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)
83 __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
85 /* we have problems to freeze the queue if it's initializing */
86 if (!blk_queue_bypass(q) || !blk_queue_init_done(q))
89 __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
91 spin_lock_irq(q->queue_lock);
92 ret = wait_event_interruptible_lock_irq(q->mq_freeze_wq,
93 !blk_queue_bypass(q) || blk_queue_dying(q),
95 /* inc usage with lock hold to avoid freeze_queue runs here */
96 if (!ret && !blk_queue_dying(q))
97 __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
98 else if (blk_queue_dying(q))
100 spin_unlock_irq(q->queue_lock);
105 static void blk_mq_queue_exit(struct request_queue *q)
107 __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
110 static void __blk_mq_drain_queue(struct request_queue *q)
115 spin_lock_irq(q->queue_lock);
116 count = percpu_counter_sum(&q->mq_usage_counter);
117 spin_unlock_irq(q->queue_lock);
121 blk_mq_run_queues(q, false);
127 * Guarantee no request is in use, so we can change any data structure of
128 * the queue afterward.
130 static void blk_mq_freeze_queue(struct request_queue *q)
134 spin_lock_irq(q->queue_lock);
135 drain = !q->bypass_depth++;
136 queue_flag_set(QUEUE_FLAG_BYPASS, q);
137 spin_unlock_irq(q->queue_lock);
140 __blk_mq_drain_queue(q);
143 void blk_mq_drain_queue(struct request_queue *q)
145 __blk_mq_drain_queue(q);
148 static void blk_mq_unfreeze_queue(struct request_queue *q)
152 spin_lock_irq(q->queue_lock);
153 if (!--q->bypass_depth) {
154 queue_flag_clear(QUEUE_FLAG_BYPASS, q);
157 WARN_ON_ONCE(q->bypass_depth < 0);
158 spin_unlock_irq(q->queue_lock);
160 wake_up_all(&q->mq_freeze_wq);
163 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
165 return blk_mq_has_free_tags(hctx->tags);
167 EXPORT_SYMBOL(blk_mq_can_queue);
169 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
170 struct request *rq, unsigned int rw_flags)
172 if (blk_queue_io_stat(q))
173 rw_flags |= REQ_IO_STAT;
175 INIT_LIST_HEAD(&rq->queuelist);
176 /* csd/requeue_work/fifo_time is initialized before use */
179 rq->cmd_flags |= rw_flags;
180 /* do not touch atomic flags, it needs atomic ops against the timer */
182 INIT_HLIST_NODE(&rq->hash);
183 RB_CLEAR_NODE(&rq->rb_node);
186 #ifdef CONFIG_BLK_CGROUP
188 set_start_time_ns(rq);
189 rq->io_start_time_ns = 0;
191 rq->nr_phys_segments = 0;
192 #if defined(CONFIG_BLK_DEV_INTEGRITY)
193 rq->nr_integrity_segments = 0;
196 /* tag was already set */
204 INIT_LIST_HEAD(&rq->timeout_list);
206 rq->end_io_data = NULL;
209 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
212 static struct request *
213 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
218 tag = blk_mq_get_tag(data);
219 if (tag != BLK_MQ_TAG_FAIL) {
220 rq = data->hctx->tags->rqs[tag];
223 if (blk_mq_tag_busy(data->hctx)) {
224 rq->cmd_flags = REQ_MQ_INFLIGHT;
225 atomic_inc(&data->hctx->nr_active);
229 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
236 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
239 struct blk_mq_ctx *ctx;
240 struct blk_mq_hw_ctx *hctx;
242 struct blk_mq_alloc_data alloc_data;
244 if (blk_mq_queue_enter(q))
247 ctx = blk_mq_get_ctx(q);
248 hctx = q->mq_ops->map_queue(q, ctx->cpu);
249 blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT,
250 reserved, ctx, hctx);
252 rq = __blk_mq_alloc_request(&alloc_data, rw);
253 if (!rq && (gfp & __GFP_WAIT)) {
254 __blk_mq_run_hw_queue(hctx);
257 ctx = blk_mq_get_ctx(q);
258 hctx = q->mq_ops->map_queue(q, ctx->cpu);
259 blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx,
261 rq = __blk_mq_alloc_request(&alloc_data, rw);
262 ctx = alloc_data.ctx;
267 EXPORT_SYMBOL(blk_mq_alloc_request);
269 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
270 struct blk_mq_ctx *ctx, struct request *rq)
272 const int tag = rq->tag;
273 struct request_queue *q = rq->q;
275 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
276 atomic_dec(&hctx->nr_active);
278 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
279 blk_mq_put_tag(hctx, tag, &ctx->last_tag);
280 blk_mq_queue_exit(q);
283 void blk_mq_free_request(struct request *rq)
285 struct blk_mq_ctx *ctx = rq->mq_ctx;
286 struct blk_mq_hw_ctx *hctx;
287 struct request_queue *q = rq->q;
289 ctx->rq_completed[rq_is_sync(rq)]++;
291 hctx = q->mq_ops->map_queue(q, ctx->cpu);
292 __blk_mq_free_request(hctx, ctx, rq);
296 * Clone all relevant state from a request that has been put on hold in
297 * the flush state machine into the preallocated flush request that hangs
298 * off the request queue.
300 * For a driver the flush request should be invisible, that's why we are
301 * impersonating the original request here.
