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.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)
85 if (percpu_ref_tryget_live(&q->mq_usage_counter))
88 ret = wait_event_interruptible(q->mq_freeze_wq,
89 !q->mq_freeze_depth || blk_queue_dying(q));
90 if (blk_queue_dying(q))
97 static void blk_mq_queue_exit(struct request_queue *q)
99 percpu_ref_put(&q->mq_usage_counter);
102 static void blk_mq_usage_counter_release(struct percpu_ref *ref)
104 struct request_queue *q =
105 container_of(ref, struct request_queue, mq_usage_counter);
107 wake_up_all(&q->mq_freeze_wq);
110 void blk_mq_freeze_queue_start(struct request_queue *q)
114 spin_lock_irq(q->queue_lock);
115 freeze = !q->mq_freeze_depth++;
116 spin_unlock_irq(q->queue_lock);
119 percpu_ref_kill(&q->mq_usage_counter);
120 blk_mq_run_hw_queues(q, false);
123 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
125 static void blk_mq_freeze_queue_wait(struct request_queue *q)
127 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->mq_usage_counter));
131 * Guarantee no request is in use, so we can change any data structure of
132 * the queue afterward.
134 void blk_mq_freeze_queue(struct request_queue *q)
136 blk_mq_freeze_queue_start(q);
137 blk_mq_freeze_queue_wait(q);
139 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
141 void blk_mq_unfreeze_queue(struct request_queue *q)
145 spin_lock_irq(q->queue_lock);
146 wake = !--q->mq_freeze_depth;
147 WARN_ON_ONCE(q->mq_freeze_depth < 0);
148 spin_unlock_irq(q->queue_lock);
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);
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)
399 struct request_queue *q = rq->q;
401 if (unlikely(blk_should_fake_timeout(q)))
403 if (!blk_mark_rq_complete(rq))
404 __blk_mq_complete_request(rq);
406 EXPORT_SYMBOL(blk_mq_complete_request);
408 int blk_mq_request_started(struct request *rq)
410 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
412 EXPORT_SYMBOL_GPL(blk_mq_request_started);
414 void blk_mq_start_request(struct request *rq)
416 struct request_queue *q = rq->q;
418 trace_block_rq_issue(q, rq);
420 rq->resid_len = blk_rq_bytes(rq);
421 if (unlikely(blk_bidi_rq(rq)))
422 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
427 * Ensure that ->deadline is visible before set the started
428 * flag and clear the completed flag.
430 smp_mb__before_atomic();
433 * Mark us as started and clear complete. Complete might have been
434 * set if requeue raced with timeout, which then marked it as
435 * complete. So be sure to clear complete again when we start
436 * the request, otherwise we'll ignore the completion event.
438 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
439 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
440 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
441 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
443 if (q->dma_drain_size && blk_rq_bytes(rq)) {
445 * Make sure space for the drain appears. We know we can do
446 * this because max_hw_segments has been adjusted to be one
447 * fewer than the device can handle.
449 rq->nr_phys_segments++;
452 EXPORT_SYMBOL(blk_mq_start_request);
454 static void __blk_mq_requeue_request(struct request *rq)
456 struct request_queue *q = rq->q;
458 trace_block_rq_requeue(q, rq);
460 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
461 if (q->dma_drain_size && blk_rq_bytes(rq))
462 rq->nr_phys_segments--;
466 void blk_mq_requeue_request(struct request *rq)
468 __blk_mq_requeue_request(rq);
470 BUG_ON(blk_queued_rq(rq));
471 blk_mq_add_to_requeue_list(rq, true);
473 EXPORT_SYMBOL(blk_mq_requeue_request);
475 static void blk_mq_requeue_work(struct work_struct *work)
477 struct request_queue *q =
478 container_of(work, struct request_queue, requeue_work);
480 struct request *rq, *next;
483 spin_lock_irqsave(&q->requeue_lock, flags);
484 list_splice_init(&q->requeue_list, &rq_list);
485 spin_unlock_irqrestore(&q->requeue_lock, flags);
487 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
488 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
491 rq->cmd_flags &= ~REQ_SOFTBARRIER;
492 list_del_init(&rq->queuelist);
493 blk_mq_insert_request(rq, true, false, false);
496 while (!list_empty(&rq_list)) {
497 rq = list_entry(rq_list.next, struct request, queuelist);
498 list_del_init(&rq->queuelist);
499 blk_mq_insert_request(rq, false, false, false);
503 * Use the start variant of queue running here, so that running
504 * the requeue work will kick stopped queues.
506 blk_mq_start_hw_queues(q);
509 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
511 struct request_queue *q = rq->q;
515 * We abuse this flag that is otherwise used by the I/O scheduler to
516 * request head insertation from the workqueue.
