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>
12 #include <linux/kmemleak.h>
14 #include <linux/init.h>
15 #include <linux/slab.h>
16 #include <linux/workqueue.h>
17 #include <linux/smp.h>
18 #include <linux/llist.h>
19 #include <linux/list_sort.h>
20 #include <linux/cpu.h>
21 #include <linux/cache.h>
22 #include <linux/sched/sysctl.h>
23 #include <linux/delay.h>
24 #include <linux/crash_dump.h>
26 #include <trace/events/block.h>
28 #include <linux/blk-mq.h>
31 #include "blk-mq-tag.h"
33 static DEFINE_MUTEX(all_q_mutex);
34 static LIST_HEAD(all_q_list);
36 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
39 * Check if any of the ctx's have pending work in this hardware queue
41 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
45 for (i = 0; i < hctx->ctx_map.size; i++)
46 if (hctx->ctx_map.map[i].word)
52 static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx,
53 struct blk_mq_ctx *ctx)
55 return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word];
58 #define CTX_TO_BIT(hctx, ctx) \
59 ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
62 * Mark this ctx as having pending work in this hardware queue
64 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
65 struct blk_mq_ctx *ctx)
67 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
69 if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word))
70 set_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
73 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
74 struct blk_mq_ctx *ctx)
76 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
78 clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
81 void blk_mq_freeze_queue_start(struct request_queue *q)
85 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
86 if (freeze_depth == 1) {
87 percpu_ref_kill(&q->q_usage_counter);
88 blk_mq_run_hw_queues(q, false);
91 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
93 static void blk_mq_freeze_queue_wait(struct request_queue *q)
95 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
99 * Guarantee no request is in use, so we can change any data structure of
100 * the queue afterward.
102 void blk_freeze_queue(struct request_queue *q)
105 * In the !blk_mq case we are only calling this to kill the
106 * q_usage_counter, otherwise this increases the freeze depth
107 * and waits for it to return to zero. For this reason there is
108 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
109 * exported to drivers as the only user for unfreeze is blk_mq.
111 blk_mq_freeze_queue_start(q);
112 blk_mq_freeze_queue_wait(q);
115 void blk_mq_freeze_queue(struct request_queue *q)
118 * ...just an alias to keep freeze and unfreeze actions balanced
119 * in the blk_mq_* namespace
123 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
125 void blk_mq_unfreeze_queue(struct request_queue *q)
129 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
130 WARN_ON_ONCE(freeze_depth < 0);
132 percpu_ref_reinit(&q->q_usage_counter);
133 wake_up_all(&q->mq_freeze_wq);
136 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
138 void blk_mq_wake_waiters(struct request_queue *q)
140 struct blk_mq_hw_ctx *hctx;
143 queue_for_each_hw_ctx(q, hctx, i)
144 if (blk_mq_hw_queue_mapped(hctx))
145 blk_mq_tag_wakeup_all(hctx->tags, true);
148 * If we are called because the queue has now been marked as
149 * dying, we need to ensure that processes currently waiting on
150 * the queue are notified as well.
152 wake_up_all(&q->mq_freeze_wq);
155 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
157 return blk_mq_has_free_tags(hctx->tags);
159 EXPORT_SYMBOL(blk_mq_can_queue);
161 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
162 struct request *rq, unsigned int rw_flags)
164 if (blk_queue_io_stat(q))
165 rw_flags |= REQ_IO_STAT;
167 INIT_LIST_HEAD(&rq->queuelist);
168 /* csd/requeue_work/fifo_time is initialized before use */
171 rq->cmd_flags |= rw_flags;
172 /* do not touch atomic flags, it needs atomic ops against the timer */
174 INIT_HLIST_NODE(&rq->hash);
175 RB_CLEAR_NODE(&rq->rb_node);
178 rq->start_time = jiffies;
179 #ifdef CONFIG_BLK_CGROUP
181 set_start_time_ns(rq);
182 rq->io_start_time_ns = 0;
184 rq->nr_phys_segments = 0;
185 #if defined(CONFIG_BLK_DEV_INTEGRITY)
186 rq->nr_integrity_segments = 0;
189 /* tag was already set */
199 INIT_LIST_HEAD(&rq->timeout_list);
203 rq->end_io_data = NULL;
206 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
209 static struct request *
210 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
215 tag = blk_mq_get_tag(data);
216 if (tag != BLK_MQ_TAG_FAIL) {
217 rq = data->hctx->tags->rqs[tag];
219 if (blk_mq_tag_busy(data->hctx)) {
220 rq->cmd_flags = REQ_MQ_INFLIGHT;
221 atomic_inc(&data->hctx->nr_active);
225 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
232 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
235 struct blk_mq_ctx *ctx;
236 struct blk_mq_hw_ctx *hctx;
238 struct blk_mq_alloc_data alloc_data;
241 ret = blk_queue_enter(q, gfp);
245 ctx = blk_mq_get_ctx(q);
246 hctx = q->mq_ops->map_queue(q, ctx->cpu);
247 blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_DIRECT_RECLAIM,
248 reserved, ctx, hctx);
250 rq = __blk_mq_alloc_request(&alloc_data, rw);
251 if (!rq && (gfp & __GFP_DIRECT_RECLAIM)) {
252 __blk_mq_run_hw_queue(hctx);
255 ctx = blk_mq_get_ctx(q);
256 hctx = q->mq_ops->map_queue(q, ctx->cpu);
257 blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx,
259 rq = __blk_mq_alloc_request(&alloc_data, rw);
260 ctx = alloc_data.ctx;
265 return ERR_PTR(-EWOULDBLOCK);
269 EXPORT_SYMBOL(blk_mq_alloc_request);
271 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
272 struct blk_mq_ctx *ctx, struct request *rq)
274 const int tag = rq->tag;
275 struct request_queue *q = rq->q;
277 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
278 atomic_dec(&hctx->nr_active);
281 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
282 blk_mq_put_tag(hctx, tag, &ctx->last_tag);
286 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
288 struct blk_mq_ctx *ctx = rq->mq_ctx;
290 ctx->rq_completed[rq_is_sync(rq)]++;
291 __blk_mq_free_request(hctx, ctx, rq);
294 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
296 void blk_mq_free_request(struct request *rq)
298 struct blk_mq_hw_ctx *hctx;
299 struct request_queue *q = rq->q;
301 hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
302 blk_mq_free_hctx_request(hctx, rq);
304 EXPORT_SYMBOL_GPL(blk_mq_free_request);
306 inline void __blk_mq_end_request(struct request *rq, int error)
308 blk_account_io_done(rq);
311 rq->end_io(rq, error);
313 if (unlikely(blk_bidi_rq(rq)))
314 blk_mq_free_request(rq->next_rq);
315 blk_mq_free_request(rq);
318 EXPORT_SYMBOL(__blk_mq_end_request);
320 void blk_mq_end_request(struct request *rq, int error)
322 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
324 __blk_mq_end_request(rq, error);
326 EXPORT_SYMBOL(blk_mq_end_request);
328 static void __blk_mq_complete_request_remote(void *data)
330 struct request *rq = data;
332 rq->q->softirq_done_fn(rq);
335 static void blk_mq_ipi_complete_request(struct request *rq)
337 struct blk_mq_ctx *ctx = rq->mq_ctx;
341 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
342 rq->q->softirq_done_fn(rq);
347 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
348 shared = cpus_share_cache(cpu, ctx->cpu);
350 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
351 rq->csd.func = __blk_mq_complete_request_remote;
354 smp_call_function_single_async(ctx->cpu, &rq->csd);
356 rq->q->softirq_done_fn(rq);
361 static void __blk_mq_complete_request(struct request *rq)
363 struct request_queue *q = rq->q;
365 if (!q->softirq_done_fn)
366 blk_mq_end_request(rq, rq->errors);
368 blk_mq_ipi_complete_request(rq);
372 * blk_mq_complete_request - end I/O on a request
373 * @rq: the request being processed
376 * Ends all I/O on a request. It does not handle partial completions.
