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>
29 #include <trace/events/block.h>
31 #define DM_MSG_PREFIX "core"
35 * ratelimit state to be used in DMXXX_LIMIT().
37 DEFINE_RATELIMIT_STATE(dm_ratelimit_state,
38 DEFAULT_RATELIMIT_INTERVAL,
39 DEFAULT_RATELIMIT_BURST);
40 EXPORT_SYMBOL(dm_ratelimit_state);
44 * Cookies are numeric values sent with CHANGE and REMOVE
45 * uevents while resuming, removing or renaming the device.
47 #define DM_COOKIE_ENV_VAR_NAME "DM_COOKIE"
48 #define DM_COOKIE_LENGTH 24
50 static const char *_name = DM_NAME;
52 static unsigned int major = 0;
53 static unsigned int _major = 0;
55 static DEFINE_IDR(_minor_idr);
57 static DEFINE_SPINLOCK(_minor_lock);
59 static void do_deferred_remove(struct work_struct *w);
61 static DECLARE_WORK(deferred_remove_work, do_deferred_remove);
63 static struct workqueue_struct *deferred_remove_workqueue;
67 * One of these is allocated per bio.
70 struct mapped_device *md;
74 unsigned long start_time;
75 spinlock_t endio_lock;
76 struct dm_stats_aux stats_aux;
80 * For request-based dm.
81 * One of these is allocated per request.
83 struct dm_rq_target_io {
84 struct mapped_device *md;
86 struct request *orig, *clone;
87 struct kthread_work work;
90 struct dm_stats_aux stats_aux;
91 unsigned long duration_jiffies;
96 * For request-based dm - the bio clones we allocate are embedded in these
99 * We allocate these with bio_alloc_bioset, using the front_pad parameter when
100 * the bioset is created - this means the bio has to come at the end of the
103 struct dm_rq_clone_bio_info {
105 struct dm_rq_target_io *tio;
109 union map_info *dm_get_rq_mapinfo(struct request *rq)
111 if (rq && rq->end_io_data)
112 return &((struct dm_rq_target_io *)rq->end_io_data)->info;
115 EXPORT_SYMBOL_GPL(dm_get_rq_mapinfo);
117 #define MINOR_ALLOCED ((void *)-1)
120 * Bits for the md->flags field.
122 #define DMF_BLOCK_IO_FOR_SUSPEND 0
123 #define DMF_SUSPENDED 1
125 #define DMF_FREEING 3
126 #define DMF_DELETING 4
127 #define DMF_NOFLUSH_SUSPENDING 5
128 #define DMF_DEFERRED_REMOVE 6
129 #define DMF_SUSPENDED_INTERNALLY 7
132 * A dummy definition to make RCU happy.
133 * struct dm_table should never be dereferenced in this file.
140 * Work processed by per-device workqueue.
142 struct mapped_device {
143 struct srcu_struct io_barrier;
144 struct mutex suspend_lock;
149 * The current mapping.
150 * Use dm_get_live_table{_fast} or take suspend_lock for
153 struct dm_table __rcu *map;
155 struct list_head table_devices;
156 struct mutex table_devices_lock;
160 struct request_queue *queue;
162 /* Protect queue and type against concurrent access. */
163 struct mutex type_lock;
165 struct target_type *immutable_target_type;
167 struct gendisk *disk;
173 * A list of ios that arrived while we were suspended.
176 wait_queue_head_t wait;
177 struct work_struct work;
178 struct bio_list deferred;
179 spinlock_t deferred_lock;
182 * Processing queue (flush)
184 struct workqueue_struct *wq;
187 * io objects are allocated from here.
198 wait_queue_head_t eventq;
200 struct list_head uevent_list;
201 spinlock_t uevent_lock; /* Protect access to uevent_list */
204 * freeze/thaw support require holding onto a super block
206 struct super_block *frozen_sb;
207 struct block_device *bdev;
209 /* forced geometry settings */
210 struct hd_geometry geometry;
212 /* kobject and completion */
213 struct dm_kobject_holder kobj_holder;
215 /* zero-length flush that will be cloned and submitted to targets */
216 struct bio flush_bio;
218 /* the number of internal suspends */
219 unsigned internal_suspend_count;
221 struct dm_stats stats;
223 struct kthread_worker kworker;
224 struct task_struct *kworker_task;
226 /* for request-based merge heuristic in dm_request_fn() */
227 unsigned seq_rq_merge_deadline_usecs;
229 sector_t last_rq_pos;
230 ktime_t last_rq_start_time;
232 /* for blk-mq request-based DM support */
233 struct blk_mq_tag_set tag_set;
237 #ifdef CONFIG_DM_MQ_DEFAULT
238 static bool use_blk_mq = true;
240 static bool use_blk_mq = false;
243 bool dm_use_blk_mq(struct mapped_device *md)
245 return md->use_blk_mq;
249 * For mempools pre-allocation at the table loading time.
251 struct dm_md_mempools {
257 struct table_device {
258 struct list_head list;
260 struct dm_dev dm_dev;
263 #define RESERVED_BIO_BASED_IOS 16
264 #define RESERVED_REQUEST_BASED_IOS 256
265 #define RESERVED_MAX_IOS 1024
266 static struct kmem_cache *_io_cache;
267 static struct kmem_cache *_rq_tio_cache;
268 static struct kmem_cache *_rq_cache;
271 * Bio-based DM's mempools' reserved IOs set by the user.
273 static unsigned reserved_bio_based_ios = RESERVED_BIO_BASED_IOS;
276 * Request-based DM's mempools' reserved IOs set by the user.
278 static unsigned reserved_rq_based_ios = RESERVED_REQUEST_BASED_IOS;
280 static unsigned __dm_get_module_param(unsigned *module_param,
281 unsigned def, unsigned max)
283 unsigned param = ACCESS_ONCE(*module_param);
284 unsigned modified_param = 0;
287 modified_param = def;
288 else if (param > max)
289 modified_param = max;
291 if (modified_param) {
292 (void)cmpxchg(module_param, param, modified_param);
293 param = modified_param;
299 unsigned dm_get_reserved_bio_based_ios(void)
301 return __dm_get_module_param(&reserved_bio_based_ios,
302 RESERVED_BIO_BASED_IOS, RESERVED_MAX_IOS);
304 EXPORT_SYMBOL_GPL(dm_get_reserved_bio_based_ios);
306 unsigned dm_get_reserved_rq_based_ios(void)
308 return __dm_get_module_param(&reserved_rq_based_ios,
309 RESERVED_REQUEST_BASED_IOS, RESERVED_MAX_IOS);
311 EXPORT_SYMBOL_GPL(dm_get_reserved_rq_based_ios);
313 static int __init local_init(void)
317 /* allocate a slab for the dm_ios */
318 _io_cache = KMEM_CACHE(dm_io, 0);
322 _rq_tio_cache = KMEM_CACHE(dm_rq_target_io, 0);
324 goto out_free_io_cache;
326 _rq_cache = kmem_cache_create("dm_clone_request", sizeof(struct request),
327 __alignof__(struct request), 0, NULL);
329 goto out_free_rq_tio_cache;
331 r = dm_uevent_init();
333 goto out_free_rq_cache;
335 deferred_remove_workqueue = alloc_workqueue("kdmremove", WQ_UNBOUND, 1);
336 if (!deferred_remove_workqueue) {
338 goto out_uevent_exit;
342 r = register_blkdev(_major, _name);
344 goto out_free_workqueue;
352 destroy_workqueue(deferred_remove_workqueue);
356 kmem_cache_destroy(_rq_cache);
357 out_free_rq_tio_cache:
358 kmem_cache_destroy(_rq_tio_cache);
360 kmem_cache_destroy(_io_cache);
365 static void local_exit(void)
367 flush_scheduled_work();
368 destroy_workqueue(deferred_remove_workqueue);
370 kmem_cache_destroy(_rq_cache);
371 kmem_cache_destroy(_rq_tio_cache);
372 kmem_cache_destroy(_io_cache);
373 unregister_blkdev(_major, _name);
378 DMINFO("cleaned up");
381 static int (*_inits[])(void) __initdata = {
392 static void (*_exits[])(void) = {
403 static int __init dm_init(void)
405 const int count = ARRAY_SIZE(_inits);
409 for (i = 0; i < count; i++) {
424 static void __exit dm_exit(void)
426 int i = ARRAY_SIZE(_exits);
432 * Should be empty by this point.
