2 * Block multiqueue core code
4 * Copyright (C) 2013-2014 Jens Axboe
5 * Copyright (C) 2013-2014 Christoph Hellwig
7 #include <linux/kernel.h>
8 #include <linux/module.h>
9 #include <linux/backing-dev.h>
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
12 #include <linux/kmemleak.h>
14 #include <linux/init.h>
15 #include <linux/slab.h>
16 #include <linux/workqueue.h>
17 #include <linux/smp.h>
18 #include <linux/llist.h>
19 #include <linux/list_sort.h>
20 #include <linux/cpu.h>
21 #include <linux/cache.h>
22 #include <linux/sched/sysctl.h>
23 #include <linux/delay.h>
24 #include <linux/crash_dump.h>
26 #include <trace/events/block.h>
28 #include <linux/blk-mq.h>
31 #include "blk-mq-tag.h"
33 static DEFINE_MUTEX(all_q_mutex);
34 static LIST_HEAD(all_q_list);
36 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
39 * Check if any of the ctx's have pending work in this hardware queue
41 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
45 for (i = 0; i < hctx->ctx_map.size; i++)
46 if (hctx->ctx_map.map[i].word)
52 static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx,
53 struct blk_mq_ctx *ctx)
55 return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word];
58 #define CTX_TO_BIT(hctx, ctx) \
59 ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
62 * Mark this ctx as having pending work in this hardware queue
64 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
65 struct blk_mq_ctx *ctx)
67 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
69 if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word))
70 set_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
73 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
74 struct blk_mq_ctx *ctx)
76 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
78 clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
81 void blk_mq_freeze_queue_start(struct request_queue *q)
85 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
86 if (freeze_depth == 1) {
87 percpu_ref_kill(&q->q_usage_counter);
88 blk_mq_run_hw_queues(q, false);
91 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
93 static void blk_mq_freeze_queue_wait(struct request_queue *q)
95 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
99 * Guarantee no request is in use, so we can change any data structure of
100 * the queue afterward.
102 void blk_freeze_queue(struct request_queue *q)
105 * In the !blk_mq case we are only calling this to kill the
106 * q_usage_counter, otherwise this increases the freeze depth
107 * and waits for it to return to zero. For this reason there is
108 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
109 * exported to drivers as the only user for unfreeze is blk_mq.
111 blk_mq_freeze_queue_start(q);
112 blk_mq_freeze_queue_wait(q);
115 void blk_mq_freeze_queue(struct request_queue *q)
118 * ...just an alias to keep freeze and unfreeze actions balanced
119 * in the blk_mq_* namespace
123 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
125 void blk_mq_unfreeze_queue(struct request_queue *q)
129 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
130 WARN_ON_ONCE(freeze_depth < 0);
132 percpu_ref_reinit(&q->q_usage_counter);
133 wake_up_all(&q->mq_freeze_wq);
136 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
138 void blk_mq_wake_waiters(struct request_queue *q)
140 struct blk_mq_hw_ctx *hctx;
143 queue_for_each_hw_ctx(q, hctx, i)
144 if (blk_mq_hw_queue_mapped(hctx))
145 blk_mq_tag_wakeup_all(hctx->tags, true);
148 * If we are called because the queue has now been marked as
149 * dying, we need to ensure that processes currently waiting on
150 * the queue are notified as well.
152 wake_up_all(&q->mq_freeze_wq);
155 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
157 return blk_mq_has_free_tags(hctx->tags);
159 EXPORT_SYMBOL(blk_mq_can_queue);
161 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
162 struct request *rq, int op,
163 unsigned int op_flags)
165 if (blk_queue_io_stat(q))
166 op_flags |= REQ_IO_STAT;
168 INIT_LIST_HEAD(&rq->queuelist);
169 /* csd/requeue_work/fifo_time is initialized before use */
172 req_set_op_attrs(rq, op, op_flags);
173 /* do not touch atomic flags, it needs atomic ops against the timer */
175 INIT_HLIST_NODE(&rq->hash);
176 RB_CLEAR_NODE(&rq->rb_node);
179 rq->start_time = jiffies;
180 #ifdef CONFIG_BLK_CGROUP
182 set_start_time_ns(rq);
183 rq->io_start_time_ns = 0;
185 rq->nr_phys_segments = 0;
186 #if defined(CONFIG_BLK_DEV_INTEGRITY)
187 rq->nr_integrity_segments = 0;
190 /* tag was already set */
200 INIT_LIST_HEAD(&rq->timeout_list);
204 rq->end_io_data = NULL;
207 ctx->rq_dispatched[rw_is_sync(op, op_flags)]++;
210 static struct request *
211 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int op, int op_flags)
216 tag = blk_mq_get_tag(data);
217 if (tag != BLK_MQ_TAG_FAIL) {
218 rq = data->hctx->tags->rqs[tag];
220 if (blk_mq_tag_busy(data->hctx)) {
221 rq->cmd_flags = REQ_MQ_INFLIGHT;
222 atomic_inc(&data->hctx->nr_active);
226 blk_mq_rq_ctx_init(data->q, data->ctx, rq, op, op_flags);
233 struct request *blk_mq_alloc_request(struct request_queue *q, int rw,
236 struct blk_mq_ctx *ctx;
237 struct blk_mq_hw_ctx *hctx;
239 struct blk_mq_alloc_data alloc_data;
242 ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
246 ctx = blk_mq_get_ctx(q);
247 hctx = q->mq_ops->map_queue(q, ctx->cpu);
248 blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx);
250 rq = __blk_mq_alloc_request(&alloc_data, rw, 0);
251 if (!rq && !(flags & BLK_MQ_REQ_NOWAIT)) {
252 __blk_mq_run_hw_queue(hctx);
255 ctx = blk_mq_get_ctx(q);
256 hctx = q->mq_ops->map_queue(q, ctx->cpu);
257 blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx);
258 rq = __blk_mq_alloc_request(&alloc_data, rw, 0);
259 ctx = alloc_data.ctx;
264 return ERR_PTR(-EWOULDBLOCK);
268 EXPORT_SYMBOL(blk_mq_alloc_request);
270 struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw,
271 unsigned int flags, unsigned int hctx_idx)
273 struct blk_mq_hw_ctx *hctx;
274 struct blk_mq_ctx *ctx;
276 struct blk_mq_alloc_data alloc_data;
280 * If the tag allocator sleeps we could get an allocation for a
281 * different hardware context. No need to complicate the low level
282 * allocator for this for the rare use case of a command tied to
285 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
286 return ERR_PTR(-EINVAL);
288 if (hctx_idx >= q->nr_hw_queues)
289 return ERR_PTR(-EIO);
291 ret = blk_queue_enter(q, true);
295 hctx = q->queue_hw_ctx[hctx_idx];
296 ctx = __blk_mq_get_ctx(q, cpumask_first(hctx->cpumask));
298 blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx);
299 rq = __blk_mq_alloc_request(&alloc_data, rw, 0);
302 return ERR_PTR(-EWOULDBLOCK);
307 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
309 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
310 struct blk_mq_ctx *ctx, struct request *rq)
312 const int tag = rq->tag;
313 struct request_queue *q = rq->q;
315 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
316 atomic_dec(&hctx->nr_active);
