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>
24 #include <trace/events/block.h>
26 #include <linux/blk-mq.h>
29 #include "blk-mq-tag.h"
31 static DEFINE_MUTEX(all_q_mutex);
32 static LIST_HEAD(all_q_list);
34 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
37 * Check if any of the ctx's have pending work in this hardware queue
39 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
43 for (i = 0; i < hctx->ctx_map.map_size; i++)
44 if (hctx->ctx_map.map[i].word)
50 static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx,
51 struct blk_mq_ctx *ctx)
53 return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word];
56 #define CTX_TO_BIT(hctx, ctx) \
57 ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
60 * Mark this ctx as having pending work in this hardware queue
62 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
63 struct blk_mq_ctx *ctx)
65 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
67 if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word))
68 set_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
71 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
72 struct blk_mq_ctx *ctx)
74 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
76 clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
79 static int blk_mq_queue_enter(struct request_queue *q)
84 if (percpu_ref_tryget_live(&q->mq_usage_counter))
87 ret = wait_event_interruptible(q->mq_freeze_wq,
88 !q->mq_freeze_depth || blk_queue_dying(q));
89 if (blk_queue_dying(q))
96 static void blk_mq_queue_exit(struct request_queue *q)
98 percpu_ref_put(&q->mq_usage_counter);
101 static void blk_mq_usage_counter_release(struct percpu_ref *ref)
103 struct request_queue *q =
104 container_of(ref, struct request_queue, mq_usage_counter);
106 wake_up_all(&q->mq_freeze_wq);
110 * Guarantee no request is in use, so we can change any data structure of
111 * the queue afterward.
113 void blk_mq_freeze_queue(struct request_queue *q)
117 spin_lock_irq(q->queue_lock);
118 freeze = !q->mq_freeze_depth++;
119 spin_unlock_irq(q->queue_lock);
122 percpu_ref_kill(&q->mq_usage_counter);
123 blk_mq_run_queues(q, false);
125 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->mq_usage_counter));
128 static void blk_mq_unfreeze_queue(struct request_queue *q)
132 spin_lock_irq(q->queue_lock);
133 wake = !--q->mq_freeze_depth;
134 WARN_ON_ONCE(q->mq_freeze_depth < 0);
135 spin_unlock_irq(q->queue_lock);
137 percpu_ref_reinit(&q->mq_usage_counter);
138 wake_up_all(&q->mq_freeze_wq);
142 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
144 return blk_mq_has_free_tags(hctx->tags);
146 EXPORT_SYMBOL(blk_mq_can_queue);
148 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
149 struct request *rq, unsigned int rw_flags)
151 if (blk_queue_io_stat(q))
152 rw_flags |= REQ_IO_STAT;
154 INIT_LIST_HEAD(&rq->queuelist);
155 /* csd/requeue_work/fifo_time is initialized before use */
158 rq->cmd_flags |= rw_flags;
159 /* do not touch atomic flags, it needs atomic ops against the timer */
161 INIT_HLIST_NODE(&rq->hash);
162 RB_CLEAR_NODE(&rq->rb_node);
165 rq->start_time = jiffies;
166 #ifdef CONFIG_BLK_CGROUP
168 set_start_time_ns(rq);
169 rq->io_start_time_ns = 0;
171 rq->nr_phys_segments = 0;
172 #if defined(CONFIG_BLK_DEV_INTEGRITY)
173 rq->nr_integrity_segments = 0;
176 /* tag was already set */
186 INIT_LIST_HEAD(&rq->timeout_list);
190 rq->end_io_data = NULL;
193 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
196 static struct request *
197 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
202 tag = blk_mq_get_tag(data);
203 if (tag != BLK_MQ_TAG_FAIL) {
204 rq = data->hctx->tags->rqs[tag];
206 if (blk_mq_tag_busy(data->hctx)) {
207 rq->cmd_flags = REQ_MQ_INFLIGHT;
208 atomic_inc(&data->hctx->nr_active);
212 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
219 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
222 struct blk_mq_ctx *ctx;
223 struct blk_mq_hw_ctx *hctx;
225 struct blk_mq_alloc_data alloc_data;
228 ret = blk_mq_queue_enter(q);
232 ctx = blk_mq_get_ctx(q);
233 hctx = q->mq_ops->map_queue(q, ctx->cpu);
234 blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT,
235 reserved, ctx, hctx);
237 rq = __blk_mq_alloc_request(&alloc_data, rw);
238 if (!rq && (gfp & __GFP_WAIT)) {
239 __blk_mq_run_hw_queue(hctx);
242 ctx = blk_mq_get_ctx(q);
243 hctx = q->mq_ops->map_queue(q, ctx->cpu);
244 blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx,
246 rq = __blk_mq_alloc_request(&alloc_data, rw);
247 ctx = alloc_data.ctx;
251 return ERR_PTR(-EWOULDBLOCK);
254 EXPORT_SYMBOL(blk_mq_alloc_request);
256 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
257 struct blk_mq_ctx *ctx, struct request *rq)
259 const int tag = rq->tag;
260 struct request_queue *q = rq->q;
262 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
263 atomic_dec(&hctx->nr_active);
266 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
267 blk_mq_put_tag(hctx, tag, &ctx->last_tag);
268 blk_mq_queue_exit(q);
271 void blk_mq_free_request(struct request *rq)
273 struct blk_mq_ctx *ctx = rq->mq_ctx;
274 struct blk_mq_hw_ctx *hctx;
275 struct request_queue *q = rq->q;
277 ctx->rq_completed[rq_is_sync(rq)]++;
279 hctx = q->mq_ops->map_queue(q, ctx->cpu);
280 __blk_mq_free_request(hctx, ctx, rq);
284 * Clone all relevant state from a request that has been put on hold in
285 * the flush state machine into the preallocated flush request that hangs
286 * off the request queue.
