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
)
83 __percpu_counter_add(&q
->mq_usage_counter
, 1, 1000000);
86 /* we have problems freezing the queue if it's initializing */
87 if (!q
->mq_freeze_depth
)
90 __percpu_counter_add(&q
->mq_usage_counter
, -1, 1000000);
92 spin_lock_irq(q
->queue_lock
);
93 ret
= wait_event_interruptible_lock_irq(q
->mq_freeze_wq
,
94 !q
->mq_freeze_depth
|| blk_queue_dying(q
),
96 /* inc usage with lock hold to avoid freeze_queue runs here */
97 if (!ret
&& !blk_queue_dying(q
))
98 __percpu_counter_add(&q
->mq_usage_counter
, 1, 1000000);
99 else if (blk_queue_dying(q
))
101 spin_unlock_irq(q
->queue_lock
);
106 static void blk_mq_queue_exit(struct request_queue
*q
)
108 __percpu_counter_add(&q
->mq_usage_counter
, -1, 1000000);
112 * Guarantee no request is in use, so we can change any data structure of
113 * the queue afterward.
115 void blk_mq_freeze_queue(struct request_queue
*q
)
117 spin_lock_irq(q
->queue_lock
);
118 q
->mq_freeze_depth
++;
119 spin_unlock_irq(q
->queue_lock
);
124 spin_lock_irq(q
->queue_lock
);
125 count
= percpu_counter_sum(&q
->mq_usage_counter
);
126 spin_unlock_irq(q
->queue_lock
);
130 blk_mq_start_hw_queues(q
);
135 static void blk_mq_unfreeze_queue(struct request_queue
*q
)
139 spin_lock_irq(q
->queue_lock
);
140 wake
= !--q
->mq_freeze_depth
;
141 WARN_ON_ONCE(q
->mq_freeze_depth
< 0);
142 spin_unlock_irq(q
->queue_lock
);
144 wake_up_all(&q
->mq_freeze_wq
);
147 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
149 return blk_mq_has_free_tags(hctx
->tags
);
151 EXPORT_SYMBOL(blk_mq_can_queue
);
153 static void blk_mq_rq_ctx_init(struct request_queue
*q
, struct blk_mq_ctx
*ctx
,
154 struct request
*rq
, unsigned int rw_flags
)
156 if (blk_queue_io_stat(q
))
157 rw_flags
|= REQ_IO_STAT
;
159 INIT_LIST_HEAD(&rq
->queuelist
);
160 /* csd/requeue_work/fifo_time is initialized before use */
163 rq
->cmd_flags
|= rw_flags
;
164 /* do not touch atomic flags, it needs atomic ops against the timer */
166 INIT_HLIST_NODE(&rq
->hash
);
167 RB_CLEAR_NODE(&rq
->rb_node
);
170 rq
->start_time
= jiffies
;
171 #ifdef CONFIG_BLK_CGROUP
173 set_start_time_ns(rq
);
174 rq
->io_start_time_ns
= 0;
176 rq
->nr_phys_segments
= 0;
177 #if defined(CONFIG_BLK_DEV_INTEGRITY)
178 rq
->nr_integrity_segments
= 0;
181 /* tag was already set */
189 INIT_LIST_HEAD(&rq
->timeout_list
);
193 rq
->end_io_data
= NULL
;
196 ctx
->rq_dispatched
[rw_is_sync(rw_flags
)]++;
199 static struct request
*
200 __blk_mq_alloc_request(struct blk_mq_alloc_data
*data
, int rw
)
205 tag
= blk_mq_get_tag(data
);
206 if (tag
!= BLK_MQ_TAG_FAIL
) {
207 rq
= data
->hctx
->tags
->rqs
[tag
];
210 if (blk_mq_tag_busy(data
->hctx
)) {
211 rq
->cmd_flags
= REQ_MQ_INFLIGHT
;
212 atomic_inc(&data
->hctx
->nr_active
);
216 blk_mq_rq_ctx_init(data
->q
, data
->ctx
, rq
, rw
);
223 struct request
*blk_mq_alloc_request(struct request_queue
*q
, int rw
, gfp_t gfp
,
226 struct blk_mq_ctx
*ctx
;
227 struct blk_mq_hw_ctx
*hctx
;
229 struct blk_mq_alloc_data alloc_data
;
231 if (blk_mq_queue_enter(q
))
234 ctx
= blk_mq_get_ctx(q
);
235 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
236 blk_mq_set_alloc_data(&alloc_data
, q
, gfp
& ~__GFP_WAIT
,
237 reserved
, ctx
, hctx
);
239 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
240 if (!rq
&& (gfp
& __GFP_WAIT
)) {
241 __blk_mq_run_hw_queue(hctx
);
244 ctx
= blk_mq_get_ctx(q
);
245 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
246 blk_mq_set_alloc_data(&alloc_data
, q
, gfp
, reserved
, ctx
,
248 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
249 ctx
= alloc_data
.ctx
;
254 EXPORT_SYMBOL(blk_mq_alloc_request
);
256 static void __blk_mq_free_request(struct blk_mq_hw_ctx
*hctx
,
257 struct blk_mq_ctx
*ctx
, struct request
*rq
)
259 const int tag
= rq
->tag
;
260 struct request_queue
*q
= rq
->q
;
262 if (rq
->cmd_flags
& REQ_MQ_INFLIGHT
)
263 atomic_dec(&hctx
->nr_active
);
265 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
266 blk_mq_put_tag(hctx
, tag
, &ctx
->last_tag
);
267 blk_mq_queue_exit(q
);
270 void blk_mq_free_request(struct request
*rq
)
272 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
273 struct blk_mq_hw_ctx
*hctx
;
274 struct request_queue
*q
= rq
->q
;
276 ctx
->rq_completed
[rq_is_sync(rq
)]++;
278 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
279 __blk_mq_free_request(hctx
, ctx
, rq
);
283 * Clone all relevant state from a request that has been put on hold in
284 * the flush state machine into the preallocated flush request that hangs
285 * off the request queue.
287 * For a driver the flush request should be invisible, that's why we are
288 * impersonating the original request here.