319 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
320 blk_mq_put_tag(hctx, tag, &ctx->last_tag);
324 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
326 struct blk_mq_ctx *ctx = rq->mq_ctx;
328 ctx->rq_completed[rq_is_sync(rq)]++;
329 __blk_mq_free_request(hctx, ctx, rq);
332 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
334 void blk_mq_free_request(struct request *rq)
336 struct blk_mq_hw_ctx *hctx;
337 struct request_queue *q = rq->q;
339 hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
340 blk_mq_free_hctx_request(hctx, rq);
342 EXPORT_SYMBOL_GPL(blk_mq_free_request);
344 inline void __blk_mq_end_request(struct request *rq, int error)
346 blk_account_io_done(rq);
349 rq->end_io(rq, error);
351 if (unlikely(blk_bidi_rq(rq)))
352 blk_mq_free_request(rq->next_rq);
353 blk_mq_free_request(rq);
356 EXPORT_SYMBOL(__blk_mq_end_request);
358 void blk_mq_end_request(struct request *rq, int error)
360 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
362 __blk_mq_end_request(rq, error);
364 EXPORT_SYMBOL(blk_mq_end_request);
366 static void __blk_mq_complete_request_remote(void *data)
368 struct request *rq = data;
370 rq->q->softirq_done_fn(rq);
373 static void blk_mq_ipi_complete_request(struct request *rq)
375 struct blk_mq_ctx *ctx = rq->mq_ctx;
379 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
380 rq->q->softirq_done_fn(rq);
385 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
386 shared = cpus_share_cache(cpu, ctx->cpu);
388 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
389 rq->csd.func = __blk_mq_complete_request_remote;
392 smp_call_function_single_async(ctx->cpu, &rq->csd);
394 rq->q->softirq_done_fn(rq);
399 static void __blk_mq_complete_request(struct request *rq)
401 struct request_queue *q = rq->q;
403 if (!q->softirq_done_fn)
404 blk_mq_end_request(rq, rq->errors);
406 blk_mq_ipi_complete_request(rq);
410 * blk_mq_complete_request - end I/O on a request
411 * @rq: the request being processed
414 * Ends all I/O on a request. It does not handle partial completions.
415 * The actual completion happens out-of-order, through a IPI handler.
417 void blk_mq_complete_request(struct request *rq, int error)
419 struct request_queue *q = rq->q;
421 if (unlikely(blk_should_fake_timeout(q)))
423 if (!blk_mark_rq_complete(rq)) {
425 __blk_mq_complete_request(rq);
428 EXPORT_SYMBOL(blk_mq_complete_request);
430 int blk_mq_request_started(struct request *rq)
432 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
434 EXPORT_SYMBOL_GPL(blk_mq_request_started);
436 void blk_mq_start_request(struct request *rq)
438 struct request_queue *q = rq->q;
440 trace_block_rq_issue(q, rq);
442 rq->resid_len = blk_rq_bytes(rq);
443 if (unlikely(blk_bidi_rq(rq)))
444 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
449 * Ensure that ->deadline is visible before set the started
450 * flag and clear the completed flag.
452 smp_mb__before_atomic();
455 * Mark us as started and clear complete. Complete might have been
456 * set if requeue raced with timeout, which then marked it as
457 * complete. So be sure to clear complete again when we start
458 * the request, otherwise we'll ignore the completion event.
460 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
461 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
462 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
463 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
465 if (q->dma_drain_size && blk_rq_bytes(rq)) {
467 * Make sure space for the drain appears. We know we can do
468 * this because max_hw_segments has been adjusted to be one
469 * fewer than the device can handle.
471 rq->nr_phys_segments++;
474 EXPORT_SYMBOL(blk_mq_start_request);
476 static void __blk_mq_requeue_request(struct request *rq)
478 struct request_queue *q = rq->q;
480 trace_block_rq_requeue(q, rq);
482 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
483 if (q->dma_drain_size && blk_rq_bytes(rq))
484 rq->nr_phys_segments--;
488 void blk_mq_requeue_request(struct request *rq)
490 __blk_mq_requeue_request(rq);
492 BUG_ON(blk_queued_rq(rq));
493 blk_mq_add_to_requeue_list(rq, true);
495 EXPORT_SYMBOL(blk_mq_requeue_request);
497 static void blk_mq_requeue_work(struct work_struct *work)
499 struct request_queue *q =
500 container_of(work, struct request_queue, requeue_work);
502 struct request *rq, *next;
505 spin_lock_irqsave(&q->requeue_lock, flags);
506 list_splice_init(&q->requeue_list, &rq_list);
507 spin_unlock_irqrestore(&q->requeue_lock, flags);
509 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
510 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
513 rq->cmd_flags &= ~REQ_SOFTBARRIER;
514 list_del_init(&rq->queuelist);
515 blk_mq_insert_request(rq, true, false, false);
518 while (!list_empty(&rq_list)) {
519 rq = list_entry(rq_list.next, struct request, queuelist);
520 list_del_init(&rq->queuelist);
521 blk_mq_insert_request(rq, false, false, false);
525 * Use the start variant of queue running here, so that running
526 * the requeue work will kick stopped queues.
528 blk_mq_start_hw_queues(q);
531 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
533 struct request_queue *q = rq->q;
537 * We abuse this flag that is otherwise used by the I/O scheduler to
538 * request head insertation from the workqueue.
540 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
542 spin_lock_irqsave(&q->requeue_lock, flags);
544 rq->cmd_flags |= REQ_SOFTBARRIER;
545 list_add(&rq->queuelist, &q->requeue_list);
547 list_add_tail(&rq->queuelist, &q->requeue_list);
549 spin_unlock_irqrestore(&q->requeue_lock, flags);
551 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
553 void blk_mq_cancel_requeue_work(struct request_queue *q)
555 cancel_work_sync(&q->requeue_work);
557 EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work);
559 void blk_mq_kick_requeue_list(struct request_queue *q)
561 kblockd_schedule_work(&q->requeue_work);
563 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
565 void blk_mq_abort_requeue_list(struct request_queue *q)
570 spin_lock_irqsave(&q->requeue_lock, flags);
571 list_splice_init(&q->requeue_list, &rq_list);
572 spin_unlock_irqrestore(&q->requeue_lock, flags);
574 while (!list_empty(&rq_list)) {
577 rq = list_first_entry(&rq_list, struct request, queuelist);
578 list_del_init(&rq->queuelist);
580 blk_mq_end_request(rq, rq->errors);
583 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
585 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
587 if (tag < tags->nr_tags)
588 return tags->rqs[tag];
592 EXPORT_SYMBOL(blk_mq_tag_to_rq);
594 struct blk_mq_timeout_data {
596 unsigned int next_set;
599 void blk_mq_rq_timed_out(struct request *req, bool reserved)
601 struct blk_mq_ops *ops = req->q->mq_ops;
602 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
605 * We know that complete is set at this point. If STARTED isn't set
606 * anymore, then the request isn't active and the "timeout" should
607 * just be ignored. This can happen due to the bitflag ordering.
