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 static int blk_mq_queue_enter(struct request_queue *q, gfp_t gfp)
86 if (percpu_ref_tryget_live(&q->mq_usage_counter))
89 if (!(gfp & __GFP_WAIT))
92 ret = wait_event_interruptible(q->mq_freeze_wq,
93 !atomic_read(&q->mq_freeze_depth) ||
95 if (blk_queue_dying(q))
102 static void blk_mq_queue_exit(struct request_queue *q)
104 percpu_ref_put(&q->mq_usage_counter);
107 static void blk_mq_usage_counter_release(struct percpu_ref *ref)
109 struct request_queue *q =
110 container_of(ref, struct request_queue, mq_usage_counter);
112 wake_up_all(&q->mq_freeze_wq);
115 void blk_mq_freeze_queue_start(struct request_queue *q)
119 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
120 if (freeze_depth == 1) {
121 percpu_ref_kill(&q->mq_usage_counter);
122 blk_mq_run_hw_queues(q, false);
125 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
127 static void blk_mq_freeze_queue_wait(struct request_queue *q)
129 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->mq_usage_counter));
133 * Guarantee no request is in use, so we can change any data structure of
134 * the queue afterward.
136 void blk_mq_freeze_queue(struct request_queue *q)
138 blk_mq_freeze_queue_start(q);
139 blk_mq_freeze_queue_wait(q);
141 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
143 void blk_mq_unfreeze_queue(struct request_queue *q)
147 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
148 WARN_ON_ONCE(freeze_depth < 0);
150 percpu_ref_reinit(&q->mq_usage_counter);
151 wake_up_all(&q->mq_freeze_wq);
154 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
156 void blk_mq_wake_waiters(struct request_queue *q)
158 struct blk_mq_hw_ctx *hctx;
161 queue_for_each_hw_ctx(q, hctx, i)
162 if (blk_mq_hw_queue_mapped(hctx))
163 blk_mq_tag_wakeup_all(hctx->tags, true);
166 * If we are called because the queue has now been marked as
167 * dying, we need to ensure that processes currently waiting on
168 * the queue are notified as well.
170 wake_up_all(&q->mq_freeze_wq);
173 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
175 return blk_mq_has_free_tags(hctx->tags);
177 EXPORT_SYMBOL(blk_mq_can_queue);
179 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
180 struct request *rq, unsigned int rw_flags)
182 if (blk_queue_io_stat(q))
183 rw_flags |= REQ_IO_STAT;
185 INIT_LIST_HEAD(&rq->queuelist);
186 /* csd/requeue_work/fifo_time is initialized before use */
189 rq->cmd_flags |= rw_flags;
190 /* do not touch atomic flags, it needs atomic ops against the timer */
192 INIT_HLIST_NODE(&rq->hash);
193 RB_CLEAR_NODE(&rq->rb_node);
196 rq->start_time = jiffies;
197 #ifdef CONFIG_BLK_CGROUP
199 set_start_time_ns(rq);
200 rq->io_start_time_ns = 0;
202 rq->nr_phys_segments = 0;
203 #if defined(CONFIG_BLK_DEV_INTEGRITY)
204 rq->nr_integrity_segments = 0;
207 /* tag was already set */
217 INIT_LIST_HEAD(&rq->timeout_list);
221 rq->end_io_data = NULL;
224 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
227 static struct request *
228 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
233 tag = blk_mq_get_tag(data);
234 if (tag != BLK_MQ_TAG_FAIL) {
235 rq = data->hctx->tags->rqs[tag];
237 if (blk_mq_tag_busy(data->hctx)) {
238 rq->cmd_flags = REQ_MQ_INFLIGHT;
239 atomic_inc(&data->hctx->nr_active);
243 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
250 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
253 struct blk_mq_ctx *ctx;
254 struct blk_mq_hw_ctx *hctx;
256 struct blk_mq_alloc_data alloc_data;
259 ret = blk_mq_queue_enter(q, gfp);
263 ctx = blk_mq_get_ctx(q);
264 hctx = q->mq_ops->map_queue(q, ctx->cpu);
265 blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT,
266 reserved, ctx, hctx);
268 rq = __blk_mq_alloc_request(&alloc_data, rw);
269 if (!rq && (gfp & __GFP_WAIT)) {
270 __blk_mq_run_hw_queue(hctx);
273 ctx = blk_mq_get_ctx(q);
274 hctx = q->mq_ops->map_queue(q, ctx->cpu);
275 blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx,
277 rq = __blk_mq_alloc_request(&alloc_data, rw);
278 ctx = alloc_data.ctx;
282 blk_mq_queue_exit(q);
283 return ERR_PTR(-EWOULDBLOCK);
287 EXPORT_SYMBOL(blk_mq_alloc_request);
289 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
290 struct blk_mq_ctx *ctx, struct request *rq)
292 const int tag = rq->tag;
293 struct request_queue *q = rq->q;
295 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
296 atomic_dec(&hctx->nr_active);
299 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
300 blk_mq_put_tag(hctx, tag, &ctx->last_tag);
301 blk_mq_queue_exit(q);
304 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
306 struct blk_mq_ctx *ctx = rq->mq_ctx;
308 ctx->rq_completed[rq_is_sync(rq)]++;
309 __blk_mq_free_request(hctx, ctx, rq);
312 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
314 void blk_mq_free_request(struct request *rq)
316 struct blk_mq_hw_ctx *hctx;
317 struct request_queue *q = rq->q;
319 hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
320 blk_mq_free_hctx_request(hctx, rq);
322 EXPORT_SYMBOL_GPL(blk_mq_free_request);
324 inline void __blk_mq_end_request(struct request *rq, int error)
326 blk_account_io_done(rq);
329 rq->end_io(rq, error);
331 if (unlikely(blk_bidi_rq(rq)))
332 blk_mq_free_request(rq->next_rq);
333 blk_mq_free_request(rq);
336 EXPORT_SYMBOL(__blk_mq_end_request);
338 void blk_mq_end_request(struct request *rq, int error)
340 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
342 __blk_mq_end_request(rq, error);
344 EXPORT_SYMBOL(blk_mq_end_request);
346 static void __blk_mq_complete_request_remote(void *data)
348 struct request *rq = data;
350 rq->q->softirq_done_fn(rq);
353 static void blk_mq_ipi_complete_request(struct request *rq)
355 struct blk_mq_ctx *ctx = rq->mq_ctx;
359 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
360 rq->q->softirq_done_fn(rq);
365 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
366 shared = cpus_share_cache(cpu, ctx->cpu);
368 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
369 rq->csd.func = __blk_mq_complete_request_remote;
372 smp_call_function_single_async(ctx->cpu, &rq->csd);
374 rq->q->softirq_done_fn(rq);
379 void __blk_mq_complete_request(struct request *rq)
381 struct request_queue *q = rq->q;
383 if (!q->softirq_done_fn)
384 blk_mq_end_request(rq, rq->errors);
386 blk_mq_ipi_complete_request(rq);
390 * blk_mq_complete_request - end I/O on a request
391 * @rq: the request being processed
394 * Ends all I/O on a request. It does not handle partial completions.
