1 #include <linux/kernel.h>
2 #include <linux/module.h>
3 #include <linux/backing-dev.h>
5 #include <linux/blkdev.h>
7 #include <linux/init.h>
8 #include <linux/slab.h>
9 #include <linux/workqueue.h>
10 #include <linux/smp.h>
11 #include <linux/llist.h>
12 #include <linux/list_sort.h>
13 #include <linux/cpu.h>
14 #include <linux/cache.h>
15 #include <linux/sched/sysctl.h>
16 #include <linux/delay.h>
18 #include <trace/events/block.h>
20 #include <linux/blk-mq.h>
23 #include "blk-mq-tag.h"
25 static DEFINE_MUTEX(all_q_mutex);
26 static LIST_HEAD(all_q_list);
28 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
30 static struct blk_mq_ctx *__blk_mq_get_ctx(struct request_queue *q,
33 return per_cpu_ptr(q->queue_ctx, cpu);
37 * This assumes per-cpu software queueing queues. They could be per-node
38 * as well, for instance. For now this is hardcoded as-is. Note that we don't
39 * care about preemption, since we know the ctx's are persistent. This does
40 * mean that we can't rely on ctx always matching the currently running CPU.
42 static struct blk_mq_ctx *blk_mq_get_ctx(struct request_queue *q)
44 return __blk_mq_get_ctx(q, get_cpu());
47 static void blk_mq_put_ctx(struct blk_mq_ctx *ctx)
53 * Check if any of the ctx's have pending work in this hardware queue
55 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
59 for (i = 0; i < hctx->ctx_map.map_size; i++)
60 if (hctx->ctx_map.map[i].word)
66 static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx,
67 struct blk_mq_ctx *ctx)
69 return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word];
72 #define CTX_TO_BIT(hctx, ctx) \
73 ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
76 * Mark this ctx as having pending work in this hardware queue
78 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
79 struct blk_mq_ctx *ctx)
81 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
83 if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word))
84 set_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
87 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
88 struct blk_mq_ctx *ctx)
90 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
92 clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
95 static struct request *__blk_mq_alloc_request(struct blk_mq_hw_ctx *hctx,
96 struct blk_mq_ctx *ctx,
97 gfp_t gfp, bool reserved)
102 tag = blk_mq_get_tag(hctx, &ctx->last_tag, gfp, reserved);
103 if (tag != BLK_MQ_TAG_FAIL) {
104 rq = hctx->tags->rqs[tag];
107 if (blk_mq_tag_busy(hctx)) {
108 rq->cmd_flags = REQ_MQ_INFLIGHT;
109 atomic_inc(&hctx->nr_active);
119 static int blk_mq_queue_enter(struct request_queue *q)
123 __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
125 /* we have problems to freeze the queue if it's initializing */
126 if (!blk_queue_bypass(q) || !blk_queue_init_done(q))
129 __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
131 spin_lock_irq(q->queue_lock);
132 ret = wait_event_interruptible_lock_irq(q->mq_freeze_wq,
133 !blk_queue_bypass(q) || blk_queue_dying(q),
135 /* inc usage with lock hold to avoid freeze_queue runs here */
136 if (!ret && !blk_queue_dying(q))
137 __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
138 else if (blk_queue_dying(q))
140 spin_unlock_irq(q->queue_lock);
145 static void blk_mq_queue_exit(struct request_queue *q)
147 __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
150 static void __blk_mq_drain_queue(struct request_queue *q)
155 spin_lock_irq(q->queue_lock);
156 count = percpu_counter_sum(&q->mq_usage_counter);
157 spin_unlock_irq(q->queue_lock);
161 blk_mq_run_queues(q, false);
167 * Guarantee no request is in use, so we can change any data structure of
168 * the queue afterward.
170 static void blk_mq_freeze_queue(struct request_queue *q)
174 spin_lock_irq(q->queue_lock);
175 drain = !q->bypass_depth++;
176 queue_flag_set(QUEUE_FLAG_BYPASS, q);
177 spin_unlock_irq(q->queue_lock);
180 __blk_mq_drain_queue(q);
183 void blk_mq_drain_queue(struct request_queue *q)
185 __blk_mq_drain_queue(q);
188 static void blk_mq_unfreeze_queue(struct request_queue *q)
192 spin_lock_irq(q->queue_lock);
193 if (!--q->bypass_depth) {
194 queue_flag_clear(QUEUE_FLAG_BYPASS, q);
197 WARN_ON_ONCE(q->bypass_depth < 0);
198 spin_unlock_irq(q->queue_lock);
200 wake_up_all(&q->mq_freeze_wq);
203 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
205 return blk_mq_has_free_tags(hctx->tags);
207 EXPORT_SYMBOL(blk_mq_can_queue);
209 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
210 struct request *rq, unsigned int rw_flags)
212 if (blk_queue_io_stat(q))
213 rw_flags |= REQ_IO_STAT;
215 INIT_LIST_HEAD(&rq->queuelist);
216 /* csd/requeue_work/fifo_time is initialized before use */
219 rq->cmd_flags |= rw_flags;
221 /* do not touch atomic flags, it needs atomic ops against the timer */
224 rq->__sector = (sector_t) -1;
227 INIT_HLIST_NODE(&rq->hash);
228 RB_CLEAR_NODE(&rq->rb_node);
229 memset(&rq->flush, 0, max(sizeof(rq->flush), sizeof(rq->elv)));
232 rq->start_time = jiffies;
233 #ifdef CONFIG_BLK_CGROUP
235 set_start_time_ns(rq);
236 rq->io_start_time_ns = 0;
238 rq->nr_phys_segments = 0;
239 #if defined(CONFIG_BLK_DEV_INTEGRITY)