434 idr_destroy(&_minor_idr);
438 * Block device functions
440 int dm_deleting_md(struct mapped_device *md)
442 return test_bit(DMF_DELETING, &md->flags);
445 static int dm_blk_open(struct block_device *bdev, fmode_t mode)
447 struct mapped_device *md;
449 spin_lock(&_minor_lock);
451 md = bdev->bd_disk->private_data;
455 if (test_bit(DMF_FREEING, &md->flags) ||
456 dm_deleting_md(md)) {
462 atomic_inc(&md->open_count);
464 spin_unlock(&_minor_lock);
466 return md ? 0 : -ENXIO;
469 static void dm_blk_close(struct gendisk *disk, fmode_t mode)
471 struct mapped_device *md;
473 spin_lock(&_minor_lock);
475 md = disk->private_data;
479 if (atomic_dec_and_test(&md->open_count) &&
480 (test_bit(DMF_DEFERRED_REMOVE, &md->flags)))
481 queue_work(deferred_remove_workqueue, &deferred_remove_work);
485 spin_unlock(&_minor_lock);
488 int dm_open_count(struct mapped_device *md)
490 return atomic_read(&md->open_count);
494 * Guarantees nothing is using the device before it's deleted.
496 int dm_lock_for_deletion(struct mapped_device *md, bool mark_deferred, bool only_deferred)
500 spin_lock(&_minor_lock);
502 if (dm_open_count(md)) {
505 set_bit(DMF_DEFERRED_REMOVE, &md->flags);
506 } else if (only_deferred && !test_bit(DMF_DEFERRED_REMOVE, &md->flags))
509 set_bit(DMF_DELETING, &md->flags);
511 spin_unlock(&_minor_lock);
516 int dm_cancel_deferred_remove(struct mapped_device *md)
520 spin_lock(&_minor_lock);
522 if (test_bit(DMF_DELETING, &md->flags))
525 clear_bit(DMF_DEFERRED_REMOVE, &md->flags);
527 spin_unlock(&_minor_lock);
532 static void do_deferred_remove(struct work_struct *w)
534 dm_deferred_remove();
537 sector_t dm_get_size(struct mapped_device *md)
539 return get_capacity(md->disk);
542 struct request_queue *dm_get_md_queue(struct mapped_device *md)
547 struct dm_stats *dm_get_stats(struct mapped_device *md)
552 static int dm_blk_getgeo(struct block_device *bdev, struct hd_geometry *geo)
554 struct mapped_device *md = bdev->bd_disk->private_data;
556 return dm_get_geometry(md, geo);
559 static int dm_get_live_table_for_ioctl(struct mapped_device *md,
560 struct dm_target **tgt, struct block_device **bdev,
561 fmode_t *mode, int *srcu_idx)
563 struct dm_table *map;
568 map = dm_get_live_table(md, srcu_idx);
569 if (!map || !dm_table_get_size(map))
572 /* We only support devices that have a single target */
573 if (dm_table_get_num_targets(map) != 1)
576 *tgt = dm_table_get_target(map, 0);
578 if (!(*tgt)->type->prepare_ioctl)
581 if (dm_suspended_md(md)) {
586 r = (*tgt)->type->prepare_ioctl(*tgt, bdev, mode);
593 dm_put_live_table(md, *srcu_idx);
594 if (r == -ENOTCONN && !fatal_signal_pending(current)) {
601 static int dm_blk_ioctl(struct block_device *bdev, fmode_t mode,
602 unsigned int cmd, unsigned long arg)
604 struct mapped_device *md = bdev->bd_disk->private_data;
605 struct dm_target *tgt;
606 struct block_device *tgt_bdev = NULL;
609 r = dm_get_live_table_for_ioctl(md, &tgt, &tgt_bdev, &mode, &srcu_idx);
615 * Target determined this ioctl is being issued against
616 * a logical partition of the parent bdev; so extra
617 * validation is needed.
619 r = scsi_verify_blk_ioctl(NULL, cmd);
624 r = __blkdev_driver_ioctl(tgt_bdev, mode, cmd, arg);
626 dm_put_live_table(md, srcu_idx);
630 static struct dm_io *alloc_io(struct mapped_device *md)
632 return mempool_alloc(md->io_pool, GFP_NOIO);
635 static void free_io(struct mapped_device *md, struct dm_io *io)
637 mempool_free(io, md->io_pool);
640 static void free_tio(struct mapped_device *md, struct dm_target_io *tio)
642 bio_put(&tio->clone);
645 static struct dm_rq_target_io *alloc_rq_tio(struct mapped_device *md,
648 return mempool_alloc(md->io_pool, gfp_mask);
651 static void free_rq_tio(struct dm_rq_target_io *tio)
653 mempool_free(tio, tio->md->io_pool);
656 static struct request *alloc_clone_request(struct mapped_device *md,
659 return mempool_alloc(md->rq_pool, gfp_mask);
662 static void free_clone_request(struct mapped_device *md, struct request *rq)
664 mempool_free(rq, md->rq_pool);
667 static int md_in_flight(struct mapped_device *md)
669 return atomic_read(&md->pending[READ]) +
670 atomic_read(&md->pending[WRITE]);
673 static void start_io_acct(struct dm_io *io)
675 struct mapped_device *md = io->md;
676 struct bio *bio = io->bio;
678 int rw = bio_data_dir(bio);
680 io->start_time = jiffies;
682 cpu = part_stat_lock();
683 part_round_stats(cpu, &dm_disk(md)->part0);
685 atomic_set(&dm_disk(md)->part0.in_flight[rw],
686 atomic_inc_return(&md->pending[rw]));
688 if (unlikely(dm_stats_used(&md->stats)))
689 dm_stats_account_io(&md->stats, bio->bi_rw, bio->bi_iter.bi_sector,
690 bio_sectors(bio), false, 0, &io->stats_aux);
693 static void end_io_acct(struct dm_io *io)
695 struct mapped_device *md = io->md;
696 struct bio *bio = io->bio;
697 unsigned long duration = jiffies - io->start_time;
699 int rw = bio_data_dir(bio);
701 generic_end_io_acct(rw, &dm_disk(md)->part0, io->start_time);
703 if (unlikely(dm_stats_used(&md->stats)))
704 dm_stats_account_io(&md->stats, bio->bi_rw, bio->bi_iter.bi_sector,
705 bio_sectors(bio), true, duration, &io->stats_aux);
708 * After this is decremented the bio must not be touched if it is
711 pending = atomic_dec_return(&md->pending[rw]);
712 atomic_set(&dm_disk(md)->part0.in_flight[rw], pending);
713 pending += atomic_read(&md->pending[rw^0x1]);
715 /* nudge anyone waiting on suspend queue */
721 * Add the bio to the list of deferred io.
723 static void queue_io(struct mapped_device *md, struct bio *bio)
727 spin_lock_irqsave(&md->deferred_lock, flags);
728 bio_list_add(&md->deferred, bio);
729 spin_unlock_irqrestore(&md->deferred_lock, flags);
730 queue_work(md->wq, &md->work);
734 * Everyone (including functions in this file), should use this
735 * function to access the md->map field, and make sure they call
736 * dm_put_live_table() when finished.
