2 * Block multiqueue core code
4 * Copyright (C) 2013-2014 Jens Axboe
5 * Copyright (C) 2013-2014 Christoph Hellwig
7 #include <linux/kernel.h>
8 #include <linux/module.h>
9 #include <linux/backing-dev.h>
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
13 #include <linux/init.h>
14 #include <linux/slab.h>
15 #include <linux/workqueue.h>
16 #include <linux/smp.h>
17 #include <linux/llist.h>
18 #include <linux/list_sort.h>
19 #include <linux/cpu.h>
20 #include <linux/cache.h>
21 #include <linux/sched/sysctl.h>
22 #include <linux/delay.h>
23 #include <linux/crash_dump.h>
25 #include <trace/events/block.h>
27 #include <linux/blk-mq.h>
30 #include "blk-mq-tag.h"
32 static DEFINE_MUTEX(all_q_mutex);
33 static LIST_HEAD(all_q_list);
35 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
38 * Check if any of the ctx's have pending work in this hardware queue
40 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
44 for (i = 0; i < hctx->ctx_map.size; i++)
45 if (hctx->ctx_map.map[i].word)
51 static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx,
52 struct blk_mq_ctx *ctx)
54 return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word];
57 #define CTX_TO_BIT(hctx, ctx) \
58 ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
61 * Mark this ctx as having pending work in this hardware queue
63 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
64 struct blk_mq_ctx *ctx)
66 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
68 if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word))
69 set_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
72 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
73 struct blk_mq_ctx *ctx)
75 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
77 clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
80 static int blk_mq_queue_enter(struct request_queue *q, gfp_t gfp)
85 if (percpu_ref_tryget_live(&q->mq_usage_counter))
88 if (!(gfp & __GFP_WAIT))
91 ret = wait_event_interruptible(q->mq_freeze_wq,
92 !atomic_read(&q->mq_freeze_depth) ||
94 if (blk_queue_dying(q))
101 static void blk_mq_queue_exit(struct request_queue *q)
103 percpu_ref_put(&q->mq_usage_counter);
106 static void blk_mq_usage_counter_release(struct percpu_ref *ref)
108 struct request_queue *q =
109 container_of(ref, struct request_queue, mq_usage_counter);
111 wake_up_all(&q->mq_freeze_wq);
114 void blk_mq_freeze_queue_start(struct request_queue *q)
118 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
119 if (freeze_depth == 1) {
120 percpu_ref_kill(&q->mq_usage_counter);
121 blk_mq_run_hw_queues(q, false);
124 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
126 static void blk_mq_freeze_queue_wait(struct request_queue *q)
128 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->mq_usage_counter));
132 * Guarantee no request is in use, so we can change any data structure of
133 * the queue afterward.
135 void blk_mq_freeze_queue(struct request_queue *q)
137 blk_mq_freeze_queue_start(q);
138 blk_mq_freeze_queue_wait(q);
140 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
142 void blk_mq_unfreeze_queue(struct request_queue *q)
146 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
147 WARN_ON_ONCE(freeze_depth < 0);
149 percpu_ref_reinit(&q->mq_usage_counter);
150 wake_up_all(&q->mq_freeze_wq);
153 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
155 void blk_mq_wake_waiters(struct request_queue *q)
157 struct blk_mq_hw_ctx *hctx;
160 queue_for_each_hw_ctx(q, hctx, i)
161 if (blk_mq_hw_queue_mapped(hctx))
162 blk_mq_tag_wakeup_all(hctx->tags, true);
165 * If we are called because the queue has now been marked as
166 * dying, we need to ensure that processes currently waiting on
167 * the queue are notified as well.
169 wake_up_all(&q->mq_freeze_wq);
172 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
174 return blk_mq_has_free_tags(hctx->tags);
176 EXPORT_SYMBOL(blk_mq_can_queue);
178 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
179 struct request *rq, unsigned int rw_flags)
181 if (blk_queue_io_stat(q))
182 rw_flags |= REQ_IO_STAT;
184 INIT_LIST_HEAD(&rq->queuelist);
185 /* csd/requeue_work/fifo_time is initialized before use */
188 rq->cmd_flags |= rw_flags;
189 /* do not touch atomic flags, it needs atomic ops against the timer */
191 INIT_HLIST_NODE(&rq->hash);
192 RB_CLEAR_NODE(&rq->rb_node);
195 rq->start_time = jiffies;
196 #ifdef CONFIG_BLK_CGROUP
198 set_start_time_ns(rq);
199 rq->io_start_time_ns = 0;
201 rq->nr_phys_segments = 0;
202 #if defined(CONFIG_BLK_DEV_INTEGRITY)
203 rq->nr_integrity_segments = 0;
206 /* tag was already set */
216 INIT_LIST_HEAD(&rq->timeout_list);
220 rq->end_io_data = NULL;
223 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
226 static struct request *
227 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
232 tag = blk_mq_get_tag(data);
233 if (tag != BLK_MQ_TAG_FAIL) {
234 rq = data->hctx->tags->rqs[tag];
236 if (blk_mq_tag_busy(data->hctx)) {
237 rq->cmd_flags = REQ_MQ_INFLIGHT;
238 atomic_inc(&data->hctx->nr_active);
242 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
249 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
252 struct blk_mq_ctx *ctx;
253 struct blk_mq_hw_ctx *hctx;
255 struct blk_mq_alloc_data alloc_data;
258 ret = blk_mq_queue_enter(q, gfp);
262 ctx = blk_mq_get_ctx(q);
263 hctx = q->mq_ops->map_queue(q, ctx->cpu);
264 blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT,
265 reserved, ctx, hctx);
267 rq = __blk_mq_alloc_request(&alloc_data, rw);
268 if (!rq && (gfp & __GFP_WAIT)) {
269 __blk_mq_run_hw_queue(hctx);
272 ctx = blk_mq_get_ctx(q);
273 hctx = q->mq_ops->map_queue(q, ctx->cpu);
274 blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx,
276 rq = __blk_mq_alloc_request(&alloc_data, rw);
