2 * Copyright (C) 2016 Red Hat, Inc. All rights reserved.
4 * This file is released under the GPL.
10 #include <linux/elevator.h> /* for rq_end_sector() */
11 #include <linux/blk-mq.h>
13 #define DM_MSG_PREFIX "core-rq"
15 #define DM_MQ_NR_HW_QUEUES 1
16 #define DM_MQ_QUEUE_DEPTH 2048
17 static unsigned dm_mq_nr_hw_queues = DM_MQ_NR_HW_QUEUES;
18 static unsigned dm_mq_queue_depth = DM_MQ_QUEUE_DEPTH;
21 * Request-based DM's mempools' reserved IOs set by the user.
23 #define RESERVED_REQUEST_BASED_IOS 256
24 static unsigned reserved_rq_based_ios = RESERVED_REQUEST_BASED_IOS;
26 static bool use_blk_mq = IS_ENABLED(CONFIG_DM_MQ_DEFAULT);
28 bool dm_use_blk_mq_default(void)
33 bool dm_use_blk_mq(struct mapped_device *md)
35 return md->use_blk_mq;
37 EXPORT_SYMBOL_GPL(dm_use_blk_mq);
39 unsigned dm_get_reserved_rq_based_ios(void)
41 return __dm_get_module_param(&reserved_rq_based_ios,
42 RESERVED_REQUEST_BASED_IOS, DM_RESERVED_MAX_IOS);
44 EXPORT_SYMBOL_GPL(dm_get_reserved_rq_based_ios);
46 static unsigned dm_get_blk_mq_nr_hw_queues(void)
48 return __dm_get_module_param(&dm_mq_nr_hw_queues, 1, 32);
51 static unsigned dm_get_blk_mq_queue_depth(void)
53 return __dm_get_module_param(&dm_mq_queue_depth,
54 DM_MQ_QUEUE_DEPTH, BLK_MQ_MAX_DEPTH);
57 int dm_request_based(struct mapped_device *md)
59 return blk_queue_stackable(md->queue);
62 static void dm_old_start_queue(struct request_queue *q)
66 spin_lock_irqsave(q->queue_lock, flags);
67 if (blk_queue_stopped(q))
69 spin_unlock_irqrestore(q->queue_lock, flags);
72 static void dm_mq_start_queue(struct request_queue *q)
74 blk_mq_start_stopped_hw_queues(q, true);
75 blk_mq_kick_requeue_list(q);
78 void dm_start_queue(struct request_queue *q)
81 dm_old_start_queue(q);
86 static void dm_old_stop_queue(struct request_queue *q)
90 spin_lock_irqsave(q->queue_lock, flags);
91 if (!blk_queue_stopped(q))
93 spin_unlock_irqrestore(q->queue_lock, flags);
96 static void dm_mq_stop_queue(struct request_queue *q)
98 if (blk_mq_queue_stopped(q))
101 blk_mq_quiesce_queue(q);
104 void dm_stop_queue(struct request_queue *q)
107 dm_old_stop_queue(q);
113 * Partial completion handling for request-based dm
115 static void end_clone_bio(struct bio *clone)
117 struct dm_rq_clone_bio_info *info =
118 container_of(clone, struct dm_rq_clone_bio_info, clone);
119 struct dm_rq_target_io *tio = info->tio;
120 struct bio *bio = info->orig;
121 unsigned int nr_bytes = info->orig->bi_iter.bi_size;
122 int error = clone->bi_error;
128 * An error has already been detected on the request.
129 * Once error occurred, just let clone->end_io() handle
135 * Don't notice the error to the upper layer yet.
136 * The error handling decision is made by the target driver,
137 * when the request is completed.
144 * I/O for the bio successfully completed.
145 * Notice the data completion to the upper layer.
149 * bios are processed from the head of the list.
150 * So the completing bio should always be rq->bio.
151 * If it's not, something wrong is happening.
153 if (tio->orig->bio != bio)
154 DMERR("bio completion is going in the middle of the request");
157 * Update the original request.
