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)
301 if (blk_mq_queue_enter(q))
304 rq = blk_mq_alloc_request_pinned(q, rw, gfp, false);
306 blk_mq_put_ctx(rq->mq_ctx);
309 EXPORT_SYMBOL(blk_mq_alloc_request);
311 struct request *blk_mq_alloc_reserved_request(struct request_queue *q, int rw,
316 if (blk_mq_queue_enter(q))
319 rq = blk_mq_alloc_request_pinned(q, rw, gfp, true);
321 blk_mq_put_ctx(rq->mq_ctx);
324 EXPORT_SYMBOL(blk_mq_alloc_reserved_request);
326 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
327 struct blk_mq_ctx *ctx, struct request *rq)
329 const int tag = rq->tag;
330 struct request_queue *q = rq->q;
332 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
333 atomic_dec(&hctx->nr_active);
335 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
336 blk_mq_put_tag(hctx, tag, &ctx->last_tag);
337 blk_mq_queue_exit(q);
340 void blk_mq_free_request(struct request *rq)
342 struct blk_mq_ctx *ctx = rq->mq_ctx;
343 struct blk_mq_hw_ctx *hctx;
344 struct request_queue *q = rq->q;
346 ctx->rq_completed[rq_is_sync(rq)]++;
348 hctx = q->mq_ops->map_queue(q, ctx->cpu);
349 __blk_mq_free_request(hctx, ctx, rq);
353 * Clone all relevant state from a request that has been put on hold in
354 * the flush state machine into the preallocated flush request that hangs
355 * off the request queue.
357 * For a driver the flush request should be invisible, that's why we are
358 * impersonating the original request here.
360 void blk_mq_clone_flush_request(struct request *flush_rq,
361 struct request *orig_rq)
363 struct blk_mq_hw_ctx *hctx =
364 orig_rq->q->mq_ops->map_queue(orig_rq->q, orig_rq->mq_ctx->cpu);
366 flush_rq->mq_ctx = orig_rq->mq_ctx;
367 flush_rq->tag = orig_rq->tag;
368 memcpy(blk_mq_rq_to_pdu(flush_rq), blk_mq_rq_to_pdu(orig_rq),
372 inline void __blk_mq_end_io(struct request *rq, int error)
374 blk_account_io_done(rq);
377 rq->end_io(rq, error);
379 if (unlikely(blk_bidi_rq(rq)))
380 blk_mq_free_request(rq->next_rq);
381 blk_mq_free_request(rq);
384 EXPORT_SYMBOL(__blk_mq_end_io);
386 void blk_mq_end_io(struct request *rq, int error)
388 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
390 __blk_mq_end_io(rq, error);
392 EXPORT_SYMBOL(blk_mq_end_io);
394 static void __blk_mq_complete_request_remote(void *data)
396 struct request *rq = data;
398 rq->q->softirq_done_fn(rq);
401 void __blk_mq_complete_request(struct request *rq)
403 struct blk_mq_ctx *ctx = rq->mq_ctx;
407 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
408 rq->q->softirq_done_fn(rq);
413 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
414 shared = cpus_share_cache(cpu, ctx->cpu);
416 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
417 rq->csd.func = __blk_mq_complete_request_remote;
420 smp_call_function_single_async(ctx->cpu, &rq->csd);
422 rq->q->softirq_done_fn(rq);
428 * blk_mq_complete_request - end I/O on a request
429 * @rq: the request being processed
432 * Ends all I/O on a request. It does not handle partial completions.
433 * The actual completion happens out-of-order, through a IPI handler.
435 void blk_mq_complete_request(struct request *rq)
437 struct request_queue *q = rq->q;
439 if (unlikely(blk_should_fake_timeout(q)))
441 if (!blk_mark_rq_complete(rq)) {
442 if (q->softirq_done_fn)
443 __blk_mq_complete_request(rq);
445 blk_mq_end_io(rq, rq->errors);
448 EXPORT_SYMBOL(blk_mq_complete_request);
450 static void blk_mq_start_request(struct request *rq, bool last)
452 struct request_queue *q = rq->q;
454 trace_block_rq_issue(q, rq);
456 rq->resid_len = blk_rq_bytes(rq);
457 if (unlikely(blk_bidi_rq(rq)))
458 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
461 * Just mark start time and set the started bit. Due to memory
462 * ordering, we know we'll see the correct deadline as long as
463 * REQ_ATOMIC_STARTED is seen. Use the default queue timeout,
464 * unless one has been set in the request.
467 rq->deadline = jiffies + q->rq_timeout;
469 rq->deadline = jiffies + rq->timeout;
472 * Mark us as started and clear complete. Complete might have been
473 * set if requeue raced with timeout, which then marked it as
474 * complete. So be sure to clear complete again when we start
475 * the request, otherwise we'll ignore the completion event.
477 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
478 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
480 if (q->dma_drain_size && blk_rq_bytes(rq)) {
482 * Make sure space for the drain appears. We know we can do
483 * this because max_hw_segments has been adjusted to be one
484 * fewer than the device can handle.
486 rq->nr_phys_segments++;
490 * Flag the last request in the series so that drivers know when IO
491 * should be kicked off, if they don't do it on a per-request basis.