277 ctx = alloc_data.ctx;
281 blk_mq_queue_exit(q);
282 return ERR_PTR(-EWOULDBLOCK);
286 EXPORT_SYMBOL(blk_mq_alloc_request);
288 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
289 struct blk_mq_ctx *ctx, struct request *rq)
291 const int tag = rq->tag;
292 struct request_queue *q = rq->q;
294 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
295 atomic_dec(&hctx->nr_active);
298 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
299 blk_mq_put_tag(hctx, tag, &ctx->last_tag);
300 blk_mq_queue_exit(q);
303 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
305 struct blk_mq_ctx *ctx = rq->mq_ctx;
307 ctx->rq_completed[rq_is_sync(rq)]++;
308 __blk_mq_free_request(hctx, ctx, rq);
311 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
313 void blk_mq_free_request(struct request *rq)
315 struct blk_mq_hw_ctx *hctx;
316 struct request_queue *q = rq->q;
318 hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
319 blk_mq_free_hctx_request(hctx, rq);
321 EXPORT_SYMBOL_GPL(blk_mq_free_request);
323 inline void __blk_mq_end_request(struct request *rq, int error)
325 blk_account_io_done(rq);
328 rq->end_io(rq, error);
330 if (unlikely(blk_bidi_rq(rq)))
331 blk_mq_free_request(rq->next_rq);
332 blk_mq_free_request(rq);
335 EXPORT_SYMBOL(__blk_mq_end_request);
337 void blk_mq_end_request(struct request *rq, int error)
339 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
341 __blk_mq_end_request(rq, error);
343 EXPORT_SYMBOL(blk_mq_end_request);
345 static void __blk_mq_complete_request_remote(void *data)
347 struct request *rq = data;
349 rq->q->softirq_done_fn(rq);
352 static void blk_mq_ipi_complete_request(struct request *rq)
354 struct blk_mq_ctx *ctx = rq->mq_ctx;
358 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
359 rq->q->softirq_done_fn(rq);
364 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
365 shared = cpus_share_cache(cpu, ctx->cpu);
367 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
368 rq->csd.func = __blk_mq_complete_request_remote;
371 smp_call_function_single_async(ctx->cpu, &rq->csd);
373 rq->q->softirq_done_fn(rq);
378 void __blk_mq_complete_request(struct request *rq)
380 struct request_queue *q = rq->q;
382 if (!q->softirq_done_fn)
383 blk_mq_end_request(rq, rq->errors);
385 blk_mq_ipi_complete_request(rq);
389 * blk_mq_complete_request - end I/O on a request
390 * @rq: the request being processed
393 * Ends all I/O on a request. It does not handle partial completions.
394 * The actual completion happens out-of-order, through a IPI handler.
396 void blk_mq_complete_request(struct request *rq)
398 struct request_queue *q = rq->q;
400 if (unlikely(blk_should_fake_timeout(q)))
402 if (!blk_mark_rq_complete(rq))
403 __blk_mq_complete_request(rq);
405 EXPORT_SYMBOL(blk_mq_complete_request);
407 int blk_mq_request_started(struct request *rq)
409 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
411 EXPORT_SYMBOL_GPL(blk_mq_request_started);
413 void blk_mq_start_request(struct request *rq)
415 struct request_queue *q = rq->q;
417 trace_block_rq_issue(q, rq);
419 rq->resid_len = blk_rq_bytes(rq);
420 if (unlikely(blk_bidi_rq(rq)))
421 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
426 * Ensure that ->deadline is visible before set the started
427 * flag and clear the completed flag.
429 smp_mb__before_atomic();
432 * Mark us as started and clear complete. Complete might have been
433 * set if requeue raced with timeout, which then marked it as
434 * complete. So be sure to clear complete again when we start
435 * the request, otherwise we'll ignore the completion event.
437 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
438 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
439 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
440 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
442 if (q->dma_drain_size && blk_rq_bytes(rq)) {
444 * Make sure space for the drain appears. We know we can do
445 * this because max_hw_segments has been adjusted to be one
446 * fewer than the device can handle.
448 rq->nr_phys_segments++;
451 EXPORT_SYMBOL(blk_mq_start_request);
453 static void __blk_mq_requeue_request(struct request *rq)
455 struct request_queue *q = rq->q;
457 trace_block_rq_requeue(q, rq);
459 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
460 if (q->dma_drain_size && blk_rq_bytes(rq))
461 rq->nr_phys_segments--;
465 void blk_mq_requeue_request(struct request *rq)
467 __blk_mq_requeue_request(rq);
469 BUG_ON(blk_queued_rq(rq));
470 blk_mq_add_to_requeue_list(rq, true);
472 EXPORT_SYMBOL(blk_mq_requeue_request);
474 static void blk_mq_requeue_work(struct work_struct *work)
476 struct request_queue *q =
477 container_of(work, struct request_queue, requeue_work);
479 struct request *rq, *next;
482 spin_lock_irqsave(&q->requeue_lock, flags);
483 list_splice_init(&q->requeue_list, &rq_list);
484 spin_unlock_irqrestore(&q->requeue_lock, flags);
486 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
487 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
490 rq->cmd_flags &= ~REQ_SOFTBARRIER;
491 list_del_init(&rq->queuelist);
492 blk_mq_insert_request(rq, true, false, false);
495 while (!list_empty(&rq_list)) {
496 rq = list_entry(rq_list.next, struct request, queuelist);
497 list_del_init(&rq->queuelist);
498 blk_mq_insert_request(rq, false, false, false);
502 * Use the start variant of queue running here, so that running
503 * the requeue work will kick stopped queues.
505 blk_mq_start_hw_queues(q);
508 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
510 struct request_queue *q = rq->q;
514 * We abuse this flag that is otherwise used by the I/O scheduler to
515 * request head insertation from the workqueue.
