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;
1009 * An error has already been detected on the request.
1010 * Once error occurred, just let clone->end_io() handle
1014 else if (bio->bi_error) {
1016 * Don't notice the error to the upper layer yet.
1017 * The error handling decision is made by the target driver,
1018 * when the request is completed.
1020 tio->error = bio->bi_error;
1025 * I/O for the bio successfully completed.
1026 * Notice the data completion to the upper layer.
1030 * bios are processed from the head of the list.
1031 * So the completing bio should always be rq->bio.
1032 * If it's not, something wrong is happening.
1034 if (tio->orig->bio != bio)
1035 DMERR("bio completion is going in the middle of the request");
1038 * Update the original request.
1039 * Do not use blk_end_request() here, because it may complete
1040 * the original request before the clone, and break the ordering.
1042 blk_update_request(tio->orig, 0, nr_bytes);
1045 static struct dm_rq_target_io *tio_from_request(struct request *rq)
1047 return (rq->q->mq_ops ? blk_mq_rq_to_pdu(rq) : rq->special);
1050 static void rq_end_stats(struct mapped_device *md, struct request *orig)
1052 if (unlikely(dm_stats_used(&md->stats))) {
1053 struct dm_rq_target_io *tio = tio_from_request(orig);
1054 tio->duration_jiffies = jiffies - tio->duration_jiffies;
1055 dm_stats_account_io(&md->stats, orig->cmd_flags, blk_rq_pos(orig),
1056 tio->n_sectors, true, tio->duration_jiffies,
1062 * Don't touch any member of the md after calling this function because
1063 * the md may be freed in dm_put() at the end of this function.
1064 * Or do dm_get() before calling this function and dm_put() later.
1066 static void rq_completed(struct mapped_device *md, int rw, bool run_queue)
1068 atomic_dec(&md->pending[rw]);
1070 /* nudge anyone waiting on suspend queue */
1071 if (!md_in_flight(md))
1075 * Run this off this callpath, as drivers could invoke end_io while
1076 * inside their request_fn (and holding the queue lock). Calling
1077 * back into ->request_fn() could deadlock attempting to grab the
1081 if (md->queue->mq_ops)
1082 blk_mq_run_hw_queues(md->queue, true);
1084 blk_run_queue_async(md->queue);
1088 * dm_put() must be at the end of this function. See the comment above
1093 static void free_rq_clone(struct request *clone)
1095 struct dm_rq_target_io *tio = clone->end_io_data;
1096 struct mapped_device *md = tio->md;
1098 blk_rq_unprep_clone(clone);
1100 if (md->type == DM_TYPE_MQ_REQUEST_BASED)
1101 /* stacked on blk-mq queue(s) */
1102 tio->ti->type->release_clone_rq(clone);
1103 else if (!md->queue->mq_ops)
1104 /* request_fn queue stacked on request_fn queue(s) */
1105 free_clone_request(md, clone);
1107 * NOTE: for the blk-mq queue stacked on request_fn queue(s) case:
1108 * no need to call free_clone_request() because we leverage blk-mq by
1109 * allocating the clone at the end of the blk-mq pdu (see: clone_rq)
1112 if (!md->queue->mq_ops)
1117 * Complete the clone and the original request.
1118 * Must be called without clone's queue lock held,
1119 * see end_clone_request() for more details.
1121 static void dm_end_request(struct request *clone, int error)
1123 int rw = rq_data_dir(clone);
1124 struct dm_rq_target_io *tio = clone->end_io_data;
1125 struct mapped_device *md = tio->md;
1126 struct request *rq = tio->orig;
1128 if (rq->cmd_type == REQ_TYPE_BLOCK_PC) {
1129 rq->errors = clone->errors;
1130 rq->resid_len = clone->resid_len;
1134 * We are using the sense buffer of the original
1136 * So setting the length of the sense data is enough.
1138 rq->sense_len = clone->sense_len;
1141 free_rq_clone(clone);
1142 rq_end_stats(md, rq);
1144 blk_end_request_all(rq, error);
1146 blk_mq_end_request(rq, error);
1147 rq_completed(md, rw, true);
1150 static void dm_unprep_request(struct request *rq)
1152 struct dm_rq_target_io *tio = tio_from_request(rq);
1153 struct request *clone = tio->clone;
1155 if (!rq->q->mq_ops) {
1157 rq->cmd_flags &= ~REQ_DONTPREP;
1161 free_rq_clone(clone);
1165 * Requeue the original request of a clone.
1167 static void old_requeue_request(struct request *rq)
1169 struct request_queue *q = rq->q;
1170 unsigned long flags;
1172 spin_lock_irqsave(q->queue_lock, flags);
1173 blk_requeue_request(q, rq);
1174 blk_run_queue_async(q);
1175 spin_unlock_irqrestore(q->queue_lock, flags);
1178 static void dm_requeue_original_request(struct mapped_device *md,
1181 int rw = rq_data_dir(rq);
1183 dm_unprep_request(rq);
1185 rq_end_stats(md, rq);
1187 old_requeue_request(rq);
1189 blk_mq_requeue_request(rq);
1190 blk_mq_kick_requeue_list(rq->q);
1193 rq_completed(md, rw, false);
1196 static void old_stop_queue(struct request_queue *q)
1198 unsigned long flags;
1200 if (blk_queue_stopped(q))
1203 spin_lock_irqsave(q->queue_lock, flags);
1205 spin_unlock_irqrestore(q->queue_lock, flags);
1208 static void stop_queue(struct request_queue *q)
1213 blk_mq_stop_hw_queues(q);
1216 static void old_start_queue(struct request_queue *q)
1218 unsigned long flags;
1220 spin_lock_irqsave(q->queue_lock, flags);
1221 if (blk_queue_stopped(q))
1223 spin_unlock_irqrestore(q->queue_lock, flags);
1226 static void start_queue(struct request_queue *q)
1231 blk_mq_start_stopped_hw_queues(q, true);
1234 static void dm_done(struct request *clone, int error, bool mapped)
1237 struct dm_rq_target_io *tio = clone->end_io_data;
1238 dm_request_endio_fn rq_end_io = NULL;
1241 rq_end_io = tio->ti->type->rq_end_io;
1243 if (mapped && rq_end_io)
1244 r = rq_end_io(tio->ti, clone, error, &tio->info);
1247 if (unlikely(r == -EREMOTEIO && (clone->cmd_flags & REQ_WRITE_SAME) &&
1248 !clone->q->limits.max_write_same_sectors))
1249 disable_write_same(tio->md);
1252 /* The target wants to complete the I/O */
1253 dm_end_request(clone, r);
1254 else if (r == DM_ENDIO_INCOMPLETE)
1255 /* The target will handle the I/O */
1257 else if (r == DM_ENDIO_REQUEUE)
1258 /* The target wants to requeue the I/O */
1259 dm_requeue_original_request(tio->md, tio->orig);
1261 DMWARN("unimplemented target endio return value: %d", r);
1267 * Request completion handler for request-based dm
1269 static void dm_softirq_done(struct request *rq)
1272 struct dm_rq_target_io *tio = tio_from_request(rq);
1273 struct request *clone = tio->clone;
1277 rq_end_stats(tio->md, rq);
1278 rw = rq_data_dir(rq);
1279 if (!rq->q->mq_ops) {
1280 blk_end_request_all(rq, tio->error);
1281 rq_completed(tio->md, rw, false);
1284 blk_mq_end_request(rq, tio->error);
1285 rq_completed(tio->md, rw, false);
1290 if (rq->cmd_flags & REQ_FAILED)
1293 dm_done(clone, tio->error, mapped);
1297 * Complete the clone and the original request with the error status
1298 * through softirq context.
1300 static void dm_complete_request(struct request *rq, int error)
1302 struct dm_rq_target_io *tio = tio_from_request(rq);
1305 blk_complete_request(rq);
1309 * Complete the not-mapped clone and the original request with the error status
1310 * through softirq context.
1311 * Target's rq_end_io() function isn't called.
1312 * This may be used when the target's map_rq() or clone_and_map_rq() functions fail.
1314 static void dm_kill_unmapped_request(struct request *rq, int error)
1316 rq->cmd_flags |= REQ_FAILED;
1317 dm_complete_request(rq, error);
1321 * Called with the clone's queue lock held (for non-blk-mq)
1323 static void end_clone_request(struct request *clone, int error)
1325 struct dm_rq_target_io *tio = clone->end_io_data;
1327 if (!clone->q->mq_ops) {
1329 * For just cleaning up the information of the queue in which
1330 * the clone was dispatched.
