2 * Copyright (C) 2001, 2002 Sistina Software (UK) Limited.
3 * Copyright (C) 2004-2008 Red Hat, Inc. All rights reserved.
5 * This file is released under the GPL.
11 #include <linux/init.h>
12 #include <linux/module.h>
13 #include <linux/mutex.h>
14 #include <linux/moduleparam.h>
15 #include <linux/blkpg.h>
16 #include <linux/bio.h>
17 #include <linux/mempool.h>
18 #include <linux/slab.h>
19 #include <linux/idr.h>
20 #include <linux/hdreg.h>
21 #include <linux/delay.h>
22 #include <linux/wait.h>
23 #include <linux/kthread.h>
24 #include <linux/ktime.h>
25 #include <linux/elevator.h> /* for rq_end_sector() */
26 #include <linux/blk-mq.h>
28 #include <trace/events/block.h>
30 #define DM_MSG_PREFIX "core"
34 * ratelimit state to be used in DMXXX_LIMIT().
36 DEFINE_RATELIMIT_STATE(dm_ratelimit_state,
37 DEFAULT_RATELIMIT_INTERVAL,
38 DEFAULT_RATELIMIT_BURST);
39 EXPORT_SYMBOL(dm_ratelimit_state);
43 * Cookies are numeric values sent with CHANGE and REMOVE
44 * uevents while resuming, removing or renaming the device.
46 #define DM_COOKIE_ENV_VAR_NAME "DM_COOKIE"
47 #define DM_COOKIE_LENGTH 24
49 static const char *_name = DM_NAME;
51 static unsigned int major = 0;
52 static unsigned int _major = 0;
54 static DEFINE_IDR(_minor_idr);
56 static DEFINE_SPINLOCK(_minor_lock);
58 static void do_deferred_remove(struct work_struct *w);
60 static DECLARE_WORK(deferred_remove_work, do_deferred_remove);
62 static struct workqueue_struct *deferred_remove_workqueue;
66 * One of these is allocated per bio.
69 struct mapped_device *md;
73 unsigned long start_time;
74 spinlock_t endio_lock;
75 struct dm_stats_aux stats_aux;
79 * For request-based dm.
80 * One of these is allocated per request.
82 struct dm_rq_target_io {
83 struct mapped_device *md;
85 struct request *orig, *clone;
86 struct kthread_work work;
89 struct dm_stats_aux stats_aux;
90 unsigned long duration_jiffies;
95 * For request-based dm - the bio clones we allocate are embedded in these
98 * We allocate these with bio_alloc_bioset, using the front_pad parameter when
99 * the bioset is created - this means the bio has to come at the end of the
102 struct dm_rq_clone_bio_info {
104 struct dm_rq_target_io *tio;
108 union map_info *dm_get_rq_mapinfo(struct request *rq)
110 if (rq && rq->end_io_data)
111 return &((struct dm_rq_target_io *)rq->end_io_data)->info;
114 EXPORT_SYMBOL_GPL(dm_get_rq_mapinfo);
116 #define MINOR_ALLOCED ((void *)-1)
119 * Bits for the md->flags field.
121 #define DMF_BLOCK_IO_FOR_SUSPEND 0
122 #define DMF_SUSPENDED 1
124 #define DMF_FREEING 3
125 #define DMF_DELETING 4
126 #define DMF_NOFLUSH_SUSPENDING 5
127 #define DMF_DEFERRED_REMOVE 6
128 #define DMF_SUSPENDED_INTERNALLY 7
131 * A dummy definition to make RCU happy.
132 * struct dm_table should never be dereferenced in this file.
139 * Work processed by per-device workqueue.
141 struct mapped_device {
142 struct srcu_struct io_barrier;
143 struct mutex suspend_lock;
148 * The current mapping.
149 * Use dm_get_live_table{_fast} or take suspend_lock for
152 struct dm_table __rcu *map;
154 struct list_head table_devices;
155 struct mutex table_devices_lock;
159 struct request_queue *queue;
161 /* Protect queue and type against concurrent access. */
162 struct mutex type_lock;
164 struct target_type *immutable_target_type;
166 struct gendisk *disk;
172 * A list of ios that arrived while we were suspended.
175 wait_queue_head_t wait;
176 struct work_struct work;
177 struct bio_list deferred;
178 spinlock_t deferred_lock;
181 * Processing queue (flush)
183 struct workqueue_struct *wq;
186 * io objects are allocated from here.
197 wait_queue_head_t eventq;
199 struct list_head uevent_list;
200 spinlock_t uevent_lock; /* Protect access to uevent_list */
203 * freeze/thaw support require holding onto a super block
205 struct super_block *frozen_sb;
206 struct block_device *bdev;
208 /* forced geometry settings */
209 struct hd_geometry geometry;
211 /* kobject and completion */
212 struct dm_kobject_holder kobj_holder;
214 /* zero-length flush that will be cloned and submitted to targets */
215 struct bio flush_bio;
217 /* the number of internal suspends */
218 unsigned internal_suspend_count;
220 struct dm_stats stats;
222 struct kthread_worker kworker;
223 struct task_struct *kworker_task;
225 /* for request-based merge heuristic in dm_request_fn() */
226 unsigned seq_rq_merge_deadline_usecs;
228 sector_t last_rq_pos;
229 ktime_t last_rq_start_time;
231 /* for blk-mq request-based DM support */
232 struct blk_mq_tag_set tag_set;
236 #ifdef CONFIG_DM_MQ_DEFAULT
237 static bool use_blk_mq = true;
239 static bool use_blk_mq = false;
242 bool dm_use_blk_mq(struct mapped_device *md)
244 return md->use_blk_mq;
248 * For mempools pre-allocation at the table loading time.
250 struct dm_md_mempools {
256 struct table_device {
257 struct list_head list;
259 struct dm_dev dm_dev;
262 #define RESERVED_BIO_BASED_IOS 16
263 #define RESERVED_REQUEST_BASED_IOS 256
264 #define RESERVED_MAX_IOS 1024
265 static struct kmem_cache *_io_cache;
266 static struct kmem_cache *_rq_tio_cache;
267 static struct kmem_cache *_rq_cache;
270 * Bio-based DM's mempools' reserved IOs set by the user.
272 static unsigned reserved_bio_based_ios = RESERVED_BIO_BASED_IOS;
275 * Request-based DM's mempools' reserved IOs set by the user.
277 static unsigned reserved_rq_based_ios = RESERVED_REQUEST_BASED_IOS;
279 static unsigned __dm_get_module_param(unsigned *module_param,
280 unsigned def, unsigned max)
282 unsigned param = ACCESS_ONCE(*module_param);
283 unsigned modified_param = 0;
286 modified_param = def;
287 else if (param > max)
288 modified_param = max;
290 if (modified_param) {
291 (void)cmpxchg(module_param, param, modified_param);
292 param = modified_param;
298 unsigned dm_get_reserved_bio_based_ios(void)
300 return __dm_get_module_param(&reserved_bio_based_ios,
301 RESERVED_BIO_BASED_IOS, RESERVED_MAX_IOS);
303 EXPORT_SYMBOL_GPL(dm_get_reserved_bio_based_ios);
305 unsigned dm_get_reserved_rq_based_ios(void)
307 return __dm_get_module_param(&reserved_rq_based_ios,
308 RESERVED_REQUEST_BASED_IOS, RESERVED_MAX_IOS);
310 EXPORT_SYMBOL_GPL(dm_get_reserved_rq_based_ios);
312 static int __init local_init(void)
316 /* allocate a slab for the dm_ios */
317 _io_cache = KMEM_CACHE(dm_io, 0);
321 _rq_tio_cache = KMEM_CACHE(dm_rq_target_io, 0);
323 goto out_free_io_cache;
325 _rq_cache = kmem_cache_create("dm_clone_request", sizeof(struct request),
326 __alignof__(struct request), 0, NULL);
328 goto out_free_rq_tio_cache;
330 r = dm_uevent_init();
332 goto out_free_rq_cache;
334 deferred_remove_workqueue = alloc_workqueue("kdmremove", WQ_UNBOUND, 1);
335 if (!deferred_remove_workqueue) {
337 goto out_uevent_exit;
341 r = register_blkdev(_major, _name);
343 goto out_free_workqueue;
351 destroy_workqueue(deferred_remove_workqueue);
355 kmem_cache_destroy(_rq_cache);
356 out_free_rq_tio_cache:
357 kmem_cache_destroy(_rq_tio_cache);
359 kmem_cache_destroy(_io_cache);
364 static void local_exit(void)
366 flush_scheduled_work();
367 destroy_workqueue(deferred_remove_workqueue);
369 kmem_cache_destroy(_rq_cache);
370 kmem_cache_destroy(_rq_tio_cache);
371 kmem_cache_destroy(_io_cache);
372 unregister_blkdev(_major, _name);
377 DMINFO("cleaned up");
380 static int (*_inits[])(void) __initdata = {
391 static void (*_exits[])(void) = {
402 static int __init dm_init(void)
404 const int count = ARRAY_SIZE(_inits);
408 for (i = 0; i < count; i++) {
423 static void __exit dm_exit(void)
425 int i = ARRAY_SIZE(_exits);
431 * Should be empty by this point.
433 idr_destroy(&_minor_idr);
437 * Block device functions
439 int dm_deleting_md(struct mapped_device *md)
441 return test_bit(DMF_DELETING, &md->flags);
444 static int dm_blk_open(struct block_device *bdev, fmode_t mode)
446 struct mapped_device *md;
448 spin_lock(&_minor_lock);
450 md = bdev->bd_disk->private_data;
454 if (test_bit(DMF_FREEING, &md->flags) ||
455 dm_deleting_md(md)) {
461 atomic_inc(&md->open_count);
463 spin_unlock(&_minor_lock);
465 return md ? 0 : -ENXIO;
468 static void dm_blk_close(struct gendisk *disk, fmode_t mode)
470 struct mapped_device *md;
472 spin_lock(&_minor_lock);
474 md = disk->private_data;
478 if (atomic_dec_and_test(&md->open_count) &&
479 (test_bit(DMF_DEFERRED_REMOVE, &md->flags)))
480 queue_work(deferred_remove_workqueue, &deferred_remove_work);
484 spin_unlock(&_minor_lock);
487 int dm_open_count(struct mapped_device *md)
489 return atomic_read(&md->open_count);
493 * Guarantees nothing is using the device before it's deleted.