303 void blk_mq_clone_flush_request(struct request *flush_rq,
304 struct request *orig_rq)
306 struct blk_mq_hw_ctx *hctx =
307 orig_rq->q->mq_ops->map_queue(orig_rq->q, orig_rq->mq_ctx->cpu);
309 flush_rq->mq_ctx = orig_rq->mq_ctx;
310 flush_rq->tag = orig_rq->tag;
311 memcpy(blk_mq_rq_to_pdu(flush_rq), blk_mq_rq_to_pdu(orig_rq),
315 inline void __blk_mq_end_io(struct request *rq, int error)
317 blk_account_io_done(rq);
320 rq->end_io(rq, error);
322 if (unlikely(blk_bidi_rq(rq)))
323 blk_mq_free_request(rq->next_rq);
324 blk_mq_free_request(rq);
327 EXPORT_SYMBOL(__blk_mq_end_io);
329 void blk_mq_end_io(struct request *rq, int error)
331 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
333 __blk_mq_end_io(rq, error);
335 EXPORT_SYMBOL(blk_mq_end_io);
337 static void __blk_mq_complete_request_remote(void *data)
339 struct request *rq = data;
341 rq->q->softirq_done_fn(rq);
344 static void blk_mq_ipi_complete_request(struct request *rq)
346 struct blk_mq_ctx *ctx = rq->mq_ctx;
350 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
351 rq->q->softirq_done_fn(rq);
356 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
357 shared = cpus_share_cache(cpu, ctx->cpu);
359 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
360 rq->csd.func = __blk_mq_complete_request_remote;
363 smp_call_function_single_async(ctx->cpu, &rq->csd);
365 rq->q->softirq_done_fn(rq);
370 void __blk_mq_complete_request(struct request *rq)
372 struct request_queue *q = rq->q;
374 if (!q->softirq_done_fn)
375 blk_mq_end_io(rq, rq->errors);
377 blk_mq_ipi_complete_request(rq);
381 * blk_mq_complete_request - end I/O on a request
382 * @rq: the request being processed
385 * Ends all I/O on a request. It does not handle partial completions.
386 * The actual completion happens out-of-order, through a IPI handler.
388 void blk_mq_complete_request(struct request *rq)
390 struct request_queue *q = rq->q;
392 if (unlikely(blk_should_fake_timeout(q)))
394 if (!blk_mark_rq_complete(rq))
395 __blk_mq_complete_request(rq);
397 EXPORT_SYMBOL(blk_mq_complete_request);
399 static void blk_mq_start_request(struct request *rq, bool last)
401 struct request_queue *q = rq->q;
403 trace_block_rq_issue(q, rq);
405 rq->resid_len = blk_rq_bytes(rq);
406 if (unlikely(blk_bidi_rq(rq)))
407 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
410 * Just mark start time and set the started bit. Due to memory
411 * ordering, we know we'll see the correct deadline as long as
412 * REQ_ATOMIC_STARTED is seen. Use the default queue timeout,
413 * unless one has been set in the request.
416 rq->deadline = jiffies + q->rq_timeout;
418 rq->deadline = jiffies + rq->timeout;
421 * Mark us as started and clear complete. Complete might have been
422 * set if requeue raced with timeout, which then marked it as
423 * complete. So be sure to clear complete again when we start
424 * the request, otherwise we'll ignore the completion event.
426 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
427 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
428 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
429 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
431 if (q->dma_drain_size && blk_rq_bytes(rq)) {
433 * Make sure space for the drain appears. We know we can do
434 * this because max_hw_segments has been adjusted to be one
435 * fewer than the device can handle.
437 rq->nr_phys_segments++;
441 * Flag the last request in the series so that drivers know when IO
442 * should be kicked off, if they don't do it on a per-request basis.
444 * Note: the flag isn't the only condition drivers should do kick off.
445 * If drive is busy, the last request might not have the bit set.
448 rq->cmd_flags |= REQ_END;
451 static void __blk_mq_requeue_request(struct request *rq)
453 struct request_queue *q = rq->q;
455 trace_block_rq_requeue(q, rq);
456 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
458 rq->cmd_flags &= ~REQ_END;
460 if (q->dma_drain_size && blk_rq_bytes(rq))
461 rq->nr_phys_segments--;
464 void blk_mq_requeue_request(struct request *rq)
466 __blk_mq_requeue_request(rq);
467 blk_clear_rq_complete(rq);
469 BUG_ON(blk_queued_rq(rq));
470 blk_mq_add_to_requeue_list(rq, true);
472 EXPORT_SYMBOL(blk_mq_requeue_request);
474 static void blk_mq_requeue_work(struct work_struct *work)
476 struct request_queue *q =
477 container_of(work, struct request_queue, requeue_work);
479 struct request *rq, *next;
482 spin_lock_irqsave(&q->requeue_lock, flags);
483 list_splice_init(&q->requeue_list, &rq_list);
484 spin_unlock_irqrestore(&q->requeue_lock, flags);
486 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
487 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
490 rq->cmd_flags &= ~REQ_SOFTBARRIER;
491 list_del_init(&rq->queuelist);
492 blk_mq_insert_request(rq, true, false, false);
495 while (!list_empty(&rq_list)) {
496 rq = list_entry(rq_list.next, struct request, queuelist);
497 list_del_init(&rq->queuelist);
498 blk_mq_insert_request(rq, false, false, false);
501 blk_mq_run_queues(q, false);
504 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
506 struct request_queue *q = rq->q;
510 * We abuse this flag that is otherwise used by the I/O scheduler to
511 * request head insertation from the workqueue.