377 * The actual completion happens out-of-order, through a IPI handler.
379 void blk_mq_complete_request(struct request *rq, int error)
381 struct request_queue *q = rq->q;
383 if (unlikely(blk_should_fake_timeout(q)))
385 if (!blk_mark_rq_complete(rq)) {
387 __blk_mq_complete_request(rq);
390 EXPORT_SYMBOL(blk_mq_complete_request);
392 int blk_mq_request_started(struct request *rq)
394 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
396 EXPORT_SYMBOL_GPL(blk_mq_request_started);
398 void blk_mq_start_request(struct request *rq)
400 struct request_queue *q = rq->q;
402 trace_block_rq_issue(q, rq);
404 rq->resid_len = blk_rq_bytes(rq);
405 if (unlikely(blk_bidi_rq(rq)))
406 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
411 * Ensure that ->deadline is visible before set the started
412 * flag and clear the completed flag.
414 smp_mb__before_atomic();
417 * Mark us as started and clear complete. Complete might have been
418 * set if requeue raced with timeout, which then marked it as
419 * complete. So be sure to clear complete again when we start
420 * the request, otherwise we'll ignore the completion event.
422 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
423 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
424 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
425 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
427 if (q->dma_drain_size && blk_rq_bytes(rq)) {
429 * Make sure space for the drain appears. We know we can do
430 * this because max_hw_segments has been adjusted to be one
431 * fewer than the device can handle.
433 rq->nr_phys_segments++;
436 EXPORT_SYMBOL(blk_mq_start_request);
438 static void __blk_mq_requeue_request(struct request *rq)
440 struct request_queue *q = rq->q;
442 trace_block_rq_requeue(q, rq);
444 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
445 if (q->dma_drain_size && blk_rq_bytes(rq))
446 rq->nr_phys_segments--;
450 void blk_mq_requeue_request(struct request *rq)
452 __blk_mq_requeue_request(rq);
454 BUG_ON(blk_queued_rq(rq));
455 blk_mq_add_to_requeue_list(rq, true);
457 EXPORT_SYMBOL(blk_mq_requeue_request);
459 static void blk_mq_requeue_work(struct work_struct *work)
461 struct request_queue *q =
462 container_of(work, struct request_queue, requeue_work);
464 struct request *rq, *next;
467 spin_lock_irqsave(&q->requeue_lock, flags);
468 list_splice_init(&q->requeue_list, &rq_list);
469 spin_unlock_irqrestore(&q->requeue_lock, flags);
471 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
472 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
475 rq->cmd_flags &= ~REQ_SOFTBARRIER;
476 list_del_init(&rq->queuelist);
477 blk_mq_insert_request(rq, true, false, false);
480 while (!list_empty(&rq_list)) {
481 rq = list_entry(rq_list.next, struct request, queuelist);
482 list_del_init(&rq->queuelist);
483 blk_mq_insert_request(rq, false, false, false);
487 * Use the start variant of queue running here, so that running
488 * the requeue work will kick stopped queues.
490 blk_mq_start_hw_queues(q);
493 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
495 struct request_queue *q = rq->q;
499 * We abuse this flag that is otherwise used by the I/O scheduler to
500 * request head insertation from the workqueue.
502 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
504 spin_lock_irqsave(&q->requeue_lock, flags);
506 rq->cmd_flags |= REQ_SOFTBARRIER;
507 list_add(&rq->queuelist, &q->requeue_list);
509 list_add_tail(&rq->queuelist, &q->requeue_list);
511 spin_unlock_irqrestore(&q->requeue_lock, flags);
513 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
515 void blk_mq_cancel_requeue_work(struct request_queue *q)
517 cancel_work_sync(&q->requeue_work);
519 EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work);
521 void blk_mq_kick_requeue_list(struct request_queue *q)
523 kblockd_schedule_work(&q->requeue_work);
525 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
527 void blk_mq_abort_requeue_list(struct request_queue *q)
532 spin_lock_irqsave(&q->requeue_lock, flags);
533 list_splice_init(&q->requeue_list, &rq_list);
534 spin_unlock_irqrestore(&q->requeue_lock, flags);
536 while (!list_empty(&rq_list)) {
539 rq = list_first_entry(&rq_list, struct request, queuelist);
540 list_del_init(&rq->queuelist);
542 blk_mq_end_request(rq, rq->errors);
545 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
547 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
549 return tags->rqs[tag];
551 EXPORT_SYMBOL(blk_mq_tag_to_rq);
553 struct blk_mq_timeout_data {
555 unsigned int next_set;
558 void blk_mq_rq_timed_out(struct request *req, bool reserved)
560 struct blk_mq_ops *ops = req->q->mq_ops;
561 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
564 * We know that complete is set at this point. If STARTED isn't set
565 * anymore, then the request isn't active and the "timeout" should
566 * just be ignored. This can happen due to the bitflag ordering.
567 * Timeout first checks if STARTED is set, and if it is, assumes
568 * the request is active. But if we race with completion, then
569 * we both flags will get cleared. So check here again, and ignore
570 * a timeout event with a request that isn't active.
