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 #ifdef CONFIG_DM_MQ_DEFAULT
27 static bool use_blk_mq = true;
29 static bool use_blk_mq = false;
32 bool dm_use_blk_mq_default(void)
37 bool dm_use_blk_mq(struct mapped_device *md)
39 return md->use_blk_mq;
41 EXPORT_SYMBOL_GPL(dm_use_blk_mq);
43 unsigned dm_get_reserved_rq_based_ios(void)
45 return __dm_get_module_param(&reserved_rq_based_ios,
46 RESERVED_REQUEST_BASED_IOS, DM_RESERVED_MAX_IOS);
48 EXPORT_SYMBOL_GPL(dm_get_reserved_rq_based_ios);
50 static unsigned dm_get_blk_mq_nr_hw_queues(void)
52 return __dm_get_module_param(&dm_mq_nr_hw_queues, 1, 32);
55 static unsigned dm_get_blk_mq_queue_depth(void)
57 return __dm_get_module_param(&dm_mq_queue_depth,
58 DM_MQ_QUEUE_DEPTH, BLK_MQ_MAX_DEPTH);
61 int dm_request_based(struct mapped_device *md)
63 return blk_queue_stackable(md->queue);
66 static void dm_old_start_queue(struct request_queue *q)
70 spin_lock_irqsave(q->queue_lock, flags);
71 if (blk_queue_stopped(q))
73 spin_unlock_irqrestore(q->queue_lock, flags);
76 static void dm_mq_start_queue(struct request_queue *q)
80 spin_lock_irqsave(q->queue_lock, flags);
81 queue_flag_clear(QUEUE_FLAG_STOPPED, q);
82 spin_unlock_irqrestore(q->queue_lock, flags);
84 blk_mq_start_stopped_hw_queues(q, true);
85 blk_mq_kick_requeue_list(q);
88 void dm_start_queue(struct request_queue *q)
91 dm_old_start_queue(q);
96 static void dm_old_stop_queue(struct request_queue *q)
100 spin_lock_irqsave(q->queue_lock, flags);
101 if (!blk_queue_stopped(q))
103 spin_unlock_irqrestore(q->queue_lock, flags);
106 static void dm_mq_stop_queue(struct request_queue *q)
110 spin_lock_irqsave(q->queue_lock, flags);
111 if (blk_queue_stopped(q)) {
112 spin_unlock_irqrestore(q->queue_lock, flags);
116 queue_flag_set(QUEUE_FLAG_STOPPED, q);
117 spin_unlock_irqrestore(q->queue_lock, flags);
119 /* Avoid that requeuing could restart the queue. */
120 blk_mq_cancel_requeue_work(q);
121 blk_mq_stop_hw_queues(q);
124 void dm_stop_queue(struct request_queue *q)
127 dm_old_stop_queue(q);
132 static struct dm_rq_target_io *alloc_old_rq_tio(struct mapped_device *md,
135 return mempool_alloc(md->io_pool, gfp_mask);
138 static void free_old_rq_tio(struct dm_rq_target_io *tio)
140 mempool_free(tio, tio->md->io_pool);
143 static struct request *alloc_old_clone_request(struct mapped_device *md,
146 return mempool_alloc(md->rq_pool, gfp_mask);
149 static void free_old_clone_request(struct mapped_device *md, struct request *rq)
151 mempool_free(rq, md->rq_pool);
155 * Partial completion handling for request-based dm
157 static void end_clone_bio(struct bio *clone)
159 struct dm_rq_clone_bio_info *info =
160 container_of(clone, struct dm_rq_clone_bio_info, clone);
161 struct dm_rq_target_io *tio = info->tio;
162 struct bio *bio = info->orig;
163 unsigned int nr_bytes = info->orig->bi_iter.bi_size;
164 int error = clone->bi_error;
170 * An error has already been detected on the request.
171 * Once error occurred, just let clone->end_io() handle
177 * Don't notice the error to the upper layer yet.