290 void blk_mq_clone_flush_request(struct request
*flush_rq
,
291 struct request
*orig_rq
)
293 struct blk_mq_hw_ctx
*hctx
=
294 orig_rq
->q
->mq_ops
->map_queue(orig_rq
->q
, orig_rq
->mq_ctx
->cpu
);
296 flush_rq
->mq_ctx
= orig_rq
->mq_ctx
;
297 flush_rq
->tag
= orig_rq
->tag
;
298 memcpy(blk_mq_rq_to_pdu(flush_rq
), blk_mq_rq_to_pdu(orig_rq
),
302 inline void __blk_mq_end_io(struct request
*rq
, int error
)
304 blk_account_io_done(rq
);
307 rq
->end_io(rq
, error
);
309 if (unlikely(blk_bidi_rq(rq
)))
310 blk_mq_free_request(rq
->next_rq
);
311 blk_mq_free_request(rq
);
314 EXPORT_SYMBOL(__blk_mq_end_io
);
316 void blk_mq_end_io(struct request
*rq
, int error
)
318 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
320 __blk_mq_end_io(rq
, error
);
322 EXPORT_SYMBOL(blk_mq_end_io
);
324 static void __blk_mq_complete_request_remote(void *data
)
326 struct request
*rq
= data
;
328 rq
->q
->softirq_done_fn(rq
);
331 static void blk_mq_ipi_complete_request(struct request
*rq
)
333 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
337 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
338 rq
->q
->softirq_done_fn(rq
);
343 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
344 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
346 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
347 rq
->csd
.func
= __blk_mq_complete_request_remote
;
350 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
352 rq
->q
->softirq_done_fn(rq
);
357 void __blk_mq_complete_request(struct request
*rq
)
359 struct request_queue
*q
= rq
->q
;
361 if (!q
->softirq_done_fn
)
362 blk_mq_end_io(rq
, rq
->errors
);
364 blk_mq_ipi_complete_request(rq
);
368 * blk_mq_complete_request - end I/O on a request
369 * @rq: the request being processed
372 * Ends all I/O on a request. It does not handle partial completions.
373 * The actual completion happens out-of-order, through a IPI handler.
375 void blk_mq_complete_request(struct request
*rq
)
377 struct request_queue
*q
= rq
->q
;
379 if (unlikely(blk_should_fake_timeout(q
)))
381 if (!blk_mark_rq_complete(rq
))
382 __blk_mq_complete_request(rq
);
384 EXPORT_SYMBOL(blk_mq_complete_request
);
386 static void blk_mq_start_request(struct request
*rq
, bool last
)
388 struct request_queue
*q
= rq
->q
;
390 trace_block_rq_issue(q
, rq
);
392 rq
->resid_len
= blk_rq_bytes(rq
);
393 if (unlikely(blk_bidi_rq(rq
)))
394 rq
->next_rq
->resid_len
= blk_rq_bytes(rq
->next_rq
);
399 * Mark us as started and clear complete. Complete might have been
400 * set if requeue raced with timeout, which then marked it as
401 * complete. So be sure to clear complete again when we start
402 * the request, otherwise we'll ignore the completion event.
404 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
405 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
406 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
407 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
409 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
411 * Make sure space for the drain appears. We know we can do
412 * this because max_hw_segments has been adjusted to be one
413 * fewer than the device can handle.
415 rq
->nr_phys_segments
++;
419 * Flag the last request in the series so that drivers know when IO
420 * should be kicked off, if they don't do it on a per-request basis.
422 * Note: the flag isn't the only condition drivers should do kick off.
423 * If drive is busy, the last request might not have the bit set.
426 rq
->cmd_flags
|= REQ_END
;
429 static void __blk_mq_requeue_request(struct request
*rq
)
431 struct request_queue
*q
= rq
->q
;
433 trace_block_rq_requeue(q
, rq
);
434 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
436 rq
->cmd_flags
&= ~REQ_END
;
438 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
439 rq
->nr_phys_segments
--;
442 void blk_mq_requeue_request(struct request
*rq
)
444 __blk_mq_requeue_request(rq
);
445 blk_clear_rq_complete(rq
);
447 BUG_ON(blk_queued_rq(rq
));
448 blk_mq_add_to_requeue_list(rq
, true);
450 EXPORT_SYMBOL(blk_mq_requeue_request
);
452 static void blk_mq_requeue_work(struct work_struct
*work
)
454 struct request_queue
*q
=
455 container_of(work
, struct request_queue
, requeue_work
);
457 struct request
*rq
, *next
;
460 spin_lock_irqsave(&q
->requeue_lock
, flags
);
461 list_splice_init(&q
->requeue_list
, &rq_list
);
462 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
464 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
465 if (!(rq
->cmd_flags
& REQ_SOFTBARRIER
))
468 rq
->cmd_flags
&= ~REQ_SOFTBARRIER
;
469 list_del_init(&rq
->queuelist
);
470 blk_mq_insert_request(rq
, true, false, false);
473 while (!list_empty(&rq_list
)) {
474 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
475 list_del_init(&rq
->queuelist
);
476 blk_mq_insert_request(rq
, false, false, false);
479 blk_mq_run_queues(q
, false);
482 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
)
484 struct request_queue
*q
= rq
->q
;
488 * We abuse this flag that is otherwise used by the I/O scheduler to
489 * request head insertation from the workqueue.
491 BUG_ON(rq
->cmd_flags
& REQ_SOFTBARRIER
);
493 spin_lock_irqsave(&q
->requeue_lock
, flags
);
495 rq
->cmd_flags
|= REQ_SOFTBARRIER
;
496 list_add(&rq
->queuelist
, &q
->requeue_list
);
498 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
500 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
502 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
504 void blk_mq_kick_requeue_list(struct request_queue
*q
)
506 kblockd_schedule_work(&q
->requeue_work
);
508 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
510 static inline bool is_flush_request(struct request
*rq
, unsigned int tag
)
512 return ((rq
->cmd_flags
& REQ_FLUSH_SEQ
) &&
513 rq
->q
->flush_rq
->tag
== tag
);
516 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
518 struct request
*rq
= tags
->rqs
[tag
];
520 if (!is_flush_request(rq
, tag
))
523 return rq
->q
->flush_rq
;
525 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
527 struct blk_mq_timeout_data
{
528 struct blk_mq_hw_ctx
*hctx
;
530 unsigned int *next_set
;
533 static void blk_mq_timeout_check(void *__data
, unsigned long *free_tags
)
535 struct blk_mq_timeout_data
*data
= __data
;
536 struct blk_mq_hw_ctx
*hctx
= data
->hctx
;
539 /* It may not be in flight yet (this is where
540 * the REQ_ATOMIC_STARTED flag comes in). The requests are
541 * statically allocated, so we know it's always safe to access the
542 * memory associated with a bit offset into ->rqs[].