572 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
576 ret = ops->timeout(req, reserved);
580 __blk_mq_complete_request(req);
582 case BLK_EH_RESET_TIMER:
584 blk_clear_rq_complete(req);
586 case BLK_EH_NOT_HANDLED:
589 printk(KERN_ERR "block: bad eh return: %d\n", ret);
594 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
595 struct request *rq, void *priv, bool reserved)
597 struct blk_mq_timeout_data *data = priv;
599 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
601 * If a request wasn't started before the queue was
602 * marked dying, kill it here or it'll go unnoticed.
604 if (unlikely(blk_queue_dying(rq->q)))
605 blk_mq_complete_request(rq, -EIO);
608 if (rq->cmd_flags & REQ_NO_TIMEOUT)
611 if (time_after_eq(jiffies, rq->deadline)) {
612 if (!blk_mark_rq_complete(rq))
613 blk_mq_rq_timed_out(rq, reserved);
614 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
615 data->next = rq->deadline;
620 static void blk_mq_rq_timer(unsigned long priv)
622 struct request_queue *q = (struct request_queue *)priv;
623 struct blk_mq_timeout_data data = {
629 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
632 data.next = blk_rq_timeout(round_jiffies_up(data.next));
633 mod_timer(&q->timeout, data.next);
635 struct blk_mq_hw_ctx *hctx;
637 queue_for_each_hw_ctx(q, hctx, i) {
638 /* the hctx may be unmapped, so check it here */
639 if (blk_mq_hw_queue_mapped(hctx))
640 blk_mq_tag_idle(hctx);
646 * Reverse check our software queue for entries that we could potentially
647 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
648 * too much time checking for merges.
650 static bool blk_mq_attempt_merge(struct request_queue *q,
651 struct blk_mq_ctx *ctx, struct bio *bio)
656 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
662 if (!blk_rq_merge_ok(rq, bio))
665 el_ret = blk_try_merge(rq, bio);
666 if (el_ret == ELEVATOR_BACK_MERGE) {
667 if (bio_attempt_back_merge(q, rq, bio)) {
672 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
673 if (bio_attempt_front_merge(q, rq, bio)) {
685 * Process software queues that have been marked busy, splicing them
686 * to the for-dispatch
688 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
690 struct blk_mq_ctx *ctx;
693 for (i = 0; i < hctx->ctx_map.size; i++) {
694 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
695 unsigned int off, bit;
701 off = i * hctx->ctx_map.bits_per_word;
703 bit = find_next_bit(&bm->word, bm->depth, bit);
704 if (bit >= bm->depth)
707 ctx = hctx->ctxs[bit + off];
708 clear_bit(bit, &bm->word);
709 spin_lock(&ctx->lock);
710 list_splice_tail_init(&ctx->rq_list, list);
711 spin_unlock(&ctx->lock);
719 * Run this hardware queue, pulling any software queues mapped to it in.
720 * Note that this function currently has various problems around ordering
721 * of IO. In particular, we'd like FIFO behaviour on handling existing
722 * items on the hctx->dispatch list. Ignore that for now.
724 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
726 struct request_queue *q = hctx->queue;
729 LIST_HEAD(driver_list);
730 struct list_head *dptr;
733 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
735 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
741 * Touch any software queue that has pending entries.
743 flush_busy_ctxs(hctx, &rq_list);
746 * If we have previous entries on our dispatch list, grab them
747 * and stuff them at the front for more fair dispatch.
749 if (!list_empty_careful(&hctx->dispatch)) {
750 spin_lock(&hctx->lock);
751 if (!list_empty(&hctx->dispatch))
752 list_splice_init(&hctx->dispatch, &rq_list);
753 spin_unlock(&hctx->lock);
757 * Start off with dptr being NULL, so we start the first request
758 * immediately, even if we have more pending.
763 * Now process all the entries, sending them to the driver.
766 while (!list_empty(&rq_list)) {
767 struct blk_mq_queue_data bd;
770 rq = list_first_entry(&rq_list, struct request, queuelist);
771 list_del_init(&rq->queuelist);
775 bd.last = list_empty(&rq_list);
777 ret = q->mq_ops->queue_rq(hctx, &bd);
779 case BLK_MQ_RQ_QUEUE_OK:
782 case BLK_MQ_RQ_QUEUE_BUSY:
783 list_add(&rq->queuelist, &rq_list);
784 __blk_mq_requeue_request(rq);
787 pr_err("blk-mq: bad return on queue: %d\n", ret);
788 case BLK_MQ_RQ_QUEUE_ERROR:
790 blk_mq_end_request(rq, rq->errors);
794 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
798 * We've done the first request. If we have more than 1
799 * left in the list, set dptr to defer issue.
801 if (!dptr && rq_list.next != rq_list.prev)
806 hctx->dispatched[0]++;
807 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
808 hctx->dispatched[ilog2(queued) + 1]++;
811 * Any items that need requeuing? Stuff them into hctx->dispatch,
812 * that is where we will continue on next queue run.
814 if (!list_empty(&rq_list)) {
815 spin_lock(&hctx->lock);
816 list_splice(&rq_list, &hctx->dispatch);
817 spin_unlock(&hctx->lock);
819 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
820 * it's possible the queue is stopped and restarted again
821 * before this. Queue restart will dispatch requests. And since
822 * requests in rq_list aren't added into hctx->dispatch yet,
823 * the requests in rq_list might get lost.
825 * blk_mq_run_hw_queue() already checks the STOPPED bit
827 blk_mq_run_hw_queue(hctx, true);
832 * It'd be great if the workqueue API had a way to pass
833 * in a mask and had some smarts for more clever placement.
834 * For now we just round-robin here, switching for every
835 * BLK_MQ_CPU_WORK_BATCH queued items.