158 * Do not use blk_end_request() here, because it may complete
159 * the original request before the clone, and break the ordering.
161 blk_update_request(tio->orig, 0, nr_bytes);
164 static struct dm_rq_target_io *tio_from_request(struct request *rq)
166 return blk_mq_rq_to_pdu(rq);
169 static void rq_end_stats(struct mapped_device *md, struct request *orig)
171 if (unlikely(dm_stats_used(&md->stats))) {
172 struct dm_rq_target_io *tio = tio_from_request(orig);
173 tio->duration_jiffies = jiffies - tio->duration_jiffies;
174 dm_stats_account_io(&md->stats, rq_data_dir(orig),
175 blk_rq_pos(orig), tio->n_sectors, true,
176 tio->duration_jiffies, &tio->stats_aux);
181 * Don't touch any member of the md after calling this function because
182 * the md may be freed in dm_put() at the end of this function.
183 * Or do dm_get() before calling this function and dm_put() later.
185 static void rq_completed(struct mapped_device *md, int rw, bool run_queue)
187 struct request_queue *q = md->queue;
190 atomic_dec(&md->pending[rw]);
192 /* nudge anyone waiting on suspend queue */
193 if (!md_in_flight(md))
197 * Run this off this callpath, as drivers could invoke end_io while
198 * inside their request_fn (and holding the queue lock). Calling
199 * back into ->request_fn() could deadlock attempting to grab the
202 if (!q->mq_ops && run_queue) {
203 spin_lock_irqsave(q->queue_lock, flags);
204 blk_run_queue_async(q);
205 spin_unlock_irqrestore(q->queue_lock, flags);
209 * dm_put() must be at the end of this function. See the comment above
215 * Complete the clone and the original request.
216 * Must be called without clone's queue lock held,
217 * see end_clone_request() for more details.
219 static void dm_end_request(struct request *clone, int error)
221 int rw = rq_data_dir(clone);
222 struct dm_rq_target_io *tio = clone->end_io_data;
223 struct mapped_device *md = tio->md;
224 struct request *rq = tio->orig;
226 blk_rq_unprep_clone(clone);
227 tio->ti->type->release_clone_rq(clone);
229 rq_end_stats(md, rq);
231 blk_end_request_all(rq, error);
233 blk_mq_end_request(rq, error);
234 rq_completed(md, rw, true);
238 * Requeue the original request of a clone.
240 static void dm_old_requeue_request(struct request *rq)
242 struct request_queue *q = rq->q;
245 spin_lock_irqsave(q->queue_lock, flags);
246 blk_requeue_request(q, rq);
247 blk_run_queue_async(q);
248 spin_unlock_irqrestore(q->queue_lock, flags);
251 static void __dm_mq_kick_requeue_list(struct request_queue *q, unsigned long msecs)
253 blk_mq_delay_kick_requeue_list(q, msecs);
256 void dm_mq_kick_requeue_list(struct mapped_device *md)
258 __dm_mq_kick_requeue_list(dm_get_md_queue(md), 0);
260 EXPORT_SYMBOL(dm_mq_kick_requeue_list);
262 static void dm_mq_delay_requeue_request(struct request *rq, unsigned long msecs)
264 blk_mq_requeue_request(rq, false);
265 __dm_mq_kick_requeue_list(rq->q, msecs);
268 static void dm_requeue_original_request(struct dm_rq_target_io *tio, bool delay_requeue)
270 struct mapped_device *md = tio->md;
271 struct request *rq = tio->orig;
272 int rw = rq_data_dir(rq);
274 rq_end_stats(md, rq);
276 blk_rq_unprep_clone(tio->clone);
277 tio->ti->type->release_clone_rq(tio->clone);
281 dm_old_requeue_request(rq);
283 dm_mq_delay_requeue_request(rq, delay_requeue ? 5000 : 0);
285 rq_completed(md, rw, false);
288 static void dm_done(struct request *clone, int error, bool mapped)
291 struct dm_rq_target_io *tio = clone->end_io_data;
292 dm_request_endio_fn rq_end_io = NULL;
295 rq_end_io = tio->ti->type->rq_end_io;
297 if (mapped && rq_end_io)
298 r = rq_end_io(tio->ti, clone, error, &tio->info);
301 if (unlikely(r == -EREMOTEIO && (req_op(clone) == REQ_OP_WRITE_SAME) &&