518 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
520 spin_lock_irqsave(&q->requeue_lock, flags);
522 rq->cmd_flags |= REQ_SOFTBARRIER;
523 list_add(&rq->queuelist, &q->requeue_list);
525 list_add_tail(&rq->queuelist, &q->requeue_list);
527 spin_unlock_irqrestore(&q->requeue_lock, flags);
529 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
531 void blk_mq_cancel_requeue_work(struct request_queue *q)
533 cancel_work_sync(&q->requeue_work);
535 EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work);
537 void blk_mq_kick_requeue_list(struct request_queue *q)
539 kblockd_schedule_work(&q->requeue_work);
541 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
543 void blk_mq_abort_requeue_list(struct request_queue *q)
548 spin_lock_irqsave(&q->requeue_lock, flags);
549 list_splice_init(&q->requeue_list, &rq_list);
550 spin_unlock_irqrestore(&q->requeue_lock, flags);
552 while (!list_empty(&rq_list)) {
555 rq = list_first_entry(&rq_list, struct request, queuelist);
556 list_del_init(&rq->queuelist);
558 blk_mq_end_request(rq, rq->errors);
561 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
563 static inline bool is_flush_request(struct request *rq,
564 struct blk_flush_queue *fq, unsigned int tag)
566 return ((rq->cmd_flags & REQ_FLUSH_SEQ) &&
567 fq->flush_rq->tag == tag);
570 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
572 struct request *rq = tags->rqs[tag];
573 /* mq_ctx of flush rq is always cloned from the corresponding req */
574 struct blk_flush_queue *fq = blk_get_flush_queue(rq->q, rq->mq_ctx);
576 if (!is_flush_request(rq, fq, tag))
581 EXPORT_SYMBOL(blk_mq_tag_to_rq);
583 struct blk_mq_timeout_data {
585 unsigned int next_set;
588 void blk_mq_rq_timed_out(struct request *req, bool reserved)
590 struct blk_mq_ops *ops = req->q->mq_ops;
591 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
594 * We know that complete is set at this point. If STARTED isn't set
595 * anymore, then the request isn't active and the "timeout" should
596 * just be ignored. This can happen due to the bitflag ordering.
597 * Timeout first checks if STARTED is set, and if it is, assumes
598 * the request is active. But if we race with completion, then
599 * we both flags will get cleared. So check here again, and ignore
600 * a timeout event with a request that isn't active.
602 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
606 ret = ops->timeout(req, reserved);
610 __blk_mq_complete_request(req);
612 case BLK_EH_RESET_TIMER:
614 blk_clear_rq_complete(req);
616 case BLK_EH_NOT_HANDLED:
619 printk(KERN_ERR "block: bad eh return: %d\n", ret);
624 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
625 struct request *rq, void *priv, bool reserved)
627 struct blk_mq_timeout_data *data = priv;
629 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
631 * If a request wasn't started before the queue was
632 * marked dying, kill it here or it'll go unnoticed.
634 if (unlikely(blk_queue_dying(rq->q))) {
636 blk_mq_complete_request(rq);
640 if (rq->cmd_flags & REQ_NO_TIMEOUT)
643 if (time_after_eq(jiffies, rq->deadline)) {
644 if (!blk_mark_rq_complete(rq))
645 blk_mq_rq_timed_out(rq, reserved);
646 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
647 data->next = rq->deadline;
652 static void blk_mq_rq_timer(unsigned long priv)
654 struct request_queue *q = (struct request_queue *)priv;
655 struct blk_mq_timeout_data data = {
659 struct blk_mq_hw_ctx *hctx;
662 queue_for_each_hw_ctx(q, hctx, i) {
664 * If not software queues are currently mapped to this
665 * hardware queue, there's nothing to check
667 if (!blk_mq_hw_queue_mapped(hctx))
670 blk_mq_tag_busy_iter(hctx, blk_mq_check_expired, &data);
674 data.next = blk_rq_timeout(round_jiffies_up(data.next));
675 mod_timer(&q->timeout, data.next);
677 queue_for_each_hw_ctx(q, hctx, i)
678 blk_mq_tag_idle(hctx);
683 * Reverse check our software queue for entries that we could potentially
684 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
685 * too much time checking for merges.
687 static bool blk_mq_attempt_merge(struct request_queue *q,
688 struct blk_mq_ctx *ctx, struct bio *bio)
693 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
699 if (!blk_rq_merge_ok(rq, bio))
702 el_ret = blk_try_merge(rq, bio);
703 if (el_ret == ELEVATOR_BACK_MERGE) {
704 if (bio_attempt_back_merge(q, rq, bio)) {
709 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
710 if (bio_attempt_front_merge(q, rq, bio)) {
722 * Process software queues that have been marked busy, splicing them
723 * to the for-dispatch
725 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
727 struct blk_mq_ctx *ctx;
730 for (i = 0; i < hctx->ctx_map.map_size; i++) {
731 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
732 unsigned int off, bit;
738 off = i * hctx->ctx_map.bits_per_word;
740 bit = find_next_bit(&bm->word, bm->depth, bit);
741 if (bit >= bm->depth)
744 ctx = hctx->ctxs[bit + off];
745 clear_bit(bit, &bm->word);
746 spin_lock(&ctx->lock);
747 list_splice_tail_init(&ctx->rq_list, list);
748 spin_unlock(&ctx->lock);
756 * Run this hardware queue, pulling any software queues mapped to it in.
757 * Note that this function currently has various problems around ordering
758 * of IO. In particular, we'd like FIFO behaviour on handling existing
759 * items on the hctx->dispatch list. Ignore that for now.
761 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
763 struct request_queue *q = hctx->queue;
766 LIST_HEAD(driver_list);
767 struct list_head *dptr;
770 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
772 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
778 * Touch any software queue that has pending entries.
780 flush_busy_ctxs(hctx, &rq_list);
783 * If we have previous entries on our dispatch list, grab them
784 * and stuff them at the front for more fair dispatch.