837 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
839 if (hctx->queue->nr_hw_queues == 1)
840 return WORK_CPU_UNBOUND;
842 if (--hctx->next_cpu_batch <= 0) {
843 int cpu = hctx->next_cpu, next_cpu;
845 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
846 if (next_cpu >= nr_cpu_ids)
847 next_cpu = cpumask_first(hctx->cpumask);
849 hctx->next_cpu = next_cpu;
850 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
855 return hctx->next_cpu;
858 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
860 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
861 !blk_mq_hw_queue_mapped(hctx)))
866 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
867 __blk_mq_run_hw_queue(hctx);
875 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
879 void blk_mq_run_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 ((!blk_mq_hctx_has_pending(hctx) &&
886 list_empty_careful(&hctx->dispatch)) ||
887 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
890 blk_mq_run_hw_queue(hctx, async);
893 EXPORT_SYMBOL(blk_mq_run_hw_queues);
895 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
897 cancel_delayed_work(&hctx->run_work);
898 cancel_delayed_work(&hctx->delay_work);
899 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
901 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
903 void blk_mq_stop_hw_queues(struct request_queue *q)
905 struct blk_mq_hw_ctx *hctx;
908 queue_for_each_hw_ctx(q, hctx, i)
909 blk_mq_stop_hw_queue(hctx);
911 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
913 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
915 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
917 blk_mq_run_hw_queue(hctx, false);
919 EXPORT_SYMBOL(blk_mq_start_hw_queue);
921 void blk_mq_start_hw_queues(struct request_queue *q)
923 struct blk_mq_hw_ctx *hctx;
926 queue_for_each_hw_ctx(q, hctx, i)
927 blk_mq_start_hw_queue(hctx);
929 EXPORT_SYMBOL(blk_mq_start_hw_queues);
931 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
933 struct blk_mq_hw_ctx *hctx;
936 queue_for_each_hw_ctx(q, hctx, i) {
937 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
940 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
941 blk_mq_run_hw_queue(hctx, async);
944 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
946 static void blk_mq_run_work_fn(struct work_struct *work)
948 struct blk_mq_hw_ctx *hctx;
950 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
952 __blk_mq_run_hw_queue(hctx);
955 static void blk_mq_delay_work_fn(struct work_struct *work)
957 struct blk_mq_hw_ctx *hctx;
959 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
961 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
962 __blk_mq_run_hw_queue(hctx);
965 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
967 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
970 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
971 &hctx->delay_work, msecs_to_jiffies(msecs));
973 EXPORT_SYMBOL(blk_mq_delay_queue);
975 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
976 struct blk_mq_ctx *ctx,
980 trace_block_rq_insert(hctx->queue, rq);
983 list_add(&rq->queuelist, &ctx->rq_list);
985 list_add_tail(&rq->queuelist, &ctx->rq_list);
988 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
989 struct request *rq, bool at_head)
991 struct blk_mq_ctx *ctx = rq->mq_ctx;
993 __blk_mq_insert_req_list(hctx, ctx, rq, at_head);
994 blk_mq_hctx_mark_pending(hctx, ctx);
997 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1000 struct request_queue *q = rq->q;
1001 struct blk_mq_hw_ctx *hctx;
1002 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
1004 current_ctx = blk_mq_get_ctx(q);
1005 if (!cpu_online(ctx->cpu))
1006 rq->mq_ctx = ctx = current_ctx;
1008 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1010 spin_lock(&ctx->lock);
1011 __blk_mq_insert_request(hctx, rq, at_head);
1012 spin_unlock(&ctx->lock);
1015 blk_mq_run_hw_queue(hctx, async);
1017 blk_mq_put_ctx(current_ctx);
1020 static void blk_mq_insert_requests(struct request_queue *q,
1021 struct blk_mq_ctx *ctx,
1022 struct list_head *list,
1027 struct blk_mq_hw_ctx *hctx;
1028 struct blk_mq_ctx *current_ctx;
1030 trace_block_unplug(q, depth, !from_schedule);
1032 current_ctx = blk_mq_get_ctx(q);
1034 if (!cpu_online(ctx->cpu))
1036 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1039 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1042 spin_lock(&ctx->lock);
1043 while (!list_empty(list)) {
1046 rq = list_first_entry(list, struct request, queuelist);
1047 list_del_init(&rq->queuelist);
1049 __blk_mq_insert_req_list(hctx, ctx, rq, false);
1051 blk_mq_hctx_mark_pending(hctx, ctx);
1052 spin_unlock(&ctx->lock);
1054 blk_mq_run_hw_queue(hctx, from_schedule);
1055 blk_mq_put_ctx(current_ctx);
1058 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1060 struct request *rqa = container_of(a, struct request, queuelist);
1061 struct request *rqb = container_of(b, struct request, queuelist);
1063 return !(rqa->mq_ctx < rqb->mq_ctx ||
1064 (rqa->mq_ctx == rqb->mq_ctx &&
1065 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1068 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1070 struct blk_mq_ctx *this_ctx;
1071 struct request_queue *this_q;
1074 LIST_HEAD(ctx_list);
1077 list_splice_init(&plug->mq_list, &list);
1079 list_sort(NULL, &list, plug_ctx_cmp);
1085 while (!list_empty(&list)) {
1086 rq = list_entry_rq(list.next);
1087 list_del_init(&rq->queuelist);
1089 if (rq->mq_ctx != this_ctx) {
1091 blk_mq_insert_requests(this_q, this_ctx,
1096 this_ctx = rq->mq_ctx;
1102 list_add_tail(&rq->queuelist, &ctx_list);
1106 * If 'this_ctx' is set, we know we have entries to complete
1107 * on 'ctx_list'. Do those.
1110 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1115 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1117 init_request_from_bio(rq, bio);
1119 if (blk_do_io_stat(rq))
1120 blk_account_io_start(rq, 1);
1123 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1125 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1126 !blk_queue_nomerges(hctx->queue);
1129 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1130 struct blk_mq_ctx *ctx,
1131 struct request *rq, struct bio *bio)
1133 if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) {
1134 blk_mq_bio_to_request(rq, bio);
1135 spin_lock(&ctx->lock);
1137 __blk_mq_insert_request(hctx, rq, false);
1138 spin_unlock(&ctx->lock);
1141 struct request_queue *q = hctx->queue;
1143 spin_lock(&ctx->lock);
1144 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1145 blk_mq_bio_to_request(rq, bio);
1149 spin_unlock(&ctx->lock);
1150 __blk_mq_free_request(hctx, ctx, rq);
1155 struct blk_map_ctx {
1156 struct blk_mq_hw_ctx *hctx;
1157 struct blk_mq_ctx *ctx;
1160 static struct request *blk_mq_map_request(struct request_queue *q,
1162 struct blk_map_ctx *data)
1164 struct blk_mq_hw_ctx *hctx;
1165 struct blk_mq_ctx *ctx;
1167 int rw = bio_data_dir(bio);
1168 struct blk_mq_alloc_data alloc_data;
1170 blk_queue_enter_live(q);
1171 ctx = blk_mq_get_ctx(q);
1172 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1174 if (rw_is_sync(bio->bi_rw))
1177 trace_block_getrq(q, bio, rw);
1178 blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1180 rq = __blk_mq_alloc_request(&alloc_data, rw);
1181 if (unlikely(!rq)) {
1182 __blk_mq_run_hw_queue(hctx);
1183 blk_mq_put_ctx(ctx);
1184 trace_block_sleeprq(q, bio, rw);
1186 ctx = blk_mq_get_ctx(q);
1187 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1188 blk_mq_set_alloc_data(&alloc_data, q,
1189 __GFP_RECLAIM|__GFP_HIGH, false, ctx, hctx);
1190 rq = __blk_mq_alloc_request(&alloc_data, rw);
1191 ctx = alloc_data.ctx;
1192 hctx = alloc_data.hctx;
1201 static int blk_mq_direct_issue_request(struct request *rq, blk_qc_t *cookie)
1204 struct request_queue *q = rq->q;
1205 struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q,
1207 struct blk_mq_queue_data bd = {
1212 blk_qc_t new_cookie = blk_tag_to_qc_t(rq->tag, hctx->queue_num);
1215 * For OK queue, we are done. For error, kill it. Any other
1216 * error (busy), just add it to our list as we previously
1219 ret = q->mq_ops->queue_rq(hctx, &bd);
1220 if (ret == BLK_MQ_RQ_QUEUE_OK) {
1221 *cookie = new_cookie;
1225 __blk_mq_requeue_request(rq);
1227 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1228 *cookie = BLK_QC_T_NONE;
1230 blk_mq_end_request(rq, rq->errors);
1238 * Multiple hardware queue variant. This will not use per-process plugs,
1239 * but will attempt to bypass the hctx queueing if we can go straight to
1240 * hardware for SYNC IO.