495 int dm_lock_for_deletion(struct mapped_device *md, bool mark_deferred, bool only_deferred)
499 spin_lock(&_minor_lock);
501 if (dm_open_count(md)) {
504 set_bit(DMF_DEFERRED_REMOVE, &md->flags);
505 } else if (only_deferred && !test_bit(DMF_DEFERRED_REMOVE, &md->flags))
508 set_bit(DMF_DELETING, &md->flags);
510 spin_unlock(&_minor_lock);
515 int dm_cancel_deferred_remove(struct mapped_device *md)
519 spin_lock(&_minor_lock);
521 if (test_bit(DMF_DELETING, &md->flags))
524 clear_bit(DMF_DEFERRED_REMOVE, &md->flags);
526 spin_unlock(&_minor_lock);
531 static void do_deferred_remove(struct work_struct *w)
533 dm_deferred_remove();
536 sector_t dm_get_size(struct mapped_device *md)
538 return get_capacity(md->disk);
541 struct request_queue *dm_get_md_queue(struct mapped_device *md)
546 struct dm_stats *dm_get_stats(struct mapped_device *md)
551 static int dm_blk_getgeo(struct block_device *bdev, struct hd_geometry *geo)
553 struct mapped_device *md = bdev->bd_disk->private_data;
555 return dm_get_geometry(md, geo);
558 static int dm_blk_ioctl(struct block_device *bdev, fmode_t mode,
559 unsigned int cmd, unsigned long arg)
561 struct mapped_device *md = bdev->bd_disk->private_data;
563 struct dm_table *map;
564 struct dm_target *tgt;
568 map = dm_get_live_table(md, &srcu_idx);
570 if (!map || !dm_table_get_size(map))
573 /* We only support devices that have a single target */
574 if (dm_table_get_num_targets(map) != 1)
577 tgt = dm_table_get_target(map, 0);
578 if (!tgt->type->ioctl)
581 if (dm_suspended_md(md)) {
586 r = tgt->type->ioctl(tgt, cmd, arg);
589 dm_put_live_table(md, srcu_idx);
591 if (r == -ENOTCONN) {
599 static struct dm_io *alloc_io(struct mapped_device *md)
601 return mempool_alloc(md->io_pool, GFP_NOIO);
604 static void free_io(struct mapped_device *md, struct dm_io *io)
606 mempool_free(io, md->io_pool);
609 static void free_tio(struct mapped_device *md, struct dm_target_io *tio)
611 bio_put(&tio->clone);
614 static struct dm_rq_target_io *alloc_rq_tio(struct mapped_device *md,
617 return mempool_alloc(md->io_pool, gfp_mask);
620 static void free_rq_tio(struct dm_rq_target_io *tio)
622 mempool_free(tio, tio->md->io_pool);
625 static struct request *alloc_clone_request(struct mapped_device *md,
628 return mempool_alloc(md->rq_pool, gfp_mask);
631 static void free_clone_request(struct mapped_device *md, struct request *rq)
633 mempool_free(rq, md->rq_pool);
636 static int md_in_flight(struct mapped_device *md)
638 return atomic_read(&md->pending[READ]) +
639 atomic_read(&md->pending[WRITE]);
642 static void start_io_acct(struct dm_io *io)
644 struct mapped_device *md = io->md;
645 struct bio *bio = io->bio;
647 int rw = bio_data_dir(bio);
649 io->start_time = jiffies;
651 cpu = part_stat_lock();
652 part_round_stats(cpu, &dm_disk(md)->part0);
654 atomic_set(&dm_disk(md)->part0.in_flight[rw],
655 atomic_inc_return(&md->pending[rw]));
657 if (unlikely(dm_stats_used(&md->stats)))
658 dm_stats_account_io(&md->stats, bio->bi_rw, bio->bi_iter.bi_sector,
659 bio_sectors(bio), false, 0, &io->stats_aux);
662 static void end_io_acct(struct dm_io *io)
664 struct mapped_device *md = io->md;
665 struct bio *bio = io->bio;
666 unsigned long duration = jiffies - io->start_time;
668 int rw = bio_data_dir(bio);
670 generic_end_io_acct(rw, &dm_disk(md)->part0, io->start_time);
672 if (unlikely(dm_stats_used(&md->stats)))
673 dm_stats_account_io(&md->stats, bio->bi_rw, bio->bi_iter.bi_sector,
674 bio_sectors(bio), true, duration, &io->stats_aux);
677 * After this is decremented the bio must not be touched if it is
680 pending = atomic_dec_return(&md->pending[rw]);
681 atomic_set(&dm_disk(md)->part0.in_flight[rw], pending);
682 pending += atomic_read(&md->pending[rw^0x1]);
684 /* nudge anyone waiting on suspend queue */
690 * Add the bio to the list of deferred io.
692 static void queue_io(struct mapped_device *md, struct bio *bio)
696 spin_lock_irqsave(&md->deferred_lock, flags);
697 bio_list_add(&md->deferred, bio);
698 spin_unlock_irqrestore(&md->deferred_lock, flags);
699 queue_work(md->wq, &md->work);
703 * Everyone (including functions in this file), should use this
704 * function to access the md->map field, and make sure they call
705 * dm_put_live_table() when finished.
707 struct dm_table *dm_get_live_table(struct mapped_device *md, int *srcu_idx) __acquires(md->io_barrier)
709 *srcu_idx = srcu_read_lock(&md->io_barrier);
711 return srcu_dereference(md->map, &md->io_barrier);
714 void dm_put_live_table(struct mapped_device *md, int srcu_idx) __releases(md->io_barrier)
716 srcu_read_unlock(&md->io_barrier, srcu_idx);
719 void dm_sync_table(struct mapped_device *md)
721 synchronize_srcu(&md->io_barrier);
722 synchronize_rcu_expedited();
726 * A fast alternative to dm_get_live_table/dm_put_live_table.
727 * The caller must not block between these two functions.
729 static struct dm_table *dm_get_live_table_fast(struct mapped_device *md) __acquires(RCU)
732 return rcu_dereference(md->map);
735 static void dm_put_live_table_fast(struct mapped_device *md) __releases(RCU)
741 * Open a table device so we can use it as a map destination.
743 static int open_table_device(struct table_device *td, dev_t dev,
744 struct mapped_device *md)
746 static char *_claim_ptr = "I belong to device-mapper";
747 struct block_device *bdev;
751 BUG_ON(td->dm_dev.bdev);
753 bdev = blkdev_get_by_dev(dev, td->dm_dev.mode | FMODE_EXCL, _claim_ptr);
755 return PTR_ERR(bdev);
757 r = bd_link_disk_holder(bdev, dm_disk(md));
759 blkdev_put(bdev, td->dm_dev.mode | FMODE_EXCL);
763 td->dm_dev.bdev = bdev;
768 * Close a table device that we've been using.
770 static void close_table_device(struct table_device *td, struct mapped_device *md)
772 if (!td->dm_dev.bdev)
775 bd_unlink_disk_holder(td->dm_dev.bdev, dm_disk(md));
776 blkdev_put(td->dm_dev.bdev, td->dm_dev.mode | FMODE_EXCL);
777 td->dm_dev.bdev = NULL;
780 static struct table_device *find_table_device(struct list_head *l, dev_t dev,
782 struct table_device *td;
784 list_for_each_entry(td, l, list)
785 if (td->dm_dev.bdev->bd_dev == dev && td->dm_dev.mode == mode)
791 int dm_get_table_device(struct mapped_device *md, dev_t dev, fmode_t mode,
792 struct dm_dev **result) {
794 struct table_device *td;
796 mutex_lock(&md->table_devices_lock);
797 td = find_table_device(&md->table_devices, dev, mode);
799 td = kmalloc(sizeof(*td), GFP_KERNEL);
801 mutex_unlock(&md->table_devices_lock);
805 td->dm_dev.mode = mode;
806 td->dm_dev.bdev = NULL;
808 if ((r = open_table_device(td, dev, md))) {
809 mutex_unlock(&md->table_devices_lock);
814 format_dev_t(td->dm_dev.name, dev);
816 atomic_set(&td->count, 0);
817 list_add(&td->list, &md->table_devices);
819 atomic_inc(&td->count);
820 mutex_unlock(&md->table_devices_lock);
822 *result = &td->dm_dev;
825 EXPORT_SYMBOL_GPL(dm_get_table_device);
827 void dm_put_table_device(struct mapped_device *md, struct dm_dev *d)
829 struct table_device *td = container_of(d, struct table_device, dm_dev);
831 mutex_lock(&md->table_devices_lock);
832 if (atomic_dec_and_test(&td->count)) {
833 close_table_device(td, md);
837 mutex_unlock(&md->table_devices_lock);
839 EXPORT_SYMBOL(dm_put_table_device);
841 static void free_table_devices(struct list_head *devices)
843 struct list_head *tmp, *next;
845 list_for_each_safe(tmp, next, devices) {
846 struct table_device *td = list_entry(tmp, struct table_device, list);
848 DMWARN("dm_destroy: %s still exists with %d references",
849 td->dm_dev.name, atomic_read(&td->count));
855 * Get the geometry associated with a dm device
857 int dm_get_geometry(struct mapped_device *md, struct hd_geometry *geo)
865 * Set the geometry of a device.
867 int dm_set_geometry(struct mapped_device *md, struct hd_geometry *geo)
869 sector_t sz = (sector_t)geo->cylinders * geo->heads * geo->sectors;
871 if (geo->start > sz) {
872 DMWARN("Start sector is beyond the geometry limits.");
881 /*-----------------------------------------------------------------
883 * A more elegant soln is in the works that uses the queue
884 * merge fn, unfortunately there are a couple of changes to
885 * the block layer that I want to make for this. So in the
886 * interests of getting something for people to use I give
887 * you this clearly demarcated crap.
888 *---------------------------------------------------------------*/
890 static int __noflush_suspending(struct mapped_device *md)
892 return test_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
896 * Decrements the number of outstanding ios that a bio has been
897 * cloned into, completing the original io if necc.
899 static void dec_pending(struct dm_io *io, int error)
904 struct mapped_device *md = io->md;
906 /* Push-back supersedes any I/O errors */
907 if (unlikely(error)) {
908 spin_lock_irqsave(&io->endio_lock, flags);
909 if (!(io->error > 0 && __noflush_suspending(md)))
911 spin_unlock_irqrestore(&io->endio_lock, flags);
914 if (atomic_dec_and_test(&io->io_count)) {
915 if (io->error == DM_ENDIO_REQUEUE) {
917 * Target requested pushing back the I/O.
919 spin_lock_irqsave(&md->deferred_lock, flags);
920 if (__noflush_suspending(md))
921 bio_list_add_head(&md->deferred, io->bio);
923 /* noflush suspend was interrupted. */
925 spin_unlock_irqrestore(&md->deferred_lock, flags);
928 io_error = io->error;
933 if (io_error == DM_ENDIO_REQUEUE)
936 if ((bio->bi_rw & REQ_FLUSH) && bio->bi_iter.bi_size) {
938 * Preflush done for flush with data, reissue
941 bio->bi_rw &= ~REQ_FLUSH;
944 /* done with normal IO or empty flush */
945 trace_block_bio_complete(md->queue, bio, io_error);
946 bio->bi_error = io_error;
952 static void disable_write_same(struct mapped_device *md)
954 struct queue_limits *limits = dm_get_queue_limits(md);
956 /* device doesn't really support WRITE SAME, disable it */
957 limits->max_write_same_sectors = 0;
960 static void clone_endio(struct bio *bio)
962 int error = bio->bi_error;
964 struct dm_target_io *tio = container_of(bio, struct dm_target_io, clone);
965 struct dm_io *io = tio->io;
966 struct mapped_device *md = tio->io->md;
967 dm_endio_fn endio = tio->ti->type->end_io;
970 r = endio(tio->ti, bio, error);
971 if (r < 0 || r == DM_ENDIO_REQUEUE)
973 * error and requeue request are handled
977 else if (r == DM_ENDIO_INCOMPLETE)
978 /* The target will handle the io */
981 DMWARN("unimplemented target endio return value: %d", r);
986 if (unlikely(r == -EREMOTEIO && (bio->bi_rw & REQ_WRITE_SAME) &&
987 !bdev_get_queue(bio->bi_bdev)->limits.max_write_same_sectors))
988 disable_write_same(md);
991 dec_pending(io, error);
995 * Partial completion handling for request-based dm
997 static void end_clone_bio(struct bio *clone)
999 struct dm_rq_clone_bio_info *info =
1000 container_of(clone, struct dm_rq_clone_bio_info, clone);
1001 struct dm_rq_target_io *tio = info->tio;
1002 struct bio *bio = info->orig;
1003 unsigned int nr_bytes = info->orig->bi_iter.bi_size;
1004 int error = clone->bi_error;
1010 * An error has already been detected on the request.