493 * Note: the flag isn't the only condition drivers should do kick off.
494 * If drive is busy, the last request might not have the bit set.
497 rq->cmd_flags |= REQ_END;
500 static void __blk_mq_requeue_request(struct request *rq)
502 struct request_queue *q = rq->q;
504 trace_block_rq_requeue(q, rq);
505 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
507 rq->cmd_flags &= ~REQ_END;
509 if (q->dma_drain_size && blk_rq_bytes(rq))
510 rq->nr_phys_segments--;
513 void blk_mq_requeue_request(struct request *rq)
515 __blk_mq_requeue_request(rq);
516 blk_clear_rq_complete(rq);
518 BUG_ON(blk_queued_rq(rq));
519 blk_mq_add_to_requeue_list(rq, true);
521 EXPORT_SYMBOL(blk_mq_requeue_request);
523 static void blk_mq_requeue_work(struct work_struct *work)
525 struct request_queue *q =
526 container_of(work, struct request_queue, requeue_work);
528 struct request *rq, *next;
531 spin_lock_irqsave(&q->requeue_lock, flags);
532 list_splice_init(&q->requeue_list, &rq_list);
533 spin_unlock_irqrestore(&q->requeue_lock, flags);
535 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
536 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
539 rq->cmd_flags &= ~REQ_SOFTBARRIER;
540 list_del_init(&rq->queuelist);
541 blk_mq_insert_request(rq, true, false, false);
544 while (!list_empty(&rq_list)) {
545 rq = list_entry(rq_list.next, struct request, queuelist);
546 list_del_init(&rq->queuelist);
547 blk_mq_insert_request(rq, false, false, false);
550 blk_mq_run_queues(q, false);
553 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
555 struct request_queue *q = rq->q;
559 * We abuse this flag that is otherwise used by the I/O scheduler to
560 * request head insertation from the workqueue.
562 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
564 spin_lock_irqsave(&q->requeue_lock, flags);
566 rq->cmd_flags |= REQ_SOFTBARRIER;
567 list_add(&rq->queuelist, &q->requeue_list);
569 list_add_tail(&rq->queuelist, &q->requeue_list);
571 spin_unlock_irqrestore(&q->requeue_lock, flags);
573 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
575 void blk_mq_kick_requeue_list(struct request_queue *q)
577 kblockd_schedule_work(&q->requeue_work);
579 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
581 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
583 return tags->rqs[tag];
585 EXPORT_SYMBOL(blk_mq_tag_to_rq);
587 struct blk_mq_timeout_data {
588 struct blk_mq_hw_ctx *hctx;
590 unsigned int *next_set;
593 static void blk_mq_timeout_check(void *__data, unsigned long *free_tags)
595 struct blk_mq_timeout_data *data = __data;
596 struct blk_mq_hw_ctx *hctx = data->hctx;
599 /* It may not be in flight yet (this is where
600 * the REQ_ATOMIC_STARTED flag comes in). The requests are
601 * statically allocated, so we know it's always safe to access the
602 * memory associated with a bit offset into ->rqs[].
608 tag = find_next_zero_bit(free_tags, hctx->tags->nr_tags, tag);
609 if (tag >= hctx->tags->nr_tags)
612 rq = blk_mq_tag_to_rq(hctx->tags, tag++);
613 if (rq->q != hctx->queue)
615 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
618 blk_rq_check_expired(rq, data->next, data->next_set);
622 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx,
624 unsigned int *next_set)
626 struct blk_mq_timeout_data data = {
629 .next_set = next_set,
633 * Ask the tagging code to iterate busy requests, so we can
634 * check them for timeout.
636 blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data);
639 static enum blk_eh_timer_return blk_mq_rq_timed_out(struct request *rq)
641 struct request_queue *q = rq->q;
644 * We know that complete is set at this point. If STARTED isn't set
645 * anymore, then the request isn't active and the "timeout" should
646 * just be ignored. This can happen due to the bitflag ordering.
647 * Timeout first checks if STARTED is set, and if it is, assumes
648 * the request is active. But if we race with completion, then
649 * we both flags will get cleared. So check here again, and ignore
650 * a timeout event with a request that isn't active.
652 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
653 return BLK_EH_NOT_HANDLED;
655 if (!q->mq_ops->timeout)
656 return BLK_EH_RESET_TIMER;
658 return q->mq_ops->timeout(rq);
661 static void blk_mq_rq_timer(unsigned long data)
663 struct request_queue *q = (struct request_queue *) data;
664 struct blk_mq_hw_ctx *hctx;
665 unsigned long next = 0;
668 queue_for_each_hw_ctx(q, hctx, i) {
670 * If not software queues are currently mapped to this
671 * hardware queue, there's nothing to check
673 if (!hctx->nr_ctx || !hctx->tags)
676 blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set);
680 next = blk_rq_timeout(round_jiffies_up(next));
681 mod_timer(&q->timeout, next);
683 queue_for_each_hw_ctx(q, hctx, i)
684 blk_mq_tag_idle(hctx);
689 * Reverse check our software queue for entries that we could potentially
690 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
691 * too much time checking for merges.