302 !clone->q->limits.max_write_same_sectors))
303 disable_write_same(tio->md);
306 /* The target wants to complete the I/O */
307 dm_end_request(clone, r);
308 else if (r == DM_ENDIO_INCOMPLETE)
309 /* The target will handle the I/O */
311 else if (r == DM_ENDIO_REQUEUE)
312 /* The target wants to requeue the I/O */
313 dm_requeue_original_request(tio, false);
315 DMWARN("unimplemented target endio return value: %d", r);
321 * Request completion handler for request-based dm
323 static void dm_softirq_done(struct request *rq)
326 struct dm_rq_target_io *tio = tio_from_request(rq);
327 struct request *clone = tio->clone;
331 rq_end_stats(tio->md, rq);
332 rw = rq_data_dir(rq);
334 blk_end_request_all(rq, tio->error);
336 blk_mq_end_request(rq, tio->error);
337 rq_completed(tio->md, rw, false);
341 if (rq->rq_flags & RQF_FAILED)
344 dm_done(clone, tio->error, mapped);
348 * Complete the clone and the original request with the error status
349 * through softirq context.
351 static void dm_complete_request(struct request *rq, int error)
353 struct dm_rq_target_io *tio = tio_from_request(rq);
357 blk_complete_request(rq);
359 blk_mq_complete_request(rq, error);
363 * Complete the not-mapped clone and the original request with the error status
364 * through softirq context.
365 * Target's rq_end_io() function isn't called.
366 * This may be used when the target's map_rq() or clone_and_map_rq() functions fail.
368 static void dm_kill_unmapped_request(struct request *rq, int error)
370 rq->rq_flags |= RQF_FAILED;
371 dm_complete_request(rq, error);
375 * Called with the clone's queue lock held (in the case of .request_fn)
377 static void end_clone_request(struct request *clone, int error)
379 struct dm_rq_target_io *tio = clone->end_io_data;
382 * Actual request completion is done in a softirq context which doesn't
383 * hold the clone's queue lock. Otherwise, deadlock could occur because:
384 * - another request may be submitted by the upper level driver
385 * of the stacking during the completion
386 * - the submission which requires queue lock may be done
387 * against this clone's queue
389 dm_complete_request(tio->orig, error);
392 static void dm_dispatch_clone_request(struct request *clone, struct request *rq)
396 if (blk_queue_io_stat(clone->q))
397 clone->rq_flags |= RQF_IO_STAT;
399 clone->start_time = jiffies;
400 r = blk_insert_cloned_request(clone->q, clone);
402 /* must complete clone in terms of original request */
403 dm_complete_request(rq, r);
406 static int dm_rq_bio_constructor(struct bio *bio, struct bio *bio_orig,
409 struct dm_rq_target_io *tio = data;
410 struct dm_rq_clone_bio_info *info =
411 container_of(bio, struct dm_rq_clone_bio_info, clone);
413 info->orig = bio_orig;
415 bio->bi_end_io = end_clone_bio;
420 static int setup_clone(struct request *clone, struct request *rq,
421 struct dm_rq_target_io *tio, gfp_t gfp_mask)
425 r = blk_rq_prep_clone(clone, rq, tio->md->bs, gfp_mask,
426 dm_rq_bio_constructor, tio);
430 clone->end_io = end_clone_request;
431 clone->end_io_data = tio;
438 static void map_tio_request(struct kthread_work *work);
440 static void init_tio(struct dm_rq_target_io *tio, struct request *rq,
441 struct mapped_device *md)
449 * Avoid initializing info for blk-mq; it passes
450 * target-specific data through info.ptr
451 * (see: dm_mq_init_request)
453 if (!md->init_tio_pdu)
454 memset(&tio->info, 0, sizeof(tio->info));
455 if (md->kworker_task)
456 kthread_init_work(&tio->work, map_tio_request);
461 * DM_MAPIO_* : the request has been processed as indicated
462 * DM_MAPIO_REQUEUE : the original request needs to be immediately requeued