178 * The error handling decision is made by the target driver,
179 * when the request is completed.
186 * I/O for the bio successfully completed.
187 * Notice the data completion to the upper layer.
191 * bios are processed from the head of the list.
192 * So the completing bio should always be rq->bio.
193 * If it's not, something wrong is happening.
195 if (tio->orig->bio != bio)
196 DMERR("bio completion is going in the middle of the request");
199 * Update the original request.
200 * Do not use blk_end_request() here, because it may complete
201 * the original request before the clone, and break the ordering.
203 blk_update_request(tio->orig, 0, nr_bytes);
206 static struct dm_rq_target_io *tio_from_request(struct request *rq)
208 return (rq->q->mq_ops ? blk_mq_rq_to_pdu(rq) : rq->special);
211 static void rq_end_stats(struct mapped_device *md, struct request *orig)
213 if (unlikely(dm_stats_used(&md->stats))) {
214 struct dm_rq_target_io *tio = tio_from_request(orig);
215 tio->duration_jiffies = jiffies - tio->duration_jiffies;
216 dm_stats_account_io(&md->stats, rq_data_dir(orig),
217 blk_rq_pos(orig), tio->n_sectors, true,
218 tio->duration_jiffies, &tio->stats_aux);
223 * Don't touch any member of the md after calling this function because
224 * the md may be freed in dm_put() at the end of this function.
225 * Or do dm_get() before calling this function and dm_put() later.
227 static void rq_completed(struct mapped_device *md, int rw, bool run_queue)
229 atomic_dec(&md->pending[rw]);
231 /* nudge anyone waiting on suspend queue */
232 if (!md_in_flight(md))
236 * Run this off this callpath, as drivers could invoke end_io while
237 * inside their request_fn (and holding the queue lock). Calling
238 * back into ->request_fn() could deadlock attempting to grab the
241 if (!md->queue->mq_ops && run_queue)
242 blk_run_queue_async(md->queue);
245 * dm_put() must be at the end of this function. See the comment above
250 static void free_rq_clone(struct request *clone)
252 struct dm_rq_target_io *tio = clone->end_io_data;
253 struct mapped_device *md = tio->md;
255 blk_rq_unprep_clone(clone);
258 * It is possible for a clone_old_rq() allocated clone to
259 * get passed in -- it may not yet have a request_queue.
260 * This is known to occur if the error target replaces
261 * a multipath target that has a request_fn queue stacked
262 * on blk-mq queue(s).
264 if (clone->q && clone->q->mq_ops)
265 /* stacked on blk-mq queue(s) */
266 tio->ti->type->release_clone_rq(clone);
267 else if (!md->queue->mq_ops)
268 /* request_fn queue stacked on request_fn queue(s) */
269 free_old_clone_request(md, clone);
271 if (!md->queue->mq_ops)
272 free_old_rq_tio(tio);
276 * Complete the clone and the original request.
277 * Must be called without clone's queue lock held,
278 * see end_clone_request() for more details.
280 static void dm_end_request(struct request *clone, int error)
282 int rw = rq_data_dir(clone);
283 struct dm_rq_target_io *tio = clone->end_io_data;
284 struct mapped_device *md = tio->md;
285 struct request *rq = tio->orig;
287 if (rq->cmd_type == REQ_TYPE_BLOCK_PC) {
288 rq->errors = clone->errors;
289 rq->resid_len = clone->resid_len;
293 * We are using the sense buffer of the original
295 * So setting the length of the sense data is enough.
297 rq->sense_len = clone->sense_len;
300 free_rq_clone(clone);
301 rq_end_stats(md, rq);
303 blk_end_request_all(rq, error);
305 blk_mq_end_request(rq, error);
306 rq_completed(md, rw, true);
309 static void dm_unprep_request(struct request *rq)
311 struct dm_rq_target_io *tio = tio_from_request(rq);
312 struct request *clone = tio->clone;
314 if (!rq->q->mq_ops) {
316 rq->cmd_flags &= ~REQ_DONTPREP;
320 free_rq_clone(clone);
321 else if (!tio->md->queue->mq_ops)
322 free_old_rq_tio(tio);
326 * Requeue the original request of a clone.