395 * The actual completion happens out-of-order, through a IPI handler.
397 void blk_mq_complete_request(struct request *rq, int error)
399 struct request_queue *q = rq->q;
401 if (unlikely(blk_should_fake_timeout(q)))
403 if (!blk_mark_rq_complete(rq)) {
405 __blk_mq_complete_request(rq);
408 EXPORT_SYMBOL(blk_mq_complete_request);
410 int blk_mq_request_started(struct request *rq)
412 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
414 EXPORT_SYMBOL_GPL(blk_mq_request_started);
416 void blk_mq_start_request(struct request *rq)
418 struct request_queue *q = rq->q;
420 trace_block_rq_issue(q, rq);
422 rq->resid_len = blk_rq_bytes(rq);
423 if (unlikely(blk_bidi_rq(rq)))
424 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
429 * Ensure that ->deadline is visible before set the started
430 * flag and clear the completed flag.
432 smp_mb__before_atomic();
435 * Mark us as started and clear complete. Complete might have been
436 * set if requeue raced with timeout, which then marked it as
437 * complete. So be sure to clear complete again when we start
438 * the request, otherwise we'll ignore the completion event.
440 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
441 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
442 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
443 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
445 if (q->dma_drain_size && blk_rq_bytes(rq)) {
447 * Make sure space for the drain appears. We know we can do
448 * this because max_hw_segments has been adjusted to be one
449 * fewer than the device can handle.
451 rq->nr_phys_segments++;
454 EXPORT_SYMBOL(blk_mq_start_request);
456 static void __blk_mq_requeue_request(struct request *rq)
458 struct request_queue *q = rq->q;
460 trace_block_rq_requeue(q, rq);
462 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
463 if (q->dma_drain_size && blk_rq_bytes(rq))
464 rq->nr_phys_segments--;
468 void blk_mq_requeue_request(struct request *rq)
470 __blk_mq_requeue_request(rq);
472 BUG_ON(blk_queued_rq(rq));
473 blk_mq_add_to_requeue_list(rq, true);
475 EXPORT_SYMBOL(blk_mq_requeue_request);
477 static void blk_mq_requeue_work(struct work_struct *work)
479 struct request_queue *q =
480 container_of(work, struct request_queue, requeue_work);
482 struct request *rq, *next;
485 spin_lock_irqsave(&q->requeue_lock, flags);
486 list_splice_init(&q->requeue_list, &rq_list);
487 spin_unlock_irqrestore(&q->requeue_lock, flags);
489 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
490 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
493 rq->cmd_flags &= ~REQ_SOFTBARRIER;
494 list_del_init(&rq->queuelist);
495 blk_mq_insert_request(rq, true, false, false);
498 while (!list_empty(&rq_list)) {
499 rq = list_entry(rq_list.next, struct request, queuelist);
500 list_del_init(&rq->queuelist);
501 blk_mq_insert_request(rq, false, false, false);
505 * Use the start variant of queue running here, so that running
506 * the requeue work will kick stopped queues.
508 blk_mq_start_hw_queues(q);
511 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
513 struct request_queue *q = rq->q;
517 * We abuse this flag that is otherwise used by the I/O scheduler to
518 * request head insertation from the workqueue.
520 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
522 spin_lock_irqsave(&q->requeue_lock, flags);
524 rq->cmd_flags |= REQ_SOFTBARRIER;
525 list_add(&rq->queuelist, &q->requeue_list);
527 list_add_tail(&rq->queuelist, &q->requeue_list);
529 spin_unlock_irqrestore(&q->requeue_lock, flags);
531 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
533 void blk_mq_cancel_requeue_work(struct request_queue *q)
535 cancel_work_sync(&q->requeue_work);
537 EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work);
539 void blk_mq_kick_requeue_list(struct request_queue *q)
541 kblockd_schedule_work(&q->requeue_work);
543 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
545 void blk_mq_abort_requeue_list(struct request_queue *q)
550 spin_lock_irqsave(&q->requeue_lock, flags);
551 list_splice_init(&q->requeue_list, &rq_list);
552 spin_unlock_irqrestore(&q->requeue_lock, flags);
554 while (!list_empty(&rq_list)) {
557 rq = list_first_entry(&rq_list, struct request, queuelist);
558 list_del_init(&rq->queuelist);
560 blk_mq_end_request(rq, rq->errors);
563 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
565 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
567 return tags->rqs[tag];
569 EXPORT_SYMBOL(blk_mq_tag_to_rq);
571 struct blk_mq_timeout_data {
573 unsigned int next_set;
576 void blk_mq_rq_timed_out(struct request *req, bool reserved)
578 struct blk_mq_ops *ops = req->q->mq_ops;
579 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
582 * We know that complete is set at this point. If STARTED isn't set
583 * anymore, then the request isn't active and the "timeout" should
584 * just be ignored. This can happen due to the bitflag ordering.
585 * Timeout first checks if STARTED is set, and if it is, assumes
586 * the request is active. But if we race with completion, then
587 * we both flags will get cleared. So check here again, and ignore
588 * a timeout event with a request that isn't active.
590 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
594 ret = ops->timeout(req, reserved);
598 __blk_mq_complete_request(req);
600 case BLK_EH_RESET_TIMER:
602 blk_clear_rq_complete(req);
604 case BLK_EH_NOT_HANDLED:
607 printk(KERN_ERR "block: bad eh return: %d\n", ret);
612 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
613 struct request *rq, void *priv, bool reserved)
615 struct blk_mq_timeout_data *data = priv;
617 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
619 * If a request wasn't started before the queue was
620 * marked dying, kill it here or it'll go unnoticed.
