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 int blk_mq_queue_enter(struct request_queue *q)
99 __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
101 /* we have problems to freeze the queue if it's initializing */
102 if (!blk_queue_bypass(q) || !blk_queue_init_done(q))
105 __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
107 spin_lock_irq(q->queue_lock);
108 ret = wait_event_interruptible_lock_irq(q->mq_freeze_wq,
109 !blk_queue_bypass(q) || blk_queue_dying(q),
111 /* inc usage with lock hold to avoid freeze_queue runs here */
112 if (!ret && !blk_queue_dying(q))
113 __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
114 else if (blk_queue_dying(q))
116 spin_unlock_irq(q->queue_lock);
121 static void blk_mq_queue_exit(struct request_queue *q)
123 __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
126 static void __blk_mq_drain_queue(struct request_queue *q)
131 spin_lock_irq(q->queue_lock);
132 count = percpu_counter_sum(&q->mq_usage_counter);
133 spin_unlock_irq(q->queue_lock);
137 blk_mq_run_queues(q, false);
143 * Guarantee no request is in use, so we can change any data structure of
144 * the queue afterward.
146 static void blk_mq_freeze_queue(struct request_queue *q)
150 spin_lock_irq(q->queue_lock);
151 drain = !q->bypass_depth++;
152 queue_flag_set(QUEUE_FLAG_BYPASS, q);
153 spin_unlock_irq(q->queue_lock);
156 __blk_mq_drain_queue(q);
159 void blk_mq_drain_queue(struct request_queue *q)
161 __blk_mq_drain_queue(q);
164 static void blk_mq_unfreeze_queue(struct request_queue *q)
168 spin_lock_irq(q->queue_lock);
169 if (!--q->bypass_depth) {
170 queue_flag_clear(QUEUE_FLAG_BYPASS, q);
173 WARN_ON_ONCE(q->bypass_depth < 0);
174 spin_unlock_irq(q->queue_lock);
176 wake_up_all(&q->mq_freeze_wq);
179 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
181 return blk_mq_has_free_tags(hctx->tags);
183 EXPORT_SYMBOL(blk_mq_can_queue);
185 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
186 struct request *rq, unsigned int rw_flags)
188 if (blk_queue_io_stat(q))
189 rw_flags |= REQ_IO_STAT;
191 INIT_LIST_HEAD(&rq->queuelist);
192 /* csd/requeue_work/fifo_time is initialized before use */
195 rq->cmd_flags |= rw_flags;
197 /* do not touch atomic flags, it needs atomic ops against the timer */
200 rq->__sector = (sector_t) -1;
203 INIT_HLIST_NODE(&rq->hash);
204 RB_CLEAR_NODE(&rq->rb_node);
205 memset(&rq->flush, 0, max(sizeof(rq->flush), sizeof(rq->elv)));
208 rq->start_time = jiffies;
209 #ifdef CONFIG_BLK_CGROUP
211 set_start_time_ns(rq);
212 rq->io_start_time_ns = 0;
214 rq->nr_phys_segments = 0;
215 #if defined(CONFIG_BLK_DEV_INTEGRITY)
216 rq->nr_integrity_segments = 0;
220 /* tag was already set */
222 memset(rq->__cmd, 0, sizeof(rq->__cmd));
224 rq->cmd_len = BLK_MAX_CDB;
232 INIT_LIST_HEAD(&rq->timeout_list);
236 rq->end_io_data = NULL;
239 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
242 static struct request *
243 __blk_mq_alloc_request(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
244 struct blk_mq_ctx *ctx, int rw, gfp_t gfp, bool reserved)
249 tag = blk_mq_get_tag(hctx, &ctx->last_tag, gfp, reserved);
250 if (tag != BLK_MQ_TAG_FAIL) {
251 rq = hctx->tags->rqs[tag];
254 if (blk_mq_tag_busy(hctx)) {
255 rq->cmd_flags = REQ_MQ_INFLIGHT;
256 atomic_inc(&hctx->nr_active);
260 blk_mq_rq_ctx_init(q, ctx, rq, rw);
267 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
270 struct blk_mq_ctx *ctx;
271 struct blk_mq_hw_ctx *hctx;
274 if (blk_mq_queue_enter(q))
277 ctx = blk_mq_get_ctx(q);
278 hctx = q->mq_ops->map_queue(q, ctx->cpu);
280 rq = __blk_mq_alloc_request(q, hctx, ctx, rw, gfp & ~__GFP_WAIT,
282 if (!rq && (gfp & __GFP_WAIT)) {
283 __blk_mq_run_hw_queue(hctx);
286 ctx = blk_mq_get_ctx(q);
287 hctx = q->mq_ops->map_queue(q, ctx->cpu);
288 rq = __blk_mq_alloc_request(q, hctx, ctx, rw, gfp, reserved);
293 EXPORT_SYMBOL(blk_mq_alloc_request);
295 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
296 struct blk_mq_ctx *ctx, struct request *rq)
298 const int tag = rq->tag;
299 struct request_queue *q = rq->q;
301 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
302 atomic_dec(&hctx->nr_active);
304 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
305 blk_mq_put_tag(hctx, tag, &ctx->last_tag);
306 blk_mq_queue_exit(q);
309 void blk_mq_free_request(struct request *rq)
311 struct blk_mq_ctx *ctx = rq->mq_ctx;
312 struct blk_mq_hw_ctx *hctx;
313 struct request_queue *q = rq->q;
315 ctx->rq_completed[rq_is_sync(rq)]++;
317 hctx = q->mq_ops->map_queue(q, ctx->cpu);
318 __blk_mq_free_request(hctx, ctx, rq);
322 * Clone all relevant state from a request that has been put on hold in
323 * the flush state machine into the preallocated flush request that hangs
324 * off the request queue.