328 static void dm_old_requeue_request(struct request *rq)
330 struct request_queue *q = rq->q;
333 spin_lock_irqsave(q->queue_lock, flags);
334 blk_requeue_request(q, rq);
335 blk_run_queue_async(q);
336 spin_unlock_irqrestore(q->queue_lock, flags);
339 static void dm_mq_requeue_request(struct request *rq)
341 struct request_queue *q = rq->q;
344 blk_mq_requeue_request(rq);
345 spin_lock_irqsave(q->queue_lock, flags);
346 if (!blk_queue_stopped(q))
347 blk_mq_kick_requeue_list(q);
348 spin_unlock_irqrestore(q->queue_lock, flags);
351 static void dm_requeue_original_request(struct mapped_device *md,
354 int rw = rq_data_dir(rq);
356 rq_end_stats(md, rq);
357 dm_unprep_request(rq);
360 dm_old_requeue_request(rq);
362 dm_mq_requeue_request(rq);
364 rq_completed(md, rw, false);
367 static void dm_done(struct request *clone, int error, bool mapped)
370 struct dm_rq_target_io *tio = clone->end_io_data;
371 dm_request_endio_fn rq_end_io = NULL;
374 rq_end_io = tio->ti->type->rq_end_io;
376 if (mapped && rq_end_io)
377 r = rq_end_io(tio->ti, clone, error, &tio->info);
380 if (unlikely(r == -EREMOTEIO && (req_op(clone) == REQ_OP_WRITE_SAME) &&
381 !clone->q->limits.max_write_same_sectors))
382 disable_write_same(tio->md);
385 /* The target wants to complete the I/O */
386 dm_end_request(clone, r);
387 else if (r == DM_ENDIO_INCOMPLETE)
388 /* The target will handle the I/O */
390 else if (r == DM_ENDIO_REQUEUE)
391 /* The target wants to requeue the I/O */
392 dm_requeue_original_request(tio->md, tio->orig);
394 DMWARN("unimplemented target endio return value: %d", r);
400 * Request completion handler for request-based dm
402 static void dm_softirq_done(struct request *rq)
405 struct dm_rq_target_io *tio = tio_from_request(rq);
406 struct request *clone = tio->clone;
410 rq_end_stats(tio->md, rq);
411 rw = rq_data_dir(rq);
412 if (!rq->q->mq_ops) {
413 blk_end_request_all(rq, tio->error);
414 rq_completed(tio->md, rw, false);
415 free_old_rq_tio(tio);
417 blk_mq_end_request(rq, tio->error);
418 rq_completed(tio->md, rw, false);
423 if (rq->cmd_flags & REQ_FAILED)
426 dm_done(clone, tio->error, mapped);
430 * Complete the clone and the original request with the error status
431 * through softirq context.
433 static void dm_complete_request(struct request *rq, int error)
435 struct dm_rq_target_io *tio = tio_from_request(rq);
439 blk_complete_request(rq);
441 blk_mq_complete_request(rq, error);
445 * Complete the not-mapped clone and the original request with the error status
446 * through softirq context.
447 * Target's rq_end_io() function isn't called.
448 * This may be used when the target's map_rq() or clone_and_map_rq() functions fail.
450 static void dm_kill_unmapped_request(struct request *rq, int error)
452 rq->cmd_flags |= REQ_FAILED;
453 dm_complete_request(rq, error);
457 * Called with the clone's queue lock held (in the case of .request_fn)
459 static void end_clone_request(struct request *clone, int error)
461 struct dm_rq_target_io *tio = clone->end_io_data;
463 if (!clone->q->mq_ops) {
465 * For just cleaning up the information of the queue in which
466 * the clone was dispatched.