622 if (unlikely(blk_queue_dying(rq->q)))
623 blk_mq_complete_request(rq, -EIO);
626 if (rq->cmd_flags & REQ_NO_TIMEOUT)
629 if (time_after_eq(jiffies, rq->deadline)) {
630 if (!blk_mark_rq_complete(rq))
631 blk_mq_rq_timed_out(rq, reserved);
632 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
633 data->next = rq->deadline;
638 static void blk_mq_rq_timer(unsigned long priv)
640 struct request_queue *q = (struct request_queue *)priv;
641 struct blk_mq_timeout_data data = {
647 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
650 data.next = blk_rq_timeout(round_jiffies_up(data.next));
651 mod_timer(&q->timeout, data.next);
653 struct blk_mq_hw_ctx *hctx;
655 queue_for_each_hw_ctx(q, hctx, i) {
656 /* the hctx may be unmapped, so check it here */
657 if (blk_mq_hw_queue_mapped(hctx))
658 blk_mq_tag_idle(hctx);
664 * Reverse check our software queue for entries that we could potentially
665 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
666 * too much time checking for merges.
668 static bool blk_mq_attempt_merge(struct request_queue *q,
669 struct blk_mq_ctx *ctx, struct bio *bio)
674 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
680 if (!blk_rq_merge_ok(rq, bio))
683 el_ret = blk_try_merge(rq, bio);
684 if (el_ret == ELEVATOR_BACK_MERGE) {
685 if (bio_attempt_back_merge(q, rq, bio)) {
690 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
691 if (bio_attempt_front_merge(q, rq, bio)) {
703 * Process software queues that have been marked busy, splicing them
704 * to the for-dispatch
706 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
708 struct blk_mq_ctx *ctx;
711 for (i = 0; i < hctx->ctx_map.size; i++) {
712 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
713 unsigned int off, bit;
719 off = i * hctx->ctx_map.bits_per_word;
721 bit = find_next_bit(&bm->word, bm->depth, bit);
722 if (bit >= bm->depth)
725 ctx = hctx->ctxs[bit + off];
726 clear_bit(bit, &bm->word);
727 spin_lock(&ctx->lock);
728 list_splice_tail_init(&ctx->rq_list, list);
729 spin_unlock(&ctx->lock);
737 * Run this hardware queue, pulling any software queues mapped to it in.
738 * Note that this function currently has various problems around ordering
739 * of IO. In particular, we'd like FIFO behaviour on handling existing
740 * items on the hctx->dispatch list. Ignore that for now.
742 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
744 struct request_queue *q = hctx->queue;
747 LIST_HEAD(driver_list);
748 struct list_head *dptr;
751 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
753 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
759 * Touch any software queue that has pending entries.
761 flush_busy_ctxs(hctx, &rq_list);
764 * If we have previous entries on our dispatch list, grab them
765 * and stuff them at the front for more fair dispatch.
767 if (!list_empty_careful(&hctx->dispatch)) {
768 spin_lock(&hctx->lock);
769 if (!list_empty(&hctx->dispatch))
770 list_splice_init(&hctx->dispatch, &rq_list);
771 spin_unlock(&hctx->lock);
775 * Start off with dptr being NULL, so we start the first request
776 * immediately, even if we have more pending.
781 * Now process all the entries, sending them to the driver.
784 while (!list_empty(&rq_list)) {
785 struct blk_mq_queue_data bd;
788 rq = list_first_entry(&rq_list, struct request, queuelist);
789 list_del_init(&rq->queuelist);
793 bd.last = list_empty(&rq_list);
795 ret = q->mq_ops->queue_rq(hctx, &bd);
797 case BLK_MQ_RQ_QUEUE_OK:
800 case BLK_MQ_RQ_QUEUE_BUSY:
801 list_add(&rq->queuelist, &rq_list);
802 __blk_mq_requeue_request(rq);
805 pr_err("blk-mq: bad return on queue: %d\n", ret);
806 case BLK_MQ_RQ_QUEUE_ERROR:
808 blk_mq_end_request(rq, rq->errors);
812 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
816 * We've done the first request. If we have more than 1
817 * left in the list, set dptr to defer issue.
819 if (!dptr && rq_list.next != rq_list.prev)
824 hctx->dispatched[0]++;
825 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
826 hctx->dispatched[ilog2(queued) + 1]++;
829 * Any items that need requeuing? Stuff them into hctx->dispatch,
830 * that is where we will continue on next queue run.
832 if (!list_empty(&rq_list)) {
833 spin_lock(&hctx->lock);
834 list_splice(&rq_list, &hctx->dispatch);
835 spin_unlock(&hctx->lock);
837 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
838 * it's possible the queue is stopped and restarted again
839 * before this. Queue restart will dispatch requests. And since
840 * requests in rq_list aren't added into hctx->dispatch yet,
841 * the requests in rq_list might get lost.
843 * blk_mq_run_hw_queue() already checks the STOPPED bit
845 blk_mq_run_hw_queue(hctx, true);
850 * It'd be great if the workqueue API had a way to pass
851 * in a mask and had some smarts for more clever placement.
852 * For now we just round-robin here, switching for every
853 * BLK_MQ_CPU_WORK_BATCH queued items.