786 if (!list_empty_careful(&hctx->dispatch)) {
787 spin_lock(&hctx->lock);
788 if (!list_empty(&hctx->dispatch))
789 list_splice_init(&hctx->dispatch, &rq_list);
790 spin_unlock(&hctx->lock);
794 * Start off with dptr being NULL, so we start the first request
795 * immediately, even if we have more pending.
800 * Now process all the entries, sending them to the driver.
803 while (!list_empty(&rq_list)) {
804 struct blk_mq_queue_data bd;
807 rq = list_first_entry(&rq_list, struct request, queuelist);
808 list_del_init(&rq->queuelist);
812 bd.last = list_empty(&rq_list);
814 ret = q->mq_ops->queue_rq(hctx, &bd);
816 case BLK_MQ_RQ_QUEUE_OK:
819 case BLK_MQ_RQ_QUEUE_BUSY:
820 list_add(&rq->queuelist, &rq_list);
821 __blk_mq_requeue_request(rq);
824 pr_err("blk-mq: bad return on queue: %d\n", ret);
825 case BLK_MQ_RQ_QUEUE_ERROR:
827 blk_mq_end_request(rq, rq->errors);
831 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
835 * We've done the first request. If we have more than 1
836 * left in the list, set dptr to defer issue.
838 if (!dptr && rq_list.next != rq_list.prev)
843 hctx->dispatched[0]++;
844 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
845 hctx->dispatched[ilog2(queued) + 1]++;
848 * Any items that need requeuing? Stuff them into hctx->dispatch,
849 * that is where we will continue on next queue run.
851 if (!list_empty(&rq_list)) {
852 spin_lock(&hctx->lock);
853 list_splice(&rq_list, &hctx->dispatch);
854 spin_unlock(&hctx->lock);
859 * It'd be great if the workqueue API had a way to pass
860 * in a mask and had some smarts for more clever placement.
861 * For now we just round-robin here, switching for every
862 * BLK_MQ_CPU_WORK_BATCH queued items.
864 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
866 if (hctx->queue->nr_hw_queues == 1)
867 return WORK_CPU_UNBOUND;
869 if (--hctx->next_cpu_batch <= 0) {
870 int cpu = hctx->next_cpu, next_cpu;
872 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
873 if (next_cpu >= nr_cpu_ids)
874 next_cpu = cpumask_first(hctx->cpumask);
876 hctx->next_cpu = next_cpu;
877 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
882 return hctx->next_cpu;
885 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
887 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
888 !blk_mq_hw_queue_mapped(hctx)))
893 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
894 __blk_mq_run_hw_queue(hctx);
902 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
906 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
908 struct blk_mq_hw_ctx *hctx;
911 queue_for_each_hw_ctx(q, hctx, i) {
912 if ((!blk_mq_hctx_has_pending(hctx) &&
913 list_empty_careful(&hctx->dispatch)) ||
914 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
917 blk_mq_run_hw_queue(hctx, async);
920 EXPORT_SYMBOL(blk_mq_run_hw_queues);
922 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
924 cancel_delayed_work(&hctx->run_work);
925 cancel_delayed_work(&hctx->delay_work);
926 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
928 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
930 void blk_mq_stop_hw_queues(struct request_queue *q)
932 struct blk_mq_hw_ctx *hctx;
935 queue_for_each_hw_ctx(q, hctx, i)
936 blk_mq_stop_hw_queue(hctx);
938 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
940 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
942 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
944 blk_mq_run_hw_queue(hctx, false);
946 EXPORT_SYMBOL(blk_mq_start_hw_queue);
948 void blk_mq_start_hw_queues(struct request_queue *q)
950 struct blk_mq_hw_ctx *hctx;
953 queue_for_each_hw_ctx(q, hctx, i)
954 blk_mq_start_hw_queue(hctx);
956 EXPORT_SYMBOL(blk_mq_start_hw_queues);
958 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
960 struct blk_mq_hw_ctx *hctx;
963 queue_for_each_hw_ctx(q, hctx, i) {
964 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
967 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
968 blk_mq_run_hw_queue(hctx, async);
971 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
973 static void blk_mq_run_work_fn(struct work_struct *work)
975 struct blk_mq_hw_ctx *hctx;
977 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
979 __blk_mq_run_hw_queue(hctx);
982 static void blk_mq_delay_work_fn(struct work_struct *work)
984 struct blk_mq_hw_ctx *hctx;
986 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
988 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
989 __blk_mq_run_hw_queue(hctx);
992 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
994 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
997 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
998 &hctx->delay_work, msecs_to_jiffies(msecs));
1000 EXPORT_SYMBOL(blk_mq_delay_queue);