467 * The clone is *NOT* freed actually here because it is alloced
468 * from dm own mempool (REQ_ALLOCED isn't set).
470 __blk_put_request(clone->q, clone);
474 * Actual request completion is done in a softirq context which doesn't
475 * hold the clone's queue lock. Otherwise, deadlock could occur because:
476 * - another request may be submitted by the upper level driver
477 * of the stacking during the completion
478 * - the submission which requires queue lock may be done
479 * against this clone's queue
481 dm_complete_request(tio->orig, error);
484 static void dm_dispatch_clone_request(struct request *clone, struct request *rq)
488 if (blk_queue_io_stat(clone->q))
489 clone->cmd_flags |= REQ_IO_STAT;
491 clone->start_time = jiffies;
492 r = blk_insert_cloned_request(clone->q, clone);
494 /* must complete clone in terms of original request */
495 dm_complete_request(rq, r);
498 static int dm_rq_bio_constructor(struct bio *bio, struct bio *bio_orig,
501 struct dm_rq_target_io *tio = data;
502 struct dm_rq_clone_bio_info *info =
503 container_of(bio, struct dm_rq_clone_bio_info, clone);
505 info->orig = bio_orig;
507 bio->bi_end_io = end_clone_bio;
512 static int setup_clone(struct request *clone, struct request *rq,
513 struct dm_rq_target_io *tio, gfp_t gfp_mask)
517 r = blk_rq_prep_clone(clone, rq, tio->md->bs, gfp_mask,
518 dm_rq_bio_constructor, tio);
522 clone->cmd = rq->cmd;
523 clone->cmd_len = rq->cmd_len;
524 clone->sense = rq->sense;
525 clone->end_io = end_clone_request;
526 clone->end_io_data = tio;
533 static struct request *clone_old_rq(struct request *rq, struct mapped_device *md,
534 struct dm_rq_target_io *tio, gfp_t gfp_mask)
537 * Create clone for use with .request_fn request_queue
539 struct request *clone;
541 clone = alloc_old_clone_request(md, gfp_mask);
545 blk_rq_init(NULL, clone);
546 if (setup_clone(clone, rq, tio, gfp_mask)) {
548 free_old_clone_request(md, clone);
555 static void map_tio_request(struct kthread_work *work);
557 static void init_tio(struct dm_rq_target_io *tio, struct request *rq,
558 struct mapped_device *md)
566 * Avoid initializing info for blk-mq; it passes
567 * target-specific data through info.ptr
568 * (see: dm_mq_init_request)
570 if (!md->init_tio_pdu)
571 memset(&tio->info, 0, sizeof(tio->info));
572 if (md->kworker_task)
573 init_kthread_work(&tio->work, map_tio_request);
576 static struct dm_rq_target_io *dm_old_prep_tio(struct request *rq,
577 struct mapped_device *md,
580 struct dm_rq_target_io *tio;
582 struct dm_table *table;
584 tio = alloc_old_rq_tio(md, gfp_mask);
588 init_tio(tio, rq, md);
590 table = dm_get_live_table(md, &srcu_idx);
592 * Must clone a request if this .request_fn DM device
593 * is stacked on .request_fn device(s).
595 if (!dm_table_all_blk_mq_devices(table)) {
596 if (!clone_old_rq(rq, md, tio, gfp_mask)) {
597 dm_put_live_table(md, srcu_idx);
598 free_old_rq_tio(tio);
602 dm_put_live_table(md, srcu_idx);
608 * Called with the queue lock held.