548 tag
= find_next_zero_bit(free_tags
, hctx
->tags
->nr_tags
, tag
);
549 if (tag
>= hctx
->tags
->nr_tags
)
552 rq
= blk_mq_tag_to_rq(hctx
->tags
, tag
++);
553 if (rq
->q
!= hctx
->queue
)
555 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
558 blk_rq_check_expired(rq
, data
->next
, data
->next_set
);
562 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx
*hctx
,
564 unsigned int *next_set
)
566 struct blk_mq_timeout_data data
= {
569 .next_set
= next_set
,
573 * Ask the tagging code to iterate busy requests, so we can
574 * check them for timeout.
576 blk_mq_tag_busy_iter(hctx
->tags
, blk_mq_timeout_check
, &data
);
579 static enum blk_eh_timer_return
blk_mq_rq_timed_out(struct request
*rq
)
581 struct request_queue
*q
= rq
->q
;
584 * We know that complete is set at this point. If STARTED isn't set
585 * anymore, then the request isn't active and the "timeout" should
586 * just be ignored. This can happen due to the bitflag ordering.
587 * Timeout first checks if STARTED is set, and if it is, assumes
588 * the request is active. But if we race with completion, then
589 * we both flags will get cleared. So check here again, and ignore
590 * a timeout event with a request that isn't active.
592 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
593 return BLK_EH_NOT_HANDLED
;
595 if (!q
->mq_ops
->timeout
)
596 return BLK_EH_RESET_TIMER
;
598 return q
->mq_ops
->timeout(rq
);
601 static void blk_mq_rq_timer(unsigned long data
)
603 struct request_queue
*q
= (struct request_queue
*) data
;
604 struct blk_mq_hw_ctx
*hctx
;
605 unsigned long next
= 0;
608 queue_for_each_hw_ctx(q
, hctx
, i
) {
610 * If not software queues are currently mapped to this
611 * hardware queue, there's nothing to check
613 if (!hctx
->nr_ctx
|| !hctx
->tags
)
616 blk_mq_hw_ctx_check_timeout(hctx
, &next
, &next_set
);
620 next
= blk_rq_timeout(round_jiffies_up(next
));
621 mod_timer(&q
->timeout
, next
);
623 queue_for_each_hw_ctx(q
, hctx
, i
)
624 blk_mq_tag_idle(hctx
);
629 * Reverse check our software queue for entries that we could potentially
630 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
631 * too much time checking for merges.
633 static bool blk_mq_attempt_merge(struct request_queue
*q
,
634 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
639 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
645 if (!blk_rq_merge_ok(rq
, bio
))
648 el_ret
= blk_try_merge(rq
, bio
);
649 if (el_ret
== ELEVATOR_BACK_MERGE
) {
650 if (bio_attempt_back_merge(q
, rq
, bio
)) {
655 } else if (el_ret
== ELEVATOR_FRONT_MERGE
) {
656 if (bio_attempt_front_merge(q
, rq
, bio
)) {
668 * Process software queues that have been marked busy, splicing them
669 * to the for-dispatch
671 static void flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
673 struct blk_mq_ctx
*ctx
;
676 for (i
= 0; i
< hctx
->ctx_map
.map_size
; i
++) {
677 struct blk_align_bitmap
*bm
= &hctx
->ctx_map
.map
[i
];
678 unsigned int off
, bit
;
684 off
= i
* hctx
->ctx_map
.bits_per_word
;
686 bit
= find_next_bit(&bm
->word
, bm
->depth
, bit
);
687 if (bit
>= bm
->depth
)
690 ctx
= hctx
->ctxs
[bit
+ off
];
691 clear_bit(bit
, &bm
->word
);
692 spin_lock(&ctx
->lock
);
693 list_splice_tail_init(&ctx
->rq_list
, list
);
694 spin_unlock(&ctx
->lock
);
702 * Run this hardware queue, pulling any software queues mapped to it in.
703 * Note that this function currently has various problems around ordering
704 * of IO. In particular, we'd like FIFO behaviour on handling existing
705 * items on the hctx->dispatch list. Ignore that for now.
707 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
709 struct request_queue
*q
= hctx
->queue
;
714 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
));
716 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)))
722 * Touch any software queue that has pending entries.
724 flush_busy_ctxs(hctx
, &rq_list
);
727 * If we have previous entries on our dispatch list, grab them
728 * and stuff them at the front for more fair dispatch.
730 if (!list_empty_careful(&hctx
->dispatch
)) {
731 spin_lock(&hctx
->lock
);
732 if (!list_empty(&hctx
->dispatch
))
733 list_splice_init(&hctx
->dispatch
, &rq_list
);
734 spin_unlock(&hctx
->lock
);
738 * Now process all the entries, sending them to the driver.
741 while (!list_empty(&rq_list
)) {
744 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
745 list_del_init(&rq
->queuelist
);
747 blk_mq_start_request(rq
, list_empty(&rq_list
));
749 ret
= q
->mq_ops
->queue_rq(hctx
, rq
);
751 case BLK_MQ_RQ_QUEUE_OK
:
754 case BLK_MQ_RQ_QUEUE_BUSY
:
755 list_add(&rq
->queuelist
, &rq_list
);
756 __blk_mq_requeue_request(rq
);
759 pr_err("blk-mq: bad return on queue: %d\n", ret
);
760 case BLK_MQ_RQ_QUEUE_ERROR
:
762 blk_mq_end_io(rq
, rq
->errors
);
766 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
771 hctx
->dispatched
[0]++;
772 else if (queued
< (1 << (BLK_MQ_MAX_DISPATCH_ORDER
- 1)))
773 hctx
->dispatched
[ilog2(queued
) + 1]++;
776 * Any items that need requeuing? Stuff them into hctx->dispatch,
777 * that is where we will continue on next queue run.