855 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
857 if (hctx->queue->nr_hw_queues == 1)
858 return WORK_CPU_UNBOUND;
860 if (--hctx->next_cpu_batch <= 0) {
861 int cpu = hctx->next_cpu, next_cpu;
863 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
864 if (next_cpu >= nr_cpu_ids)
865 next_cpu = cpumask_first(hctx->cpumask);
867 hctx->next_cpu = next_cpu;
868 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
873 return hctx->next_cpu;
876 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
878 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
879 !blk_mq_hw_queue_mapped(hctx)))
884 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
885 __blk_mq_run_hw_queue(hctx);
893 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
897 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
899 struct blk_mq_hw_ctx *hctx;
902 queue_for_each_hw_ctx(q, hctx, i) {
903 if ((!blk_mq_hctx_has_pending(hctx) &&
904 list_empty_careful(&hctx->dispatch)) ||
905 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
908 blk_mq_run_hw_queue(hctx, async);
911 EXPORT_SYMBOL(blk_mq_run_hw_queues);
913 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
915 cancel_delayed_work(&hctx->run_work);
916 cancel_delayed_work(&hctx->delay_work);
917 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
919 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
921 void blk_mq_stop_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_stop_hw_queue(hctx);
929 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
931 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
933 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
935 blk_mq_run_hw_queue(hctx, false);
937 EXPORT_SYMBOL(blk_mq_start_hw_queue);
939 void blk_mq_start_hw_queues(struct request_queue *q)
941 struct blk_mq_hw_ctx *hctx;
944 queue_for_each_hw_ctx(q, hctx, i)
945 blk_mq_start_hw_queue(hctx);
947 EXPORT_SYMBOL(blk_mq_start_hw_queues);
949 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
951 struct blk_mq_hw_ctx *hctx;
954 queue_for_each_hw_ctx(q, hctx, i) {
955 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
958 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
959 blk_mq_run_hw_queue(hctx, async);
962 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
964 static void blk_mq_run_work_fn(struct work_struct *work)
966 struct blk_mq_hw_ctx *hctx;
968 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
970 __blk_mq_run_hw_queue(hctx);
973 static void blk_mq_delay_work_fn(struct work_struct *work)
975 struct blk_mq_hw_ctx *hctx;
977 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
979 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
980 __blk_mq_run_hw_queue(hctx);
983 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
985 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
988 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
989 &hctx->delay_work, msecs_to_jiffies(msecs));
991 EXPORT_SYMBOL(blk_mq_delay_queue);
993 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
994 struct request *rq, bool at_head)
996 struct blk_mq_ctx *ctx = rq->mq_ctx;
998 trace_block_rq_insert(hctx->queue, rq);
1001 list_add(&rq->queuelist, &ctx->rq_list);
1003 list_add_tail(&rq->queuelist, &ctx->rq_list);
1005 blk_mq_hctx_mark_pending(hctx, ctx);
1008 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1011 struct request_queue *q = rq->q;
1012 struct blk_mq_hw_ctx *hctx;
1013 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
1015 current_ctx = blk_mq_get_ctx(q);
1016 if (!cpu_online(ctx->cpu))
1017 rq->mq_ctx = ctx = current_ctx;
1019 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1021 spin_lock(&ctx->lock);
1022 __blk_mq_insert_request(hctx, rq, at_head);
1023 spin_unlock(&ctx->lock);
1026 blk_mq_run_hw_queue(hctx, async);
1028 blk_mq_put_ctx(current_ctx);
1031 static void blk_mq_insert_requests(struct request_queue *q,
1032 struct blk_mq_ctx *ctx,
1033 struct list_head *list,
1038 struct blk_mq_hw_ctx *hctx;
1039 struct blk_mq_ctx *current_ctx;
1041 trace_block_unplug(q, depth, !from_schedule);
1043 current_ctx = blk_mq_get_ctx(q);
1045 if (!cpu_online(ctx->cpu))
1047 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1050 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1053 spin_lock(&ctx->lock);
1054 while (!list_empty(list)) {
1057 rq = list_first_entry(list, struct request, queuelist);
1058 list_del_init(&rq->queuelist);
1060 __blk_mq_insert_request(hctx, rq, false);
1062 spin_unlock(&ctx->lock);
1064 blk_mq_run_hw_queue(hctx, from_schedule);
1065 blk_mq_put_ctx(current_ctx);
1068 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1070 struct request *rqa = container_of(a, struct request, queuelist);
1071 struct request *rqb = container_of(b, struct request, queuelist);
1073 return !(rqa->mq_ctx < rqb->mq_ctx ||
1074 (rqa->mq_ctx == rqb->mq_ctx &&
1075 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1078 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1080 struct blk_mq_ctx *this_ctx;
1081 struct request_queue *this_q;
1084 LIST_HEAD(ctx_list);
1087 list_splice_init(&plug->mq_list, &list);
1089 list_sort(NULL, &list, plug_ctx_cmp);
1095 while (!list_empty(&list)) {
1096 rq = list_entry_rq(list.next);
1097 list_del_init(&rq->queuelist);
1099 if (rq->mq_ctx != this_ctx) {
1101 blk_mq_insert_requests(this_q, this_ctx,
1106 this_ctx = rq->mq_ctx;
1112 list_add_tail(&rq->queuelist, &ctx_list);
1116 * If 'this_ctx' is set, we know we have entries to complete
1117 * on 'ctx_list'. Do those.
1120 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1125 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1127 init_request_from_bio(rq, bio);
1129 if (blk_do_io_stat(rq))
1130 blk_account_io_start(rq, 1);
1133 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1135 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1136 !blk_queue_nomerges(hctx->queue);
1139 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1140 struct blk_mq_ctx *ctx,
1141 struct request *rq, struct bio *bio)
1143 if (!hctx_allow_merges(hctx)) {
1144 blk_mq_bio_to_request(rq, bio);
1145 spin_lock(&ctx->lock);
1147 __blk_mq_insert_request(hctx, rq, false);
1148 spin_unlock(&ctx->lock);
1151 struct request_queue *q = hctx->queue;
1153 spin_lock(&ctx->lock);
1154 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1155 blk_mq_bio_to_request(rq, bio);
1159 spin_unlock(&ctx->lock);
1160 __blk_mq_free_request(hctx, ctx, rq);
1165 struct blk_map_ctx {
1166 struct blk_mq_hw_ctx *hctx;
1167 struct blk_mq_ctx *ctx;
1170 static struct request *blk_mq_map_request(struct request_queue *q,
1172 struct blk_map_ctx *data)
1174 struct blk_mq_hw_ctx *hctx;
1175 struct blk_mq_ctx *ctx;
1177 int rw = bio_data_dir(bio);
1178 struct blk_mq_alloc_data alloc_data;
1180 if (unlikely(blk_mq_queue_enter(q, GFP_KERNEL))) {
1185 ctx = blk_mq_get_ctx(q);
1186 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1188 if (rw_is_sync(bio->bi_rw))
1191 trace_block_getrq(q, bio, rw);
1192 blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1194 rq = __blk_mq_alloc_request(&alloc_data, rw);
1195 if (unlikely(!rq)) {
1196 __blk_mq_run_hw_queue(hctx);
1197 blk_mq_put_ctx(ctx);
1198 trace_block_sleeprq(q, bio, rw);
1200 ctx = blk_mq_get_ctx(q);
1201 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1202 blk_mq_set_alloc_data(&alloc_data, q,
1203 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1204 rq = __blk_mq_alloc_request(&alloc_data, rw);
1205 ctx = alloc_data.ctx;
1206 hctx = alloc_data.hctx;
1215 static int blk_mq_direct_issue_request(struct request *rq)
1218 struct request_queue *q = rq->q;
1219 struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q,
1221 struct blk_mq_queue_data bd = {
1228 * For OK queue, we are done. For error, kill it. Any other
1229 * error (busy), just add it to our list as we previously
1232 ret = q->mq_ops->queue_rq(hctx, &bd);
1233 if (ret == BLK_MQ_RQ_QUEUE_OK)
1236 __blk_mq_requeue_request(rq);
1238 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1240 blk_mq_end_request(rq, rq->errors);
1248 * Multiple hardware queue variant. This will not use per-process plugs,
1249 * but will attempt to bypass the hctx queueing if we can go straight to
1250 * hardware for SYNC IO.