326 * For a driver the flush request should be invisible, that's why we are
327 * impersonating the original request here.
329 void blk_mq_clone_flush_request(struct request *flush_rq,
330 struct request *orig_rq)
332 struct blk_mq_hw_ctx *hctx =
333 orig_rq->q->mq_ops->map_queue(orig_rq->q, orig_rq->mq_ctx->cpu);
335 flush_rq->mq_ctx = orig_rq->mq_ctx;
336 flush_rq->tag = orig_rq->tag;
337 memcpy(blk_mq_rq_to_pdu(flush_rq), blk_mq_rq_to_pdu(orig_rq),
341 inline void __blk_mq_end_io(struct request *rq, int error)
343 blk_account_io_done(rq);
346 rq->end_io(rq, error);
348 if (unlikely(blk_bidi_rq(rq)))
349 blk_mq_free_request(rq->next_rq);
350 blk_mq_free_request(rq);
353 EXPORT_SYMBOL(__blk_mq_end_io);
355 void blk_mq_end_io(struct request *rq, int error)
357 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
359 __blk_mq_end_io(rq, error);
361 EXPORT_SYMBOL(blk_mq_end_io);
363 static void __blk_mq_complete_request_remote(void *data)
365 struct request *rq = data;
367 rq->q->softirq_done_fn(rq);
370 void __blk_mq_complete_request(struct request *rq)
372 struct blk_mq_ctx *ctx = rq->mq_ctx;
376 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
377 rq->q->softirq_done_fn(rq);
382 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
383 shared = cpus_share_cache(cpu, ctx->cpu);
385 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
386 rq->csd.func = __blk_mq_complete_request_remote;
389 smp_call_function_single_async(ctx->cpu, &rq->csd);
391 rq->q->softirq_done_fn(rq);
397 * blk_mq_complete_request - end I/O on a request
398 * @rq: the request being processed
401 * Ends all I/O on a request. It does not handle partial completions.
402 * The actual completion happens out-of-order, through a IPI handler.
404 void blk_mq_complete_request(struct request *rq)
406 struct request_queue *q = rq->q;
408 if (unlikely(blk_should_fake_timeout(q)))
410 if (!blk_mark_rq_complete(rq)) {
411 if (q->softirq_done_fn)
412 __blk_mq_complete_request(rq);
414 blk_mq_end_io(rq, rq->errors);
417 EXPORT_SYMBOL(blk_mq_complete_request);
419 static void blk_mq_start_request(struct request *rq, bool last)
421 struct request_queue *q = rq->q;
423 trace_block_rq_issue(q, rq);
425 rq->resid_len = blk_rq_bytes(rq);
426 if (unlikely(blk_bidi_rq(rq)))
427 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
430 * Just mark start time and set the started bit. Due to memory
431 * ordering, we know we'll see the correct deadline as long as
432 * REQ_ATOMIC_STARTED is seen. Use the default queue timeout,
433 * unless one has been set in the request.
436 rq->deadline = jiffies + q->rq_timeout;
438 rq->deadline = jiffies + rq->timeout;
441 * Mark us as started and clear complete. Complete might have been
442 * set if requeue raced with timeout, which then marked it as
443 * complete. So be sure to clear complete again when we start
444 * the request, otherwise we'll ignore the completion event.
446 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
447 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
449 if (q->dma_drain_size && blk_rq_bytes(rq)) {
451 * Make sure space for the drain appears. We know we can do
452 * this because max_hw_segments has been adjusted to be one
453 * fewer than the device can handle.
455 rq->nr_phys_segments++;
459 * Flag the last request in the series so that drivers know when IO
460 * should be kicked off, if they don't do it on a per-request basis.
462 * Note: the flag isn't the only condition drivers should do kick off.
463 * If drive is busy, the last request might not have the bit set.
466 rq->cmd_flags |= REQ_END;
469 static void __blk_mq_requeue_request(struct request *rq)
471 struct request_queue *q = rq->q;
473 trace_block_rq_requeue(q, rq);
474 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
476 rq->cmd_flags &= ~REQ_END;
478 if (q->dma_drain_size && blk_rq_bytes(rq))
479 rq->nr_phys_segments--;
482 void blk_mq_requeue_request(struct request *rq)
484 __blk_mq_requeue_request(rq);
485 blk_clear_rq_complete(rq);
487 BUG_ON(blk_queued_rq(rq));
488 blk_mq_add_to_requeue_list(rq, true);
490 EXPORT_SYMBOL(blk_mq_requeue_request);
492 static void blk_mq_requeue_work(struct work_struct *work)
494 struct request_queue *q =
495 container_of(work, struct request_queue, requeue_work);
497 struct request *rq, *next;
500 spin_lock_irqsave(&q->requeue_lock, flags);
501 list_splice_init(&q->requeue_list, &rq_list);
502 spin_unlock_irqrestore(&q->requeue_lock, flags);
504 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
505 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
508 rq->cmd_flags &= ~REQ_SOFTBARRIER;
509 list_del_init(&rq->queuelist);
510 blk_mq_insert_request(rq, true, false, false);
513 while (!list_empty(&rq_list)) {
514 rq = list_entry(rq_list.next, struct request, queuelist);
515 list_del_init(&rq->queuelist);
516 blk_mq_insert_request(rq, false, false, false);
519 blk_mq_run_queues(q, false);
522 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
524 struct request_queue *q = rq->q;
528 * We abuse this flag that is otherwise used by the I/O scheduler to
529 * request head insertation from the workqueue.