513 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
515 spin_lock_irqsave(&q->requeue_lock, flags);
517 rq->cmd_flags |= REQ_SOFTBARRIER;
518 list_add(&rq->queuelist, &q->requeue_list);
520 list_add_tail(&rq->queuelist, &q->requeue_list);
522 spin_unlock_irqrestore(&q->requeue_lock, flags);
524 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
526 void blk_mq_kick_requeue_list(struct request_queue *q)
528 kblockd_schedule_work(&q->requeue_work);
530 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
532 static inline bool is_flush_request(struct request *rq, unsigned int tag)
534 return ((rq->cmd_flags & REQ_FLUSH_SEQ) &&
535 rq->q->flush_rq->tag == tag);
538 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
540 struct request *rq = tags->rqs[tag];
542 if (!is_flush_request(rq, tag))
545 return rq->q->flush_rq;
547 EXPORT_SYMBOL(blk_mq_tag_to_rq);
549 struct blk_mq_timeout_data {
550 struct blk_mq_hw_ctx *hctx;
552 unsigned int *next_set;
555 static void blk_mq_timeout_check(void *__data, unsigned long *free_tags)
557 struct blk_mq_timeout_data *data = __data;
558 struct blk_mq_hw_ctx *hctx = data->hctx;
561 /* It may not be in flight yet (this is where
562 * the REQ_ATOMIC_STARTED flag comes in). The requests are
563 * statically allocated, so we know it's always safe to access the
564 * memory associated with a bit offset into ->rqs[].
570 tag = find_next_zero_bit(free_tags, hctx->tags->nr_tags, tag);
571 if (tag >= hctx->tags->nr_tags)
574 rq = blk_mq_tag_to_rq(hctx->tags, tag++);
575 if (rq->q != hctx->queue)
577 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
580 blk_rq_check_expired(rq, data->next, data->next_set);
584 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx,
586 unsigned int *next_set)
588 struct blk_mq_timeout_data data = {
591 .next_set = next_set,
595 * Ask the tagging code to iterate busy requests, so we can
596 * check them for timeout.
598 blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data);
601 static enum blk_eh_timer_return blk_mq_rq_timed_out(struct request *rq)
603 struct request_queue *q = rq->q;
606 * We know that complete is set at this point. If STARTED isn't set
607 * anymore, then the request isn't active and the "timeout" should
608 * just be ignored. This can happen due to the bitflag ordering.
609 * Timeout first checks if STARTED is set, and if it is, assumes
610 * the request is active. But if we race with completion, then
611 * we both flags will get cleared. So check here again, and ignore
612 * a timeout event with a request that isn't active.
614 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
615 return BLK_EH_NOT_HANDLED;
617 if (!q->mq_ops->timeout)
618 return BLK_EH_RESET_TIMER;
620 return q->mq_ops->timeout(rq);
623 static void blk_mq_rq_timer(unsigned long data)
625 struct request_queue *q = (struct request_queue *) data;
626 struct blk_mq_hw_ctx *hctx;
627 unsigned long next = 0;
630 queue_for_each_hw_ctx(q, hctx, i) {
632 * If not software queues are currently mapped to this
633 * hardware queue, there's nothing to check
635 if (!hctx->nr_ctx || !hctx->tags)
638 blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set);
642 next = blk_rq_timeout(round_jiffies_up(next));
643 mod_timer(&q->timeout, next);
645 queue_for_each_hw_ctx(q, hctx, i)
646 blk_mq_tag_idle(hctx);
651 * Reverse check our software queue for entries that we could potentially
652 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
653 * too much time checking for merges.
655 static bool blk_mq_attempt_merge(struct request_queue *q,
656 struct blk_mq_ctx *ctx, struct bio *bio)
661 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
667 if (!blk_rq_merge_ok(rq, bio))
670 el_ret = blk_try_merge(rq, bio);
671 if (el_ret == ELEVATOR_BACK_MERGE) {
672 if (bio_attempt_back_merge(q, rq, bio)) {
677 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
678 if (bio_attempt_front_merge(q, rq, bio)) {
690 * Process software queues that have been marked busy, splicing them
691 * to the for-dispatch
693 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
695 struct blk_mq_ctx *ctx;
698 for (i = 0; i < hctx->ctx_map.map_size; i++) {
699 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
700 unsigned int off, bit;
706 off = i * hctx->ctx_map.bits_per_word;
708 bit = find_next_bit(&bm->word, bm->depth, bit);
709 if (bit >= bm->depth)
712 ctx = hctx->ctxs[bit + off];
713 clear_bit(bit, &bm->word);
714 spin_lock(&ctx->lock);
715 list_splice_tail_init(&ctx->rq_list, list);
716 spin_unlock(&ctx->lock);
724 * Run this hardware queue, pulling any software queues mapped to it in.
725 * Note that this function currently has various problems around ordering
726 * of IO. In particular, we'd like FIFO behaviour on handling existing
727 * items on the hctx->dispatch list. Ignore that for now.
729 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
731 struct request_queue *q = hctx->queue;
736 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
738 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
744 * Touch any software queue that has pending entries.
746 flush_busy_ctxs(hctx, &rq_list);
749 * If we have previous entries on our dispatch list, grab them
750 * and stuff them at the front for more fair dispatch.
752 if (!list_empty_careful(&hctx->dispatch)) {
753 spin_lock(&hctx->lock);
754 if (!list_empty(&hctx->dispatch))
755 list_splice_init(&hctx->dispatch, &rq_list);
756 spin_unlock(&hctx->lock);
760 * Now process all the entries, sending them to the driver.