240 rq->nr_integrity_segments = 0;
244 /* tag was already set */
246 memset(rq->__cmd, 0, sizeof(rq->__cmd));
248 rq->cmd_len = BLK_MAX_CDB;
256 INIT_LIST_HEAD(&rq->timeout_list);
260 rq->end_io_data = NULL;
263 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
266 static struct request *blk_mq_alloc_request_pinned(struct request_queue *q,
273 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
274 struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q, ctx->cpu);
276 rq = __blk_mq_alloc_request(hctx, ctx, gfp & ~__GFP_WAIT,
279 blk_mq_rq_ctx_init(q, ctx, rq, rw);
283 if (gfp & __GFP_WAIT) {
284 __blk_mq_run_hw_queue(hctx);
291 blk_mq_wait_for_tags(hctx, reserved);
297 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
302 if (blk_mq_queue_enter(q))
305 rq = blk_mq_alloc_request_pinned(q, rw, gfp, reserved);
307 blk_mq_put_ctx(rq->mq_ctx);
310 EXPORT_SYMBOL(blk_mq_alloc_request);
312 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
313 struct blk_mq_ctx *ctx, struct request *rq)
315 const int tag = rq->tag;
316 struct request_queue *q = rq->q;
318 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
319 atomic_dec(&hctx->nr_active);
321 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
322 blk_mq_put_tag(hctx, tag, &ctx->last_tag);
323 blk_mq_queue_exit(q);
326 void blk_mq_free_request(struct request *rq)
328 struct blk_mq_ctx *ctx = rq->mq_ctx;
329 struct blk_mq_hw_ctx *hctx;
330 struct request_queue *q = rq->q;
332 ctx->rq_completed[rq_is_sync(rq)]++;
334 hctx = q->mq_ops->map_queue(q, ctx->cpu);
335 __blk_mq_free_request(hctx, ctx, rq);
339 * Clone all relevant state from a request that has been put on hold in
340 * the flush state machine into the preallocated flush request that hangs
341 * off the request queue.
343 * For a driver the flush request should be invisible, that's why we are
344 * impersonating the original request here.
346 void blk_mq_clone_flush_request(struct request *flush_rq,
347 struct request *orig_rq)
349 struct blk_mq_hw_ctx *hctx =
350 orig_rq->q->mq_ops->map_queue(orig_rq->q, orig_rq->mq_ctx->cpu);
352 flush_rq->mq_ctx = orig_rq->mq_ctx;
353 flush_rq->tag = orig_rq->tag;
354 memcpy(blk_mq_rq_to_pdu(flush_rq), blk_mq_rq_to_pdu(orig_rq),
358 inline void __blk_mq_end_io(struct request *rq, int error)
360 blk_account_io_done(rq);
363 rq->end_io(rq, error);
365 if (unlikely(blk_bidi_rq(rq)))
366 blk_mq_free_request(rq->next_rq);
367 blk_mq_free_request(rq);
370 EXPORT_SYMBOL(__blk_mq_end_io);
372 void blk_mq_end_io(struct request *rq, int error)
374 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
376 __blk_mq_end_io(rq, error);
378 EXPORT_SYMBOL(blk_mq_end_io);
380 static void __blk_mq_complete_request_remote(void *data)
382 struct request *rq = data;
384 rq->q->softirq_done_fn(rq);
387 void __blk_mq_complete_request(struct request *rq)
389 struct blk_mq_ctx *ctx = rq->mq_ctx;
393 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
394 rq->q->softirq_done_fn(rq);
399 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
400 shared = cpus_share_cache(cpu, ctx->cpu);
402 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
403 rq->csd.func = __blk_mq_complete_request_remote;
406 smp_call_function_single_async(ctx->cpu, &rq->csd);
408 rq->q->softirq_done_fn(rq);
414 * blk_mq_complete_request - end I/O on a request
415 * @rq: the request being processed
418 * Ends all I/O on a request. It does not handle partial completions.
419 * The actual completion happens out-of-order, through a IPI handler.
421 void blk_mq_complete_request(struct request *rq)
423 struct request_queue *q = rq->q;
425 if (unlikely(blk_should_fake_timeout(q)))
427 if (!blk_mark_rq_complete(rq)) {
428 if (q->softirq_done_fn)
429 __blk_mq_complete_request(rq);
431 blk_mq_end_io(rq, rq->errors);
434 EXPORT_SYMBOL(blk_mq_complete_request);
436 static void blk_mq_start_request(struct request *rq, bool last)
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);
447 * Just mark start time and set the started bit. Due to memory
448 * ordering, we know we'll see the correct deadline as long as
449 * REQ_ATOMIC_STARTED is seen. Use the default queue timeout,
450 * unless one has been set in the request.
453 rq->deadline = jiffies + q->rq_timeout;
455 rq->deadline = jiffies + rq->timeout;
458 * Mark us as started and clear complete. Complete might have been
459 * set if requeue raced with timeout, which then marked it as
460 * complete. So be sure to clear complete again when we start
461 * the request, otherwise we'll ignore the completion event.
463 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
464 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
466 if (q->dma_drain_size && blk_rq_bytes(rq)) {
468 * Make sure space for the drain appears. We know we can do
469 * this because max_hw_segments has been adjusted to be one
470 * fewer than the device can handle.
472 rq->nr_phys_segments++;
476 * Flag the last request in the series so that drivers know when IO
477 * should be kicked off, if they don't do it on a per-request basis.
479 * Note: the flag isn't the only condition drivers should do kick off.
480 * If drive is busy, the last request might not have the bit set.