531 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
533 spin_lock_irqsave(&q->requeue_lock, flags);
535 rq->cmd_flags |= REQ_SOFTBARRIER;
536 list_add(&rq->queuelist, &q->requeue_list);
538 list_add_tail(&rq->queuelist, &q->requeue_list);
540 spin_unlock_irqrestore(&q->requeue_lock, flags);
542 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
544 void blk_mq_kick_requeue_list(struct request_queue *q)
546 kblockd_schedule_work(&q->requeue_work);
548 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
550 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
552 return tags->rqs[tag];
554 EXPORT_SYMBOL(blk_mq_tag_to_rq);
556 struct blk_mq_timeout_data {
557 struct blk_mq_hw_ctx *hctx;
559 unsigned int *next_set;
562 static void blk_mq_timeout_check(void *__data, unsigned long *free_tags)
564 struct blk_mq_timeout_data *data = __data;
565 struct blk_mq_hw_ctx *hctx = data->hctx;
568 /* It may not be in flight yet (this is where
569 * the REQ_ATOMIC_STARTED flag comes in). The requests are
570 * statically allocated, so we know it's always safe to access the
571 * memory associated with a bit offset into ->rqs[].
577 tag = find_next_zero_bit(free_tags, hctx->tags->nr_tags, tag);
578 if (tag >= hctx->tags->nr_tags)
581 rq = blk_mq_tag_to_rq(hctx->tags, tag++);
582 if (rq->q != hctx->queue)
584 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
587 blk_rq_check_expired(rq, data->next, data->next_set);
591 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx,
593 unsigned int *next_set)
595 struct blk_mq_timeout_data data = {
598 .next_set = next_set,
602 * Ask the tagging code to iterate busy requests, so we can
603 * check them for timeout.
605 blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data);
608 static enum blk_eh_timer_return blk_mq_rq_timed_out(struct request *rq)
610 struct request_queue *q = rq->q;
613 * We know that complete is set at this point. If STARTED isn't set
614 * anymore, then the request isn't active and the "timeout" should
615 * just be ignored. This can happen due to the bitflag ordering.
616 * Timeout first checks if STARTED is set, and if it is, assumes
617 * the request is active. But if we race with completion, then
618 * we both flags will get cleared. So check here again, and ignore
619 * a timeout event with a request that isn't active.
621 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
622 return BLK_EH_NOT_HANDLED;
624 if (!q->mq_ops->timeout)
625 return BLK_EH_RESET_TIMER;
627 return q->mq_ops->timeout(rq);
630 static void blk_mq_rq_timer(unsigned long data)
632 struct request_queue *q = (struct request_queue *) data;
633 struct blk_mq_hw_ctx *hctx;
634 unsigned long next = 0;
637 queue_for_each_hw_ctx(q, hctx, i) {
639 * If not software queues are currently mapped to this
640 * hardware queue, there's nothing to check
642 if (!hctx->nr_ctx || !hctx->tags)
645 blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set);
649 next = blk_rq_timeout(round_jiffies_up(next));
650 mod_timer(&q->timeout, next);
652 queue_for_each_hw_ctx(q, hctx, i)
653 blk_mq_tag_idle(hctx);
658 * Reverse check our software queue for entries that we could potentially
659 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
660 * too much time checking for merges.
662 static bool blk_mq_attempt_merge(struct request_queue *q,
663 struct blk_mq_ctx *ctx, struct bio *bio)
668 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
674 if (!blk_rq_merge_ok(rq, bio))
677 el_ret = blk_try_merge(rq, bio);
678 if (el_ret == ELEVATOR_BACK_MERGE) {
679 if (bio_attempt_back_merge(q, rq, bio)) {
684 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
685 if (bio_attempt_front_merge(q, rq, bio)) {
697 * Process software queues that have been marked busy, splicing them
698 * to the for-dispatch
700 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
702 struct blk_mq_ctx *ctx;
705 for (i = 0; i < hctx->ctx_map.map_size; i++) {
706 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
707 unsigned int off, bit;
713 off = i * hctx->ctx_map.bits_per_word;
715 bit = find_next_bit(&bm->word, bm->depth, bit);
716 if (bit >= bm->depth)
719 ctx = hctx->ctxs[bit + off];
720 clear_bit(bit, &bm->word);
721 spin_lock(&ctx->lock);
722 list_splice_tail_init(&ctx->rq_list, list);
723 spin_unlock(&ctx->lock);
731 * Run this hardware queue, pulling any software queues mapped to it in.
732 * Note that this function currently has various problems around ordering
733 * of IO. In particular, we'd like FIFO behaviour on handling existing
734 * items on the hctx->dispatch list. Ignore that for now.
736 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
738 struct request_queue *q = hctx->queue;
743 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
745 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
751 * Touch any software queue that has pending entries.
753 flush_busy_ctxs(hctx, &rq_list);
756 * If we have previous entries on our dispatch list, grab them
757 * and stuff them at the front for more fair dispatch.
759 if (!list_empty_careful(&hctx->dispatch)) {
760 spin_lock(&hctx->lock);
761 if (!list_empty(&hctx->dispatch))
762 list_splice_init(&hctx->dispatch, &rq_list);
763 spin_unlock(&hctx->lock);
767 * Now process all the entries, sending them to the driver.