608 * Timeout first checks if STARTED is set, and if it is, assumes
609 * the request is active. But if we race with completion, then
610 * we both flags will get cleared. So check here again, and ignore
611 * a timeout event with a request that isn't active.
613 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
617 ret = ops->timeout(req, reserved);
621 __blk_mq_complete_request(req);
623 case BLK_EH_RESET_TIMER:
625 blk_clear_rq_complete(req);
627 case BLK_EH_NOT_HANDLED:
630 printk(KERN_ERR "block: bad eh return: %d\n", ret);
635 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
636 struct request *rq, void *priv, bool reserved)
638 struct blk_mq_timeout_data *data = priv;
640 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
642 * If a request wasn't started before the queue was
643 * marked dying, kill it here or it'll go unnoticed.
645 if (unlikely(blk_queue_dying(rq->q))) {
647 blk_mq_end_request(rq, rq->errors);
652 if (time_after_eq(jiffies, rq->deadline)) {
653 if (!blk_mark_rq_complete(rq))
654 blk_mq_rq_timed_out(rq, reserved);
655 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
656 data->next = rq->deadline;
661 static void blk_mq_timeout_work(struct work_struct *work)
663 struct request_queue *q =
664 container_of(work, struct request_queue, timeout_work);
665 struct blk_mq_timeout_data data = {
671 if (blk_queue_enter(q, true))
674 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
677 data.next = blk_rq_timeout(round_jiffies_up(data.next));
678 mod_timer(&q->timeout, data.next);
680 struct blk_mq_hw_ctx *hctx;
682 queue_for_each_hw_ctx(q, hctx, i) {
683 /* the hctx may be unmapped, so check it here */
684 if (blk_mq_hw_queue_mapped(hctx))
685 blk_mq_tag_idle(hctx);
692 * Reverse check our software queue for entries that we could potentially
693 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
694 * too much time checking for merges.
696 static bool blk_mq_attempt_merge(struct request_queue *q,
697 struct blk_mq_ctx *ctx, struct bio *bio)
702 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
708 if (!blk_rq_merge_ok(rq, bio))
711 el_ret = blk_try_merge(rq, bio);
712 if (el_ret == ELEVATOR_BACK_MERGE) {
713 if (bio_attempt_back_merge(q, rq, bio)) {
718 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
719 if (bio_attempt_front_merge(q, rq, bio)) {
731 * Process software queues that have been marked busy, splicing them
732 * to the for-dispatch
734 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
736 struct blk_mq_ctx *ctx;
739 for (i = 0; i < hctx->ctx_map.size; i++) {
740 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
741 unsigned int off, bit;
747 off = i * hctx->ctx_map.bits_per_word;
749 bit = find_next_bit(&bm->word, bm->depth, bit);
750 if (bit >= bm->depth)
753 ctx = hctx->ctxs[bit + off];
754 clear_bit(bit, &bm->word);
755 spin_lock(&ctx->lock);
756 list_splice_tail_init(&ctx->rq_list, list);
757 spin_unlock(&ctx->lock);
765 * Run this hardware queue, pulling any software queues mapped to it in.
766 * Note that this function currently has various problems around ordering
767 * of IO. In particular, we'd like FIFO behaviour on handling existing
768 * items on the hctx->dispatch list. Ignore that for now.
770 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
772 struct request_queue *q = hctx->queue;
775 LIST_HEAD(driver_list);
776 struct list_head *dptr;
779 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
781 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
787 * Touch any software queue that has pending entries.
789 flush_busy_ctxs(hctx, &rq_list);
792 * If we have previous entries on our dispatch list, grab them
793 * and stuff them at the front for more fair dispatch.
795 if (!list_empty_careful(&hctx->dispatch)) {
796 spin_lock(&hctx->lock);
797 if (!list_empty(&hctx->dispatch))
798 list_splice_init(&hctx->dispatch, &rq_list);
799 spin_unlock(&hctx->lock);
803 * Start off with dptr being NULL, so we start the first request
804 * immediately, even if we have more pending.
809 * Now process all the entries, sending them to the driver.
812 while (!list_empty(&rq_list)) {
813 struct blk_mq_queue_data bd;
816 rq = list_first_entry(&rq_list, struct request, queuelist);
817 list_del_init(&rq->queuelist);
821 bd.last = list_empty(&rq_list);
823 ret = q->mq_ops->queue_rq(hctx, &bd);
825 case BLK_MQ_RQ_QUEUE_OK:
828 case BLK_MQ_RQ_QUEUE_BUSY:
829 list_add(&rq->queuelist, &rq_list);
830 __blk_mq_requeue_request(rq);
833 pr_err("blk-mq: bad return on queue: %d\n", ret);
834 case BLK_MQ_RQ_QUEUE_ERROR:
836 blk_mq_end_request(rq, rq->errors);
840 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
844 * We've done the first request. If we have more than 1
845 * left in the list, set dptr to defer issue.
847 if (!dptr && rq_list.next != rq_list.prev)
852 hctx->dispatched[0]++;
853 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
854 hctx->dispatched[ilog2(queued) + 1]++;
857 * Any items that need requeuing? Stuff them into hctx->dispatch,
858 * that is where we will continue on next queue run.
860 if (!list_empty(&rq_list)) {
861 spin_lock(&hctx->lock);
862 list_splice(&rq_list, &hctx->dispatch);
863 spin_unlock(&hctx->lock);
865 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
866 * it's possible the queue is stopped and restarted again
867 * before this. Queue restart will dispatch requests. And since
868 * requests in rq_list aren't added into hctx->dispatch yet,
869 * the requests in rq_list might get lost.
871 * blk_mq_run_hw_queue() already checks the STOPPED bit
873 blk_mq_run_hw_queue(hctx, true);
878 * It'd be great if the workqueue API had a way to pass
879 * in a mask and had some smarts for more clever placement.
880 * For now we just round-robin here, switching for every
881 * BLK_MQ_CPU_WORK_BATCH queued items.