1242 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1244 const int is_sync = rw_is_sync(bio->bi_rw);
1245 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1246 struct blk_map_ctx data;
1248 unsigned int request_count = 0;
1249 struct blk_plug *plug;
1250 struct request *same_queue_rq = NULL;
1253 blk_queue_bounce(q, &bio);
1255 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1257 return BLK_QC_T_NONE;
1260 blk_queue_split(q, &bio, q->bio_split);
1262 if (!is_flush_fua && !blk_queue_nomerges(q)) {
1263 if (blk_attempt_plug_merge(q, bio, &request_count,
1265 return BLK_QC_T_NONE;
1267 request_count = blk_plug_queued_count(q);
1269 rq = blk_mq_map_request(q, bio, &data);
1271 return BLK_QC_T_NONE;
1273 cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num);
1275 if (unlikely(is_flush_fua)) {
1276 blk_mq_bio_to_request(rq, bio);
1277 blk_insert_flush(rq);
1281 plug = current->plug;
1283 * If the driver supports defer issued based on 'last', then
1284 * queue it up like normal since we can potentially save some
1287 if (((plug && !blk_queue_nomerges(q)) || is_sync) &&
1288 !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1289 struct request *old_rq = NULL;
1291 blk_mq_bio_to_request(rq, bio);
1294 * we do limited pluging. If bio can be merged, do merge.
1295 * Otherwise the existing request in the plug list will be
1296 * issued. So the plug list will have one request at most
1300 * The plug list might get flushed before this. If that
1301 * happens, same_queue_rq is invalid and plug list is empty
1303 if (same_queue_rq && !list_empty(&plug->mq_list)) {
1304 old_rq = same_queue_rq;
1305 list_del_init(&old_rq->queuelist);
1307 list_add_tail(&rq->queuelist, &plug->mq_list);
1308 } else /* is_sync */
1310 blk_mq_put_ctx(data.ctx);
1313 if (!blk_mq_direct_issue_request(old_rq, &cookie))
1315 blk_mq_insert_request(old_rq, false, true, true);
1319 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1321 * For a SYNC request, send it to the hardware immediately. For
1322 * an ASYNC request, just ensure that we run it later on. The
1323 * latter allows for merging opportunities and more efficient
1327 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1329 blk_mq_put_ctx(data.ctx);
1335 * Single hardware queue variant. This will attempt to use any per-process
1336 * plug for merging and IO deferral.
1338 static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio)
1340 const int is_sync = rw_is_sync(bio->bi_rw);
1341 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1342 struct blk_plug *plug;
1343 unsigned int request_count = 0;
1344 struct blk_map_ctx data;
1348 blk_queue_bounce(q, &bio);
1350 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1352 return BLK_QC_T_NONE;
1355 blk_queue_split(q, &bio, q->bio_split);
1357 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1358 blk_attempt_plug_merge(q, bio, &request_count, NULL))
1359 return BLK_QC_T_NONE;
1361 rq = blk_mq_map_request(q, bio, &data);
1363 return BLK_QC_T_NONE;
1365 cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num);
1367 if (unlikely(is_flush_fua)) {
1368 blk_mq_bio_to_request(rq, bio);
1369 blk_insert_flush(rq);
1374 * A task plug currently exists. Since this is completely lockless,
1375 * utilize that to temporarily store requests until the task is
1376 * either done or scheduled away.
1378 plug = current->plug;
1380 blk_mq_bio_to_request(rq, bio);
1382 trace_block_plug(q);
1383 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1384 blk_flush_plug_list(plug, false);
1385 trace_block_plug(q);
1387 list_add_tail(&rq->queuelist, &plug->mq_list);
1388 blk_mq_put_ctx(data.ctx);
1392 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1394 * For a SYNC request, send it to the hardware immediately. For
1395 * an ASYNC request, just ensure that we run it later on. The
1396 * latter allows for merging opportunities and more efficient
1400 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1403 blk_mq_put_ctx(data.ctx);
1408 * Default mapping to a software queue, since we use one per CPU.
1410 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1412 return q->queue_hw_ctx[q->mq_map[cpu]];
1414 EXPORT_SYMBOL(blk_mq_map_queue);
1416 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1417 struct blk_mq_tags *tags, unsigned int hctx_idx)
1421 if (tags->rqs && set->ops->exit_request) {
1424 for (i = 0; i < tags->nr_tags; i++) {
1427 set->ops->exit_request(set->driver_data, tags->rqs[i],
1429 tags->rqs[i] = NULL;
1433 while (!list_empty(&tags->page_list)) {
1434 page = list_first_entry(&tags->page_list, struct page, lru);
1435 list_del_init(&page->lru);
1437 * Remove kmemleak object previously allocated in
1438 * blk_mq_init_rq_map().
1440 kmemleak_free(page_address(page));
1441 __free_pages(page, page->private);
1446 blk_mq_free_tags(tags);
1449 static size_t order_to_size(unsigned int order)
1451 return (size_t)PAGE_SIZE << order;
1454 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1455 unsigned int hctx_idx)
1457 struct blk_mq_tags *tags;
1458 unsigned int i, j, entries_per_page, max_order = 4;
1459 size_t rq_size, left;
1461 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1463 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1467 INIT_LIST_HEAD(&tags->page_list);
1469 tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1470 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1473 blk_mq_free_tags(tags);
1478 * rq_size is the size of the request plus driver payload, rounded
1479 * to the cacheline size
1481 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1483 left = rq_size * set->queue_depth;
1485 for (i = 0; i < set->queue_depth; ) {
1486 int this_order = max_order;
1491 while (left < order_to_size(this_order - 1) && this_order)
1495 page = alloc_pages_node(set->numa_node,
1496 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1502 if (order_to_size(this_order) < rq_size)
1509 page->private = this_order;
1510 list_add_tail(&page->lru, &tags->page_list);
1512 p = page_address(page);
1514 * Allow kmemleak to scan these pages as they contain pointers
1515 * to additional allocations like via ops->init_request().