770 while (!list_empty(&rq_list)) {
773 rq = list_first_entry(&rq_list, struct request, queuelist);
774 list_del_init(&rq->queuelist);
776 blk_mq_start_request(rq, list_empty(&rq_list));
778 ret = q->mq_ops->queue_rq(hctx, rq);
780 case BLK_MQ_RQ_QUEUE_OK:
783 case BLK_MQ_RQ_QUEUE_BUSY:
784 list_add(&rq->queuelist, &rq_list);
785 __blk_mq_requeue_request(rq);
788 pr_err("blk-mq: bad return on queue: %d\n", ret);
789 case BLK_MQ_RQ_QUEUE_ERROR:
791 blk_mq_end_io(rq, rq->errors);
795 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
800 hctx->dispatched[0]++;
801 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
802 hctx->dispatched[ilog2(queued) + 1]++;
805 * Any items that need requeuing? Stuff them into hctx->dispatch,
806 * that is where we will continue on next queue run.
808 if (!list_empty(&rq_list)) {
809 spin_lock(&hctx->lock);
810 list_splice(&rq_list, &hctx->dispatch);
811 spin_unlock(&hctx->lock);
816 * It'd be great if the workqueue API had a way to pass
817 * in a mask and had some smarts for more clever placement.
818 * For now we just round-robin here, switching for every
819 * BLK_MQ_CPU_WORK_BATCH queued items.
821 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
823 int cpu = hctx->next_cpu;
825 if (--hctx->next_cpu_batch <= 0) {
828 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
829 if (next_cpu >= nr_cpu_ids)
830 next_cpu = cpumask_first(hctx->cpumask);
832 hctx->next_cpu = next_cpu;
833 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
839 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
841 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
844 if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask))
845 __blk_mq_run_hw_queue(hctx);
846 else if (hctx->queue->nr_hw_queues == 1)
847 kblockd_schedule_delayed_work(&hctx->run_work, 0);
851 cpu = blk_mq_hctx_next_cpu(hctx);
852 kblockd_schedule_delayed_work_on(cpu, &hctx->run_work, 0);
856 void blk_mq_run_queues(struct request_queue *q, bool async)
858 struct blk_mq_hw_ctx *hctx;
861 queue_for_each_hw_ctx(q, hctx, i) {
862 if ((!blk_mq_hctx_has_pending(hctx) &&
863 list_empty_careful(&hctx->dispatch)) ||
864 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
868 blk_mq_run_hw_queue(hctx, async);
872 EXPORT_SYMBOL(blk_mq_run_queues);
874 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
876 cancel_delayed_work(&hctx->run_work);
877 cancel_delayed_work(&hctx->delay_work);
878 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
880 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
882 void blk_mq_stop_hw_queues(struct request_queue *q)
884 struct blk_mq_hw_ctx *hctx;
887 queue_for_each_hw_ctx(q, hctx, i)
888 blk_mq_stop_hw_queue(hctx);
890 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
892 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
894 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
897 __blk_mq_run_hw_queue(hctx);
900 EXPORT_SYMBOL(blk_mq_start_hw_queue);
902 void blk_mq_start_hw_queues(struct request_queue *q)
904 struct blk_mq_hw_ctx *hctx;
907 queue_for_each_hw_ctx(q, hctx, i)
908 blk_mq_start_hw_queue(hctx);
910 EXPORT_SYMBOL(blk_mq_start_hw_queues);
913 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
915 struct blk_mq_hw_ctx *hctx;
918 queue_for_each_hw_ctx(q, hctx, i) {
919 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
922 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
924 blk_mq_run_hw_queue(hctx, async);
928 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
930 static void blk_mq_run_work_fn(struct work_struct *work)
932 struct blk_mq_hw_ctx *hctx;
934 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
936 __blk_mq_run_hw_queue(hctx);
939 static void blk_mq_delay_work_fn(struct work_struct *work)
941 struct blk_mq_hw_ctx *hctx;
943 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
945 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
946 __blk_mq_run_hw_queue(hctx);
949 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
951 unsigned long tmo = msecs_to_jiffies(msecs);
953 if (hctx->queue->nr_hw_queues == 1)
954 kblockd_schedule_delayed_work(&hctx->delay_work, tmo);
958 cpu = blk_mq_hctx_next_cpu(hctx);
959 kblockd_schedule_delayed_work_on(cpu, &hctx->delay_work, tmo);
962 EXPORT_SYMBOL(blk_mq_delay_queue);
964 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
965 struct request *rq, bool at_head)
967 struct blk_mq_ctx *ctx = rq->mq_ctx;
969 trace_block_rq_insert(hctx->queue, rq);
972 list_add(&rq->queuelist, &ctx->rq_list);
974 list_add_tail(&rq->queuelist, &ctx->rq_list);
976 blk_mq_hctx_mark_pending(hctx, ctx);
979 * We do this early, to ensure we are on the right CPU.