288 * For a driver the flush request should be invisible, that's why we are
289 * impersonating the original request here.
291 void blk_mq_clone_flush_request(struct request *flush_rq,
292 struct request *orig_rq)
294 struct blk_mq_hw_ctx *hctx =
295 orig_rq->q->mq_ops->map_queue(orig_rq->q, orig_rq->mq_ctx->cpu);
297 flush_rq->mq_ctx = orig_rq->mq_ctx;
298 flush_rq->tag = orig_rq->tag;
299 memcpy(blk_mq_rq_to_pdu(flush_rq), blk_mq_rq_to_pdu(orig_rq),
303 inline void __blk_mq_end_io(struct request *rq, int error)
305 blk_account_io_done(rq);
308 rq->end_io(rq, error);
310 if (unlikely(blk_bidi_rq(rq)))
311 blk_mq_free_request(rq->next_rq);
312 blk_mq_free_request(rq);
315 EXPORT_SYMBOL(__blk_mq_end_io);
317 void blk_mq_end_io(struct request *rq, int error)
319 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
321 __blk_mq_end_io(rq, error);
323 EXPORT_SYMBOL(blk_mq_end_io);
325 static void __blk_mq_complete_request_remote(void *data)
327 struct request *rq = data;
329 rq->q->softirq_done_fn(rq);
332 static void blk_mq_ipi_complete_request(struct request *rq)
334 struct blk_mq_ctx *ctx = rq->mq_ctx;
338 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
339 rq->q->softirq_done_fn(rq);
344 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
345 shared = cpus_share_cache(cpu, ctx->cpu);
347 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
348 rq->csd.func = __blk_mq_complete_request_remote;
351 smp_call_function_single_async(ctx->cpu, &rq->csd);
353 rq->q->softirq_done_fn(rq);
358 void __blk_mq_complete_request(struct request *rq)
360 struct request_queue *q = rq->q;
362 if (!q->softirq_done_fn)
363 blk_mq_end_io(rq, rq->errors);
365 blk_mq_ipi_complete_request(rq);
369 * blk_mq_complete_request - end I/O on a request
370 * @rq: the request being processed
373 * Ends all I/O on a request. It does not handle partial completions.
374 * The actual completion happens out-of-order, through a IPI handler.
376 void blk_mq_complete_request(struct request *rq)
378 struct request_queue *q = rq->q;
380 if (unlikely(blk_should_fake_timeout(q)))
382 if (!blk_mark_rq_complete(rq))
383 __blk_mq_complete_request(rq);
385 EXPORT_SYMBOL(blk_mq_complete_request);
387 void blk_mq_start_request(struct request *rq)
389 struct request_queue *q = rq->q;
391 trace_block_rq_issue(q, rq);
393 rq->resid_len = blk_rq_bytes(rq);
394 if (unlikely(blk_bidi_rq(rq)))
395 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
400 * Ensure that ->deadline is visible before set the started
401 * flag and clear the completed flag.
403 smp_mb__before_atomic();
406 * Mark us as started and clear complete. Complete might have been
407 * set if requeue raced with timeout, which then marked it as
408 * complete. So be sure to clear complete again when we start
409 * the request, otherwise we'll ignore the completion event.
411 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
412 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
413 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
414 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
416 if (q->dma_drain_size && blk_rq_bytes(rq)) {
418 * Make sure space for the drain appears. We know we can do
419 * this because max_hw_segments has been adjusted to be one
420 * fewer than the device can handle.
422 rq->nr_phys_segments++;
425 EXPORT_SYMBOL(blk_mq_start_request);
427 static void __blk_mq_requeue_request(struct request *rq)
429 struct request_queue *q = rq->q;
431 trace_block_rq_requeue(q, rq);
433 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
434 if (q->dma_drain_size && blk_rq_bytes(rq))
435 rq->nr_phys_segments--;
439 void blk_mq_requeue_request(struct request *rq)
441 __blk_mq_requeue_request(rq);
442 blk_clear_rq_complete(rq);
444 BUG_ON(blk_queued_rq(rq));
445 blk_mq_add_to_requeue_list(rq, true);
447 EXPORT_SYMBOL(blk_mq_requeue_request);
449 static void blk_mq_requeue_work(struct work_struct *work)
451 struct request_queue *q =
452 container_of(work, struct request_queue, requeue_work);
454 struct request *rq, *next;
457 spin_lock_irqsave(&q->requeue_lock, flags);
458 list_splice_init(&q->requeue_list, &rq_list);
459 spin_unlock_irqrestore(&q->requeue_lock, flags);
461 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
462 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
465 rq->cmd_flags &= ~REQ_SOFTBARRIER;
466 list_del_init(&rq->queuelist);
467 blk_mq_insert_request(rq, true, false, false);
470 while (!list_empty(&rq_list)) {
471 rq = list_entry(rq_list.next, struct request, queuelist);
472 list_del_init(&rq->queuelist);
473 blk_mq_insert_request(rq, false, false, false);
477 * Use the start variant of queue running here, so that running
478 * the requeue work will kick stopped queues.
480 blk_mq_start_hw_queues(q);
483 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
485 struct request_queue *q = rq->q;
489 * We abuse this flag that is otherwise used by the I/O scheduler to
490 * request head insertation from the workqueue.