483 rq->cmd_flags |= REQ_END;
486 static void __blk_mq_requeue_request(struct request *rq)
488 struct request_queue *q = rq->q;
490 trace_block_rq_requeue(q, rq);
491 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
493 rq->cmd_flags &= ~REQ_END;
495 if (q->dma_drain_size && blk_rq_bytes(rq))
496 rq->nr_phys_segments--;
499 void blk_mq_requeue_request(struct request *rq)
501 __blk_mq_requeue_request(rq);
502 blk_clear_rq_complete(rq);
504 BUG_ON(blk_queued_rq(rq));
505 blk_mq_add_to_requeue_list(rq, true);
507 EXPORT_SYMBOL(blk_mq_requeue_request);
509 static void blk_mq_requeue_work(struct work_struct *work)
511 struct request_queue *q =
512 container_of(work, struct request_queue, requeue_work);
514 struct request *rq, *next;
517 spin_lock_irqsave(&q->requeue_lock, flags);
518 list_splice_init(&q->requeue_list, &rq_list);
519 spin_unlock_irqrestore(&q->requeue_lock, flags);
521 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
522 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
525 rq->cmd_flags &= ~REQ_SOFTBARRIER;
526 list_del_init(&rq->queuelist);
527 blk_mq_insert_request(rq, true, false, false);
530 while (!list_empty(&rq_list)) {
531 rq = list_entry(rq_list.next, struct request, queuelist);
532 list_del_init(&rq->queuelist);
533 blk_mq_insert_request(rq, false, false, false);
536 blk_mq_run_queues(q, false);
539 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
541 struct request_queue *q = rq->q;
545 * We abuse this flag that is otherwise used by the I/O scheduler to
546 * request head insertation from the workqueue.
548 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
550 spin_lock_irqsave(&q->requeue_lock, flags);
552 rq->cmd_flags |= REQ_SOFTBARRIER;
553 list_add(&rq->queuelist, &q->requeue_list);
555 list_add_tail(&rq->queuelist, &q->requeue_list);
557 spin_unlock_irqrestore(&q->requeue_lock, flags);
559 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
561 void blk_mq_kick_requeue_list(struct request_queue *q)
563 kblockd_schedule_work(&q->requeue_work);
565 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
567 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
569 return tags->rqs[tag];
571 EXPORT_SYMBOL(blk_mq_tag_to_rq);
573 struct blk_mq_timeout_data {
574 struct blk_mq_hw_ctx *hctx;
576 unsigned int *next_set;
579 static void blk_mq_timeout_check(void *__data, unsigned long *free_tags)
581 struct blk_mq_timeout_data *data = __data;
582 struct blk_mq_hw_ctx *hctx = data->hctx;
585 /* It may not be in flight yet (this is where
586 * the REQ_ATOMIC_STARTED flag comes in). The requests are
587 * statically allocated, so we know it's always safe to access the
588 * memory associated with a bit offset into ->rqs[].
594 tag = find_next_zero_bit(free_tags, hctx->tags->nr_tags, tag);
595 if (tag >= hctx->tags->nr_tags)
598 rq = blk_mq_tag_to_rq(hctx->tags, tag++);
599 if (rq->q != hctx->queue)
601 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
604 blk_rq_check_expired(rq, data->next, data->next_set);
608 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx,
610 unsigned int *next_set)
612 struct blk_mq_timeout_data data = {
615 .next_set = next_set,
619 * Ask the tagging code to iterate busy requests, so we can
620 * check them for timeout.
622 blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data);
625 static enum blk_eh_timer_return blk_mq_rq_timed_out(struct request *rq)
627 struct request_queue *q = rq->q;
630 * We know that complete is set at this point. If STARTED isn't set
631 * anymore, then the request isn't active and the "timeout" should
632 * just be ignored. This can happen due to the bitflag ordering.
633 * Timeout first checks if STARTED is set, and if it is, assumes
634 * the request is active. But if we race with completion, then
635 * we both flags will get cleared. So check here again, and ignore
636 * a timeout event with a request that isn't active.
638 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
639 return BLK_EH_NOT_HANDLED;
641 if (!q->mq_ops->timeout)
642 return BLK_EH_RESET_TIMER;
644 return q->mq_ops->timeout(rq);
647 static void blk_mq_rq_timer(unsigned long data)
649 struct request_queue *q = (struct request_queue *) data;
650 struct blk_mq_hw_ctx *hctx;
651 unsigned long next = 0;
654 queue_for_each_hw_ctx(q, hctx, i) {
656 * If not software queues are currently mapped to this
657 * hardware queue, there's nothing to check
659 if (!hctx->nr_ctx || !hctx->tags)
662 blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set);
666 next = blk_rq_timeout(round_jiffies_up(next));
667 mod_timer(&q->timeout, next);
669 queue_for_each_hw_ctx(q, hctx, i)
670 blk_mq_tag_idle(hctx);
675 * Reverse check our software queue for entries that we could potentially
676 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
677 * too much time checking for merges.
679 static bool blk_mq_attempt_merge(struct request_queue *q,
680 struct blk_mq_ctx *ctx, struct bio *bio)
685 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
691 if (!blk_rq_merge_ok(rq, bio))
694 el_ret = blk_try_merge(rq, bio);
695 if (el_ret == ELEVATOR_BACK_MERGE) {
696 if (bio_attempt_back_merge(q, rq, bio)) {
701 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
702 if (bio_attempt_front_merge(q, rq, bio)) {
714 * Process software queues that have been marked busy, splicing them
715 * to the for-dispatch
717 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
719 struct blk_mq_ctx *ctx;
722 for (i = 0; i < hctx->ctx_map.map_size; i++) {
723 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
724 unsigned int off, bit;
730 off = i * hctx->ctx_map.bits_per_word;
732 bit = find_next_bit(&bm->word, bm->depth, bit);
733 if (bit >= bm->depth)
736 ctx = hctx->ctxs[bit + off];
737 clear_bit(bit, &bm->word);
738 spin_lock(&ctx->lock);
739 list_splice_tail_init(&ctx->rq_list, list);
740 spin_unlock(&ctx->lock);
748 * Run this hardware queue, pulling any software queues mapped to it in.
749 * Note that this function currently has various problems around ordering
750 * of IO. In particular, we'd like FIFO behaviour on handling existing
751 * items on the hctx->dispatch list. Ignore that for now.