984 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
987 struct request_queue *q = rq->q;
988 struct blk_mq_hw_ctx *hctx;
989 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
991 current_ctx = blk_mq_get_ctx(q);
992 if (!cpu_online(ctx->cpu))
993 rq->mq_ctx = ctx = current_ctx;
995 hctx = q->mq_ops->map_queue(q, ctx->cpu);
997 if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA) &&
998 !(rq->cmd_flags & (REQ_FLUSH_SEQ))) {
999 blk_insert_flush(rq);
1001 spin_lock(&ctx->lock);
1002 __blk_mq_insert_request(hctx, rq, at_head);
1003 spin_unlock(&ctx->lock);
1007 blk_mq_run_hw_queue(hctx, async);
1009 blk_mq_put_ctx(current_ctx);
1012 static void blk_mq_insert_requests(struct request_queue *q,
1013 struct blk_mq_ctx *ctx,
1014 struct list_head *list,
1019 struct blk_mq_hw_ctx *hctx;
1020 struct blk_mq_ctx *current_ctx;
1022 trace_block_unplug(q, depth, !from_schedule);
1024 current_ctx = blk_mq_get_ctx(q);
1026 if (!cpu_online(ctx->cpu))
1028 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1031 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1034 spin_lock(&ctx->lock);
1035 while (!list_empty(list)) {
1038 rq = list_first_entry(list, struct request, queuelist);
1039 list_del_init(&rq->queuelist);
1041 __blk_mq_insert_request(hctx, rq, false);
1043 spin_unlock(&ctx->lock);
1045 blk_mq_run_hw_queue(hctx, from_schedule);
1046 blk_mq_put_ctx(current_ctx);
1049 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1051 struct request *rqa = container_of(a, struct request, queuelist);
1052 struct request *rqb = container_of(b, struct request, queuelist);
1054 return !(rqa->mq_ctx < rqb->mq_ctx ||
1055 (rqa->mq_ctx == rqb->mq_ctx &&
1056 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1059 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1061 struct blk_mq_ctx *this_ctx;
1062 struct request_queue *this_q;
1065 LIST_HEAD(ctx_list);
1068 list_splice_init(&plug->mq_list, &list);
1070 list_sort(NULL, &list, plug_ctx_cmp);
1076 while (!list_empty(&list)) {
1077 rq = list_entry_rq(list.next);
1078 list_del_init(&rq->queuelist);
1080 if (rq->mq_ctx != this_ctx) {
1082 blk_mq_insert_requests(this_q, this_ctx,
1087 this_ctx = rq->mq_ctx;
1093 list_add_tail(&rq->queuelist, &ctx_list);
1097 * If 'this_ctx' is set, we know we have entries to complete
1098 * on 'ctx_list'. Do those.
1101 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1106 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1108 init_request_from_bio(rq, bio);
1109 blk_account_io_start(rq, 1);
1112 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1113 struct blk_mq_ctx *ctx,
1114 struct request *rq, struct bio *bio)
1116 struct request_queue *q = hctx->queue;
1118 if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE)) {
1119 blk_mq_bio_to_request(rq, bio);
1120 spin_lock(&ctx->lock);
1122 __blk_mq_insert_request(hctx, rq, false);
1123 spin_unlock(&ctx->lock);
1126 spin_lock(&ctx->lock);
1127 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1128 blk_mq_bio_to_request(rq, bio);
1132 spin_unlock(&ctx->lock);
1133 __blk_mq_free_request(hctx, ctx, rq);
1138 struct blk_map_ctx {
1139 struct blk_mq_hw_ctx *hctx;
1140 struct blk_mq_ctx *ctx;
1143 static struct request *blk_mq_map_request(struct request_queue *q,
1145 struct blk_map_ctx *data)
1147 struct blk_mq_hw_ctx *hctx;
1148 struct blk_mq_ctx *ctx;
1150 int rw = bio_data_dir(bio);
1152 if (unlikely(blk_mq_queue_enter(q))) {
1153 bio_endio(bio, -EIO);
1157 ctx = blk_mq_get_ctx(q);
1158 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1160 if (rw_is_sync(bio->bi_rw))
1163 trace_block_getrq(q, bio, rw);
1164 rq = __blk_mq_alloc_request(q, hctx, ctx, rw, GFP_ATOMIC, false);
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 rq = __blk_mq_alloc_request(q, hctx, ctx, rw,
1173 __GFP_WAIT|GFP_ATOMIC, false);
1183 * Multiple hardware queue variant. This will not use per-process plugs,
1184 * but will attempt to bypass the hctx queueing if we can go straight to
1185 * hardware for SYNC IO.
1187 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1189 const int is_sync = rw_is_sync(bio->bi_rw);
1190 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1191 struct blk_map_ctx data;
1194 blk_queue_bounce(q, &bio);
1196 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1197 bio_endio(bio, -EIO);
1201 rq = blk_mq_map_request(q, bio, &data);
1205 if (unlikely(is_flush_fua)) {
1206 blk_mq_bio_to_request(rq, bio);
1207 blk_insert_flush(rq);
1214 blk_mq_bio_to_request(rq, bio);
1215 blk_mq_start_request(rq, true);
1218 * For OK queue, we are done. For error, kill it. Any other
1219 * error (busy), just add it to our list as we previously
1222 ret = q->mq_ops->queue_rq(data.hctx, rq);
1223 if (ret == BLK_MQ_RQ_QUEUE_OK)
1226 __blk_mq_requeue_request(rq);
1228 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1230 blk_mq_end_io(rq, rq->errors);
1236 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1238 * For a SYNC request, send it to the hardware immediately. For
1239 * an ASYNC request, just ensure that we run it later on. The
1240 * latter allows for merging opportunities and more efficient
1244 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1247 blk_mq_put_ctx(data.ctx);
1251 * Single hardware queue variant. This will attempt to use any per-process
1252 * plug for merging and IO deferral.
1254 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1256 const int is_sync = rw_is_sync(bio->bi_rw);
1257 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1258 unsigned int use_plug, request_count = 0;
1259 struct blk_map_ctx data;
1263 * If we have multiple hardware queues, just go directly to
1264 * one of those for sync IO.
1266 use_plug = !is_flush_fua && !is_sync;
1268 blk_queue_bounce(q, &bio);
1270 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1271 bio_endio(bio, -EIO);
1275 if (use_plug && !blk_queue_nomerges(q) &&
1276 blk_attempt_plug_merge(q, bio, &request_count))
1279 rq = blk_mq_map_request(q, bio, &data);
1281 if (unlikely(is_flush_fua)) {
1282 blk_mq_bio_to_request(rq, bio);
1283 blk_insert_flush(rq);
1288 * A task plug currently exists. Since this is completely lockless,
1289 * utilize that to temporarily store requests until the task is
1290 * either done or scheduled away.