1331 * The clone is *NOT* freed actually here because it is alloced
1332 * from dm own mempool (REQ_ALLOCED isn't set).
1334 __blk_put_request(clone->q, clone);
1338 * Actual request completion is done in a softirq context which doesn't
1339 * hold the clone's queue lock. Otherwise, deadlock could occur because:
1340 * - another request may be submitted by the upper level driver
1341 * of the stacking during the completion
1342 * - the submission which requires queue lock may be done
1343 * against this clone's queue
1345 dm_complete_request(tio->orig, error);
1349 * Return maximum size of I/O possible at the supplied sector up to the current
1352 static sector_t max_io_len_target_boundary(sector_t sector, struct dm_target *ti)
1354 sector_t target_offset = dm_target_offset(ti, sector);
1356 return ti->len - target_offset;
1359 static sector_t max_io_len(sector_t sector, struct dm_target *ti)
1361 sector_t len = max_io_len_target_boundary(sector, ti);
1362 sector_t offset, max_len;
1365 * Does the target need to split even further?
1367 if (ti->max_io_len) {
1368 offset = dm_target_offset(ti, sector);
1369 if (unlikely(ti->max_io_len & (ti->max_io_len - 1)))
1370 max_len = sector_div(offset, ti->max_io_len);
1372 max_len = offset & (ti->max_io_len - 1);
1373 max_len = ti->max_io_len - max_len;
1382 int dm_set_target_max_io_len(struct dm_target *ti, sector_t len)
1384 if (len > UINT_MAX) {
1385 DMERR("Specified maximum size of target IO (%llu) exceeds limit (%u)",
1386 (unsigned long long)len, UINT_MAX);
1387 ti->error = "Maximum size of target IO is too large";
1391 ti->max_io_len = (uint32_t) len;
1395 EXPORT_SYMBOL_GPL(dm_set_target_max_io_len);
1398 * A target may call dm_accept_partial_bio only from the map routine. It is
1399 * allowed for all bio types except REQ_FLUSH.
1401 * dm_accept_partial_bio informs the dm that the target only wants to process
1402 * additional n_sectors sectors of the bio and the rest of the data should be
1403 * sent in a next bio.
1405 * A diagram that explains the arithmetics:
1406 * +--------------------+---------------+-------+
1408 * +--------------------+---------------+-------+
1410 * <-------------- *tio->len_ptr --------------->
1411 * <------- bi_size ------->
1414 * Region 1 was already iterated over with bio_advance or similar function.
1415 * (it may be empty if the target doesn't use bio_advance)
1416 * Region 2 is the remaining bio size that the target wants to process.
1417 * (it may be empty if region 1 is non-empty, although there is no reason
1419 * The target requires that region 3 is to be sent in the next bio.
1421 * If the target wants to receive multiple copies of the bio (via num_*bios, etc),
1422 * the partially processed part (the sum of regions 1+2) must be the same for all
1423 * copies of the bio.
1425 void dm_accept_partial_bio(struct bio *bio, unsigned n_sectors)
1427 struct dm_target_io *tio = container_of(bio, struct dm_target_io, clone);
1428 unsigned bi_size = bio->bi_iter.bi_size >> SECTOR_SHIFT;
1429 BUG_ON(bio->bi_rw & REQ_FLUSH);
1430 BUG_ON(bi_size > *tio->len_ptr);
1431 BUG_ON(n_sectors > bi_size);
1432 *tio->len_ptr -= bi_size - n_sectors;
1433 bio->bi_iter.bi_size = n_sectors << SECTOR_SHIFT;
1435 EXPORT_SYMBOL_GPL(dm_accept_partial_bio);
1437 static void __map_bio(struct dm_target_io *tio)
1441 struct mapped_device *md;
1442 struct bio *clone = &tio->clone;
1443 struct dm_target *ti = tio->ti;
1445 clone->bi_end_io = clone_endio;
1448 * Map the clone. If r == 0 we don't need to do
1449 * anything, the target has assumed ownership of
1452 atomic_inc(&tio->io->io_count);
1453 sector = clone->bi_iter.bi_sector;
1454 r = ti->type->map(ti, clone);
1455 if (r == DM_MAPIO_REMAPPED) {
1456 /* the bio has been remapped so dispatch it */
1458 trace_block_bio_remap(bdev_get_queue(clone->bi_bdev), clone,
1459 tio->io->bio->bi_bdev->bd_dev, sector);
1461 generic_make_request(clone);
1462 } else if (r < 0 || r == DM_MAPIO_REQUEUE) {
1463 /* error the io and bail out, or requeue it if needed */
1465 dec_pending(tio->io, r);
1467 } else if (r != DM_MAPIO_SUBMITTED) {
1468 DMWARN("unimplemented target map return value: %d", r);
1474 struct mapped_device *md;
1475 struct dm_table *map;
1479 unsigned sector_count;
1482 static void bio_setup_sector(struct bio *bio, sector_t sector, unsigned len)
1484 bio->bi_iter.bi_sector = sector;
1485 bio->bi_iter.bi_size = to_bytes(len);
1489 * Creates a bio that consists of range of complete bvecs.
1491 static void clone_bio(struct dm_target_io *tio, struct bio *bio,
1492 sector_t sector, unsigned len)
1494 struct bio *clone = &tio->clone;
1496 __bio_clone_fast(clone, bio);
1498 if (bio_integrity(bio))
1499 bio_integrity_clone(clone, bio, GFP_NOIO);
1501 bio_advance(clone, to_bytes(sector - clone->bi_iter.bi_sector));
1502 clone->bi_iter.bi_size = to_bytes(len);
1504 if (bio_integrity(bio))
1505 bio_integrity_trim(clone, 0, len);
1508 static struct dm_target_io *alloc_tio(struct clone_info *ci,
1509 struct dm_target *ti,
1510 unsigned target_bio_nr)
1512 struct dm_target_io *tio;
1515 clone = bio_alloc_bioset(GFP_NOIO, 0, ci->md->bs);
1516 tio = container_of(clone, struct dm_target_io, clone);
1520 tio->target_bio_nr = target_bio_nr;
1525 static void __clone_and_map_simple_bio(struct clone_info *ci,
1526 struct dm_target *ti,
1527 unsigned target_bio_nr, unsigned *len)
1529 struct dm_target_io *tio = alloc_tio(ci, ti, target_bio_nr);
1530 struct bio *clone = &tio->clone;
1534 __bio_clone_fast(clone, ci->bio);
1536 bio_setup_sector(clone, ci->sector, *len);
1541 static void __send_duplicate_bios(struct clone_info *ci, struct dm_target *ti,
1542 unsigned num_bios, unsigned *len)
1544 unsigned target_bio_nr;
1546 for (target_bio_nr = 0; target_bio_nr < num_bios; target_bio_nr++)
1547 __clone_and_map_simple_bio(ci, ti, target_bio_nr, len);
1550 static int __send_empty_flush(struct clone_info *ci)
1552 unsigned target_nr = 0;
1553 struct dm_target *ti;
1555 BUG_ON(bio_has_data(ci->bio));
1556 while ((ti = dm_table_get_target(ci->map, target_nr++)))
1557 __send_duplicate_bios(ci, ti, ti->num_flush_bios, NULL);
1562 static void __clone_and_map_data_bio(struct clone_info *ci, struct dm_target *ti,
1563 sector_t sector, unsigned *len)
1565 struct bio *bio = ci->bio;
1566 struct dm_target_io *tio;
1567 unsigned target_bio_nr;
1568 unsigned num_target_bios = 1;
1571 * Does the target want to receive duplicate copies of the bio?
1573 if (bio_data_dir(bio) == WRITE && ti->num_write_bios)
1574 num_target_bios = ti->num_write_bios(ti, bio);
1576 for (target_bio_nr = 0; target_bio_nr < num_target_bios; target_bio_nr++) {
1577 tio = alloc_tio(ci, ti, target_bio_nr);
1579 clone_bio(tio, bio, sector, *len);
1584 typedef unsigned (*get_num_bios_fn)(struct dm_target *ti);
1586 static unsigned get_num_discard_bios(struct dm_target *ti)
1588 return ti->num_discard_bios;
1591 static unsigned get_num_write_same_bios(struct dm_target *ti)
1593 return ti->num_write_same_bios;
1596 typedef bool (*is_split_required_fn)(struct dm_target *ti);
1598 static bool is_split_required_for_discard(struct dm_target *ti)
1600 return ti->split_discard_bios;
1603 static int __send_changing_extent_only(struct clone_info *ci,
1604 get_num_bios_fn get_num_bios,
1605 is_split_required_fn is_split_required)
1607 struct dm_target *ti;
1612 ti = dm_table_find_target(ci->map, ci->sector);
1613 if (!dm_target_is_valid(ti))
1617 * Even though the device advertised support for this type of
1618 * request, that does not mean every target supports it, and
1619 * reconfiguration might also have changed that since the
1620 * check was performed.