1517 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_KERNEL);
1518 entries_per_page = order_to_size(this_order) / rq_size;
1519 to_do = min(entries_per_page, set->queue_depth - i);
1520 left -= to_do * rq_size;
1521 for (j = 0; j < to_do; j++) {
1523 if (set->ops->init_request) {
1524 if (set->ops->init_request(set->driver_data,
1525 tags->rqs[i], hctx_idx, i,
1527 tags->rqs[i] = NULL;
1539 blk_mq_free_rq_map(set, tags, hctx_idx);
1543 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1548 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1550 unsigned int bpw = 8, total, num_maps, i;
1552 bitmap->bits_per_word = bpw;
1554 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1555 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1561 for (i = 0; i < num_maps; i++) {
1562 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1563 total -= bitmap->map[i].depth;
1569 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1571 struct request_queue *q = hctx->queue;
1572 struct blk_mq_ctx *ctx;
1576 * Move ctx entries to new CPU, if this one is going away.
1578 ctx = __blk_mq_get_ctx(q, cpu);
1580 spin_lock(&ctx->lock);
1581 if (!list_empty(&ctx->rq_list)) {
1582 list_splice_init(&ctx->rq_list, &tmp);
1583 blk_mq_hctx_clear_pending(hctx, ctx);
1585 spin_unlock(&ctx->lock);
1587 if (list_empty(&tmp))
1590 ctx = blk_mq_get_ctx(q);
1591 spin_lock(&ctx->lock);
1593 while (!list_empty(&tmp)) {
1596 rq = list_first_entry(&tmp, struct request, queuelist);
1598 list_move_tail(&rq->queuelist, &ctx->rq_list);
1601 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1602 blk_mq_hctx_mark_pending(hctx, ctx);
1604 spin_unlock(&ctx->lock);
1606 blk_mq_run_hw_queue(hctx, true);
1607 blk_mq_put_ctx(ctx);
1611 static int blk_mq_hctx_notify(void *data, unsigned long action,
1614 struct blk_mq_hw_ctx *hctx = data;
1616 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1617 return blk_mq_hctx_cpu_offline(hctx, cpu);
1620 * In case of CPU online, tags may be reallocated
1621 * in blk_mq_map_swqueue() after mapping is updated.
1627 /* hctx->ctxs will be freed in queue's release handler */
1628 static void blk_mq_exit_hctx(struct request_queue *q,
1629 struct blk_mq_tag_set *set,
1630 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1632 unsigned flush_start_tag = set->queue_depth;
1634 blk_mq_tag_idle(hctx);
1636 if (set->ops->exit_request)
1637 set->ops->exit_request(set->driver_data,
1638 hctx->fq->flush_rq, hctx_idx,
1639 flush_start_tag + hctx_idx);
1641 if (set->ops->exit_hctx)
1642 set->ops->exit_hctx(hctx, hctx_idx);
1644 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1645 blk_free_flush_queue(hctx->fq);
1646 blk_mq_free_bitmap(&hctx->ctx_map);
1649 static void blk_mq_exit_hw_queues(struct request_queue *q,
1650 struct blk_mq_tag_set *set, int nr_queue)
1652 struct blk_mq_hw_ctx *hctx;
1655 queue_for_each_hw_ctx(q, hctx, i) {
1658 blk_mq_exit_hctx(q, set, hctx, i);
1662 static void blk_mq_free_hw_queues(struct request_queue *q,
1663 struct blk_mq_tag_set *set)
1665 struct blk_mq_hw_ctx *hctx;
1668 queue_for_each_hw_ctx(q, hctx, i)
1669 free_cpumask_var(hctx->cpumask);
1672 static int blk_mq_init_hctx(struct request_queue *q,
1673 struct blk_mq_tag_set *set,
1674 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1677 unsigned flush_start_tag = set->queue_depth;
1679 node = hctx->numa_node;
1680 if (node == NUMA_NO_NODE)
1681 node = hctx->numa_node = set->numa_node;
1683 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1684 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1685 spin_lock_init(&hctx->lock);
1686 INIT_LIST_HEAD(&hctx->dispatch);
1688 hctx->queue_num = hctx_idx;
1689 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1691 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1692 blk_mq_hctx_notify, hctx);
1693 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1695 hctx->tags = set->tags[hctx_idx];
1698 * Allocate space for all possible cpus to avoid allocation at
1701 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1704 goto unregister_cpu_notifier;
1706 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1711 if (set->ops->init_hctx &&
1712 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1715 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1719 if (set->ops->init_request &&
1720 set->ops->init_request(set->driver_data,
1721 hctx->fq->flush_rq, hctx_idx,
1722 flush_start_tag + hctx_idx, node))
1730 if (set->ops->exit_hctx)
1731 set->ops->exit_hctx(hctx, hctx_idx);
1733 blk_mq_free_bitmap(&hctx->ctx_map);
1736 unregister_cpu_notifier:
1737 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1742 static int blk_mq_init_hw_queues(struct request_queue *q,
1743 struct blk_mq_tag_set *set)
1745 struct blk_mq_hw_ctx *hctx;
1749 * Initialize hardware queues
1751 queue_for_each_hw_ctx(q, hctx, i) {
1752 if (blk_mq_init_hctx(q, set, hctx, i))
1756 if (i == q->nr_hw_queues)
1762 blk_mq_exit_hw_queues(q, set, i);
1767 static void blk_mq_init_cpu_queues(struct request_queue *q,
1768 unsigned int nr_hw_queues)
1772 for_each_possible_cpu(i) {
1773 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1774 struct blk_mq_hw_ctx *hctx;
1776 memset(__ctx, 0, sizeof(*__ctx));
1778 spin_lock_init(&__ctx->lock);
1779 INIT_LIST_HEAD(&__ctx->rq_list);
1782 /* If the cpu isn't online, the cpu is mapped to first hctx */
1786 hctx = q->mq_ops->map_queue(q, i);
1789 * Set local node, IFF we have more than one hw queue. If
1790 * not, we remain on the home node of the device
1792 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1793 hctx->numa_node = cpu_to_node(i);
1797 static void blk_mq_map_swqueue(struct request_queue *q,
1798 const struct cpumask *online_mask)
1801 struct blk_mq_hw_ctx *hctx;
1802 struct blk_mq_ctx *ctx;
1803 struct blk_mq_tag_set *set = q->tag_set;
1806 * Avoid others reading imcomplete hctx->cpumask through sysfs
1808 mutex_lock(&q->sysfs_lock);
1810 queue_for_each_hw_ctx(q, hctx, i) {
1811 cpumask_clear(hctx->cpumask);
1816 * Map software to hardware queues
1818 queue_for_each_ctx(q, ctx, i) {
1819 /* If the cpu isn't online, the cpu is mapped to first hctx */
1820 if (!cpumask_test_cpu(i, online_mask))
1823 hctx = q->mq_ops->map_queue(q, i);
1824 cpumask_set_cpu(i, hctx->cpumask);
1825 ctx->index_hw = hctx->nr_ctx;
1826 hctx->ctxs[hctx->nr_ctx++] = ctx;
1829 mutex_unlock(&q->sysfs_lock);
1831 queue_for_each_hw_ctx(q, hctx, i) {
1832 struct blk_mq_ctxmap *map = &hctx->ctx_map;
1835 * If no software queues are mapped to this hardware queue,
1836 * disable it and free the request entries.
