporarily store requests until the task is
1257 * either done or scheduled away.
1260 struct blk_plug
*plug
= current
->plug
;
1263 blk_mq_bio_to_request(rq
, bio
);
1264 if (list_empty(&plug
->mq_list
))
1265 trace_block_plug(q
);
1266 else if (request_count
>= BLK_MAX_REQUEST_COUNT
) {
1267 blk_flush_plug_list(plug
, false);
1268 trace_block_plug(q
);
1270 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1271 blk_mq_put_ctx(data
.ctx
);
1276 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1278 * For a SYNC request, send it to the hardware immediately. For
1279 * an ASYNC request, just ensure that we run it later on. The
1280 * latter allows for merging opportunities and more efficient
1284 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1287 blk_mq_put_ctx(data
.ctx
);
1291 * Default mapping to a software queue, since we use one per CPU.
1293 struct blk_mq_hw_ctx
*blk_mq_map_queue(struct request_queue
*q
, const int cpu
)
1295 return q
->queue_hw_ctx
[q
->mq_map
[cpu
]];
1297 EXPORT_SYMBOL(blk_mq_map_queue
);
1299 static void blk_mq_free_rq_map(struct blk_mq_tag_set
*set
,
1300 struct blk_mq_tags
*tags
, unsigned int hctx_idx
)
1304 if (tags
->rqs
&& set
->ops
->exit_request
) {
1307 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1310 set
->ops
->exit_request(set
->driver_data
, tags
->rqs
[i
],
1315 while (!list_empty(&tags
->page_list
)) {
1316 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1317 list_del_init(&page
->lru
);
1318 __free_pages(page
, page
->private);
1323 blk_mq_free_tags(tags
);
1326 static size_t order_to_size(unsigned int order
)
1328 return (size_t)PAGE_SIZE
<< order
;
1331 static struct blk_mq_tags
*blk_mq_init_rq_map(struct blk_mq_tag_set
*set
,
1332 unsigned int hctx_idx
)
1334 struct blk_mq_tags
*tags
;
1335 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1336 size_t rq_size
, left
;
1338 tags
= blk_mq_init_tags(set
->queue_depth
, set
->reserved_tags
,
1343 INIT_LIST_HEAD(&tags
->page_list
);
1345 tags
->rqs
= kmalloc_node(set
->queue_depth
* sizeof(struct request
*),
1346 GFP_KERNEL
, set
->numa_node
);
1348 blk_mq_free_tags(tags
);
1353 * rq_size is the size of the request plus driver payload, rounded
1354 * to the cacheline size
1356 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1358 left
= rq_size
* set
->queue_depth
;
1360 for (i
= 0; i
< set
->queue_depth
; ) {
1361 int this_order
= max_order
;
1366 while (left
< order_to_size(this_order
- 1) && this_order
)
1370 page
= alloc_pages_node(set
->numa_node
, GFP_KERNEL
,
1376 if (order_to_size(this_order
) < rq_size
)
1383 page
->private = this_order
;
1384 list_add_tail(&page
->lru
, &tags
->page_list
);
1386 p
= page_address(page
);
1387 entries_per_page
= order_to_size(this_order
) / rq_size
;
1388 to_do
= min(entries_per_page
, set
->queue_depth
- i
);
1389 left
-= to_do
* rq_size
;
1390 for (j
= 0; j
< to_do
; j
++) {
1392 if (set
->ops
->init_request
) {
1393 if (set
->ops
->init_request(set
->driver_data
,
1394 tags
->rqs
[i
], hctx_idx
, i
,
1407 pr_warn("%s: failed to allocate requests\n", __func__
);
1408 blk_mq_free_rq_map(set
, tags
, hctx_idx
);
1412 static void blk_mq_free_bitmap(struct blk_mq_ctxmap
*bitmap
)
1417 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap
*bitmap
, int node
)
1419 unsigned int bpw
= 8, total
, num_maps
, i
;
1421 bitmap
->bits_per_word
= bpw
;
1423 num_maps
= ALIGN(nr_cpu_ids
, bpw
) / bpw
;
1424 bitmap
->map
= kzalloc_node(num_maps
* sizeof(struct blk_align_bitmap
),
1429 bitmap
->map_size
= num_maps
;
1432 for (i
= 0; i
< num_maps
; i
++) {
1433 bitmap
->map
[i
].depth
= min(total
, bitmap
->bits_per_word
);
1434 total
-= bitmap
->map
[i
].depth
;
1440 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx
*hctx
, int cpu
)
1442 struct request_queue
*q
= hctx
->queue
;
1443 struct blk_mq_ctx
*ctx
;
1447 * Move ctx entries to new CPU, if this one is going away.