693 static bool blk_mq_attempt_merge(struct request_queue *q,
694 struct blk_mq_ctx *ctx, struct bio *bio)
699 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
705 if (!blk_rq_merge_ok(rq, bio))
708 el_ret = blk_try_merge(rq, bio);
709 if (el_ret == ELEVATOR_BACK_MERGE) {
710 if (bio_attempt_back_merge(q, rq, bio)) {
715 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
716 if (bio_attempt_front_merge(q, rq, bio)) {
728 * Process software queues that have been marked busy, splicing them
729 * to the for-dispatch
731 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
733 struct blk_mq_ctx *ctx;
736 for (i = 0; i < hctx->ctx_map.map_size; i++) {
737 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
738 unsigned int off, bit;
744 off = i * hctx->ctx_map.bits_per_word;
746 bit = find_next_bit(&bm->word, bm->depth, bit);
747 if (bit >= bm->depth)
750 ctx = hctx->ctxs[bit + off];
751 clear_bit(bit, &bm->word);
752 spin_lock(&ctx->lock);
753 list_splice_tail_init(&ctx->rq_list, list);
754 spin_unlock(&ctx->lock);
762 * Run this hardware queue, pulling any software queues mapped to it in.
763 * Note that this function currently has various problems around ordering
764 * of IO. In particular, we'd like FIFO behaviour on handling existing
765 * items on the hctx->dispatch list. Ignore that for now.
767 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
769 struct request_queue *q = hctx->queue;
774 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
776 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
782 * Touch any software queue that has pending entries.
784 flush_busy_ctxs(hctx, &rq_list);
787 * If we have previous entries on our dispatch list, grab them
788 * and stuff them at the front for more fair dispatch.
790 if (!list_empty_careful(&hctx->dispatch)) {
791 spin_lock(&hctx->lock);
792 if (!list_empty(&hctx->dispatch))
793 list_splice_init(&hctx->dispatch, &rq_list);
794 spin_unlock(&hctx->lock);
798 * Now process all the entries, sending them to the driver.
801 while (!list_empty(&rq_list)) {
804 rq = list_first_entry(&rq_list, struct request, queuelist);
805 list_del_init(&rq->queuelist);
807 blk_mq_start_request(rq, list_empty(&rq_list));
809 ret = q->mq_ops->queue_rq(hctx, rq);
811 case BLK_MQ_RQ_QUEUE_OK:
814 case BLK_MQ_RQ_QUEUE_BUSY:
815 list_add(&rq->queuelist, &rq_list);
816 __blk_mq_requeue_request(rq);
819 pr_err("blk-mq: bad return on queue: %d\n", ret);
820 case BLK_MQ_RQ_QUEUE_ERROR:
822 blk_mq_end_io(rq, rq->errors);
826 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
831 hctx->dispatched[0]++;
832 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
833 hctx->dispatched[ilog2(queued) + 1]++;
836 * Any items that need requeuing? Stuff them into hctx->dispatch,
837 * that is where we will continue on next queue run.
839 if (!list_empty(&rq_list)) {
840 spin_lock(&hctx->lock);
841 list_splice(&rq_list, &hctx->dispatch);
842 spin_unlock(&hctx->lock);
847 * It'd be great if the workqueue API had a way to pass
848 * in a mask and had some smarts for more clever placement.
849 * For now we just round-robin here, switching for every
850 * BLK_MQ_CPU_WORK_BATCH queued items.
852 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
854 int cpu = hctx->next_cpu;
856 if (--hctx->next_cpu_batch <= 0) {
859 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
860 if (next_cpu >= nr_cpu_ids)
861 next_cpu = cpumask_first(hctx->cpumask);
863 hctx->next_cpu = next_cpu;
864 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
870 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
872 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
875 if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask))
876 __blk_mq_run_hw_queue(hctx);
877 else if (hctx->queue->nr_hw_queues == 1)
878 kblockd_schedule_delayed_work(&hctx->run_work, 0);
882 cpu = blk_mq_hctx_next_cpu(hctx);
883 kblockd_schedule_delayed_work_on(cpu, &hctx->run_work, 0);
887 void blk_mq_run_queues(struct request_queue *q, bool async)
889 struct blk_mq_hw_ctx *hctx;
892 queue_for_each_hw_ctx(q, hctx, i) {
893 if ((!blk_mq_hctx_has_pending(hctx) &&
894 list_empty_careful(&hctx->dispatch)) ||
895 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
899 blk_mq_run_hw_queue(hctx, async);
903 EXPORT_SYMBOL(blk_mq_run_queues);
905 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
907 cancel_delayed_work(&hctx->run_work);
908 cancel_delayed_work(&hctx->delay_work);
909 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