1838 if (!hctx->nr_ctx) {
1840 blk_mq_free_rq_map(set, set->tags[i], i);
1841 set->tags[i] = NULL;
1847 /* unmapped hw queue can be remapped after CPU topo changed */
1849 set->tags[i] = blk_mq_init_rq_map(set, i);
1850 hctx->tags = set->tags[i];
1851 WARN_ON(!hctx->tags);
1854 * Set the map size to the number of mapped software queues.
1855 * This is more accurate and more efficient than looping
1856 * over all possibly mapped software queues.
1858 map->size = DIV_ROUND_UP(hctx->nr_ctx, map->bits_per_word);
1861 * Initialize batch roundrobin counts
1863 hctx->next_cpu = cpumask_first(hctx->cpumask);
1864 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1867 queue_for_each_ctx(q, ctx, i) {
1868 if (!cpumask_test_cpu(i, online_mask))
1871 hctx = q->mq_ops->map_queue(q, i);
1872 cpumask_set_cpu(i, hctx->tags->cpumask);
1876 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
1878 struct blk_mq_hw_ctx *hctx;
1881 queue_for_each_hw_ctx(q, hctx, i) {
1883 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1885 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1889 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
1891 struct request_queue *q;
1893 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1894 blk_mq_freeze_queue(q);
1895 queue_set_hctx_shared(q, shared);
1896 blk_mq_unfreeze_queue(q);
1900 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1902 struct blk_mq_tag_set *set = q->tag_set;
1904 mutex_lock(&set->tag_list_lock);
1905 list_del_init(&q->tag_set_list);
1906 if (list_is_singular(&set->tag_list)) {
1907 /* just transitioned to unshared */
1908 set->flags &= ~BLK_MQ_F_TAG_SHARED;
1909 /* update existing queue */
1910 blk_mq_update_tag_set_depth(set, false);
1912 mutex_unlock(&set->tag_list_lock);
1915 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1916 struct request_queue *q)
1920 mutex_lock(&set->tag_list_lock);
1922 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
1923 if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
1924 set->flags |= BLK_MQ_F_TAG_SHARED;
1925 /* update existing queue */
1926 blk_mq_update_tag_set_depth(set, true);
1928 if (set->flags & BLK_MQ_F_TAG_SHARED)
1929 queue_set_hctx_shared(q, true);
1930 list_add_tail(&q->tag_set_list, &set->tag_list);
1932 mutex_unlock(&set->tag_list_lock);
1936 * It is the actual release handler for mq, but we do it from
1937 * request queue's release handler for avoiding use-after-free
1938 * and headache because q->mq_kobj shouldn't have been introduced,
1939 * but we can't group ctx/kctx kobj without it.
1941 void blk_mq_release(struct request_queue *q)
1943 struct blk_mq_hw_ctx *hctx;
1946 /* hctx kobj stays in hctx */
1947 queue_for_each_hw_ctx(q, hctx, i) {
1957 kfree(q->queue_hw_ctx);
1959 /* ctx kobj stays in queue_ctx */
1960 free_percpu(q->queue_ctx);
1963 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1965 struct request_queue *uninit_q, *q;
1967 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1969 return ERR_PTR(-ENOMEM);
1971 q = blk_mq_init_allocated_queue(set, uninit_q);
1973 blk_cleanup_queue(uninit_q);
1977 EXPORT_SYMBOL(blk_mq_init_queue);
1979 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
1980 struct request_queue *q)
1982 struct blk_mq_hw_ctx **hctxs;
1983 struct blk_mq_ctx __percpu *ctx;
1987 ctx = alloc_percpu(struct blk_mq_ctx);
1989 return ERR_PTR(-ENOMEM);
1991 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1997 map = blk_mq_make_queue_map(set);
2001 for (i = 0; i < set->nr_hw_queues; i++) {
2002 int node = blk_mq_hw_queue_to_node(map, i);
2004 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
2009 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2013 atomic_set(&hctxs[i]->nr_active, 0);
2014 hctxs[i]->numa_node = node;
2015 hctxs[i]->queue_num = i;
2018 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
2019 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2021 q->nr_queues = nr_cpu_ids;
2022 q->nr_hw_queues = set->nr_hw_queues;
2026 q->queue_hw_ctx = hctxs;
2028 q->mq_ops = set->ops;
2029 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2031 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2032 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2034 q->sg_reserved_size = INT_MAX;
2036 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
2037 INIT_LIST_HEAD(&q->requeue_list);
2038 spin_lock_init(&q->requeue_lock);
2040 if (q->nr_hw_queues > 1)
2041 blk_queue_make_request(q, blk_mq_make_request);
2043 blk_queue_make_request(q, blk_sq_make_request);
2046 * Do this after blk_queue_make_request() overrides it...