1252 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1254 const int is_sync = rw_is_sync(bio->bi_rw);
1255 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1256 struct blk_map_ctx data;
1258 unsigned int request_count = 0;
1259 struct blk_plug *plug;
1260 struct request *same_queue_rq = NULL;
1262 blk_queue_bounce(q, &bio);
1264 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1269 blk_queue_split(q, &bio, q->bio_split);
1271 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1272 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1275 rq = blk_mq_map_request(q, bio, &data);
1279 if (unlikely(is_flush_fua)) {
1280 blk_mq_bio_to_request(rq, bio);
1281 blk_insert_flush(rq);
1285 plug = current->plug;
1287 * If the driver supports defer issued based on 'last', then
1288 * queue it up like normal since we can potentially save some
1291 if (((plug && !blk_queue_nomerges(q)) || is_sync) &&
1292 !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1293 struct request *old_rq = NULL;
1295 blk_mq_bio_to_request(rq, bio);
1298 * we do limited pluging. If bio can be merged, do merge.
1299 * Otherwise the existing request in the plug list will be
1300 * issued. So the plug list will have one request at most
1304 * The plug list might get flushed before this. If that
1305 * happens, same_queue_rq is invalid and plug list is empty
1307 if (same_queue_rq && !list_empty(&plug->mq_list)) {
1308 old_rq = same_queue_rq;
1309 list_del_init(&old_rq->queuelist);
1311 list_add_tail(&rq->queuelist, &plug->mq_list);
1312 } else /* is_sync */
1314 blk_mq_put_ctx(data.ctx);
1317 if (!blk_mq_direct_issue_request(old_rq))
1319 blk_mq_insert_request(old_rq, false, true, true);
1323 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1325 * For a SYNC request, send it to the hardware immediately. For
1326 * an ASYNC request, just ensure that we run it later on. The
1327 * latter allows for merging opportunities and more efficient
1331 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1333 blk_mq_put_ctx(data.ctx);
1337 * Single hardware queue variant. This will attempt to use any per-process
1338 * plug for merging and IO deferral.
1340 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1342 const int is_sync = rw_is_sync(bio->bi_rw);
1343 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1344 struct blk_plug *plug;
1345 unsigned int request_count = 0;
1346 struct blk_map_ctx data;
1349 blk_queue_bounce(q, &bio);
1351 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1356 blk_queue_split(q, &bio, q->bio_split);
1358 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1359 blk_attempt_plug_merge(q, bio, &request_count, NULL))
1362 rq = blk_mq_map_request(q, bio, &data);
1366 if (unlikely(is_flush_fua)) {
1367 blk_mq_bio_to_request(rq, bio);
1368 blk_insert_flush(rq);
1373 * A task plug currently exists. Since this is completely lockless,
1374 * utilize that to temporarily store requests until the task is
1375 * either done or scheduled away.
1377 plug = current->plug;
1379 blk_mq_bio_to_request(rq, bio);
1380 if (list_empty(&plug->mq_list))
1381 trace_block_plug(q);
1382 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1383 blk_flush_plug_list(plug, false);
1384 trace_block_plug(q);
1386 list_add_tail(&rq->queuelist, &plug->mq_list);
1387 blk_mq_put_ctx(data.ctx);
1391 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1393 * For a SYNC request, send it to the hardware immediately. For
1394 * an ASYNC request, just ensure that we run it later on. The
1395 * latter allows for merging opportunities and more efficient
1399 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1402 blk_mq_put_ctx(data.ctx);
1406 * Default mapping to a software queue, since we use one per CPU.
1408 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1410 return q->queue_hw_ctx[q->mq_map[cpu]];
1412 EXPORT_SYMBOL(blk_mq_map_queue);
1414 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1415 struct blk_mq_tags *tags, unsigned int hctx_idx)
1419 if (tags->rqs && set->ops->exit_request) {
1422 for (i = 0; i < tags->nr_tags; i++) {
1425 set->ops->exit_request(set->driver_data, tags->rqs[i],
1427 tags->rqs[i] = NULL;
1431 while (!list_empty(&tags->page_list)) {
1432 page = list_first_entry(&tags->page_list, struct page, lru);
1433 list_del_init(&page->lru);
1435 * Remove kmemleak object previously allocated in
1436 * blk_mq_init_rq_map().
1438 kmemleak_free(page_address(page));
1439 __free_pages(page, page->private);
1444 blk_mq_free_tags(tags);
1447 static size_t order_to_size(unsigned int order)
1449 return (size_t)PAGE_SIZE << order;
1452 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1453 unsigned int hctx_idx)
1455 struct blk_mq_tags *tags;
1456 unsigned int i, j, entries_per_page, max_order = 4;
1457 size_t rq_size, left;
1459 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1461 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1465 INIT_LIST_HEAD(&tags->page_list);
1467 tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1468 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1471 blk_mq_free_tags(tags);
1476 * rq_size is the size of the request plus driver payload, rounded
1477 * to the cacheline size
1479 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1481 left = rq_size * set->queue_depth;
1483 for (i = 0; i < set->queue_depth; ) {
1484 int this_order = max_order;
1489 while (left < order_to_size(this_order - 1) && this_order)
1493 page = alloc_pages_node(set->numa_node,
1494 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1500 if (order_to_size(this_order) < rq_size)
1507 page->private = this_order;
1508 list_add_tail(&page->lru, &tags->page_list);
1510 p = page_address(page);
1512 * Allow kmemleak to scan these pages as they contain pointers
1513 * to additional allocations like via ops->init_request().