492 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
494 spin_lock_irqsave(&q->requeue_lock, flags);
496 rq->cmd_flags |= REQ_SOFTBARRIER;
497 list_add(&rq->queuelist, &q->requeue_list);
499 list_add_tail(&rq->queuelist, &q->requeue_list);
501 spin_unlock_irqrestore(&q->requeue_lock, flags);
503 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
505 void blk_mq_kick_requeue_list(struct request_queue *q)
507 kblockd_schedule_work(&q->requeue_work);
509 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
511 static inline bool is_flush_request(struct request *rq, unsigned int tag)
513 return ((rq->cmd_flags & REQ_FLUSH_SEQ) &&
514 rq->q->flush_rq->tag == tag);
517 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
519 struct request *rq = tags->rqs[tag];
521 if (!is_flush_request(rq, tag))
524 return rq->q->flush_rq;
526 EXPORT_SYMBOL(blk_mq_tag_to_rq);
528 struct blk_mq_timeout_data {
529 struct blk_mq_hw_ctx *hctx;
531 unsigned int *next_set;
534 static void blk_mq_timeout_check(void *__data, unsigned long *free_tags)
536 struct blk_mq_timeout_data *data = __data;
537 struct blk_mq_hw_ctx *hctx = data->hctx;
540 /* It may not be in flight yet (this is where
541 * the REQ_ATOMIC_STARTED flag comes in). The requests are
542 * statically allocated, so we know it's always safe to access the
543 * memory associated with a bit offset into ->rqs[].
549 tag = find_next_zero_bit(free_tags, hctx->tags->nr_tags, tag);
550 if (tag >= hctx->tags->nr_tags)
553 rq = blk_mq_tag_to_rq(hctx->tags, tag++);
554 if (rq->q != hctx->queue)
556 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
559 blk_rq_check_expired(rq, data->next, data->next_set);
563 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx,
565 unsigned int *next_set)
567 struct blk_mq_timeout_data data = {
570 .next_set = next_set,
574 * Ask the tagging code to iterate busy requests, so we can
575 * check them for timeout.
577 blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data);
580 static enum blk_eh_timer_return blk_mq_rq_timed_out(struct request *rq)
582 struct request_queue *q = rq->q;
585 * We know that complete is set at this point. If STARTED isn't set
586 * anymore, then the request isn't active and the "timeout" should
587 * just be ignored. This can happen due to the bitflag ordering.
588 * Timeout first checks if STARTED is set, and if it is, assumes
589 * the request is active. But if we race with completion, then
590 * we both flags will get cleared. So check here again, and ignore
591 * a timeout event with a request that isn't active.
593 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
594 return BLK_EH_NOT_HANDLED;
596 if (!q->mq_ops->timeout)
597 return BLK_EH_RESET_TIMER;
599 return q->mq_ops->timeout(rq);
602 static void blk_mq_rq_timer(unsigned long data)
604 struct request_queue *q = (struct request_queue *) data;
605 struct blk_mq_hw_ctx *hctx;
606 unsigned long next = 0;
609 queue_for_each_hw_ctx(q, hctx, i) {
611 * If not software queues are currently mapped to this
612 * hardware queue, there's nothing to check
614 if (!hctx->nr_ctx || !hctx->tags)
617 blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set);
621 next = blk_rq_timeout(round_jiffies_up(next));
622 mod_timer(&q->timeout, next);
624 queue_for_each_hw_ctx(q, hctx, i)
625 blk_mq_tag_idle(hctx);
630 * Reverse check our software queue for entries that we could potentially
631 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
632 * too much time checking for merges.
634 static bool blk_mq_attempt_merge(struct request_queue *q,
635 struct blk_mq_ctx *ctx, struct bio *bio)
640 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
646 if (!blk_rq_merge_ok(rq, bio))
649 el_ret = blk_try_merge(rq, bio);
650 if (el_ret == ELEVATOR_BACK_MERGE) {
651 if (bio_attempt_back_merge(q, rq, bio)) {
656 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
657 if (bio_attempt_front_merge(q, rq, bio)) {
669 * Process software queues that have been marked busy, splicing them
670 * to the for-dispatch
672 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
674 struct blk_mq_ctx *ctx;
677 for (i = 0; i < hctx->ctx_map.map_size; i++) {
678 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
679 unsigned int off, bit;
685 off = i * hctx->ctx_map.bits_per_word;
687 bit = find_next_bit(&bm->word, bm->depth, bit);
688 if (bit >= bm->depth)
691 ctx = hctx->ctxs[bit + off];
692 clear_bit(bit, &bm->word);
693 spin_lock(&ctx->lock);
694 list_splice_tail_init(&ctx->rq_list, list);
695 spin_unlock(&ctx->lock);
703 * Run this hardware queue, pulling any software queues mapped to it in.
704 * Note that this function currently has various problems around ordering
705 * of IO. In particular, we'd like FIFO behaviour on handling existing
706 * items on the hctx->dispatch list. Ignore that for now.
708 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
710 struct request_queue *q = hctx->queue;
715 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
717 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
723 * Touch any software queue that has pending entries.
725 flush_busy_ctxs(hctx, &rq_list);
728 * If we have previous entries on our dispatch list, grab them
729 * and stuff them at the front for more fair dispatch.
731 if (!list_empty_careful(&hctx->dispatch)) {
732 spin_lock(&hctx->lock);
733 if (!list_empty(&hctx->dispatch))
734 list_splice_init(&hctx->dispatch, &rq_list);
735 spin_unlock(&hctx->lock);
739 * Now process all the entries, sending them to the driver.