1002 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
1003 struct request *rq, bool at_head)
1005 struct blk_mq_ctx *ctx = rq->mq_ctx;
1007 trace_block_rq_insert(hctx->queue, rq);
1010 list_add(&rq->queuelist, &ctx->rq_list);
1012 list_add_tail(&rq->queuelist, &ctx->rq_list);
1014 blk_mq_hctx_mark_pending(hctx, ctx);
1017 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1020 struct request_queue *q = rq->q;
1021 struct blk_mq_hw_ctx *hctx;
1022 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
1024 current_ctx = blk_mq_get_ctx(q);
1025 if (!cpu_online(ctx->cpu))
1026 rq->mq_ctx = ctx = current_ctx;
1028 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1030 spin_lock(&ctx->lock);
1031 __blk_mq_insert_request(hctx, rq, at_head);
1032 spin_unlock(&ctx->lock);
1035 blk_mq_run_hw_queue(hctx, async);
1037 blk_mq_put_ctx(current_ctx);
1040 static void blk_mq_insert_requests(struct request_queue *q,
1041 struct blk_mq_ctx *ctx,
1042 struct list_head *list,
1047 struct blk_mq_hw_ctx *hctx;
1048 struct blk_mq_ctx *current_ctx;
1050 trace_block_unplug(q, depth, !from_schedule);
1052 current_ctx = blk_mq_get_ctx(q);
1054 if (!cpu_online(ctx->cpu))
1056 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1059 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1062 spin_lock(&ctx->lock);
1063 while (!list_empty(list)) {
1066 rq = list_first_entry(list, struct request, queuelist);
1067 list_del_init(&rq->queuelist);
1069 __blk_mq_insert_request(hctx, rq, false);
1071 spin_unlock(&ctx->lock);
1073 blk_mq_run_hw_queue(hctx, from_schedule);
1074 blk_mq_put_ctx(current_ctx);
1077 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1079 struct request *rqa = container_of(a, struct request, queuelist);
1080 struct request *rqb = container_of(b, struct request, queuelist);
1082 return !(rqa->mq_ctx < rqb->mq_ctx ||
1083 (rqa->mq_ctx == rqb->mq_ctx &&
1084 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1087 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1089 struct blk_mq_ctx *this_ctx;
1090 struct request_queue *this_q;
1093 LIST_HEAD(ctx_list);
1096 list_splice_init(&plug->mq_list, &list);
1098 list_sort(NULL, &list, plug_ctx_cmp);
1104 while (!list_empty(&list)) {
1105 rq = list_entry_rq(list.next);
1106 list_del_init(&rq->queuelist);
1108 if (rq->mq_ctx != this_ctx) {
1110 blk_mq_insert_requests(this_q, this_ctx,
1115 this_ctx = rq->mq_ctx;
1121 list_add_tail(&rq->queuelist, &ctx_list);
1125 * If 'this_ctx' is set, we know we have entries to complete
1126 * on 'ctx_list'. Do those.
1129 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1134 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1136 init_request_from_bio(rq, bio);
1138 if (blk_do_io_stat(rq))
1139 blk_account_io_start(rq, 1);
1142 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1144 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1145 !blk_queue_nomerges(hctx->queue);
1148 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1149 struct blk_mq_ctx *ctx,
1150 struct request *rq, struct bio *bio)
1152 if (!hctx_allow_merges(hctx)) {
1153 blk_mq_bio_to_request(rq, bio);
1154 spin_lock(&ctx->lock);
1156 __blk_mq_insert_request(hctx, rq, false);
1157 spin_unlock(&ctx->lock);
1160 struct request_queue *q = hctx->queue;
1162 spin_lock(&ctx->lock);
1163 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1164 blk_mq_bio_to_request(rq, bio);
1168 spin_unlock(&ctx->lock);
1169 __blk_mq_free_request(hctx, ctx, rq);
1174 struct blk_map_ctx {
1175 struct blk_mq_hw_ctx *hctx;
1176 struct blk_mq_ctx *ctx;
1179 static struct request *blk_mq_map_request(struct request_queue *q,
1181 struct blk_map_ctx *data)
1183 struct blk_mq_hw_ctx *hctx;
1184 struct blk_mq_ctx *ctx;
1186 int rw = bio_data_dir(bio);
1187 struct blk_mq_alloc_data alloc_data;
1189 if (unlikely(blk_mq_queue_enter(q))) {
1190 bio_endio(bio, -EIO);
1194 ctx = blk_mq_get_ctx(q);
1195 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1197 if (rw_is_sync(bio->bi_rw))
1200 trace_block_getrq(q, bio, rw);
1201 blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1203 rq = __blk_mq_alloc_request(&alloc_data, rw);
1204 if (unlikely(!rq)) {
1205 __blk_mq_run_hw_queue(hctx);
1206 blk_mq_put_ctx(ctx);
1207 trace_block_sleeprq(q, bio, rw);
1209 ctx = blk_mq_get_ctx(q);
1210 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1211 blk_mq_set_alloc_data(&alloc_data, q,
1212 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1213 rq = __blk_mq_alloc_request(&alloc_data, rw);
1214 ctx = alloc_data.ctx;
1215 hctx = alloc_data.hctx;
1225 * Multiple hardware queue variant. This will not use per-process plugs,
1226 * but will attempt to bypass the hctx queueing if we can go straight to
1227 * hardware for SYNC IO.