517 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
519 spin_lock_irqsave(&q->requeue_lock, flags);
521 rq->cmd_flags |= REQ_SOFTBARRIER;
522 list_add(&rq->queuelist, &q->requeue_list);
524 list_add_tail(&rq->queuelist, &q->requeue_list);
526 spin_unlock_irqrestore(&q->requeue_lock, flags);
528 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
530 void blk_mq_cancel_requeue_work(struct request_queue *q)
532 cancel_work_sync(&q->requeue_work);
534 EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work);
536 void blk_mq_kick_requeue_list(struct request_queue *q)
538 kblockd_schedule_work(&q->requeue_work);
540 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
542 void blk_mq_abort_requeue_list(struct request_queue *q)
547 spin_lock_irqsave(&q->requeue_lock, flags);
548 list_splice_init(&q->requeue_list, &rq_list);
549 spin_unlock_irqrestore(&q->requeue_lock, flags);
551 while (!list_empty(&rq_list)) {
554 rq = list_first_entry(&rq_list, struct request, queuelist);
555 list_del_init(&rq->queuelist);
557 blk_mq_end_request(rq, rq->errors);
560 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
562 static inline bool is_flush_request(struct request *rq,
563 struct blk_flush_queue *fq, unsigned int tag)
565 return ((rq->cmd_flags & REQ_FLUSH_SEQ) &&
566 fq->flush_rq->tag == tag);
569 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
571 struct request *rq = tags->rqs[tag];
572 /* mq_ctx of flush rq is always cloned from the corresponding req */
573 struct blk_flush_queue *fq = blk_get_flush_queue(rq->q, rq->mq_ctx);
575 if (!is_flush_request(rq, fq, tag))
580 EXPORT_SYMBOL(blk_mq_tag_to_rq);
582 struct blk_mq_timeout_data {
584 unsigned int next_set;
587 void blk_mq_rq_timed_out(struct request *req, bool reserved)
589 struct blk_mq_ops *ops = req->q->mq_ops;
590 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
593 * We know that complete is set at this point. If STARTED isn't set
594 * anymore, then the request isn't active and the "timeout" should
595 * just be ignored. This can happen due to the bitflag ordering.
596 * Timeout first checks if STARTED is set, and if it is, assumes
597 * the request is active. But if we race with completion, then
598 * we both flags will get cleared. So check here again, and ignore
599 * a timeout event with a request that isn't active.
601 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
605 ret = ops->timeout(req, reserved);
609 __blk_mq_complete_request(req);
611 case BLK_EH_RESET_TIMER:
613 blk_clear_rq_complete(req);
615 case BLK_EH_NOT_HANDLED:
618 printk(KERN_ERR "block: bad eh return: %d\n", ret);
623 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
624 struct request *rq, void *priv, bool reserved)
626 struct blk_mq_timeout_data *data = priv;
628 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
630 * If a request wasn't started before the queue was
631 * marked dying, kill it here or it'll go unnoticed.
633 if (unlikely(blk_queue_dying(rq->q))) {
635 blk_mq_complete_request(rq);
639 if (rq->cmd_flags & REQ_NO_TIMEOUT)
642 if (time_after_eq(jiffies, rq->deadline)) {
643 if (!blk_mark_rq_complete(rq))
644 blk_mq_rq_timed_out(rq, reserved);
645 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
646 data->next = rq->deadline;
651 static void blk_mq_rq_timer(unsigned long priv)
653 struct request_queue *q = (struct request_queue *)priv;
654 struct blk_mq_timeout_data data = {
658 struct blk_mq_hw_ctx *hctx;
661 queue_for_each_hw_ctx(q, hctx, i) {
663 * If not software queues are currently mapped to this
664 * hardware queue, there's nothing to check
666 if (!blk_mq_hw_queue_mapped(hctx))
669 blk_mq_tag_busy_iter(hctx, blk_mq_check_expired, &data);
673 data.next = blk_rq_timeout(round_jiffies_up(data.next));
674 mod_timer(&q->timeout, data.next);
676 queue_for_each_hw_ctx(q, hctx, i) {
677 /* the hctx may be unmapped, so check it here */
678 if (blk_mq_hw_queue_mapped(hctx))
679 blk_mq_tag_idle(hctx);
685 * Reverse check our software queue for entries that we could potentially
686 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
687 * too much time checking for merges.
689 static bool blk_mq_attempt_merge(struct request_queue *q,
690 struct blk_mq_ctx *ctx, struct bio *bio)
695 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
701 if (!blk_rq_merge_ok(rq, bio))
704 el_ret = blk_try_merge(rq, bio);
705 if (el_ret == ELEVATOR_BACK_MERGE) {
706 if (bio_attempt_back_merge(q, rq, bio)) {
711 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
712 if (bio_attempt_front_merge(q, rq, bio)) {
724 * Process software queues that have been marked busy, splicing them
725 * to the for-dispatch
727 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
729 struct blk_mq_ctx *ctx;
732 for (i = 0; i < hctx->ctx_map.size; i++) {
733 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
734 unsigned int off, bit;
740 off = i * hctx->ctx_map.bits_per_word;
742 bit = find_next_bit(&bm->word, bm->depth, bit);
743 if (bit >= bm->depth)
746 ctx = hctx->ctxs[bit + off];
747 clear_bit(bit, &bm->word);
748 spin_lock(&ctx->lock);
749 list_splice_tail_init(&ctx->rq_list, list);
750 spin_unlock(&ctx->lock);
758 * Run this hardware queue, pulling any software queues mapped to it in.
759 * Note that this function currently has various problems around ordering
760 * of IO. In particular, we'd like FIFO behaviour on handling existing
761 * items on the hctx->dispatch list. Ignore that for now.
763 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
765 struct request_queue *q = hctx->queue;
768 LIST_HEAD(driver_list);
769 struct list_head *dptr;
772 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
774 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
780 * Touch any software queue that has pending entries.
782 flush_busy_ctxs(hctx, &rq_list);
785 * If we have previous entries on our dispatch list, grab them
786 * and stuff them at the front for more fair dispatch.
788 if (!list_empty_careful(&hctx->dispatch)) {
789 spin_lock(&hctx->lock);
790 if (!list_empty(&hctx->dispatch))
791 list_splice_init(&hctx->dispatch, &rq_list);
792 spin_unlock(&hctx->lock);
796 * Start off with dptr being NULL, so we start the first request
797 * immediately, even if we have more pending.
802 * Now process all the entries, sending them to the driver.
805 while (!list_empty(&rq_list)) {
806 struct blk_mq_queue_data bd;
809 rq = list_first_entry(&rq_list, struct request, queuelist);
810 list_del_init(&rq->queuelist);
814 bd.last = list_empty(&rq_list);
816 ret = q->mq_ops->queue_rq(hctx, &bd);
818 case BLK_MQ_RQ_QUEUE_OK:
821 case BLK_MQ_RQ_QUEUE_BUSY:
822 list_add(&rq->queuelist, &rq_list);
823 __blk_mq_requeue_request(rq);
826 pr_err("blk-mq: bad return on queue: %d\n", ret);
827 case BLK_MQ_RQ_QUEUE_ERROR:
829 blk_mq_end_request(rq, rq->errors);
833 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
837 * We've done the first request. If we have more than 1
838 * left in the list, set dptr to defer issue.
840 if (!dptr && rq_list.next != rq_list.prev)
845 hctx->dispatched[0]++;
846 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
847 hctx->dispatched[ilog2(queued) + 1]++;
850 * Any items that need requeuing? Stuff them into hctx->dispatch,
851 * that is where we will continue on next queue run.