763 while (!list_empty(&rq_list)) {
766 rq = list_first_entry(&rq_list, struct request, queuelist);
767 list_del_init(&rq->queuelist);
769 blk_mq_start_request(rq, list_empty(&rq_list));
771 ret = q->mq_ops->queue_rq(hctx, rq);
773 case BLK_MQ_RQ_QUEUE_OK:
776 case BLK_MQ_RQ_QUEUE_BUSY:
777 list_add(&rq->queuelist, &rq_list);
778 __blk_mq_requeue_request(rq);
781 pr_err("blk-mq: bad return on queue: %d\n", ret);
782 case BLK_MQ_RQ_QUEUE_ERROR:
784 blk_mq_end_io(rq, rq->errors);
788 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
793 hctx->dispatched[0]++;
794 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
795 hctx->dispatched[ilog2(queued) + 1]++;
798 * Any items that need requeuing? Stuff them into hctx->dispatch,
799 * that is where we will continue on next queue run.
801 if (!list_empty(&rq_list)) {
802 spin_lock(&hctx->lock);
803 list_splice(&rq_list, &hctx->dispatch);
804 spin_unlock(&hctx->lock);
809 * It'd be great if the workqueue API had a way to pass
810 * in a mask and had some smarts for more clever placement.
811 * For now we just round-robin here, switching for every
812 * BLK_MQ_CPU_WORK_BATCH queued items.
814 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
816 int cpu = hctx->next_cpu;
818 if (--hctx->next_cpu_batch <= 0) {
821 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
822 if (next_cpu >= nr_cpu_ids)
823 next_cpu = cpumask_first(hctx->cpumask);
825 hctx->next_cpu = next_cpu;
826 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
832 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
834 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
837 if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask))
838 __blk_mq_run_hw_queue(hctx);
839 else if (hctx->queue->nr_hw_queues == 1)
840 kblockd_schedule_delayed_work(&hctx->run_work, 0);
844 cpu = blk_mq_hctx_next_cpu(hctx);
845 kblockd_schedule_delayed_work_on(cpu, &hctx->run_work, 0);
849 void blk_mq_run_queues(struct request_queue *q, bool async)
851 struct blk_mq_hw_ctx *hctx;
854 queue_for_each_hw_ctx(q, hctx, i) {
855 if ((!blk_mq_hctx_has_pending(hctx) &&
856 list_empty_careful(&hctx->dispatch)) ||
857 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
861 blk_mq_run_hw_queue(hctx, async);
865 EXPORT_SYMBOL(blk_mq_run_queues);
867 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
869 cancel_delayed_work(&hctx->run_work);
870 cancel_delayed_work(&hctx->delay_work);
871 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
873 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
875 void blk_mq_stop_hw_queues(struct request_queue *q)
877 struct blk_mq_hw_ctx *hctx;
880 queue_for_each_hw_ctx(q, hctx, i)
881 blk_mq_stop_hw_queue(hctx);
883 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
885 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
887 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
890 __blk_mq_run_hw_queue(hctx);
893 EXPORT_SYMBOL(blk_mq_start_hw_queue);
895 void blk_mq_start_hw_queues(struct request_queue *q)
897 struct blk_mq_hw_ctx *hctx;
900 queue_for_each_hw_ctx(q, hctx, i)
901 blk_mq_start_hw_queue(hctx);
903 EXPORT_SYMBOL(blk_mq_start_hw_queues);
906 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
908 struct blk_mq_hw_ctx *hctx;
911 queue_for_each_hw_ctx(q, hctx, i) {
912 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
915 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
917 blk_mq_run_hw_queue(hctx, async);
921 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
923 static void blk_mq_run_work_fn(struct work_struct *work)
925 struct blk_mq_hw_ctx *hctx;
927 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
929 __blk_mq_run_hw_queue(hctx);
932 static void blk_mq_delay_work_fn(struct work_struct *work)
934 struct blk_mq_hw_ctx *hctx;
936 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
938 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
939 __blk_mq_run_hw_queue(hctx);
942 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
944 unsigned long tmo = msecs_to_jiffies(msecs);
946 if (hctx->queue->nr_hw_queues == 1)
947 kblockd_schedule_delayed_work(&hctx->delay_work, tmo);
951 cpu = blk_mq_hctx_next_cpu(hctx);
952 kblockd_schedule_delayed_work_on(cpu, &hctx->delay_work, tmo);
955 EXPORT_SYMBOL(blk_mq_delay_queue);
957 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
958 struct request *rq, bool at_head)
960 struct blk_mq_ctx *ctx = rq->mq_ctx;
962 trace_block_rq_insert(hctx->queue, rq);
965 list_add(&rq->queuelist, &ctx->rq_list);
967 list_add_tail(&rq->queuelist, &ctx->rq_list);
969 blk_mq_hctx_mark_pending(hctx, ctx);
972 * We do this early, to ensure we are on the right CPU.