1293 struct blk_plug *plug = current->plug;
1296 blk_mq_bio_to_request(rq, bio);
1297 if (list_empty(&plug->mq_list))
1298 trace_block_plug(q);
1299 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1300 blk_flush_plug_list(plug, false);
1301 trace_block_plug(q);
1303 list_add_tail(&rq->queuelist, &plug->mq_list);
1304 blk_mq_put_ctx(data.ctx);
1309 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1311 * For a SYNC request, send it to the hardware immediately. For
1312 * an ASYNC request, just ensure that we run it later on. The
1313 * latter allows for merging opportunities and more efficient
1317 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1320 blk_mq_put_ctx(data.ctx);
1324 * Default mapping to a software queue, since we use one per CPU.
1326 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1328 return q->queue_hw_ctx[q->mq_map[cpu]];
1330 EXPORT_SYMBOL(blk_mq_map_queue);
1332 struct blk_mq_hw_ctx *blk_mq_alloc_single_hw_queue(struct blk_mq_tag_set *set,
1333 unsigned int hctx_index,
1336 return kzalloc_node(sizeof(struct blk_mq_hw_ctx), GFP_KERNEL, node);
1338 EXPORT_SYMBOL(blk_mq_alloc_single_hw_queue);
1340 void blk_mq_free_single_hw_queue(struct blk_mq_hw_ctx *hctx,
1341 unsigned int hctx_index)
1345 EXPORT_SYMBOL(blk_mq_free_single_hw_queue);
1347 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1348 struct blk_mq_tags *tags, unsigned int hctx_idx)
1352 if (tags->rqs && set->ops->exit_request) {
1355 for (i = 0; i < tags->nr_tags; i++) {
1358 set->ops->exit_request(set->driver_data, tags->rqs[i],
1363 while (!list_empty(&tags->page_list)) {
1364 page = list_first_entry(&tags->page_list, struct page, lru);
1365 list_del_init(&page->lru);
1366 __free_pages(page, page->private);
1371 blk_mq_free_tags(tags);
1374 static size_t order_to_size(unsigned int order)
1376 return (size_t)PAGE_SIZE << order;
1379 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1380 unsigned int hctx_idx)
1382 struct blk_mq_tags *tags;
1383 unsigned int i, j, entries_per_page, max_order = 4;
1384 size_t rq_size, left;
1386 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1391 INIT_LIST_HEAD(&tags->page_list);
1393 tags->rqs = kmalloc_node(set->queue_depth * sizeof(struct request *),
1394 GFP_KERNEL, set->numa_node);
1396 blk_mq_free_tags(tags);
1401 * rq_size is the size of the request plus driver payload, rounded
1402 * to the cacheline size
1404 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1406 left = rq_size * set->queue_depth;
1408 for (i = 0; i < set->queue_depth; ) {
1409 int this_order = max_order;
1414 while (left < order_to_size(this_order - 1) && this_order)
1418 page = alloc_pages_node(set->numa_node, GFP_KERNEL,
1424 if (order_to_size(this_order) < rq_size)
1431 page->private = this_order;
1432 list_add_tail(&page->lru, &tags->page_list);
1434 p = page_address(page);
1435 entries_per_page = order_to_size(this_order) / rq_size;
1436 to_do = min(entries_per_page, set->queue_depth - i);
1437 left -= to_do * rq_size;
1438 for (j = 0; j < to_do; j++) {
1440 if (set->ops->init_request) {
1441 if (set->ops->init_request(set->driver_data,
1442 tags->rqs[i], hctx_idx, i,
1455 pr_warn("%s: failed to allocate requests\n", __func__);
1456 blk_mq_free_rq_map(set, tags, hctx_idx);
1460 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1465 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1467 unsigned int bpw = 8, total, num_maps, i;
1469 bitmap->bits_per_word = bpw;
1471 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1472 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1477 bitmap->map_size = num_maps;
1480 for (i = 0; i < num_maps; i++) {
1481 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1482 total -= bitmap->map[i].depth;
1488 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1490 struct request_queue *q = hctx->queue;
1491 struct blk_mq_ctx *ctx;
1495 * Move ctx entries to new CPU, if this one is going away.