853 if (!list_empty(&rq_list)) {
854 spin_lock(&hctx->lock);
855 list_splice(&rq_list, &hctx->dispatch);
856 spin_unlock(&hctx->lock);
858 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
859 * it's possible the queue is stopped and restarted again
860 * before this. Queue restart will dispatch requests. And since
861 * requests in rq_list aren't added into hctx->dispatch yet,
862 * the requests in rq_list might get lost.
864 * blk_mq_run_hw_queue() already checks the STOPPED bit
866 blk_mq_run_hw_queue(hctx, true);
871 * It'd be great if the workqueue API had a way to pass
872 * in a mask and had some smarts for more clever placement.
873 * For now we just round-robin here, switching for every
874 * BLK_MQ_CPU_WORK_BATCH queued items.
876 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
878 if (hctx->queue->nr_hw_queues == 1)
879 return WORK_CPU_UNBOUND;
881 if (--hctx->next_cpu_batch <= 0) {
882 int cpu = hctx->next_cpu, next_cpu;
884 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
885 if (next_cpu >= nr_cpu_ids)
886 next_cpu = cpumask_first(hctx->cpumask);
888 hctx->next_cpu = next_cpu;
889 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
894 return hctx->next_cpu;
897 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
899 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
900 !blk_mq_hw_queue_mapped(hctx)))
905 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
906 __blk_mq_run_hw_queue(hctx);
914 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
918 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
920 struct blk_mq_hw_ctx *hctx;
923 queue_for_each_hw_ctx(q, hctx, i) {
924 if ((!blk_mq_hctx_has_pending(hctx) &&
925 list_empty_careful(&hctx->dispatch)) ||
926 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
929 blk_mq_run_hw_queue(hctx, async);
932 EXPORT_SYMBOL(blk_mq_run_hw_queues);
934 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
936 cancel_delayed_work(&hctx->run_work);
937 cancel_delayed_work(&hctx->delay_work);
938 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
940 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
942 void blk_mq_stop_hw_queues(struct request_queue *q)
944 struct blk_mq_hw_ctx *hctx;
947 queue_for_each_hw_ctx(q, hctx, i)
948 blk_mq_stop_hw_queue(hctx);
950 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
952 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
954 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
956 blk_mq_run_hw_queue(hctx, false);
958 EXPORT_SYMBOL(blk_mq_start_hw_queue);
960 void blk_mq_start_hw_queues(struct request_queue *q)
962 struct blk_mq_hw_ctx *hctx;
965 queue_for_each_hw_ctx(q, hctx, i)
966 blk_mq_start_hw_queue(hctx);
968 EXPORT_SYMBOL(blk_mq_start_hw_queues);
970 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
972 struct blk_mq_hw_ctx *hctx;
975 queue_for_each_hw_ctx(q, hctx, i) {
976 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
979 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
980 blk_mq_run_hw_queue(hctx, async);
983 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
985 static void blk_mq_run_work_fn(struct work_struct *work)
987 struct blk_mq_hw_ctx *hctx;
989 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
991 __blk_mq_run_hw_queue(hctx);
994 static void blk_mq_delay_work_fn(struct work_struct *work)
996 struct blk_mq_hw_ctx *hctx;
998 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
1000 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
1001 __blk_mq_run_hw_queue(hctx);
1004 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1006 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1009 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1010 &hctx->delay_work, msecs_to_jiffies(msecs));
1012 EXPORT_SYMBOL(blk_mq_delay_queue);
1014 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
1015 struct request *rq, bool at_head)
1017 struct blk_mq_ctx *ctx = rq->mq_ctx;
1019 trace_block_rq_insert(hctx->queue, rq);
1022 list_add(&rq->queuelist, &ctx->rq_list);
1024 list_add_tail(&rq->queuelist, &ctx->rq_list);
1026 blk_mq_hctx_mark_pending(hctx, ctx);
1029 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1032 struct request_queue *q = rq->q;
1033 struct blk_mq_hw_ctx *hctx;
1034 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
1036 current_ctx = blk_mq_get_ctx(q);
1037 if (!cpu_online(ctx->cpu))
1038 rq->mq_ctx = ctx = current_ctx;
1040 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1042 spin_lock(&ctx->lock);
1043 __blk_mq_insert_request(hctx, rq, at_head);
1044 spin_unlock(&ctx->lock);
1047 blk_mq_run_hw_queue(hctx, async);
1049 blk_mq_put_ctx(current_ctx);
1052 static void blk_mq_insert_requests(struct request_queue *q,
1053 struct blk_mq_ctx *ctx,
1054 struct list_head *list,
1059 struct blk_mq_hw_ctx *hctx;
1060 struct blk_mq_ctx *current_ctx;
1062 trace_block_unplug(q, depth, !from_schedule);
1064 current_ctx = blk_mq_get_ctx(q);
1066 if (!cpu_online(ctx->cpu))
1068 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1071 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1074 spin_lock(&ctx->lock);
1075 while (!list_empty(list)) {
1078 rq = list_first_entry(list, struct request, queuelist);
1079 list_del_init(&rq->queuelist);
1081 __blk_mq_insert_request(hctx, rq, false);
1083 spin_unlock(&ctx->lock);
1085 blk_mq_run_hw_queue(hctx, from_schedule);
1086 blk_mq_put_ctx(current_ctx);
1089 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1091 struct request *rqa = container_of(a, struct request, queuelist);
1092 struct request *rqb = container_of(b, struct request, queuelist);
1094 return !(rqa->mq_ctx < rqb->mq_ctx ||
1095 (rqa->mq_ctx == rqb->mq_ctx &&
1096 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1099 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1101 struct blk_mq_ctx *this_ctx;
1102 struct request_queue *this_q;
1105 LIST_HEAD(ctx_list);
1108 list_splice_init(&plug->mq_list, &list);
1110 list_sort(NULL, &list, plug_ctx_cmp);
1116 while (!list_empty(&list)) {
1117 rq = list_entry_rq(list.next);
1118 list_del_init(&rq->queuelist);
1120 if (rq->mq_ctx != this_ctx) {
1122 blk_mq_insert_requests(this_q, this_ctx,
1127 this_ctx = rq->mq_ctx;
1133 list_add_tail(&rq->queuelist, &ctx_list);
1137 * If 'this_ctx' is set, we know we have entries to complete
1138 * on 'ctx_list'. Do those.