1515 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_KERNEL);
1516 entries_per_page = order_to_size(this_order) / rq_size;
1517 to_do = min(entries_per_page, set->queue_depth - i);
1518 left -= to_do * rq_size;
1519 for (j = 0; j < to_do; j++) {
1521 if (set->ops->init_request) {
1522 if (set->ops->init_request(set->driver_data,
1523 tags->rqs[i], hctx_idx, i,
1525 tags->rqs[i] = NULL;
1537 blk_mq_free_rq_map(set, tags, hctx_idx);
1541 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1546 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1548 unsigned int bpw = 8, total, num_maps, i;
1550 bitmap->bits_per_word = bpw;
1552 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1553 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1559 for (i = 0; i < num_maps; i++) {
1560 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1561 total -= bitmap->map[i].depth;
1567 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1569 struct request_queue *q = hctx->queue;
1570 struct blk_mq_ctx *ctx;
1574 * Move ctx entries to new CPU, if this one is going away.
1576 ctx = __blk_mq_get_ctx(q, cpu);
1578 spin_lock(&ctx->lock);
1579 if (!list_empty(&ctx->rq_list)) {
1580 list_splice_init(&ctx->rq_list, &tmp);
1581 blk_mq_hctx_clear_pending(hctx, ctx);
1583 spin_unlock(&ctx->lock);
1585 if (list_empty(&tmp))
1588 ctx = blk_mq_get_ctx(q);
1589 spin_lock(&ctx->lock);
1591 while (!list_empty(&tmp)) {
1594 rq = list_first_entry(&tmp, struct request, queuelist);
1596 list_move_tail(&rq->queuelist, &ctx->rq_list);
1599 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1600 blk_mq_hctx_mark_pending(hctx, ctx);
1602 spin_unlock(&ctx->lock);
1604 blk_mq_run_hw_queue(hctx, true);
1605 blk_mq_put_ctx(ctx);
1609 static int blk_mq_hctx_notify(void *data, unsigned long action,
1612 struct blk_mq_hw_ctx *hctx = data;
1614 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1615 return blk_mq_hctx_cpu_offline(hctx, cpu);
1618 * In case of CPU online, tags may be reallocated
1619 * in blk_mq_map_swqueue() after mapping is updated.
1625 /* hctx->ctxs will be freed in queue's release handler */
1626 static void blk_mq_exit_hctx(struct request_queue *q,
1627 struct blk_mq_tag_set *set,
1628 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1630 unsigned flush_start_tag = set->queue_depth;
1632 blk_mq_tag_idle(hctx);
1634 if (set->ops->exit_request)
1635 set->ops->exit_request(set->driver_data,
1636 hctx->fq->flush_rq, hctx_idx,
1637 flush_start_tag + hctx_idx);
1639 if (set->ops->exit_hctx)
1640 set->ops->exit_hctx(hctx, hctx_idx);
1642 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1643 blk_free_flush_queue(hctx->fq);
1644 blk_mq_free_bitmap(&hctx->ctx_map);
1647 static void blk_mq_exit_hw_queues(struct request_queue *q,
1648 struct blk_mq_tag_set *set, int nr_queue)
1650 struct blk_mq_hw_ctx *hctx;
1653 queue_for_each_hw_ctx(q, hctx, i) {
1656 blk_mq_exit_hctx(q, set, hctx, i);
1660 static void blk_mq_free_hw_queues(struct request_queue *q,
1661 struct blk_mq_tag_set *set)
1663 struct blk_mq_hw_ctx *hctx;
1666 queue_for_each_hw_ctx(q, hctx, i)
1667 free_cpumask_var(hctx->cpumask);
1670 static int blk_mq_init_hctx(struct request_queue *q,
1671 struct blk_mq_tag_set *set,
1672 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1675 unsigned flush_start_tag = set->queue_depth;
1677 node = hctx->numa_node;
1678 if (node == NUMA_NO_NODE)
1679 node = hctx->numa_node = set->numa_node;
1681 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1682 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1683 spin_lock_init(&hctx->lock);
1684 INIT_LIST_HEAD(&hctx->dispatch);
1686 hctx->queue_num = hctx_idx;
1687 hctx->flags = set->flags;
1689 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1690 blk_mq_hctx_notify, hctx);
1691 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1693 hctx->tags = set->tags[hctx_idx];
1696 * Allocate space for all possible cpus to avoid allocation at
1699 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1702 goto unregister_cpu_notifier;
1704 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1709 if (set->ops->init_hctx &&
1710 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1713 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1717 if (set->ops->init_request &&
1718 set->ops->init_request(set->driver_data,
1719 hctx->fq->flush_rq, hctx_idx,
1720 flush_start_tag + hctx_idx, node))
1728 if (set->ops->exit_hctx)
1729 set->ops->exit_hctx(hctx, hctx_idx);
1731 blk_mq_free_bitmap(&hctx->ctx_map);
1734 unregister_cpu_notifier:
1735 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1740 static int blk_mq_init_hw_queues(struct request_queue *q,
1741 struct blk_mq_tag_set *set)
1743 struct blk_mq_hw_ctx *hctx;
1747 * Initialize hardware queues
1749 queue_for_each_hw_ctx(q, hctx, i) {
1750 if (blk_mq_init_hctx(q, set, hctx, i))
1754 if (i == q->nr_hw_queues)
1760 blk_mq_exit_hw_queues(q, set, i);
1765 static void blk_mq_init_cpu_queues(struct request_queue *q,
1766 unsigned int nr_hw_queues)
1770 for_each_possible_cpu(i) {
1771 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1772 struct blk_mq_hw_ctx *hctx;
1774 memset(__ctx, 0, sizeof(*__ctx));
1776 spin_lock_init(&__ctx->lock);
1777 INIT_LIST_HEAD(&__ctx->rq_list);
1780 /* If the cpu isn't online, the cpu is mapped to first hctx */
1784 hctx = q->mq_ops->map_queue(q, i);
1787 * Set local node, IFF we have more than one hw queue. If
1788 * not, we remain on the home node of the device
1790 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1791 hctx->numa_node = cpu_to_node(i);
1795 static void blk_mq_map_swqueue(struct request_queue *q,
1796 const struct cpumask *online_mask)
1799 struct blk_mq_hw_ctx *hctx;
1800 struct blk_mq_ctx *ctx;
1801 struct blk_mq_tag_set *set = q->tag_set;
1804 * Avoid others reading imcomplete hctx->cpumask through sysfs
1806 mutex_lock(&q->sysfs_lock);
1808 queue_for_each_hw_ctx(q, hctx, i) {
1809 cpumask_clear(hctx->cpumask);
1814 * Map software to hardware queues
1816 queue_for_each_ctx(q, ctx, i) {
1817 /* If the cpu isn't online, the cpu is mapped to first hctx */
1818 if (!cpumask_test_cpu(i, online_mask))
1821 hctx = q->mq_ops->map_queue(q, i);
1822 cpumask_set_cpu(i, hctx->cpumask);
1823 ctx->index_hw = hctx->nr_ctx;
1824 hctx->ctxs[hctx->nr_ctx++] = ctx;
1827 mutex_unlock(&q->sysfs_lock);
1829 queue_for_each_hw_ctx(q, hctx, i) {
1830 struct blk_mq_ctxmap *map = &hctx->ctx_map;
1833 * If no software queues are mapped to this hardware queue,
1834 * disable it and free the request entries.