1229 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1231 const int is_sync = rw_is_sync(bio->bi_rw);
1232 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1233 struct blk_map_ctx data;
1236 blk_queue_bounce(q, &bio);
1238 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1239 bio_endio(bio, -EIO);
1243 rq = blk_mq_map_request(q, bio, &data);
1247 if (unlikely(is_flush_fua)) {
1248 blk_mq_bio_to_request(rq, bio);
1249 blk_insert_flush(rq);
1254 * If the driver supports defer issued based on 'last', then
1255 * queue it up like normal since we can potentially save some
1258 if (is_sync && !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1259 struct blk_mq_queue_data bd = {
1266 blk_mq_bio_to_request(rq, bio);
1269 * For OK queue, we are done. For error, kill it. Any other
1270 * error (busy), just add it to our list as we previously
1273 ret = q->mq_ops->queue_rq(data.hctx, &bd);
1274 if (ret == BLK_MQ_RQ_QUEUE_OK)
1277 __blk_mq_requeue_request(rq);
1279 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1281 blk_mq_end_request(rq, rq->errors);
1287 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1289 * For a SYNC request, send it to the hardware immediately. For
1290 * an ASYNC request, just ensure that we run it later on. The
1291 * latter allows for merging opportunities and more efficient
1295 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1298 blk_mq_put_ctx(data.ctx);
1302 * Single hardware queue variant. This will attempt to use any per-process
1303 * plug for merging and IO deferral.
1305 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1307 const int is_sync = rw_is_sync(bio->bi_rw);
1308 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1309 unsigned int use_plug, request_count = 0;
1310 struct blk_map_ctx data;
1314 * If we have multiple hardware queues, just go directly to
1315 * one of those for sync IO.
1317 use_plug = !is_flush_fua && !is_sync;
1319 blk_queue_bounce(q, &bio);
1321 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1322 bio_endio(bio, -EIO);
1326 if (use_plug && !blk_queue_nomerges(q) &&
1327 blk_attempt_plug_merge(q, bio, &request_count))
1330 rq = blk_mq_map_request(q, bio, &data);
1334 if (unlikely(is_flush_fua)) {
1335 blk_mq_bio_to_request(rq, bio);
1336 blk_insert_flush(rq);
1341 * A task plug currently exists. Since this is completely lockless,
1342 * utilize that to temporarily store requests until the task is
1343 * either done or scheduled away.
1346 struct blk_plug *plug = current->plug;
1349 blk_mq_bio_to_request(rq, bio);
1350 if (list_empty(&plug->mq_list))
1351 trace_block_plug(q);
1352 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1353 blk_flush_plug_list(plug, false);
1354 trace_block_plug(q);
1356 list_add_tail(&rq->queuelist, &plug->mq_list);
1357 blk_mq_put_ctx(data.ctx);
1362 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1364 * For a SYNC request, send it to the hardware immediately. For
1365 * an ASYNC request, just ensure that we run it later on. The
1366 * latter allows for merging opportunities and more efficient
1370 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1373 blk_mq_put_ctx(data.ctx);
1377 * Default mapping to a software queue, since we use one per CPU.
1379 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1381 return q->queue_hw_ctx[q->mq_map[cpu]];
1383 EXPORT_SYMBOL(blk_mq_map_queue);
1385 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1386 struct blk_mq_tags *tags, unsigned int hctx_idx)
1390 if (tags->rqs && set->ops->exit_request) {
1393 for (i = 0; i < tags->nr_tags; i++) {
1396 set->ops->exit_request(set->driver_data, tags->rqs[i],
1398 tags->rqs[i] = NULL;
1402 while (!list_empty(&tags->page_list)) {
1403 page = list_first_entry(&tags->page_list, struct page, lru);
1404 list_del_init(&page->lru);
1405 __free_pages(page, page->private);
1410 blk_mq_free_tags(tags);
1413 static size_t order_to_size(unsigned int order)
1415 return (size_t)PAGE_SIZE << order;
1418 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1419 unsigned int hctx_idx)
1421 struct blk_mq_tags *tags;
1422 unsigned int i, j, entries_per_page, max_order = 4;
1423 size_t rq_size, left;
1425 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1427 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1431 INIT_LIST_HEAD(&tags->page_list);
1433 tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1434 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1437 blk_mq_free_tags(tags);
1442 * rq_size is the size of the request plus driver payload, rounded
1443 * to the cacheline size
1445 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1447 left = rq_size * set->queue_depth;
1449 for (i = 0; i < set->queue_depth; ) {
1450 int this_order = max_order;
1455 while (left < order_to_size(this_order - 1) && this_order)
1459 page = alloc_pages_node(set->numa_node,
1460 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1466 if (order_to_size(this_order) < rq_size)
1473 page->private = this_order;
1474 list_add_tail(&page->lru, &tags->page_list);
1476 p = page_address(page);
1477 entries_per_page = order_to_size(this_order) / rq_size;
1478 to_do = min(entries_per_page, set->queue_depth - i);
1479 left -= to_do * rq_size;
1480 for (j = 0; j < to_do; j++) {
1482 tags->rqs[i]->atomic_flags = 0;
1483 tags->rqs[i]->cmd_flags = 0;
1484 if (set->ops->init_request) {
1485 if (set->ops->init_request(set->driver_data,
1486 tags->rqs[i], hctx_idx, i,
1488 tags->rqs[i] = NULL;
1501 blk_mq_free_rq_map(set, tags, hctx_idx);
1505 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1510 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1512 unsigned int bpw = 8, total, num_maps, i;
1514 bitmap->bits_per_word = bpw;
1516 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1517 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1522 bitmap->map_size = num_maps;
1525 for (i = 0; i < num_maps; i++) {
1526 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1527 total -= bitmap->map[i].depth;
1533 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1535 struct request_queue *q = hctx->queue;
1536 struct blk_mq_ctx *ctx;
1540 * Move ctx entries to new CPU, if this one is going away.