610 static int dm_old_prep_fn(struct request_queue *q, struct request *rq)
612 struct mapped_device *md = q->queuedata;
613 struct dm_rq_target_io *tio;
615 if (unlikely(rq->special)) {
616 DMWARN("Already has something in rq->special.");
620 tio = dm_old_prep_tio(rq, md, GFP_ATOMIC);
622 return BLKPREP_DEFER;
625 rq->cmd_flags |= REQ_DONTPREP;
632 * 0 : the request has been processed
633 * DM_MAPIO_REQUEUE : the original request needs to be requeued
634 * < 0 : the request was completed due to failure
636 static int map_request(struct dm_rq_target_io *tio, struct request *rq,
637 struct mapped_device *md)
640 struct dm_target *ti = tio->ti;
641 struct request *clone = NULL;
645 r = ti->type->map_rq(ti, clone, &tio->info);
647 r = ti->type->clone_and_map_rq(ti, rq, &tio->info, &clone);
649 /* The target wants to complete the I/O */
650 dm_kill_unmapped_request(rq, r);
653 if (r != DM_MAPIO_REMAPPED)
655 if (setup_clone(clone, rq, tio, GFP_ATOMIC)) {
657 ti->type->release_clone_rq(clone);
658 return DM_MAPIO_REQUEUE;
663 case DM_MAPIO_SUBMITTED:
664 /* The target has taken the I/O to submit by itself later */
666 case DM_MAPIO_REMAPPED:
667 /* The target has remapped the I/O so dispatch it */
668 trace_block_rq_remap(clone->q, clone, disk_devt(dm_disk(md)),
670 dm_dispatch_clone_request(clone, rq);
672 case DM_MAPIO_REQUEUE:
673 /* The target wants to requeue the I/O */
674 dm_requeue_original_request(md, tio->orig);
678 DMWARN("unimplemented target map return value: %d", r);
682 /* The target wants to complete the I/O */
683 dm_kill_unmapped_request(rq, r);
690 static void dm_start_request(struct mapped_device *md, struct request *orig)
692 if (!orig->q->mq_ops)
693 blk_start_request(orig);
695 blk_mq_start_request(orig);
696 atomic_inc(&md->pending[rq_data_dir(orig)]);
698 if (md->seq_rq_merge_deadline_usecs) {
699 md->last_rq_pos = rq_end_sector(orig);
700 md->last_rq_rw = rq_data_dir(orig);
701 md->last_rq_start_time = ktime_get();
704 if (unlikely(dm_stats_used(&md->stats))) {
705 struct dm_rq_target_io *tio = tio_from_request(orig);
706 tio->duration_jiffies = jiffies;
707 tio->n_sectors = blk_rq_sectors(orig);
708 dm_stats_account_io(&md->stats, rq_data_dir(orig),
709 blk_rq_pos(orig), tio->n_sectors, false, 0,
714 * Hold the md reference here for the in-flight I/O.
715 * We can't rely on the reference count by device opener,
716 * because the device may be closed during the request completion
717 * when all bios are completed.
718 * See the comment in rq_completed() too.
723 static void map_tio_request(struct kthread_work *work)
725 struct dm_rq_target_io *tio = container_of(work, struct dm_rq_target_io, work);
726 struct request *rq = tio->orig;
727 struct mapped_device *md = tio->md;
729 if (map_request(tio, rq, md) == DM_MAPIO_REQUEUE)
730 dm_requeue_original_request(md, rq);
733 ssize_t dm_attr_rq_based_seq_io_merge_deadline_show(struct mapped_device *md, char *buf)
735 return sprintf(buf, "%u\n", md->seq_rq_merge_deadline_usecs);
738 #define MAX_SEQ_RQ_MERGE_DEADLINE_USECS 100000
740 ssize_t dm_attr_rq_based_seq_io_merge_deadline_store(struct mapped_device *md,
741 const char *buf, size_t count)
745 if (dm_get_md_type(md) != DM_TYPE_REQUEST_BASED)
748 if (kstrtouint(buf, 10, &deadline))
751 if (deadline > MAX_SEQ_RQ_MERGE_DEADLINE_USECS)
752 deadline = MAX_SEQ_RQ_MERGE_DEADLINE_USECS;
754 md->seq_rq_merge_deadline_usecs = deadline;
759 static bool dm_old_request_peeked_before_merge_deadline(struct mapped_device *md)
763 if (!md->seq_rq_merge_deadline_usecs)
766 kt_deadline = ns_to_ktime((u64)md->seq_rq_merge_deadline_usecs * NSEC_PER_USEC);
767 kt_deadline = ktime_add_safe(md->last_rq_start_time, kt_deadline);
769 return !ktime_after(ktime_get(), kt_deadline);
773 * q->request_fn for old request-based dm.