1449 ctx
= __blk_mq_get_ctx(q
, cpu
);
1451 spin_lock(&ctx
->lock
);
1452 if (!list_empty(&ctx
->rq_list
)) {
1453 list_splice_init(&ctx
->rq_list
, &tmp
);
1454 blk_mq_hctx_clear_pending(hctx
, ctx
);
1456 spin_unlock(&ctx
->lock
);
1458 if (list_empty(&tmp
))
1461 ctx
= blk_mq_get_ctx(q
);
1462 spin_lock(&ctx
->lock
);
1464 while (!list_empty(&tmp
)) {
1467 rq
= list_first_entry(&tmp
, struct request
, queuelist
);
1469 list_move_tail(&rq
->queuelist
, &ctx
->rq_list
);
1472 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1473 blk_mq_hctx_mark_pending(hctx
, ctx
);
1475 spin_unlock(&ctx
->lock
);
1477 blk_mq_run_hw_queue(hctx
, true);
1478 blk_mq_put_ctx(ctx
);
1482 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx
*hctx
, int cpu
)
1484 struct request_queue
*q
= hctx
->queue
;
1485 struct blk_mq_tag_set
*set
= q
->tag_set
;
1487 if (set
->tags
[hctx
->queue_num
])
1490 set
->tags
[hctx
->queue_num
] = blk_mq_init_rq_map(set
, hctx
->queue_num
);
1491 if (!set
->tags
[hctx
->queue_num
])
1494 hctx
->tags
= set
->tags
[hctx
->queue_num
];
1498 static int blk_mq_hctx_notify(void *data
, unsigned long action
,
1501 struct blk_mq_hw_ctx
*hctx
= data
;
1503 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
)
1504 return blk_mq_hctx_cpu_offline(hctx
, cpu
);
1505 else if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
)
1506 return blk_mq_hctx_cpu_online(hctx
, cpu
);
1511 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1512 struct blk_mq_tag_set
*set
, int nr_queue
)
1514 struct blk_mq_hw_ctx
*hctx
;
1517 queue_for_each_hw_ctx(q
, hctx
, i
) {
1521 blk_mq_tag_idle(hctx
);
1523 if (set
->ops
->exit_hctx
)
1524 set
->ops
->exit_hctx(hctx
, i
);
1526 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1528 blk_mq_free_bitmap(&hctx
->ctx_map
);
1533 static void blk_mq_free_hw_queues(struct request_queue
*q
,
1534 struct blk_mq_tag_set
*set
)
1536 struct blk_mq_hw_ctx
*hctx
;
1539 queue_for_each_hw_ctx(q
, hctx
, i
) {
1540 free_cpumask_var(hctx
->cpumask
);
1545 static int blk_mq_init_hw_queues(struct request_queue
*q
,
1546 struct blk_mq_tag_set
*set
)
1548 struct blk_mq_hw_ctx
*hctx
;
1552 * Initialize hardware queues
1554 queue_for_each_hw_ctx(q
, hctx
, i
) {
1557 node
= hctx
->numa_node
;
1558 if (node
== NUMA_NO_NODE
)
1559 node
= hctx
->numa_node
= set
->numa_node
;
1561 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1562 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1563 spin_lock_init(&hctx
->lock
);
1564 INIT_LIST_HEAD(&hctx
->dispatch
);
1566 hctx
->queue_num
= i
;
1567 hctx
->flags
= set
->flags
;
1568 hctx
->cmd_size
= set
->cmd_size
;
1570 blk_mq_init_cpu_notifier(&hctx
->cpu_notifier
,
1571 blk_mq_hctx_notify
, hctx
);
1572 blk_mq_register_cpu_notifier(&hctx
->cpu_notifier
);
1574 hctx
->tags
= set
->tags
[i
];
1577 * Allocate space for all possible cpus to avoid allocation in
1580 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1585 if (blk_mq_alloc_bitmap(&hctx
->ctx_map
, node
))
1590 if (set
->ops
->init_hctx
&&
1591 set
->ops
->init_hctx(hctx
, set
->driver_data
, i
))
1595 if (i
== q
->nr_hw_queues
)
1601 blk_mq_exit_hw_queues(q
, set
, i
);
1606 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1607 unsigned int nr_hw_queues
)
1611 for_each_possible_cpu(i
) {
1612 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1613 struct blk_mq_hw_ctx
*hctx
;
1615 memset(__ctx
, 0, sizeof(*__ctx
));
1617 spin_lock_init(&__ctx
->lock
);
1618 INIT_LIST_HEAD(&__ctx
->rq_list
);
1621 /* If the cpu isn't online, the cpu is mapped to first hctx */
1625 hctx
= q
->mq_ops
->map_queue(q
, i
);
1626 cpumask_set_cpu(i
, hctx
->cpumask
);
1630 * Set local node, IFF we have more than one hw queue. If
1631 * not, we remain on the home node of the device
1633 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1634 hctx
->numa_node
= cpu_to_node(i