2048 q->nr_requests = set->queue_depth;
2050 if (set->ops->complete)
2051 blk_queue_softirq_done(q, set->ops->complete);
2053 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2055 if (blk_mq_init_hw_queues(q, set))
2059 mutex_lock(&all_q_mutex);
2061 list_add_tail(&q->all_q_node, &all_q_list);
2062 blk_mq_add_queue_tag_set(set, q);
2063 blk_mq_map_swqueue(q, cpu_online_mask);
2065 mutex_unlock(&all_q_mutex);
2072 for (i = 0; i < set->nr_hw_queues; i++) {
2075 free_cpumask_var(hctxs[i]->cpumask);
2082 return ERR_PTR(-ENOMEM);
2084 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2086 void blk_mq_free_queue(struct request_queue *q)
2088 struct blk_mq_tag_set *set = q->tag_set;
2090 mutex_lock(&all_q_mutex);
2091 list_del_init(&q->all_q_node);
2092 mutex_unlock(&all_q_mutex);
2094 blk_mq_del_queue_tag_set(q);
2096 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2097 blk_mq_free_hw_queues(q, set);
2100 /* Basically redo blk_mq_init_queue with queue frozen */
2101 static void blk_mq_queue_reinit(struct request_queue *q,
2102 const struct cpumask *online_mask)
2104 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2106 blk_mq_sysfs_unregister(q);
2108 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues, online_mask);
2111 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2112 * we should change hctx numa_node according to new topology (this
2113 * involves free and re-allocate memory, worthy doing?)
2116 blk_mq_map_swqueue(q, online_mask);
2118 blk_mq_sysfs_register(q);
2121 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2122 unsigned long action, void *hcpu)
2124 struct request_queue *q;
2125 int cpu = (unsigned long)hcpu;
2127 * New online cpumask which is going to be set in this hotplug event.
2128 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2129 * one-by-one and dynamically allocating this could result in a failure.
2131 static struct cpumask online_new;
2134 * Before hotadded cpu starts handling requests, new mappings must
2135 * be established. Otherwise, these requests in hw queue might
2136 * never be dispatched.
2138 * For example, there is a single hw queue (hctx) and two CPU queues
2139 * (ctx0 for CPU0, and ctx1 for CPU1).
2141 * Now CPU1 is just onlined and a request is inserted into
2142 * ctx1->rq_list and set bit0 in pending bitmap as ctx1->index_hw is
2145 * And then while running hw queue, flush_busy_ctxs() finds bit0 is
2146 * set in pending bitmap and tries to retrieve requests in
2147 * hctx->ctxs[0]->rq_list. But htx->ctxs[0] is a pointer to ctx0,
2148 * so the request in ctx1->rq_list is ignored.
2150 switch (action & ~CPU_TASKS_FROZEN) {
2152 case CPU_UP_CANCELED:
2153 cpumask_copy(&online_new, cpu_online_mask);
2155 case CPU_UP_PREPARE:
2156 cpumask_copy(&online_new, cpu_online_mask);
2157 cpumask_set_cpu(cpu, &online_new);
2163 mutex_lock(&all_q_mutex);
2166 * We need to freeze and reinit all existing queues. Freezing
2167 * involves synchronous wait for an RCU grace period and doing it
2168 * one by one may take a long time. Start freezing all queues in
2169 * one swoop and then wait for the completions so that freezing can
2170 * take place in parallel.
2172 list_for_each_entry(q, &all_q_list, all_q_node)
2173 blk_mq_freeze_queue_start(q);
2174 list_for_each_entry(q, &all_q_list, all_q_node) {
2175 blk_mq_freeze_queue_wait(q);
2178 * timeout handler can't touch hw queue during the
2181 del_timer_sync(&q->timeout);
2184 list_for_each_entry(q, &all_q_list, all_q_node)
2185 blk_mq_queue_reinit(q, &online_new);
2187 list_for_each_entry(q, &all_q_list, all_q_node)
2188 blk_mq_unfreeze_queue(q);
2190 mutex_unlock(&all_q_mutex);
2194 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2198 for (i = 0; i < set->nr_hw_queues; i++) {
2199 set->tags[i] = blk_mq_init_rq_map(set, i);
2208 blk_mq_free_rq_map(set, set->tags[i], i);
2214 * Allocate the request maps associated with this tag_set. Note that this
2215 * may reduce the depth asked for, if memory is tight. set->queue_depth
2216 * will be updated to reflect the allocated depth.
2218 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2223 depth = set->queue_depth;
2225 err = __blk_mq_alloc_rq_maps(set);
2229 set->queue_depth >>= 1;
2230 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2234 } while (set->queue_depth);
2236 if (!set->queue_depth || err) {
2237 pr_err("blk-mq: failed to allocate request map\n");
2241 if (depth != set->queue_depth)
2242 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2243 depth, set->queue_depth);
2248 struct cpumask *blk_mq_tags_cpumask(struct blk_mq_tags *tags)
2250 return tags->cpumask;
2252 EXPORT_SYMBOL_GPL(blk_mq_tags_cpumask);
2255 * Alloc a tag set to be associated with one or more request queues.
2256 * May fail with EINVAL for various error conditions. May adjust the
2257 * requested depth down, if if it too large. In that case, the set
2258 * value will be stored in set->queue_depth.
2260 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2262 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2264 if (!set->nr_hw_queues)
2266 if (!set->queue_depth)
2268 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2271 if (!set->ops->queue_rq || !set->ops->map_queue)
2274 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2275 pr_info("blk-mq: reduced tag depth to %u\n",
2277 set->queue_depth = BLK_MQ_MAX_DEPTH;
2281 * If a crashdump is active, then we are potentially in a very
2282 * memory constrained environment. Limit us to 1 queue and
2283 * 64 tags to prevent using too much memory.
2285 if (is_kdump_kernel()) {
2286 set->nr_hw_queues = 1;
2287 set->queue_depth = min(64U, set->queue_depth);
2290 set->tags = kmalloc_node(set->nr_hw_queues *
2291 sizeof(struct blk_mq_tags *),
2292 GFP_KERNEL, set->numa_node);
2296 if (blk_mq_alloc_rq_maps(set))
2299 mutex_init(&set->tag_list_lock);
2300 INIT_LIST_HEAD(&set->tag_list);
2308 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2310 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2314 for (i = 0; i < set->nr_hw_queues; i++) {
2316 blk_mq_free_rq_map(set, set->tags[i], i);
2322 EXPORT_SYMBOL(blk_mq_free_tag_set);
2324 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2326 struct blk_mq_tag_set *set = q->tag_set;
2327 struct blk_mq_hw_ctx *hctx;
2330 if (!set || nr > set->queue_depth)
2334 queue_for_each_hw_ctx(q, hctx, i) {
2335 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2341 q->nr_requests = nr;
2346 void blk_mq_disable_hotplug(void)
2348 mutex_lock(&all_q_mutex);
2351 void blk_mq_enable_hotplug(void)
2353 mutex_unlock(&all_q_mutex);
2356 static int __init blk_mq_init(void)
2360 hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2364 subsys_initcall(blk_mq_init);