463 * < 0 : the request was completed due to failure
465 static int map_request(struct dm_rq_target_io *tio)
468 struct dm_target *ti = tio->ti;
469 struct mapped_device *md = tio->md;
470 struct request *rq = tio->orig;
471 struct request *clone = NULL;
473 r = ti->type->clone_and_map_rq(ti, rq, &tio->info, &clone);
475 case DM_MAPIO_SUBMITTED:
476 /* The target has taken the I/O to submit by itself later */
478 case DM_MAPIO_REMAPPED:
479 if (setup_clone(clone, rq, tio, GFP_ATOMIC)) {
481 ti->type->release_clone_rq(clone);
482 return DM_MAPIO_REQUEUE;
485 /* The target has remapped the I/O so dispatch it */
486 trace_block_rq_remap(clone->q, clone, disk_devt(dm_disk(md)),
488 dm_dispatch_clone_request(clone, rq);
490 case DM_MAPIO_REQUEUE:
491 /* The target wants to requeue the I/O */
493 case DM_MAPIO_DELAY_REQUEUE:
494 /* The target wants to requeue the I/O after a delay */
495 dm_requeue_original_request(tio, true);
499 DMWARN("unimplemented target map return value: %d", r);
503 /* The target wants to complete the I/O */
504 dm_kill_unmapped_request(rq, r);
510 static void dm_start_request(struct mapped_device *md, struct request *orig)
512 if (!orig->q->mq_ops)
513 blk_start_request(orig);
515 blk_mq_start_request(orig);
516 atomic_inc(&md->pending[rq_data_dir(orig)]);
518 if (md->seq_rq_merge_deadline_usecs) {
519 md->last_rq_pos = rq_end_sector(orig);
520 md->last_rq_rw = rq_data_dir(orig);
521 md->last_rq_start_time = ktime_get();
524 if (unlikely(dm_stats_used(&md->stats))) {
525 struct dm_rq_target_io *tio = tio_from_request(orig);
526 tio->duration_jiffies = jiffies;
527 tio->n_sectors = blk_rq_sectors(orig);
528 dm_stats_account_io(&md->stats, rq_data_dir(orig),
529 blk_rq_pos(orig), tio->n_sectors, false, 0,
534 * Hold the md reference here for the in-flight I/O.
535 * We can't rely on the reference count by device opener,
536 * because the device may be closed during the request completion
537 * when all bios are completed.
538 * See the comment in rq_completed() too.
543 static int __dm_rq_init_rq(struct mapped_device *md, struct request *rq)
545 struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
548 * Must initialize md member of tio, otherwise it won't
549 * be available in dm_mq_queue_rq.
553 if (md->init_tio_pdu) {
554 /* target-specific per-io data is immediately after the tio */
555 tio->info.ptr = tio + 1;
561 static int dm_rq_init_rq(struct request_queue *q, struct request *rq, gfp_t gfp)
563 return __dm_rq_init_rq(q->rq_alloc_data, rq);
566 static void map_tio_request(struct kthread_work *work)
568 struct dm_rq_target_io *tio = container_of(work, struct dm_rq_target_io, work);
570 if (map_request(tio) == DM_MAPIO_REQUEUE)
571 dm_requeue_original_request(tio, false);
574 ssize_t dm_attr_rq_based_seq_io_merge_deadline_show(struct mapped_device *md, char *buf)
576 return sprintf(buf, "%u\n", md->seq_rq_merge_deadline_usecs);
579 #define MAX_SEQ_RQ_MERGE_DEADLINE_USECS 100000
581 ssize_t dm_attr_rq_based_seq_io_merge_deadline_store(struct mapped_device *md,
582 const char *buf, size_t count)
586 if (dm_get_md_type(md) != DM_TYPE_REQUEST_BASED)
589 if (kstrtouint(buf, 10, &deadline))
592 if (deadline > MAX_SEQ_RQ_MERGE_DEADLINE_USECS)
593 deadline = MAX_SEQ_RQ_MERGE_DEADLINE_USECS;
595 md->seq_rq_merge_deadline_usecs = deadline;
600 static bool dm_old_request_peeked_before_merge_deadline(struct mapped_device *md)
604 if (!md->seq_rq_merge_deadline_usecs)
607 kt_deadline = ns_to_ktime((u64)md->seq_rq_merge_deadline_usecs * NSEC_PER_USEC);
608 kt_deadline = ktime_add_safe(md->last_rq_start_time, kt_deadline);
610 return !ktime_after(ktime_get(), kt_deadline);
614 * q->request_fn for old request-based dm.