1141 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1146 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1148 init_request_from_bio(rq, bio);
1150 if (blk_do_io_stat(rq))
1151 blk_account_io_start(rq, 1);
1154 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1156 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1157 !blk_queue_nomerges(hctx->queue);
1160 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1161 struct blk_mq_ctx *ctx,
1162 struct request *rq, struct bio *bio)
1164 if (!hctx_allow_merges(hctx)) {
1165 blk_mq_bio_to_request(rq, bio);
1166 spin_lock(&ctx->lock);
1168 __blk_mq_insert_request(hctx, rq, false);
1169 spin_unlock(&ctx->lock);
1172 struct request_queue *q = hctx->queue;
1174 spin_lock(&ctx->lock);
1175 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1176 blk_mq_bio_to_request(rq, bio);
1180 spin_unlock(&ctx->lock);
1181 __blk_mq_free_request(hctx, ctx, rq);
1186 struct blk_map_ctx {
1187 struct blk_mq_hw_ctx *hctx;
1188 struct blk_mq_ctx *ctx;
1191 static struct request *blk_mq_map_request(struct request_queue *q,
1193 struct blk_map_ctx *data)
1195 struct blk_mq_hw_ctx *hctx;
1196 struct blk_mq_ctx *ctx;
1198 int rw = bio_data_dir(bio);
1199 struct blk_mq_alloc_data alloc_data;
1201 if (unlikely(blk_mq_queue_enter(q, GFP_KERNEL))) {
1206 ctx = blk_mq_get_ctx(q);
1207 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1209 if (rw_is_sync(bio->bi_rw))
1212 trace_block_getrq(q, bio, rw);
1213 blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1215 rq = __blk_mq_alloc_request(&alloc_data, rw);
1216 if (unlikely(!rq)) {
1217 __blk_mq_run_hw_queue(hctx);
1218 blk_mq_put_ctx(ctx);
1219 trace_block_sleeprq(q, bio, rw);
1221 ctx = blk_mq_get_ctx(q);
1222 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1223 blk_mq_set_alloc_data(&alloc_data, q,
1224 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1225 rq = __blk_mq_alloc_request(&alloc_data, rw);
1226 ctx = alloc_data.ctx;
1227 hctx = alloc_data.hctx;
1236 static int blk_mq_direct_issue_request(struct request *rq)
1239 struct request_queue *q = rq->q;
1240 struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q,
1242 struct blk_mq_queue_data bd = {
1249 * For OK queue, we are done. For error, kill it. Any other
1250 * error (busy), just add it to our list as we previously
1253 ret = q->mq_ops->queue_rq(hctx, &bd);
1254 if (ret == BLK_MQ_RQ_QUEUE_OK)
1257 __blk_mq_requeue_request(rq);
1259 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1261 blk_mq_end_request(rq, rq->errors);
1269 * Multiple hardware queue variant. This will not use per-process plugs,
1270 * but will attempt to bypass the hctx queueing if we can go straight to
1271 * hardware for SYNC IO.
1273 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1275 const int is_sync = rw_is_sync(bio->bi_rw);
1276 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1277 struct blk_map_ctx data;
1279 unsigned int request_count = 0;
1280 struct blk_plug *plug;
1281 struct request *same_queue_rq = NULL;
1283 blk_queue_bounce(q, &bio);
1285 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1290 blk_queue_split(q, &bio, q->bio_split);
1292 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1293 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1296 rq = blk_mq_map_request(q, bio, &data);
1300 if (unlikely(is_flush_fua)) {
1301 blk_mq_bio_to_request(rq, bio);
1302 blk_insert_flush(rq);
1306 plug = current->plug;
1308 * If the driver supports defer issued based on 'last', then
1309 * queue it up like normal since we can potentially save some
1312 if (((plug && !blk_queue_nomerges(q)) || is_sync) &&
1313 !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1314 struct request *old_rq = NULL;
1316 blk_mq_bio_to_request(rq, bio);
1319 * we do limited pluging. If bio can be merged, do merge.
1320 * Otherwise the existing request in the plug list will be
1321 * issued. So the plug list will have one request at most
1325 * The plug list might get flushed before this. If that
1326 * happens, same_queue_rq is invalid and plug list is empty
1328 if (same_queue_rq && !list_empty(&plug->mq_list)) {
1329 old_rq = same_queue_rq;
1330 list_del_init(&old_rq->queuelist);
1332 list_add_tail(&rq->queuelist, &plug->mq_list);
1333 } else /* is_sync */
1335 blk_mq_put_ctx(data.ctx);
1338 if (!blk_mq_direct_issue_request(old_rq))
1340 blk_mq_insert_request(old_rq, false, true, true);
1344 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1346 * For a SYNC request, send it to the hardware immediately. For
1347 * an ASYNC request, just ensure that we run it later on. The
1348 * latter allows for merging opportunities and more efficient
1352 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1354 blk_mq_put_ctx(data.ctx);
1358 * Single hardware queue variant. This will attempt to use any per-process
1359 * plug for merging and IO deferral.
1361 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1363 const int is_sync = rw_is_sync(bio->bi_rw);
1364 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1365 struct blk_plug *plug;
1366 unsigned int request_count = 0;
1367 struct blk_map_ctx data;
1370 blk_queue_bounce(q, &bio);
1372 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1377 blk_queue_split(q, &bio, q->bio_split);
1379 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1380 blk_attempt_plug_merge(q, bio, &request_count, NULL))
1383 rq = blk_mq_map_request(q, bio, &data);
1387 if (unlikely(is_flush_fua)) {
1388 blk_mq_bio_to_request(rq, bio);
1389 blk_insert_flush(rq);
1394 * A task plug currently exists. Since this is completely lockless,
1395 * utilize that to temporarily store requests until the task is
1396 * either done or scheduled away.
1398 plug = current->plug;
1400 blk_mq_bio_to_request(rq, bio);
1401 if (list_empty(&plug->mq_list))
1402 trace_block_plug(q);
1403 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1404 blk_flush_plug_list(plug, false);
1405 trace_block_plug(q);
1407 list_add_tail(&rq->queuelist, &plug->mq_list);
1408 blk_mq_put_ctx(data.ctx);
1412 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1414 * For a SYNC request, send it to the hardware immediately. For
1415 * an ASYNC request, just ensure that we run it later on. The
1416 * latter allows for merging opportunities and more efficient
1420 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1423 blk_mq_put_ctx(data.ctx);
1427 * Default mapping to a software queue, since we use one per CPU.