753 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
755 struct request_queue *q = hctx->queue;
760 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
762 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
768 * Touch any software queue that has pending entries.
770 flush_busy_ctxs(hctx, &rq_list);
773 * If we have previous entries on our dispatch list, grab them
774 * and stuff them at the front for more fair dispatch.
776 if (!list_empty_careful(&hctx->dispatch)) {
777 spin_lock(&hctx->lock);
778 if (!list_empty(&hctx->dispatch))
779 list_splice_init(&hctx->dispatch, &rq_list);
780 spin_unlock(&hctx->lock);
784 * Now process all the entries, sending them to the driver.
787 while (!list_empty(&rq_list)) {
790 rq = list_first_entry(&rq_list, struct request, queuelist);
791 list_del_init(&rq->queuelist);
793 blk_mq_start_request(rq, list_empty(&rq_list));
795 ret = q->mq_ops->queue_rq(hctx, rq);
797 case BLK_MQ_RQ_QUEUE_OK:
800 case BLK_MQ_RQ_QUEUE_BUSY:
801 list_add(&rq->queuelist, &rq_list);
802 __blk_mq_requeue_request(rq);
805 pr_err("blk-mq: bad return on queue: %d\n", ret);
806 case BLK_MQ_RQ_QUEUE_ERROR:
808 blk_mq_end_io(rq, rq->errors);
812 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
817 hctx->dispatched[0]++;
818 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
819 hctx->dispatched[ilog2(queued) + 1]++;
822 * Any items that need requeuing? Stuff them into hctx->dispatch,
823 * that is where we will continue on next queue run.
825 if (!list_empty(&rq_list)) {
826 spin_lock(&hctx->lock);
827 list_splice(&rq_list, &hctx->dispatch);
828 spin_unlock(&hctx->lock);
833 * It'd be great if the workqueue API had a way to pass
834 * in a mask and had some smarts for more clever placement.
835 * For now we just round-robin here, switching for every
836 * BLK_MQ_CPU_WORK_BATCH queued items.
838 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
840 int cpu = hctx->next_cpu;
842 if (--hctx->next_cpu_batch <= 0) {
845 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
846 if (next_cpu >= nr_cpu_ids)
847 next_cpu = cpumask_first(hctx->cpumask);
849 hctx->next_cpu = next_cpu;
850 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
856 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
858 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
861 if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask))
862 __blk_mq_run_hw_queue(hctx);
863 else if (hctx->queue->nr_hw_queues == 1)
864 kblockd_schedule_delayed_work(&hctx->run_work, 0);
868 cpu = blk_mq_hctx_next_cpu(hctx);
869 kblockd_schedule_delayed_work_on(cpu, &hctx->run_work, 0);
873 void blk_mq_run_queues(struct request_queue *q, bool async)
875 struct blk_mq_hw_ctx *hctx;
878 queue_for_each_hw_ctx(q, hctx, i) {
879 if ((!blk_mq_hctx_has_pending(hctx) &&
880 list_empty_careful(&hctx->dispatch)) ||
881 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
885 blk_mq_run_hw_queue(hctx, async);
889 EXPORT_SYMBOL(blk_mq_run_queues);
891 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
893 cancel_delayed_work(&hctx->run_work);
894 cancel_delayed_work(&hctx->delay_work);
895 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
897 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
899 void blk_mq_stop_hw_queues(struct request_queue *q)
901 struct blk_mq_hw_ctx *hctx;
904 queue_for_each_hw_ctx(q, hctx, i)
905 blk_mq_stop_hw_queue(hctx);
907 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
909 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
911 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
914 __blk_mq_run_hw_queue(hctx);
917 EXPORT_SYMBOL(blk_mq_start_hw_queue);
919 void blk_mq_start_hw_queues(struct request_queue *q)
921 struct blk_mq_hw_ctx *hctx;
924 queue_for_each_hw_ctx(q, hctx, i)
925 blk_mq_start_hw_queue(hctx);
927 EXPORT_SYMBOL(blk_mq_start_hw_queues);
930 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
932 struct blk_mq_hw_ctx *hctx;
935 queue_for_each_hw_ctx(q, hctx, i) {
936 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
939 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
941 blk_mq_run_hw_queue(hctx, async);
945 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
947 static void blk_mq_run_work_fn(struct work_struct *work)
949 struct blk_mq_hw_ctx *hctx;
951 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
953 __blk_mq_run_hw_queue(hctx);
956 static void blk_mq_delay_work_fn(struct work_struct *work)
958 struct blk_mq_hw_ctx *hctx;
960 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
962 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
963 __blk_mq_run_hw_queue(hctx);
966 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
968 unsigned long tmo = msecs_to_jiffies(msecs);
970 if (hctx->queue->nr_hw_queues == 1)
971 kblockd_schedule_delayed_work(&hctx->delay_work, tmo);
975 cpu = blk_mq_hctx_next_cpu(hctx);
976 kblockd_schedule_delayed_work_on(cpu, &hctx->delay_work, tmo);
979 EXPORT_SYMBOL(blk_mq_delay_queue);
981 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
982 struct request *rq, bool at_head)
984 struct blk_mq_ctx *ctx = rq->mq_ctx;
986 trace_block_rq_insert(hctx->queue, rq);
989 list_add(&rq->queuelist, &ctx->rq_list);
991 list_add_tail(&rq->queuelist, &ctx->rq_list);
993 blk_mq_hctx_mark_pending(hctx, ctx);
996 * We do this early, to ensure we are on the right CPU.