1011 * Once error occurred, just let clone->end_io() handle
1017 * Don't notice the error to the upper layer yet.
1018 * The error handling decision is made by the target driver,
1019 * when the request is completed.
1026 * I/O for the bio successfully completed.
1027 * Notice the data completion to the upper layer.
1031 * bios are processed from the head of the list.
1032 * So the completing bio should always be rq->bio.
1033 * If it's not, something wrong is happening.
1035 if (tio->orig->bio != bio)
1036 DMERR("bio completion is going in the middle of the request");
1039 * Update the original request.
1040 * Do not use blk_end_request() here, because it may complete
1041 * the original request before the clone, and break the ordering.
1043 blk_update_request(tio->orig, 0, nr_bytes);
1046 static struct dm_rq_target_io *tio_from_request(struct request *rq)
1048 return (rq->q->mq_ops ? blk_mq_rq_to_pdu(rq) : rq->special);
1051 static void rq_end_stats(struct mapped_device *md, struct request *orig)
1053 if (unlikely(dm_stats_used(&md->stats))) {
1054 struct dm_rq_target_io *tio = tio_from_request(orig);
1055 tio->duration_jiffies = jiffies - tio->duration_jiffies;
1056 dm_stats_account_io(&md->stats, orig->cmd_flags, blk_rq_pos(orig),
1057 tio->n_sectors, true, tio->duration_jiffies,
1063 * Don't touch any member of the md after calling this function because
1064 * the md may be freed in dm_put() at the end of this function.
1065 * Or do dm_get() before calling this function and dm_put() later.
1067 static void rq_completed(struct mapped_device *md, int rw, bool run_queue)
1069 atomic_dec(&md->pending[rw]);
1071 /* nudge anyone waiting on suspend queue */
1072 if (!md_in_flight(md))
1076 * Run this off this callpath, as drivers could invoke end_io while
1077 * inside their request_fn (and holding the queue lock). Calling
1078 * back into ->request_fn() could deadlock attempting to grab the
1082 if (md->queue->mq_ops)
1083 blk_mq_run_hw_queues(md->queue, true);
1085 blk_run_queue_async(md->queue);
1089 * dm_put() must be at the end of this function. See the comment above
1094 static void free_rq_clone(struct request *clone)
1096 struct dm_rq_target_io *tio = clone->end_io_data;
1097 struct mapped_device *md = tio->md;
1099 blk_rq_unprep_clone(clone);
1101 if (md->type == DM_TYPE_MQ_REQUEST_BASED)
1102 /* stacked on blk-mq queue(s) */
1103 tio->ti->type->release_clone_rq(clone);
1104 else if (!md->queue->mq_ops)
1105 /* request_fn queue stacked on request_fn queue(s) */
1106 free_clone_request(md, clone);
1108 * NOTE: for the blk-mq queue stacked on request_fn queue(s) case:
1109 * no need to call free_clone_request() because we leverage blk-mq by
1110 * allocating the clone at the end of the blk-mq pdu (see: clone_rq)
1113 if (!md->queue->mq_ops)
1118 * Complete the clone and the original request.
1119 * Must be called without clone's queue lock held,
1120 * see end_clone_request() for more details.
1122 static void dm_end_request(struct request *clone, int error)
1124 int rw = rq_data_dir(clone);
1125 struct dm_rq_target_io *tio = clone->end_io_data;
1126 struct mapped_device *md = tio->md;
1127 struct request *rq = tio->orig;
1129 if (rq->cmd_type == REQ_TYPE_BLOCK_PC) {
1130 rq->errors = clone->errors;
1131 rq->resid_len = clone->resid_len;
1135 * We are using the sense buffer of the original
1137 * So setting the length of the sense data is enough.
1139 rq->sense_len = clone->sense_len;
1142 free_rq_clone(clone);
1143 rq_end_stats(md, rq);
1145 blk_end_request_all(rq, error);
1147 blk_mq_end_request(rq, error);
1148 rq_completed(md, rw, true);
1151 static void dm_unprep_request(struct request *rq)
1153 struct dm_rq_target_io *tio = tio_from_request(rq);
1154 struct request *clone = tio->clone;
1156 if (!rq->q->mq_ops) {
1158 rq->cmd_flags &= ~REQ_DONTPREP;
1162 free_rq_clone(clone);
1166 * Requeue the original request of a clone.
1168 static void old_requeue_request(struct request *rq)
1170 struct request_queue *q = rq->q;
1171 unsigned long flags;
1173 spin_lock_irqsave(q->queue_lock, flags);
1174 blk_requeue_request(q, rq);
1175 blk_run_queue_async(q);
1176 spin_unlock_irqrestore(q->queue_lock, flags);
1179 static void dm_requeue_original_request(struct mapped_device *md,
1182 int rw = rq_data_dir(rq);
1184 dm_unprep_request(rq);
1186 rq_end_stats(md, rq);
1188 old_requeue_request(rq);
1190 blk_mq_requeue_request(rq);
1191 blk_mq_kick_requeue_list(rq->q);
1194 rq_completed(md, rw, false);
1197 static void old_stop_queue(struct request_queue *q)
1199 unsigned long flags;
1201 if (blk_queue_stopped(q))
1204 spin_lock_irqsave(q->queue_lock, flags);
1206 spin_unlock_irqrestore(q->queue_lock, flags);
1209 static void stop_queue(struct request_queue *q)
1214 blk_mq_stop_hw_queues(q);
1217 static void old_start_queue(struct request_queue *q)
1219 unsigned long flags;
1221 spin_lock_irqsave(q->queue_lock, flags);
1222 if (blk_queue_stopped(q))
1224 spin_unlock_irqrestore(q->queue_lock, flags);
1227 static void start_queue(struct request_queue *q)
1232 blk_mq_start_stopped_hw_queues(q, true);
1235 static void dm_done(struct request *clone, int error, bool mapped)
1238 struct dm_rq_target_io *tio = clone->end_io_data;
1239 dm_request_endio_fn rq_end_io = NULL;
1242 rq_end_io = tio->ti->type->rq_end_io;
1244 if (mapped && rq_end_io)
1245 r = rq_end_io(tio->ti, clone, error, &tio->info);
1248 if (unlikely(r == -EREMOTEIO && (clone->cmd_flags & REQ_WRITE_SAME) &&
1249 !clone->q->limits.max_write_same_sectors))
1250 disable_write_same(tio->md);
1253 /* The target wants to complete the I/O */
1254 dm_end_request(clone, r);
1255 else if (r == DM_ENDIO_INCOMPLETE)
1256 /* The target will handle the I/O */
1258 else if (r == DM_ENDIO_REQUEUE)
1259 /* The target wants to requeue the I/O */
1260 dm_requeue_original_request(tio->md, tio->orig);
1262 DMWARN("unimplemented target endio return value: %d", r);
1268 * Request completion handler for request-based dm
1270 static void dm_softirq_done(struct request *rq)
1273 struct dm_rq_target_io *tio = tio_from_request(rq);
1274 struct request *clone = tio->clone;
1278 rq_end_stats(tio->md, rq);
1279 rw = rq_data_dir(rq);
1280 if (!rq->q->mq_ops) {
1281 blk_end_request_all(rq, tio->error);
1282 rq_completed(tio->md, rw, false);
1285 blk_mq_end_request(rq, tio->error);
1286 rq_completed(tio->md, rw, false);
1291 if (rq->cmd_flags & REQ_FAILED)
1294 dm_done(clone, tio->error, mapped);
1298 * Complete the clone and the original request with the error status
1299 * through softirq context.
1301 static void dm_complete_request(struct request *rq, int error)
1303 struct dm_rq_target_io *tio = tio_from_request(rq);
1306 blk_complete_request(rq);
1310 * Complete the not-mapped clone and the original request with the error status
1311 * through softirq context.
1312 * Target's rq_end_io() function isn't called.
1313 * This may be used when the target's map_rq() or clone_and_map_rq() functions fail.
1315 static void dm_kill_unmapped_request(struct request *rq, int error)
1317 rq->cmd_flags |= REQ_FAILED;
1318 dm_complete_request(rq, error);
1322 * Called with the clone's queue lock held (for non-blk-mq)
1324 static void end_clone_request(struct request *clone, int error)
1326 struct dm_rq_target_io *tio = clone->end_io_data;
1328 if (!clone->q->mq_ops) {
1330 * For just cleaning up the information of the queue in which
1331 * the clone was dispatched.
1332 * The clone is *NOT* freed actually here because it is alloced
1333 * from dm own mempool (REQ_ALLOCED isn't set).
1335 __blk_put_request(clone->q, clone);
1339 * Actual request completion is done in a softirq context which doesn't
1340 * hold the clone's queue lock. Otherwise, deadlock could occur because:
1341 * - another request may be submitted by the upper level driver
1342 * of the stacking during the completion
1343 * - the submission which requires queue lock may be done
1344 * against this clone's queue
1346 dm_complete_request(tio->orig, error);
1350 * Return maximum size of I/O possible at the supplied sector up to the current
1353 static sector_t max_io_len_target_boundary(sector_t sector, struct dm_target *ti)
1355 sector_t target_offset = dm_target_offset(ti, sector);
1357 return ti->len - target_offset;
1360 static sector_t max_io_len(sector_t sector, struct dm_target *ti)
1362 sector_t len = max_io_len_target_boundary(sector, ti);
1363 sector_t offset, max_len;
1366 * Does the target need to split even further?
1368 if (ti->max_io_len) {
1369 offset = dm_target_offset(ti, sector);
1370 if (unlikely(ti->max_io_len & (ti->max_io_len - 1)))
1371 max_len = sector_div(offset, ti->max_io_len);
1373 max_len = offset & (ti->max_io_len - 1);
1374 max_len = ti->max_io_len - max_len;
1383 int dm_set_target_max_io_len(struct dm_target *ti, sector_t len)
1385 if (len > UINT_MAX) {
1386 DMERR("Specified maximum size of target IO (%llu) exceeds limit (%u)",
1387 (unsigned long long)len, UINT_MAX);
1388 ti->error = "Maximum size of target IO is too large";
1392 ti->max_io_len = (uint32_t) len;
1396 EXPORT_SYMBOL_GPL(dm_set_target_max_io_len);
1399 * A target may call dm_accept_partial_bio only from the map routine. It is
1400 * allowed for all bio types except REQ_FLUSH.
1402 * dm_accept_partial_bio informs the dm that the target only wants to process
1403 * additional n_sectors sectors of the bio and the rest of the data should be
1404 * sent in a next bio.
1406 * A diagram that explains the arithmetics:
1407 * +--------------------+---------------+-------+
1409 * +--------------------+---------------+-------+
1411 * <-------------- *tio->len_ptr --------------->
1412 * <------- bi_size ------->
1415 * Region 1 was already iterated over with bio_advance or similar function.