1429 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1431 return q->queue_hw_ctx[q->mq_map[cpu]];
1433 EXPORT_SYMBOL(blk_mq_map_queue);
1435 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1436 struct blk_mq_tags *tags, unsigned int hctx_idx)
1440 if (tags->rqs && set->ops->exit_request) {
1443 for (i = 0; i < tags->nr_tags; i++) {
1446 set->ops->exit_request(set->driver_data, tags->rqs[i],
1448 tags->rqs[i] = NULL;
1452 while (!list_empty(&tags->page_list)) {
1453 page = list_first_entry(&tags->page_list, struct page, lru);
1454 list_del_init(&page->lru);
1455 __free_pages(page, page->private);
1460 blk_mq_free_tags(tags);
1463 static size_t order_to_size(unsigned int order)
1465 return (size_t)PAGE_SIZE << order;
1468 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1469 unsigned int hctx_idx)
1471 struct blk_mq_tags *tags;
1472 unsigned int i, j, entries_per_page, max_order = 4;
1473 size_t rq_size, left;
1475 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1477 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1481 INIT_LIST_HEAD(&tags->page_list);
1483 tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1484 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1487 blk_mq_free_tags(tags);
1492 * rq_size is the size of the request plus driver payload, rounded
1493 * to the cacheline size
1495 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1497 left = rq_size * set->queue_depth;
1499 for (i = 0; i < set->queue_depth; ) {
1500 int this_order = max_order;
1505 while (left < order_to_size(this_order - 1) && this_order)
1509 page = alloc_pages_node(set->numa_node,
1510 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1516 if (order_to_size(this_order) < rq_size)
1523 page->private = this_order;
1524 list_add_tail(&page->lru, &tags->page_list);
1526 p = page_address(page);
1527 entries_per_page = order_to_size(this_order) / rq_size;
1528 to_do = min(entries_per_page, set->queue_depth - i);
1529 left -= to_do * rq_size;
1530 for (j = 0; j < to_do; j++) {
1532 if (set->ops->init_request) {
1533 if (set->ops->init_request(set->driver_data,
1534 tags->rqs[i], hctx_idx, i,
1536 tags->rqs[i] = NULL;
1548 blk_mq_free_rq_map(set, tags, hctx_idx);
1552 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1557 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1559 unsigned int bpw = 8, total, num_maps, i;
1561 bitmap->bits_per_word = bpw;
1563 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1564 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1570 for (i = 0; i < num_maps; i++) {
1571 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1572 total -= bitmap->map[i].depth;
1578 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1580 struct request_queue *q = hctx->queue;
1581 struct blk_mq_ctx *ctx;
1585 * Move ctx entries to new CPU, if this one is going away.
1587 ctx = __blk_mq_get_ctx(q, cpu);
1589 spin_lock(&ctx->lock);
1590 if (!list_empty(&ctx->rq_list)) {
1591 list_splice_init(&ctx->rq_list, &tmp);
1592 blk_mq_hctx_clear_pending(hctx, ctx);
1594 spin_unlock(&ctx->lock);
1596 if (list_empty(&tmp))
1599 ctx = blk_mq_get_ctx(q);
1600 spin_lock(&ctx->lock);
1602 while (!list_empty(&tmp)) {
1605 rq = list_first_entry(&tmp, struct request, queuelist);
1607 list_move_tail(&rq->queuelist, &ctx->rq_list);
1610 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1611 blk_mq_hctx_mark_pending(hctx, ctx);
1613 spin_unlock(&ctx->lock);
1615 blk_mq_run_hw_queue(hctx, true);
1616 blk_mq_put_ctx(ctx);
1620 static int blk_mq_hctx_notify(void *data, unsigned long action,
1623 struct blk_mq_hw_ctx *hctx = data;
1625 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1626 return blk_mq_hctx_cpu_offline(hctx, cpu);
1629 * In case of CPU online, tags may be reallocated
1630 * in blk_mq_map_swqueue() after mapping is updated.
1636 /* hctx->ctxs will be freed in queue's release handler */
1637 static void blk_mq_exit_hctx(struct request_queue *q,
1638 struct blk_mq_tag_set *set,
1639 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1641 unsigned flush_start_tag = set->queue_depth;
1643 blk_mq_tag_idle(hctx);
1645 if (set->ops->exit_request)
1646 set->ops->exit_request(set->driver_data,
1647 hctx->fq->flush_rq, hctx_idx,
1648 flush_start_tag + hctx_idx);
1650 if (set->ops->exit_hctx)
1651 set->ops->exit_hctx(hctx, hctx_idx);
1653 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1654 blk_free_flush_queue(hctx->fq);
1655 blk_mq_free_bitmap(&hctx->ctx_map);
1658 static void blk_mq_exit_hw_queues(struct request_queue *q,
1659 struct blk_mq_tag_set *set, int nr_queue)
1661 struct blk_mq_hw_ctx *hctx;
1664 queue_for_each_hw_ctx(q, hctx, i) {
1667 blk_mq_exit_hctx(q, set, hctx, i);
1671 static void blk_mq_free_hw_queues(struct request_queue *q,
1672 struct blk_mq_tag_set *set)
1674 struct blk_mq_hw_ctx *hctx;
1677 queue_for_each_hw_ctx(q, hctx, i)
1678 free_cpumask_var(hctx->cpumask);
1681 static int blk_mq_init_hctx(struct request_queue *q,
1682 struct blk_mq_tag_set *set,
1683 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1686 unsigned flush_start_tag = set->queue_depth;
1688 node = hctx->numa_node;
1689 if (node == NUMA_NO_NODE)
1690 node = hctx->numa_node = set->numa_node;
1692 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1693 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1694 spin_lock_init(&hctx->lock);
1695 INIT_LIST_HEAD(&hctx->dispatch);
1697 hctx->queue_num = hctx_idx;
1698 hctx->flags = set->flags;
1700 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1701 blk_mq_hctx_notify, hctx);
1702 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1704 hctx->tags = set->tags[hctx_idx];
1707 * Allocate space for all possible cpus to avoid allocation at
1710 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1713 goto unregister_cpu_notifier;
1715 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1720 if (set->ops->init_hctx &&
1721 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1724 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1728 if (set->ops->init_request &&
1729 set->ops->init_request(set->driver_data,
1730 hctx->fq->flush_rq, hctx_idx,