977 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
980 struct request_queue *q = rq->q;
981 struct blk_mq_hw_ctx *hctx;
982 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
984 current_ctx = blk_mq_get_ctx(q);
985 if (!cpu_online(ctx->cpu))
986 rq->mq_ctx = ctx = current_ctx;
988 hctx = q->mq_ops->map_queue(q, ctx->cpu);
990 if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA) &&
991 !(rq->cmd_flags & (REQ_FLUSH_SEQ))) {
992 blk_insert_flush(rq);
994 spin_lock(&ctx->lock);
995 __blk_mq_insert_request(hctx, rq, at_head);
996 spin_unlock(&ctx->lock);
1000 blk_mq_run_hw_queue(hctx, async);
1002 blk_mq_put_ctx(current_ctx);
1005 static void blk_mq_insert_requests(struct request_queue *q,
1006 struct blk_mq_ctx *ctx,
1007 struct list_head *list,
1012 struct blk_mq_hw_ctx *hctx;
1013 struct blk_mq_ctx *current_ctx;
1015 trace_block_unplug(q, depth, !from_schedule);
1017 current_ctx = blk_mq_get_ctx(q);
1019 if (!cpu_online(ctx->cpu))
1021 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1024 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1027 spin_lock(&ctx->lock);
1028 while (!list_empty(list)) {
1031 rq = list_first_entry(list, struct request, queuelist);
1032 list_del_init(&rq->queuelist);
1034 __blk_mq_insert_request(hctx, rq, false);
1036 spin_unlock(&ctx->lock);
1038 blk_mq_run_hw_queue(hctx, from_schedule);
1039 blk_mq_put_ctx(current_ctx);
1042 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1044 struct request *rqa = container_of(a, struct request, queuelist);
1045 struct request *rqb = container_of(b, struct request, queuelist);
1047 return !(rqa->mq_ctx < rqb->mq_ctx ||
1048 (rqa->mq_ctx == rqb->mq_ctx &&
1049 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1052 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1054 struct blk_mq_ctx *this_ctx;
1055 struct request_queue *this_q;
1058 LIST_HEAD(ctx_list);
1061 list_splice_init(&plug->mq_list, &list);
1063 list_sort(NULL, &list, plug_ctx_cmp);
1069 while (!list_empty(&list)) {
1070 rq = list_entry_rq(list.next);
1071 list_del_init(&rq->queuelist);
1073 if (rq->mq_ctx != this_ctx) {
1075 blk_mq_insert_requests(this_q, this_ctx,
1080 this_ctx = rq->mq_ctx;
1086 list_add_tail(&rq->queuelist, &ctx_list);
1090 * If 'this_ctx' is set, we know we have entries to complete
1091 * on 'ctx_list'. Do those.
1094 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1099 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1101 init_request_from_bio(rq, bio);
1103 if (blk_do_io_stat(rq)) {
1104 rq->start_time = jiffies;
1105 blk_account_io_start(rq, 1);
1109 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1110 struct blk_mq_ctx *ctx,
1111 struct request *rq, struct bio *bio)
1113 struct request_queue *q = hctx->queue;
1115 if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE)) {
1116 blk_mq_bio_to_request(rq, bio);
1117 spin_lock(&ctx->lock);
1119 __blk_mq_insert_request(hctx, rq, false);
1120 spin_unlock(&ctx->lock);
1123 spin_lock(&ctx->lock);
1124 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1125 blk_mq_bio_to_request(rq, bio);
1129 spin_unlock(&ctx->lock);
1130 __blk_mq_free_request(hctx, ctx, rq);
1135 struct blk_map_ctx {
1136 struct blk_mq_hw_ctx *hctx;
1137 struct blk_mq_ctx *ctx;
1140 static struct request *blk_mq_map_request(struct request_queue *q,
1142 struct blk_map_ctx *data)
1144 struct blk_mq_hw_ctx *hctx;
1145 struct blk_mq_ctx *ctx;
1147 int rw = bio_data_dir(bio);
1148 struct blk_mq_alloc_data alloc_data;
1150 if (unlikely(blk_mq_queue_enter(q))) {
1151 bio_endio(bio, -EIO);
1155 ctx = blk_mq_get_ctx(q);
1156 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1158 if (rw_is_sync(bio->bi_rw))
1161 trace_block_getrq(q, bio, rw);
1162 blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1164 rq = __blk_mq_alloc_request(&alloc_data, rw);
1165 if (unlikely(!rq)) {
1166 __blk_mq_run_hw_queue(hctx);
1167 blk_mq_put_ctx(ctx);
1168 trace_block_sleeprq(q, bio, rw);
1170 ctx = blk_mq_get_ctx(q);
1171 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1172 blk_mq_set_alloc_data(&alloc_data, q,
1173 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1174 rq = __blk_mq_alloc_request(&alloc_data, rw);
1175 ctx = alloc_data.ctx;
1176 hctx = alloc_data.hctx;
1186 * Multiple hardware queue variant. This will not use per-process plugs,
1187 * but will attempt to bypass the hctx queueing if we can go straight to
1188 * hardware for SYNC IO.
1190 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1192 const int is_sync = rw_is_sync(bio->bi_rw);
1193 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1194 struct blk_map_ctx data;
1197 blk_queue_bounce(q, &bio);
1199 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1200 bio_endio(bio, -EIO);
1204 rq = blk_mq_map_request(q, bio, &data);
1208 if (unlikely(is_flush_fua)) {
1209 blk_mq_bio_to_request(rq, bio);
1210 blk_insert_flush(rq);
1217 blk_mq_bio_to_request(rq, bio);
1218 blk_mq_start_request(rq, true);
1222 * For OK queue, we are done. For error, kill it. Any other
1223 * error (busy), just add it to our list as we previously
1226 ret = q->mq_ops->queue_rq(data.hctx, rq);
1227 if (ret == BLK_MQ_RQ_QUEUE_OK)
1230 __blk_mq_requeue_request(rq);
1232 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1234 blk_mq_end_io(rq, rq->errors);
1240 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1242 * For a SYNC request, send it to the hardware immediately. For
1243 * an ASYNC request, just ensure that we run it later on. The
1244 * latter allows for merging opportunities and more efficient
1248 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1251 blk_mq_put_ctx(data.ctx);
1255 * Single hardware queue variant. This will attempt to use any per-process
1256 * plug for merging and IO deferral.
1258 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1260 const int is_sync = rw_is_sync(bio->bi_rw);
1261 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1262 unsigned int use_plug, request_count = 0;
1263 struct blk_map_ctx data;
1267 * If we have multiple hardware queues, just go directly to
1268 * one of those for sync IO.