742 while (!list_empty(&rq_list)) {
745 rq = list_first_entry(&rq_list, struct request, queuelist);
746 list_del_init(&rq->queuelist);
748 ret = q->mq_ops->queue_rq(hctx, rq, list_empty(&rq_list));
750 case BLK_MQ_RQ_QUEUE_OK:
753 case BLK_MQ_RQ_QUEUE_BUSY:
754 list_add(&rq->queuelist, &rq_list);
755 __blk_mq_requeue_request(rq);
758 pr_err("blk-mq: bad return on queue: %d\n", ret);
759 case BLK_MQ_RQ_QUEUE_ERROR:
761 blk_mq_end_io(rq, rq->errors);
765 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
770 hctx->dispatched[0]++;
771 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
772 hctx->dispatched[ilog2(queued) + 1]++;
775 * Any items that need requeuing? Stuff them into hctx->dispatch,
776 * that is where we will continue on next queue run.
778 if (!list_empty(&rq_list)) {
779 spin_lock(&hctx->lock);
780 list_splice(&rq_list, &hctx->dispatch);
781 spin_unlock(&hctx->lock);
786 * It'd be great if the workqueue API had a way to pass
787 * in a mask and had some smarts for more clever placement.
788 * For now we just round-robin here, switching for every
789 * BLK_MQ_CPU_WORK_BATCH queued items.
791 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
793 int cpu = hctx->next_cpu;
795 if (--hctx->next_cpu_batch <= 0) {
798 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
799 if (next_cpu >= nr_cpu_ids)
800 next_cpu = cpumask_first(hctx->cpumask);
802 hctx->next_cpu = next_cpu;
803 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
809 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
811 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
814 if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask))
815 __blk_mq_run_hw_queue(hctx);
816 else if (hctx->queue->nr_hw_queues == 1)
817 kblockd_schedule_delayed_work(&hctx->run_work, 0);
821 cpu = blk_mq_hctx_next_cpu(hctx);
822 kblockd_schedule_delayed_work_on(cpu, &hctx->run_work, 0);
826 void blk_mq_run_queues(struct request_queue *q, bool async)
828 struct blk_mq_hw_ctx *hctx;
831 queue_for_each_hw_ctx(q, hctx, i) {
832 if ((!blk_mq_hctx_has_pending(hctx) &&
833 list_empty_careful(&hctx->dispatch)) ||
834 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
838 blk_mq_run_hw_queue(hctx, async);
842 EXPORT_SYMBOL(blk_mq_run_queues);
844 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
846 cancel_delayed_work(&hctx->run_work);
847 cancel_delayed_work(&hctx->delay_work);
848 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
850 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
852 void blk_mq_stop_hw_queues(struct request_queue *q)
854 struct blk_mq_hw_ctx *hctx;
857 queue_for_each_hw_ctx(q, hctx, i)
858 blk_mq_stop_hw_queue(hctx);
860 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
862 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
864 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
867 blk_mq_run_hw_queue(hctx, false);
870 EXPORT_SYMBOL(blk_mq_start_hw_queue);
872 void blk_mq_start_hw_queues(struct request_queue *q)
874 struct blk_mq_hw_ctx *hctx;
877 queue_for_each_hw_ctx(q, hctx, i)
878 blk_mq_start_hw_queue(hctx);
880 EXPORT_SYMBOL(blk_mq_start_hw_queues);
883 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
885 struct blk_mq_hw_ctx *hctx;
888 queue_for_each_hw_ctx(q, hctx, i) {
889 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
892 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
894 blk_mq_run_hw_queue(hctx, async);
898 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
900 static void blk_mq_run_work_fn(struct work_struct *work)
902 struct blk_mq_hw_ctx *hctx;
904 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
906 __blk_mq_run_hw_queue(hctx);
909 static void blk_mq_delay_work_fn(struct work_struct *work)
911 struct blk_mq_hw_ctx *hctx;
913 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
915 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
916 __blk_mq_run_hw_queue(hctx);
919 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
921 unsigned long tmo = msecs_to_jiffies(msecs);
923 if (hctx->queue->nr_hw_queues == 1)
924 kblockd_schedule_delayed_work(&hctx->delay_work, tmo);
928 cpu = blk_mq_hctx_next_cpu(hctx);
929 kblockd_schedule_delayed_work_on(cpu, &hctx->delay_work, tmo);
932 EXPORT_SYMBOL(blk_mq_delay_queue);
934 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
935 struct request *rq, bool at_head)
937 struct blk_mq_ctx *ctx = rq->mq_ctx;
939 trace_block_rq_insert(hctx->queue, rq);
942 list_add(&rq->queuelist, &ctx->rq_list);
944 list_add_tail(&rq->queuelist, &ctx->rq_list);
946 blk_mq_hctx_mark_pending(hctx, ctx);
949 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
952 struct request_queue *q = rq->q;
953 struct blk_mq_hw_ctx *hctx;
954 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
956 current_ctx = blk_mq_get_ctx(q);
957 if (!cpu_online(ctx->cpu))
958 rq->mq_ctx = ctx = current_ctx;
960 hctx = q->mq_ops->map_queue(q, ctx->cpu);
962 spin_lock(&ctx->lock);
963 __blk_mq_insert_request(hctx, rq, at_head);
964 spin_unlock(&ctx->lock);
967 blk_mq_run_hw_queue(hctx, async);
969 blk_mq_put_ctx(current_ctx);
972 static void blk_mq_insert_requests(struct request_queue *q,
973 struct blk_mq_ctx *ctx,
974 struct list_head *list,
979 struct blk_mq_hw_ctx *hctx;
980 struct blk_mq_ctx *current_ctx;
982 trace_block_unplug(q, depth, !from_schedule);
984 current_ctx = blk_mq_get_ctx(q);
986 if (!cpu_online(ctx->cpu))
988 hctx = q->mq_ops->map_queue(q, ctx->cpu);
991 * preemption doesn't flush plug list, so it's possible ctx->cpu is
994 spin_lock(&ctx->lock);
995 while (!list_empty(list)) {
998 rq = list_first_entry(list, struct request, queuelist);
999 list_del_init(&rq->queuelist);
1001 __blk_mq_insert_request(hctx, rq, false);
1003 spin_unlock(&ctx->lock);
1005 blk_mq_run_hw_queue(hctx, from_schedule);
1006 blk_mq_put_ctx(current_ctx);
1009 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1011 struct request *rqa = container_of(a, struct request, queuelist);
1012 struct request *rqb = container_of(b, struct request, queuelist);
1014 return !(rqa->mq_ctx < rqb->mq_ctx ||
1015 (rqa->mq_ctx == rqb->mq_ctx &&
1016 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1019 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1021 struct blk_mq_ctx *this_ctx;
1022 struct request_queue *this_q;
1025 LIST_HEAD(ctx_list);
1028 list_splice_init(&plug->mq_list, &list);
1030 list_sort(NULL, &list, plug_ctx_cmp);
1036 while (!list_empty(&list)) {
1037 rq = list_entry_rq(list.next);
1038 list_del_init(&rq->queuelist);
1040 if (rq->mq_ctx != this_ctx) {
1042 blk_mq_insert_requests(this_q, this_ctx,
1047 this_ctx = rq->mq_ctx;
1053 list_add_tail(&rq->queuelist, &ctx_list);
1057 * If 'this_ctx' is set, we know we have entries to complete
1058 * on 'ctx_list'. Do those.