1731 flush_start_tag + hctx_idx, node))
1739 if (set->ops->exit_hctx)
1740 set->ops->exit_hctx(hctx, hctx_idx);
1742 blk_mq_free_bitmap(&hctx->ctx_map);
1745 unregister_cpu_notifier:
1746 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1751 static int blk_mq_init_hw_queues(struct request_queue *q,
1752 struct blk_mq_tag_set *set)
1754 struct blk_mq_hw_ctx *hctx;
1758 * Initialize hardware queues
1760 queue_for_each_hw_ctx(q, hctx, i) {
1761 if (blk_mq_init_hctx(q, set, hctx, i))
1765 if (i == q->nr_hw_queues)
1771 blk_mq_exit_hw_queues(q, set, i);
1776 static void blk_mq_init_cpu_queues(struct request_queue *q,
1777 unsigned int nr_hw_queues)
1781 for_each_possible_cpu(i) {
1782 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1783 struct blk_mq_hw_ctx *hctx;
1785 memset(__ctx, 0, sizeof(*__ctx));
1787 spin_lock_init(&__ctx->lock);
1788 INIT_LIST_HEAD(&__ctx->rq_list);
1791 /* If the cpu isn't online, the cpu is mapped to first hctx */
1795 hctx = q->mq_ops->map_queue(q, i);
1798 * Set local node, IFF we have more than one hw queue. If
1799 * not, we remain on the home node of the device
1801 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1802 hctx->numa_node = cpu_to_node(i);
1806 static void blk_mq_map_swqueue(struct request_queue *q)
1809 struct blk_mq_hw_ctx *hctx;
1810 struct blk_mq_ctx *ctx;
1811 struct blk_mq_tag_set *set = q->tag_set;
1813 queue_for_each_hw_ctx(q, hctx, i) {
1814 cpumask_clear(hctx->cpumask);
1819 * Map software to hardware queues
1821 queue_for_each_ctx(q, ctx, i) {
1822 /* If the cpu isn't online, the cpu is mapped to first hctx */
1826 hctx = q->mq_ops->map_queue(q, i);
1827 cpumask_set_cpu(i, hctx->cpumask);
1828 cpumask_set_cpu(i, hctx->tags->cpumask);
1829 ctx->index_hw = hctx->nr_ctx;
1830 hctx->ctxs[hctx->nr_ctx++] = ctx;
1833 queue_for_each_hw_ctx(q, hctx, i) {
1834 struct blk_mq_ctxmap *map = &hctx->ctx_map;
1837 * If no software queues are mapped to this hardware queue,
1838 * disable it and free the request entries.
1840 if (!hctx->nr_ctx) {
1842 blk_mq_free_rq_map(set, set->tags[i], i);
1843 set->tags[i] = NULL;
1849 /* unmapped hw queue can be remapped after CPU topo changed */
1851 set->tags[i] = blk_mq_init_rq_map(set, i);
1852 hctx->tags = set->tags[i];
1853 WARN_ON(!hctx->tags);
1856 * Set the map size to the number of mapped software queues.
1857 * This is more accurate and more efficient than looping
1858 * over all possibly mapped software queues.
1860 map->size = DIV_ROUND_UP(hctx->nr_ctx, map->bits_per_word);
1863 * Initialize batch roundrobin counts
1865 hctx->next_cpu = cpumask_first(hctx->cpumask);
1866 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1870 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1872 struct blk_mq_hw_ctx *hctx;
1873 struct request_queue *q;
1877 if (set->tag_list.next == set->tag_list.prev)
1882 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1883 blk_mq_freeze_queue(q);
1885 queue_for_each_hw_ctx(q, hctx, i) {
1887 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1889 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1891 blk_mq_unfreeze_queue(q);
1895 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1897 struct blk_mq_tag_set *set = q->tag_set;
1899 mutex_lock(&set->tag_list_lock);
1900 list_del_init(&q->tag_set_list);
1901 blk_mq_update_tag_set_depth(set);
1902 mutex_unlock(&set->tag_list_lock);
1905 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1906 struct request_queue *q)
1910 mutex_lock(&set->tag_list_lock);
1911 list_add_tail(&q->tag_set_list, &set->tag_list);
1912 blk_mq_update_tag_set_depth(set);
1913 mutex_unlock(&set->tag_list_lock);
1917 * It is the actual release handler for mq, but we do it from
1918 * request queue's release handler for avoiding use-after-free
1919 * and headache because q->mq_kobj shouldn't have been introduced,
1920 * but we can't group ctx/kctx kobj without it.
1922 void blk_mq_release(struct request_queue *q)
1924 struct blk_mq_hw_ctx *hctx;
1927 /* hctx kobj stays in hctx */
1928 queue_for_each_hw_ctx(q, hctx, i) {
1935 kfree(q->queue_hw_ctx);
1937 /* ctx kobj stays in queue_ctx */
1938 free_percpu(q->queue_ctx);
1941 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1943 struct request_queue *uninit_q, *q;
1945 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1947 return ERR_PTR(-ENOMEM);
1949 q = blk_mq_init_allocated_queue(set, uninit_q);
1951 blk_cleanup_queue(uninit_q);
1955 EXPORT_SYMBOL(blk_mq_init_queue);
1957 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
1958 struct request_queue *q)
1960 struct blk_mq_hw_ctx **hctxs;
1961 struct blk_mq_ctx __percpu *ctx;
1965 ctx = alloc_percpu(struct blk_mq_ctx);
1967 return ERR_PTR(-ENOMEM);
1969 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1975 map = blk_mq_make_queue_map(set);
1979 for (i = 0; i < set->nr_hw_queues; i++) {
1980 int node = blk_mq_hw_queue_to_node(map, i);
1982 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1987 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
1991 atomic_set(&hctxs[i]->nr_active, 0);
1992 hctxs[i]->numa_node = node;
1993 hctxs[i]->queue_num = i;
1997 * Init percpu_ref in atomic mode so that it's faster to shutdown.
1998 * See blk_register_queue() for details.
2000 if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release,
2001 PERCPU_REF_INIT_ATOMIC, GFP_KERNEL))
2004 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
2005 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2007 q->nr_queues = nr_cpu_ids;
2008 q->nr_hw_queues = set->nr_hw_queues;
2012 q->queue_hw_ctx = hctxs;
2014 q->mq_ops = set->ops;
2015 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2017 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2018 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2020 q->sg_reserved_size = INT_MAX;
2022 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
2023 INIT_LIST_HEAD(&q->requeue_list);
2024 spin_lock_init(&q->requeue_lock);
2026 if (q->nr_hw_queues > 1)
2027 blk_queue_make_request(q, blk_mq_make_request);
2029 blk_queue_make_request(q, blk_sq_make_request);
2032 * Do this after blk_queue_make_request() overrides it...