1270 use_plug = !is_flush_fua && !is_sync;
1272 blk_queue_bounce(q, &bio);
1274 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1275 bio_endio(bio, -EIO);
1279 if (use_plug && !blk_queue_nomerges(q) &&
1280 blk_attempt_plug_merge(q, bio, &request_count))
1283 rq = blk_mq_map_request(q, bio, &data);
1287 if (unlikely(is_flush_fua)) {
1288 blk_mq_bio_to_request(rq, bio);
1289 blk_insert_flush(rq);
1294 * A task plug currently exists. Since this is completely lockless,
1295 * utilize that to temporarily store requests until the task is
1296 * either done or scheduled away.
1299 struct blk_plug *plug = current->plug;
1302 blk_mq_bio_to_request(rq, bio);
1303 if (list_empty(&plug->mq_list))
1304 trace_block_plug(q);
1305 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1306 blk_flush_plug_list(plug, false);
1307 trace_block_plug(q);
1309 list_add_tail(&rq->queuelist, &plug->mq_list);
1310 blk_mq_put_ctx(data.ctx);
1315 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1317 * For a SYNC request, send it to the hardware immediately. For
1318 * an ASYNC request, just ensure that we run it later on. The
1319 * latter allows for merging opportunities and more efficient
1323 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1326 blk_mq_put_ctx(data.ctx);
1330 * Default mapping to a software queue, since we use one per CPU.
1332 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1334 return q->queue_hw_ctx[q->mq_map[cpu]];
1336 EXPORT_SYMBOL(blk_mq_map_queue);
1338 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1339 struct blk_mq_tags *tags, unsigned int hctx_idx)
1343 if (tags->rqs && set->ops->exit_request) {
1346 for (i = 0; i < tags->nr_tags; i++) {
1349 set->ops->exit_request(set->driver_data, tags->rqs[i],
1354 while (!list_empty(&tags->page_list)) {
1355 page = list_first_entry(&tags->page_list, struct page, lru);
1356 list_del_init(&page->lru);
1357 __free_pages(page, page->private);
1362 blk_mq_free_tags(tags);
1365 static size_t order_to_size(unsigned int order)
1367 return (size_t)PAGE_SIZE << order;
1370 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1371 unsigned int hctx_idx)
1373 struct blk_mq_tags *tags;
1374 unsigned int i, j, entries_per_page, max_order = 4;
1375 size_t rq_size, left;
1377 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1382 INIT_LIST_HEAD(&tags->page_list);
1384 tags->rqs = kmalloc_node(set->queue_depth * sizeof(struct request *),
1385 GFP_KERNEL, set->numa_node);
1387 blk_mq_free_tags(tags);
1392 * rq_size is the size of the request plus driver payload, rounded
1393 * to the cacheline size
1395 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1397 left = rq_size * set->queue_depth;
1399 for (i = 0; i < set->queue_depth; ) {
1400 int this_order = max_order;
1405 while (left < order_to_size(this_order - 1) && this_order)
1409 page = alloc_pages_node(set->numa_node, GFP_KERNEL,
1415 if (order_to_size(this_order) < rq_size)
1422 page->private = this_order;
1423 list_add_tail(&page->lru, &tags->page_list);
1425 p = page_address(page);
1426 entries_per_page = order_to_size(this_order) / rq_size;
1427 to_do = min(entries_per_page, set->queue_depth - i);
1428 left -= to_do * rq_size;
1429 for (j = 0; j < to_do; j++) {
1431 if (set->ops->init_request) {
1432 if (set->ops->init_request(set->driver_data,
1433 tags->rqs[i], hctx_idx, i,
1446 pr_warn("%s: failed to allocate requests\n", __func__);
1447 blk_mq_free_rq_map(set, tags, hctx_idx);
1451 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1456 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1458 unsigned int bpw = 8, total, num_maps, i;
1460 bitmap->bits_per_word = bpw;
1462 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1463 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1468 bitmap->map_size = num_maps;
1471 for (i = 0; i < num_maps; i++) {
1472 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1473 total -= bitmap->map[i].depth;
1479 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1481 struct request_queue *q = hctx->queue;
1482 struct blk_mq_ctx *ctx;
1486 * Move ctx entries to new CPU, if this one is going away.