1001 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1004 struct request_queue *q = rq->q;
1005 struct blk_mq_hw_ctx *hctx;
1006 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
1008 current_ctx = blk_mq_get_ctx(q);
1009 if (!cpu_online(ctx->cpu))
1010 rq->mq_ctx = ctx = current_ctx;
1012 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1014 if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA) &&
1015 !(rq->cmd_flags & (REQ_FLUSH_SEQ))) {
1016 blk_insert_flush(rq);
1018 spin_lock(&ctx->lock);
1019 __blk_mq_insert_request(hctx, rq, at_head);
1020 spin_unlock(&ctx->lock);
1024 blk_mq_run_hw_queue(hctx, async);
1026 blk_mq_put_ctx(current_ctx);
1029 static void blk_mq_insert_requests(struct request_queue *q,
1030 struct blk_mq_ctx *ctx,
1031 struct list_head *list,
1036 struct blk_mq_hw_ctx *hctx;
1037 struct blk_mq_ctx *current_ctx;
1039 trace_block_unplug(q, depth, !from_schedule);
1041 current_ctx = blk_mq_get_ctx(q);
1043 if (!cpu_online(ctx->cpu))
1045 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1048 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1051 spin_lock(&ctx->lock);
1052 while (!list_empty(list)) {
1055 rq = list_first_entry(list, struct request, queuelist);
1056 list_del_init(&rq->queuelist);
1058 __blk_mq_insert_request(hctx, rq, false);
1060 spin_unlock(&ctx->lock);
1062 blk_mq_run_hw_queue(hctx, from_schedule);
1063 blk_mq_put_ctx(current_ctx);
1066 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1068 struct request *rqa = container_of(a, struct request, queuelist);
1069 struct request *rqb = container_of(b, struct request, queuelist);
1071 return !(rqa->mq_ctx < rqb->mq_ctx ||
1072 (rqa->mq_ctx == rqb->mq_ctx &&
1073 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1076 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1078 struct blk_mq_ctx *this_ctx;
1079 struct request_queue *this_q;
1082 LIST_HEAD(ctx_list);
1085 list_splice_init(&plug->mq_list, &list);
1087 list_sort(NULL, &list, plug_ctx_cmp);
1093 while (!list_empty(&list)) {
1094 rq = list_entry_rq(list.next);
1095 list_del_init(&rq->queuelist);
1097 if (rq->mq_ctx != this_ctx) {
1099 blk_mq_insert_requests(this_q, this_ctx,
1104 this_ctx = rq->mq_ctx;
1110 list_add_tail(&rq->queuelist, &ctx_list);
1114 * If 'this_ctx' is set, we know we have entries to complete
1115 * on 'ctx_list'. Do those.
1118 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1123 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1125 init_request_from_bio(rq, bio);
1126 blk_account_io_start(rq, 1);
1129 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1130 struct blk_mq_ctx *ctx,
1131 struct request *rq, struct bio *bio)
1133 struct request_queue *q = hctx->queue;
1135 if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE)) {
1136 blk_mq_bio_to_request(rq, bio);
1137 spin_lock(&ctx->lock);
1139 __blk_mq_insert_request(hctx, rq, false);
1140 spin_unlock(&ctx->lock);
1143 spin_lock(&ctx->lock);
1144 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1145 blk_mq_bio_to_request(rq, bio);
1149 spin_unlock(&ctx->lock);
1150 __blk_mq_free_request(hctx, ctx, rq);
1155 struct blk_map_ctx {
1156 struct blk_mq_hw_ctx *hctx;
1157 struct blk_mq_ctx *ctx;
1160 static struct request *blk_mq_map_request(struct request_queue *q,
1162 struct blk_map_ctx *data)
1164 struct blk_mq_hw_ctx *hctx;
1165 struct blk_mq_ctx *ctx;
1167 int rw = bio_data_dir(bio);
1169 if (unlikely(blk_mq_queue_enter(q))) {
1170 bio_endio(bio, -EIO);
1174 ctx = blk_mq_get_ctx(q);
1175 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1177 if (rw_is_sync(bio->bi_rw))
1180 trace_block_getrq(q, bio, rw);
1181 rq = __blk_mq_alloc_request(hctx, ctx, GFP_ATOMIC, false);
1183 blk_mq_rq_ctx_init(q, ctx, rq, rw);
1185 blk_mq_put_ctx(ctx);
1186 trace_block_sleeprq(q, bio, rw);
1187 rq = blk_mq_alloc_request_pinned(q, rw, __GFP_WAIT|GFP_ATOMIC,
1190 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1200 * Multiple hardware queue variant. This will not use per-process plugs,
1201 * but will attempt to bypass the hctx queueing if we can go straight to
1202 * hardware for SYNC IO.
1204 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1206 const int is_sync = rw_is_sync(bio->bi_rw);
1207 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1208 struct blk_map_ctx data;
1211 blk_queue_bounce(q, &bio);
1213 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1214 bio_endio(bio, -EIO);
1218 rq = blk_mq_map_request(q, bio, &data);
1222 if (unlikely(is_flush_fua)) {
1223 blk_mq_bio_to_request(rq, bio);
1224 blk_insert_flush(rq);
1231 blk_mq_bio_to_request(rq, bio);
1232 blk_mq_start_request(rq, true);
1235 * For OK queue, we are done. For error, kill it. Any other
1236 * error (busy), just add it to our list as we previously
1239 ret = q->mq_ops->queue_rq(data.hctx, rq);
1240 if (ret == BLK_MQ_RQ_QUEUE_OK)
1243 __blk_mq_requeue_request(rq);
1245 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1247 blk_mq_end_io(rq, rq->errors);
1253 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1255 * For a SYNC request, send it to the hardware immediately. For
1256 * an ASYNC request, just ensure that we run it later on. The
1257 * latter allows for merging opportunities and more efficient
1261 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1264 blk_mq_put_ctx(data.ctx);
1268 * Single hardware queue variant. This will attempt to use any per-process
1269 * plug for merging and IO deferral.
1271 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1273 const int is_sync = rw_is_sync(bio->bi_rw);
1274 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1275 unsigned int use_plug, request_count = 0;
1276 struct blk_map_ctx data;
1280 * If we have multiple hardware queues, just go directly to
1281 * one of those for sync IO.