1622 num_bios = get_num_bios ? get_num_bios(ti) : 0;
1626 if (is_split_required && !is_split_required(ti))
1627 len = min((sector_t)ci->sector_count, max_io_len_target_boundary(ci->sector, ti));
1629 len = min((sector_t)ci->sector_count, max_io_len(ci->sector, ti));
1631 __send_duplicate_bios(ci, ti, num_bios, &len);
1634 } while (ci->sector_count -= len);
1639 static int __send_discard(struct clone_info *ci)
1641 return __send_changing_extent_only(ci, get_num_discard_bios,
1642 is_split_required_for_discard);
1645 static int __send_write_same(struct clone_info *ci)
1647 return __send_changing_extent_only(ci, get_num_write_same_bios, NULL);
1651 * Select the correct strategy for processing a non-flush bio.
1653 static int __split_and_process_non_flush(struct clone_info *ci)
1655 struct bio *bio = ci->bio;
1656 struct dm_target *ti;
1659 if (unlikely(bio->bi_rw & REQ_DISCARD))
1660 return __send_discard(ci);
1661 else if (unlikely(bio->bi_rw & REQ_WRITE_SAME))
1662 return __send_write_same(ci);
1664 ti = dm_table_find_target(ci->map, ci->sector);
1665 if (!dm_target_is_valid(ti))
1668 len = min_t(sector_t, max_io_len(ci->sector, ti), ci->sector_count);
1670 __clone_and_map_data_bio(ci, ti, ci->sector, &len);
1673 ci->sector_count -= len;
1679 * Entry point to split a bio into clones and submit them to the targets.
1681 static void __split_and_process_bio(struct mapped_device *md,
1682 struct dm_table *map, struct bio *bio)
1684 struct clone_info ci;
1687 if (unlikely(!map)) {
1694 ci.io = alloc_io(md);
1696 atomic_set(&ci.io->io_count, 1);
1699 spin_lock_init(&ci.io->endio_lock);
1700 ci.sector = bio->bi_iter.bi_sector;
1702 start_io_acct(ci.io);
1704 if (bio->bi_rw & REQ_FLUSH) {
1705 ci.bio = &ci.md->flush_bio;
1706 ci.sector_count = 0;
1707 error = __send_empty_flush(&ci);
1708 /* dec_pending submits any data associated with flush */
1711 ci.sector_count = bio_sectors(bio);
1712 while (ci.sector_count && !error)
1713 error = __split_and_process_non_flush(&ci);
1716 /* drop the extra reference count */
1717 dec_pending(ci.io, error);
1719 /*-----------------------------------------------------------------
1721 *---------------------------------------------------------------*/
1724 * The request function that just remaps the bio built up by
1727 static void dm_make_request(struct request_queue *q, struct bio *bio)
1729 int rw = bio_data_dir(bio);
1730 struct mapped_device *md = q->queuedata;
1732 struct dm_table *map;
1734 map = dm_get_live_table(md, &srcu_idx);
1736 blk_queue_split(q, &bio, q->bio_split);
1738 generic_start_io_acct(rw, bio_sectors(bio), &dm_disk(md)->part0);
1740 /* if we're suspended, we have to queue this io for later */
1741 if (unlikely(test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags))) {
1742 dm_put_live_table(md, srcu_idx);
1744 if (bio_rw(bio) != READA)
1751 __split_and_process_bio(md, map, bio);
1752 dm_put_live_table(md, srcu_idx);
1756 int dm_request_based(struct mapped_device *md)
1758 return blk_queue_stackable(md->queue);
1761 static void dm_dispatch_clone_request(struct request *clone, struct request *rq)
1765 if (blk_queue_io_stat(clone->q))
1766 clone->cmd_flags |= REQ_IO_STAT;
1768 clone->start_time = jiffies;
1769 r = blk_insert_cloned_request(clone->q, clone);
1771 /* must complete clone in terms of original request */
1772 dm_complete_request(rq, r);
1775 static int dm_rq_bio_constructor(struct bio *bio, struct bio *bio_orig,
1778 struct dm_rq_target_io *tio = data;
1779 struct dm_rq_clone_bio_info *info =
1780 container_of(bio, struct dm_rq_clone_bio_info, clone);
1782 info->orig = bio_orig;
1784 bio->bi_end_io = end_clone_bio;
1789 static int setup_clone(struct request *clone, struct request *rq,
1790 struct dm_rq_target_io *tio, gfp_t gfp_mask)
1794 r = blk_rq_prep_clone(clone, rq, tio->md->bs, gfp_mask,
1795 dm_rq_bio_constructor, tio);
1799 clone->cmd = rq->cmd;
1800 clone->cmd_len = rq->cmd_len;
1801 clone->sense = rq->sense;
1802 clone->end_io = end_clone_request;
1803 clone->end_io_data = tio;
1810 static struct request *clone_rq(struct request *rq, struct mapped_device *md,
1811 struct dm_rq_target_io *tio, gfp_t gfp_mask)
1814 * Do not allocate a clone if tio->clone was already set
1815 * (see: dm_mq_queue_rq).
1817 bool alloc_clone = !tio->clone;
1818 struct request *clone;
1821 clone = alloc_clone_request(md, gfp_mask);
1827 blk_rq_init(NULL, clone);
1828 if (setup_clone(clone, rq, tio, gfp_mask)) {
1831 free_clone_request(md, clone);
1838 static void map_tio_request(struct kthread_work *work);
1840 static void init_tio(struct dm_rq_target_io *tio, struct request *rq,
1841 struct mapped_device *md)
1848 memset(&tio->info, 0, sizeof(tio->info));
1849 if (md->kworker_task)
1850 init_kthread_work(&tio->work, map_tio_request);
1853 static struct dm_rq_target_io *prep_tio(struct request *rq,
1854 struct mapped_device *md, gfp_t gfp_mask)
1856 struct dm_rq_target_io *tio;
1858 struct dm_table *table;
1860 tio = alloc_rq_tio(md, gfp_mask);
1864 init_tio(tio, rq, md);
1866 table = dm_get_live_table(md, &srcu_idx);
1867 if (!dm_table_mq_request_based(table)) {
1868 if (!clone_rq(rq, md, tio, gfp_mask)) {
1869 dm_put_live_table(md, srcu_idx);
1874 dm_put_live_table(md, srcu_idx);
1880 * Called with the queue lock held.
1882 static int dm_prep_fn(struct request_queue *q, struct request *rq)
1884 struct mapped_device *md = q->queuedata;
1885 struct dm_rq_target_io *tio;
1887 if (unlikely(rq->special)) {
1888 DMWARN("Already has something in rq->special.");
1889 return BLKPREP_KILL;
1892 tio = prep_tio(rq, md, GFP_ATOMIC);
1894 return BLKPREP_DEFER;
1897 rq->cmd_flags |= REQ_DONTPREP;
1904 * 0 : the request has been processed
1905 * DM_MAPIO_REQUEUE : the original request needs to be requeued
1906 * < 0 : the request was completed due to failure
1908 static int map_request(struct dm_rq_target_io *tio, struct request *rq,
1909 struct mapped_device *md)
1912 struct dm_target *ti = tio->ti;
1913 struct request *clone = NULL;
1917 r = ti->type->map_rq(ti, clone, &tio->info);
1919 r = ti->type->clone_and_map_rq(ti, rq, &tio->info, &clone);
1921 /* The target wants to complete the I/O */
1922 dm_kill_unmapped_request(rq, r);
1925 if (r != DM_MAPIO_REMAPPED)
1927 if (setup_clone(clone, rq, tio, GFP_ATOMIC)) {
1929 ti->type->release_clone_rq(clone);
1930 return DM_MAPIO_REQUEUE;
1935 case DM_MAPIO_SUBMITTED:
1936 /* The target has taken the I/O to submit by itself later */
1938 case DM_MAPIO_REMAPPED:
1939 /* The target has remapped the I/O so dispatch it */
1940 trace_block_rq_remap(clone->q, clone, disk_devt(dm_disk(md)),
1942 dm_dispatch_clone_request(clone, rq);
1944 case DM_MAPIO_REQUEUE:
1945 /* The target wants to requeue the I/O */
1946 dm_requeue_original_request(md, tio->orig);
1950 DMWARN("unimplemented target map return value: %d", r);
1954 /* The target wants to complete the I/O */
1955 dm_kill_unmapped_request(rq, r);
1962 static void map_tio_request(struct kthread_work *work)
1964 struct dm_rq_target_io *tio = container_of(work, struct dm_rq_target_io, work);
1965 struct request *rq = tio->orig;
1966 struct mapped_device *md = tio->md;
1968 if (map_request(tio, rq, md) == DM_MAPIO_REQUEUE)
1969 dm_requeue_original_request(md, rq);
1972 static void dm_start_request(struct mapped_device *md, struct request *orig)
1974 if (!orig->q->mq_ops)
1975 blk_start_request(orig);
1977 blk_mq_start_request(orig);
1978 atomic_inc(&md->pending[rq_data_dir(orig)]);
1980 if (md->seq_rq_merge_deadline_usecs) {
1981 md->last_rq_pos = rq_end_sector(orig);
1982 md->last_rq_rw = rq_data_dir(orig);
1983 md->last_rq_start_time = ktime_get();
1986 if (unlikely(dm_stats_used(&md->stats))) {
1987 struct dm_rq_target_io *tio = tio_from_request(orig);
1988 tio->duration_jiffies = jiffies;
1989 tio->n_sectors = blk_rq_sectors(orig);
1990 dm_stats_account_io(&md->stats, orig->cmd_flags, blk_rq_pos(orig),
1991 tio->n_sectors, false, 0, &tio->stats_aux);
1995 * Hold the md reference here for the in-flight I/O.