779 if (!list_empty(&rq_list
)) {
780 spin_lock(&hctx
->lock
);
781 list_splice(&rq_list
, &hctx
->dispatch
);
782 spin_unlock(&hctx
->lock
);
787 * It'd be great if the workqueue API had a way to pass
788 * in a mask and had some smarts for more clever placement.
789 * For now we just round-robin here, switching for every
790 * BLK_MQ_CPU_WORK_BATCH queued items.
792 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
794 int cpu
= hctx
->next_cpu
;
796 if (--hctx
->next_cpu_batch
<= 0) {
799 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
800 if (next_cpu
>= nr_cpu_ids
)
801 next_cpu
= cpumask_first(hctx
->cpumask
);
803 hctx
->next_cpu
= next_cpu
;
804 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
810 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
812 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)))
815 if (!async
&& cpumask_test_cpu(smp_processor_id(), hctx
->cpumask
))
816 __blk_mq_run_hw_queue(hctx
);
817 else if (hctx
->queue
->nr_hw_queues
== 1)
818 kblockd_schedule_delayed_work(&hctx
->run_work
, 0);
822 cpu
= blk_mq_hctx_next_cpu(hctx
);
823 kblockd_schedule_delayed_work_on(cpu
, &hctx
->run_work
, 0);
827 void blk_mq_run_queues(struct request_queue
*q
, bool async
)
829 struct blk_mq_hw_ctx
*hctx
;
832 queue_for_each_hw_ctx(q
, hctx
, i
) {
833 if ((!blk_mq_hctx_has_pending(hctx
) &&
834 list_empty_careful(&hctx
->dispatch
)) ||
835 test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
839 blk_mq_run_hw_queue(hctx
, async
);
843 EXPORT_SYMBOL(blk_mq_run_queues
);
845 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
847 cancel_delayed_work(&hctx
->run_work
);
848 cancel_delayed_work(&hctx
->delay_work
);
849 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
851 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
853 void blk_mq_stop_hw_queues(struct request_queue
*q
)
855 struct blk_mq_hw_ctx
*hctx
;
858 queue_for_each_hw_ctx(q
, hctx
, i
)
859 blk_mq_stop_hw_queue(hctx
);
861 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
863 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
865 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
868 blk_mq_run_hw_queue(hctx
, false);
871 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
873 void blk_mq_start_hw_queues(struct request_queue
*q
)
875 struct blk_mq_hw_ctx
*hctx
;
878 queue_for_each_hw_ctx(q
, hctx
, i
)
879 blk_mq_start_hw_queue(hctx
);
881 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
884 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
886 struct blk_mq_hw_ctx
*hctx
;
889 queue_for_each_hw_ctx(q
, hctx
, i
) {
890 if (!test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
893 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
895 blk_mq_run_hw_queue(hctx
, async
);
899 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
901 static void blk_mq_run_work_fn(struct work_struct
*work
)
903 struct blk_mq_hw_ctx
*hctx
;
905 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
907 __blk_mq_run_hw_queue(hctx
);
910 static void blk_mq_delay_work_fn(struct work_struct
*work
)
912 struct blk_mq_hw_ctx
*hctx
;
914 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
916 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
917 __blk_mq_run_hw_queue(hctx
);
920 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
922 unsigned long tmo
= msecs_to_jiffies(msecs
);
924 if (hctx
->queue
->nr_hw_queues
== 1)
925 kblockd_schedule_delayed_work(&hctx
->delay_work
, tmo
);
929 cpu
= blk_mq_hctx_next_cpu(hctx
);
930 kblockd_schedule_delayed_work_on(cpu
, &hctx
->delay_work
, tmo
);
933 EXPORT_SYMBOL(blk_mq_delay_queue
);
935 static void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
,
936 struct request
*rq
, bool at_head
)
938 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
940 trace_block_rq_insert(hctx
->queue
, rq
);
943 list_add(&rq
->queuelist
, &ctx
->rq_list
);
945 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
947 blk_mq_hctx_mark_pending(hctx
, ctx
);
950 void blk_mq_insert_request(struct request
*rq
, bool at_head
, bool run_queue
,
953 struct request_queue
*q
= rq
->q
;
954 struct blk_mq_hw_ctx
*hctx
;
955 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
, *current_ctx
;
957 current_ctx
= blk_mq_get_ctx(q
);
958 if (!cpu_online(ctx
->cpu
))
959 rq
->mq_ctx
= ctx
= current_ctx
;
961 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
963 if (rq
->cmd_flags
& (REQ_FLUSH
| REQ_FUA
) &&
964 !(rq
->cmd_flags
& (REQ_FLUSH_SEQ
))) {
965 blk_insert_flush(rq
);
967 spin_lock(&ctx
->lock
);
968 __blk_mq_insert_request(hctx
, rq
, at_head
);
969 spin_unlock(&ctx
->lock
);
973 blk_mq_run_hw_queue(hctx
, async
);
975 blk_mq_put_ctx(current_ctx
);
978 static void blk_mq_insert_requests(struct request_queue
*q
,
979 struct blk_mq_ctx
*ctx
,
980 struct list_head
*list
,
985 struct blk_mq_hw_ctx
*hctx
;
986 struct blk_mq_ctx
*current_ctx
;
988 trace_block_unplug(q
, depth
, !from_schedule
);
990 current_ctx
= blk_mq_get_ctx(q
);
992 if (!cpu_online(ctx
->cpu
))
994 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
997 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1000 spin_lock(&ctx
->lock
);
1001 while (!list_empty(list
)) {
1004 rq
= list_first_entry(list
, struct request
, queuelist
);
1005 list_del_init(&rq
->queuelist
);
1007 __blk_mq_insert_request(hctx
, rq
, false);
1009 spin_unlock(&ctx
->lock
);
1011 blk_mq_run_hw_queue(hctx
, from_schedule
);
1012 blk_mq_put_ctx(current_ctx
);
1015 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1017 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1018 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1020 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1021 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1022 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1025 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1027 struct blk_mq_ctx
*this_ctx
;
1028 struct request_queue
*this_q
;
1031 LIST_HEAD(ctx_list
);
1034 list_splice_init(&plug
->mq_list
, &list
);
1036 list_sort(NULL
, &list
, plug_ctx_cmp
);
1042 while (!list_empty(&list
)) {
1043 rq
= list_entry_rq(list
.next
);
1044 list_del_init(&rq
->queuelist
);
1046 if (rq
->mq_ctx
!= this_ctx
) {
1048 blk_mq_insert_requests(this_q
, this_ctx
,
1053 this_ctx
= rq
->mq_ctx
;
1059 list_add_tail(&rq
->queuelist
, &ctx_list
);
1063 * If 'this_ctx' is set, we know we have entries to complete
1064 * on 'ctx_list'. Do those.