883 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
885 if (hctx->queue->nr_hw_queues == 1)
886 return WORK_CPU_UNBOUND;
888 if (--hctx->next_cpu_batch <= 0) {
889 int cpu = hctx->next_cpu, next_cpu;
891 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
892 if (next_cpu >= nr_cpu_ids)
893 next_cpu = cpumask_first(hctx->cpumask);
895 hctx->next_cpu = next_cpu;
896 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
901 return hctx->next_cpu;
904 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
906 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
907 !blk_mq_hw_queue_mapped(hctx)))
912 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
913 __blk_mq_run_hw_queue(hctx);
921 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
925 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
927 struct blk_mq_hw_ctx *hctx;
930 queue_for_each_hw_ctx(q, hctx, i) {
931 if ((!blk_mq_hctx_has_pending(hctx) &&
932 list_empty_careful(&hctx->dispatch)) ||
933 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
936 blk_mq_run_hw_queue(hctx, async);
939 EXPORT_SYMBOL(blk_mq_run_hw_queues);
941 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
943 cancel_delayed_work(&hctx->run_work);
944 cancel_delayed_work(&hctx->delay_work);
945 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
947 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
949 void blk_mq_stop_hw_queues(struct request_queue *q)
951 struct blk_mq_hw_ctx *hctx;
954 queue_for_each_hw_ctx(q, hctx, i)
955 blk_mq_stop_hw_queue(hctx);
957 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
959 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
961 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
963 blk_mq_run_hw_queue(hctx, false);
965 EXPORT_SYMBOL(blk_mq_start_hw_queue);
967 void blk_mq_start_hw_queues(struct request_queue *q)
969 struct blk_mq_hw_ctx *hctx;
972 queue_for_each_hw_ctx(q, hctx, i)
973 blk_mq_start_hw_queue(hctx);
975 EXPORT_SYMBOL(blk_mq_start_hw_queues);
977 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
979 struct blk_mq_hw_ctx *hctx;
982 queue_for_each_hw_ctx(q, hctx, i) {
983 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
986 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
987 blk_mq_run_hw_queue(hctx, async);
990 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
992 static void blk_mq_run_work_fn(struct work_struct *work)
994 struct blk_mq_hw_ctx *hctx;
996 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
998 __blk_mq_run_hw_queue(hctx);
1001 static void blk_mq_delay_work_fn(struct work_struct *work)
1003 struct blk_mq_hw_ctx *hctx;
1005 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
1007 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
1008 __blk_mq_run_hw_queue(hctx);
1011 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1013 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1016 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1017 &hctx->delay_work, msecs_to_jiffies(msecs));
1019 EXPORT_SYMBOL(blk_mq_delay_queue);
1021 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1022 struct blk_mq_ctx *ctx,
1026 trace_block_rq_insert(hctx->queue, rq);
1029 list_add(&rq->queuelist, &ctx->rq_list);
1031 list_add_tail(&rq->queuelist, &ctx->rq_list);
1034 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
1035 struct request *rq, bool at_head)
1037 struct blk_mq_ctx *ctx = rq->mq_ctx;
1039 __blk_mq_insert_req_list(hctx, ctx, rq, at_head);
1040 blk_mq_hctx_mark_pending(hctx, ctx);
1043 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1046 struct request_queue *q = rq->q;
1047 struct blk_mq_hw_ctx *hctx;
1048 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
1050 current_ctx = blk_mq_get_ctx(q);
1051 if (!cpu_online(ctx->cpu))
1052 rq->mq_ctx = ctx = current_ctx;
1054 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1056 spin_lock(&ctx->lock);
1057 __blk_mq_insert_request(hctx, rq, at_head);
1058 spin_unlock(&ctx->lock);
1061 blk_mq_run_hw_queue(hctx, async);
1063 blk_mq_put_ctx(current_ctx);
1066 static void blk_mq_insert_requests(struct request_queue *q,
1067 struct blk_mq_ctx *ctx,
1068 struct list_head *list,
1073 struct blk_mq_hw_ctx *hctx;
1074 struct blk_mq_ctx *current_ctx;
1076 trace_block_unplug(q, depth, !from_schedule);
1078 current_ctx = blk_mq_get_ctx(q);
1080 if (!cpu_online(ctx->cpu))
1082 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1085 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1088 spin_lock(&ctx->lock);
1089 while (!list_empty(list)) {
1092 rq = list_first_entry(list, struct request, queuelist);
1093 list_del_init(&rq->queuelist);
1095 __blk_mq_insert_req_list(hctx, ctx, rq, false);
1097 blk_mq_hctx_mark_pending(hctx, ctx);
1098 spin_unlock(&ctx->lock);
1100 blk_mq_run_hw_queue(hctx, from_schedule);
1101 blk_mq_put_ctx(current_ctx);
1104 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1106 struct request *rqa = container_of(a, struct request, queuelist);
1107 struct request *rqb = container_of(b, struct request, queuelist);
1109 return !(rqa->mq_ctx < rqb->mq_ctx ||
1110 (rqa->mq_ctx == rqb->mq_ctx &&
1111 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1114 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1116 struct blk_mq_ctx *this_ctx;
1117 struct request_queue *this_q;
1120 LIST_HEAD(ctx_list);
1123 list_splice_init(&plug->mq_list, &list);
1125 list_sort(NULL, &list, plug_ctx_cmp);
1131 while (!list_empty(&list)) {
1132 rq = list_entry_rq(list.next);
1133 list_del_init(&rq->queuelist);
1135 if (rq->mq_ctx != this_ctx) {
1137 blk_mq_insert_requests(this_q, this_ctx,
1142 this_ctx = rq->mq_ctx;
1148 list_add_tail(&rq->queuelist, &ctx_list);
1152 * If 'this_ctx' is set, we know we have entries to complete
1153 * on 'ctx_list'. Do those.
1156 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1161 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1163 init_request_from_bio(rq, bio);
1165 blk_account_io_start(rq, 1);
1168 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1170 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1171 !blk_queue_nomerges(hctx->queue);
1174 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1175 struct blk_mq_ctx *ctx,
1176 struct request *rq, struct bio *bio)
1178 if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) {
1179 blk_mq_bio_to_request(rq, bio);
1180 spin_lock(&ctx->lock);
1182 __blk_mq_insert_request(hctx, rq, false);
1183 spin_unlock(&ctx->lock);
1186 struct request_queue *q = hctx->queue;
1188 spin_lock(&ctx->lock);
1189 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1190 blk_mq_bio_to_request(rq, bio);
1194 spin_unlock(&ctx->lock);
1195 __blk_mq_free_request(hctx, ctx, rq);
1200 struct blk_map_ctx {
1201 struct blk_mq_hw_ctx *hctx;
1202 struct blk_mq_ctx *ctx;
1205 static struct request *blk_mq_map_request(struct request_queue *q,
1207 struct blk_map_ctx *data)
1209 struct blk_mq_hw_ctx *hctx;
1210 struct blk_mq_ctx *ctx;
1212 int op = bio_data_dir(bio);
1214 struct blk_mq_alloc_data alloc_data;
1216 blk_queue_enter_live(q);
1217 ctx = blk_mq_get_ctx(q);
1218 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1220 if (rw_is_sync(bio_op(bio), bio->bi_rw))
1221 op_flags |= REQ_SYNC;
1223 trace_block_getrq(q, bio, op);
1224 blk_mq_set_alloc_data(&alloc_data, q, BLK_MQ_REQ_NOWAIT, ctx, hctx);
1225 rq = __blk_mq_alloc_request(&alloc_data, op, op_flags);
1226 if (unlikely(!rq)) {
1227 __blk_mq_run_hw_queue(hctx);
1228 blk_mq_put_ctx(ctx);
1229 trace_block_sleeprq(q, bio, op);
1231 ctx = blk_mq_get_ctx(q);
1232 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1233 blk_mq_set_alloc_data(&alloc_data, q, 0, ctx, hctx);
1234 rq = __blk_mq_alloc_request(&alloc_data, op, op_flags);
1235 ctx = alloc_data.ctx;
1236 hctx = alloc_data.hctx;
1245 static int blk_mq_direct_issue_request(struct request *rq, blk_qc_t *cookie)
1248 struct request_queue *q = rq->q;
1249 struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q,
1251 struct blk_mq_queue_data bd = {
1256 blk_qc_t new_cookie = blk_tag_to_qc_t(rq->tag, hctx->queue_num);
1259 * For OK queue, we are done. For error, kill it. Any other
1260 * error (busy), just add it to our list as we previously
1263 ret = q->mq_ops->queue_rq(hctx, &bd);
1264 if (ret == BLK_MQ_RQ_QUEUE_OK) {
1265 *cookie = new_cookie;
1269 __blk_mq_requeue_request(rq);
1271 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1272 *cookie = BLK_QC_T_NONE;
1274 blk_mq_end_request(rq, rq->errors);
1282 * Multiple hardware queue variant. This will not use per-process plugs,
1283 * but will attempt to bypass the hctx queueing if we can go straight to
1284 * hardware for SYNC IO.