615 * Called with the queue lock held.
617 static void dm_old_request_fn(struct request_queue *q)
619 struct mapped_device *md = q->queuedata;
620 struct dm_target *ti = md->immutable_target;
622 struct dm_rq_target_io *tio;
627 struct dm_table *map = dm_get_live_table(md, &srcu_idx);
629 if (unlikely(!map)) {
630 dm_put_live_table(md, srcu_idx);
633 ti = dm_table_find_target(map, pos);
634 dm_put_live_table(md, srcu_idx);
638 * For suspend, check blk_queue_stopped() and increment
639 * ->pending within a single queue_lock not to increment the
640 * number of in-flight I/Os after the queue is stopped in
643 while (!blk_queue_stopped(q)) {
644 rq = blk_peek_request(q);
648 /* always use block 0 to find the target for flushes for now */
650 if (req_op(rq) != REQ_OP_FLUSH)
651 pos = blk_rq_pos(rq);
653 if ((dm_old_request_peeked_before_merge_deadline(md) &&
654 md_in_flight(md) && rq->bio && !bio_multiple_segments(rq->bio) &&
655 md->last_rq_pos == pos && md->last_rq_rw == rq_data_dir(rq)) ||
656 (ti->type->busy && ti->type->busy(ti))) {
657 blk_delay_queue(q, 10);
661 dm_start_request(md, rq);
663 tio = tio_from_request(rq);
664 init_tio(tio, rq, md);
665 /* Establish tio->ti before queuing work (map_tio_request) */
667 kthread_queue_work(&md->kworker, &tio->work);
668 BUG_ON(!irqs_disabled());
673 * Fully initialize a .request_fn request-based queue.
675 int dm_old_init_request_queue(struct mapped_device *md, struct dm_table *t)
677 struct dm_target *immutable_tgt;
679 /* Fully initialize the queue */
680 md->queue->cmd_size = sizeof(struct dm_rq_target_io);
681 md->queue->rq_alloc_data = md;
682 md->queue->request_fn = dm_old_request_fn;
683 md->queue->init_rq_fn = dm_rq_init_rq;
685 immutable_tgt = dm_table_get_immutable_target(t);
686 if (immutable_tgt && immutable_tgt->per_io_data_size) {
687 /* any target-specific per-io data is immediately after the tio */
688 md->queue->cmd_size += immutable_tgt->per_io_data_size;
689 md->init_tio_pdu = true;
691 if (blk_init_allocated_queue(md->queue) < 0)
694 /* disable dm_old_request_fn's merge heuristic by default */
695 md->seq_rq_merge_deadline_usecs = 0;
697 dm_init_normal_md_queue(md);
698 blk_queue_softirq_done(md->queue, dm_softirq_done);
700 /* Initialize the request-based DM worker thread */
701 kthread_init_worker(&md->kworker);
702 md->kworker_task = kthread_run(kthread_worker_fn, &md->kworker,
703 "kdmwork-%s", dm_device_name(md));
704 if (IS_ERR(md->kworker_task)) {
705 int error = PTR_ERR(md->kworker_task);
706 md->kworker_task = NULL;
710 elv_register_queue(md->queue);
715 static int dm_mq_init_request(void *data, struct request *rq,
716 unsigned int hctx_idx, unsigned int request_idx,
717 unsigned int numa_node)
719 return __dm_rq_init_rq(data, rq);
722 static int dm_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
723 const struct blk_mq_queue_data *bd)
725 struct request *rq = bd->rq;
726 struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
727 struct mapped_device *md = tio->md;
728 struct dm_target *ti = md->immutable_target;
732 struct dm_table *map = dm_get_live_table(md, &srcu_idx);
734 ti = dm_table_find_target(map, 0);
735 dm_put_live_table(md, srcu_idx);
738 if (ti->type->busy && ti->type->busy(ti))
739 return BLK_MQ_RQ_QUEUE_BUSY;
741 dm_start_request(md, rq);
743 /* Init tio using md established in .init_request */
744 init_tio(tio, rq, md);
747 * Establish tio->ti before calling map_request().