1996 * We can't rely on the reference count by device opener,
1997 * because the device may be closed during the request completion
1998 * when all bios are completed.
1999 * See the comment in rq_completed() too.
2004 #define MAX_SEQ_RQ_MERGE_DEADLINE_USECS 100000
2006 ssize_t dm_attr_rq_based_seq_io_merge_deadline_show(struct mapped_device *md, char *buf)
2008 return sprintf(buf, "%u\n", md->seq_rq_merge_deadline_usecs);
2011 ssize_t dm_attr_rq_based_seq_io_merge_deadline_store(struct mapped_device *md,
2012 const char *buf, size_t count)
2016 if (!dm_request_based(md) || md->use_blk_mq)
2019 if (kstrtouint(buf, 10, &deadline))
2022 if (deadline > MAX_SEQ_RQ_MERGE_DEADLINE_USECS)
2023 deadline = MAX_SEQ_RQ_MERGE_DEADLINE_USECS;
2025 md->seq_rq_merge_deadline_usecs = deadline;
2030 static bool dm_request_peeked_before_merge_deadline(struct mapped_device *md)
2032 ktime_t kt_deadline;
2034 if (!md->seq_rq_merge_deadline_usecs)
2037 kt_deadline = ns_to_ktime((u64)md->seq_rq_merge_deadline_usecs * NSEC_PER_USEC);
2038 kt_deadline = ktime_add_safe(md->last_rq_start_time, kt_deadline);
2040 return !ktime_after(ktime_get(), kt_deadline);
2044 * q->request_fn for request-based dm.
2045 * Called with the queue lock held.
2047 static void dm_request_fn(struct request_queue *q)
2049 struct mapped_device *md = q->queuedata;
2051 struct dm_table *map = dm_get_live_table(md, &srcu_idx);
2052 struct dm_target *ti;
2054 struct dm_rq_target_io *tio;
2058 * For suspend, check blk_queue_stopped() and increment
2059 * ->pending within a single queue_lock not to increment the
2060 * number of in-flight I/Os after the queue is stopped in
2063 while (!blk_queue_stopped(q)) {
2064 rq = blk_peek_request(q);
2068 /* always use block 0 to find the target for flushes for now */
2070 if (!(rq->cmd_flags & REQ_FLUSH))
2071 pos = blk_rq_pos(rq);
2073 ti = dm_table_find_target(map, pos);
2074 if (!dm_target_is_valid(ti)) {
2076 * Must perform setup, that rq_completed() requires,
2077 * before calling dm_kill_unmapped_request
2079 DMERR_LIMIT("request attempted access beyond the end of device");
2080 dm_start_request(md, rq);
2081 dm_kill_unmapped_request(rq, -EIO);
2085 if (dm_request_peeked_before_merge_deadline(md) &&
2086 md_in_flight(md) && rq->bio && rq->bio->bi_vcnt == 1 &&
2087 md->last_rq_pos == pos && md->last_rq_rw == rq_data_dir(rq))
2090 if (ti->type->busy && ti->type->busy(ti))
2093 dm_start_request(md, rq);
2095 tio = tio_from_request(rq);
2096 /* Establish tio->ti before queuing work (map_tio_request) */
2098 queue_kthread_work(&md->kworker, &tio->work);
2099 BUG_ON(!irqs_disabled());
2105 blk_delay_queue(q, HZ / 100);
2107 dm_put_live_table(md, srcu_idx);
2110 static int dm_any_congested(void *congested_data, int bdi_bits)
2113 struct mapped_device *md = congested_data;
2114 struct dm_table *map;
2116 if (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
2117 map = dm_get_live_table_fast(md);
2120 * Request-based dm cares about only own queue for
2121 * the query about congestion status of request_queue
2123 if (dm_request_based(md))
2124 r = md->queue->backing_dev_info.wb.state &
2127 r = dm_table_any_congested(map, bdi_bits);
2129 dm_put_live_table_fast(md);
2135 /*-----------------------------------------------------------------
2136 * An IDR is used to keep track of allocated minor numbers.
2137 *---------------------------------------------------------------*/
2138 static void free_minor(int minor)
2140 spin_lock(&_minor_lock);
2141 idr_remove(&_minor_idr, minor);
2142 spin_unlock(&_minor_lock);
2146 * See if the device with a specific minor # is free.
2148 static int specific_minor(int minor)
2152 if (minor >= (1 << MINORBITS))
2155 idr_preload(GFP_KERNEL);
2156 spin_lock(&_minor_lock);
2158 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, minor, minor + 1, GFP_NOWAIT);
2160 spin_unlock(&_minor_lock);
2163 return r == -ENOSPC ? -EBUSY : r;
2167 static int next_free_minor(int *minor)
2171 idr_preload(GFP_KERNEL);
2172 spin_lock(&_minor_lock);
2174 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, 0, 1 << MINORBITS, GFP_NOWAIT);
2176 spin_unlock(&_minor_lock);
2184 static const struct block_device_operations dm_blk_dops;
2186 static void dm_wq_work(struct work_struct *work);
2188 static void dm_init_md_queue(struct mapped_device *md)
2191 * Request-based dm devices cannot be stacked on top of bio-based dm
2192 * devices. The type of this dm device may not have been decided yet.
2193 * The type is decided at the first table loading time.
2194 * To prevent problematic device stacking, clear the queue flag
2195 * for request stacking support until then.
2197 * This queue is new, so no concurrency on the queue_flags.
2199 queue_flag_clear_unlocked(QUEUE_FLAG_STACKABLE, md->queue);
2202 static void dm_init_old_md_queue(struct mapped_device *md)
2204 md->use_blk_mq = false;
2205 dm_init_md_queue(md);
2208 * Initialize aspects of queue that aren't relevant for blk-mq
2210 md->queue->queuedata = md;
2211 md->queue->backing_dev_info.congested_fn = dm_any_congested;
2212 md->queue->backing_dev_info.congested_data = md;
2214 blk_queue_bounce_limit(md->queue, BLK_BOUNCE_ANY);
2217 static void cleanup_mapped_device(struct mapped_device *md)
2220 destroy_workqueue(md->wq);
2221 if (md->kworker_task)
2222 kthread_stop(md->kworker_task);
2224 mempool_destroy(md->io_pool);
2226 mempool_destroy(md->rq_pool);
2228 bioset_free(md->bs);
2230 cleanup_srcu_struct(&md->io_barrier);
2233 spin_lock(&_minor_lock);
2234 md->disk->private_data = NULL;
2235 spin_unlock(&_minor_lock);
2236 if (blk_get_integrity(md->disk))
2237 blk_integrity_unregister(md->disk);
2238 del_gendisk(md->disk);
2243 blk_cleanup_queue(md->queue);
2252 * Allocate and initialise a blank device with a given minor.