911 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
913 void blk_mq_stop_hw_queues(struct request_queue *q)
915 struct blk_mq_hw_ctx *hctx;
918 queue_for_each_hw_ctx(q, hctx, i)
919 blk_mq_stop_hw_queue(hctx);
921 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
923 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
925 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
928 __blk_mq_run_hw_queue(hctx);
931 EXPORT_SYMBOL(blk_mq_start_hw_queue);
933 void blk_mq_start_hw_queues(struct request_queue *q)
935 struct blk_mq_hw_ctx *hctx;
938 queue_for_each_hw_ctx(q, hctx, i)
939 blk_mq_start_hw_queue(hctx);
941 EXPORT_SYMBOL(blk_mq_start_hw_queues);
944 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
946 struct blk_mq_hw_ctx *hctx;
949 queue_for_each_hw_ctx(q, hctx, i) {
950 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
953 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
955 blk_mq_run_hw_queue(hctx, async);
959 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
961 static void blk_mq_run_work_fn(struct work_struct *work)
963 struct blk_mq_hw_ctx *hctx;
965 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
967 __blk_mq_run_hw_queue(hctx);
970 static void blk_mq_delay_work_fn(struct work_struct *work)
972 struct blk_mq_hw_ctx *hctx;
974 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
976 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
977 __blk_mq_run_hw_queue(hctx);
980 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
982 unsigned long tmo = msecs_to_jiffies(msecs);
984 if (hctx->queue->nr_hw_queues == 1)
985 kblockd_schedule_delayed_work(&hctx->delay_work, tmo);
989 cpu = blk_mq_hctx_next_cpu(hctx);
990 kblockd_schedule_delayed_work_on(cpu, &hctx->delay_work, tmo);
993 EXPORT_SYMBOL(blk_mq_delay_queue);
995 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
996 struct request *rq, bool at_head)
998 struct blk_mq_ctx *ctx = rq->mq_ctx;
1000 trace_block_rq_insert(hctx->queue, rq);
1003 list_add(&rq->queuelist, &ctx->rq_list);
1005 list_add_tail(&rq->queuelist, &ctx->rq_list);
1007 blk_mq_hctx_mark_pending(hctx, ctx);
1010 * We do this early, to ensure we are on the right CPU.
1015 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1018 struct request_queue *q = rq->q;
1019 struct blk_mq_hw_ctx *hctx;
1020 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
1022 current_ctx = blk_mq_get_ctx(q);
1023 if (!cpu_online(ctx->cpu))
1024 rq->mq_ctx = ctx = current_ctx;
1026 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1028 if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA) &&
1029 !(rq->cmd_flags & (REQ_FLUSH_SEQ))) {
1030 blk_insert_flush(rq);
1032 spin_lock(&ctx->lock);
1033 __blk_mq_insert_request(hctx, rq, at_head);
1034 spin_unlock(&ctx->lock);
1038 blk_mq_run_hw_queue(hctx, async);
1040 blk_mq_put_ctx(current_ctx);
1043 static void blk_mq_insert_requests(struct request_queue *q,
1044 struct blk_mq_ctx *ctx,
1045 struct list_head *list,
1050 struct blk_mq_hw_ctx *hctx;
1051 struct blk_mq_ctx *current_ctx;
1053 trace_block_unplug(q, depth, !from_schedule);
1055 current_ctx = blk_mq_get_ctx(q);
1057 if (!cpu_online(ctx->cpu))
1059 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1062 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1065 spin_lock(&ctx->lock);
1066 while (!list_empty(list)) {
1069 rq = list_first_entry(list, struct request, queuelist);
1070 list_del_init(&rq->queuelist);
1072 __blk_mq_insert_request(hctx, rq, false);
1074 spin_unlock(&ctx->lock);
1076 blk_mq_run_hw_queue(hctx, from_schedule);
1077 blk_mq_put_ctx(current_ctx);
1080 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1082 struct request *rqa = container_of(a, struct request, queuelist);
1083 struct request *rqb = container_of(b, struct request, queuelist);
1085 return !(rqa->mq_ctx < rqb->mq_ctx ||
1086 (rqa->mq_ctx == rqb->mq_ctx &&
1087 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1090 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1092 struct blk_mq_ctx *this_ctx;
1093 struct request_queue *this_q;
1096 LIST_HEAD(ctx_list);
1099 list_splice_init(&plug->mq_list, &list);
1101 list_sort(NULL, &list, plug_ctx_cmp);
1107 while (!list_empty(&list)) {
1108 rq = list_entry_rq(list.next);
1109 list_del_init(&rq->queuelist);
1111 if (rq->mq_ctx != this_ctx) {
1113 blk_mq_insert_requests(this_q, this_ctx,
1118 this_ctx = rq->mq_ctx;
1124 list_add_tail(&rq->queuelist, &ctx_list);
1128 * If 'this_ctx' is set, we know we have entries to complete
1129 * on 'ctx_list'. Do those.