1061 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1066 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1068 init_request_from_bio(rq, bio);
1070 if (blk_do_io_stat(rq))
1071 blk_account_io_start(rq, 1);
1074 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1076 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1077 !blk_queue_nomerges(hctx->queue);
1080 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1081 struct blk_mq_ctx *ctx,
1082 struct request *rq, struct bio *bio)
1084 if (!hctx_allow_merges(hctx)) {
1085 blk_mq_bio_to_request(rq, bio);
1086 spin_lock(&ctx->lock);
1088 __blk_mq_insert_request(hctx, rq, false);
1089 spin_unlock(&ctx->lock);
1092 struct request_queue *q = hctx->queue;
1094 spin_lock(&ctx->lock);
1095 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1096 blk_mq_bio_to_request(rq, bio);
1100 spin_unlock(&ctx->lock);
1101 __blk_mq_free_request(hctx, ctx, rq);
1106 struct blk_map_ctx {
1107 struct blk_mq_hw_ctx *hctx;
1108 struct blk_mq_ctx *ctx;
1111 static struct request *blk_mq_map_request(struct request_queue *q,
1113 struct blk_map_ctx *data)
1115 struct blk_mq_hw_ctx *hctx;
1116 struct blk_mq_ctx *ctx;
1118 int rw = bio_data_dir(bio);
1119 struct blk_mq_alloc_data alloc_data;
1121 if (unlikely(blk_mq_queue_enter(q))) {
1122 bio_endio(bio, -EIO);
1126 ctx = blk_mq_get_ctx(q);
1127 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1129 if (rw_is_sync(bio->bi_rw))
1132 trace_block_getrq(q, bio, rw);
1133 blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1135 rq = __blk_mq_alloc_request(&alloc_data, rw);
1136 if (unlikely(!rq)) {
1137 __blk_mq_run_hw_queue(hctx);
1138 blk_mq_put_ctx(ctx);
1139 trace_block_sleeprq(q, bio, rw);
1141 ctx = blk_mq_get_ctx(q);
1142 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1143 blk_mq_set_alloc_data(&alloc_data, q,
1144 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1145 rq = __blk_mq_alloc_request(&alloc_data, rw);
1146 ctx = alloc_data.ctx;
1147 hctx = alloc_data.hctx;
1157 * Multiple hardware queue variant. This will not use per-process plugs,
1158 * but will attempt to bypass the hctx queueing if we can go straight to
1159 * hardware for SYNC IO.
1161 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1163 const int is_sync = rw_is_sync(bio->bi_rw);
1164 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1165 struct blk_map_ctx data;
1168 blk_queue_bounce(q, &bio);
1170 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1171 bio_endio(bio, -EIO);
1175 rq = blk_mq_map_request(q, bio, &data);
1179 if (unlikely(is_flush_fua)) {
1180 blk_mq_bio_to_request(rq, bio);
1181 blk_insert_flush(rq);
1188 blk_mq_bio_to_request(rq, bio);
1191 * For OK queue, we are done. For error, kill it. Any other
1192 * error (busy), just add it to our list as we previously
1195 ret = q->mq_ops->queue_rq(data.hctx, rq, true);
1196 if (ret == BLK_MQ_RQ_QUEUE_OK)
1199 __blk_mq_requeue_request(rq);
1201 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1203 blk_mq_end_io(rq, rq->errors);
1209 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1211 * For a SYNC request, send it to the hardware immediately. For
1212 * an ASYNC request, just ensure that we run it later on. The
1213 * latter allows for merging opportunities and more efficient
1217 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1220 blk_mq_put_ctx(data.ctx);
1224 * Single hardware queue variant. This will attempt to use any per-process
1225 * plug for merging and IO deferral.
1227 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1229 const int is_sync = rw_is_sync(bio->bi_rw);
1230 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1231 unsigned int use_plug, request_count = 0;
1232 struct blk_map_ctx data;
1236 * If we have multiple hardware queues, just go directly to
1237 * one of those for sync IO.
1239 use_plug = !is_flush_fua && !is_sync;
1241 blk_queue_bounce(q, &bio);
1243 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1244 bio_endio(bio, -EIO);
1248 if (use_plug && !blk_queue_nomerges(q) &&
1249 blk_attempt_plug_merge(q, bio, &request_count))
1252 rq = blk_mq_map_request(q, bio, &data);
1256 if (unlikely(is_flush_fua)) {
1257 blk_mq_bio_to_request(rq, bio);
1258 blk_insert_flush(rq);
1263 * A task plug currently exists. Since this is completely lockless,
1264 * utilize that to temporarily store requests until the task is
1265 * either done or scheduled away.