1416 * (it may be empty if the target doesn't use bio_advance)
1417 * Region 2 is the remaining bio size that the target wants to process.
1418 * (it may be empty if region 1 is non-empty, although there is no reason
1420 * The target requires that region 3 is to be sent in the next bio.
1422 * If the target wants to receive multiple copies of the bio (via num_*bios, etc),
1423 * the partially processed part (the sum of regions 1+2) must be the same for all
1424 * copies of the bio.
1426 void dm_accept_partial_bio(struct bio *bio, unsigned n_sectors)
1428 struct dm_target_io *tio = container_of(bio, struct dm_target_io, clone);
1429 unsigned bi_size = bio->bi_iter.bi_size >> SECTOR_SHIFT;
1430 BUG_ON(bio->bi_rw & REQ_FLUSH);
1431 BUG_ON(bi_size > *tio->len_ptr);
1432 BUG_ON(n_sectors > bi_size);
1433 *tio->len_ptr -= bi_size - n_sectors;
1434 bio->bi_iter.bi_size = n_sectors << SECTOR_SHIFT;
1436 EXPORT_SYMBOL_GPL(dm_accept_partial_bio);
1438 static void __map_bio(struct dm_target_io *tio)
1442 struct mapped_device *md;
1443 struct bio *clone = &tio->clone;
1444 struct dm_target *ti = tio->ti;
1446 clone->bi_end_io = clone_endio;
1449 * Map the clone. If r == 0 we don't need to do
1450 * anything, the target has assumed ownership of
1453 atomic_inc(&tio->io->io_count);
1454 sector = clone->bi_iter.bi_sector;
1455 r = ti->type->map(ti, clone);
1456 if (r == DM_MAPIO_REMAPPED) {
1457 /* the bio has been remapped so dispatch it */
1459 trace_block_bio_remap(bdev_get_queue(clone->bi_bdev), clone,
1460 tio->io->bio->bi_bdev->bd_dev, sector);
1462 generic_make_request(clone);
1463 } else if (r < 0 || r == DM_MAPIO_REQUEUE) {
1464 /* error the io and bail out, or requeue it if needed */
1466 dec_pending(tio->io, r);
1468 } else if (r != DM_MAPIO_SUBMITTED) {
1469 DMWARN("unimplemented target map return value: %d", r);
1475 struct mapped_device *md;
1476 struct dm_table *map;
1480 unsigned sector_count;
1483 static void bio_setup_sector(struct bio *bio, sector_t sector, unsigned len)
1485 bio->bi_iter.bi_sector = sector;
1486 bio->bi_iter.bi_size = to_bytes(len);
1490 * Creates a bio that consists of range of complete bvecs.
1492 static void clone_bio(struct dm_target_io *tio, struct bio *bio,
1493 sector_t sector, unsigned len)
1495 struct bio *clone = &tio->clone;
1497 __bio_clone_fast(clone, bio);
1499 if (bio_integrity(bio))
1500 bio_integrity_clone(clone, bio, GFP_NOIO);
1502 bio_advance(clone, to_bytes(sector - clone->bi_iter.bi_sector));
1503 clone->bi_iter.bi_size = to_bytes(len);
1505 if (bio_integrity(bio))
1506 bio_integrity_trim(clone, 0, len);
1509 static struct dm_target_io *alloc_tio(struct clone_info *ci,
1510 struct dm_target *ti,
1511 unsigned target_bio_nr)
1513 struct dm_target_io *tio;
1516 clone = bio_alloc_bioset(GFP_NOIO, 0, ci->md->bs);
1517 tio = container_of(clone, struct dm_target_io, clone);
1521 tio->target_bio_nr = target_bio_nr;
1526 static void __clone_and_map_simple_bio(struct clone_info *ci,
1527 struct dm_target *ti,
1528 unsigned target_bio_nr, unsigned *len)
1530 struct dm_target_io *tio = alloc_tio(ci, ti, target_bio_nr);
1531 struct bio *clone = &tio->clone;
1535 __bio_clone_fast(clone, ci->bio);
1537 bio_setup_sector(clone, ci->sector, *len);
1542 static void __send_duplicate_bios(struct clone_info *ci, struct dm_target *ti,
1543 unsigned num_bios, unsigned *len)
1545 unsigned target_bio_nr;
1547 for (target_bio_nr = 0; target_bio_nr < num_bios; target_bio_nr++)
1548 __clone_and_map_simple_bio(ci, ti, target_bio_nr, len);
1551 static int __send_empty_flush(struct clone_info *ci)
1553 unsigned target_nr = 0;
1554 struct dm_target *ti;
1556 BUG_ON(bio_has_data(ci->bio));
1557 while ((ti = dm_table_get_target(ci->map, target_nr++)))
1558 __send_duplicate_bios(ci, ti, ti->num_flush_bios, NULL);
1563 static void __clone_and_map_data_bio(struct clone_info *ci, struct dm_target *ti,
1564 sector_t sector, unsigned *len)
1566 struct bio *bio = ci->bio;
1567 struct dm_target_io *tio;
1568 unsigned target_bio_nr;
1569 unsigned num_target_bios = 1;
1572 * Does the target want to receive duplicate copies of the bio?
1574 if (bio_data_dir(bio) == WRITE && ti->num_write_bios)
1575 num_target_bios = ti->num_write_bios(ti, bio);
1577 for (target_bio_nr = 0; target_bio_nr < num_target_bios; target_bio_nr++) {
1578 tio = alloc_tio(ci, ti, target_bio_nr);
1580 clone_bio(tio, bio, sector, *len);
1585 typedef unsigned (*get_num_bios_fn)(struct dm_target *ti);
1587 static unsigned get_num_discard_bios(struct dm_target *ti)
1589 return ti->num_discard_bios;
1592 static unsigned get_num_write_same_bios(struct dm_target *ti)
1594 return ti->num_write_same_bios;
1597 typedef bool (*is_split_required_fn)(struct dm_target *ti);
1599 static bool is_split_required_for_discard(struct dm_target *ti)
1601 return ti->split_discard_bios;
1604 static int __send_changing_extent_only(struct clone_info *ci,
1605 get_num_bios_fn get_num_bios,
1606 is_split_required_fn is_split_required)
1608 struct dm_target *ti;
1613 ti = dm_table_find_target(ci->map, ci->sector);
1614 if (!dm_target_is_valid(ti))
1618 * Even though the device advertised support for this type of
1619 * request, that does not mean every target supports it, and
1620 * reconfiguration might also have changed that since the
1621 * check was performed.
1623 num_bios = get_num_bios ? get_num_bios(ti) : 0;
1627 if (is_split_required && !is_split_required(ti))
1628 len = min((sector_t)ci->sector_count, max_io_len_target_boundary(ci->sector, ti));
1630 len = min((sector_t)ci->sector_count, max_io_len(ci->sector, ti));
1632 __send_duplicate_bios(ci, ti, num_bios, &len);
1635 } while (ci->sector_count -= len);
1640 static int __send_discard(struct clone_info *ci)
1642 return __send_changing_extent_only(ci, get_num_discard_bios,
1643 is_split_required_for_discard);
1646 static int __send_write_same(struct clone_info *ci)
1648 return __send_changing_extent_only(ci, get_num_write_same_bios, NULL);
1652 * Select the correct strategy for processing a non-flush bio.
1654 static int __split_and_process_non_flush(struct clone_info *ci)
1656 struct bio *bio = ci->bio;
1657 struct dm_target *ti;
1660 if (unlikely(bio->bi_rw & REQ_DISCARD))
1661 return __send_discard(ci);
1662 else if (unlikely(bio->bi_rw & REQ_WRITE_SAME))
1663 return __send_write_same(ci);
1665 ti = dm_table_find_target(ci->map, ci->sector);
1666 if (!dm_target_is_valid(ti))
1669 len = min_t(sector_t, max_io_len(ci->sector, ti), ci->sector_count);
1671 __clone_and_map_data_bio(ci, ti, ci->sector, &len);
1674 ci->sector_count -= len;
1680 * Entry point to split a bio into clones and submit them to the targets.
1682 static void __split_and_process_bio(struct mapped_device *md,
1683 struct dm_table *map, struct bio *bio)
1685 struct clone_info ci;
1688 if (unlikely(!map)) {
1695 ci.io = alloc_io(md);
1697 atomic_set(&ci.io->io_count, 1);
1700 spin_lock_init(&ci.io->endio_lock);
1701 ci.sector = bio->bi_iter.bi_sector;
1703 start_io_acct(ci.io);
1705 if (bio->bi_rw & REQ_FLUSH) {
1706 ci.bio = &ci.md->flush_bio;
1707 ci.sector_count = 0;
1708 error = __send_empty_flush(&ci);
1709 /* dec_pending submits any data associated with flush */
1712 ci.sector_count = bio_sectors(bio);
1713 while (ci.sector_count && !error)
1714 error = __split_and_process_non_flush(&ci);
1717 /* drop the extra reference count */
1718 dec_pending(ci.io, error);
1720 /*-----------------------------------------------------------------
1722 *---------------------------------------------------------------*/
1725 * The request function that just remaps the bio built up by
1728 static void dm_make_request(struct request_queue *q, struct bio *bio)
1730 int rw = bio_data_dir(bio);
1731 struct mapped_device *md = q->queuedata;
1733 struct dm_table *map;
1735 map = dm_get_live_table(md, &srcu_idx);
1737 blk_queue_split(q, &bio, q->bio_split);
1739 generic_start_io_acct(rw, bio_sectors(bio), &dm_disk(md)->part0);
1741 /* if we're suspended, we have to queue this io for later */
1742 if (unlikely(test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags))) {
1743 dm_put_live_table(md, srcu_idx);
1745 if (bio_rw(bio) != READA)
1752 __split_and_process_bio(md, map, bio);
1753 dm_put_live_table(md, srcu_idx);
1757 int dm_request_based(struct mapped_device *md)
1759 return blk_queue_stackable(md->queue);
1762 static void dm_dispatch_clone_request(struct request *clone, struct request *rq)
1766 if (blk_queue_io_stat(clone->q))
1767 clone->cmd_flags |= REQ_IO_STAT;
1769 clone->start_time = jiffies;
1770 r = blk_insert_cloned_request(clone->q, clone);
1772 /* must complete clone in terms of original request */
1773 dm_complete_request(rq, r);
1776 static int dm_rq_bio_constructor(struct bio *bio, struct bio *bio_orig,
1779 struct dm_rq_target_io *tio = data;
1780 struct dm_rq_clone_bio_info *info =
1781 container_of(bio, struct dm_rq_clone_bio_info, clone);
1783 info->orig = bio_orig;
1785 bio->bi_end_io = end_clone_bio;
1790 static int setup_clone(struct request *clone, struct request *rq,
1791 struct dm_rq_target_io *tio, gfp_t gfp_mask)
1795 r = blk_rq_prep_clone(clone, rq, tio->md->bs, gfp_mask,
1796 dm_rq_bio_constructor, tio);
1800 clone->cmd = rq->cmd;
1801 clone->cmd_len = rq->cmd_len;
1802 clone->sense = rq->sense;
1803 clone->end_io = end_clone_request;
1804 clone->end_io_data = tio;
1811 static struct request *clone_rq(struct request *rq, struct mapped_device *md,
1812 struct dm_rq_target_io *tio, gfp_t gfp_mask)
1815 * Do not allocate a clone if tio->clone was already set
1816 * (see: dm_mq_queue_rq).