1542 ctx = __blk_mq_get_ctx(q, cpu);
1544 spin_lock(&ctx->lock);
1545 if (!list_empty(&ctx->rq_list)) {
1546 list_splice_init(&ctx->rq_list, &tmp);
1547 blk_mq_hctx_clear_pending(hctx, ctx);
1549 spin_unlock(&ctx->lock);
1551 if (list_empty(&tmp))
1554 ctx = blk_mq_get_ctx(q);
1555 spin_lock(&ctx->lock);
1557 while (!list_empty(&tmp)) {
1560 rq = list_first_entry(&tmp, struct request, queuelist);
1562 list_move_tail(&rq->queuelist, &ctx->rq_list);
1565 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1566 blk_mq_hctx_mark_pending(hctx, ctx);
1568 spin_unlock(&ctx->lock);
1570 blk_mq_run_hw_queue(hctx, true);
1571 blk_mq_put_ctx(ctx);
1575 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1577 struct request_queue *q = hctx->queue;
1578 struct blk_mq_tag_set *set = q->tag_set;
1580 if (set->tags[hctx->queue_num])
1583 set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1584 if (!set->tags[hctx->queue_num])
1587 hctx->tags = set->tags[hctx->queue_num];
1591 static int blk_mq_hctx_notify(void *data, unsigned long action,
1594 struct blk_mq_hw_ctx *hctx = data;
1596 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1597 return blk_mq_hctx_cpu_offline(hctx, cpu);
1598 else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1599 return blk_mq_hctx_cpu_online(hctx, cpu);
1604 static void blk_mq_exit_hctx(struct request_queue *q,
1605 struct blk_mq_tag_set *set,
1606 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1608 unsigned flush_start_tag = set->queue_depth;
1610 blk_mq_tag_idle(hctx);
1612 if (set->ops->exit_request)
1613 set->ops->exit_request(set->driver_data,
1614 hctx->fq->flush_rq, hctx_idx,
1615 flush_start_tag + hctx_idx);
1617 if (set->ops->exit_hctx)
1618 set->ops->exit_hctx(hctx, hctx_idx);
1620 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1621 blk_free_flush_queue(hctx->fq);
1623 blk_mq_free_bitmap(&hctx->ctx_map);
1626 static void blk_mq_exit_hw_queues(struct request_queue *q,
1627 struct blk_mq_tag_set *set, int nr_queue)
1629 struct blk_mq_hw_ctx *hctx;
1632 queue_for_each_hw_ctx(q, hctx, i) {
1635 blk_mq_exit_hctx(q, set, hctx, i);
1639 static void blk_mq_free_hw_queues(struct request_queue *q,
1640 struct blk_mq_tag_set *set)
1642 struct blk_mq_hw_ctx *hctx;
1645 queue_for_each_hw_ctx(q, hctx, i)
1646 free_cpumask_var(hctx->cpumask);
1649 static int blk_mq_init_hctx(struct request_queue *q,
1650 struct blk_mq_tag_set *set,
1651 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1654 unsigned flush_start_tag = set->queue_depth;
1656 node = hctx->numa_node;
1657 if (node == NUMA_NO_NODE)
1658 node = hctx->numa_node = set->numa_node;
1660 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1661 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1662 spin_lock_init(&hctx->lock);
1663 INIT_LIST_HEAD(&hctx->dispatch);
1665 hctx->queue_num = hctx_idx;
1666 hctx->flags = set->flags;
1668 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1669 blk_mq_hctx_notify, hctx);
1670 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1672 hctx->tags = set->tags[hctx_idx];
1675 * Allocate space for all possible cpus to avoid allocation at
1678 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1681 goto unregister_cpu_notifier;
1683 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1688 if (set->ops->init_hctx &&
1689 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1692 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1696 if (set->ops->init_request &&
1697 set->ops->init_request(set->driver_data,
1698 hctx->fq->flush_rq, hctx_idx,
1699 flush_start_tag + hctx_idx, node))
1707 if (set->ops->exit_hctx)
1708 set->ops->exit_hctx(hctx, hctx_idx);
1710 blk_mq_free_bitmap(&hctx->ctx_map);
1713 unregister_cpu_notifier:
1714 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1719 static int blk_mq_init_hw_queues(struct request_queue *q,
1720 struct blk_mq_tag_set *set)
1722 struct blk_mq_hw_ctx *hctx;
1726 * Initialize hardware queues
1728 queue_for_each_hw_ctx(q, hctx, i) {
1729 if (blk_mq_init_hctx(q, set, hctx, i))
1733 if (i == q->nr_hw_queues)
1739 blk_mq_exit_hw_queues(q, set, i);
1744 static void blk_mq_init_cpu_queues(struct request_queue *q,
1745 unsigned int nr_hw_queues)
1749 for_each_possible_cpu(i) {
1750 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1751 struct blk_mq_hw_ctx *hctx;
1753 memset(__ctx, 0, sizeof(*__ctx));
1755 spin_lock_init(&__ctx->lock);
1756 INIT_LIST_HEAD(&__ctx->rq_list);
1759 /* If the cpu isn't online, the cpu is mapped to first hctx */
1763 hctx = q->mq_ops->map_queue(q, i);
1764 cpumask_set_cpu(i, hctx->cpumask);
1768 * Set local node, IFF we have more than one hw queue. If
1769 * not, we remain on the home node of the device
1771 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1772 hctx->numa_node = cpu_to_node(i);
1776 static void blk_mq_map_swqueue(struct request_queue *q)
1779 struct blk_mq_hw_ctx *hctx;
1780 struct blk_mq_ctx *ctx;
1782 queue_for_each_hw_ctx(q, hctx, i) {
1783 cpumask_clear(hctx->cpumask);
1788 * Map software to hardware queues
1790 queue_for_each_ctx(q, ctx, i) {
1791 /* If the cpu isn't online, the cpu is mapped to first hctx */
1795 hctx = q->mq_ops->map_queue(q, i);
1796 cpumask_set_cpu(i, hctx->cpumask);
1797 ctx->index_hw = hctx->nr_ctx;
1798 hctx->ctxs[hctx->nr_ctx++] = ctx;
1801 queue_for_each_hw_ctx(q, hctx, i) {
1803 * If no software queues are mapped to this hardware queue,
1804 * disable it and free the request entries.