1067 blk_mq_insert_requests(this_q
, this_ctx
, &ctx_list
, depth
,
1072 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1074 init_request_from_bio(rq
, bio
);
1076 if (blk_do_io_stat(rq
))
1077 blk_account_io_start(rq
, 1);
1080 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx
*hctx
,
1081 struct blk_mq_ctx
*ctx
,
1082 struct request
*rq
, struct bio
*bio
)
1084 struct request_queue
*q
= hctx
->queue
;
1086 if (!(hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
)) {
1087 blk_mq_bio_to_request(rq
, bio
);
1088 spin_lock(&ctx
->lock
);
1090 __blk_mq_insert_request(hctx
, rq
, false);
1091 spin_unlock(&ctx
->lock
);
1094 spin_lock(&ctx
->lock
);
1095 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1096 blk_mq_bio_to_request(rq
, bio
);
1100 spin_unlock(&ctx
->lock
);
1101 __blk_mq_free_request(hctx
, ctx
, rq
);
1106 struct blk_map_ctx
{
1107 struct blk_mq_hw_ctx
*hctx
;
1108 struct blk_mq_ctx
*ctx
;
1111 static struct request
*blk_mq_map_request(struct request_queue
*q
,
1113 struct blk_map_ctx
*data
)
1115 struct blk_mq_hw_ctx
*hctx
;
1116 struct blk_mq_ctx
*ctx
;
1118 int rw
= bio_data_dir(bio
);
1119 struct blk_mq_alloc_data alloc_data
;
1121 if (unlikely(blk_mq_queue_enter(q
))) {
1122 bio_endio(bio
, -EIO
);
1126 ctx
= blk_mq_get_ctx(q
);
1127 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1129 if (rw_is_sync(bio
->bi_rw
))
1132 trace_block_getrq(q
, bio
, rw
);
1133 blk_mq_set_alloc_data(&alloc_data
, q
, GFP_ATOMIC
, false, ctx
,
1135 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
1136 if (unlikely(!rq
)) {
1137 __blk_mq_run_hw_queue(hctx
);
1138 blk_mq_put_ctx(ctx
);
1139 trace_block_sleeprq(q
, bio
, rw
);
1141 ctx
= blk_mq_get_ctx(q
);
1142 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1143 blk_mq_set_alloc_data(&alloc_data
, q
,
1144 __GFP_WAIT
|GFP_ATOMIC
, false, ctx
, hctx
);
1145 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
1146 ctx
= alloc_data
.ctx
;
1147 hctx
= alloc_data
.hctx
;
1157 * Multiple hardware queue variant. This will not use per-process plugs,
1158 * but will attempt to bypass the hctx queueing if we can go straight to
1159 * hardware for SYNC IO.
1161 static void blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1163 const int is_sync
= rw_is_sync(bio
->bi_rw
);
1164 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
1165 struct blk_map_ctx data
;
1168 blk_queue_bounce(q
, &bio
);
1170 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1171 bio_endio(bio
, -EIO
);
1175 rq
= blk_mq_map_request(q
, bio
, &data
);
1179 if (unlikely(is_flush_fua
)) {
1180 blk_mq_bio_to_request(rq
, bio
);
1181 blk_insert_flush(rq
);
1188 blk_mq_bio_to_request(rq
, bio
);
1189 blk_mq_start_request(rq
, true);
1192 * For OK queue, we are done. For error, kill it. Any other
1193 * error (busy), just add it to our list as we previously
1196 ret
= q
->mq_ops
->queue_rq(data
.hctx
, rq
);
1197 if (ret
== BLK_MQ_RQ_QUEUE_OK
)
1200 __blk_mq_requeue_request(rq
);
1202 if (ret
== BLK_MQ_RQ_QUEUE_ERROR
) {
1204 blk_mq_end_io(rq
, rq
->errors
);
1210 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1212 * For a SYNC request, send it to the hardware immediately. For
1213 * an ASYNC request, just ensure that we run it later on. The
1214 * latter allows for merging opportunities and more efficient
1218 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1221 blk_mq_put_ctx(data
.ctx
);
1225 * Single hardware queue variant. This will attempt to use any per-process
1226 * plug for merging and IO deferral.
1228 static void blk_sq_make_request(struct request_queue
*q
, struct bio
*bio
)
1230 const int is_sync
= rw_is_sync(bio
->bi_rw
);
1231 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
1232 unsigned int use_plug
, request_count
= 0;
1233 struct blk_map_ctx data
;
1237 * If we have multiple hardware queues, just go directly to
1238 * one of those for sync IO.
1240 use_plug
= !is_flush_fua
&& !is_sync
;
1242 blk_queue_bounce(q
, &bio
);
1244 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1245 bio_endio(bio
, -EIO
);
1249 if (use_plug
&& !blk_queue_nomerges(q
) &&
1250 blk_attempt_plug_merge(q
, bio
, &request_count
))
1253 rq
= blk_mq_map_request(q
, bio
, &data
);
1257 if (unlikely(is_flush_fua
)) {
1258 blk_mq_bio_to_request(rq
, bio
);
1259 blk_insert_flush(rq
);
1264 * A task plug currently exists. Since this is completely lockless,
1265 * utilize that to temporarily store requests until the task is
1266 * either done or scheduled away.