2254 static struct mapped_device *alloc_dev(int minor)
2257 struct mapped_device *md = kzalloc(sizeof(*md), GFP_KERNEL);
2261 DMWARN("unable to allocate device, out of memory.");
2265 if (!try_module_get(THIS_MODULE))
2266 goto bad_module_get;
2268 /* get a minor number for the dev */
2269 if (minor == DM_ANY_MINOR)
2270 r = next_free_minor(&minor);
2272 r = specific_minor(minor);
2276 r = init_srcu_struct(&md->io_barrier);
2278 goto bad_io_barrier;
2280 md->use_blk_mq = use_blk_mq;
2281 md->type = DM_TYPE_NONE;
2282 mutex_init(&md->suspend_lock);
2283 mutex_init(&md->type_lock);
2284 mutex_init(&md->table_devices_lock);
2285 spin_lock_init(&md->deferred_lock);
2286 atomic_set(&md->holders, 1);
2287 atomic_set(&md->open_count, 0);
2288 atomic_set(&md->event_nr, 0);
2289 atomic_set(&md->uevent_seq, 0);
2290 INIT_LIST_HEAD(&md->uevent_list);
2291 INIT_LIST_HEAD(&md->table_devices);
2292 spin_lock_init(&md->uevent_lock);
2294 md->queue = blk_alloc_queue(GFP_KERNEL);
2298 dm_init_md_queue(md);
2300 md->disk = alloc_disk(1);
2304 atomic_set(&md->pending[0], 0);
2305 atomic_set(&md->pending[1], 0);
2306 init_waitqueue_head(&md->wait);
2307 INIT_WORK(&md->work, dm_wq_work);
2308 init_waitqueue_head(&md->eventq);
2309 init_completion(&md->kobj_holder.completion);
2310 md->kworker_task = NULL;
2312 md->disk->major = _major;
2313 md->disk->first_minor = minor;
2314 md->disk->fops = &dm_blk_dops;
2315 md->disk->queue = md->queue;
2316 md->disk->private_data = md;
2317 sprintf(md->disk->disk_name, "dm-%d", minor);
2319 format_dev_t(md->name, MKDEV(_major, minor));
2321 md->wq = alloc_workqueue("kdmflush", WQ_MEM_RECLAIM, 0);
2325 md->bdev = bdget_disk(md->disk, 0);
2329 bio_init(&md->flush_bio);
2330 md->flush_bio.bi_bdev = md->bdev;
2331 md->flush_bio.bi_rw = WRITE_FLUSH;
2333 dm_stats_init(&md->stats);
2335 /* Populate the mapping, nobody knows we exist yet */
2336 spin_lock(&_minor_lock);
2337 old_md = idr_replace(&_minor_idr, md, minor);
2338 spin_unlock(&_minor_lock);
2340 BUG_ON(old_md != MINOR_ALLOCED);
2345 cleanup_mapped_device(md);
2349 module_put(THIS_MODULE);
2355 static void unlock_fs(struct mapped_device *md);
2357 static void free_dev(struct mapped_device *md)
2359 int minor = MINOR(disk_devt(md->disk));
2363 cleanup_mapped_device(md);
2365 blk_mq_free_tag_set(&md->tag_set);
2367 free_table_devices(&md->table_devices);
2368 dm_stats_cleanup(&md->stats);
2371 module_put(THIS_MODULE);
2375 static void __bind_mempools(struct mapped_device *md, struct dm_table *t)
2377 struct dm_md_mempools *p = dm_table_get_md_mempools(t);
2380 /* The md already has necessary mempools. */
2381 if (dm_table_get_type(t) == DM_TYPE_BIO_BASED) {
2383 * Reload bioset because front_pad may have changed
2384 * because a different table was loaded.
2386 bioset_free(md->bs);
2391 * There's no need to reload with request-based dm
2392 * because the size of front_pad doesn't change.
2393 * Note for future: If you are to reload bioset,
2394 * prep-ed requests in the queue may refer
2395 * to bio from the old bioset, so you must walk
2396 * through the queue to unprep.
2401 BUG_ON(!p || md->io_pool || md->rq_pool || md->bs);
2403 md->io_pool = p->io_pool;
2405 md->rq_pool = p->rq_pool;
2411 /* mempool bind completed, no longer need any mempools in the table */
2412 dm_table_free_md_mempools(t);
2416 * Bind a table to the device.
2418 static void event_callback(void *context)
2420 unsigned long flags;
2422 struct mapped_device *md = (struct mapped_device *) context;
2424 spin_lock_irqsave(&md->uevent_lock, flags);
2425 list_splice_init(&md->uevent_list, &uevents);
2426 spin_unlock_irqrestore(&md->uevent_lock, flags);
2428 dm_send_uevents(&uevents, &disk_to_dev(md->disk)->kobj);
2430 atomic_inc(&md->event_nr);
2431 wake_up(&md->eventq);
2435 * Protected by md->suspend_lock obtained by dm_swap_table().
2437 static void __set_size(struct mapped_device *md, sector_t size)
2439 set_capacity(md->disk, size);
2441 i_size_write(md->bdev->bd_inode, (loff_t)size << SECTOR_SHIFT);
2445 * Returns old map, which caller must destroy.
2447 static struct dm_table *__bind(struct mapped_device *md, struct dm_table *t,
2448 struct queue_limits *limits)
2450 struct dm_table *old_map;
2451 struct request_queue *q = md->queue;
2454 size = dm_table_get_size(t);
2457 * Wipe any geometry if the size of the table changed.
2459 if (size != dm_get_size(md))
2460 memset(&md->geometry, 0, sizeof(md->geometry));
2462 __set_size(md, size);
2464 dm_table_event_callback(t, event_callback, md);
2467 * The queue hasn't been stopped yet, if the old table type wasn't
2468 * for request-based during suspension. So stop it to prevent
2469 * I/O mapping before resume.
2470 * This must be done before setting the queue restrictions,
2471 * because request-based dm may be run just after the setting.
2473 if (dm_table_request_based(t))
2476 __bind_mempools(md, t);
2478 old_map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2479 rcu_assign_pointer(md->map, t);
2480 md->immutable_target_type = dm_table_get_immutable_target_type(t);
2482 dm_table_set_restrictions(t, q, limits);
2490 * Returns unbound table for the caller to free.
2492 static struct dm_table *__unbind(struct mapped_device *md)
2494 struct dm_table *map = rcu_dereference_protected(md->map, 1);
2499 dm_table_event_callback(map, NULL, NULL);
2500 RCU_INIT_POINTER(md->map, NULL);
2507 * Constructor for a new device.
2509 int dm_create(int minor, struct mapped_device **result)
2511 struct mapped_device *md;
2513 md = alloc_dev(minor);
2524 * Functions to manage md->type.
2525 * All are required to hold md->type_lock.
2527 void dm_lock_md_type(struct mapped_device *md)
2529 mutex_lock(&md->type_lock);
2532 void dm_unlock_md_type(struct mapped_device *md)
2534 mutex_unlock(&md->type_lock);
2537 void dm_set_md_type(struct mapped_device *md, unsigned type)
2539 BUG_ON(!mutex_is_locked(&md->type_lock));
2543 unsigned dm_get_md_type(struct mapped_device *md)
2545 BUG_ON(!mutex_is_locked(&md->type_lock));
2549 struct target_type *dm_get_immutable_target_type(struct mapped_device *md)
2551 return md->immutable_target_type;
2555 * The queue_limits are only valid as long as you have a reference
2558 struct queue_limits *dm_get_queue_limits(struct mapped_device *md)
2560 BUG_ON(!atomic_read(&md->holders));
2561 return &md->queue->limits;
2563 EXPORT_SYMBOL_GPL(dm_get_queue_limits);
2565 static void init_rq_based_worker_thread(struct mapped_device *md)
2567 /* Initialize the request-based DM worker thread */
2568 init_kthread_worker(&md->kworker);
2569 md->kworker_task = kthread_run(kthread_worker_fn, &md->kworker,
2570 "kdmwork-%s", dm_device_name(md));
2574 * Fully initialize a request-based queue (->elevator, ->request_fn, etc).