1283 use_plug = !is_flush_fua && !is_sync;
1285 blk_queue_bounce(q, &bio);
1287 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1288 bio_endio(bio, -EIO);
1292 if (use_plug && !blk_queue_nomerges(q) &&
1293 blk_attempt_plug_merge(q, bio, &request_count))
1296 rq = blk_mq_map_request(q, bio, &data);
1298 if (unlikely(is_flush_fua)) {
1299 blk_mq_bio_to_request(rq, bio);
1300 blk_insert_flush(rq);
1305 * A task plug currently exists. Since this is completely lockless,
1306 * utilize that to temporarily store requests until the task is
1307 * either done or scheduled away.
1310 struct blk_plug *plug = current->plug;
1313 blk_mq_bio_to_request(rq, bio);
1314 if (list_empty(&plug->mq_list))
1315 trace_block_plug(q);
1316 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1317 blk_flush_plug_list(plug, false);
1318 trace_block_plug(q);
1320 list_add_tail(&rq->queuelist, &plug->mq_list);
1321 blk_mq_put_ctx(data.ctx);
1326 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1328 * For a SYNC request, send it to the hardware immediately. For
1329 * an ASYNC request, just ensure that we run it later on. The
1330 * latter allows for merging opportunities and more efficient
1334 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1337 blk_mq_put_ctx(data.ctx);
1341 * Default mapping to a software queue, since we use one per CPU.
1343 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1345 return q->queue_hw_ctx[q->mq_map[cpu]];
1347 EXPORT_SYMBOL(blk_mq_map_queue);
1349 struct blk_mq_hw_ctx *blk_mq_alloc_single_hw_queue(struct blk_mq_tag_set *set,
1350 unsigned int hctx_index,
1353 return kzalloc_node(sizeof(struct blk_mq_hw_ctx), GFP_KERNEL, node);
1355 EXPORT_SYMBOL(blk_mq_alloc_single_hw_queue);
1357 void blk_mq_free_single_hw_queue(struct blk_mq_hw_ctx *hctx,
1358 unsigned int hctx_index)
1362 EXPORT_SYMBOL(blk_mq_free_single_hw_queue);
1364 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1365 struct blk_mq_tags *tags, unsigned int hctx_idx)
1369 if (tags->rqs && set->ops->exit_request) {
1372 for (i = 0; i < tags->nr_tags; i++) {
1375 set->ops->exit_request(set->driver_data, tags->rqs[i],
1380 while (!list_empty(&tags->page_list)) {
1381 page = list_first_entry(&tags->page_list, struct page, lru);
1382 list_del_init(&page->lru);
1383 __free_pages(page, page->private);
1388 blk_mq_free_tags(tags);
1391 static size_t order_to_size(unsigned int order)
1393 return (size_t)PAGE_SIZE << order;
1396 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1397 unsigned int hctx_idx)
1399 struct blk_mq_tags *tags;
1400 unsigned int i, j, entries_per_page, max_order = 4;
1401 size_t rq_size, left;
1403 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1408 INIT_LIST_HEAD(&tags->page_list);
1410 tags->rqs = kmalloc_node(set->queue_depth * sizeof(struct request *),
1411 GFP_KERNEL, set->numa_node);
1413 blk_mq_free_tags(tags);
1418 * rq_size is the size of the request plus driver payload, rounded
1419 * to the cacheline size
1421 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1423 left = rq_size * set->queue_depth;
1425 for (i = 0; i < set->queue_depth; ) {
1426 int this_order = max_order;
1431 while (left < order_to_size(this_order - 1) && this_order)
1435 page = alloc_pages_node(set->numa_node, GFP_KERNEL,
1441 if (order_to_size(this_order) < rq_size)
1448 page->private = this_order;
1449 list_add_tail(&page->lru, &tags->page_list);
1451 p = page_address(page);
1452 entries_per_page = order_to_size(this_order) / rq_size;
1453 to_do = min(entries_per_page, set->queue_depth - i);
1454 left -= to_do * rq_size;
1455 for (j = 0; j < to_do; j++) {
1457 if (set->ops->init_request) {
1458 if (set->ops->init_request(set->driver_data,
1459 tags->rqs[i], hctx_idx, i,
1472 pr_warn("%s: failed to allocate requests\n", __func__);
1473 blk_mq_free_rq_map(set, tags, hctx_idx);
1477 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1482 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1484 unsigned int bpw = 8, total, num_maps, i;
1486 bitmap->bits_per_word = bpw;
1488 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1489 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1494 bitmap->map_size = num_maps;
1497 for (i = 0; i < num_maps; i++) {
1498 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1499 total -= bitmap->map[i].depth;
1505 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1507 struct request_queue *q = hctx->queue;
1508 struct blk_mq_ctx *ctx;
1512 * Move ctx entries to new CPU, if this one is going away.