1269 struct blk_plug
*plug
= current
->plug
;
1272 blk_mq_bio_to_request(rq
, bio
);
1273 if (list_empty(&plug
->mq_list
))
1274 trace_block_plug(q
);
1275 else if (request_count
>= BLK_MAX_REQUEST_COUNT
) {
1276 blk_flush_plug_list(plug
, false);
1277 trace_block_plug(q
);
1279 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1280 blk_mq_put_ctx(data
.ctx
);
1285 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1287 * For a SYNC request, send it to the hardware immediately. For
1288 * an ASYNC request, just ensure that we run it later on. The
1289 * latter allows for merging opportunities and more efficient
1293 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1296 blk_mq_put_ctx(data
.ctx
);
1300 * Default mapping to a software queue, since we use one per CPU.
1302 struct blk_mq_hw_ctx
*blk_mq_map_queue(struct request_queue
*q
, const int cpu
)
1304 return q
->queue_hw_ctx
[q
->mq_map
[cpu
]];
1306 EXPORT_SYMBOL(blk_mq_map_queue
);
1308 static void blk_mq_free_rq_map(struct blk_mq_tag_set
*set
,
1309 struct blk_mq_tags
*tags
, unsigned int hctx_idx
)
1313 if (tags
->rqs
&& set
->ops
->exit_request
) {
1316 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1319 set
->ops
->exit_request(set
->driver_data
, tags
->rqs
[i
],
1324 while (!list_empty(&tags
->page_list
)) {
1325 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1326 list_del_init(&page
->lru
);
1327 __free_pages(page
, page
->private);
1332 blk_mq_free_tags(tags
);
1335 static size_t order_to_size(unsigned int order
)
1337 return (size_t)PAGE_SIZE
<< order
;
1340 static struct blk_mq_tags
*blk_mq_init_rq_map(struct blk_mq_tag_set
*set
,
1341 unsigned int hctx_idx
)
1343 struct blk_mq_tags
*tags
;
1344 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1345 size_t rq_size
, left
;
1347 tags
= blk_mq_init_tags(set
->queue_depth
, set
->reserved_tags
,
1352 INIT_LIST_HEAD(&tags
->page_list
);
1354 tags
->rqs
= kmalloc_node(set
->queue_depth
* sizeof(struct request
*),
1355 GFP_KERNEL
, set
->numa_node
);
1357 blk_mq_free_tags(tags
);
1362 * rq_size is the size of the request plus driver payload, rounded
1363 * to the cacheline size
1365 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1367 left
= rq_size
* set
->queue_depth
;
1369 for (i
= 0; i
< set
->queue_depth
; ) {
1370 int this_order
= max_order
;
1375 while (left
< order_to_size(this_order
- 1) && this_order
)
1379 page
= alloc_pages_node(set
->numa_node
, GFP_KERNEL
,
1385 if (order_to_size(this_order
) < rq_size
)
1392 page
->private = this_order
;
1393 list_add_tail(&page
->lru
, &tags
->page_list
);
1395 p
= page_address(page
);
1396 entries_per_page
= order_to_size(this_order
) / rq_size
;
1397 to_do
= min(entries_per_page
, set
->queue_depth
- i
);
1398 left
-= to_do
* rq_size
;
1399 for (j
= 0; j
< to_do
; j
++) {
1401 if (set
->ops
->init_request
) {
1402 if (set
->ops
->init_request(set
->driver_data
,
1403 tags
->rqs
[i
], hctx_idx
, i
,
1416 pr_warn("%s: failed to allocate requests\n", __func__
);
1417 blk_mq_free_rq_map(set
, tags
, hctx_idx
);
1421 static void blk_mq_free_bitmap(struct blk_mq_ctxmap
*bitmap
)
1426 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap
*bitmap
, int node
)
1428 unsigned int bpw
= 8, total
, num_maps
, i
;
1430 bitmap
->bits_per_word
= bpw
;
1432 num_maps
= ALIGN(nr_cpu_ids
, bpw
) / bpw
;
1433 bitmap
->map
= kzalloc_node(num_maps
* sizeof(struct blk_align_bitmap
),
1438 bitmap
->map_size
= num_maps
;
1441 for (i
= 0; i
< num_maps
; i
++) {
1442 bitmap
->map
[i
].depth
= min(total
, bitmap
->bits_per_word
);
1443 total
-= bitmap
->map
[i
].depth
;
1449 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx
*hctx
, int cpu
)
1451 struct request_queue
*q
= hctx
->queue
;
1452 struct blk_mq_ctx
*ctx
;
1456 * Move ctx entries to new CPU, if this one is going away.
1458 ctx
= __blk_mq_get_ctx(q
, cpu
);
1460 spin_lock(&ctx
->lock
);
1461 if (!list_empty(&ctx
->rq_list
)) {
1462 list_splice_init(&ctx
->rq_list
, &tmp
);
1463 blk_mq_hctx_clear_pending(hctx
, ctx
);
1465 spin_unlock(&ctx
->lock
);
1467 if (list_empty(&tmp
))
1470 ctx
= blk_mq_get_ctx(q
);
1471 spin_lock(&ctx
->lock
);
1473 while (!list_empty(&tmp
)) {
1476 rq
= list_first_entry(&tmp
, struct request
, queuelist
);
1478 list_move_tail(&rq
->queuelist
, &ctx
->rq_list
);
1481 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1482 blk_mq_hctx_mark_pending(hctx
, ctx
);
1484 spin_unlock(&ctx
->lock
);
1486 blk_mq_run_hw_queue(hctx
, true);
1487 blk_mq_put_ctx(ctx
);
1491 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx
*hctx
, int cpu
)
1493 struct request_queue
*q
= hctx
->queue
;
1494 struct blk_mq_tag_set
*set
= q
->tag_set
;
1496 if (set
->tags
[hctx
->queue_num
])
1499 set
->tags
[hctx
->queue_num
] = blk_mq_init_rq_map(set
, hctx
->queue_num
);
1500 if (!set
->tags
[hctx
->queue_num
])
1503 hctx
->tags
= set
->tags
[hctx
->queue_num
];
1507 static int blk_mq_hctx_notify(void *data
, unsigned long action
,
1510 struct blk_mq_hw_ctx
*hctx
= data
;
1512 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
)
1513 return blk_mq_hctx_cpu_offline(hctx
, cpu
);
1514 else if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
)
1515 return blk_mq_hctx_cpu_online(hctx
, cpu
);
1520 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1521 struct blk_mq_tag_set
*set
, int nr_queue
)
1523 struct blk_mq_hw_ctx
*hctx
;
1526 queue_for_each_hw_ctx(q
, hctx
, i
) {
1530 blk_mq_tag_idle(hctx
);
1532 if (set
->ops
->exit_hctx
)
1533 set
->ops
->exit_hctx(hctx
, i
);
1535 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1537 blk_mq_free_bitmap(&hctx
->ctx_map
);
1542 static void blk_mq_free_hw_queues(struct request_queue
*q
,
1543 struct blk_mq_tag_set
*set
)
1545 struct blk_mq_hw_ctx
*hctx
;