2034 q->nr_requests = set->queue_depth;
2036 if (set->ops->complete)
2037 blk_queue_softirq_done(q, set->ops->complete);
2039 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2041 if (blk_mq_init_hw_queues(q, set))
2044 mutex_lock(&all_q_mutex);
2045 list_add_tail(&q->all_q_node, &all_q_list);
2046 mutex_unlock(&all_q_mutex);
2048 blk_mq_add_queue_tag_set(set, q);
2050 blk_mq_map_swqueue(q);
2056 for (i = 0; i < set->nr_hw_queues; i++) {
2059 free_cpumask_var(hctxs[i]->cpumask);
2066 return ERR_PTR(-ENOMEM);
2068 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2070 void blk_mq_free_queue(struct request_queue *q)
2072 struct blk_mq_tag_set *set = q->tag_set;
2074 blk_mq_del_queue_tag_set(q);
2076 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2077 blk_mq_free_hw_queues(q, set);
2079 percpu_ref_exit(&q->mq_usage_counter);
2085 mutex_lock(&all_q_mutex);
2086 list_del_init(&q->all_q_node);
2087 mutex_unlock(&all_q_mutex);
2090 /* Basically redo blk_mq_init_queue with queue frozen */
2091 static void blk_mq_queue_reinit(struct request_queue *q)
2093 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2095 blk_mq_sysfs_unregister(q);
2097 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
2100 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2101 * we should change hctx numa_node according to new topology (this
2102 * involves free and re-allocate memory, worthy doing?)
2105 blk_mq_map_swqueue(q);
2107 blk_mq_sysfs_register(q);
2110 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2111 unsigned long action, void *hcpu)
2113 struct request_queue *q;
2116 * Before new mappings are established, hotadded cpu might already
2117 * start handling requests. This doesn't break anything as we map
2118 * offline CPUs to first hardware queue. We will re-init the queue
2119 * below to get optimal settings.
2121 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
2122 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
2125 mutex_lock(&all_q_mutex);
2128 * We need to freeze and reinit all existing queues. Freezing
2129 * involves synchronous wait for an RCU grace period and doing it
2130 * one by one may take a long time. Start freezing all queues in
2131 * one swoop and then wait for the completions so that freezing can
2132 * take place in parallel.
2134 list_for_each_entry(q, &all_q_list, all_q_node)
2135 blk_mq_freeze_queue_start(q);
2136 list_for_each_entry(q, &all_q_list, all_q_node) {
2137 blk_mq_freeze_queue_wait(q);
2140 * timeout handler can't touch hw queue during the
2143 del_timer_sync(&q->timeout);
2146 list_for_each_entry(q, &all_q_list, all_q_node)
2147 blk_mq_queue_reinit(q);
2149 list_for_each_entry(q, &all_q_list, all_q_node)
2150 blk_mq_unfreeze_queue(q);
2152 mutex_unlock(&all_q_mutex);
2156 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2160 for (i = 0; i < set->nr_hw_queues; i++) {
2161 set->tags[i] = blk_mq_init_rq_map(set, i);
2170 blk_mq_free_rq_map(set, set->tags[i], i);
2176 * Allocate the request maps associated with this tag_set. Note that this
2177 * may reduce the depth asked for, if memory is tight. set->queue_depth
2178 * will be updated to reflect the allocated depth.
2180 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2185 depth = set->queue_depth;
2187 err = __blk_mq_alloc_rq_maps(set);
2191 set->queue_depth >>= 1;
2192 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2196 } while (set->queue_depth);
2198 if (!set->queue_depth || err) {
2199 pr_err("blk-mq: failed to allocate request map\n");
2203 if (depth != set->queue_depth)
2204 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2205 depth, set->queue_depth);
2210 struct cpumask *blk_mq_tags_cpumask(struct blk_mq_tags *tags)
2212 return tags->cpumask;
2214 EXPORT_SYMBOL_GPL(blk_mq_tags_cpumask);
2217 * Alloc a tag set to be associated with one or more request queues.
2218 * May fail with EINVAL for various error conditions. May adjust the
2219 * requested depth down, if if it too large. In that case, the set
2220 * value will be stored in set->queue_depth.
2222 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2224 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2226 if (!set->nr_hw_queues)
2228 if (!set->queue_depth)
2230 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2233 if (!set->ops->queue_rq || !set->ops->map_queue)
2236 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2237 pr_info("blk-mq: reduced tag depth to %u\n",
2239 set->queue_depth = BLK_MQ_MAX_DEPTH;
2243 * If a crashdump is active, then we are potentially in a very
2244 * memory constrained environment. Limit us to 1 queue and
2245 * 64 tags to prevent using too much memory.
2247 if (is_kdump_kernel()) {
2248 set->nr_hw_queues = 1;
2249 set->queue_depth = min(64U, set->queue_depth);
2252 set->tags = kmalloc_node(set->nr_hw_queues *
2253 sizeof(struct blk_mq_tags *),
2254 GFP_KERNEL, set->numa_node);
2258 if (blk_mq_alloc_rq_maps(set))
2261 mutex_init(&set->tag_list_lock);
2262 INIT_LIST_HEAD(&set->tag_list);
2270 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2272 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2276 for (i = 0; i < set->nr_hw_queues; i++) {
2278 blk_mq_free_rq_map(set, set->tags[i], i);
2279 free_cpumask_var(set->tags[i]->cpumask);
2286 EXPORT_SYMBOL(blk_mq_free_tag_set);
2288 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2290 struct blk_mq_tag_set *set = q->tag_set;
2291 struct blk_mq_hw_ctx *hctx;
2294 if (!set || nr > set->queue_depth)
2298 queue_for_each_hw_ctx(q, hctx, i) {
2299 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2305 q->nr_requests = nr;
2310 void blk_mq_disable_hotplug(void)
2312 mutex_lock(&all_q_mutex);
2315 void blk_mq_enable_hotplug(void)
2317 mutex_unlock(&all_q_mutex);
2320 static int __init blk_mq_init(void)
2324 hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2328 subsys_initcall(blk_mq_init);