751 /* Direct call is fine since .queue_rq allows allocations */
752 if (map_request(tio) == DM_MAPIO_REQUEUE) {
753 /* Undo dm_start_request() before requeuing */
754 rq_end_stats(md, rq);
755 rq_completed(md, rq_data_dir(rq), false);
756 return BLK_MQ_RQ_QUEUE_BUSY;
759 return BLK_MQ_RQ_QUEUE_OK;
762 static struct blk_mq_ops dm_mq_ops = {
763 .queue_rq = dm_mq_queue_rq,
764 .complete = dm_softirq_done,
765 .init_request = dm_mq_init_request,
768 int dm_mq_init_request_queue(struct mapped_device *md, struct dm_table *t)
770 struct request_queue *q;
771 struct dm_target *immutable_tgt;
774 if (!dm_table_all_blk_mq_devices(t)) {
775 DMERR("request-based dm-mq may only be stacked on blk-mq device(s)");
779 md->tag_set = kzalloc_node(sizeof(struct blk_mq_tag_set), GFP_KERNEL, md->numa_node_id);
783 md->tag_set->ops = &dm_mq_ops;
784 md->tag_set->queue_depth = dm_get_blk_mq_queue_depth();
785 md->tag_set->numa_node = md->numa_node_id;
786 md->tag_set->flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
787 md->tag_set->nr_hw_queues = dm_get_blk_mq_nr_hw_queues();
788 md->tag_set->driver_data = md;
790 md->tag_set->cmd_size = sizeof(struct dm_rq_target_io);
791 immutable_tgt = dm_table_get_immutable_target(t);
792 if (immutable_tgt && immutable_tgt->per_io_data_size) {
793 /* any target-specific per-io data is immediately after the tio */
794 md->tag_set->cmd_size += immutable_tgt->per_io_data_size;
795 md->init_tio_pdu = true;
798 err = blk_mq_alloc_tag_set(md->tag_set);
800 goto out_kfree_tag_set;
802 q = blk_mq_init_allocated_queue(md->tag_set, md->queue);
807 dm_init_md_queue(md);
809 /* backfill 'mq' sysfs registration normally done in blk_register_queue */
810 blk_mq_register_dev(disk_to_dev(md->disk), q);
815 blk_mq_free_tag_set(md->tag_set);
822 void dm_mq_cleanup_mapped_device(struct mapped_device *md)
825 blk_mq_free_tag_set(md->tag_set);
830 module_param(reserved_rq_based_ios, uint, S_IRUGO | S_IWUSR);
831 MODULE_PARM_DESC(reserved_rq_based_ios, "Reserved IOs in request-based mempools");
833 module_param(use_blk_mq, bool, S_IRUGO | S_IWUSR);
834 MODULE_PARM_DESC(use_blk_mq, "Use block multiqueue for request-based DM devices");
836 module_param(dm_mq_nr_hw_queues, uint, S_IRUGO | S_IWUSR);
837 MODULE_PARM_DESC(dm_mq_nr_hw_queues, "Number of hardware queues for request-based dm-mq devices");
839 module_param(dm_mq_queue_depth, uint, S_IRUGO | S_IWUSR);
840 MODULE_PARM_DESC(dm_mq_queue_depth, "Queue depth for request-based dm-mq devices");