2576 static int dm_init_request_based_queue(struct mapped_device *md)
2578 struct request_queue *q = NULL;
2580 /* Fully initialize the queue */
2581 q = blk_init_allocated_queue(md->queue, dm_request_fn, NULL);
2585 /* disable dm_request_fn's merge heuristic by default */
2586 md->seq_rq_merge_deadline_usecs = 0;
2589 dm_init_old_md_queue(md);
2590 blk_queue_softirq_done(md->queue, dm_softirq_done);
2591 blk_queue_prep_rq(md->queue, dm_prep_fn);
2593 init_rq_based_worker_thread(md);
2595 elv_register_queue(md->queue);
2600 static int dm_mq_init_request(void *data, struct request *rq,
2601 unsigned int hctx_idx, unsigned int request_idx,
2602 unsigned int numa_node)
2604 struct mapped_device *md = data;
2605 struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
2608 * Must initialize md member of tio, otherwise it won't
2609 * be available in dm_mq_queue_rq.
2616 static int dm_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
2617 const struct blk_mq_queue_data *bd)
2619 struct request *rq = bd->rq;
2620 struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
2621 struct mapped_device *md = tio->md;
2623 struct dm_table *map = dm_get_live_table(md, &srcu_idx);
2624 struct dm_target *ti;
2627 /* always use block 0 to find the target for flushes for now */
2629 if (!(rq->cmd_flags & REQ_FLUSH))
2630 pos = blk_rq_pos(rq);
2632 ti = dm_table_find_target(map, pos);
2633 if (!dm_target_is_valid(ti)) {
2634 dm_put_live_table(md, srcu_idx);
2635 DMERR_LIMIT("request attempted access beyond the end of device");
2637 * Must perform setup, that rq_completed() requires,
2638 * before returning BLK_MQ_RQ_QUEUE_ERROR
2640 dm_start_request(md, rq);
2641 return BLK_MQ_RQ_QUEUE_ERROR;
2643 dm_put_live_table(md, srcu_idx);
2645 if (ti->type->busy && ti->type->busy(ti))
2646 return BLK_MQ_RQ_QUEUE_BUSY;
2648 dm_start_request(md, rq);
2650 /* Init tio using md established in .init_request */
2651 init_tio(tio, rq, md);
2654 * Establish tio->ti before queuing work (map_tio_request)
2655 * or making direct call to map_request().
2659 /* Clone the request if underlying devices aren't blk-mq */
2660 if (dm_table_get_type(map) == DM_TYPE_REQUEST_BASED) {
2661 /* clone request is allocated at the end of the pdu */
2662 tio->clone = (void *)blk_mq_rq_to_pdu(rq) + sizeof(struct dm_rq_target_io);
2663 (void) clone_rq(rq, md, tio, GFP_ATOMIC);
2664 queue_kthread_work(&md->kworker, &tio->work);
2666 /* Direct call is fine since .queue_rq allows allocations */
2667 if (map_request(tio, rq, md) == DM_MAPIO_REQUEUE) {
2668 /* Undo dm_start_request() before requeuing */
2669 rq_end_stats(md, rq);
2670 rq_completed(md, rq_data_dir(rq), false);
2671 return BLK_MQ_RQ_QUEUE_BUSY;
2675 return BLK_MQ_RQ_QUEUE_OK;
2678 static struct blk_mq_ops dm_mq_ops = {
2679 .queue_rq = dm_mq_queue_rq,
2680 .map_queue = blk_mq_map_queue,
2681 .complete = dm_softirq_done,
2682 .init_request = dm_mq_init_request,
2685 static int dm_init_request_based_blk_mq_queue(struct mapped_device *md)
2687 unsigned md_type = dm_get_md_type(md);
2688 struct request_queue *q;
2691 memset(&md->tag_set, 0, sizeof(md->tag_set));
2692 md->tag_set.ops = &dm_mq_ops;
2693 md->tag_set.queue_depth = BLKDEV_MAX_RQ;
2694 md->tag_set.numa_node = NUMA_NO_NODE;
2695 md->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
2696 md->tag_set.nr_hw_queues = 1;
2697 if (md_type == DM_TYPE_REQUEST_BASED) {
2698 /* make the memory for non-blk-mq clone part of the pdu */
2699 md->tag_set.cmd_size = sizeof(struct dm_rq_target_io) + sizeof(struct request);
2701 md->tag_set.cmd_size = sizeof(struct dm_rq_target_io);
2702 md->tag_set.driver_data = md;
2704 err = blk_mq_alloc_tag_set(&md->tag_set);
2708 q = blk_mq_init_allocated_queue(&md->tag_set, md->queue);
2714 dm_init_md_queue(md);
2716 /* backfill 'mq' sysfs registration normally done in blk_register_queue */
2717 blk_mq_register_disk(md->disk);
2719 if (md_type == DM_TYPE_REQUEST_BASED)
2720 init_rq_based_worker_thread(md);
2725 blk_mq_free_tag_set(&md->tag_set);
2729 static unsigned filter_md_type(unsigned type, struct mapped_device *md)
2731 if (type == DM_TYPE_BIO_BASED)
2734 return !md->use_blk_mq ? DM_TYPE_REQUEST_BASED : DM_TYPE_MQ_REQUEST_BASED;
2738 * Setup the DM device's queue based on md's type
2740 int dm_setup_md_queue(struct mapped_device *md)
2743 unsigned md_type = filter_md_type(dm_get_md_type(md), md);
2746 case DM_TYPE_REQUEST_BASED:
2747 r = dm_init_request_based_queue(md);
2749 DMWARN("Cannot initialize queue for request-based mapped device");
2753 case DM_TYPE_MQ_REQUEST_BASED:
2754 r = dm_init_request_based_blk_mq_queue(md);
2756 DMWARN("Cannot initialize queue for request-based blk-mq mapped device");
2760 case DM_TYPE_BIO_BASED:
2761 dm_init_old_md_queue(md);
2762 blk_queue_make_request(md->queue, dm_make_request);
2769 struct mapped_device *dm_get_md(dev_t dev)
2771 struct mapped_device *md;
2772 unsigned minor = MINOR(dev);
2774 if (MAJOR(dev) != _major || minor >= (1 << MINORBITS))
2777 spin_lock(&_minor_lock);
2779 md = idr_find(&_minor_idr, minor);
2781 if ((md == MINOR_ALLOCED ||
2782 (MINOR(disk_devt(dm_disk(md))) != minor) ||
2783 dm_deleting_md(md) ||
2784 test_bit(DMF_FREEING, &md->flags))) {
2792 spin_unlock(&_minor_lock);
2796 EXPORT_SYMBOL_GPL(dm_get_md);
2798 void *dm_get_mdptr(struct mapped_device *md)
2800 return md->interface_ptr;
2803 void dm_set_mdptr(struct mapped_device *md, void *ptr)
2805 md->interface_ptr = ptr;
2808 void dm_get(struct mapped_device *md)
2810 atomic_inc(&md->holders);
2811 BUG_ON(test_bit(DMF_FREEING, &md->flags));
2814 int dm_hold(struct mapped_device *md)
2816 spin_lock(&_minor_lock);
2817 if (test_bit(DMF_FREEING, &md->flags)) {
2818 spin_unlock(&_minor_lock);
2822 spin_unlock(&_minor_lock);
2825 EXPORT_SYMBOL_GPL(dm_hold);
2827 const char *dm_device_name(struct mapped_device *md)
2831 EXPORT_SYMBOL_GPL(dm_device_name);
2833 static void __dm_destroy(struct mapped_device *md, bool wait)
2835 struct dm_table *map;
2840 map = dm_get_live_table(md, &srcu_idx);
2842 spin_lock(&_minor_lock);
2843 idr_replace(&_minor_idr, MINOR_ALLOCED, MINOR(disk_devt(dm_disk(md))));
2844 set_bit(DMF_FREEING, &md->flags);
2845 spin_unlock(&_minor_lock);
2847 if (dm_request_based(md) && md->kworker_task)
2848 flush_kthread_worker(&md->kworker);
2851 * Take suspend_lock so that presuspend and postsuspend methods
2852 * do not race with internal suspend.
2854 mutex_lock(&md->suspend_lock);
2855 if (!dm_suspended_md(md)) {
2856 dm_table_presuspend_targets(map);
2857 dm_table_postsuspend_targets(map);
2859 mutex_unlock(&md->suspend_lock);
2861 /* dm_put_live_table must be before msleep, otherwise deadlock is possible */
2862 dm_put_live_table(md, srcu_idx);
2865 * Rare, but there may be I/O requests still going to complete,
2866 * for example. Wait for all references to disappear.
2867 * No one should increment the reference count of the mapped_device,
2868 * after the mapped_device state becomes DMF_FREEING.