1497 ctx = __blk_mq_get_ctx(q, cpu);
1499 spin_lock(&ctx->lock);
1500 if (!list_empty(&ctx->rq_list)) {
1501 list_splice_init(&ctx->rq_list, &tmp);
1502 blk_mq_hctx_clear_pending(hctx, ctx);
1504 spin_unlock(&ctx->lock);
1506 if (list_empty(&tmp))
1509 ctx = blk_mq_get_ctx(q);
1510 spin_lock(&ctx->lock);
1512 while (!list_empty(&tmp)) {
1515 rq = list_first_entry(&tmp, struct request, queuelist);
1517 list_move_tail(&rq->queuelist, &ctx->rq_list);
1520 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1521 blk_mq_hctx_mark_pending(hctx, ctx);
1523 spin_unlock(&ctx->lock);
1525 blk_mq_run_hw_queue(hctx, true);
1526 blk_mq_put_ctx(ctx);
1530 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1532 struct request_queue *q = hctx->queue;
1533 struct blk_mq_tag_set *set = q->tag_set;
1535 if (set->tags[hctx->queue_num])
1538 set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1539 if (!set->tags[hctx->queue_num])
1542 hctx->tags = set->tags[hctx->queue_num];
1546 static int blk_mq_hctx_notify(void *data, unsigned long action,
1549 struct blk_mq_hw_ctx *hctx = data;
1551 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1552 return blk_mq_hctx_cpu_offline(hctx, cpu);
1553 else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1554 return blk_mq_hctx_cpu_online(hctx, cpu);
1559 static void blk_mq_exit_hw_queues(struct request_queue *q,
1560 struct blk_mq_tag_set *set, int nr_queue)
1562 struct blk_mq_hw_ctx *hctx;
1565 queue_for_each_hw_ctx(q, hctx, i) {
1569 if (set->ops->exit_hctx)
1570 set->ops->exit_hctx(hctx, i);
1572 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1574 blk_mq_free_bitmap(&hctx->ctx_map);
1579 static void blk_mq_free_hw_queues(struct request_queue *q,
1580 struct blk_mq_tag_set *set)
1582 struct blk_mq_hw_ctx *hctx;
1585 queue_for_each_hw_ctx(q, hctx, i) {
1586 free_cpumask_var(hctx->cpumask);
1587 set->ops->free_hctx(hctx, i);
1591 static int blk_mq_init_hw_queues(struct request_queue *q,
1592 struct blk_mq_tag_set *set)
1594 struct blk_mq_hw_ctx *hctx;
1598 * Initialize hardware queues
1600 queue_for_each_hw_ctx(q, hctx, i) {
1603 node = hctx->numa_node;
1604 if (node == NUMA_NO_NODE)
1605 node = hctx->numa_node = set->numa_node;
1607 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1608 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1609 spin_lock_init(&hctx->lock);
1610 INIT_LIST_HEAD(&hctx->dispatch);
1612 hctx->queue_num = i;
1613 hctx->flags = set->flags;
1614 hctx->cmd_size = set->cmd_size;
1616 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1617 blk_mq_hctx_notify, hctx);
1618 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1620 hctx->tags = set->tags[i];
1623 * Allocate space for all possible cpus to avoid allocation in
1626 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1631 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1636 if (set->ops->init_hctx &&
1637 set->ops->init_hctx(hctx, set->driver_data, i))
1641 if (i == q->nr_hw_queues)
1647 blk_mq_exit_hw_queues(q, set, i);
1652 static void blk_mq_init_cpu_queues(struct request_queue *q,
1653 unsigned int nr_hw_queues)
1657 for_each_possible_cpu(i) {
1658 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1659 struct blk_mq_hw_ctx *hctx;
1661 memset(__ctx, 0, sizeof(*__ctx));
1663 spin_lock_init(&__ctx->lock);
1664 INIT_LIST_HEAD(&__ctx->rq_list);
1667 /* If the cpu isn't online, the cpu is mapped to first hctx */
1671 hctx = q->mq_ops->map_queue(q, i);
1672 cpumask_set_cpu(i, hctx->cpumask);
1676 * Set local node, IFF we have more than one hw queue. If
1677 * not, we remain on the home node of the device
1679 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1680 hctx->numa_node = cpu_to_node(i);
1684 static void blk_mq_map_swqueue(struct request_queue *q)
1687 struct blk_mq_hw_ctx *hctx;
1688 struct blk_mq_ctx *ctx;
1690 queue_for_each_hw_ctx(q, hctx, i) {
1691 cpumask_clear(hctx->cpumask);
1696 * Map software to hardware queues
1698 queue_for_each_ctx(q, ctx, i) {
1699 /* If the cpu isn't online, the cpu is mapped to first hctx */
1703 hctx = q->mq_ops->map_queue(q, i);
1704 cpumask_set_cpu(i, hctx->cpumask);
1705 ctx->index_hw = hctx->nr_ctx;
1706 hctx->ctxs[hctx->nr_ctx++] = ctx;
1709 queue_for_each_hw_ctx(q, hctx, i) {
1711 * If not software queues are mapped to this hardware queue,
1712 * disable it and free the request entries
1714 if (!hctx->nr_ctx) {
1715 struct blk_mq_tag_set *set = q->tag_set;
1718 blk_mq_free_rq_map(set, set->tags[i], i);
1719 set->tags[i] = NULL;
1726 * Initialize batch roundrobin counts
1728 hctx->next_cpu = cpumask_first(hctx->cpumask);
1729 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1733 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1735 struct blk_mq_hw_ctx *hctx;
1736 struct request_queue *q;
1740 if (set->tag_list.next == set->tag_list.prev)
1745 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1746 blk_mq_freeze_queue(q);
1748 queue_for_each_hw_ctx(q, hctx, i) {
1750 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1752 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1754 blk_mq_unfreeze_queue(q);
1758 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1760 struct blk_mq_tag_set *set = q->tag_set;
1762 blk_mq_freeze_queue(q);
1764 mutex_lock(&set->tag_list_lock);
1765 list_del_init(&q->tag_set_list);
1766 blk_mq_update_tag_set_depth(set);
1767 mutex_unlock(&set->tag_list_lock);
1769 blk_mq_unfreeze_queue(q);
1772 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1773 struct request_queue *q)
1777 mutex_lock(&set->tag_list_lock);
1778 list_add_tail(&q->tag_set_list, &set->tag_list);
1779 blk_mq_update_tag_set_depth(set);
1780 mutex_unlock(&set->tag_list_lock);
1783 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1785 struct blk_mq_hw_ctx **hctxs;
1786 struct blk_mq_ctx *ctx;
1787 struct request_queue *q;
1791 ctx = alloc_percpu(struct blk_mq_ctx);
1793 return ERR_PTR(-ENOMEM);
1795 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1801 map = blk_mq_make_queue_map(set);
1805 for (i = 0; i < set->nr_hw_queues; i++) {
1806 int node = blk_mq_hw_queue_to_node(map, i);
1808 hctxs[i] = set->ops->alloc_hctx(set, i, node);
1812 if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL))
1815 atomic_set(&hctxs[i]->nr_active, 0);
1816 hctxs[i]->numa_node = node;
1817 hctxs[i]->queue_num = i;
1820 q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1824 if (percpu_counter_init(&q->mq_usage_counter, 0))
1827 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1828 blk_queue_rq_timeout(q, 30000);
1830 q->nr_queues = nr_cpu_ids;
1831 q->nr_hw_queues = set->nr_hw_queues;
1835 q->queue_hw_ctx = hctxs;
1837 q->mq_ops = set->ops;
1838 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1840 q->sg_reserved_size = INT_MAX;
1842 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1843 INIT_LIST_HEAD(&q->requeue_list);
1844 spin_lock_init(&q->requeue_lock);
1846 if (q->nr_hw_queues > 1)
1847 blk_queue_make_request(q, blk_mq_make_request);
1849 blk_queue_make_request(q, blk_sq_make_request);
1851 blk_queue_rq_timed_out(q, blk_mq_rq_timed_out);
1853 blk_queue_rq_timeout(q, set->timeout);
1856 * Do this after blk_queue_make_request() overrides it...