1806 if (!hctx->nr_ctx) {
1807 struct blk_mq_tag_set *set = q->tag_set;
1810 blk_mq_free_rq_map(set, set->tags[i], i);
1811 set->tags[i] = NULL;
1818 * Initialize batch roundrobin counts
1820 hctx->next_cpu = cpumask_first(hctx->cpumask);
1821 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1825 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1827 struct blk_mq_hw_ctx *hctx;
1828 struct request_queue *q;
1832 if (set->tag_list.next == set->tag_list.prev)
1837 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1838 blk_mq_freeze_queue(q);
1840 queue_for_each_hw_ctx(q, hctx, i) {
1842 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1844 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1846 blk_mq_unfreeze_queue(q);
1850 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1852 struct blk_mq_tag_set *set = q->tag_set;
1854 mutex_lock(&set->tag_list_lock);
1855 list_del_init(&q->tag_set_list);
1856 blk_mq_update_tag_set_depth(set);
1857 mutex_unlock(&set->tag_list_lock);
1860 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1861 struct request_queue *q)
1865 mutex_lock(&set->tag_list_lock);
1866 list_add_tail(&q->tag_set_list, &set->tag_list);
1867 blk_mq_update_tag_set_depth(set);
1868 mutex_unlock(&set->tag_list_lock);
1872 * It is the actual release handler for mq, but we do it from
1873 * request queue's release handler for avoiding use-after-free
1874 * and headache because q->mq_kobj shouldn't have been introduced,
1875 * but we can't group ctx/kctx kobj without it.
1877 void blk_mq_release(struct request_queue *q)
1879 struct blk_mq_hw_ctx *hctx;
1882 /* hctx kobj stays in hctx */
1883 queue_for_each_hw_ctx(q, hctx, i)
1886 kfree(q->queue_hw_ctx);
1888 /* ctx kobj stays in queue_ctx */
1889 free_percpu(q->queue_ctx);
1892 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1894 struct request_queue *uninit_q, *q;
1896 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1898 return ERR_PTR(-ENOMEM);
1900 q = blk_mq_init_allocated_queue(set, uninit_q);
1902 blk_cleanup_queue(uninit_q);
1906 EXPORT_SYMBOL(blk_mq_init_queue);
1908 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
1909 struct request_queue *q)
1911 struct blk_mq_hw_ctx **hctxs;
1912 struct blk_mq_ctx __percpu *ctx;
1916 ctx = alloc_percpu(struct blk_mq_ctx);
1918 return ERR_PTR(-ENOMEM);
1920 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1926 map = blk_mq_make_queue_map(set);
1930 for (i = 0; i < set->nr_hw_queues; i++) {
1931 int node = blk_mq_hw_queue_to_node(map, i);
1933 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1938 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
1942 atomic_set(&hctxs[i]->nr_active, 0);
1943 hctxs[i]->numa_node = node;
1944 hctxs[i]->queue_num = i;
1948 * Init percpu_ref in atomic mode so that it's faster to shutdown.
1949 * See blk_register_queue() for details.
1951 if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release,
1952 PERCPU_REF_INIT_ATOMIC, GFP_KERNEL))
1955 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1956 blk_queue_rq_timeout(q, 30000);
1958 q->nr_queues = nr_cpu_ids;
1959 q->nr_hw_queues = set->nr_hw_queues;
1963 q->queue_hw_ctx = hctxs;
1965 q->mq_ops = set->ops;
1966 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1968 if (!(set->flags & BLK_MQ_F_SG_MERGE))
1969 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
1971 q->sg_reserved_size = INT_MAX;
1973 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1974 INIT_LIST_HEAD(&q->requeue_list);
1975 spin_lock_init(&q->requeue_lock);
1977 if (q->nr_hw_queues > 1)
1978 blk_queue_make_request(q, blk_mq_make_request);
1980 blk_queue_make_request(q, blk_sq_make_request);
1983 blk_queue_rq_timeout(q, set->timeout);
1986 * Do this after blk_queue_make_request() overrides it...