1268 struct blk_plug *plug = current->plug;
1271 blk_mq_bio_to_request(rq, bio);
1272 if (list_empty(&plug->mq_list))
1273 trace_block_plug(q);
1274 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1275 blk_flush_plug_list(plug, false);
1276 trace_block_plug(q);
1278 list_add_tail(&rq->queuelist, &plug->mq_list);
1279 blk_mq_put_ctx(data.ctx);
1284 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1286 * For a SYNC request, send it to the hardware immediately. For
1287 * an ASYNC request, just ensure that we run it later on. The
1288 * latter allows for merging opportunities and more efficient
1292 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1295 blk_mq_put_ctx(data.ctx);
1299 * Default mapping to a software queue, since we use one per CPU.
1301 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1303 return q->queue_hw_ctx[q->mq_map[cpu]];
1305 EXPORT_SYMBOL(blk_mq_map_queue);
1307 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1308 struct blk_mq_tags *tags, unsigned int hctx_idx)
1312 if (tags->rqs && set->ops->exit_request) {
1315 for (i = 0; i < tags->nr_tags; i++) {
1318 set->ops->exit_request(set->driver_data, tags->rqs[i],
1320 tags->rqs[i] = NULL;
1324 while (!list_empty(&tags->page_list)) {
1325 page = list_first_entry(&tags->page_list, struct page, lru);
1326 list_del_init(&page->lru);
1327 __free_pages(page, page->private);
1332 blk_mq_free_tags(tags);
1335 static size_t order_to_size(unsigned int order)
1337 return (size_t)PAGE_SIZE << order;
1340 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1341 unsigned int hctx_idx)
1343 struct blk_mq_tags *tags;
1344 unsigned int i, j, entries_per_page, max_order = 4;
1345 size_t rq_size, left;
1347 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1352 INIT_LIST_HEAD(&tags->page_list);
1354 tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1355 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1358 blk_mq_free_tags(tags);
1363 * rq_size is the size of the request plus driver payload, rounded
1364 * to the cacheline size
1366 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1368 left = rq_size * set->queue_depth;
1370 for (i = 0; i < set->queue_depth; ) {
1371 int this_order = max_order;
1376 while (left < order_to_size(this_order - 1) && this_order)
1380 page = alloc_pages_node(set->numa_node,
1381 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1387 if (order_to_size(this_order) < rq_size)
1394 page->private = this_order;
1395 list_add_tail(&page->lru, &tags->page_list);
1397 p = page_address(page);
1398 entries_per_page = order_to_size(this_order) / rq_size;
1399 to_do = min(entries_per_page, set->queue_depth - i);
1400 left -= to_do * rq_size;
1401 for (j = 0; j < to_do; j++) {
1403 tags->rqs[i]->atomic_flags = 0;
1404 tags->rqs[i]->cmd_flags = 0;
1405 if (set->ops->init_request) {
1406 if (set->ops->init_request(set->driver_data,
1407 tags->rqs[i], hctx_idx, i,
1409 tags->rqs[i] = NULL;
1422 blk_mq_free_rq_map(set, tags, hctx_idx);
1426 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1431 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1433 unsigned int bpw = 8, total, num_maps, i;
1435 bitmap->bits_per_word = bpw;
1437 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1438 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1443 bitmap->map_size = num_maps;
1446 for (i = 0; i < num_maps; i++) {
1447 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1448 total -= bitmap->map[i].depth;
1454 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1456 struct request_queue *q = hctx->queue;
1457 struct blk_mq_ctx *ctx;
1461 * Move ctx entries to new CPU, if this one is going away.
1463 ctx = __blk_mq_get_ctx(q, cpu);
1465 spin_lock(&ctx->lock);
1466 if (!list_empty(&ctx->rq_list)) {
1467 list_splice_init(&ctx->rq_list, &tmp);
1468 blk_mq_hctx_clear_pending(hctx, ctx);
1470 spin_unlock(&ctx->lock);
1472 if (list_empty(&tmp))
1475 ctx = blk_mq_get_ctx(q);
1476 spin_lock(&ctx->lock);
1478 while (!list_empty(&tmp)) {
1481 rq = list_first_entry(&tmp, struct request, queuelist);
1483 list_move_tail(&rq->queuelist, &ctx->rq_list);
1486 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1487 blk_mq_hctx_mark_pending(hctx, ctx);
1489 spin_unlock(&ctx->lock);
1491 blk_mq_run_hw_queue(hctx, true);
1492 blk_mq_put_ctx(ctx);
1496 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1498 struct request_queue *q = hctx->queue;
1499 struct blk_mq_tag_set *set = q->tag_set;
1501 if (set->tags[hctx->queue_num])
1504 set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1505 if (!set->tags[hctx->queue_num])
1508 hctx->tags = set->tags[hctx->queue_num];
1512 static int blk_mq_hctx_notify(void *data, unsigned long action,
1515 struct blk_mq_hw_ctx *hctx = data;
1517 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1518 return blk_mq_hctx_cpu_offline(hctx, cpu);
1519 else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1520 return blk_mq_hctx_cpu_online(hctx, cpu);
1525 static void blk_mq_exit_hw_queues(struct request_queue *q,
1526 struct blk_mq_tag_set *set, int nr_queue)
1528 struct blk_mq_hw_ctx *hctx;
1531 queue_for_each_hw_ctx(q, hctx, i) {
1535 blk_mq_tag_idle(hctx);
1537 if (set->ops->exit_hctx)
1538 set->ops->exit_hctx(hctx, i);
1540 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1542 blk_mq_free_bitmap(&hctx->ctx_map);
1547 static void blk_mq_free_hw_queues(struct request_queue *q,
1548 struct blk_mq_tag_set *set)