1548 queue_for_each_hw_ctx(q
, hctx
, i
) {
1549 free_cpumask_var(hctx
->cpumask
);
1554 static int blk_mq_init_hw_queues(struct request_queue
*q
,
1555 struct blk_mq_tag_set
*set
)
1557 struct blk_mq_hw_ctx
*hctx
;
1561 * Initialize hardware queues
1563 queue_for_each_hw_ctx(q
, hctx
, i
) {
1566 node
= hctx
->numa_node
;
1567 if (node
== NUMA_NO_NODE
)
1568 node
= hctx
->numa_node
= set
->numa_node
;
1570 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1571 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1572 spin_lock_init(&hctx
->lock
);
1573 INIT_LIST_HEAD(&hctx
->dispatch
);
1575 hctx
->queue_num
= i
;
1576 hctx
->flags
= set
->flags
;
1577 hctx
->cmd_size
= set
->cmd_size
;
1579 blk_mq_init_cpu_notifier(&hctx
->cpu_notifier
,
1580 blk_mq_hctx_notify
, hctx
);
1581 blk_mq_register_cpu_notifier(&hctx
->cpu_notifier
);
1583 hctx
->tags
= set
->tags
[i
];
1586 * Allocate space for all possible cpus to avoid allocation in
1589 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1594 if (blk_mq_alloc_bitmap(&hctx
->ctx_map
, node
))
1599 if (set
->ops
->init_hctx
&&
1600 set
->ops
->init_hctx(hctx
, set
->driver_data
, i
))
1604 if (i
== q
->nr_hw_queues
)
1610 blk_mq_exit_hw_queues(q
, set
, i
);
1615 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1616 unsigned int nr_hw_queues
)
1620 for_each_possible_cpu(i
) {
1621 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1622 struct blk_mq_hw_ctx
*hctx
;
1624 memset(__ctx
, 0, sizeof(*__ctx
));
1626 spin_lock_init(&__ctx
->lock
);
1627 INIT_LIST_HEAD(&__ctx
->rq_list
);
1630 /* If the cpu isn't online, the cpu is mapped to first hctx */
1634 hctx
= q
->mq_ops
->map_queue(q
, i
);
1635 cpumask_set_cpu(i
, hctx
->cpumask
);
1639 * Set local node, IFF we have more than one hw queue. If
1640 * not, we remain on the home node of the device
1642 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1643 hctx
->numa_node
= cpu_to_node(i
);
1647 static void blk_mq_map_swqueue(struct request_queue
*q
)
1650 struct blk_mq_hw_ctx
*hctx
;
1651 struct blk_mq_ctx
*ctx
;
1653 queue_for_each_hw_ctx(q
, hctx
, i
) {
1654 cpumask_clear(hctx
->cpumask
);
1659 * Map software to hardware queues
1661 queue_for_each_ctx(q
, ctx
, i
) {
1662 /* If the cpu isn't online, the cpu is mapped to first hctx */
1666 hctx
= q
->mq_ops
->map_queue(q
, i
);
1667 cpumask_set_cpu(i
, hctx
->cpumask
);
1668 ctx
->index_hw
= hctx
->nr_ctx
;
1669 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
1672 queue_for_each_hw_ctx(q
, hctx
, i
) {
1674 * If not software queues are mapped to this hardware queue,
1675 * disable it and free the request entries
1677 if (!hctx
->nr_ctx
) {
1678 struct blk_mq_tag_set
*set
= q
->tag_set
;
1681 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1682 set
->tags
[i
] = NULL
;
1689 * Initialize batch roundrobin counts
1691 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
1692 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1696 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
)
1698 struct blk_mq_hw_ctx
*hctx
;
1699 struct request_queue
*q
;
1703 if (set
->tag_list
.next
== set
->tag_list
.prev
)
1708 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
1709 blk_mq_freeze_queue(q
);
1711 queue_for_each_hw_ctx(q
, hctx
, i
) {
1713 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
1715 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
1717 blk_mq_unfreeze_queue(q
);
1721 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
1723 struct blk_mq_tag_set
*set
= q
->tag_set
;
1725 blk_mq_freeze_queue(q
);
1727 mutex_lock(&set
->tag_list_lock
);
1728 list_del_init(&q
->tag_set_list
);
1729 blk_mq_update_tag_set_depth(set
);
1730 mutex_unlock(&set
->tag_list_lock
);
1732 blk_mq_unfreeze_queue(q
);
1735 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
1736 struct request_queue
*q
)
1740 mutex_lock(&set
->tag_list_lock
);
1741 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
1742 blk_mq_update_tag_set_depth(set
);
1743 mutex_unlock(&set
->tag_list_lock
);
1746 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
1748 struct blk_mq_hw_ctx
**hctxs
;
1749 struct blk_mq_ctx __percpu
*ctx
;
1750 struct request_queue
*q
;
1754 ctx
= alloc_percpu(struct blk_mq_ctx
);
1756 return ERR_PTR(-ENOMEM
);
1758 hctxs
= kmalloc_node(set
->nr_hw_queues
* sizeof(*hctxs
), GFP_KERNEL
,
1764 map
= blk_mq_make_queue_map(set
);
1768 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1769 int node
= blk_mq_hw_queue_to_node(map
, i
);
1771 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
1776 if (!zalloc_cpumask_var(&hctxs
[i
]->cpumask
, GFP_KERNEL
))
1779 atomic_set(&hctxs
[i
]->nr_active
, 0);
1780 hctxs
[i
]->numa_node
= node
;
1781 hctxs
[i
]->queue_num
= i
;
1784 q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
1788 if (percpu_counter_init(&q
->mq_usage_counter
, 0))
1791 setup_timer(&q
->timeout
, blk_mq_rq_timer
, (unsigned long) q
);
1792 blk_queue_rq_timeout(q
, 30000);
1794 q
->nr_queues
= nr_cpu_ids
;
1795 q
->nr_hw_queues
= set
->nr_hw_queues
;
1799 q
->queue_hw_ctx
= hctxs
;
1801 q
->mq_ops
= set
->ops
;
1802 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
1804 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
1805 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
1807 q
->sg_reserved_size
= INT_MAX
;
1809 INIT_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
1810 INIT_LIST_HEAD(&q
->requeue_list
);
1811 spin_lock_init(&q
->requeue_lock
);
1813 if (q
->nr_hw_queues
> 1)
1814 blk_queue_make_request(q
, blk_mq_make_request
);
1816 blk_queue_make_request(q
, blk_sq_make_request
);
1818 blk_queue_rq_timed_out(q
, blk_mq_rq_timed_out
);
1820 blk_queue_rq_timeout(q
, set
->timeout
);
1823 * Do this after blk_queue_make_request() overrides it...