1836 if (!hctx->nr_ctx) {
1838 blk_mq_free_rq_map(set, set->tags[i], i);
1839 set->tags[i] = NULL;
1845 /* unmapped hw queue can be remapped after CPU topo changed */
1847 set->tags[i] = blk_mq_init_rq_map(set, i);
1848 hctx->tags = set->tags[i];
1849 WARN_ON(!hctx->tags);
1852 * Set the map size to the number of mapped software queues.
1853 * This is more accurate and more efficient than looping
1854 * over all possibly mapped software queues.
1856 map->size = DIV_ROUND_UP(hctx->nr_ctx, map->bits_per_word);
1859 * Initialize batch roundrobin counts
1861 hctx->next_cpu = cpumask_first(hctx->cpumask);
1862 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1865 queue_for_each_ctx(q, ctx, i) {
1866 if (!cpumask_test_cpu(i, online_mask))
1869 hctx = q->mq_ops->map_queue(q, i);
1870 cpumask_set_cpu(i, hctx->tags->cpumask);
1874 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1876 struct blk_mq_hw_ctx *hctx;
1877 struct request_queue *q;
1881 if (set->tag_list.next == set->tag_list.prev)
1886 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1887 blk_mq_freeze_queue(q);
1889 queue_for_each_hw_ctx(q, hctx, i) {
1891 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1893 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1895 blk_mq_unfreeze_queue(q);
1899 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1901 struct blk_mq_tag_set *set = q->tag_set;
1903 mutex_lock(&set->tag_list_lock);
1904 list_del_init(&q->tag_set_list);
1905 blk_mq_update_tag_set_depth(set);
1906 mutex_unlock(&set->tag_list_lock);
1909 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1910 struct request_queue *q)
1914 mutex_lock(&set->tag_list_lock);
1915 list_add_tail(&q->tag_set_list, &set->tag_list);
1916 blk_mq_update_tag_set_depth(set);
1917 mutex_unlock(&set->tag_list_lock);
1921 * It is the actual release handler for mq, but we do it from
1922 * request queue's release handler for avoiding use-after-free
1923 * and headache because q->mq_kobj shouldn't have been introduced,
1924 * but we can't group ctx/kctx kobj without it.
1926 void blk_mq_release(struct request_queue *q)
1928 struct blk_mq_hw_ctx *hctx;
1931 /* hctx kobj stays in hctx */
1932 queue_for_each_hw_ctx(q, hctx, i) {
1942 kfree(q->queue_hw_ctx);
1944 /* ctx kobj stays in queue_ctx */
1945 free_percpu(q->queue_ctx);
1948 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1950 struct request_queue *uninit_q, *q;
1952 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1954 return ERR_PTR(-ENOMEM);
1956 q = blk_mq_init_allocated_queue(set, uninit_q);
1958 blk_cleanup_queue(uninit_q);
1962 EXPORT_SYMBOL(blk_mq_init_queue);
1964 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
1965 struct request_queue *q)
1967 struct blk_mq_hw_ctx **hctxs;
1968 struct blk_mq_ctx __percpu *ctx;
1972 ctx = alloc_percpu(struct blk_mq_ctx);
1974 return ERR_PTR(-ENOMEM);
1976 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1982 map = blk_mq_make_queue_map(set);
1986 for (i = 0; i < set->nr_hw_queues; i++) {
1987 int node = blk_mq_hw_queue_to_node(map, i);
1989 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1994 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
1998 atomic_set(&hctxs[i]->nr_active, 0);
1999 hctxs[i]->numa_node = node;
2000 hctxs[i]->queue_num = i;
2004 * Init percpu_ref in atomic mode so that it's faster to shutdown.
2005 * See blk_register_queue() for details.
2007 if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release,
2008 PERCPU_REF_INIT_ATOMIC, GFP_KERNEL))
2011 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
2012 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2014 q->nr_queues = nr_cpu_ids;
2015 q->nr_hw_queues = set->nr_hw_queues;
2019 q->queue_hw_ctx = hctxs;
2021 q->mq_ops = set->ops;
2022 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2024 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2025 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2027 q->sg_reserved_size = INT_MAX;
2029 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
2030 INIT_LIST_HEAD(&q->requeue_list);
2031 spin_lock_init(&q->requeue_lock);
2033 if (q->nr_hw_queues > 1)
2034 blk_queue_make_request(q, blk_mq_make_request);
2036 blk_queue_make_request(q, blk_sq_make_request);
2039 * Do this after blk_queue_make_request() overrides it...
2041 q->nr_requests = set->queue_depth;
2043 if (set->ops->complete)
2044 blk_queue_softirq_done(q, set->ops->complete);
2046 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2048 if (blk_mq_init_hw_queues(q, set))
2052 mutex_lock(&all_q_mutex);
2054 list_add_tail(&q->all_q_node, &all_q_list);
2055 blk_mq_add_queue_tag_set(set, q);
2056 blk_mq_map_swqueue(q, cpu_online_mask);
2058 mutex_unlock(&all_q_mutex);
2065 for (i = 0; i < set->nr_hw_queues; i++) {
2068 free_cpumask_var(hctxs[i]->cpumask);
2075 return ERR_PTR(-ENOMEM);
2077 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2079 void blk_mq_free_queue(struct request_queue *q)
2081 struct blk_mq_tag_set *set = q->tag_set;
2083 mutex_lock(&all_q_mutex);
2084 list_del_init(&q->all_q_node);
2085 mutex_unlock(&all_q_mutex);
2087 blk_mq_del_queue_tag_set(q);
2089 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2090 blk_mq_free_hw_queues(q, set);
2092 percpu_ref_exit(&q->mq_usage_counter);
2095 /* Basically redo blk_mq_init_queue with queue frozen */
2096 static void blk_mq_queue_reinit(struct request_queue *q,
2097 const struct cpumask *online_mask)
2099 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2101 blk_mq_sysfs_unregister(q);
2103 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues, online_mask);
2106 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2107 * we should change hctx numa_node according to new topology (this
2108 * involves free and re-allocate memory, worthy doing?)