1550 struct blk_mq_hw_ctx *hctx;
1553 queue_for_each_hw_ctx(q, hctx, i) {
1554 free_cpumask_var(hctx->cpumask);
1559 static int blk_mq_init_hw_queues(struct request_queue *q,
1560 struct blk_mq_tag_set *set)
1562 struct blk_mq_hw_ctx *hctx;
1566 * Initialize hardware queues
1568 queue_for_each_hw_ctx(q, hctx, i) {
1571 node = hctx->numa_node;
1572 if (node == NUMA_NO_NODE)
1573 node = hctx->numa_node = set->numa_node;
1575 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1576 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1577 spin_lock_init(&hctx->lock);
1578 INIT_LIST_HEAD(&hctx->dispatch);
1580 hctx->queue_num = i;
1581 hctx->flags = set->flags;
1582 hctx->cmd_size = set->cmd_size;
1584 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1585 blk_mq_hctx_notify, hctx);
1586 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1588 hctx->tags = set->tags[i];
1591 * Allocate space for all possible cpus to avoid allocation at
1594 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1599 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1604 if (set->ops->init_hctx &&
1605 set->ops->init_hctx(hctx, set->driver_data, i))
1609 if (i == q->nr_hw_queues)
1615 blk_mq_exit_hw_queues(q, set, i);
1620 static void blk_mq_init_cpu_queues(struct request_queue *q,
1621 unsigned int nr_hw_queues)
1625 for_each_possible_cpu(i) {
1626 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1627 struct blk_mq_hw_ctx *hctx;
1629 memset(__ctx, 0, sizeof(*__ctx));
1631 spin_lock_init(&__ctx->lock);
1632 INIT_LIST_HEAD(&__ctx->rq_list);
1635 /* If the cpu isn't online, the cpu is mapped to first hctx */
1639 hctx = q->mq_ops->map_queue(q, i);
1640 cpumask_set_cpu(i, hctx->cpumask);
1644 * Set local node, IFF we have more than one hw queue. If
1645 * not, we remain on the home node of the device
1647 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1648 hctx->numa_node = cpu_to_node(i);
1652 static void blk_mq_map_swqueue(struct request_queue *q)
1655 struct blk_mq_hw_ctx *hctx;
1656 struct blk_mq_ctx *ctx;
1658 queue_for_each_hw_ctx(q, hctx, i) {
1659 cpumask_clear(hctx->cpumask);
1664 * Map software to hardware queues
1666 queue_for_each_ctx(q, ctx, i) {
1667 /* If the cpu isn't online, the cpu is mapped to first hctx */
1671 hctx = q->mq_ops->map_queue(q, i);
1672 cpumask_set_cpu(i, hctx->cpumask);
1673 ctx->index_hw = hctx->nr_ctx;
1674 hctx->ctxs[hctx->nr_ctx++] = ctx;
1677 queue_for_each_hw_ctx(q, hctx, i) {
1679 * If no software queues are mapped to this hardware queue,
1680 * disable it and free the request entries.
1682 if (!hctx->nr_ctx) {
1683 struct blk_mq_tag_set *set = q->tag_set;
1686 blk_mq_free_rq_map(set, set->tags[i], i);
1687 set->tags[i] = NULL;
1694 * Initialize batch roundrobin counts
1696 hctx->next_cpu = cpumask_first(hctx->cpumask);
1697 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1701 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1703 struct blk_mq_hw_ctx *hctx;
1704 struct request_queue *q;
1708 if (set->tag_list.next == set->tag_list.prev)
1713 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1714 blk_mq_freeze_queue(q);
1716 queue_for_each_hw_ctx(q, hctx, i) {
1718 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1720 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1722 blk_mq_unfreeze_queue(q);
1726 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1728 struct blk_mq_tag_set *set = q->tag_set;
1730 mutex_lock(&set->tag_list_lock);
1731 list_del_init(&q->tag_set_list);
1732 blk_mq_update_tag_set_depth(set);
1733 mutex_unlock(&set->tag_list_lock);
1736 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1737 struct request_queue *q)
1741 mutex_lock(&set->tag_list_lock);
1742 list_add_tail(&q->tag_set_list, &set->tag_list);
1743 blk_mq_update_tag_set_depth(set);
1744 mutex_unlock(&set->tag_list_lock);
1747 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1749 struct blk_mq_hw_ctx **hctxs;
1750 struct blk_mq_ctx __percpu *ctx;
1751 struct request_queue *q;
1755 ctx = alloc_percpu(struct blk_mq_ctx);
1757 return ERR_PTR(-ENOMEM);
1759 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1765 map = blk_mq_make_queue_map(set);
1769 for (i = 0; i < set->nr_hw_queues; i++) {
1770 int node = blk_mq_hw_queue_to_node(map, i);
1772 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1777 if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL))
1780 atomic_set(&hctxs[i]->nr_active, 0);
1781 hctxs[i]->numa_node = node;
1782 hctxs[i]->queue_num = i;
1785 q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1789 if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release))
1792 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1793 blk_queue_rq_timeout(q, 30000);
1795 q->nr_queues = nr_cpu_ids;
1796 q->nr_hw_queues = set->nr_hw_queues;
1800 q->queue_hw_ctx = hctxs;
1802 q->mq_ops = set->ops;
1803 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1805 if (!(set->flags & BLK_MQ_F_SG_MERGE))
1806 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
1808 q->sg_reserved_size = INT_MAX;
1810 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1811 INIT_LIST_HEAD(&q->requeue_list);
1812 spin_lock_init(&q->requeue_lock);
1814 if (q->nr_hw_queues > 1)
1815 blk_queue_make_request(q, blk_mq_make_request);
1817 blk_queue_make_request(q, blk_sq_make_request);
1819 blk_queue_rq_timed_out(q, blk_mq_rq_timed_out);
1821 blk_queue_rq_timeout(q, set->timeout);
1824 * Do this after blk_queue_make_request() overrides it...