774 * Called with the queue lock held.
776 static void dm_old_request_fn(struct request_queue *q)
778 struct mapped_device *md = q->queuedata;
779 struct dm_target *ti = md->immutable_target;
781 struct dm_rq_target_io *tio;
786 struct dm_table *map = dm_get_live_table(md, &srcu_idx);
788 ti = dm_table_find_target(map, pos);
789 dm_put_live_table(md, srcu_idx);
793 * For suspend, check blk_queue_stopped() and increment
794 * ->pending within a single queue_lock not to increment the
795 * number of in-flight I/Os after the queue is stopped in
798 while (!blk_queue_stopped(q)) {
799 rq = blk_peek_request(q);
803 /* always use block 0 to find the target for flushes for now */
805 if (req_op(rq) != REQ_OP_FLUSH)
806 pos = blk_rq_pos(rq);
808 if ((dm_old_request_peeked_before_merge_deadline(md) &&
809 md_in_flight(md) && rq->bio && rq->bio->bi_vcnt == 1 &&
810 md->last_rq_pos == pos && md->last_rq_rw == rq_data_dir(rq)) ||
811 (ti->type->busy && ti->type->busy(ti))) {
812 blk_delay_queue(q, 10);
816 dm_start_request(md, rq);
818 tio = tio_from_request(rq);
819 /* Establish tio->ti before queuing work (map_tio_request) */
821 queue_kthread_work(&md->kworker, &tio->work);
822 BUG_ON(!irqs_disabled());
827 * Fully initialize a .request_fn request-based queue.
829 int dm_old_init_request_queue(struct mapped_device *md)
831 /* Fully initialize the queue */
832 if (!blk_init_allocated_queue(md->queue, dm_old_request_fn, NULL))
835 /* disable dm_old_request_fn's merge heuristic by default */
836 md->seq_rq_merge_deadline_usecs = 0;
838 dm_init_normal_md_queue(md);
839 blk_queue_softirq_done(md->queue, dm_softirq_done);
840 blk_queue_prep_rq(md->queue, dm_old_prep_fn);
842 /* Initialize the request-based DM worker thread */
843 init_kthread_worker(&md->kworker);
844 md->kworker_task = kthread_run(kthread_worker_fn, &md->kworker,
845 "kdmwork-%s", dm_device_name(md));
846 if (IS_ERR(md->kworker_task))
847 return PTR_ERR(md->kworker_task);
849 elv_register_queue(md->queue);
854 static int dm_mq_init_request(void *data, struct request *rq,
855 unsigned int hctx_idx, unsigned int request_idx,
856 unsigned int numa_node)
858 struct mapped_device *md = data;
859 struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
862 * Must initialize md member of tio, otherwise it won't
863 * be available in dm_mq_queue_rq.
867 if (md->init_tio_pdu) {
868 /* target-specific per-io data is immediately after the tio */
869 tio->info.ptr = tio + 1;
875 static int dm_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
876 const struct blk_mq_queue_data *bd)
878 struct request *rq = bd->rq;
879 struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
880 struct mapped_device *md = tio->md;
881 struct dm_target *ti = md->immutable_target;
885 struct dm_table *map = dm_get_live_table(md, &srcu_idx);
887 ti = dm_table_find_target(map, 0);
888 dm_put_live_table(md, srcu_idx);
892 * On suspend dm_stop_queue() handles stopping the blk-mq
893 * request_queue BUT: even though the hw_queues are marked
894 * BLK_MQ_S_STOPPED at that point there is still a race that
895 * is allowing block/blk-mq.c to call ->queue_rq against a
896 * hctx that it really shouldn't. The following check guards
897 * against this rarity (albeit _not_ race-free).