1818 bool alloc_clone = !tio->clone;
1819 struct request *clone;
1822 clone = alloc_clone_request(md, gfp_mask);
1828 blk_rq_init(NULL, clone);
1829 if (setup_clone(clone, rq, tio, gfp_mask)) {
1832 free_clone_request(md, clone);
1839 static void map_tio_request(struct kthread_work *work);
1841 static void init_tio(struct dm_rq_target_io *tio, struct request *rq,
1842 struct mapped_device *md)
1849 memset(&tio->info, 0, sizeof(tio->info));
1850 if (md->kworker_task)
1851 init_kthread_work(&tio->work, map_tio_request);
1854 static struct dm_rq_target_io *prep_tio(struct request *rq,
1855 struct mapped_device *md, gfp_t gfp_mask)
1857 struct dm_rq_target_io *tio;
1859 struct dm_table *table;
1861 tio = alloc_rq_tio(md, gfp_mask);
1865 init_tio(tio, rq, md);
1867 table = dm_get_live_table(md, &srcu_idx);
1868 if (!dm_table_mq_request_based(table)) {
1869 if (!clone_rq(rq, md, tio, gfp_mask)) {
1870 dm_put_live_table(md, srcu_idx);
1875 dm_put_live_table(md, srcu_idx);
1881 * Called with the queue lock held.
1883 static int dm_prep_fn(struct request_queue *q, struct request *rq)
1885 struct mapped_device *md = q->queuedata;
1886 struct dm_rq_target_io *tio;
1888 if (unlikely(rq->special)) {
1889 DMWARN("Already has something in rq->special.");
1890 return BLKPREP_KILL;
1893 tio = prep_tio(rq, md, GFP_ATOMIC);
1895 return BLKPREP_DEFER;
1898 rq->cmd_flags |= REQ_DONTPREP;
1905 * 0 : the request has been processed
1906 * DM_MAPIO_REQUEUE : the original request needs to be requeued
1907 * < 0 : the request was completed due to failure
1909 static int map_request(struct dm_rq_target_io *tio, struct request *rq,
1910 struct mapped_device *md)
1913 struct dm_target *ti = tio->ti;
1914 struct request *clone = NULL;
1918 r = ti->type->map_rq(ti, clone, &tio->info);
1920 r = ti->type->clone_and_map_rq(ti, rq, &tio->info, &clone);
1922 /* The target wants to complete the I/O */
1923 dm_kill_unmapped_request(rq, r);
1926 if (r != DM_MAPIO_REMAPPED)
1928 if (setup_clone(clone, rq, tio, GFP_ATOMIC)) {
1930 ti->type->release_clone_rq(clone);
1931 return DM_MAPIO_REQUEUE;
1936 case DM_MAPIO_SUBMITTED:
1937 /* The target has taken the I/O to submit by itself later */
1939 case DM_MAPIO_REMAPPED:
1940 /* The target has remapped the I/O so dispatch it */
1941 trace_block_rq_remap(clone->q, clone, disk_devt(dm_disk(md)),
1943 dm_dispatch_clone_request(clone, rq);
1945 case DM_MAPIO_REQUEUE:
1946 /* The target wants to requeue the I/O */
1947 dm_requeue_original_request(md, tio->orig);
1951 DMWARN("unimplemented target map return value: %d", r);
1955 /* The target wants to complete the I/O */
1956 dm_kill_unmapped_request(rq, r);
1963 static void map_tio_request(struct kthread_work *work)
1965 struct dm_rq_target_io *tio = container_of(work, struct dm_rq_target_io, work);
1966 struct request *rq = tio->orig;
1967 struct mapped_device *md = tio->md;
1969 if (map_request(tio, rq, md) == DM_MAPIO_REQUEUE)
1970 dm_requeue_original_request(md, rq);
1973 static void dm_start_request(struct mapped_device *md, struct request *orig)
1975 if (!orig->q->mq_ops)
1976 blk_start_request(orig);
1978 blk_mq_start_request(orig);
1979 atomic_inc(&md->pending[rq_data_dir(orig)]);
1981 if (md->seq_rq_merge_deadline_usecs) {
1982 md->last_rq_pos = rq_end_sector(orig);
1983 md->last_rq_rw = rq_data_dir(orig);
1984 md->last_rq_start_time = ktime_get();
1987 if (unlikely(dm_stats_used(&md->stats))) {
1988 struct dm_rq_target_io *tio = tio_from_request(orig);
1989 tio->duration_jiffies = jiffies;
1990 tio->n_sectors = blk_rq_sectors(orig);
1991 dm_stats_account_io(&md->stats, orig->cmd_flags, blk_rq_pos(orig),
1992 tio->n_sectors, false, 0, &tio->stats_aux);
1996 * Hold the md reference here for the in-flight I/O.
1997 * We can't rely on the reference count by device opener,
1998 * because the device may be closed during the request completion
1999 * when all bios are completed.
2000 * See the comment in rq_completed() too.
2005 #define MAX_SEQ_RQ_MERGE_DEADLINE_USECS 100000
2007 ssize_t dm_attr_rq_based_seq_io_merge_deadline_show(struct mapped_device *md, char *buf)
2009 return sprintf(buf, "%u\n", md->seq_rq_merge_deadline_usecs);
2012 ssize_t dm_attr_rq_based_seq_io_merge_deadline_store(struct mapped_device *md,
2013 const char *buf, size_t count)
2017 if (!dm_request_based(md) || md->use_blk_mq)
2020 if (kstrtouint(buf, 10, &deadline))
2023 if (deadline > MAX_SEQ_RQ_MERGE_DEADLINE_USECS)
2024 deadline = MAX_SEQ_RQ_MERGE_DEADLINE_USECS;
2026 md->seq_rq_merge_deadline_usecs = deadline;
2031 static bool dm_request_peeked_before_merge_deadline(struct mapped_device *md)
2033 ktime_t kt_deadline;
2035 if (!md->seq_rq_merge_deadline_usecs)
2038 kt_deadline = ns_to_ktime((u64)md->seq_rq_merge_deadline_usecs * NSEC_PER_USEC);
2039 kt_deadline = ktime_add_safe(md->last_rq_start_time, kt_deadline);
2041 return !ktime_after(ktime_get(), kt_deadline);
2045 * q->request_fn for request-based dm.
2046 * Called with the queue lock held.
2048 static void dm_request_fn(struct request_queue *q)
2050 struct mapped_device *md = q->queuedata;
2052 struct dm_table *map = dm_get_live_table(md, &srcu_idx);
2053 struct dm_target *ti;
2055 struct dm_rq_target_io *tio;
2059 * For suspend, check blk_queue_stopped() and increment
2060 * ->pending within a single queue_lock not to increment the
2061 * number of in-flight I/Os after the queue is stopped in
2064 while (!blk_queue_stopped(q)) {
2065 rq = blk_peek_request(q);
2069 /* always use block 0 to find the target for flushes for now */
2071 if (!(rq->cmd_flags & REQ_FLUSH))
2072 pos = blk_rq_pos(rq);
2074 ti = dm_table_find_target(map, pos);
2075 if (!dm_target_is_valid(ti)) {
2077 * Must perform setup, that rq_completed() requires,
2078 * before calling dm_kill_unmapped_request
2080 DMERR_LIMIT("request attempted access beyond the end of device");
2081 dm_start_request(md, rq);
2082 dm_kill_unmapped_request(rq, -EIO);
2086 if (dm_request_peeked_before_merge_deadline(md) &&
2087 md_in_flight(md) && rq->bio && rq->bio->bi_vcnt == 1 &&
2088 md->last_rq_pos == pos && md->last_rq_rw == rq_data_dir(rq))
2091 if (ti->type->busy && ti->type->busy(ti))
2094 dm_start_request(md, rq);
2096 tio = tio_from_request(rq);
2097 /* Establish tio->ti before queuing work (map_tio_request) */
2099 queue_kthread_work(&md->kworker, &tio->work);
2100 BUG_ON(!irqs_disabled());
2106 blk_delay_queue(q, HZ / 100);
2108 dm_put_live_table(md, srcu_idx);
2111 static int dm_any_congested(void *congested_data, int bdi_bits)
2114 struct mapped_device *md = congested_data;
2115 struct dm_table *map;
2117 if (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
2118 map = dm_get_live_table_fast(md);
2121 * Request-based dm cares about only own queue for
2122 * the query about congestion status of request_queue
2124 if (dm_request_based(md))
2125 r = md->queue->backing_dev_info.wb.state &
2128 r = dm_table_any_congested(map, bdi_bits);
2130 dm_put_live_table_fast(md);
2136 /*-----------------------------------------------------------------
2137 * An IDR is used to keep track of allocated minor numbers.
2138 *---------------------------------------------------------------*/
2139 static void free_minor(int minor)
2141 spin_lock(&_minor_lock);
2142 idr_remove(&_minor_idr, minor);
2143 spin_unlock(&_minor_lock);
2147 * See if the device with a specific minor # is free.
2149 static int specific_minor(int minor)
2153 if (minor >= (1 << MINORBITS))
2156 idr_preload(GFP_KERNEL);
2157 spin_lock(&_minor_lock);
2159 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, minor, minor + 1, GFP_NOWAIT);
2161 spin_unlock(&_minor_lock);
2164 return r == -ENOSPC ? -EBUSY : r;
2168 static int next_free_minor(int *minor)
2172 idr_preload(GFP_KERNEL);
2173 spin_lock(&_minor_lock);
2175 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, 0, 1 << MINORBITS, GFP_NOWAIT);
2177 spin_unlock(&_minor_lock);
2185 static const struct block_device_operations dm_blk_dops;
2187 static void dm_wq_work(struct work_struct *work);
2189 static void dm_init_md_queue(struct mapped_device *md)
2192 * Request-based dm devices cannot be stacked on top of bio-based dm
2193 * devices. The type of this dm device may not have been decided yet.
2194 * The type is decided at the first table loading time.
2195 * To prevent problematic device stacking, clear the queue flag
2196 * for request stacking support until then.
2198 * This queue is new, so no concurrency on the queue_flags.
2200 queue_flag_clear_unlocked(QUEUE_FLAG_STACKABLE, md->queue);
2203 static void dm_init_old_md_queue(struct mapped_device *md)
2205 md->use_blk_mq = false;
2206 dm_init_md_queue(md);
2209 * Initialize aspects of queue that aren't relevant for blk-mq
2211 md->queue->queuedata = md;
2212 md->queue->backing_dev_info.congested_fn = dm_any_congested;
2213 md->queue->backing_dev_info.congested_data = md;
2215 blk_queue_bounce_limit(md->queue, BLK_BOUNCE_ANY);
2218 static void cleanup_mapped_device(struct mapped_device *md)
2221 destroy_workqueue(md->wq);
2222 if (md->kworker_task)
2223 kthread_stop(md->kworker_task);
2225 mempool_destroy(md->io_pool);
2227 mempool_destroy(md->rq_pool);
2229 bioset_free(md->bs);
2231 cleanup_srcu_struct(&md->io_barrier);
2234 spin_lock(&_minor_lock);
2235 md->disk->private_data = NULL;
2236 spin_unlock(&_minor_lock);
2237 if (blk_get_integrity(md->disk))
2238 blk_integrity_unregister(md->disk);
2239 del_gendisk(md->disk);
2244 blk_cleanup_queue(md->queue);
2253 * Allocate and initialise a blank device with a given minor.