1514 ctx = __blk_mq_get_ctx(q, cpu);
1516 spin_lock(&ctx->lock);
1517 if (!list_empty(&ctx->rq_list)) {
1518 list_splice_init(&ctx->rq_list, &tmp);
1519 blk_mq_hctx_clear_pending(hctx, ctx);
1521 spin_unlock(&ctx->lock);
1523 if (list_empty(&tmp))
1526 ctx = blk_mq_get_ctx(q);
1527 spin_lock(&ctx->lock);
1529 while (!list_empty(&tmp)) {
1532 rq = list_first_entry(&tmp, struct request, queuelist);
1534 list_move_tail(&rq->queuelist, &ctx->rq_list);
1537 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1538 blk_mq_hctx_mark_pending(hctx, ctx);
1540 spin_unlock(&ctx->lock);
1542 blk_mq_run_hw_queue(hctx, true);
1543 blk_mq_put_ctx(ctx);
1547 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1549 struct request_queue *q = hctx->queue;
1550 struct blk_mq_tag_set *set = q->tag_set;
1552 if (set->tags[hctx->queue_num])
1555 set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1556 if (!set->tags[hctx->queue_num])
1559 hctx->tags = set->tags[hctx->queue_num];
1563 static int blk_mq_hctx_notify(void *data, unsigned long action,
1566 struct blk_mq_hw_ctx *hctx = data;
1568 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1569 return blk_mq_hctx_cpu_offline(hctx, cpu);
1570 else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1571 return blk_mq_hctx_cpu_online(hctx, cpu);
1576 static void blk_mq_exit_hw_queues(struct request_queue *q,
1577 struct blk_mq_tag_set *set, int nr_queue)
1579 struct blk_mq_hw_ctx *hctx;
1582 queue_for_each_hw_ctx(q, hctx, i) {
1586 if (set->ops->exit_hctx)
1587 set->ops->exit_hctx(hctx, i);
1589 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1591 blk_mq_free_bitmap(&hctx->ctx_map);
1596 static void blk_mq_free_hw_queues(struct request_queue *q,
1597 struct blk_mq_tag_set *set)
1599 struct blk_mq_hw_ctx *hctx;
1602 queue_for_each_hw_ctx(q, hctx, i) {
1603 free_cpumask_var(hctx->cpumask);
1604 set->ops->free_hctx(hctx, i);
1608 static int blk_mq_init_hw_queues(struct request_queue *q,
1609 struct blk_mq_tag_set *set)
1611 struct blk_mq_hw_ctx *hctx;
1615 * Initialize hardware queues
1617 queue_for_each_hw_ctx(q, hctx, i) {
1620 node = hctx->numa_node;
1621 if (node == NUMA_NO_NODE)
1622 node = hctx->numa_node = set->numa_node;
1624 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1625 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1626 spin_lock_init(&hctx->lock);
1627 INIT_LIST_HEAD(&hctx->dispatch);
1629 hctx->queue_num = i;
1630 hctx->flags = set->flags;
1631 hctx->cmd_size = set->cmd_size;
1633 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1634 blk_mq_hctx_notify, hctx);
1635 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1637 hctx->tags = set->tags[i];
1640 * Allocate space for all possible cpus to avoid allocation in
1643 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1648 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1653 if (set->ops->init_hctx &&
1654 set->ops->init_hctx(hctx, set->driver_data, i))
1658 if (i == q->nr_hw_queues)
1664 blk_mq_exit_hw_queues(q, set, i);
1669 static void blk_mq_init_cpu_queues(struct request_queue *q,
1670 unsigned int nr_hw_queues)
1674 for_each_possible_cpu(i) {
1675 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1676 struct blk_mq_hw_ctx *hctx;
1678 memset(__ctx, 0, sizeof(*__ctx));
1680 spin_lock_init(&__ctx->lock);
1681 INIT_LIST_HEAD(&__ctx->rq_list);
1684 /* If the cpu isn't online, the cpu is mapped to first hctx */
1688 hctx = q->mq_ops->map_queue(q, i);
1689 cpumask_set_cpu(i, hctx->cpumask);
1693 * Set local node, IFF we have more than one hw queue. If
1694 * not, we remain on the home node of the device
1696 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1697 hctx->numa_node = cpu_to_node(i);
1701 static void blk_mq_map_swqueue(struct request_queue *q)
1704 struct blk_mq_hw_ctx *hctx;
1705 struct blk_mq_ctx *ctx;
1707 queue_for_each_hw_ctx(q, hctx, i) {
1708 cpumask_clear(hctx->cpumask);
1713 * Map software to hardware queues
1715 queue_for_each_ctx(q, ctx, i) {
1716 /* If the cpu isn't online, the cpu is mapped to first hctx */
1720 hctx = q->mq_ops->map_queue(q, i);
1721 cpumask_set_cpu(i, hctx->cpumask);
1722 ctx->index_hw = hctx->nr_ctx;
1723 hctx->ctxs[hctx->nr_ctx++] = ctx;
1726 queue_for_each_hw_ctx(q, hctx, i) {
1728 * If not software queues are mapped to this hardware queue,
1729 * disable it and free the request entries
1731 if (!hctx->nr_ctx) {
1732 struct blk_mq_tag_set *set = q->tag_set;
1735 blk_mq_free_rq_map(set, set->tags[i], i);
1736 set->tags[i] = NULL;
1743 * Initialize batch roundrobin counts
1745 hctx->next_cpu = cpumask_first(hctx->cpumask);
1746 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1750 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1752 struct blk_mq_hw_ctx *hctx;
1753 struct request_queue *q;
1757 if (set->tag_list.next == set->tag_list.prev)
1762 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1763 blk_mq_freeze_queue(q);
1765 queue_for_each_hw_ctx(q, hctx, i) {
1767 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1769 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1771 blk_mq_unfreeze_queue(q);
1775 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1777 struct blk_mq_tag_set *set = q->tag_set;
1779 blk_mq_freeze_queue(q);
1781 mutex_lock(&set->tag_list_lock);
1782 list_del_init(&q->tag_set_list);
1783 blk_mq_update_tag_set_depth(set);
1784 mutex_unlock(&set->tag_list_lock);
1786 blk_mq_unfreeze_queue(q);
1789 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1790 struct request_queue *q)
1794 mutex_lock(&set->tag_list_lock);
1795 list_add_tail(&q->tag_set_list, &set->tag_list);
1796 blk_mq_update_tag_set_depth(set);
1797 mutex_unlock(&set->tag_list_lock);
1800 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1802 struct blk_mq_hw_ctx **hctxs;
1803 struct blk_mq_ctx *ctx;
1804 struct request_queue *q;
1808 ctx = alloc_percpu(struct blk_mq_ctx);
1810 return ERR_PTR(-ENOMEM);
1812 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1818 map = blk_mq_make_queue_map(set);
1822 for (i = 0; i < set->nr_hw_queues; i++) {
1823 int node = blk_mq_hw_queue_to_node(map, i);
1825 hctxs[i] = set->ops->alloc_hctx(set, i, node);
1829 if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL))
1832 atomic_set(&hctxs[i]->nr_active, 0);
1833 hctxs[i]->numa_node = node;
1834 hctxs[i]->queue_num = i;
1837 q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1841 if (percpu_counter_init(&q->mq_usage_counter, 0))
1844 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1845 blk_queue_rq_timeout(q, 30000);
1847 q->nr_queues = nr_cpu_ids;
1848 q->nr_hw_queues = set->nr_hw_queues;
1852 q->queue_hw_ctx = hctxs;
1854 q->mq_ops = set->ops;
1855 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1857 q->sg_reserved_size = INT_MAX;
1859 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1860 INIT_LIST_HEAD(&q->requeue_list);
1861 spin_lock_init(&q->requeue_lock);
1863 if (q->nr_hw_queues > 1)
1864 blk_queue_make_request(q, blk_mq_make_request);
1866 blk_queue_make_request(q, blk_sq_make_request);
1868 blk_queue_rq_timed_out(q, blk_mq_rq_timed_out);
1870 blk_queue_rq_timeout(q, set->timeout);
1873 * Do this after blk_queue_make_request() overrides it...