1488 ctx = __blk_mq_get_ctx(q, cpu);
1490 spin_lock(&ctx->lock);
1491 if (!list_empty(&ctx->rq_list)) {
1492 list_splice_init(&ctx->rq_list, &tmp);
1493 blk_mq_hctx_clear_pending(hctx, ctx);
1495 spin_unlock(&ctx->lock);
1497 if (list_empty(&tmp))
1500 ctx = blk_mq_get_ctx(q);
1501 spin_lock(&ctx->lock);
1503 while (!list_empty(&tmp)) {
1506 rq = list_first_entry(&tmp, struct request, queuelist);
1508 list_move_tail(&rq->queuelist, &ctx->rq_list);
1511 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1512 blk_mq_hctx_mark_pending(hctx, ctx);
1514 spin_unlock(&ctx->lock);
1516 blk_mq_run_hw_queue(hctx, true);
1517 blk_mq_put_ctx(ctx);
1521 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1523 struct request_queue *q = hctx->queue;
1524 struct blk_mq_tag_set *set = q->tag_set;
1526 if (set->tags[hctx->queue_num])
1529 set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1530 if (!set->tags[hctx->queue_num])
1533 hctx->tags = set->tags[hctx->queue_num];
1537 static int blk_mq_hctx_notify(void *data, unsigned long action,
1540 struct blk_mq_hw_ctx *hctx = data;
1542 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1543 return blk_mq_hctx_cpu_offline(hctx, cpu);
1544 else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1545 return blk_mq_hctx_cpu_online(hctx, cpu);
1550 static void blk_mq_exit_hw_queues(struct request_queue *q,
1551 struct blk_mq_tag_set *set, int nr_queue)
1553 struct blk_mq_hw_ctx *hctx;
1556 queue_for_each_hw_ctx(q, hctx, i) {
1560 blk_mq_tag_idle(hctx);
1562 if (set->ops->exit_hctx)
1563 set->ops->exit_hctx(hctx, i);
1565 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1567 blk_mq_free_bitmap(&hctx->ctx_map);
1572 static void blk_mq_free_hw_queues(struct request_queue *q,
1573 struct blk_mq_tag_set *set)
1575 struct blk_mq_hw_ctx *hctx;
1578 queue_for_each_hw_ctx(q, hctx, i) {
1579 free_cpumask_var(hctx->cpumask);
1584 static int blk_mq_init_hw_queues(struct request_queue *q,
1585 struct blk_mq_tag_set *set)
1587 struct blk_mq_hw_ctx *hctx;
1591 * Initialize hardware queues
1593 queue_for_each_hw_ctx(q, hctx, i) {
1596 node = hctx->numa_node;
1597 if (node == NUMA_NO_NODE)
1598 node = hctx->numa_node = set->numa_node;
1600 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1601 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1602 spin_lock_init(&hctx->lock);
1603 INIT_LIST_HEAD(&hctx->dispatch);
1605 hctx->queue_num = i;
1606 hctx->flags = set->flags;
1607 hctx->cmd_size = set->cmd_size;
1609 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1610 blk_mq_hctx_notify, hctx);
1611 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1613 hctx->tags = set->tags[i];
1616 * Allocate space for all possible cpus to avoid allocation in
1619 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1624 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1629 if (set->ops->init_hctx &&
1630 set->ops->init_hctx(hctx, set->driver_data, i))
1634 if (i == q->nr_hw_queues)
1640 blk_mq_exit_hw_queues(q, set, i);
1645 static void blk_mq_init_cpu_queues(struct request_queue *q,
1646 unsigned int nr_hw_queues)
1650 for_each_possible_cpu(i) {
1651 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1652 struct blk_mq_hw_ctx *hctx;
1654 memset(__ctx, 0, sizeof(*__ctx));
1656 spin_lock_init(&__ctx->lock);
1657 INIT_LIST_HEAD(&__ctx->rq_list);
1660 /* If the cpu isn't online, the cpu is mapped to first hctx */
1664 hctx = q->mq_ops->map_queue(q, i);
1665 cpumask_set_cpu(i, hctx->cpumask);
1669 * Set local node, IFF we have more than one hw queue. If
1670 * not, we remain on the home node of the device
1672 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1673 hctx->numa_node = cpu_to_node(i);
1677 static void blk_mq_map_swqueue(struct request_queue *q)
1680 struct blk_mq_hw_ctx *hctx;
1681 struct blk_mq_ctx *ctx;
1683 queue_for_each_hw_ctx(q, hctx, i) {
1684 cpumask_clear(hctx->cpumask);
1689 * Map software to hardware queues
1691 queue_for_each_ctx(q, ctx, i) {
1692 /* If the cpu isn't online, the cpu is mapped to first hctx */
1696 hctx = q->mq_ops->map_queue(q, i);
1697 cpumask_set_cpu(i, hctx->cpumask);
1698 ctx->index_hw = hctx->nr_ctx;
1699 hctx->ctxs[hctx->nr_ctx++] = ctx;
1702 queue_for_each_hw_ctx(q, hctx, i) {
1704 * If not software queues are mapped to this hardware queue,
1705 * disable it and free the request entries
1707 if (!hctx->nr_ctx) {
1708 struct blk_mq_tag_set *set = q->tag_set;
1711 blk_mq_free_rq_map(set, set->tags[i], i);
1712 set->tags[i] = NULL;
1719 * Initialize batch roundrobin counts
1721 hctx->next_cpu = cpumask_first(hctx->cpumask);
1722 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1726 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1728 struct blk_mq_hw_ctx *hctx;
1729 struct request_queue *q;
1733 if (set->tag_list.next == set->tag_list.prev)
1738 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1739 blk_mq_freeze_queue(q);
1741 queue_for_each_hw_ctx(q, hctx, i) {
1743 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1745 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1747 blk_mq_unfreeze_queue(q);
1751 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1753 struct blk_mq_tag_set *set = q->tag_set;
1755 blk_mq_freeze_queue(q);
1757 mutex_lock(&set->tag_list_lock);
1758 list_del_init(&q->tag_set_list);
1759 blk_mq_update_tag_set_depth(set);
1760 mutex_unlock(&set->tag_list_lock);
1762 blk_mq_unfreeze_queue(q);
1765 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1766 struct request_queue *q)
1770 mutex_lock(&set->tag_list_lock);
1771 list_add_tail(&q->tag_set_list, &set->tag_list);
1772 blk_mq_update_tag_set_depth(set);
1773 mutex_unlock(&set->tag_list_lock);
1776 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1778 struct blk_mq_hw_ctx **hctxs;
1779 struct blk_mq_ctx __percpu *ctx;
1780 struct request_queue *q;
1784 ctx = alloc_percpu(struct blk_mq_ctx);
1786 return ERR_PTR(-ENOMEM);
1788 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1794 map = blk_mq_make_queue_map(set);
1798 for (i = 0; i < set->nr_hw_queues; i++) {
1799 int node = blk_mq_hw_queue_to_node(map, i);
1801 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1806 if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL))
1809 atomic_set(&hctxs[i]->nr_active, 0);
1810 hctxs[i]->numa_node = node;
1811 hctxs[i]->queue_num = i;
1814 q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1818 if (percpu_counter_init(&q->mq_usage_counter, 0))
1821 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1822 blk_queue_rq_timeout(q, 30000);
1824 q->nr_queues = nr_cpu_ids;
1825 q->nr_hw_queues = set->nr_hw_queues;
1829 q->queue_hw_ctx = hctxs;
1831 q->mq_ops = set->ops;
1832 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1834 if (!(set->flags & BLK_MQ_F_SG_MERGE))
1835 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
1837 q->sg_reserved_size = INT_MAX;
1839 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1840 INIT_LIST_HEAD(&q->requeue_list);
1841 spin_lock_init(&q->requeue_lock);
1843 if (q->nr_hw_queues > 1)
1844 blk_queue_make_request(q, blk_mq_make_request);
1846 blk_queue_make_request(q, blk_sq_make_request);
1848 blk_queue_rq_timed_out(q, blk_mq_rq_timed_out);
1850 blk_queue_rq_timeout(q, set->timeout);
1853 * Do this after blk_queue_make_request() overrides it...