2255 static struct mapped_device *alloc_dev(int minor)
2258 struct mapped_device *md = kzalloc(sizeof(*md), GFP_KERNEL);
2262 DMWARN("unable to allocate device, out of memory.");
2266 if (!try_module_get(THIS_MODULE))
2267 goto bad_module_get;
2269 /* get a minor number for the dev */
2270 if (minor == DM_ANY_MINOR)
2271 r = next_free_minor(&minor);
2273 r = specific_minor(minor);
2277 r = init_srcu_struct(&md->io_barrier);
2279 goto bad_io_barrier;
2281 md->use_blk_mq = use_blk_mq;
2282 md->type = DM_TYPE_NONE;
2283 mutex_init(&md->suspend_lock);
2284 mutex_init(&md->type_lock);
2285 mutex_init(&md->table_devices_lock);
2286 spin_lock_init(&md->deferred_lock);
2287 atomic_set(&md->holders, 1);
2288 atomic_set(&md->open_count, 0);
2289 atomic_set(&md->event_nr, 0);
2290 atomic_set(&md->uevent_seq, 0);
2291 INIT_LIST_HEAD(&md->uevent_list);
2292 INIT_LIST_HEAD(&md->table_devices);
2293 spin_lock_init(&md->uevent_lock);
2295 md->queue = blk_alloc_queue(GFP_KERNEL);
2299 dm_init_md_queue(md);
2301 md->disk = alloc_disk(1);
2305 atomic_set(&md->pending[0], 0);
2306 atomic_set(&md->pending[1], 0);
2307 init_waitqueue_head(&md->wait);
2308 INIT_WORK(&md->work, dm_wq_work);
2309 init_waitqueue_head(&md->eventq);
2310 init_completion(&md->kobj_holder.completion);
2311 md->kworker_task = NULL;
2313 md->disk->major = _major;
2314 md->disk->first_minor = minor;
2315 md->disk->fops = &dm_blk_dops;
2316 md->disk->queue = md->queue;
2317 md->disk->private_data = md;
2318 sprintf(md->disk->disk_name, "dm-%d", minor);
2320 format_dev_t(md->name, MKDEV(_major, minor));
2322 md->wq = alloc_workqueue("kdmflush", WQ_MEM_RECLAIM, 0);
2326 md->bdev = bdget_disk(md->disk, 0);
2330 bio_init(&md->flush_bio);
2331 md->flush_bio.bi_bdev = md->bdev;
2332 md->flush_bio.bi_rw = WRITE_FLUSH;
2334 dm_stats_init(&md->stats);
2336 /* Populate the mapping, nobody knows we exist yet */
2337 spin_lock(&_minor_lock);
2338 old_md = idr_replace(&_minor_idr, md, minor);
2339 spin_unlock(&_minor_lock);
2341 BUG_ON(old_md != MINOR_ALLOCED);
2346 cleanup_mapped_device(md);
2350 module_put(THIS_MODULE);
2356 static void unlock_fs(struct mapped_device *md);
2358 static void free_dev(struct mapped_device *md)
2360 int minor = MINOR(disk_devt(md->disk));
2364 cleanup_mapped_device(md);
2366 blk_mq_free_tag_set(&md->tag_set);
2368 free_table_devices(&md->table_devices);
2369 dm_stats_cleanup(&md->stats);
2372 module_put(THIS_MODULE);
2376 static void __bind_mempools(struct mapped_device *md, struct dm_table *t)
2378 struct dm_md_mempools *p = dm_table_get_md_mempools(t);
2381 /* The md already has necessary mempools. */
2382 if (dm_table_get_type(t) == DM_TYPE_BIO_BASED) {
2384 * Reload bioset because front_pad may have changed
2385 * because a different table was loaded.
2387 bioset_free(md->bs);
2392 * There's no need to reload with request-based dm
2393 * because the size of front_pad doesn't change.
2394 * Note for future: If you are to reload bioset,
2395 * prep-ed requests in the queue may refer
2396 * to bio from the old bioset, so you must walk
2397 * through the queue to unprep.
2402 BUG_ON(!p || md->io_pool || md->rq_pool || md->bs);
2404 md->io_pool = p->io_pool;
2406 md->rq_pool = p->rq_pool;
2412 /* mempool bind completed, no longer need any mempools in the table */
2413 dm_table_free_md_mempools(t);
2417 * Bind a table to the device.
2419 static void event_callback(void *context)
2421 unsigned long flags;
2423 struct mapped_device *md = (struct mapped_device *) context;
2425 spin_lock_irqsave(&md->uevent_lock, flags);
2426 list_splice_init(&md->uevent_list, &uevents);
2427 spin_unlock_irqrestore(&md->uevent_lock, flags);
2429 dm_send_uevents(&uevents, &disk_to_dev(md->disk)->kobj);
2431 atomic_inc(&md->event_nr);
2432 wake_up(&md->eventq);
2436 * Protected by md->suspend_lock obtained by dm_swap_table().
2438 static void __set_size(struct mapped_device *md, sector_t size)
2440 set_capacity(md->disk, size);
2442 i_size_write(md->bdev->bd_inode, (loff_t)size << SECTOR_SHIFT);
2446 * Returns old map, which caller must destroy.
2448 static struct dm_table *__bind(struct mapped_device *md, struct dm_table *t,
2449 struct queue_limits *limits)
2451 struct dm_table *old_map;
2452 struct request_queue *q = md->queue;
2455 size = dm_table_get_size(t);
2458 * Wipe any geometry if the size of the table changed.
2460 if (size != dm_get_size(md))
2461 memset(&md->geometry, 0, sizeof(md->geometry));
2463 __set_size(md, size);
2465 dm_table_event_callback(t, event_callback, md);
2468 * The queue hasn't been stopped yet, if the old table type wasn't
2469 * for request-based during suspension. So stop it to prevent
2470 * I/O mapping before resume.
2471 * This must be done before setting the queue restrictions,
2472 * because request-based dm may be run just after the setting.
2474 if (dm_table_request_based(t))
2477 __bind_mempools(md, t);
2479 old_map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2480 rcu_assign_pointer(md->map, t);
2481 md->immutable_target_type = dm_table_get_immutable_target_type(t);
2483 dm_table_set_restrictions(t, q, limits);
2491 * Returns unbound table for the caller to free.
2493 static struct dm_table *__unbind(struct mapped_device *md)
2495 struct dm_table *map = rcu_dereference_protected(md->map, 1);
2500 dm_table_event_callback(map, NULL, NULL);
2501 RCU_INIT_POINTER(md->map, NULL);
2508 * Constructor for a new device.
2510 int dm_create(int minor, struct mapped_device **result)
2512 struct mapped_device *md;
2514 md = alloc_dev(minor);
2525 * Functions to manage md->type.
2526 * All are required to hold md->type_lock.
2528 void dm_lock_md_type(struct mapped_device *md)
2530 mutex_lock(&md->type_lock);
2533 void dm_unlock_md_type(struct mapped_device *md)
2535 mutex_unlock(&md->type_lock);
2538 void dm_set_md_type(struct mapped_device *md, unsigned type)
2540 BUG_ON(!mutex_is_locked(&md->type_lock));
2544 unsigned dm_get_md_type(struct mapped_device *md)
2546 BUG_ON(!mutex_is_locked(&md->type_lock));
2550 struct target_type *dm_get_immutable_target_type(struct mapped_device *md)
2552 return md->immutable_target_type;
2556 * The queue_limits are only valid as long as you have a reference
2559 struct queue_limits *dm_get_queue_limits(struct mapped_device *md)
2561 BUG_ON(!atomic_read(&md->holders));
2562 return &md->queue->limits;
2564 EXPORT_SYMBOL_GPL(dm_get_queue_limits);
2566 static void init_rq_based_worker_thread(struct mapped_device *md)
2568 /* Initialize the request-based DM worker thread */
2569 init_kthread_worker(&md->kworker);
2570 md->kworker_task = kthread_run(kthread_worker_fn, &md->kworker,
2571 "kdmwork-%s", dm_device_name(md));
2575 * Fully initialize a request-based queue (->elevator, ->request_fn, etc).
2577 static int dm_init_request_based_queue(struct mapped_device *md)
2579 struct request_queue *q = NULL;
2581 /* Fully initialize the queue */
2582 q = blk_init_allocated_queue(md->queue, dm_request_fn, NULL);
2586 /* disable dm_request_fn's merge heuristic by default */
2587 md->seq_rq_merge_deadline_usecs = 0;
2590 dm_init_old_md_queue(md);
2591 blk_queue_softirq_done(md->queue, dm_softirq_done);
2592 blk_queue_prep_rq(md->queue, dm_prep_fn);
2594 init_rq_based_worker_thread(md);
2596 elv_register_queue(md->queue);
2601 static int dm_mq_init_request(void *data, struct request *rq,
2602 unsigned int hctx_idx, unsigned int request_idx,
2603 unsigned int numa_node)
2605 struct mapped_device *md = data;
2606 struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
2609 * Must initialize md member of tio, otherwise it won't
2610 * be available in dm_mq_queue_rq.
2617 static int dm_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
2618 const struct blk_mq_queue_data *bd)
2620 struct request *rq = bd->rq;
2621 struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
2622 struct mapped_device *md = tio->md;
2624 struct dm_table *map = dm_get_live_table(md, &srcu_idx);
2625 struct dm_target *ti;
2628 /* always use block 0 to find the target for flushes for now */
2630 if (!(rq->cmd_flags & REQ_FLUSH))
2631 pos = blk_rq_pos(rq);
2633 ti = dm_table_find_target(map, pos);
2634 if (!dm_target_is_valid(ti)) {
2635 dm_put_live_table(md, srcu_idx);
2636 DMERR_LIMIT("request attempted access beyond the end of device");
2638 * Must perform setup, that rq_completed() requires,
2639 * before returning BLK_MQ_RQ_QUEUE_ERROR
2641 dm_start_request(md, rq);
2642 return BLK_MQ_RQ_QUEUE_ERROR;
2644 dm_put_live_table(md, srcu_idx);
2646 if (ti->type->busy && ti->type->busy(ti))
2647 return BLK_MQ_RQ_QUEUE_BUSY;
2649 dm_start_request(md, rq);
2651 /* Init tio using md established in .init_request */
2652 init_tio(tio, rq, md);
2655 * Establish tio->ti before queuing work (map_tio_request)
2656 * or making direct call to map_request().