738 struct dm_table *dm_get_live_table(struct mapped_device *md, int *srcu_idx) __acquires(md->io_barrier)
740 *srcu_idx = srcu_read_lock(&md->io_barrier);
742 return srcu_dereference(md->map, &md->io_barrier);
745 void dm_put_live_table(struct mapped_device *md, int srcu_idx) __releases(md->io_barrier)
747 srcu_read_unlock(&md->io_barrier, srcu_idx);
750 void dm_sync_table(struct mapped_device *md)
752 synchronize_srcu(&md->io_barrier);
753 synchronize_rcu_expedited();
757 * A fast alternative to dm_get_live_table/dm_put_live_table.
758 * The caller must not block between these two functions.
760 static struct dm_table *dm_get_live_table_fast(struct mapped_device *md) __acquires(RCU)
763 return rcu_dereference(md->map);
766 static void dm_put_live_table_fast(struct mapped_device *md) __releases(RCU)
772 * Open a table device so we can use it as a map destination.
774 static int open_table_device(struct table_device *td, dev_t dev,
775 struct mapped_device *md)
777 static char *_claim_ptr = "I belong to device-mapper";
778 struct block_device *bdev;
782 BUG_ON(td->dm_dev.bdev);
784 bdev = blkdev_get_by_dev(dev, td->dm_dev.mode | FMODE_EXCL, _claim_ptr);
786 return PTR_ERR(bdev);
788 r = bd_link_disk_holder(bdev, dm_disk(md));
790 blkdev_put(bdev, td->dm_dev.mode | FMODE_EXCL);
794 td->dm_dev.bdev = bdev;
799 * Close a table device that we've been using.
801 static void close_table_device(struct table_device *td, struct mapped_device *md)
803 if (!td->dm_dev.bdev)
806 bd_unlink_disk_holder(td->dm_dev.bdev, dm_disk(md));
807 blkdev_put(td->dm_dev.bdev, td->dm_dev.mode | FMODE_EXCL);
808 td->dm_dev.bdev = NULL;
811 static struct table_device *find_table_device(struct list_head *l, dev_t dev,
813 struct table_device *td;
815 list_for_each_entry(td, l, list)
816 if (td->dm_dev.bdev->bd_dev == dev && td->dm_dev.mode == mode)
822 int dm_get_table_device(struct mapped_device *md, dev_t dev, fmode_t mode,
823 struct dm_dev **result) {
825 struct table_device *td;
827 mutex_lock(&md->table_devices_lock);
828 td = find_table_device(&md->table_devices, dev, mode);
830 td = kmalloc(sizeof(*td), GFP_KERNEL);
832 mutex_unlock(&md->table_devices_lock);
836 td->dm_dev.mode = mode;
837 td->dm_dev.bdev = NULL;
839 if ((r = open_table_device(td, dev, md))) {
840 mutex_unlock(&md->table_devices_lock);
845 format_dev_t(td->dm_dev.name, dev);
847 atomic_set(&td->count, 0);
848 list_add(&td->list, &md->table_devices);
850 atomic_inc(&td->count);
851 mutex_unlock(&md->table_devices_lock);
853 *result = &td->dm_dev;
856 EXPORT_SYMBOL_GPL(dm_get_table_device);
858 void dm_put_table_device(struct mapped_device *md, struct dm_dev *d)
860 struct table_device *td = container_of(d, struct table_device, dm_dev);
862 mutex_lock(&md->table_devices_lock);
863 if (atomic_dec_and_test(&td->count)) {
864 close_table_device(td, md);
868 mutex_unlock(&md->table_devices_lock);
870 EXPORT_SYMBOL(dm_put_table_device);
872 static void free_table_devices(struct list_head *devices)
874 struct list_head *tmp, *next;
876 list_for_each_safe(tmp, next, devices) {
877 struct table_device *td = list_entry(tmp, struct table_device, list);
879 DMWARN("dm_destroy: %s still exists with %d references",
880 td->dm_dev.name, atomic_read(&td->count));
886 * Get the geometry associated with a dm device
888 int dm_get_geometry(struct mapped_device *md, struct hd_geometry *geo)
896 * Set the geometry of a device.
898 int dm_set_geometry(struct mapped_device *md, struct hd_geometry *geo)
900 sector_t sz = (sector_t)geo->cylinders * geo->heads * geo->sectors;
902 if (geo->start > sz) {
903 DMWARN("Start sector is beyond the geometry limits.");
912 /*-----------------------------------------------------------------
914 * A more elegant soln is in the works that uses the queue
915 * merge fn, unfortunately there are a couple of changes to
916 * the block layer that I want to make for this. So in the
917 * interests of getting something for people to use I give
918 * you this clearly demarcated crap.
919 *---------------------------------------------------------------*/
921 static int __noflush_suspending(struct mapped_device *md)
923 return test_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
927 * Decrements the number of outstanding ios that a bio has been
928 * cloned into, completing the original io if necc.
930 static void dec_pending(struct dm_io *io, int error)
935 struct mapped_device *md = io->md;
937 /* Push-back supersedes any I/O errors */
938 if (unlikely(error)) {
939 spin_lock_irqsave(&io->endio_lock, flags);
940 if (!(io->error > 0 && __noflush_suspending(md)))
942 spin_unlock_irqrestore(&io->endio_lock, flags);
945 if (atomic_dec_and_test(&io->io_count)) {
946 if (io->error == DM_ENDIO_REQUEUE) {
948 * Target requested pushing back the I/O.
950 spin_lock_irqsave(&md->deferred_lock, flags);
951 if (__noflush_suspending(md))
952 bio_list_add_head(&md->deferred, io->bio);
954 /* noflush suspend was interrupted. */
956 spin_unlock_irqrestore(&md->deferred_lock, flags);
959 io_error = io->error;
964 if (io_error == DM_ENDIO_REQUEUE)
967 if ((bio->bi_rw & REQ_FLUSH) && bio->bi_iter.bi_size) {
969 * Preflush done for flush with data, reissue
972 bio->bi_rw &= ~REQ_FLUSH;
975 /* done with normal IO or empty flush */
976 trace_block_bio_complete(md->queue, bio, io_error);
977 bio->bi_error = io_error;
983 static void disable_write_same(struct mapped_device *md)
985 struct queue_limits *limits = dm_get_queue_limits(md);
987 /* device doesn't really support WRITE SAME, disable it */
988 limits->max_write_same_sectors = 0;
991 static void clone_endio(struct bio *bio)
993 int error = bio->bi_error;
995 struct dm_target_io *tio = container_of(bio, struct dm_target_io, clone);
996 struct dm_io *io = tio->io;
997 struct mapped_device *md = tio->io->md;
998 dm_endio_fn endio = tio->ti->type->end_io;
1001 r = endio(tio->ti, bio, error);
1002 if (r < 0 || r == DM_ENDIO_REQUEUE)
1004 * error and requeue request are handled
1008 else if (r == DM_ENDIO_INCOMPLETE)
1009 /* The target will handle the io */
1012 DMWARN("unimplemented target endio return value: %d", r);
1017 if (unlikely(r == -EREMOTEIO && (bio->bi_rw & REQ_WRITE_SAME) &&
1018 !bdev_get_queue(bio->bi_bdev)->limits.max_write_same_sectors))
1019 disable_write_same(md);
1022 dec_pending(io, error);
1026 * Partial completion handling for request-based dm
1028 static void end_clone_bio(struct bio *clone)
1030 struct dm_rq_clone_bio_info *info =
1031 container_of(clone, struct dm_rq_clone_bio_info, clone);
1032 struct dm_rq_target_io *tio = info->tio;
1033 struct bio *bio = info->orig;
1034 unsigned int nr_bytes = info->orig->bi_iter.bi_size;
1035 int error = clone->bi_error;
1041 * An error has already been detected on the request.