1286 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1288 const int is_sync = rw_is_sync(bio_op(bio), bio->bi_rw);
1289 const int is_flush_fua = bio->bi_rw & (REQ_PREFLUSH | REQ_FUA);
1290 struct blk_map_ctx data;
1292 unsigned int request_count = 0;
1293 struct blk_plug *plug;
1294 struct request *same_queue_rq = NULL;
1297 blk_queue_bounce(q, &bio);
1299 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1301 return BLK_QC_T_NONE;
1304 blk_queue_split(q, &bio, q->bio_split);
1306 if (!is_flush_fua && !blk_queue_nomerges(q)) {
1307 if (blk_attempt_plug_merge(q, bio, &request_count,
1309 return BLK_QC_T_NONE;
1311 request_count = blk_plug_queued_count(q);
1313 rq = blk_mq_map_request(q, bio, &data);
1315 return BLK_QC_T_NONE;
1317 cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num);
1319 if (unlikely(is_flush_fua)) {
1320 blk_mq_bio_to_request(rq, bio);
1321 blk_insert_flush(rq);
1325 plug = current->plug;
1327 * If the driver supports defer issued based on 'last', then
1328 * queue it up like normal since we can potentially save some
1331 if (((plug && !blk_queue_nomerges(q)) || is_sync) &&
1332 !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1333 struct request *old_rq = NULL;
1335 blk_mq_bio_to_request(rq, bio);
1338 * We do limited pluging. If the bio can be merged, do that.
1339 * Otherwise the existing request in the plug list will be
1340 * issued. So the plug list will have one request at most
1344 * The plug list might get flushed before this. If that
1345 * happens, same_queue_rq is invalid and plug list is
1348 if (same_queue_rq && !list_empty(&plug->mq_list)) {
1349 old_rq = same_queue_rq;
1350 list_del_init(&old_rq->queuelist);
1352 list_add_tail(&rq->queuelist, &plug->mq_list);
1353 } else /* is_sync */
1355 blk_mq_put_ctx(data.ctx);
1358 if (!blk_mq_direct_issue_request(old_rq, &cookie))
1360 blk_mq_insert_request(old_rq, false, true, true);
1364 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1366 * For a SYNC request, send it to the hardware immediately. For
1367 * an ASYNC request, just ensure that we run it later on. The
1368 * latter allows for merging opportunities and more efficient
1372 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1374 blk_mq_put_ctx(data.ctx);
1380 * Single hardware queue variant. This will attempt to use any per-process
1381 * plug for merging and IO deferral.
1383 static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio)
1385 const int is_sync = rw_is_sync(bio_op(bio), bio->bi_rw);
1386 const int is_flush_fua = bio->bi_rw & (REQ_PREFLUSH | REQ_FUA);
1387 struct blk_plug *plug;
1388 unsigned int request_count = 0;
1389 struct blk_map_ctx data;
1393 blk_queue_bounce(q, &bio);
1395 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1397 return BLK_QC_T_NONE;
1400 blk_queue_split(q, &bio, q->bio_split);
1402 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1403 blk_attempt_plug_merge(q, bio, &request_count, NULL))
1404 return BLK_QC_T_NONE;
1406 rq = blk_mq_map_request(q, bio, &data);
1408 return BLK_QC_T_NONE;
1410 cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num);
1412 if (unlikely(is_flush_fua)) {
1413 blk_mq_bio_to_request(rq, bio);
1414 blk_insert_flush(rq);
1419 * A task plug currently exists. Since this is completely lockless,
1420 * utilize that to temporarily store requests until the task is
1421 * either done or scheduled away.
1423 plug = current->plug;
1425 blk_mq_bio_to_request(rq, bio);
1427 trace_block_plug(q);
1429 blk_mq_put_ctx(data.ctx);
1431 if (request_count >= BLK_MAX_REQUEST_COUNT) {
1432 blk_flush_plug_list(plug, false);
1433 trace_block_plug(q);
1436 list_add_tail(&rq->queuelist, &plug->mq_list);
1440 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1442 * For a SYNC request, send it to the hardware immediately. For
1443 * an ASYNC request, just ensure that we run it later on. The
1444 * latter allows for merging opportunities and more efficient
1448 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1451 blk_mq_put_ctx(data.ctx);
1456 * Default mapping to a software queue, since we use one per CPU.
1458 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1460 return q->queue_hw_ctx[q->mq_map[cpu]];
1462 EXPORT_SYMBOL(blk_mq_map_queue);
1464 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1465 struct blk_mq_tags *tags, unsigned int hctx_idx)
1469 if (tags->rqs && set->ops->exit_request) {
1472 for (i = 0; i < tags->nr_tags; i++) {
1475 set->ops->exit_request(set->driver_data, tags->rqs[i],
1477 tags->rqs[i] = NULL;
1481 while (!list_empty(&tags->page_list)) {
1482 page = list_first_entry(&tags->page_list, struct page, lru);
1483 list_del_init(&page->lru);
1485 * Remove kmemleak object previously allocated in
1486 * blk_mq_init_rq_map().