2871 while (atomic_read(&md->holders))
2873 else if (atomic_read(&md->holders))
2874 DMWARN("%s: Forcibly removing mapped_device still in use! (%d users)",
2875 dm_device_name(md), atomic_read(&md->holders));
2878 dm_table_destroy(__unbind(md));
2882 void dm_destroy(struct mapped_device *md)
2884 __dm_destroy(md, true);
2887 void dm_destroy_immediate(struct mapped_device *md)
2889 __dm_destroy(md, false);
2892 void dm_put(struct mapped_device *md)
2894 atomic_dec(&md->holders);
2896 EXPORT_SYMBOL_GPL(dm_put);
2898 static int dm_wait_for_completion(struct mapped_device *md, int interruptible)
2901 DECLARE_WAITQUEUE(wait, current);
2903 add_wait_queue(&md->wait, &wait);
2906 set_current_state(interruptible);
2908 if (!md_in_flight(md))
2911 if (interruptible == TASK_INTERRUPTIBLE &&
2912 signal_pending(current)) {
2919 set_current_state(TASK_RUNNING);
2921 remove_wait_queue(&md->wait, &wait);
2927 * Process the deferred bios
2929 static void dm_wq_work(struct work_struct *work)
2931 struct mapped_device *md = container_of(work, struct mapped_device,
2935 struct dm_table *map;
2937 map = dm_get_live_table(md, &srcu_idx);
2939 while (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
2940 spin_lock_irq(&md->deferred_lock);
2941 c = bio_list_pop(&md->deferred);
2942 spin_unlock_irq(&md->deferred_lock);
2947 if (dm_request_based(md))
2948 generic_make_request(c);
2950 __split_and_process_bio(md, map, c);
2953 dm_put_live_table(md, srcu_idx);
2956 static void dm_queue_flush(struct mapped_device *md)
2958 clear_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2959 smp_mb__after_atomic();
2960 queue_work(md->wq, &md->work);
2964 * Swap in a new table, returning the old one for the caller to destroy.
2966 struct dm_table *dm_swap_table(struct mapped_device *md, struct dm_table *table)
2968 struct dm_table *live_map = NULL, *map = ERR_PTR(-EINVAL);
2969 struct queue_limits limits;
2972 mutex_lock(&md->suspend_lock);
2974 /* device must be suspended */
2975 if (!dm_suspended_md(md))
2979 * If the new table has no data devices, retain the existing limits.
2980 * This helps multipath with queue_if_no_path if all paths disappear,
2981 * then new I/O is queued based on these limits, and then some paths
2984 if (dm_table_has_no_data_devices(table)) {
2985 live_map = dm_get_live_table_fast(md);
2987 limits = md->queue->limits;
2988 dm_put_live_table_fast(md);
2992 r = dm_calculate_queue_limits(table, &limits);
2999 map = __bind(md, table, &limits);
3002 mutex_unlock(&md->suspend_lock);
3007 * Functions to lock and unlock any filesystem running on the
3010 static int lock_fs(struct mapped_device *md)
3014 WARN_ON(md->frozen_sb);
3016 md->frozen_sb = freeze_bdev(md->bdev);
3017 if (IS_ERR(md->frozen_sb)) {
3018 r = PTR_ERR(md->frozen_sb);
3019 md->frozen_sb = NULL;
3023 set_bit(DMF_FROZEN, &md->flags);
3028 static void unlock_fs(struct mapped_device *md)
3030 if (!test_bit(DMF_FROZEN, &md->flags))
3033 thaw_bdev(md->bdev, md->frozen_sb);
3034 md->frozen_sb = NULL;
3035 clear_bit(DMF_FROZEN, &md->flags);
3039 * If __dm_suspend returns 0, the device is completely quiescent
3040 * now. There is no request-processing activity. All new requests
3041 * are being added to md->deferred list.
3043 * Caller must hold md->suspend_lock
3045 static int __dm_suspend(struct mapped_device *md, struct dm_table *map,
3046 unsigned suspend_flags, int interruptible)
3048 bool do_lockfs = suspend_flags & DM_SUSPEND_LOCKFS_FLAG;
3049 bool noflush = suspend_flags & DM_SUSPEND_NOFLUSH_FLAG;
3053 * DMF_NOFLUSH_SUSPENDING must be set before presuspend.
3054 * This flag is cleared before dm_suspend returns.
3057 set_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
3060 * This gets reverted if there's an error later and the targets
3061 * provide the .presuspend_undo hook.
3063 dm_table_presuspend_targets(map);
3066 * Flush I/O to the device.
3067 * Any I/O submitted after lock_fs() may not be flushed.
3068 * noflush takes precedence over do_lockfs.
3069 * (lock_fs() flushes I/Os and waits for them to complete.)
3071 if (!noflush && do_lockfs) {
3074 dm_table_presuspend_undo_targets(map);
3080 * Here we must make sure that no processes are submitting requests
3081 * to target drivers i.e. no one may be executing
3082 * __split_and_process_bio. This is called from dm_request and
3085 * To get all processes out of __split_and_process_bio in dm_request,
3086 * we take the write lock. To prevent any process from reentering
3087 * __split_and_process_bio from dm_request and quiesce the thread
3088 * (dm_wq_work), we set BMF_BLOCK_IO_FOR_SUSPEND and call
3089 * flush_workqueue(md->wq).
3091 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
3093 synchronize_srcu(&md->io_barrier);
3096 * Stop md->queue before flushing md->wq in case request-based
3097 * dm defers requests to md->wq from md->queue.
3099 if (dm_request_based(md)) {
3100 stop_queue(md->queue);
3101 if (md->kworker_task)
3102 flush_kthread_worker(&md->kworker);
3105 flush_workqueue(md->wq);
3108 * At this point no more requests are entering target request routines.
3109 * We call dm_wait_for_completion to wait for all existing requests
3112 r = dm_wait_for_completion(md, interruptible);
3115 clear_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
3117 synchronize_srcu(&md->io_barrier);
3119 /* were we interrupted ? */
3123 if (dm_request_based(md))
3124 start_queue(md->queue);
3127 dm_table_presuspend_undo_targets(map);
3128 /* pushback list is already flushed, so skip flush */
3135 * We need to be able to change a mapping table under a mounted
3136 * filesystem. For example we might want to move some data in
3137 * the background. Before the table can be swapped with
3138 * dm_bind_table, dm_suspend must be called to flush any in
3139 * flight bios and ensure that any further io gets deferred.
3142 * Suspend mechanism in request-based dm.
3144 * 1. Flush all I/Os by lock_fs() if needed.
3145 * 2. Stop dispatching any I/O by stopping the request_queue.
3146 * 3. Wait for all in-flight I/Os to be completed or requeued.
3148 * To abort suspend, start the request_queue.
3150 int dm_suspend(struct mapped_device *md, unsigned suspend_flags)
3152 struct dm_table *map = NULL;
3156 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
3158 if (dm_suspended_md(md)) {
3163 if (dm_suspended_internally_md(md)) {
3164 /* already internally suspended, wait for internal resume */
3165 mutex_unlock(&md->suspend_lock);
3166 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
3172 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3174 r = __dm_suspend(md, map, suspend_flags, TASK_INTERRUPTIBLE);
3178 set_bit(DMF_SUSPENDED, &md->flags);
3180 dm_table_postsuspend_targets(map);
3183 mutex_unlock(&md->suspend_lock);
3187 static int __dm_resume(struct mapped_device *md, struct dm_table *map)
3190 int r = dm_table_resume_targets(map);
3198 * Flushing deferred I/Os must be done after targets are resumed
3199 * so that mapping of targets can work correctly.
3200 * Request-based dm is queueing the deferred I/Os in its request_queue.
3202 if (dm_request_based(md))
3203 start_queue(md->queue);
3210 int dm_resume(struct mapped_device *md)
3213 struct dm_table *map = NULL;
3216 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
3218 if (!dm_suspended_md(md))
3221 if (dm_suspended_internally_md(md)) {
3222 /* already internally suspended, wait for internal resume */
3223 mutex_unlock(&md->suspend_lock);
3224 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
3230 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3231 if (!map || !dm_table_get_size(map))
3234 r = __dm_resume(md, map);
3238 clear_bit(DMF_SUSPENDED, &md->flags);
3242 mutex_unlock(&md->suspend_lock);
3248 * Internal suspend/resume works like userspace-driven suspend. It waits
3249 * until all bios finish and prevents issuing new bios to the target drivers.
3250 * It may be used only from the kernel.