1042 * Once error occurred, just let clone->end_io() handle
1048 * Don't notice the error to the upper layer yet.
1049 * The error handling decision is made by the target driver,
1050 * when the request is completed.
1057 * I/O for the bio successfully completed.
1058 * Notice the data completion to the upper layer.
1062 * bios are processed from the head of the list.
1063 * So the completing bio should always be rq->bio.
1064 * If it's not, something wrong is happening.
1066 if (tio->orig->bio != bio)
1067 DMERR("bio completion is going in the middle of the request");
1070 * Update the original request.
1071 * Do not use blk_end_request() here, because it may complete
1072 * the original request before the clone, and break the ordering.
1074 blk_update_request(tio->orig, 0, nr_bytes);
1077 static struct dm_rq_target_io *tio_from_request(struct request *rq)
1079 return (rq->q->mq_ops ? blk_mq_rq_to_pdu(rq) : rq->special);
1082 static void rq_end_stats(struct mapped_device *md, struct request *orig)
1084 if (unlikely(dm_stats_used(&md->stats))) {
1085 struct dm_rq_target_io *tio = tio_from_request(orig);
1086 tio->duration_jiffies = jiffies - tio->duration_jiffies;
1087 dm_stats_account_io(&md->stats, orig->cmd_flags, blk_rq_pos(orig),
1088 tio->n_sectors, true, tio->duration_jiffies,
1094 * Don't touch any member of the md after calling this function because
1095 * the md may be freed in dm_put() at the end of this function.
1096 * Or do dm_get() before calling this function and dm_put() later.
1098 static void rq_completed(struct mapped_device *md, int rw, bool run_queue)
1100 atomic_dec(&md->pending[rw]);
1102 /* nudge anyone waiting on suspend queue */
1103 if (!md_in_flight(md))
1107 * Run this off this callpath, as drivers could invoke end_io while
1108 * inside their request_fn (and holding the queue lock). Calling
1109 * back into ->request_fn() could deadlock attempting to grab the
1112 if (!md->queue->mq_ops && run_queue)
1113 blk_run_queue_async(md->queue);
1116 * dm_put() must be at the end of this function. See the comment above
1121 static void free_rq_clone(struct request *clone)
1123 struct dm_rq_target_io *tio = clone->end_io_data;
1124 struct mapped_device *md = tio->md;
1126 blk_rq_unprep_clone(clone);
1128 if (md->type == DM_TYPE_MQ_REQUEST_BASED)
1129 /* stacked on blk-mq queue(s) */
1130 tio->ti->type->release_clone_rq(clone);
1131 else if (!md->queue->mq_ops)
1132 /* request_fn queue stacked on request_fn queue(s) */
1133 free_clone_request(md, clone);
1135 * NOTE: for the blk-mq queue stacked on request_fn queue(s) case:
1136 * no need to call free_clone_request() because we leverage blk-mq by
1137 * allocating the clone at the end of the blk-mq pdu (see: clone_rq)
1140 if (!md->queue->mq_ops)
1145 * Complete the clone and the original request.
1146 * Must be called without clone's queue lock held,
1147 * see end_clone_request() for more details.
1149 static void dm_end_request(struct request *clone, int error)
1151 int rw = rq_data_dir(clone);
1152 struct dm_rq_target_io *tio = clone->end_io_data;
1153 struct mapped_device *md = tio->md;
1154 struct request *rq = tio->orig;
1156 if (rq->cmd_type == REQ_TYPE_BLOCK_PC) {
1157 rq->errors = clone->errors;
1158 rq->resid_len = clone->resid_len;
1162 * We are using the sense buffer of the original
1164 * So setting the length of the sense data is enough.
1166 rq->sense_len = clone->sense_len;
1169 free_rq_clone(clone);
1170 rq_end_stats(md, rq);
1172 blk_end_request_all(rq, error);
1174 blk_mq_end_request(rq, error);
1175 rq_completed(md, rw, true);
1178 static void dm_unprep_request(struct request *rq)
1180 struct dm_rq_target_io *tio = tio_from_request(rq);
1181 struct request *clone = tio->clone;
1183 if (!rq->q->mq_ops) {
1185 rq->cmd_flags &= ~REQ_DONTPREP;
1189 free_rq_clone(clone);
1190 else if (!tio->md->queue->mq_ops)
1195 * Requeue the original request of a clone.
1197 static void old_requeue_request(struct request *rq)
1199 struct request_queue *q = rq->q;
1200 unsigned long flags;
1202 spin_lock_irqsave(q->queue_lock, flags);
1203 blk_requeue_request(q, rq);
1204 blk_run_queue_async(q);
1205 spin_unlock_irqrestore(q->queue_lock, flags);
1208 static void dm_requeue_original_request(struct mapped_device *md,
1211 int rw = rq_data_dir(rq);
1213 rq_end_stats(md, rq);
1214 dm_unprep_request(rq);
1217 old_requeue_request(rq);
1219 blk_mq_requeue_request(rq);
1220 blk_mq_kick_requeue_list(rq->q);
1223 rq_completed(md, rw, false);
1226 static void old_stop_queue(struct request_queue *q)
1228 unsigned long flags;
1230 if (blk_queue_stopped(q))
1233 spin_lock_irqsave(q->queue_lock, flags);
1235 spin_unlock_irqrestore(q->queue_lock, flags);
1238 static void stop_queue(struct request_queue *q)
1243 blk_mq_stop_hw_queues(q);
1246 static void old_start_queue(struct request_queue *q)
1248 unsigned long flags;
1250 spin_lock_irqsave(q->queue_lock, flags);
1251 if (blk_queue_stopped(q))
1253 spin_unlock_irqrestore(q->queue_lock, flags);
1256 static void start_queue(struct request_queue *q)
1261 blk_mq_start_stopped_hw_queues(q, true);
1264 static void dm_done(struct request *clone, int error, bool mapped)
1267 struct dm_rq_target_io *tio = clone->end_io_data;
1268 dm_request_endio_fn rq_end_io = NULL;
1271 rq_end_io = tio->ti->type->rq_end_io;
1273 if (mapped && rq_end_io)
1274 r = rq_end_io(tio->ti, clone, error, &tio->info);
1277 if (unlikely(r == -EREMOTEIO && (clone->cmd_flags & REQ_WRITE_SAME) &&
1278 !clone->q->limits.max_write_same_sectors))
1279 disable_write_same(tio->md);
1282 /* The target wants to complete the I/O */
1283 dm_end_request(clone, r);
1284 else if (r == DM_ENDIO_INCOMPLETE)
1285 /* The target will handle the I/O */
1287 else if (r == DM_ENDIO_REQUEUE)
1288 /* The target wants to requeue the I/O */
1289 dm_requeue_original_request(tio->md, tio->orig);
1291 DMWARN("unimplemented target endio return value: %d", r);
1297 * Request completion handler for request-based dm
1299 static void dm_softirq_done(struct request *rq)
1302 struct dm_rq_target_io *tio = tio_from_request(rq);
1303 struct request *clone = tio->clone;
1307 rq_end_stats(tio->md, rq);
1308 rw = rq_data_dir(rq);
1309 if (!rq->q->mq_ops) {
1310 blk_end_request_all(rq, tio->error);
1311 rq_completed(tio->md, rw, false);
1314 blk_mq_end_request(rq, tio->error);
1315 rq_completed(tio->md, rw, false);
1320 if (rq->cmd_flags & REQ_FAILED)
1323 dm_done(clone, tio->error, mapped);
1327 * Complete the clone and the original request with the error status
1328 * through softirq context.