1488 kmemleak_free(page_address(page));
1489 __free_pages(page, page->private);
1494 blk_mq_free_tags(tags);
1497 static size_t order_to_size(unsigned int order)
1499 return (size_t)PAGE_SIZE << order;
1502 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1503 unsigned int hctx_idx)
1505 struct blk_mq_tags *tags;
1506 unsigned int i, j, entries_per_page, max_order = 4;
1507 size_t rq_size, left;
1509 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1511 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1515 INIT_LIST_HEAD(&tags->page_list);
1517 tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1518 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1521 blk_mq_free_tags(tags);
1526 * rq_size is the size of the request plus driver payload, rounded
1527 * to the cacheline size
1529 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1531 left = rq_size * set->queue_depth;
1533 for (i = 0; i < set->queue_depth; ) {
1534 int this_order = max_order;
1539 while (this_order && left < order_to_size(this_order - 1))
1543 page = alloc_pages_node(set->numa_node,
1544 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1550 if (order_to_size(this_order) < rq_size)
1557 page->private = this_order;
1558 list_add_tail(&page->lru, &tags->page_list);
1560 p = page_address(page);
1562 * Allow kmemleak to scan these pages as they contain pointers
1563 * to additional allocations like via ops->init_request().
1565 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_KERNEL);
1566 entries_per_page = order_to_size(this_order) / rq_size;
1567 to_do = min(entries_per_page, set->queue_depth - i);
1568 left -= to_do * rq_size;
1569 for (j = 0; j < to_do; j++) {
1571 if (set->ops->init_request) {
1572 if (set->ops->init_request(set->driver_data,
1573 tags->rqs[i], hctx_idx, i,
1575 tags->rqs[i] = NULL;
1587 blk_mq_free_rq_map(set, tags, hctx_idx);
1591 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1596 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1598 unsigned int bpw = 8, total, num_maps, i;
1600 bitmap->bits_per_word = bpw;
1602 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1603 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1609 for (i = 0; i < num_maps; i++) {
1610 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1611 total -= bitmap->map[i].depth;
1617 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1619 struct request_queue *q = hctx->queue;
1620 struct blk_mq_ctx *ctx;
1624 * Move ctx entries to new CPU, if this one is going away.
1626 ctx = __blk_mq_get_ctx(q, cpu);
1628 spin_lock(&ctx->lock);
1629 if (!list_empty(&ctx->rq_list)) {
1630 list_splice_init(&ctx->rq_list, &tmp);
1631 blk_mq_hctx_clear_pending(hctx, ctx);
1633 spin_unlock(&ctx->lock);
1635 if (list_empty(&tmp))
1638 ctx = blk_mq_get_ctx(q);
1639 spin_lock(&ctx->lock);
1641 while (!list_empty(&tmp)) {
1644 rq = list_first_entry(&tmp, struct request, queuelist);
1646 list_move_tail(&rq->queuelist, &ctx->rq_list);
1649 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1650 blk_mq_hctx_mark_pending(hctx, ctx);
1652 spin_unlock(&ctx->lock);
1654 blk_mq_run_hw_queue(hctx, true);
1655 blk_mq_put_ctx(ctx);
1659 static int blk_mq_hctx_notify(void *data, unsigned long action,
1662 struct blk_mq_hw_ctx *hctx = data;
1664 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1665 return blk_mq_hctx_cpu_offline(hctx, cpu);
1668 * In case of CPU online, tags may be reallocated
1669 * in blk_mq_map_swqueue() after mapping is updated.
1675 /* hctx->ctxs will be freed in queue's release handler */
1676 static void blk_mq_exit_hctx(struct request_queue *q,
1677 struct blk_mq_tag_set *set,
1678 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1680 unsigned flush_start_tag = set->queue_depth;
1682 blk_mq_tag_idle(hctx);
1684 if (set->ops->exit_request)
1685 set->ops->exit_request(set->driver_data,
1686 hctx->fq->flush_rq, hctx_idx,
1687 flush_start_tag + hctx_idx);
1689 if (set->ops->exit_hctx)
1690 set->ops->exit_hctx(hctx, hctx_idx);
1692 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1693 blk_free_flush_queue(hctx->fq);
1694 blk_mq_free_bitmap(&hctx->ctx_map);
1697 static void blk_mq_exit_hw_queues(struct request_queue *q,
1698 struct blk_mq_tag_set *set, int nr_queue)
1700 struct blk_mq_hw_ctx *hctx;
1703 queue_for_each_hw_ctx(q, hctx, i) {
1706 blk_mq_exit_hctx(q, set, hctx, i);
1710 static void blk_mq_free_hw_queues(struct request_queue *q,
1711 struct blk_mq_tag_set *set)
1713 struct blk_mq_hw_ctx *hctx;
1716 queue_for_each_hw_ctx(q, hctx, i)
1717 free_cpumask_var(hctx->cpumask);
1720 static int blk_mq_init_hctx(struct request_queue *q,
1721 struct blk_mq_tag_set *set,
1722 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1725 unsigned flush_start_tag = set->queue_depth;
1727 node = hctx->numa_node;
1728 if (node == NUMA_NO_NODE)
1729 node = hctx->numa_node = set->numa_node;
1731 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1732 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1733 spin_lock_init(&hctx->lock);
1734 INIT_LIST_HEAD(&hctx->dispatch);
1736 hctx->queue_num = hctx_idx;
1737 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1739 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1740 blk_mq_hctx_notify, hctx);
1741 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1743 hctx->tags = set->tags[hctx_idx];
1746 * Allocate space for all possible cpus to avoid allocation at
1749 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1752 goto unregister_cpu_notifier;
1754 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1759 if (set->ops->init_hctx &&
1760 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1763 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1767 if (set->ops->init_request &&
1768 set->ops->init_request(set->driver_data,
1769 hctx->fq->flush_rq, hctx_idx,
1770 flush_start_tag + hctx_idx, node))
1778 if (set->ops->exit_hctx)
1779 set->ops->exit_hctx(hctx, hctx_idx);
1781 blk_mq_free_bitmap(&hctx->ctx_map);
1784 unregister_cpu_notifier:
1785 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1790 static void blk_mq_init_cpu_queues(struct request_queue *q,
1791 unsigned int nr_hw_queues)
1795 for_each_possible_cpu(i) {
1796 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1797 struct blk_mq_hw_ctx *hctx;
1799 memset(__ctx, 0, sizeof(*__ctx));
1801 spin_lock_init(&__ctx->lock);
1802 INIT_LIST_HEAD(&__ctx->rq_list);
1805 /* If the cpu isn't online, the cpu is mapped to first hctx */
1809 hctx = q->mq_ops->map_queue(q, i);
1812 * Set local node, IFF we have more than one hw queue. If
1813 * not, we remain on the home node of the device
1815 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1816 hctx->numa_node = local_memory_node(cpu_to_node(i));
1820 static void blk_mq_map_swqueue(struct request_queue *q,
1821 const struct cpumask *online_mask)
1824 struct blk_mq_hw_ctx *hctx;
1825 struct blk_mq_ctx *ctx;
1826 struct blk_mq_tag_set *set = q->tag_set;
1829 * Avoid others reading imcomplete hctx->cpumask through sysfs
1831 mutex_lock(&q->sysfs_lock);
1833 queue_for_each_hw_ctx(q, hctx, i) {
1834 cpumask_clear(hctx->cpumask);
1839 * Map software to hardware queues
1841 for_each_possible_cpu(i) {
1842 /* If the cpu isn't online, the cpu is mapped to first hctx */
1843 if (!cpumask_test_cpu(i, online_mask))
1846 ctx = per_cpu_ptr(q->queue_ctx, i);
1847 hctx = q->mq_ops->map_queue(q, i);
1849 cpumask_set_cpu(i, hctx->cpumask);
1850 ctx->index_hw = hctx->nr_ctx;
1851 hctx->ctxs[hctx->nr_ctx++] = ctx;
1854 mutex_unlock(&q->sysfs_lock);
1856 queue_for_each_hw_ctx(q, hctx, i) {
1857 struct blk_mq_ctxmap *map = &hctx->ctx_map;
1860 * If no software queues are mapped to this hardware queue,
1861 * disable it and free the request entries.