);
1638 static void blk_mq_map_swqueue(struct request_queue
*q
)
1641 struct blk_mq_hw_ctx
*hctx
;
1642 struct blk_mq_ctx
*ctx
;
1644 queue_for_each_hw_ctx(q
, hctx
, i
) {
1645 cpumask_clear(hctx
->cpumask
);
1650 * Map software to hardware queues
1652 queue_for_each_ctx(q
, ctx
, i
) {
1653 /* If the cpu isn't online, the cpu is mapped to first hctx */
1657 hctx
= q
->mq_ops
->map_queue(q
, i
);
1658 cpumask_set_cpu(i
, hctx
->cpumask
);
1659 ctx
->index_hw
= hctx
->nr_ctx
;
1660 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
1663 queue_for_each_hw_ctx(q
, hctx
, i
) {
1665 * If not software queues are mapped to this hardware queue,
1666 * disable it and free the request entries
1668 if (!hctx
->nr_ctx
) {
1669 struct blk_mq_tag_set
*set
= q
->tag_set
;
1672 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1673 set
->tags
[i
] = NULL
;
1680 * Initialize batch roundrobin counts
1682 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
1683 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1687 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
)
1689 struct blk_mq_hw_ctx
*hctx
;
1690 struct request_queue
*q
;
1694 if (set
->tag_list
.next
== set
->tag_list
.prev
)
1699 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
1700 blk_mq_freeze_queue(q
);
1702 queue_for_each_hw_ctx(q
, hctx
, i
) {
1704 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
1706 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
1708 blk_mq_unfreeze_queue(q
);
1712 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
1714 struct blk_mq_tag_set
*set
= q
->tag_set
;
1716 blk_mq_freeze_queue(q
);
1718 mutex_lock(&set
->tag_list_lock
);
1719 list_del_init(&q
->tag_set_list
);
1720 blk_mq_update_tag_set_depth(set
);
1721 mutex_unlock(&set
->tag_list_lock
);
1723 blk_mq_unfreeze_queue(q
);
1726 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
1727 struct request_queue
*q
)
1731 mutex_lock(&set
->tag_list_lock
);
1732 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
1733 blk_mq_update_tag_set_depth(set
);
1734 mutex_unlock(&set
->tag_list_lock
);
1737 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
1739 struct blk_mq_hw_ctx
**hctxs
;
1740 struct blk_mq_ctx __percpu
*ctx
;
1741 struct request_queue
*q
;
1745 ctx
= alloc_percpu(struct blk_mq_ctx
);
1747 return ERR_PTR(-ENOMEM
);
1749 hctxs
= kmalloc_node(set
->nr_hw_queues
* sizeof(*hctxs
), GFP_KERNEL
,
1755 map
= blk_mq_make_queue_map(set
);
1759 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1760 int node
= blk_mq_hw_queue_to_node(map
, i
);
1762 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
1767 if (!zalloc_cpumask_var(&hctxs
[i
]->cpumask
, GFP_KERNEL
))
1770 atomic_set(&hctxs
[i
]->nr_active
, 0);
1771 hctxs
[i
]->numa_node
= node
;
1772 hctxs
[i
]->queue_num
= i
;
1775 q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
1779 if (percpu_ref_init(&q
->mq_usage_counter
, blk_mq_usage_counter_release
))
1782 setup_timer(&q
->timeout
, blk_mq_rq_timer
, (unsigned long) q
);
1783 blk_queue_rq_timeout(q
, 30000);
1785 q
->nr_queues
= nr_cpu_ids
;
1786 q
->nr_hw_queues
= set
->nr_hw_queues
;
1790 q
->queue_hw_ctx
= hctxs
;
1792 q
->mq_ops
= set
->ops
;
1793 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
1795 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
1796 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
1798 q
->sg_reserved_size
= INT_MAX
;
1800 INIT_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
1801 INIT_LIST_HEAD(&q
->requeue_list
);
1802 spin_lock_init(&q
->requeue_lock
);
1804 if (q
->nr_hw_queues
> 1)
1805 blk_queue_make_request(q
, blk_mq_make_request
);
1807 blk_queue_make_request(q
, blk_sq_make_request
);
1809 blk_queue_rq_timed_out(q
, blk_mq_rq_timed_out
);
1811 blk_queue_rq_timeout(q
, set
->timeout
);
1814 * Do this after blk_queue_make_request() overrides it...