1132 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1137 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1139 init_request_from_bio(rq, bio);
1140 blk_account_io_start(rq, 1);
1143 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1144 struct blk_mq_ctx *ctx,
1145 struct request *rq, struct bio *bio)
1147 struct request_queue *q = hctx->queue;
1149 if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE)) {
1150 blk_mq_bio_to_request(rq, bio);
1151 spin_lock(&ctx->lock);
1153 __blk_mq_insert_request(hctx, rq, false);
1154 spin_unlock(&ctx->lock);
1157 spin_lock(&ctx->lock);
1158 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1159 blk_mq_bio_to_request(rq, bio);
1163 spin_unlock(&ctx->lock);
1164 __blk_mq_free_request(hctx, ctx, rq);
1169 struct blk_map_ctx {
1170 struct blk_mq_hw_ctx *hctx;
1171 struct blk_mq_ctx *ctx;
1174 static struct request *blk_mq_map_request(struct request_queue *q,
1176 struct blk_map_ctx *data)
1178 struct blk_mq_hw_ctx *hctx;
1179 struct blk_mq_ctx *ctx;
1181 int rw = bio_data_dir(bio);
1183 if (unlikely(blk_mq_queue_enter(q))) {
1184 bio_endio(bio, -EIO);
1188 ctx = blk_mq_get_ctx(q);
1189 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1191 if (rw_is_sync(bio->bi_rw))
1194 trace_block_getrq(q, bio, rw);
1195 rq = __blk_mq_alloc_request(hctx, ctx, GFP_ATOMIC, false);
1197 blk_mq_rq_ctx_init(q, ctx, rq, rw);
1199 blk_mq_put_ctx(ctx);
1200 trace_block_sleeprq(q, bio, rw);
1201 rq = blk_mq_alloc_request_pinned(q, rw, __GFP_WAIT|GFP_ATOMIC,
1204 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1214 * Multiple hardware queue variant. This will not use per-process plugs,
1215 * but will attempt to bypass the hctx queueing if we can go straight to
1216 * hardware for SYNC IO.
1218 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1220 const int is_sync = rw_is_sync(bio->bi_rw);
1221 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1222 struct blk_map_ctx data;
1225 blk_queue_bounce(q, &bio);
1227 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1228 bio_endio(bio, -EIO);
1232 rq = blk_mq_map_request(q, bio, &data);
1236 if (unlikely(is_flush_fua)) {
1237 blk_mq_bio_to_request(rq, bio);
1238 blk_insert_flush(rq);
1245 blk_mq_bio_to_request(rq, bio);
1246 blk_mq_start_request(rq, true);
1249 * For OK queue, we are done. For error, kill it. Any other
1250 * error (busy), just add it to our list as we previously
1253 ret = q->mq_ops->queue_rq(data.hctx, rq);
1254 if (ret == BLK_MQ_RQ_QUEUE_OK)
1257 __blk_mq_requeue_request(rq);
1259 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1261 blk_mq_end_io(rq, rq->errors);
1267 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1269 * For a SYNC request, send it to the hardware immediately. For
1270 * an ASYNC request, just ensure that we run it later on. The
1271 * latter allows for merging opportunities and more efficient
1275 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1278 blk_mq_put_ctx(data.ctx);
1282 * Single hardware queue variant. This will attempt to use any per-process
1283 * plug for merging and IO deferral.
1285 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1287 const int is_sync = rw_is_sync(bio->bi_rw);
1288 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1289 unsigned int use_plug, request_count = 0;
1290 struct blk_map_ctx data;
1294 * If we have multiple hardware queues, just go directly to
1295 * one of those for sync IO.
1297 use_plug = !is_flush_fua && !is_sync;
1299 blk_queue_bounce(q, &bio);
1301 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1302 bio_endio(bio, -EIO);
1306 if (use_plug && !blk_queue_nomerges(q) &&
1307 blk_attempt_plug_merge(q, bio, &request_count))
1310 rq = blk_mq_map_request(q, bio, &data);
1312 if (unlikely(is_flush_fua)) {
1313 blk_mq_bio_to_request(rq, bio);
1314 blk_insert_flush(rq);
1319 * A task plug currently exists. Since this is completely lockless,
1320 * utilize that to temporarily store requests until the task is
1321 * either done or scheduled away.
1324 struct blk_plug *plug = current->plug;
1327 blk_mq_bio_to_request(rq, bio);
1328 if (list_empty(&plug->mq_list))
1329 trace_block_plug(q);
1330 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1331 blk_flush_plug_list(plug, false);
1332 trace_block_plug(q);
1334 list_add_tail(&rq->queuelist, &plug->mq_list);
1335 blk_mq_put_ctx(data.ctx);
1340 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1342 * For a SYNC request, send it to the hardware immediately. For
1343 * an ASYNC request, just ensure that we run it later on. The
1344 * latter allows for merging opportunities and more efficient
1348 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1351 blk_mq_put_ctx(data.ctx);
1355 * Default mapping to a software queue, since we use one per CPU.