2660 /* Clone the request if underlying devices aren't blk-mq */
2661 if (dm_table_get_type(map) == DM_TYPE_REQUEST_BASED) {
2662 /* clone request is allocated at the end of the pdu */
2663 tio->clone = (void *)blk_mq_rq_to_pdu(rq) + sizeof(struct dm_rq_target_io);
2664 (void) clone_rq(rq, md, tio, GFP_ATOMIC);
2665 queue_kthread_work(&md->kworker, &tio->work);
2667 /* Direct call is fine since .queue_rq allows allocations */
2668 if (map_request(tio, rq, md) == DM_MAPIO_REQUEUE) {
2669 /* Undo dm_start_request() before requeuing */
2670 rq_end_stats(md, rq);
2671 rq_completed(md, rq_data_dir(rq), false);
2672 return BLK_MQ_RQ_QUEUE_BUSY;
2676 return BLK_MQ_RQ_QUEUE_OK;
2679 static struct blk_mq_ops dm_mq_ops = {
2680 .queue_rq = dm_mq_queue_rq,
2681 .map_queue = blk_mq_map_queue,
2682 .complete = dm_softirq_done,
2683 .init_request = dm_mq_init_request,
2686 static int dm_init_request_based_blk_mq_queue(struct mapped_device *md)
2688 unsigned md_type = dm_get_md_type(md);
2689 struct request_queue *q;
2692 memset(&md->tag_set, 0, sizeof(md->tag_set));
2693 md->tag_set.ops = &dm_mq_ops;
2694 md->tag_set.queue_depth = BLKDEV_MAX_RQ;
2695 md->tag_set.numa_node = NUMA_NO_NODE;
2696 md->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
2697 md->tag_set.nr_hw_queues = 1;
2698 if (md_type == DM_TYPE_REQUEST_BASED) {
2699 /* make the memory for non-blk-mq clone part of the pdu */
2700 md->tag_set.cmd_size = sizeof(struct dm_rq_target_io) + sizeof(struct request);
2702 md->tag_set.cmd_size = sizeof(struct dm_rq_target_io);
2703 md->tag_set.driver_data = md;
2705 err = blk_mq_alloc_tag_set(&md->tag_set);
2709 q = blk_mq_init_allocated_queue(&md->tag_set, md->queue);
2715 dm_init_md_queue(md);
2717 /* backfill 'mq' sysfs registration normally done in blk_register_queue */
2718 blk_mq_register_disk(md->disk);
2720 if (md_type == DM_TYPE_REQUEST_BASED)
2721 init_rq_based_worker_thread(md);
2726 blk_mq_free_tag_set(&md->tag_set);
2730 static unsigned filter_md_type(unsigned type, struct mapped_device *md)
2732 if (type == DM_TYPE_BIO_BASED)
2735 return !md->use_blk_mq ? DM_TYPE_REQUEST_BASED : DM_TYPE_MQ_REQUEST_BASED;
2739 * Setup the DM device's queue based on md's type
2741 int dm_setup_md_queue(struct mapped_device *md)
2744 unsigned md_type = filter_md_type(dm_get_md_type(md), md);
2747 case DM_TYPE_REQUEST_BASED:
2748 r = dm_init_request_based_queue(md);
2750 DMWARN("Cannot initialize queue for request-based mapped device");
2754 case DM_TYPE_MQ_REQUEST_BASED:
2755 r = dm_init_request_based_blk_mq_queue(md);
2757 DMWARN("Cannot initialize queue for request-based blk-mq mapped device");
2761 case DM_TYPE_BIO_BASED:
2762 dm_init_old_md_queue(md);
2763 blk_queue_make_request(md->queue, dm_make_request);
2770 struct mapped_device *dm_get_md(dev_t dev)
2772 struct mapped_device *md;
2773 unsigned minor = MINOR(dev);
2775 if (MAJOR(dev) != _major || minor >= (1 << MINORBITS))
2778 spin_lock(&_minor_lock);
2780 md = idr_find(&_minor_idr, minor);
2782 if ((md == MINOR_ALLOCED ||
2783 (MINOR(disk_devt(dm_disk(md))) != minor) ||
2784 dm_deleting_md(md) ||
2785 test_bit(DMF_FREEING, &md->flags))) {
2793 spin_unlock(&_minor_lock);
2797 EXPORT_SYMBOL_GPL(dm_get_md);
2799 void *dm_get_mdptr(struct mapped_device *md)
2801 return md->interface_ptr;
2804 void dm_set_mdptr(struct mapped_device *md, void *ptr)
2806 md->interface_ptr = ptr;
2809 void dm_get(struct mapped_device *md)
2811 atomic_inc(&md->holders);
2812 BUG_ON(test_bit(DMF_FREEING, &md->flags));
2815 int dm_hold(struct mapped_device *md)
2817 spin_lock(&_minor_lock);
2818 if (test_bit(DMF_FREEING, &md->flags)) {
2819 spin_unlock(&_minor_lock);
2823 spin_unlock(&_minor_lock);
2826 EXPORT_SYMBOL_GPL(dm_hold);
2828 const char *dm_device_name(struct mapped_device *md)
2832 EXPORT_SYMBOL_GPL(dm_device_name);
2834 static void __dm_destroy(struct mapped_device *md, bool wait)
2836 struct dm_table *map;
2841 spin_lock(&_minor_lock);
2842 idr_replace(&_minor_idr, MINOR_ALLOCED, MINOR(disk_devt(dm_disk(md))));
2843 set_bit(DMF_FREEING, &md->flags);
2844 spin_unlock(&_minor_lock);
2846 if (dm_request_based(md) && md->kworker_task)
2847 flush_kthread_worker(&md->kworker);
2850 * Take suspend_lock so that presuspend and postsuspend methods
2851 * do not race with internal suspend.
2853 mutex_lock(&md->suspend_lock);
2854 map = dm_get_live_table(md, &srcu_idx);
2855 if (!dm_suspended_md(md)) {
2856 dm_table_presuspend_targets(map);
2857 dm_table_postsuspend_targets(map);
2859 /* dm_put_live_table must be before msleep, otherwise deadlock is possible */
2860 dm_put_live_table(md, srcu_idx);
2861 mutex_unlock(&md->suspend_lock);
2864 * Rare, but there may be I/O requests still going to complete,
2865 * for example. Wait for all references to disappear.
2866 * No one should increment the reference count of the mapped_device,
2867 * after the mapped_device state becomes DMF_FREEING.
2870 while (atomic_read(&md->holders))
2872 else if (atomic_read(&md->holders))
2873 DMWARN("%s: Forcibly removing mapped_device still in use! (%d users)",
2874 dm_device_name(md), atomic_read(&md->holders));
2877 dm_table_destroy(__unbind(md));
2881 void dm_destroy(struct mapped_device *md)
2883 __dm_destroy(md, true);
2886 void dm_destroy_immediate(struct mapped_device *md)
2888 __dm_destroy(md, false);
2891 void dm_put(struct mapped_device *md)
2893 atomic_dec(&md->holders);
2895 EXPORT_SYMBOL_GPL(dm_put);
2897 static int dm_wait_for_completion(struct mapped_device *md, int interruptible)
2900 DECLARE_WAITQUEUE(wait, current);
2902 add_wait_queue(&md->wait, &wait);
2905 set_current_state(interruptible);
2907 if (!md_in_flight(md))
2910 if (interruptible == TASK_INTERRUPTIBLE &&
2911 signal_pending(current)) {
2918 set_current_state(TASK_RUNNING);
2920 remove_wait_queue(&md->wait, &wait);
2926 * Process the deferred bios
2928 static void dm_wq_work(struct work_struct *work)
2930 struct mapped_device *md = container_of(work, struct mapped_device,
2934 struct dm_table *map;
2936 map = dm_get_live_table(md, &srcu_idx);
2938 while (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
2939 spin_lock_irq(&md->deferred_lock);
2940 c = bio_list_pop(&md->deferred);
2941 spin_unlock_irq(&md->deferred_lock);
2946 if (dm_request_based(md))
2947 generic_make_request(c);
2949 __split_and_process_bio(md, map, c);
2952 dm_put_live_table(md, srcu_idx);
2955 static void dm_queue_flush(struct mapped_device *md)
2957 clear_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2958 smp_mb__after_atomic();
2959 queue_work(md->wq, &md->work);
2963 * Swap in a new table, returning the old one for the caller to destroy.
2965 struct dm_table *dm_swap_table(struct mapped_device *md, struct dm_table *table)
2967 struct dm_table *live_map = NULL, *map = ERR_PTR(-EINVAL);
2968 struct queue_limits limits;
2971 mutex_lock(&md->suspend_lock);
2973 /* device must be suspended */
2974 if (!dm_suspended_md(md))
2978 * If the new table has no data devices, retain the existing limits.
2979 * This helps multipath with queue_if_no_path if all paths disappear,
2980 * then new I/O is queued based on these limits, and then some paths
2983 if (dm_table_has_no_data_devices(table)) {
2984 live_map = dm_get_live_table_fast(md);
2986 limits = md->queue->limits;
2987 dm_put_live_table_fast(md);
2991 r = dm_calculate_queue_limits(table, &limits);
2998 map = __bind(md, table, &limits);
3001 mutex_unlock(&md->suspend_lock);
3006 * Functions to lock and unlock any filesystem running on the
3009 static int lock_fs(struct mapped_device *md)
3013 WARN_ON(md->frozen_sb);
3015 md->frozen_sb = freeze_bdev(md->bdev);
3016 if (IS_ERR(md->frozen_sb)) {
3017 r = PTR_ERR(md->frozen_sb);
3018 md->frozen_sb = NULL;
3022 set_bit(DMF_FROZEN, &md->flags);
3027 static void unlock_fs(struct mapped_device *md)
3029 if (!test_bit(DMF_FROZEN, &md->flags))
3032 thaw_bdev(md->bdev, md->frozen_sb);
3033 md->frozen_sb = NULL;
3034 clear_bit(DMF_FROZEN, &md->flags);
3038 * If __dm_suspend returns 0, the device is completely quiescent
3039 * now. There is no request-processing activity. All new requests
3040 * are being added to md->deferred list.
3042 * Caller must hold md->suspend_lock
3044 static int __dm_suspend(struct mapped_device *md, struct dm_table *map,
3045 unsigned suspend_flags, int interruptible)
3047 bool do_lockfs = suspend_flags & DM_SUSPEND_LOCKFS_FLAG;
3048 bool noflush = suspend_flags & DM_SUSPEND_NOFLUSH_FLAG;
3052 * DMF_NOFLUSH_SUSPENDING must be set before presuspend.
3053 * This flag is cleared before dm_suspend returns.
3056 set_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
3059 * This gets reverted if there's an error later and the targets
3060 * provide the .presuspend_undo hook.
3062 dm_table_presuspend_targets(map);
3065 * Flush I/O to the device.
3066 * Any I/O submitted after lock_fs() may not be flushed.
3067 * noflush takes precedence over do_lockfs.
3068 * (lock_fs() flushes I/Os and waits for them to complete.)
3070 if (!noflush && do_lockfs) {
3073 dm_table_presuspend_undo_targets(map);
3079 * Here we must make sure that no processes are submitting requests
3080 * to target drivers i.e. no one may be executing
3081 * __split_and_process_bio. This is called from dm_request and
3084 * To get all processes out of __split_and_process_bio in dm_request,
3085 * we take the write lock. To prevent any process from reentering
3086 * __split_and_process_bio from dm_request and quiesce the thread
3087 * (dm_wq_work), we set BMF_BLOCK_IO_FOR_SUSPEND and call
3088 * flush_workqueue(md->wq).
3090 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
3092 synchronize_srcu(&md->io_barrier);
3095 * Stop md->queue before flushing md->wq in case request-based
3096 * dm defers requests to md->wq from md->queue.
3098 if (dm_request_based(md)) {
3099 stop_queue(md->queue);
3100 if (md->kworker_task)
3101 flush_kthread_worker(&md->kworker);
3104 flush_workqueue(md->wq);
3107 * At this point no more requests are entering target request routines.