1863 if (!hctx->nr_ctx) {
1865 blk_mq_free_rq_map(set, set->tags[i], i);
1866 set->tags[i] = NULL;
1872 /* unmapped hw queue can be remapped after CPU topo changed */
1874 set->tags[i] = blk_mq_init_rq_map(set, i);
1875 hctx->tags = set->tags[i];
1876 WARN_ON(!hctx->tags);
1878 cpumask_copy(hctx->tags->cpumask, hctx->cpumask);
1880 * Set the map size to the number of mapped software queues.
1881 * This is more accurate and more efficient than looping
1882 * over all possibly mapped software queues.
1884 map->size = DIV_ROUND_UP(hctx->nr_ctx, map->bits_per_word);
1887 * Initialize batch roundrobin counts
1889 hctx->next_cpu = cpumask_first(hctx->cpumask);
1890 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1894 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
1896 struct blk_mq_hw_ctx *hctx;
1899 queue_for_each_hw_ctx(q, hctx, i) {
1901 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1903 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1907 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
1909 struct request_queue *q;
1911 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1912 blk_mq_freeze_queue(q);
1913 queue_set_hctx_shared(q, shared);
1914 blk_mq_unfreeze_queue(q);
1918 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1920 struct blk_mq_tag_set *set = q->tag_set;
1922 mutex_lock(&set->tag_list_lock);
1923 list_del_init(&q->tag_set_list);
1924 if (list_is_singular(&set->tag_list)) {
1925 /* just transitioned to unshared */
1926 set->flags &= ~BLK_MQ_F_TAG_SHARED;
1927 /* update existing queue */
1928 blk_mq_update_tag_set_depth(set, false);
1930 mutex_unlock(&set->tag_list_lock);
1933 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1934 struct request_queue *q)
1938 mutex_lock(&set->tag_list_lock);
1940 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
1941 if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
1942 set->flags |= BLK_MQ_F_TAG_SHARED;
1943 /* update existing queue */
1944 blk_mq_update_tag_set_depth(set, true);
1946 if (set->flags & BLK_MQ_F_TAG_SHARED)
1947 queue_set_hctx_shared(q, true);
1948 list_add_tail(&q->tag_set_list, &set->tag_list);
1950 mutex_unlock(&set->tag_list_lock);
1954 * It is the actual release handler for mq, but we do it from
1955 * request queue's release handler for avoiding use-after-free
1956 * and headache because q->mq_kobj shouldn't have been introduced,
1957 * but we can't group ctx/kctx kobj without it.
1959 void blk_mq_release(struct request_queue *q)
1961 struct blk_mq_hw_ctx *hctx;
1964 /* hctx kobj stays in hctx */
1965 queue_for_each_hw_ctx(q, hctx, i) {
1975 kfree(q->queue_hw_ctx);
1977 /* ctx kobj stays in queue_ctx */
1978 free_percpu(q->queue_ctx);
1981 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1983 struct request_queue *uninit_q, *q;
1985 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1987 return ERR_PTR(-ENOMEM);
1989 q = blk_mq_init_allocated_queue(set, uninit_q);
1991 blk_cleanup_queue(uninit_q);
1995 EXPORT_SYMBOL(blk_mq_init_queue);
1997 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
1998 struct request_queue *q)
2001 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2003 blk_mq_sysfs_unregister(q);
2004 for (i = 0; i < set->nr_hw_queues; i++) {
2010 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2011 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
2016 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2023 atomic_set(&hctxs[i]->nr_active, 0);
2024 hctxs[i]->numa_node = node;
2025 hctxs[i]->queue_num = i;
2027 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2028 free_cpumask_var(hctxs[i]->cpumask);
2033 blk_mq_hctx_kobj_init(hctxs[i]);
2035 for (j = i; j < q->nr_hw_queues; j++) {
2036 struct blk_mq_hw_ctx *hctx = hctxs[j];
2040 blk_mq_free_rq_map(set, hctx->tags, j);
2041 set->tags[j] = NULL;
2043 blk_mq_exit_hctx(q, set, hctx, j);
2044 free_cpumask_var(hctx->cpumask);
2045 kobject_put(&hctx->kobj);
2052 q->nr_hw_queues = i;
2053 blk_mq_sysfs_register(q);
2056 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2057 struct request_queue *q)
2059 /* mark the queue as mq asap */
2060 q->mq_ops = set->ops;
2062 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2066 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2067 GFP_KERNEL, set->numa_node);
2068 if (!q->queue_hw_ctx)
2071 q->mq_map = blk_mq_make_queue_map(set);
2075 blk_mq_realloc_hw_ctxs(set, q);
2076 if (!q->nr_hw_queues)
2079 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2080 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2082 q->nr_queues = nr_cpu_ids;
2084 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2086 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2087 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2089 q->sg_reserved_size = INT_MAX;
2091 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
2092 INIT_LIST_HEAD(&q->requeue_list);
2093 spin_lock_init(&q->requeue_lock);
2095 if (q->nr_hw_queues > 1)
2096 blk_queue_make_request(q, blk_mq_make_request);
2098 blk_queue_make_request(q, blk_sq_make_request);
2101 * Do this after blk_queue_make_request() overrides it...
2103 q->nr_requests = set->queue_depth;
2105 if (set->ops->complete)
2106 blk_queue_softirq_done(q, set->ops->complete);
2108 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2111 mutex_lock(&all_q_mutex);
2113 list_add_tail(&q->all_q_node, &all_q_list);
2114 blk_mq_add_queue_tag_set(set, q);
2115 blk_mq_map_swqueue(q, cpu_online_mask);
2117 mutex_unlock(&all_q_mutex);
2125 kfree(q->queue_hw_ctx);
2127 free_percpu(q->queue_ctx);
2130 return ERR_PTR(-ENOMEM);
2132 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2134 void blk_mq_free_queue(struct request_queue *q)
2136 struct blk_mq_tag_set *set = q->tag_set;
2138 mutex_lock(&all_q_mutex);
2139 list_del_init(&q->all_q_node);
2140 mutex_unlock(&all_q_mutex);
2142 blk_mq_del_queue_tag_set(q);
2144 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2145 blk_mq_free_hw_queues(q, set);
2148 /* Basically redo blk_mq_init_queue with queue frozen */
2149 static void blk_mq_queue_reinit(struct request_queue *q,
2150 const struct cpumask *online_mask)
2152 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2154 blk_mq_sysfs_unregister(q);
2156 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues, online_mask);
2159 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2160 * we should change hctx numa_node according to new topology (this
2161 * involves free and re-allocate memory, worthy doing?)