1816 q
->nr_requests
= set
->queue_depth
;
1818 if (set
->ops
->complete
)
1819 blk_queue_softirq_done(q
, set
->ops
->complete
);
1821 blk_mq_init_flush(q
);
1822 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
1824 q
->flush_rq
= kzalloc(round_up(sizeof(struct request
) +
1825 set
->cmd_size
, cache_line_size()),
1830 if (blk_mq_init_hw_queues(q
, set
))
1833 mutex_lock(&all_q_mutex
);
1834 list_add_tail(&q
->all_q_node
, &all_q_list
);
1835 mutex_unlock(&all_q_mutex
);
1837 blk_mq_add_queue_tag_set(set
, q
);
1839 blk_mq_map_swqueue(q
);
1846 blk_cleanup_queue(q
);
1849 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1852 free_cpumask_var(hctxs
[i
]->cpumask
);
1859 return ERR_PTR(-ENOMEM
);
1861 EXPORT_SYMBOL(blk_mq_init_queue
);
1863 void blk_mq_free_queue(struct request_queue
*q
)
1865 struct blk_mq_tag_set
*set
= q
->tag_set
;
1867 blk_mq_del_queue_tag_set(q
);
1869 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
1870 blk_mq_free_hw_queues(q
, set
);
1872 percpu_ref_exit(&q
->mq_usage_counter
);
1874 free_percpu(q
->queue_ctx
);
1875 kfree(q
->queue_hw_ctx
);
1878 q
->queue_ctx
= NULL
;
1879 q
->queue_hw_ctx
= NULL
;
1882 mutex_lock(&all_q_mutex
);
1883 list_del_init(&q
->all_q_node
);
1884 mutex_unlock(&all_q_mutex
);
1887 /* Basically redo blk_mq_init_queue with queue frozen */
1888 static void blk_mq_queue_reinit(struct request_queue
*q
)
1890 blk_mq_freeze_queue(q
);
1892 blk_mq_sysfs_unregister(q
);
1894 blk_mq_update_queue_map(q
->mq_map
, q
->nr_hw_queues
);
1897 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1898 * we should change hctx numa_node according to new topology (this
1899 * involves free and re-allocate memory, worthy doing?)
1902 blk_mq_map_swqueue(q
);
1904 blk_mq_sysfs_register(q
);
1906 blk_mq_unfreeze_queue(q
);
1909 static int blk_mq_queue_reinit_notify(struct notifier_block
*nb
,
1910 unsigned long action
, void *hcpu
)
1912 struct request_queue
*q
;
1915 * Before new mappings are established, hotadded cpu might already
1916 * start handling requests. This doesn't break anything as we map
1917 * offline CPUs to first hardware queue. We will re-init the queue
1918 * below to get optimal settings.
1920 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
&&
1921 action
!= CPU_ONLINE
&& action
!= CPU_ONLINE_FROZEN
)
1924 mutex_lock(&all_q_mutex
);
1925 list_for_each_entry(q
, &all_q_list
, all_q_node
)
1926 blk_mq_queue_reinit(q
);
1927 mutex_unlock(&all_q_mutex
);
1932 * Alloc a tag set to be associated with one or more request queues.
1933 * May fail with EINVAL for various error conditions. May adjust the
1934 * requested depth down, if if it too large. In that case, the set
1935 * value will be stored in set->queue_depth.
1937 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
1941 if (!set
->nr_hw_queues
)
1943 if (!set
->queue_depth
)
1945 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
1948 if (!set
->nr_hw_queues
|| !set
->ops
->queue_rq
|| !set
->ops
->map_queue
)
1951 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
1952 pr_info("blk-mq: reduced tag depth to %u\n",
1954 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
1957 set
->tags
= kmalloc_node(set
->nr_hw_queues
*
1958 sizeof(struct blk_mq_tags
*),
1959 GFP_KERNEL
, set
->numa_node
);
1963 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1964 set
->tags
[i
] = blk_mq_init_rq_map(set
, i
);
1969 mutex_init(&set
->tag_list_lock
);
1970 INIT_LIST_HEAD(&set
->tag_list
);
1976 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1980 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
1982 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
1986 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1988 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1993 EXPORT_SYMBOL(blk_mq_free_tag_set
);
1995 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
1997 struct blk_mq_tag_set
*set
= q
->tag_set
;
1998 struct blk_mq_hw_ctx
*hctx
;
2001 if (!set
|| nr
> set
->queue_depth
)
2005 queue_for_each_hw_ctx(q
, hctx
, i
) {
2006 ret
= blk_mq_tag_update_depth(hctx
->tags
, nr
);
2012 q
->nr_requests
= nr
;
2017 void blk_mq_disable_hotplug(void)
2019 mutex_lock(&all_q_mutex
);
2022 void blk_mq_enable_hotplug(void)
2024 mutex_unlock(&all_q_mutex
);
2027 static int __init
blk_mq_init(void)
2031 hotcpu_notifier(blk_mq_queue_reinit_notify
, 0);
2035 subsys_initcall(blk_mq_init
);