1858 q->nr_requests = set->queue_depth;
1860 if (set->ops->complete)
1861 blk_queue_softirq_done(q, set->ops->complete);
1863 blk_mq_init_flush(q);
1864 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1866 q->flush_rq = kzalloc(round_up(sizeof(struct request) +
1867 set->cmd_size, cache_line_size()),
1872 if (blk_mq_init_hw_queues(q, set))
1875 mutex_lock(&all_q_mutex);
1876 list_add_tail(&q->all_q_node, &all_q_list);
1877 mutex_unlock(&all_q_mutex);
1879 blk_mq_add_queue_tag_set(set, q);
1881 blk_mq_map_swqueue(q);
1888 blk_cleanup_queue(q);
1891 for (i = 0; i < set->nr_hw_queues; i++) {
1894 free_cpumask_var(hctxs[i]->cpumask);
1895 set->ops->free_hctx(hctxs[i], i);
1901 return ERR_PTR(-ENOMEM);
1903 EXPORT_SYMBOL(blk_mq_init_queue);
1905 void blk_mq_free_queue(struct request_queue *q)
1907 struct blk_mq_tag_set *set = q->tag_set;
1909 blk_mq_del_queue_tag_set(q);
1911 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
1912 blk_mq_free_hw_queues(q, set);
1914 percpu_counter_destroy(&q->mq_usage_counter);
1916 free_percpu(q->queue_ctx);
1917 kfree(q->queue_hw_ctx);
1920 q->queue_ctx = NULL;
1921 q->queue_hw_ctx = NULL;
1924 mutex_lock(&all_q_mutex);
1925 list_del_init(&q->all_q_node);
1926 mutex_unlock(&all_q_mutex);
1929 /* Basically redo blk_mq_init_queue with queue frozen */
1930 static void blk_mq_queue_reinit(struct request_queue *q)
1932 blk_mq_freeze_queue(q);
1934 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1937 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1938 * we should change hctx numa_node according to new topology (this
1939 * involves free and re-allocate memory, worthy doing?)
1942 blk_mq_map_swqueue(q);
1944 blk_mq_unfreeze_queue(q);
1947 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1948 unsigned long action, void *hcpu)
1950 struct request_queue *q;
1953 * Before new mappings are established, hotadded cpu might already
1954 * start handling requests. This doesn't break anything as we map
1955 * offline CPUs to first hardware queue. We will re-init the queue
1956 * below to get optimal settings.
1958 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1959 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1962 mutex_lock(&all_q_mutex);
1963 list_for_each_entry(q, &all_q_list, all_q_node)
1964 blk_mq_queue_reinit(q);
1965 mutex_unlock(&all_q_mutex);
1969 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
1973 if (!set->nr_hw_queues)
1975 if (!set->queue_depth || set->queue_depth > BLK_MQ_MAX_DEPTH)
1977 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
1980 if (!set->nr_hw_queues ||
1981 !set->ops->queue_rq || !set->ops->map_queue ||
1982 !set->ops->alloc_hctx || !set->ops->free_hctx)
1986 set->tags = kmalloc_node(set->nr_hw_queues *
1987 sizeof(struct blk_mq_tags *),
1988 GFP_KERNEL, set->numa_node);
1992 for (i = 0; i < set->nr_hw_queues; i++) {
1993 set->tags[i] = blk_mq_init_rq_map(set, i);
1998 mutex_init(&set->tag_list_lock);
1999 INIT_LIST_HEAD(&set->tag_list);
2005 blk_mq_free_rq_map(set, set->tags[i], i);
2009 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2011 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2015 for (i = 0; i < set->nr_hw_queues; i++) {
2017 blk_mq_free_rq_map(set, set->tags[i], i);
2022 EXPORT_SYMBOL(blk_mq_free_tag_set);
2024 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2026 struct blk_mq_tag_set *set = q->tag_set;
2027 struct blk_mq_hw_ctx *hctx;
2030 if (!set || nr > set->queue_depth)
2034 queue_for_each_hw_ctx(q, hctx, i) {
2035 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2041 q->nr_requests = nr;
2046 void blk_mq_disable_hotplug(void)
2048 mutex_lock(&all_q_mutex);
2051 void blk_mq_enable_hotplug(void)
2053 mutex_unlock(&all_q_mutex);
2056 static int __init blk_mq_init(void)
2060 /* Must be called after percpu_counter_hotcpu_callback() */
2061 hotcpu_notifier(blk_mq_queue_reinit_notify, -10);
2065 subsys_initcall(blk_mq_init);