1357 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1359 return q->queue_hw_ctx[q->mq_map[cpu]];
1361 EXPORT_SYMBOL(blk_mq_map_queue);
1363 struct blk_mq_hw_ctx *blk_mq_alloc_single_hw_queue(struct blk_mq_tag_set *set,
1364 unsigned int hctx_index,
1367 return kzalloc_node(sizeof(struct blk_mq_hw_ctx), GFP_KERNEL, node);
1369 EXPORT_SYMBOL(blk_mq_alloc_single_hw_queue);
1371 void blk_mq_free_single_hw_queue(struct blk_mq_hw_ctx *hctx,
1372 unsigned int hctx_index)
1376 EXPORT_SYMBOL(blk_mq_free_single_hw_queue);
1378 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1379 struct blk_mq_tags *tags, unsigned int hctx_idx)
1383 if (tags->rqs && set->ops->exit_request) {
1386 for (i = 0; i < tags->nr_tags; i++) {
1389 set->ops->exit_request(set->driver_data, tags->rqs[i],
1394 while (!list_empty(&tags->page_list)) {
1395 page = list_first_entry(&tags->page_list, struct page, lru);
1396 list_del_init(&page->lru);
1397 __free_pages(page, page->private);
1402 blk_mq_free_tags(tags);
1405 static size_t order_to_size(unsigned int order)
1407 return (size_t)PAGE_SIZE << order;
1410 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1411 unsigned int hctx_idx)
1413 struct blk_mq_tags *tags;
1414 unsigned int i, j, entries_per_page, max_order = 4;
1415 size_t rq_size, left;
1417 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1422 INIT_LIST_HEAD(&tags->page_list);
1424 tags->rqs = kmalloc_node(set->queue_depth * sizeof(struct request *),
1425 GFP_KERNEL, set->numa_node);
1427 blk_mq_free_tags(tags);
1432 * rq_size is the size of the request plus driver payload, rounded
1433 * to the cacheline size
1435 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1437 left = rq_size * set->queue_depth;
1439 for (i = 0; i < set->queue_depth; ) {
1440 int this_order = max_order;
1445 while (left < order_to_size(this_order - 1) && this_order)
1449 page = alloc_pages_node(set->numa_node, GFP_KERNEL,
1455 if (order_to_size(this_order) < rq_size)
1462 page->private = this_order;
1463 list_add_tail(&page->lru, &tags->page_list);
1465 p = page_address(page);
1466 entries_per_page = order_to_size(this_order) / rq_size;
1467 to_do = min(entries_per_page, set->queue_depth - i);
1468 left -= to_do * rq_size;
1469 for (j = 0; j < to_do; j++) {
1471 if (set->ops->init_request) {
1472 if (set->ops->init_request(set->driver_data,
1473 tags->rqs[i], hctx_idx, i,
1486 pr_warn("%s: failed to allocate requests\n", __func__);
1487 blk_mq_free_rq_map(set, tags, hctx_idx);
1491 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1496 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1498 unsigned int bpw = 8, total, num_maps, i;
1500 bitmap->bits_per_word = bpw;
1502 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1503 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1508 bitmap->map_size = num_maps;
1511 for (i = 0; i < num_maps; i++) {
1512 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1513 total -= bitmap->map[i].depth;
1519 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1521 struct request_queue *q = hctx->queue;
1522 struct blk_mq_ctx *ctx;
1526 * Move ctx entries to new CPU, if this one is going away.
1528 ctx = __blk_mq_get_ctx(q, cpu);
1530 spin_lock(&ctx->lock);
1531 if (!list_empty(&ctx->rq_list)) {
1532 list_splice_init(&ctx->rq_list, &tmp);
1533 blk_mq_hctx_clear_pending(hctx, ctx);
1535 spin_unlock(&ctx->lock);
1537 if (list_empty(&tmp))
1540 ctx = blk_mq_get_ctx(q);
1541 spin_lock(&ctx->lock);
1543 while (!list_empty(&tmp)) {
1546 rq = list_first_entry(&tmp, struct request, queuelist);
1548 list_move_tail(&rq->queuelist, &ctx->rq_list);
1551 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1552 blk_mq_hctx_mark_pending(hctx, ctx);
1554 spin_unlock(&ctx->lock);
1556 blk_mq_run_hw_queue(hctx, true);
1557 blk_mq_put_ctx(ctx);
1561 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1563 struct request_queue *q = hctx->queue;
1564 struct blk_mq_tag_set *set = q->tag_set;
1566 if (set->tags[hctx->queue_num])
1569 set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1570 if (!set->tags[hctx->queue_num])
1573 hctx->tags = set->tags[hctx->queue_num];
1577 static int blk_mq_hctx_notify(void *data, unsigned long action,
1580 struct blk_mq_hw_ctx *hctx = data;
1582 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1583 return blk_mq_hctx_cpu_offline(hctx, cpu);
1584 else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1585 return blk_mq_hctx_cpu_online(hctx, cpu);
1590 static void blk_mq_exit_hw_queues(struct request_queue *q,
1591 struct blk_mq_tag_set *set, int nr_queue)
1593 struct blk_mq_hw_ctx *hctx;
1596 queue_for_each_hw_ctx(q, hctx, i) {
1600 if (set->ops->exit_hctx)
1601 set->ops->exit_hctx(hctx, i);
1603 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1605 blk_mq_free_bitmap(&hctx->ctx_map);
1610 static void blk_mq_free_hw_queues(struct request_queue *q,
1611 struct blk_mq_tag_set *set)
1613 struct blk_mq_hw_ctx *hctx;
1616 queue_for_each_hw_ctx(q, hctx, i) {
1617 free_cpumask_var(hctx->cpumask);
1618 set->ops->free_hctx(hctx, i);
1622 static int blk_mq_init_hw_queues(struct request_queue *q,
1623 struct blk_mq_tag_set *set)