1330 static void dm_complete_request(struct request *rq, int error)
1332 struct dm_rq_target_io *tio = tio_from_request(rq);
1336 blk_complete_request(rq);
1338 blk_mq_complete_request(rq, error);
1342 * Complete the not-mapped clone and the original request with the error status
1343 * through softirq context.
1344 * Target's rq_end_io() function isn't called.
1345 * This may be used when the target's map_rq() or clone_and_map_rq() functions fail.
1347 static void dm_kill_unmapped_request(struct request *rq, int error)
1349 rq->cmd_flags |= REQ_FAILED;
1350 dm_complete_request(rq, error);
1354 * Called with the clone's queue lock held (for non-blk-mq)
1356 static void end_clone_request(struct request *clone, int error)
1358 struct dm_rq_target_io *tio = clone->end_io_data;
1360 if (!clone->q->mq_ops) {
1362 * For just cleaning up the information of the queue in which
1363 * the clone was dispatched.
1364 * The clone is *NOT* freed actually here because it is alloced
1365 * from dm own mempool (REQ_ALLOCED isn't set).
1367 __blk_put_request(clone->q, clone);
1371 * Actual request completion is done in a softirq context which doesn't
1372 * hold the clone's queue lock. Otherwise, deadlock could occur because:
1373 * - another request may be submitted by the upper level driver
1374 * of the stacking during the completion
1375 * - the submission which requires queue lock may be done
1376 * against this clone's queue
1378 dm_complete_request(tio->orig, error);
1382 * Return maximum size of I/O possible at the supplied sector up to the current
1385 static sector_t max_io_len_target_boundary(sector_t sector, struct dm_target *ti)
1387 sector_t target_offset = dm_target_offset(ti, sector);
1389 return ti->len - target_offset;
1392 static sector_t max_io_len(sector_t sector, struct dm_target *ti)
1394 sector_t len = max_io_len_target_boundary(sector, ti);
1395 sector_t offset, max_len;
1398 * Does the target need to split even further?
1400 if (ti->max_io_len) {
1401 offset = dm_target_offset(ti, sector);
1402 if (unlikely(ti->max_io_len & (ti->max_io_len - 1)))
1403 max_len = sector_div(offset, ti->max_io_len);
1405 max_len = offset & (ti->max_io_len - 1);
1406 max_len = ti->max_io_len - max_len;
1415 int dm_set_target_max_io_len(struct dm_target *ti, sector_t len)
1417 if (len > UINT_MAX) {
1418 DMERR("Specified maximum size of target IO (%llu) exceeds limit (%u)",
1419 (unsigned long long)len, UINT_MAX);
1420 ti->error = "Maximum size of target IO is too large";
1424 ti->max_io_len = (uint32_t) len;
1428 EXPORT_SYMBOL_GPL(dm_set_target_max_io_len);
1431 * A target may call dm_accept_partial_bio only from the map routine. It is
1432 * allowed for all bio types except REQ_FLUSH.
1434 * dm_accept_partial_bio informs the dm that the target only wants to process
1435 * additional n_sectors sectors of the bio and the rest of the data should be
1436 * sent in a next bio.
1438 * A diagram that explains the arithmetics:
1439 * +--------------------+---------------+-------+
1441 * +--------------------+---------------+-------+
1443 * <-------------- *tio->len_ptr --------------->
1444 * <------- bi_size ------->
1447 * Region 1 was already iterated over with bio_advance or similar function.
1448 * (it may be empty if the target doesn't use bio_advance)
1449 * Region 2 is the remaining bio size that the target wants to process.
1450 * (it may be empty if region 1 is non-empty, although there is no reason
1452 * The target requires that region 3 is to be sent in the next bio.
1454 * If the target wants to receive multiple copies of the bio (via num_*bios, etc),
1455 * the partially processed part (the sum of regions 1+2) must be the same for all
1456 * copies of the bio.
1458 void dm_accept_partial_bio(struct bio *bio, unsigned n_sectors)
1460 struct dm_target_io *tio = container_of(bio, struct dm_target_io, clone);
1461 unsigned bi_size = bio->bi_iter.bi_size >> SECTOR_SHIFT;
1462 BUG_ON(bio->bi_rw & REQ_FLUSH);
1463 BUG_ON(bi_size > *tio->len_ptr);
1464 BUG_ON(n_sectors > bi_size);
1465 *tio->len_ptr -= bi_size - n_sectors;
1466 bio->bi_iter.bi_size = n_sectors << SECTOR_SHIFT;
1468 EXPORT_SYMBOL_GPL(dm_accept_partial_bio);
1471 * Flush current->bio_list when the target map method blocks.
1472 * This fixes deadlocks in snapshot and possibly in other targets.
1475 struct blk_plug plug;
1476 struct blk_plug_cb cb;
1479 static void flush_current_bio_list(struct blk_plug_cb *cb, bool from_schedule)
1481 struct dm_offload *o = container_of(cb, struct dm_offload, cb);
1482 struct bio_list list;
1485 INIT_LIST_HEAD(&o->cb.list);
1487 if (unlikely(!current->bio_list))
1490 list = *current->bio_list;
1491 bio_list_init(current->bio_list);
1493 while ((bio = bio_list_pop(&list))) {
1494 struct bio_set *bs = bio->bi_pool;
1495 if (unlikely(!bs) || bs == fs_bio_set) {
1496 bio_list_add(current->bio_list, bio);
1500 spin_lock(&bs->rescue_lock);
1501 bio_list_add(&bs->rescue_list, bio);
1502 queue_work(bs->rescue_workqueue, &bs->rescue_work);
1503 spin_unlock(&bs->rescue_lock);
1507 static void dm_offload_start(struct dm_offload *o)
1509 blk_start_plug(&o->plug);
1510 o->cb.callback = flush_current_bio_list;
1511 list_add(&o->cb.list, ¤t->plug->cb_list);
1514 static void dm_offload_end(struct dm_offload *o)
1516 list_del(&o->cb.list);
1517 blk_finish_plug(&o->plug);
1520 static void __map_bio(struct dm_target_io *tio)
1524 struct mapped_device *md;
1525 struct dm_offload o;
1526 struct bio *clone = &tio->clone;
1527 struct dm_target *ti = tio->ti;
1529 clone->bi_end_io = clone_endio;
1532 * Map the clone. If r == 0 we don't need to do
1533 * anything, the target has assumed ownership of
1536 atomic_inc(&tio->io->io_count);
1537 sector = clone->bi_iter.bi_sector;
1539 dm_offload_start(&o);
1540 r = ti->type->map(ti, clone);
1543 if (r == DM_MAPIO_REMAPPED) {
1544 /* the bio has been remapped so dispatch it */
1546 trace_block_bio_remap(bdev_get_queue(clone->bi_bdev), clone,
1547 tio->io->bio->bi_bdev->bd_dev, sector);
1549 generic_make_request(clone);
1550 } else if (r < 0 || r == DM_MAPIO_REQUEUE) {
1551 /* error the io and bail out, or requeue it if needed */
1553 dec_pending(tio->io, r);
1555 } else if (r != DM_MAPIO_SUBMITTED) {
1556 DMWARN("unimplemented target map return value: %d", r);
1562 struct mapped_device *md;
1563 struct dm_table *map;
1567 unsigned sector_count;
1570 static void bio_setup_sector(struct bio *bio, sector_t sector, unsigned len)
1572 bio->bi_iter.bi_sector = sector;
1573 bio->bi_iter.bi_size = to_bytes(len);
1577 * Creates a bio that consists of range of complete bvecs.