2111 blk_mq_map_swqueue(q, online_mask);
2113 blk_mq_sysfs_register(q);
2116 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2117 unsigned long action, void *hcpu)
2119 struct request_queue *q;
2120 int cpu = (unsigned long)hcpu;
2122 * New online cpumask which is going to be set in this hotplug event.
2123 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2124 * one-by-one and dynamically allocating this could result in a failure.
2126 static struct cpumask online_new;
2129 * Before hotadded cpu starts handling requests, new mappings must
2130 * be established. Otherwise, these requests in hw queue might
2131 * never be dispatched.
2133 * For example, there is a single hw queue (hctx) and two CPU queues
2134 * (ctx0 for CPU0, and ctx1 for CPU1).
2136 * Now CPU1 is just onlined and a request is inserted into
2137 * ctx1->rq_list and set bit0 in pending bitmap as ctx1->index_hw is
2140 * And then while running hw queue, flush_busy_ctxs() finds bit0 is
2141 * set in pending bitmap and tries to retrieve requests in
2142 * hctx->ctxs[0]->rq_list. But htx->ctxs[0] is a pointer to ctx0,
2143 * so the request in ctx1->rq_list is ignored.
2145 switch (action & ~CPU_TASKS_FROZEN) {
2147 case CPU_UP_CANCELED:
2148 cpumask_copy(&online_new, cpu_online_mask);
2150 case CPU_UP_PREPARE:
2151 cpumask_copy(&online_new, cpu_online_mask);
2152 cpumask_set_cpu(cpu, &online_new);
2158 mutex_lock(&all_q_mutex);
2161 * We need to freeze and reinit all existing queues. Freezing
2162 * involves synchronous wait for an RCU grace period and doing it
2163 * one by one may take a long time. Start freezing all queues in
2164 * one swoop and then wait for the completions so that freezing can
2165 * take place in parallel.
2167 list_for_each_entry(q, &all_q_list, all_q_node)
2168 blk_mq_freeze_queue_start(q);
2169 list_for_each_entry(q, &all_q_list, all_q_node) {
2170 blk_mq_freeze_queue_wait(q);
2173 * timeout handler can't touch hw queue during the
2176 del_timer_sync(&q->timeout);
2179 list_for_each_entry(q, &all_q_list, all_q_node)
2180 blk_mq_queue_reinit(q, &online_new);
2182 list_for_each_entry(q, &all_q_list, all_q_node)
2183 blk_mq_unfreeze_queue(q);
2185 mutex_unlock(&all_q_mutex);
2189 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2193 for (i = 0; i < set->nr_hw_queues; i++) {
2194 set->tags[i] = blk_mq_init_rq_map(set, i);
2203 blk_mq_free_rq_map(set, set->tags[i], i);
2209 * Allocate the request maps associated with this tag_set. Note that this
2210 * may reduce the depth asked for, if memory is tight. set->queue_depth
2211 * will be updated to reflect the allocated depth.
2213 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2218 depth = set->queue_depth;
2220 err = __blk_mq_alloc_rq_maps(set);
2224 set->queue_depth >>= 1;
2225 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2229 } while (set->queue_depth);
2231 if (!set->queue_depth || err) {
2232 pr_err("blk-mq: failed to allocate request map\n");
2236 if (depth != set->queue_depth)
2237 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2238 depth, set->queue_depth);
2243 struct cpumask *blk_mq_tags_cpumask(struct blk_mq_tags *tags)
2245 return tags->cpumask;
2247 EXPORT_SYMBOL_GPL(blk_mq_tags_cpumask);
2250 * Alloc a tag set to be associated with one or more request queues.
2251 * May fail with EINVAL for various error conditions. May adjust the
2252 * requested depth down, if if it too large. In that case, the set
2253 * value will be stored in set->queue_depth.
2255 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2257 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2259 if (!set->nr_hw_queues)
2261 if (!set->queue_depth)
2263 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2266 if (!set->ops->queue_rq || !set->ops->map_queue)
2269 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2270 pr_info("blk-mq: reduced tag depth to %u\n",
2272 set->queue_depth = BLK_MQ_MAX_DEPTH;
2276 * If a crashdump is active, then we are potentially in a very
2277 * memory constrained environment. Limit us to 1 queue and
2278 * 64 tags to prevent using too much memory.
2280 if (is_kdump_kernel()) {
2281 set->nr_hw_queues = 1;
2282 set->queue_depth = min(64U, set->queue_depth);
2285 set->tags = kmalloc_node(set->nr_hw_queues *
2286 sizeof(struct blk_mq_tags *),
2287 GFP_KERNEL, set->numa_node);
2291 if (blk_mq_alloc_rq_maps(set))
2294 mutex_init(&set->tag_list_lock);
2295 INIT_LIST_HEAD(&set->tag_list);
2303 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2305 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2309 for (i = 0; i < set->nr_hw_queues; i++) {
2311 blk_mq_free_rq_map(set, set->tags[i], i);
2312 free_cpumask_var(set->tags[i]->cpumask);
2319 EXPORT_SYMBOL(blk_mq_free_tag_set);
2321 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2323 struct blk_mq_tag_set *set = q->tag_set;
2324 struct blk_mq_hw_ctx *hctx;
2327 if (!set || nr > set->queue_depth)
2331 queue_for_each_hw_ctx(q, hctx, i) {
2332 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2338 q->nr_requests = nr;
2343 void blk_mq_disable_hotplug(void)
2345 mutex_lock(&all_q_mutex);
2348 void blk_mq_enable_hotplug(void)
2350 mutex_unlock(&all_q_mutex);
2353 static int __init blk_mq_init(void)
2357 hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2361 subsys_initcall(blk_mq_init);