1825 q
->nr_requests
= set
->queue_depth
;
1827 if (set
->ops
->complete
)
1828 blk_queue_softirq_done(q
, set
->ops
->complete
);
1830 blk_mq_init_flush(q
);
1831 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
1833 q
->flush_rq
= kzalloc(round_up(sizeof(struct request
) +
1834 set
->cmd_size
, cache_line_size()),
1839 if (blk_mq_init_hw_queues(q
, set
))
1842 mutex_lock(&all_q_mutex
);
1843 list_add_tail(&q
->all_q_node
, &all_q_list
);
1844 mutex_unlock(&all_q_mutex
);
1846 blk_mq_add_queue_tag_set(set
, q
);
1848 blk_mq_map_swqueue(q
);
1855 blk_cleanup_queue(q
);
1858 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1861 free_cpumask_var(hctxs
[i
]->cpumask
);
1868 return ERR_PTR(-ENOMEM
);
1870 EXPORT_SYMBOL(blk_mq_init_queue
);
1872 void blk_mq_free_queue(struct request_queue
*q
)
1874 struct blk_mq_tag_set
*set
= q
->tag_set
;
1876 blk_mq_del_queue_tag_set(q
);
1878 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
1879 blk_mq_free_hw_queues(q
, set
);
1881 percpu_counter_destroy(&q
->mq_usage_counter
);
1883 free_percpu(q
->queue_ctx
);
1884 kfree(q
->queue_hw_ctx
);
1887 q
->queue_ctx
= NULL
;
1888 q
->queue_hw_ctx
= NULL
;
1891 mutex_lock(&all_q_mutex
);
1892 list_del_init(&q
->all_q_node
);
1893 mutex_unlock(&all_q_mutex
);
1896 /* Basically redo blk_mq_init_queue with queue frozen */
1897 static void blk_mq_queue_reinit(struct request_queue
*q
)
1899 blk_mq_freeze_queue(q
);
1901 blk_mq_sysfs_unregister(q
);
1903 blk_mq_update_queue_map(q
->mq_map
, q
->nr_hw_queues
);
1906 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1907 * we should change hctx numa_node according to new topology (this
1908 * involves free and re-allocate memory, worthy doing?)
1911 blk_mq_map_swqueue(q
);
1913 blk_mq_sysfs_register(q
);
1915 blk_mq_unfreeze_queue(q
);
1918 static int blk_mq_queue_reinit_notify(struct notifier_block
*nb
,
1919 unsigned long action
, void *hcpu
)
1921 struct request_queue
*q
;
1924 * Before new mappings are established, hotadded cpu might already
1925 * start handling requests. This doesn't break anything as we map
1926 * offline CPUs to first hardware queue. We will re-init the queue
1927 * below to get optimal settings.
1929 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
&&
1930 action
!= CPU_ONLINE
&& action
!= CPU_ONLINE_FROZEN
)
1933 mutex_lock(&all_q_mutex
);
1934 list_for_each_entry(q
, &all_q_list
, all_q_node
)
1935 blk_mq_queue_reinit(q
);
1936 mutex_unlock(&all_q_mutex
);
1941 * Alloc a tag set to be associated with one or more request queues.
1942 * May fail with EINVAL for various error conditions. May adjust the
1943 * requested depth down, if if it too large. In that case, the set
1944 * value will be stored in set->queue_depth.
1946 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
1950 if (!set
->nr_hw_queues
)
1952 if (!set
->queue_depth
)
1954 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
1957 if (!set
->nr_hw_queues
|| !set
->ops
->queue_rq
|| !set
->ops
->map_queue
)
1960 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
1961 pr_info("blk-mq: reduced tag depth to %u\n",
1963 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
1966 set
->tags
= kmalloc_node(set
->nr_hw_queues
*
1967 sizeof(struct blk_mq_tags
*),
1968 GFP_KERNEL
, set
->numa_node
);
1972 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1973 set
->tags
[i
] = blk_mq_init_rq_map(set
, i
);
1978 mutex_init(&set
->tag_list_lock
);
1979 INIT_LIST_HEAD(&set
->tag_list
);
1985 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1989 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
1991 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
1995 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1997 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2002 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2004 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2006 struct blk_mq_tag_set
*set
= q
->tag_set
;
2007 struct blk_mq_hw_ctx
*hctx
;
2010 if (!set
|| nr
> set
->queue_depth
)
2014 queue_for_each_hw_ctx(q
, hctx
, i
) {
2015 ret
= blk_mq_tag_update_depth(hctx
->tags
, nr
);
2021 q
->nr_requests
= nr
;
2026 void blk_mq_disable_hotplug(void)
2028 mutex_lock(&all_q_mutex
);
2031 void blk_mq_enable_hotplug(void)
2033 mutex_unlock(&all_q_mutex
);
2036 static int __init
blk_mq_init(void)
2040 /* Must be called after percpu_counter_hotcpu_callback() */
2041 hotcpu_notifier(blk_mq_queue_reinit_notify
, -10);
2045 subsys_initcall(blk_mq_init
);