1579 static void clone_bio(struct dm_target_io *tio, struct bio *bio,
1580 sector_t sector, unsigned len)
1582 struct bio *clone = &tio->clone;
1584 __bio_clone_fast(clone, bio);
1586 if (bio_integrity(bio))
1587 bio_integrity_clone(clone, bio, GFP_NOIO);
1589 bio_advance(clone, to_bytes(sector - clone->bi_iter.bi_sector));
1590 clone->bi_iter.bi_size = to_bytes(len);
1592 if (bio_integrity(bio))
1593 bio_integrity_trim(clone, 0, len);
1596 static struct dm_target_io *alloc_tio(struct clone_info *ci,
1597 struct dm_target *ti,
1598 unsigned target_bio_nr)
1600 struct dm_target_io *tio;
1603 clone = bio_alloc_bioset(GFP_NOIO, 0, ci->md->bs);
1604 tio = container_of(clone, struct dm_target_io, clone);
1608 tio->target_bio_nr = target_bio_nr;
1613 static void __clone_and_map_simple_bio(struct clone_info *ci,
1614 struct dm_target *ti,
1615 unsigned target_bio_nr, unsigned *len)
1617 struct dm_target_io *tio = alloc_tio(ci, ti, target_bio_nr);
1618 struct bio *clone = &tio->clone;
1622 __bio_clone_fast(clone, ci->bio);
1624 bio_setup_sector(clone, ci->sector, *len);
1629 static void __send_duplicate_bios(struct clone_info *ci, struct dm_target *ti,
1630 unsigned num_bios, unsigned *len)
1632 unsigned target_bio_nr;
1634 for (target_bio_nr = 0; target_bio_nr < num_bios; target_bio_nr++)
1635 __clone_and_map_simple_bio(ci, ti, target_bio_nr, len);
1638 static int __send_empty_flush(struct clone_info *ci)
1640 unsigned target_nr = 0;
1641 struct dm_target *ti;
1643 BUG_ON(bio_has_data(ci->bio));
1644 while ((ti = dm_table_get_target(ci->map, target_nr++)))
1645 __send_duplicate_bios(ci, ti, ti->num_flush_bios, NULL);
1650 static void __clone_and_map_data_bio(struct clone_info *ci, struct dm_target *ti,
1651 sector_t sector, unsigned *len)
1653 struct bio *bio = ci->bio;
1654 struct dm_target_io *tio;
1655 unsigned target_bio_nr;
1656 unsigned num_target_bios = 1;
1659 * Does the target want to receive duplicate copies of the bio?
1661 if (bio_data_dir(bio) == WRITE && ti->num_write_bios)
1662 num_target_bios = ti->num_write_bios(ti, bio);
1664 for (target_bio_nr = 0; target_bio_nr < num_target_bios; target_bio_nr++) {
1665 tio = alloc_tio(ci, ti, target_bio_nr);
1667 clone_bio(tio, bio, sector, *len);
1672 typedef unsigned (*get_num_bios_fn)(struct dm_target *ti);
1674 static unsigned get_num_discard_bios(struct dm_target *ti)
1676 return ti->num_discard_bios;
1679 static unsigned get_num_write_same_bios(struct dm_target *ti)
1681 return ti->num_write_same_bios;
1684 typedef bool (*is_split_required_fn)(struct dm_target *ti);
1686 static bool is_split_required_for_discard(struct dm_target *ti)
1688 return ti->split_discard_bios;
1691 static int __send_changing_extent_only(struct clone_info *ci,
1692 get_num_bios_fn get_num_bios,
1693 is_split_required_fn is_split_required)
1695 struct dm_target *ti;
1700 ti = dm_table_find_target(ci->map, ci->sector);
1701 if (!dm_target_is_valid(ti))
1705 * Even though the device advertised support for this type of
1706 * request, that does not mean every target supports it, and
1707 * reconfiguration might also have changed that since the
1708 * check was performed.
1710 num_bios = get_num_bios ? get_num_bios(ti) : 0;
1714 if (is_split_required && !is_split_required(ti))
1715 len = min((sector_t)ci->sector_count, max_io_len_target_boundary(ci->sector, ti));
1717 len = min((sector_t)ci->sector_count, max_io_len(ci->sector, ti));
1719 __send_duplicate_bios(ci, ti, num_bios, &len);
1722 } while (ci->sector_count -= len);
1727 static int __send_discard(struct clone_info *ci)
1729 return __send_changing_extent_only(ci, get_num_discard_bios,
1730 is_split_required_for_discard);
1733 static int __send_write_same(struct clone_info *ci)
1735 return __send_changing_extent_only(ci, get_num_write_same_bios, NULL);
1739 * Select the correct strategy for processing a non-flush bio.
1741 static int __split_and_process_non_flush(struct clone_info *ci)
1743 struct bio *bio = ci->bio;
1744 struct dm_target *ti;
1747 if (unlikely(bio->bi_rw & REQ_DISCARD))
1748 return __send_discard(ci);
1749 else if (unlikely(bio->bi_rw & REQ_WRITE_SAME))
1750 return __send_write_same(ci);
1752 ti = dm_table_find_target(ci->map, ci->sector);
1753 if (!dm_target_is_valid(ti))
1756 len = min_t(sector_t, max_io_len(ci->sector, ti), ci->sector_count);
1758 __clone_and_map_data_bio(ci, ti, ci->sector, &len);
1761 ci->sector_count -= len;
1767 * Entry point to split a bio into clones and submit them to the targets.
1769 static void __split_and_process_bio(struct mapped_device *md,
1770 struct dm_table *map, struct bio *bio)
1772 struct clone_info ci;
1775 if (unlikely(!map)) {
1782 ci.io = alloc_io(md);
1784 atomic_set(&ci.io->io_count, 1);
1787 spin_lock_init(&ci.io->endio_lock);
1788 ci.sector = bio->bi_iter.bi_sector;
1790 start_io_acct(ci.io);
1792 if (bio->bi_rw & REQ_FLUSH) {
1793 ci.bio = &ci.md->flush_bio;
1794 ci.sector_count = 0;
1795 error = __send_empty_flush(&ci);
1796 /* dec_pending submits any data associated with flush */
1799 ci.sector_count = bio_sectors(bio);
1800 while (ci.sector_count && !error)
1801 error = __split_and_process_non_flush(&ci);
1804 /* drop the extra reference count */
1805 dec_pending(ci.io, error);
1807 /*-----------------------------------------------------------------
1809 *---------------------------------------------------------------*/
1812 * The request function that just remaps the bio built up by
1815 static blk_qc_t dm_make_request(struct request_queue *q, struct bio *bio)
1817 int rw = bio_data_dir(bio);
1818 struct mapped_device *md = q->queuedata;
1820 struct dm_table *map;
1822 map = dm_get_live_table(md, &srcu_idx);
1824 generic_start_io_acct(rw, bio_sectors(bio), &dm_disk(md)->part0);
1826 /* if we're suspended, we have to queue this io for later */
1827 if (unlikely(test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags))) {
1828 dm_put_live_table(md, srcu_idx);
1830 if (bio_rw(bio) != READA)
1834 return BLK_QC_T_NONE;
1837 __split_and_process_bio(md, map, bio);
1838 dm_put_live_table(md, srcu_idx);
1839 return BLK_QC_T_NONE;
1842 int dm_request_based(struct mapped_device *md)
1844 return blk_queue_stackable(md->queue);
1847 static void dm_dispatch_clone_request(struct request *clone, struct request *rq)
1851 if (blk_queue_io_stat(clone->q))
1852 clone->cmd_flags |= REQ_IO_STAT;
1854 clone->start_time = jiffies;
1855 r = blk_insert_cloned_request(clone->q, clone);
1857 /* must complete clone in terms of original request */
1858 dm_complete_request(rq, r);
1861 static int dm_rq_bio_constructor(struct bio *bio, struct bio *bio_orig,
1864 struct dm_rq_target_io *tio = data;
1865 struct dm_rq_clone_bio_info *info =
1866 container_of(bio, struct dm_rq_clone_bio_info, clone);
1868 info->orig = bio_orig;
1870 bio->bi_end_io = end_clone_bio;
1875 static int setup_clone(struct request *clone, struct request *rq,
1876 struct dm_rq_target_io *tio, gfp_t gfp_mask)
1880 r = blk_rq_prep_clone(clone, rq, tio->md->bs, gfp_mask,
1881 dm_rq_bio_constructor, tio);
1885 clone->cmd = rq->cmd;
1886 clone->cmd_len = rq->cmd_len;
1887 clone->sense = rq->sense;
1888 clone->end_io = end_clone_request;
1889 clone->end_io_data = tio;
1896 static struct request *clone_rq(struct request *rq, struct mapped_device *md,
1897 struct dm_rq_target_io *tio, gfp_t gfp_mask)
1900 * Do not allocate a clone if tio->clone was already set
1901 * (see: dm_mq_queue_rq).