2164 blk_mq_map_swqueue(q, online_mask);
2166 blk_mq_sysfs_register(q);
2169 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2170 unsigned long action, void *hcpu)
2172 struct request_queue *q;
2173 int cpu = (unsigned long)hcpu;
2175 * New online cpumask which is going to be set in this hotplug event.
2176 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2177 * one-by-one and dynamically allocating this could result in a failure.
2179 static struct cpumask online_new;
2182 * Before hotadded cpu starts handling requests, new mappings must
2183 * be established. Otherwise, these requests in hw queue might
2184 * never be dispatched.
2186 * For example, there is a single hw queue (hctx) and two CPU queues
2187 * (ctx0 for CPU0, and ctx1 for CPU1).
2189 * Now CPU1 is just onlined and a request is inserted into
2190 * ctx1->rq_list and set bit0 in pending bitmap as ctx1->index_hw is
2193 * And then while running hw queue, flush_busy_ctxs() finds bit0 is
2194 * set in pending bitmap and tries to retrieve requests in
2195 * hctx->ctxs[0]->rq_list. But htx->ctxs[0] is a pointer to ctx0,
2196 * so the request in ctx1->rq_list is ignored.
2198 switch (action & ~CPU_TASKS_FROZEN) {
2200 case CPU_UP_CANCELED:
2201 cpumask_copy(&online_new, cpu_online_mask);
2203 case CPU_UP_PREPARE:
2204 cpumask_copy(&online_new, cpu_online_mask);
2205 cpumask_set_cpu(cpu, &online_new);
2211 mutex_lock(&all_q_mutex);
2214 * We need to freeze and reinit all existing queues. Freezing
2215 * involves synchronous wait for an RCU grace period and doing it
2216 * one by one may take a long time. Start freezing all queues in
2217 * one swoop and then wait for the completions so that freezing can
2218 * take place in parallel.
2220 list_for_each_entry(q, &all_q_list, all_q_node)
2221 blk_mq_freeze_queue_start(q);
2222 list_for_each_entry(q, &all_q_list, all_q_node) {
2223 blk_mq_freeze_queue_wait(q);
2226 * timeout handler can't touch hw queue during the
2229 del_timer_sync(&q->timeout);
2232 list_for_each_entry(q, &all_q_list, all_q_node)
2233 blk_mq_queue_reinit(q, &online_new);
2235 list_for_each_entry(q, &all_q_list, all_q_node)
2236 blk_mq_unfreeze_queue(q);
2238 mutex_unlock(&all_q_mutex);
2242 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2246 for (i = 0; i < set->nr_hw_queues; i++) {
2247 set->tags[i] = blk_mq_init_rq_map(set, i);
2256 blk_mq_free_rq_map(set, set->tags[i], i);
2262 * Allocate the request maps associated with this tag_set. Note that this
2263 * may reduce the depth asked for, if memory is tight. set->queue_depth
2264 * will be updated to reflect the allocated depth.
2266 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2271 depth = set->queue_depth;
2273 err = __blk_mq_alloc_rq_maps(set);
2277 set->queue_depth >>= 1;
2278 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2282 } while (set->queue_depth);
2284 if (!set->queue_depth || err) {
2285 pr_err("blk-mq: failed to allocate request map\n");
2289 if (depth != set->queue_depth)
2290 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2291 depth, set->queue_depth);
2296 struct cpumask *blk_mq_tags_cpumask(struct blk_mq_tags *tags)
2298 return tags->cpumask;
2300 EXPORT_SYMBOL_GPL(blk_mq_tags_cpumask);
2303 * Alloc a tag set to be associated with one or more request queues.
2304 * May fail with EINVAL for various error conditions. May adjust the
2305 * requested depth down, if if it too large. In that case, the set
2306 * value will be stored in set->queue_depth.
2308 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2310 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2312 if (!set->nr_hw_queues)
2314 if (!set->queue_depth)
2316 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2319 if (!set->ops->queue_rq || !set->ops->map_queue)
2322 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2323 pr_info("blk-mq: reduced tag depth to %u\n",
2325 set->queue_depth = BLK_MQ_MAX_DEPTH;
2329 * If a crashdump is active, then we are potentially in a very
2330 * memory constrained environment. Limit us to 1 queue and
2331 * 64 tags to prevent using too much memory.
2333 if (is_kdump_kernel()) {
2334 set->nr_hw_queues = 1;
2335 set->queue_depth = min(64U, set->queue_depth);
2338 * There is no use for more h/w queues than cpus.
2340 if (set->nr_hw_queues > nr_cpu_ids)
2341 set->nr_hw_queues = nr_cpu_ids;
2343 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2344 GFP_KERNEL, set->numa_node);
2348 if (blk_mq_alloc_rq_maps(set))
2351 mutex_init(&set->tag_list_lock);
2352 INIT_LIST_HEAD(&set->tag_list);
2360 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2362 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2366 for (i = 0; i < nr_cpu_ids; i++) {
2368 blk_mq_free_rq_map(set, set->tags[i], i);
2374 EXPORT_SYMBOL(blk_mq_free_tag_set);
2376 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2378 struct blk_mq_tag_set *set = q->tag_set;
2379 struct blk_mq_hw_ctx *hctx;
2382 if (!set || nr > set->queue_depth)
2386 queue_for_each_hw_ctx(q, hctx, i) {
2389 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2395 q->nr_requests = nr;
2400 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2402 struct request_queue *q;
2404 if (nr_hw_queues > nr_cpu_ids)
2405 nr_hw_queues = nr_cpu_ids;
2406 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2409 list_for_each_entry(q, &set->tag_list, tag_set_list)
2410 blk_mq_freeze_queue(q);
2412 set->nr_hw_queues = nr_hw_queues;
2413 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2414 blk_mq_realloc_hw_ctxs(set, q);
2416 if (q->nr_hw_queues > 1)
2417 blk_queue_make_request(q, blk_mq_make_request);
2419 blk_queue_make_request(q, blk_sq_make_request);
2421 blk_mq_queue_reinit(q, cpu_online_mask);
2424 list_for_each_entry(q, &set->tag_list, tag_set_list)
2425 blk_mq_unfreeze_queue(q);
2427 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2429 void blk_mq_disable_hotplug(void)
2431 mutex_lock(&all_q_mutex);
2434 void blk_mq_enable_hotplug(void)
2436 mutex_unlock(&all_q_mutex);
2439 static int __init blk_mq_init(void)
2443 hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2447 subsys_initcall(blk_mq_init);