3253 static void __dm_internal_suspend(struct mapped_device *md, unsigned suspend_flags)
3255 struct dm_table *map = NULL;
3257 if (md->internal_suspend_count++)
3258 return; /* nested internal suspend */
3260 if (dm_suspended_md(md)) {
3261 set_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3262 return; /* nest suspend */
3265 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3268 * Using TASK_UNINTERRUPTIBLE because only NOFLUSH internal suspend is
3269 * supported. Properly supporting a TASK_INTERRUPTIBLE internal suspend
3270 * would require changing .presuspend to return an error -- avoid this
3271 * until there is a need for more elaborate variants of internal suspend.
3273 (void) __dm_suspend(md, map, suspend_flags, TASK_UNINTERRUPTIBLE);
3275 set_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3277 dm_table_postsuspend_targets(map);
3280 static void __dm_internal_resume(struct mapped_device *md)
3282 BUG_ON(!md->internal_suspend_count);
3284 if (--md->internal_suspend_count)
3285 return; /* resume from nested internal suspend */
3287 if (dm_suspended_md(md))
3288 goto done; /* resume from nested suspend */
3291 * NOTE: existing callers don't need to call dm_table_resume_targets
3292 * (which may fail -- so best to avoid it for now by passing NULL map)
3294 (void) __dm_resume(md, NULL);
3297 clear_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3298 smp_mb__after_atomic();
3299 wake_up_bit(&md->flags, DMF_SUSPENDED_INTERNALLY);
3302 void dm_internal_suspend_noflush(struct mapped_device *md)
3304 mutex_lock(&md->suspend_lock);
3305 __dm_internal_suspend(md, DM_SUSPEND_NOFLUSH_FLAG);
3306 mutex_unlock(&md->suspend_lock);
3308 EXPORT_SYMBOL_GPL(dm_internal_suspend_noflush);
3310 void dm_internal_resume(struct mapped_device *md)
3312 mutex_lock(&md->suspend_lock);
3313 __dm_internal_resume(md);
3314 mutex_unlock(&md->suspend_lock);
3316 EXPORT_SYMBOL_GPL(dm_internal_resume);
3319 * Fast variants of internal suspend/resume hold md->suspend_lock,
3320 * which prevents interaction with userspace-driven suspend.
3323 void dm_internal_suspend_fast(struct mapped_device *md)
3325 mutex_lock(&md->suspend_lock);
3326 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
3329 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
3330 synchronize_srcu(&md->io_barrier);
3331 flush_workqueue(md->wq);
3332 dm_wait_for_completion(md, TASK_UNINTERRUPTIBLE);
3334 EXPORT_SYMBOL_GPL(dm_internal_suspend_fast);
3336 void dm_internal_resume_fast(struct mapped_device *md)
3338 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
3344 mutex_unlock(&md->suspend_lock);
3346 EXPORT_SYMBOL_GPL(dm_internal_resume_fast);
3348 /*-----------------------------------------------------------------
3349 * Event notification.
3350 *---------------------------------------------------------------*/
3351 int dm_kobject_uevent(struct mapped_device *md, enum kobject_action action,
3354 char udev_cookie[DM_COOKIE_LENGTH];
3355 char *envp[] = { udev_cookie, NULL };
3358 return kobject_uevent(&disk_to_dev(md->disk)->kobj, action);
3360 snprintf(udev_cookie, DM_COOKIE_LENGTH, "%s=%u",
3361 DM_COOKIE_ENV_VAR_NAME, cookie);
3362 return kobject_uevent_env(&disk_to_dev(md->disk)->kobj,
3367 uint32_t dm_next_uevent_seq(struct mapped_device *md)
3369 return atomic_add_return(1, &md->uevent_seq);
3372 uint32_t dm_get_event_nr(struct mapped_device *md)
3374 return atomic_read(&md->event_nr);
3377 int dm_wait_event(struct mapped_device *md, int event_nr)
3379 return wait_event_interruptible(md->eventq,
3380 (event_nr != atomic_read(&md->event_nr)));
3383 void dm_uevent_add(struct mapped_device *md, struct list_head *elist)
3385 unsigned long flags;
3387 spin_lock_irqsave(&md->uevent_lock, flags);
3388 list_add(elist, &md->uevent_list);
3389 spin_unlock_irqrestore(&md->uevent_lock, flags);
3393 * The gendisk is only valid as long as you have a reference
3396 struct gendisk *dm_disk(struct mapped_device *md)
3400 EXPORT_SYMBOL_GPL(dm_disk);
3402 struct kobject *dm_kobject(struct mapped_device *md)
3404 return &md->kobj_holder.kobj;
3407 struct mapped_device *dm_get_from_kobject(struct kobject *kobj)
3409 struct mapped_device *md;
3411 md = container_of(kobj, struct mapped_device, kobj_holder.kobj);
3413 if (test_bit(DMF_FREEING, &md->flags) ||
3421 int dm_suspended_md(struct mapped_device *md)
3423 return test_bit(DMF_SUSPENDED, &md->flags);
3426 int dm_suspended_internally_md(struct mapped_device *md)
3428 return test_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3431 int dm_test_deferred_remove_flag(struct mapped_device *md)
3433 return test_bit(DMF_DEFERRED_REMOVE, &md->flags);
3436 int dm_suspended(struct dm_target *ti)
3438 return dm_suspended_md(dm_table_get_md(ti->table));
3440 EXPORT_SYMBOL_GPL(dm_suspended);
3442 int dm_noflush_suspending(struct dm_target *ti)
3444 return __noflush_suspending(dm_table_get_md(ti->table));
3446 EXPORT_SYMBOL_GPL(dm_noflush_suspending);
3448 struct dm_md_mempools *dm_alloc_md_mempools(struct mapped_device *md, unsigned type,
3449 unsigned integrity, unsigned per_bio_data_size)
3451 struct dm_md_mempools *pools = kzalloc(sizeof(*pools), GFP_KERNEL);
3452 struct kmem_cache *cachep = NULL;
3453 unsigned int pool_size = 0;
3454 unsigned int front_pad;
3459 type = filter_md_type(type, md);
3462 case DM_TYPE_BIO_BASED:
3464 pool_size = dm_get_reserved_bio_based_ios();
3465 front_pad = roundup(per_bio_data_size, __alignof__(struct dm_target_io)) + offsetof(struct dm_target_io, clone);
3467 case DM_TYPE_REQUEST_BASED:
3468 cachep = _rq_tio_cache;
3469 pool_size = dm_get_reserved_rq_based_ios();
3470 pools->rq_pool = mempool_create_slab_pool(pool_size, _rq_cache);
3471 if (!pools->rq_pool)
3473 /* fall through to setup remaining rq-based pools */
3474 case DM_TYPE_MQ_REQUEST_BASED:
3476 pool_size = dm_get_reserved_rq_based_ios();
3477 front_pad = offsetof(struct dm_rq_clone_bio_info, clone);
3478 /* per_bio_data_size is not used. See __bind_mempools(). */
3479 WARN_ON(per_bio_data_size != 0);
3486 pools->io_pool = mempool_create_slab_pool(pool_size, cachep);
3487 if (!pools->io_pool)
3491 pools->bs = bioset_create_nobvec(pool_size, front_pad);
3495 if (integrity && bioset_integrity_create(pools->bs, pool_size))
3501 dm_free_md_mempools(pools);
3506 void dm_free_md_mempools(struct dm_md_mempools *pools)
3512 mempool_destroy(pools->io_pool);
3515 mempool_destroy(pools->rq_pool);
3518 bioset_free(pools->bs);
3523 static const struct block_device_operations dm_blk_dops = {
3524 .open = dm_blk_open,
3525 .release = dm_blk_close,
3526 .ioctl = dm_blk_ioctl,
3527 .getgeo = dm_blk_getgeo,
3528 .owner = THIS_MODULE
3534 module_init(dm_init);
3535 module_exit(dm_exit);
3537 module_param(major, uint, 0);
3538 MODULE_PARM_DESC(major, "The major number of the device mapper");
3540 module_param(reserved_bio_based_ios, uint, S_IRUGO | S_IWUSR);
3541 MODULE_PARM_DESC(reserved_bio_based_ios, "Reserved IOs in bio-based mempools");
3543 module_param(reserved_rq_based_ios, uint, S_IRUGO | S_IWUSR);
3544 MODULE_PARM_DESC(reserved_rq_based_ios, "Reserved IOs in request-based mempools");
3546 module_param(use_blk_mq, bool, S_IRUGO | S_IWUSR);
3547 MODULE_PARM_DESC(use_blk_mq, "Use block multiqueue for request-based DM devices");
3549 MODULE_DESCRIPTION(DM_NAME " driver");
3550 MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
3551 MODULE_LICENSE("GPL");