1875 q->nr_requests = set->queue_depth;
1877 if (set->ops->complete)
1878 blk_queue_softirq_done(q, set->ops->complete);
1880 blk_mq_init_flush(q);
1881 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1883 q->flush_rq = kzalloc(round_up(sizeof(struct request) +
1884 set->cmd_size, cache_line_size()),
1889 if (blk_mq_init_hw_queues(q, set))
1892 mutex_lock(&all_q_mutex);
1893 list_add_tail(&q->all_q_node, &all_q_list);
1894 mutex_unlock(&all_q_mutex);
1896 blk_mq_add_queue_tag_set(set, q);
1898 blk_mq_map_swqueue(q);
1905 blk_cleanup_queue(q);
1908 for (i = 0; i < set->nr_hw_queues; i++) {
1911 free_cpumask_var(hctxs[i]->cpumask);
1912 set->ops->free_hctx(hctxs[i], i);
1918 return ERR_PTR(-ENOMEM);
1920 EXPORT_SYMBOL(blk_mq_init_queue);
1922 void blk_mq_free_queue(struct request_queue *q)
1924 struct blk_mq_tag_set *set = q->tag_set;
1926 blk_mq_del_queue_tag_set(q);
1928 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
1929 blk_mq_free_hw_queues(q, set);
1931 percpu_counter_destroy(&q->mq_usage_counter);
1933 free_percpu(q->queue_ctx);
1934 kfree(q->queue_hw_ctx);
1937 q->queue_ctx = NULL;
1938 q->queue_hw_ctx = NULL;
1941 mutex_lock(&all_q_mutex);
1942 list_del_init(&q->all_q_node);
1943 mutex_unlock(&all_q_mutex);
1946 /* Basically redo blk_mq_init_queue with queue frozen */
1947 static void blk_mq_queue_reinit(struct request_queue *q)
1949 blk_mq_freeze_queue(q);
1951 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1954 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1955 * we should change hctx numa_node according to new topology (this
1956 * involves free and re-allocate memory, worthy doing?)
1959 blk_mq_map_swqueue(q);
1961 blk_mq_unfreeze_queue(q);
1964 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1965 unsigned long action, void *hcpu)
1967 struct request_queue *q;
1970 * Before new mappings are established, hotadded cpu might already
1971 * start handling requests. This doesn't break anything as we map
1972 * offline CPUs to first hardware queue. We will re-init the queue
1973 * below to get optimal settings.
1975 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1976 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1979 mutex_lock(&all_q_mutex);
1980 list_for_each_entry(q, &all_q_list, all_q_node)
1981 blk_mq_queue_reinit(q);
1982 mutex_unlock(&all_q_mutex);
1986 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
1990 if (!set->nr_hw_queues)
1992 if (!set->queue_depth || set->queue_depth > BLK_MQ_MAX_DEPTH)
1994 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
1997 if (!set->nr_hw_queues ||
1998 !set->ops->queue_rq || !set->ops->map_queue ||
1999 !set->ops->alloc_hctx || !set->ops->free_hctx)
2003 set->tags = kmalloc_node(set->nr_hw_queues *
2004 sizeof(struct blk_mq_tags *),
2005 GFP_KERNEL, set->numa_node);
2009 for (i = 0; i < set->nr_hw_queues; i++) {
2010 set->tags[i] = blk_mq_init_rq_map(set, i);
2015 mutex_init(&set->tag_list_lock);
2016 INIT_LIST_HEAD(&set->tag_list);
2022 blk_mq_free_rq_map(set, set->tags[i], i);
2026 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2028 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2032 for (i = 0; i < set->nr_hw_queues; i++) {
2034 blk_mq_free_rq_map(set, set->tags[i], i);
2039 EXPORT_SYMBOL(blk_mq_free_tag_set);
2041 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2043 struct blk_mq_tag_set *set = q->tag_set;
2044 struct blk_mq_hw_ctx *hctx;
2047 if (!set || nr > set->queue_depth)
2051 queue_for_each_hw_ctx(q, hctx, i) {
2052 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2058 q->nr_requests = nr;
2063 void blk_mq_disable_hotplug(void)
2065 mutex_lock(&all_q_mutex);
2068 void blk_mq_enable_hotplug(void)
2070 mutex_unlock(&all_q_mutex);
2073 static int __init blk_mq_init(void)
2077 /* Must be called after percpu_counter_hotcpu_callback() */
2078 hotcpu_notifier(blk_mq_queue_reinit_notify, -10);
2082 subsys_initcall(blk_mq_init);