1903 bool alloc_clone = !tio->clone;
1904 struct request *clone;
1907 clone = alloc_clone_request(md, gfp_mask);
1913 blk_rq_init(NULL, clone);
1914 if (setup_clone(clone, rq, tio, gfp_mask)) {
1917 free_clone_request(md, clone);
1924 static void map_tio_request(struct kthread_work *work);
1926 static void init_tio(struct dm_rq_target_io *tio, struct request *rq,
1927 struct mapped_device *md)
1934 memset(&tio->info, 0, sizeof(tio->info));
1935 if (md->kworker_task)
1936 init_kthread_work(&tio->work, map_tio_request);
1939 static struct dm_rq_target_io *prep_tio(struct request *rq,
1940 struct mapped_device *md, gfp_t gfp_mask)
1942 struct dm_rq_target_io *tio;
1944 struct dm_table *table;
1946 tio = alloc_rq_tio(md, gfp_mask);
1950 init_tio(tio, rq, md);
1952 table = dm_get_live_table(md, &srcu_idx);
1953 if (!dm_table_mq_request_based(table)) {
1954 if (!clone_rq(rq, md, tio, gfp_mask)) {
1955 dm_put_live_table(md, srcu_idx);
1960 dm_put_live_table(md, srcu_idx);
1966 * Called with the queue lock held.
1968 static int dm_prep_fn(struct request_queue *q, struct request *rq)
1970 struct mapped_device *md = q->queuedata;
1971 struct dm_rq_target_io *tio;
1973 if (unlikely(rq->special)) {
1974 DMWARN("Already has something in rq->special.");
1975 return BLKPREP_KILL;
1978 tio = prep_tio(rq, md, GFP_ATOMIC);
1980 return BLKPREP_DEFER;
1983 rq->cmd_flags |= REQ_DONTPREP;
1990 * 0 : the request has been processed
1991 * DM_MAPIO_REQUEUE : the original request needs to be requeued
1992 * < 0 : the request was completed due to failure
1994 static int map_request(struct dm_rq_target_io *tio, struct request *rq,
1995 struct mapped_device *md)
1998 struct dm_target *ti = tio->ti;
1999 struct request *clone = NULL;
2003 r = ti->type->map_rq(ti, clone, &tio->info);
2005 r = ti->type->clone_and_map_rq(ti, rq, &tio->info, &clone);
2007 /* The target wants to complete the I/O */
2008 dm_kill_unmapped_request(rq, r);
2011 if (r != DM_MAPIO_REMAPPED)
2013 if (setup_clone(clone, rq, tio, GFP_ATOMIC)) {
2015 ti->type->release_clone_rq(clone);
2016 return DM_MAPIO_REQUEUE;
2021 case DM_MAPIO_SUBMITTED:
2022 /* The target has taken the I/O to submit by itself later */
2024 case DM_MAPIO_REMAPPED:
2025 /* The target has remapped the I/O so dispatch it */
2026 trace_block_rq_remap(clone->q, clone, disk_devt(dm_disk(md)),
2028 dm_dispatch_clone_request(clone, rq);
2030 case DM_MAPIO_REQUEUE:
2031 /* The target wants to requeue the I/O */
2032 dm_requeue_original_request(md, tio->orig);
2036 DMWARN("unimplemented target map return value: %d", r);
2040 /* The target wants to complete the I/O */
2041 dm_kill_unmapped_request(rq, r);
2048 static void map_tio_request(struct kthread_work *work)
2050 struct dm_rq_target_io *tio = container_of(work, struct dm_rq_target_io, work);
2051 struct request *rq = tio->orig;
2052 struct mapped_device *md = tio->md;
2054 if (map_request(tio, rq, md) == DM_MAPIO_REQUEUE)
2055 dm_requeue_original_request(md, rq);
2058 static void dm_start_request(struct mapped_device *md, struct request *orig)
2060 if (!orig->q->mq_ops)
2061 blk_start_request(orig);
2063 blk_mq_start_request(orig);
2064 atomic_inc(&md->pending[rq_data_dir(orig)]);
2066 if (md->seq_rq_merge_deadline_usecs) {
2067 md->last_rq_pos = rq_end_sector(orig);
2068 md->last_rq_rw = rq_data_dir(orig);
2069 md->last_rq_start_time = ktime_get();
2072 if (unlikely(dm_stats_used(&md->stats))) {
2073 struct dm_rq_target_io *tio = tio_from_request(orig);
2074 tio->duration_jiffies = jiffies;
2075 tio->n_sectors = blk_rq_sectors(orig);
2076 dm_stats_account_io(&md->stats, orig->cmd_flags, blk_rq_pos(orig),
2077 tio->n_sectors, false, 0, &tio->stats_aux);
2081 * Hold the md reference here for the in-flight I/O.
2082 * We can't rely on the reference count by device opener,
2083 * because the device may be closed during the request completion
2084 * when all bios are completed.
2085 * See the comment in rq_completed() too.
2090 #define MAX_SEQ_RQ_MERGE_DEADLINE_USECS 100000
2092 ssize_t dm_attr_rq_based_seq_io_merge_deadline_show(struct mapped_device *md, char *buf)
2094 return sprintf(buf, "%u\n", md->seq_rq_merge_deadline_usecs);
2097 ssize_t dm_attr_rq_based_seq_io_merge_deadline_store(struct mapped_device *md,
2098 const char *buf, size_t count)
2102 if (!dm_request_based(md) || md->use_blk_mq)
2105 if (kstrtouint(buf, 10, &deadline))
2108 if (deadline > MAX_SEQ_RQ_MERGE_DEADLINE_USECS)
2109 deadline = MAX_SEQ_RQ_MERGE_DEADLINE_USECS;
2111 md->seq_rq_merge_deadline_usecs = deadline;
2116 static bool dm_request_peeked_before_merge_deadline(struct mapped_device *md)
2118 ktime_t kt_deadline;
2120 if (!md->seq_rq_merge_deadline_usecs)
2123 kt_deadline = ns_to_ktime((u64)md->seq_rq_merge_deadline_usecs * NSEC_PER_USEC);
2124 kt_deadline = ktime_add_safe(md->last_rq_start_time, kt_deadline);
2126 return !ktime_after(ktime_get(), kt_deadline);
2130 * q->request_fn for request-based dm.
2131 * Called with the queue lock held.