1826 q->nr_requests = set->queue_depth;
1828 if (set->ops->complete)
1829 blk_queue_softirq_done(q, set->ops->complete);
1831 blk_mq_init_flush(q);
1832 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1834 q->flush_rq = kzalloc(round_up(sizeof(struct request) +
1835 set->cmd_size, cache_line_size()),
1840 if (blk_mq_init_hw_queues(q, set))
1843 mutex_lock(&all_q_mutex);
1844 list_add_tail(&q->all_q_node, &all_q_list);
1845 mutex_unlock(&all_q_mutex);
1847 blk_mq_add_queue_tag_set(set, q);
1849 blk_mq_map_swqueue(q);
1856 blk_cleanup_queue(q);
1859 for (i = 0; i < set->nr_hw_queues; i++) {
1862 free_cpumask_var(hctxs[i]->cpumask);
1869 return ERR_PTR(-ENOMEM);
1871 EXPORT_SYMBOL(blk_mq_init_queue);
1873 void blk_mq_free_queue(struct request_queue *q)
1875 struct blk_mq_tag_set *set = q->tag_set;
1877 blk_mq_del_queue_tag_set(q);
1879 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
1880 blk_mq_free_hw_queues(q, set);
1882 percpu_ref_exit(&q->mq_usage_counter);
1884 free_percpu(q->queue_ctx);
1885 kfree(q->queue_hw_ctx);
1888 q->queue_ctx = NULL;
1889 q->queue_hw_ctx = NULL;
1892 mutex_lock(&all_q_mutex);
1893 list_del_init(&q->all_q_node);
1894 mutex_unlock(&all_q_mutex);
1897 /* Basically redo blk_mq_init_queue with queue frozen */
1898 static void blk_mq_queue_reinit(struct request_queue *q)
1900 blk_mq_freeze_queue(q);
1902 blk_mq_sysfs_unregister(q);
1904 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1907 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1908 * we should change hctx numa_node according to new topology (this
1909 * involves free and re-allocate memory, worthy doing?)
1912 blk_mq_map_swqueue(q);
1914 blk_mq_sysfs_register(q);
1916 blk_mq_unfreeze_queue(q);
1919 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1920 unsigned long action, void *hcpu)
1922 struct request_queue *q;
1925 * Before new mappings are established, hotadded cpu might already
1926 * start handling requests. This doesn't break anything as we map
1927 * offline CPUs to first hardware queue. We will re-init the queue
1928 * below to get optimal settings.
1930 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1931 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1934 mutex_lock(&all_q_mutex);
1935 list_for_each_entry(q, &all_q_list, all_q_node)
1936 blk_mq_queue_reinit(q);
1937 mutex_unlock(&all_q_mutex);
1941 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
1945 for (i = 0; i < set->nr_hw_queues; i++) {
1946 set->tags[i] = blk_mq_init_rq_map(set, i);
1955 blk_mq_free_rq_map(set, set->tags[i], i);
1961 * Allocate the request maps associated with this tag_set. Note that this
1962 * may reduce the depth asked for, if memory is tight. set->queue_depth
1963 * will be updated to reflect the allocated depth.
1965 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
1970 depth = set->queue_depth;
1972 err = __blk_mq_alloc_rq_maps(set);
1976 set->queue_depth >>= 1;
1977 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
1981 } while (set->queue_depth);
1983 if (!set->queue_depth || err) {
1984 pr_err("blk-mq: failed to allocate request map\n");
1988 if (depth != set->queue_depth)
1989 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
1990 depth, set->queue_depth);
1996 * Alloc a tag set to be associated with one or more request queues.
1997 * May fail with EINVAL for various error conditions. May adjust the
1998 * requested depth down, if if it too large. In that case, the set
1999 * value will be stored in set->queue_depth.
2001 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2003 if (!set->nr_hw_queues)
2005 if (!set->queue_depth)
2007 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2010 if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue)
2013 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2014 pr_info("blk-mq: reduced tag depth to %u\n",
2016 set->queue_depth = BLK_MQ_MAX_DEPTH;
2019 set->tags = kmalloc_node(set->nr_hw_queues *
2020 sizeof(struct blk_mq_tags *),
2021 GFP_KERNEL, set->numa_node);
2025 if (blk_mq_alloc_rq_maps(set))
2028 mutex_init(&set->tag_list_lock);
2029 INIT_LIST_HEAD(&set->tag_list);
2037 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2039 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2043 for (i = 0; i < set->nr_hw_queues; i++) {
2045 blk_mq_free_rq_map(set, set->tags[i], i);
2051 EXPORT_SYMBOL(blk_mq_free_tag_set);
2053 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2055 struct blk_mq_tag_set *set = q->tag_set;
2056 struct blk_mq_hw_ctx *hctx;
2059 if (!set || nr > set->queue_depth)
2063 queue_for_each_hw_ctx(q, hctx, i) {
2064 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2070 q->nr_requests = nr;
2075 void blk_mq_disable_hotplug(void)
2077 mutex_lock(&all_q_mutex);
2080 void blk_mq_enable_hotplug(void)
2082 mutex_unlock(&all_q_mutex);
2085 static int __init blk_mq_init(void)
2089 hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2093 subsys_initcall(blk_mq_init);