1988 q->nr_requests = set->queue_depth;
1990 if (set->ops->complete)
1991 blk_queue_softirq_done(q, set->ops->complete);
1993 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1995 if (blk_mq_init_hw_queues(q, set))
1998 mutex_lock(&all_q_mutex);
1999 list_add_tail(&q->all_q_node, &all_q_list);
2000 mutex_unlock(&all_q_mutex);
2002 blk_mq_add_queue_tag_set(set, q);
2004 blk_mq_map_swqueue(q);
2010 for (i = 0; i < set->nr_hw_queues; i++) {
2013 free_cpumask_var(hctxs[i]->cpumask);
2020 return ERR_PTR(-ENOMEM);
2022 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2024 void blk_mq_free_queue(struct request_queue *q)
2026 struct blk_mq_tag_set *set = q->tag_set;
2028 blk_mq_del_queue_tag_set(q);
2030 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2031 blk_mq_free_hw_queues(q, set);
2033 percpu_ref_exit(&q->mq_usage_counter);
2039 mutex_lock(&all_q_mutex);
2040 list_del_init(&q->all_q_node);
2041 mutex_unlock(&all_q_mutex);
2044 /* Basically redo blk_mq_init_queue with queue frozen */
2045 static void blk_mq_queue_reinit(struct request_queue *q)
2047 WARN_ON_ONCE(!q->mq_freeze_depth);
2049 blk_mq_sysfs_unregister(q);
2051 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
2054 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2055 * we should change hctx numa_node according to new topology (this
2056 * involves free and re-allocate memory, worthy doing?)
2059 blk_mq_map_swqueue(q);
2061 blk_mq_sysfs_register(q);
2064 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2065 unsigned long action, void *hcpu)
2067 struct request_queue *q;
2070 * Before new mappings are established, hotadded cpu might already
2071 * start handling requests. This doesn't break anything as we map
2072 * offline CPUs to first hardware queue. We will re-init the queue
2073 * below to get optimal settings.
2075 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
2076 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
2079 mutex_lock(&all_q_mutex);
2082 * We need to freeze and reinit all existing queues. Freezing
2083 * involves synchronous wait for an RCU grace period and doing it
2084 * one by one may take a long time. Start freezing all queues in
2085 * one swoop and then wait for the completions so that freezing can
2086 * take place in parallel.
2088 list_for_each_entry(q, &all_q_list, all_q_node)
2089 blk_mq_freeze_queue_start(q);
2090 list_for_each_entry(q, &all_q_list, all_q_node)
2091 blk_mq_freeze_queue_wait(q);
2093 list_for_each_entry(q, &all_q_list, all_q_node)
2094 blk_mq_queue_reinit(q);
2096 list_for_each_entry(q, &all_q_list, all_q_node)
2097 blk_mq_unfreeze_queue(q);
2099 mutex_unlock(&all_q_mutex);
2103 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2107 for (i = 0; i < set->nr_hw_queues; i++) {
2108 set->tags[i] = blk_mq_init_rq_map(set, i);
2117 blk_mq_free_rq_map(set, set->tags[i], i);
2123 * Allocate the request maps associated with this tag_set. Note that this
2124 * may reduce the depth asked for, if memory is tight. set->queue_depth
2125 * will be updated to reflect the allocated depth.
2127 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2132 depth = set->queue_depth;
2134 err = __blk_mq_alloc_rq_maps(set);
2138 set->queue_depth >>= 1;
2139 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2143 } while (set->queue_depth);
2145 if (!set->queue_depth || err) {
2146 pr_err("blk-mq: failed to allocate request map\n");
2150 if (depth != set->queue_depth)
2151 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2152 depth, set->queue_depth);
2158 * Alloc a tag set to be associated with one or more request queues.
2159 * May fail with EINVAL for various error conditions. May adjust the
2160 * requested depth down, if if it too large. In that case, the set
2161 * value will be stored in set->queue_depth.
2163 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2165 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2167 if (!set->nr_hw_queues)
2169 if (!set->queue_depth)
2171 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2174 if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue)
2177 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2178 pr_info("blk-mq: reduced tag depth to %u\n",
2180 set->queue_depth = BLK_MQ_MAX_DEPTH;
2184 * If a crashdump is active, then we are potentially in a very
2185 * memory constrained environment. Limit us to 1 queue and
2186 * 64 tags to prevent using too much memory.
2188 if (is_kdump_kernel()) {
2189 set->nr_hw_queues = 1;
2190 set->queue_depth = min(64U, set->queue_depth);
2193 set->tags = kmalloc_node(set->nr_hw_queues *
2194 sizeof(struct blk_mq_tags *),
2195 GFP_KERNEL, set->numa_node);
2199 if (blk_mq_alloc_rq_maps(set))
2202 mutex_init(&set->tag_list_lock);
2203 INIT_LIST_HEAD(&set->tag_list);
2211 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2213 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2217 for (i = 0; i < set->nr_hw_queues; i++) {
2219 blk_mq_free_rq_map(set, set->tags[i], i);
2225 EXPORT_SYMBOL(blk_mq_free_tag_set);
2227 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2229 struct blk_mq_tag_set *set = q->tag_set;
2230 struct blk_mq_hw_ctx *hctx;
2233 if (!set || nr > set->queue_depth)
2237 queue_for_each_hw_ctx(q, hctx, i) {
2238 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2244 q->nr_requests = nr;
2249 void blk_mq_disable_hotplug(void)
2251 mutex_lock(&all_q_mutex);
2254 void blk_mq_enable_hotplug(void)
2256 mutex_unlock(&all_q_mutex);
2259 static int __init blk_mq_init(void)
2263 hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2267 subsys_initcall(blk_mq_init);