1625 struct blk_mq_hw_ctx *hctx;
1629 * Initialize hardware queues
1631 queue_for_each_hw_ctx(q, hctx, i) {
1634 node = hctx->numa_node;
1635 if (node == NUMA_NO_NODE)
1636 node = hctx->numa_node = set->numa_node;
1638 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1639 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1640 spin_lock_init(&hctx->lock);
1641 INIT_LIST_HEAD(&hctx->dispatch);
1643 hctx->queue_num = i;
1644 hctx->flags = set->flags;
1645 hctx->cmd_size = set->cmd_size;
1647 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1648 blk_mq_hctx_notify, hctx);
1649 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1651 hctx->tags = set->tags[i];
1654 * Allocate space for all possible cpus to avoid allocation in
1657 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1662 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1667 if (set->ops->init_hctx &&
1668 set->ops->init_hctx(hctx, set->driver_data, i))
1672 if (i == q->nr_hw_queues)
1678 blk_mq_exit_hw_queues(q, set, i);
1683 static void blk_mq_init_cpu_queues(struct request_queue *q,
1684 unsigned int nr_hw_queues)
1688 for_each_possible_cpu(i) {
1689 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1690 struct blk_mq_hw_ctx *hctx;
1692 memset(__ctx, 0, sizeof(*__ctx));
1694 spin_lock_init(&__ctx->lock);
1695 INIT_LIST_HEAD(&__ctx->rq_list);
1698 /* If the cpu isn't online, the cpu is mapped to first hctx */
1702 hctx = q->mq_ops->map_queue(q, i);
1703 cpumask_set_cpu(i, hctx->cpumask);
1707 * Set local node, IFF we have more than one hw queue. If
1708 * not, we remain on the home node of the device
1710 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1711 hctx->numa_node = cpu_to_node(i);
1715 static void blk_mq_map_swqueue(struct request_queue *q)
1718 struct blk_mq_hw_ctx *hctx;
1719 struct blk_mq_ctx *ctx;
1721 queue_for_each_hw_ctx(q, hctx, i) {
1722 cpumask_clear(hctx->cpumask);
1727 * Map software to hardware queues
1729 queue_for_each_ctx(q, ctx, i) {
1730 /* If the cpu isn't online, the cpu is mapped to first hctx */
1734 hctx = q->mq_ops->map_queue(q, i);
1735 cpumask_set_cpu(i, hctx->cpumask);
1736 ctx->index_hw = hctx->nr_ctx;
1737 hctx->ctxs[hctx->nr_ctx++] = ctx;
1740 queue_for_each_hw_ctx(q, hctx, i) {
1742 * If not software queues are mapped to this hardware queue,
1743 * disable it and free the request entries
1745 if (!hctx->nr_ctx) {
1746 struct blk_mq_tag_set *set = q->tag_set;
1749 blk_mq_free_rq_map(set, set->tags[i], i);
1750 set->tags[i] = NULL;
1757 * Initialize batch roundrobin counts
1759 hctx->next_cpu = cpumask_first(hctx->cpumask);
1760 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1764 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1766 struct blk_mq_hw_ctx *hctx;
1767 struct request_queue *q;
1771 if (set->tag_list.next == set->tag_list.prev)
1776 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1777 blk_mq_freeze_queue(q);
1779 queue_for_each_hw_ctx(q, hctx, i) {
1781 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1783 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1785 blk_mq_unfreeze_queue(q);
1789 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1791 struct blk_mq_tag_set *set = q->tag_set;
1793 blk_mq_freeze_queue(q);
1795 mutex_lock(&set->tag_list_lock);
1796 list_del_init(&q->tag_set_list);
1797 blk_mq_update_tag_set_depth(set);
1798 mutex_unlock(&set->tag_list_lock);
1800 blk_mq_unfreeze_queue(q);
1803 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1804 struct request_queue *q)
1808 mutex_lock(&set->tag_list_lock);
1809 list_add_tail(&q->tag_set_list, &set->tag_list);
1810 blk_mq_update_tag_set_depth(set);
1811 mutex_unlock(&set->tag_list_lock);
1814 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1816 struct blk_mq_hw_ctx **hctxs;
1817 struct blk_mq_ctx *ctx;
1818 struct request_queue *q;
1822 ctx = alloc_percpu(struct blk_mq_ctx);
1824 return ERR_PTR(-ENOMEM);
1826 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1832 map = blk_mq_make_queue_map(set);
1836 for (i = 0; i < set->nr_hw_queues; i++) {
1837 int node = blk_mq_hw_queue_to_node(map, i);
1839 hctxs[i] = set->ops->alloc_hctx(set, i, node);
1843 if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL))
1846 atomic_set(&hctxs[i]->nr_active, 0);
1847 hctxs[i]->numa_node = node;
1848 hctxs[i]->queue_num = i;
1851 q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1855 if (percpu_counter_init(&q->mq_usage_counter, 0))
1858 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1859 blk_queue_rq_timeout(q, 30000);
1861 q->nr_queues = nr_cpu_ids;
1862 q->nr_hw_queues = set->nr_hw_queues;
1866 q->queue_hw_ctx = hctxs;
1868 q->mq_ops = set->ops;
1869 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1871 q->sg_reserved_size = INT_MAX;
1873 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1874 INIT_LIST_HEAD(&q->requeue_list);
1875 spin_lock_init(&q->requeue_lock);
1877 if (q->nr_hw_queues > 1)
1878 blk_queue_make_request(q, blk_mq_make_request);
1880 blk_queue_make_request(q, blk_sq_make_request);
1882 blk_queue_rq_timed_out(q, blk_mq_rq_timed_out);
1884 blk_queue_rq_timeout(q, set->timeout);
1887 * Do this after blk_queue_make_request() overrides it...