3108 * We call dm_wait_for_completion to wait for all existing requests
3111 r = dm_wait_for_completion(md, interruptible);
3114 clear_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
3116 synchronize_srcu(&md->io_barrier);
3118 /* were we interrupted ? */
3122 if (dm_request_based(md))
3123 start_queue(md->queue);
3126 dm_table_presuspend_undo_targets(map);
3127 /* pushback list is already flushed, so skip flush */
3134 * We need to be able to change a mapping table under a mounted
3135 * filesystem. For example we might want to move some data in
3136 * the background. Before the table can be swapped with
3137 * dm_bind_table, dm_suspend must be called to flush any in
3138 * flight bios and ensure that any further io gets deferred.
3141 * Suspend mechanism in request-based dm.
3143 * 1. Flush all I/Os by lock_fs() if needed.
3144 * 2. Stop dispatching any I/O by stopping the request_queue.
3145 * 3. Wait for all in-flight I/Os to be completed or requeued.
3147 * To abort suspend, start the request_queue.
3149 int dm_suspend(struct mapped_device *md, unsigned suspend_flags)
3151 struct dm_table *map = NULL;
3155 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
3157 if (dm_suspended_md(md)) {
3162 if (dm_suspended_internally_md(md)) {
3163 /* already internally suspended, wait for internal resume */
3164 mutex_unlock(&md->suspend_lock);
3165 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
3171 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3173 r = __dm_suspend(md, map, suspend_flags, TASK_INTERRUPTIBLE);
3177 set_bit(DMF_SUSPENDED, &md->flags);
3179 dm_table_postsuspend_targets(map);
3182 mutex_unlock(&md->suspend_lock);
3186 static int __dm_resume(struct mapped_device *md, struct dm_table *map)
3189 int r = dm_table_resume_targets(map);
3197 * Flushing deferred I/Os must be done after targets are resumed
3198 * so that mapping of targets can work correctly.
3199 * Request-based dm is queueing the deferred I/Os in its request_queue.
3201 if (dm_request_based(md))
3202 start_queue(md->queue);
3209 int dm_resume(struct mapped_device *md)
3212 struct dm_table *map = NULL;
3215 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
3217 if (!dm_suspended_md(md))
3220 if (dm_suspended_internally_md(md)) {
3221 /* already internally suspended, wait for internal resume */
3222 mutex_unlock(&md->suspend_lock);
3223 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
3229 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3230 if (!map || !dm_table_get_size(map))
3233 r = __dm_resume(md, map);
3237 clear_bit(DMF_SUSPENDED, &md->flags);
3241 mutex_unlock(&md->suspend_lock);
3247 * Internal suspend/resume works like userspace-driven suspend. It waits
3248 * until all bios finish and prevents issuing new bios to the target drivers.
3249 * It may be used only from the kernel.
3252 static void __dm_internal_suspend(struct mapped_device *md, unsigned suspend_flags)
3254 struct dm_table *map = NULL;
3256 if (md->internal_suspend_count++)
3257 return; /* nested internal suspend */
3259 if (dm_suspended_md(md)) {
3260 set_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3261 return; /* nest suspend */
3264 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3267 * Using TASK_UNINTERRUPTIBLE because only NOFLUSH internal suspend is
3268 * supported. Properly supporting a TASK_INTERRUPTIBLE internal suspend
3269 * would require changing .presuspend to return an error -- avoid this
3270 * until there is a need for more elaborate variants of internal suspend.
3272 (void) __dm_suspend(md, map, suspend_flags, TASK_UNINTERRUPTIBLE);
3274 set_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3276 dm_table_postsuspend_targets(map);
3279 static void __dm_internal_resume(struct mapped_device *md)
3281 BUG_ON(!md->internal_suspend_count);
3283 if (--md->internal_suspend_count)
3284 return; /* resume from nested internal suspend */
3286 if (dm_suspended_md(md))
3287 goto done; /* resume from nested suspend */
3290 * NOTE: existing callers don't need to call dm_table_resume_targets
3291 * (which may fail -- so best to avoid it for now by passing NULL map)
3293 (void) __dm_resume(md, NULL);
3296 clear_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3297 smp_mb__after_atomic();
3298 wake_up_bit(&md->flags, DMF_SUSPENDED_INTERNALLY);
3301 void dm_internal_suspend_noflush(struct mapped_device *md)
3303 mutex_lock(&md->suspend_lock);
3304 __dm_internal_suspend(md, DM_SUSPEND_NOFLUSH_FLAG);
3305 mutex_unlock(&md->suspend_lock);
3307 EXPORT_SYMBOL_GPL(dm_internal_suspend_noflush);
3309 void dm_internal_resume(struct mapped_device *md)
3311 mutex_lock(&md->suspend_lock);
3312 __dm_internal_resume(md);
3313 mutex_unlock(&md->suspend_lock);
3315 EXPORT_SYMBOL_GPL(dm_internal_resume);
3318 * Fast variants of internal suspend/resume hold md->suspend_lock,
3319 * which prevents interaction with userspace-driven suspend.
3322 void dm_internal_suspend_fast(struct mapped_device *md)
3324 mutex_lock(&md->suspend_lock);
3325 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
3328 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
3329 synchronize_srcu(&md->io_barrier);
3330 flush_workqueue(md->wq);
3331 dm_wait_for_completion(md, TASK_UNINTERRUPTIBLE);
3333 EXPORT_SYMBOL_GPL(dm_internal_suspend_fast);
3335 void dm_internal_resume_fast(struct mapped_device *md)
3337 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
3343 mutex_unlock(&md->suspend_lock);
3345 EXPORT_SYMBOL_GPL(dm_internal_resume_fast);
3347 /*-----------------------------------------------------------------
3348 * Event notification.
3349 *---------------------------------------------------------------*/
3350 int dm_kobject_uevent(struct mapped_device *md, enum kobject_action action,
3353 char udev_cookie[DM_COOKIE_LENGTH];
3354 char *envp[] = { udev_cookie, NULL };
3357 return kobject_uevent(&disk_to_dev(md->disk)->kobj, action);
3359 snprintf(udev_cookie, DM_COOKIE_LENGTH, "%s=%u",
3360 DM_COOKIE_ENV_VAR_NAME, cookie);
3361 return kobject_uevent_env(&disk_to_dev(md->disk)->kobj,
3366 uint32_t dm_next_uevent_seq(struct mapped_device *md)
3368 return atomic_add_return(1, &md->uevent_seq);
3371 uint32_t dm_get_event_nr(struct mapped_device *md)
3373 return atomic_read(&md->event_nr);
3376 int dm_wait_event(struct mapped_device *md, int event_nr)
3378 return wait_event_interruptible(md->eventq,
3379 (event_nr != atomic_read(&md->event_nr)));
3382 void dm_uevent_add(struct mapped_device *md, struct list_head *elist)
3384 unsigned long flags;
3386 spin_lock_irqsave(&md->uevent_lock, flags);
3387 list_add(elist, &md->uevent_list);
3388 spin_unlock_irqrestore(&md->uevent_lock, flags);
3392 * The gendisk is only valid as long as you have a reference
3395 struct gendisk *dm_disk(struct mapped_device *md)
3399 EXPORT_SYMBOL_GPL(dm_disk);
3401 struct kobject *dm_kobject(struct mapped_device *md)
3403 return &md->kobj_holder.kobj;
3406 struct mapped_device *dm_get_from_kobject(struct kobject *kobj)
3408 struct mapped_device *md;
3410 md = container_of(kobj, struct mapped_device, kobj_holder.kobj);
3412 if (test_bit(DMF_FREEING, &md->flags) ||
3420 int dm_suspended_md(struct mapped_device *md)
3422 return test_bit(DMF_SUSPENDED, &md->flags);
3425 int dm_suspended_internally_md(struct mapped_device *md)
3427 return test_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3430 int dm_test_deferred_remove_flag(struct mapped_device *md)
3432 return test_bit(DMF_DEFERRED_REMOVE, &md->flags);
3435 int dm_suspended(struct dm_target *ti)
3437 return dm_suspended_md(dm_table_get_md(ti->table));
3439 EXPORT_SYMBOL_GPL(dm_suspended);
3441 int dm_noflush_suspending(struct dm_target *ti)
3443 return __noflush_suspending(dm_table_get_md(ti->table));
3445 EXPORT_SYMBOL_GPL(dm_noflush_suspending);
3447 struct dm_md_mempools *dm_alloc_md_mempools(struct mapped_device *md, unsigned type,
3448 unsigned integrity, unsigned per_bio_data_size)
3450 struct dm_md_mempools *pools = kzalloc(sizeof(*pools), GFP_KERNEL);
3451 struct kmem_cache *cachep = NULL;
3452 unsigned int pool_size = 0;
3453 unsigned int front_pad;
3458 type = filter_md_type(type, md);
3461 case DM_TYPE_BIO_BASED:
3463 pool_size = dm_get_reserved_bio_based_ios();
3464 front_pad = roundup(per_bio_data_size, __alignof__(struct dm_target_io)) + offsetof(struct dm_target_io, clone);
3466 case DM_TYPE_REQUEST_BASED:
3467 cachep = _rq_tio_cache;
3468 pool_size = dm_get_reserved_rq_based_ios();
3469 pools->rq_pool = mempool_create_slab_pool(pool_size, _rq_cache);
3470 if (!pools->rq_pool)
3472 /* fall through to setup remaining rq-based pools */
3473 case DM_TYPE_MQ_REQUEST_BASED:
3475 pool_size = dm_get_reserved_rq_based_ios();
3476 front_pad = offsetof(struct dm_rq_clone_bio_info, clone);
3477 /* per_bio_data_size is not used. See __bind_mempools(). */
3478 WARN_ON(per_bio_data_size != 0);
3485 pools->io_pool = mempool_create_slab_pool(pool_size, cachep);
3486 if (!pools->io_pool)
3490 pools->bs = bioset_create_nobvec(pool_size, front_pad);
3494 if (integrity && bioset_integrity_create(pools->bs, pool_size))
3500 dm_free_md_mempools(pools);
3505 void dm_free_md_mempools(struct dm_md_mempools *pools)
3511 mempool_destroy(pools->io_pool);
3514 mempool_destroy(pools->rq_pool);
3517 bioset_free(pools->bs);
3522 static const struct block_device_operations dm_blk_dops = {
3523 .open = dm_blk_open,
3524 .release = dm_blk_close,
3525 .ioctl = dm_blk_ioctl,
3526 .getgeo = dm_blk_getgeo,
3527 .owner = THIS_MODULE
3533 module_init(dm_init);
3534 module_exit(dm_exit);
3536 module_param(major, uint, 0);
3537 MODULE_PARM_DESC(major, "The major number of the device mapper");
3539 module_param(reserved_bio_based_ios, uint, S_IRUGO | S_IWUSR);
3540 MODULE_PARM_DESC(reserved_bio_based_ios, "Reserved IOs in bio-based mempools");
3542 module_param(reserved_rq_based_ios, uint, S_IRUGO | S_IWUSR);
3543 MODULE_PARM_DESC(reserved_rq_based_ios, "Reserved IOs in request-based mempools");
3545 module_param(use_blk_mq, bool, S_IRUGO | S_IWUSR);
3546 MODULE_PARM_DESC(use_blk_mq, "Use block multiqueue for request-based DM devices");
3548 MODULE_DESCRIPTION(DM_NAME " driver");
3549 MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
3550 MODULE_LICENSE("GPL");