1855 q->nr_requests = set->queue_depth;
1857 if (set->ops->complete)
1858 blk_queue_softirq_done(q, set->ops->complete);
1860 blk_mq_init_flush(q);
1861 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1863 q->flush_rq = kzalloc(round_up(sizeof(struct request) +
1864 set->cmd_size, cache_line_size()),
1869 if (blk_mq_init_hw_queues(q, set))
1872 mutex_lock(&all_q_mutex);
1873 list_add_tail(&q->all_q_node, &all_q_list);
1874 mutex_unlock(&all_q_mutex);
1876 blk_mq_add_queue_tag_set(set, q);
1878 blk_mq_map_swqueue(q);
1885 blk_cleanup_queue(q);
1888 for (i = 0; i < set->nr_hw_queues; i++) {
1891 free_cpumask_var(hctxs[i]->cpumask);
1898 return ERR_PTR(-ENOMEM);
1900 EXPORT_SYMBOL(blk_mq_init_queue);
1902 void blk_mq_free_queue(struct request_queue *q)
1904 struct blk_mq_tag_set *set = q->tag_set;
1906 blk_mq_del_queue_tag_set(q);
1908 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
1909 blk_mq_free_hw_queues(q, set);
1911 percpu_counter_destroy(&q->mq_usage_counter);
1913 free_percpu(q->queue_ctx);
1914 kfree(q->queue_hw_ctx);
1917 q->queue_ctx = NULL;
1918 q->queue_hw_ctx = NULL;
1921 mutex_lock(&all_q_mutex);
1922 list_del_init(&q->all_q_node);
1923 mutex_unlock(&all_q_mutex);
1926 /* Basically redo blk_mq_init_queue with queue frozen */
1927 static void blk_mq_queue_reinit(struct request_queue *q)
1929 blk_mq_freeze_queue(q);
1931 blk_mq_sysfs_unregister(q);
1933 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1936 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1937 * we should change hctx numa_node according to new topology (this
1938 * involves free and re-allocate memory, worthy doing?)
1941 blk_mq_map_swqueue(q);
1943 blk_mq_sysfs_register(q);
1945 blk_mq_unfreeze_queue(q);
1948 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1949 unsigned long action, void *hcpu)
1951 struct request_queue *q;
1954 * Before new mappings are established, hotadded cpu might already
1955 * start handling requests. This doesn't break anything as we map
1956 * offline CPUs to first hardware queue. We will re-init the queue
1957 * below to get optimal settings.
1959 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1960 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1963 mutex_lock(&all_q_mutex);
1964 list_for_each_entry(q, &all_q_list, all_q_node)
1965 blk_mq_queue_reinit(q);
1966 mutex_unlock(&all_q_mutex);
1970 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
1974 if (!set->nr_hw_queues)
1976 if (!set->queue_depth || set->queue_depth > BLK_MQ_MAX_DEPTH)
1978 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
1981 if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue)
1985 set->tags = kmalloc_node(set->nr_hw_queues *
1986 sizeof(struct blk_mq_tags *),
1987 GFP_KERNEL, set->numa_node);
1991 for (i = 0; i < set->nr_hw_queues; i++) {
1992 set->tags[i] = blk_mq_init_rq_map(set, i);
1997 mutex_init(&set->tag_list_lock);
1998 INIT_LIST_HEAD(&set->tag_list);
2004 blk_mq_free_rq_map(set, set->tags[i], i);
2008 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2010 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2014 for (i = 0; i < set->nr_hw_queues; i++) {
2016 blk_mq_free_rq_map(set, set->tags[i], i);
2021 EXPORT_SYMBOL(blk_mq_free_tag_set);
2023 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2025 struct blk_mq_tag_set *set = q->tag_set;
2026 struct blk_mq_hw_ctx *hctx;
2029 if (!set || nr > set->queue_depth)
2033 queue_for_each_hw_ctx(q, hctx, i) {
2034 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2040 q->nr_requests = nr;
2045 void blk_mq_disable_hotplug(void)
2047 mutex_lock(&all_q_mutex);
2050 void blk_mq_enable_hotplug(void)
2052 mutex_unlock(&all_q_mutex);
2055 static int __init blk_mq_init(void)
2059 /* Must be called after percpu_counter_hotcpu_callback() */
2060 hotcpu_notifier(blk_mq_queue_reinit_notify, -10);
2064 subsys_initcall(blk_mq_init);