1889 q->nr_requests = set->queue_depth;
1891 if (set->ops->complete)
1892 blk_queue_softirq_done(q, set->ops->complete);
1894 blk_mq_init_flush(q);
1895 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1897 q->flush_rq = kzalloc(round_up(sizeof(struct request) +
1898 set->cmd_size, cache_line_size()),
1903 if (blk_mq_init_hw_queues(q, set))
1906 mutex_lock(&all_q_mutex);
1907 list_add_tail(&q->all_q_node, &all_q_list);
1908 mutex_unlock(&all_q_mutex);
1910 blk_mq_add_queue_tag_set(set, q);
1912 blk_mq_map_swqueue(q);
1919 blk_cleanup_queue(q);
1922 for (i = 0; i < set->nr_hw_queues; i++) {
1925 free_cpumask_var(hctxs[i]->cpumask);
1926 set->ops->free_hctx(hctxs[i], i);
1932 return ERR_PTR(-ENOMEM);
1934 EXPORT_SYMBOL(blk_mq_init_queue);
1936 void blk_mq_free_queue(struct request_queue *q)
1938 struct blk_mq_tag_set *set = q->tag_set;
1940 blk_mq_del_queue_tag_set(q);
1942 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
1943 blk_mq_free_hw_queues(q, set);
1945 percpu_counter_destroy(&q->mq_usage_counter);
1947 free_percpu(q->queue_ctx);
1948 kfree(q->queue_hw_ctx);
1951 q->queue_ctx = NULL;
1952 q->queue_hw_ctx = NULL;
1955 mutex_lock(&all_q_mutex);
1956 list_del_init(&q->all_q_node);
1957 mutex_unlock(&all_q_mutex);
1960 /* Basically redo blk_mq_init_queue with queue frozen */
1961 static void blk_mq_queue_reinit(struct request_queue *q)
1963 blk_mq_freeze_queue(q);
1965 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1968 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1969 * we should change hctx numa_node according to new topology (this
1970 * involves free and re-allocate memory, worthy doing?)
1973 blk_mq_map_swqueue(q);
1975 blk_mq_unfreeze_queue(q);
1978 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1979 unsigned long action, void *hcpu)
1981 struct request_queue *q;
1984 * Before new mappings are established, hotadded cpu might already
1985 * start handling requests. This doesn't break anything as we map
1986 * offline CPUs to first hardware queue. We will re-init the queue
1987 * below to get optimal settings.
1989 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1990 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1993 mutex_lock(&all_q_mutex);
1994 list_for_each_entry(q, &all_q_list, all_q_node)
1995 blk_mq_queue_reinit(q);
1996 mutex_unlock(&all_q_mutex);
2000 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2004 if (!set->nr_hw_queues)
2006 if (!set->queue_depth || set->queue_depth > BLK_MQ_MAX_DEPTH)
2008 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2011 if (!set->nr_hw_queues ||
2012 !set->ops->queue_rq || !set->ops->map_queue ||
2013 !set->ops->alloc_hctx || !set->ops->free_hctx)
2017 set->tags = kmalloc_node(set->nr_hw_queues *
2018 sizeof(struct blk_mq_tags *),
2019 GFP_KERNEL, set->numa_node);
2023 for (i = 0; i < set->nr_hw_queues; i++) {
2024 set->tags[i] = blk_mq_init_rq_map(set, i);
2029 mutex_init(&set->tag_list_lock);
2030 INIT_LIST_HEAD(&set->tag_list);
2036 blk_mq_free_rq_map(set, set->tags[i], i);
2040 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2042 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2046 for (i = 0; i < set->nr_hw_queues; i++) {
2048 blk_mq_free_rq_map(set, set->tags[i], i);
2053 EXPORT_SYMBOL(blk_mq_free_tag_set);
2055 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2057 struct blk_mq_tag_set *set = q->tag_set;
2058 struct blk_mq_hw_ctx *hctx;
2061 if (!set || nr > set->queue_depth)
2065 queue_for_each_hw_ctx(q, hctx, i) {
2066 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2072 q->nr_requests = nr;
2077 void blk_mq_disable_hotplug(void)
2079 mutex_lock(&all_q_mutex);
2082 void blk_mq_enable_hotplug(void)
2084 mutex_unlock(&all_q_mutex);
2087 static int __init blk_mq_init(void)
2091 /* Must be called after percpu_counter_hotcpu_callback() */
2092 hotcpu_notifier(blk_mq_queue_reinit_notify, -10);
2096 subsys_initcall(blk_mq_init);