899 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
900 return BLK_MQ_RQ_QUEUE_BUSY;
902 if (ti->type->busy && ti->type->busy(ti))
903 return BLK_MQ_RQ_QUEUE_BUSY;
905 dm_start_request(md, rq);
907 /* Init tio using md established in .init_request */
908 init_tio(tio, rq, md);
911 * Establish tio->ti before calling map_request().
915 /* Direct call is fine since .queue_rq allows allocations */
916 if (map_request(tio, rq, md) == DM_MAPIO_REQUEUE) {
917 /* Undo dm_start_request() before requeuing */
918 rq_end_stats(md, rq);
919 rq_completed(md, rq_data_dir(rq), false);
920 return BLK_MQ_RQ_QUEUE_BUSY;
923 return BLK_MQ_RQ_QUEUE_OK;
926 static struct blk_mq_ops dm_mq_ops = {
927 .queue_rq = dm_mq_queue_rq,
928 .map_queue = blk_mq_map_queue,
929 .complete = dm_softirq_done,
930 .init_request = dm_mq_init_request,
933 int dm_mq_init_request_queue(struct mapped_device *md, struct dm_table *t)
935 struct request_queue *q;
936 struct dm_target *immutable_tgt;
939 if (!dm_table_all_blk_mq_devices(t)) {
940 DMERR("request-based dm-mq may only be stacked on blk-mq device(s)");
944 md->tag_set = kzalloc_node(sizeof(struct blk_mq_tag_set), GFP_KERNEL, md->numa_node_id);
948 md->tag_set->ops = &dm_mq_ops;
949 md->tag_set->queue_depth = dm_get_blk_mq_queue_depth();
950 md->tag_set->numa_node = md->numa_node_id;
951 md->tag_set->flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
952 md->tag_set->nr_hw_queues = dm_get_blk_mq_nr_hw_queues();
953 md->tag_set->driver_data = md;
955 md->tag_set->cmd_size = sizeof(struct dm_rq_target_io);
956 immutable_tgt = dm_table_get_immutable_target(t);
957 if (immutable_tgt && immutable_tgt->per_io_data_size) {
958 /* any target-specific per-io data is immediately after the tio */
959 md->tag_set->cmd_size += immutable_tgt->per_io_data_size;
960 md->init_tio_pdu = true;
963 err = blk_mq_alloc_tag_set(md->tag_set);
965 goto out_kfree_tag_set;
967 q = blk_mq_init_allocated_queue(md->tag_set, md->queue);
972 dm_init_md_queue(md);
974 /* backfill 'mq' sysfs registration normally done in blk_register_queue */
975 blk_mq_register_disk(md->disk);
980 blk_mq_free_tag_set(md->tag_set);
987 void dm_mq_cleanup_mapped_device(struct mapped_device *md)
990 blk_mq_free_tag_set(md->tag_set);
995 module_param(reserved_rq_based_ios, uint, S_IRUGO | S_IWUSR);
996 MODULE_PARM_DESC(reserved_rq_based_ios, "Reserved IOs in request-based mempools");
998 module_param(use_blk_mq, bool, S_IRUGO | S_IWUSR);
999 MODULE_PARM_DESC(use_blk_mq, "Use block multiqueue for request-based DM devices");
1001 module_param(dm_mq_nr_hw_queues, uint, S_IRUGO | S_IWUSR);
1002 MODULE_PARM_DESC(dm_mq_nr_hw_queues, "Number of hardware queues for request-based dm-mq devices");
1004 module_param(dm_mq_queue_depth, uint, S_IRUGO | S_IWUSR);
1005 MODULE_PARM_DESC(dm_mq_queue_depth, "Queue depth for request-based dm-mq devices");