2133 static void dm_request_fn(struct request_queue *q)
2135 struct mapped_device *md = q->queuedata;
2137 struct dm_table *map = dm_get_live_table(md, &srcu_idx);
2138 struct dm_target *ti;
2140 struct dm_rq_target_io *tio;
2144 * For suspend, check blk_queue_stopped() and increment
2145 * ->pending within a single queue_lock not to increment the
2146 * number of in-flight I/Os after the queue is stopped in
2149 while (!blk_queue_stopped(q)) {
2150 rq = blk_peek_request(q);
2154 /* always use block 0 to find the target for flushes for now */
2156 if (!(rq->cmd_flags & REQ_FLUSH))
2157 pos = blk_rq_pos(rq);
2159 ti = dm_table_find_target(map, pos);
2160 if (!dm_target_is_valid(ti)) {
2162 * Must perform setup, that rq_completed() requires,
2163 * before calling dm_kill_unmapped_request
2165 DMERR_LIMIT("request attempted access beyond the end of device");
2166 dm_start_request(md, rq);
2167 dm_kill_unmapped_request(rq, -EIO);
2171 if (dm_request_peeked_before_merge_deadline(md) &&
2172 md_in_flight(md) && rq->bio && rq->bio->bi_vcnt == 1 &&
2173 md->last_rq_pos == pos && md->last_rq_rw == rq_data_dir(rq))
2176 if (ti->type->busy && ti->type->busy(ti))
2179 dm_start_request(md, rq);
2181 tio = tio_from_request(rq);
2182 /* Establish tio->ti before queuing work (map_tio_request) */
2184 queue_kthread_work(&md->kworker, &tio->work);
2185 BUG_ON(!irqs_disabled());
2191 blk_delay_queue(q, HZ / 100);
2193 dm_put_live_table(md, srcu_idx);
2196 static int dm_any_congested(void *congested_data, int bdi_bits)
2199 struct mapped_device *md = congested_data;
2200 struct dm_table *map;
2202 if (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
2203 map = dm_get_live_table_fast(md);
2206 * Request-based dm cares about only own queue for
2207 * the query about congestion status of request_queue
2209 if (dm_request_based(md))
2210 r = md->queue->backing_dev_info.wb.state &
2213 r = dm_table_any_congested(map, bdi_bits);
2215 dm_put_live_table_fast(md);
2221 /*-----------------------------------------------------------------
2222 * An IDR is used to keep track of allocated minor numbers.
2223 *---------------------------------------------------------------*/
2224 static void free_minor(int minor)
2226 spin_lock(&_minor_lock);
2227 idr_remove(&_minor_idr, minor);
2228 spin_unlock(&_minor_lock);
2232 * See if the device with a specific minor # is free.
2234 static int specific_minor(int minor)
2238 if (minor >= (1 << MINORBITS))
2241 idr_preload(GFP_KERNEL);
2242 spin_lock(&_minor_lock);
2244 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, minor, minor + 1, GFP_NOWAIT);
2246 spin_unlock(&_minor_lock);
2249 return r == -ENOSPC ? -EBUSY : r;
2253 static int next_free_minor(int *minor)
2257 idr_preload(GFP_KERNEL);
2258 spin_lock(&_minor_lock);
2260 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, 0, 1 << MINORBITS, GFP_NOWAIT);
2262 spin_unlock(&_minor_lock);
2270 static const struct block_device_operations dm_blk_dops;
2272 static void dm_wq_work(struct work_struct *work);
2274 static void dm_init_md_queue(struct mapped_device *md)
2277 * Request-based dm devices cannot be stacked on top of bio-based dm
2278 * devices. The type of this dm device may not have been decided yet.
2279 * The type is decided at the first table loading time.
2280 * To prevent problematic device stacking, clear the queue flag
2281 * for request stacking support until then.
2283 * This queue is new, so no concurrency on the queue_flags.
2285 queue_flag_clear_unlocked(QUEUE_FLAG_STACKABLE, md->queue);
2288 * Initialize data that will only be used by a non-blk-mq DM queue
2289 * - must do so here (in alloc_dev callchain) before queue is used
2291 md->queue->queuedata = md;
2292 md->queue->backing_dev_info.congested_data = md;
2295 static void dm_init_old_md_queue(struct mapped_device *md)
2297 md->use_blk_mq = false;
2298 dm_init_md_queue(md);
2301 * Initialize aspects of queue that aren't relevant for blk-mq
2303 md->queue->backing_dev_info.congested_fn = dm_any_congested;
2304 blk_queue_bounce_limit(md->queue, BLK_BOUNCE_ANY);
2307 static void cleanup_mapped_device(struct mapped_device *md)
2310 destroy_workqueue(md->wq);
2311 if (md->kworker_task)
2312 kthread_stop(md->kworker_task);
2313 mempool_destroy(md->io_pool);
2314 mempool_destroy(md->rq_pool);
2316 bioset_free(md->bs);
2319 spin_lock(&_minor_lock);
2320 md->disk->private_data = NULL;
2321 spin_unlock(&_minor_lock);
2322 del_gendisk(md->disk);
2327 blk_cleanup_queue(md->queue);
2329 cleanup_srcu_struct(&md->io_barrier);
2338 * Allocate and initialise a blank device with a given minor.
2340 static struct mapped_device *alloc_dev(int minor)
2343 struct mapped_device *md = kzalloc(sizeof(*md), GFP_KERNEL);
2347 DMWARN("unable to allocate device, out of memory.");
2351 if (!try_module_get(THIS_MODULE))
2352 goto bad_module_get;
2354 /* get a minor number for the dev */
2355 if (minor == DM_ANY_MINOR)
2356 r = next_free_minor(&minor);
2358 r = specific_minor(minor);
2362 r = init_srcu_struct(&md->io_barrier);
2364 goto bad_io_barrier;
2366 md->use_blk_mq = use_blk_mq;
2367 md->type = DM_TYPE_NONE;
2368 mutex_init(&md->suspend_lock);
2369 mutex_init(&md->type_lock);
2370 mutex_init(&md->table_devices_lock);
2371 spin_lock_init(&md->deferred_lock);
2372 atomic_set(&md->holders, 1);
2373 atomic_set(&md->open_count, 0);
2374 atomic_set(&md->event_nr, 0);
2375 atomic_set(&md->uevent_seq, 0);
2376 INIT_LIST_HEAD(&md->uevent_list);
2377 INIT_LIST_HEAD(&md->table_devices);
2378 spin_lock_init(&md->uevent_lock);
2380 md->queue = blk_alloc_queue(GFP_KERNEL);
2384 dm_init_md_queue(md);
2386 md->disk = alloc_disk(1);
2390 atomic_set(&md->pending[0], 0);
2391 atomic_set(&md->pending[1], 0);
2392 init_waitqueue_head(&md->wait);
2393 INIT_WORK(&md->work, dm_wq_work);
2394 init_waitqueue_head(&md->eventq);
2395 init_completion(&md->kobj_holder.completion);
2396 md->kworker_task = NULL;
2398 md->disk->major = _major;
2399 md->disk->first_minor = minor;
2400 md->disk->fops = &dm_blk_dops;
2401 md->disk->queue = md->queue;
2402 md->disk->private_data = md;
2403 sprintf(md->disk->disk_name, "dm-%d", minor);
2405 format_dev_t(md->name, MKDEV(_major, minor));
2407 md->wq = alloc_workqueue("kdmflush", WQ_MEM_RECLAIM, 0);
2411 md->bdev = bdget_disk(md->disk, 0);
2415 bio_init(&md->flush_bio);
2416 md->flush_bio.bi_bdev = md->bdev;
2417 md->flush_bio.bi_rw = WRITE_FLUSH;
2419 dm_stats_init(&md->stats);
2421 /* Populate the mapping, nobody knows we exist yet */
2422 spin_lock(&_minor_lock);
2423 old_md = idr_replace(&_minor_idr, md, minor);
2424 spin_unlock(&_minor_lock);
2426 BUG_ON(old_md != MINOR_ALLOCED);
2431 cleanup_mapped_device(md);
2435 module_put(THIS_MODULE);
2441 static void unlock_fs(struct mapped_device *md);
2443 static void free_dev(struct mapped_device *md)
2445 int minor = MINOR(disk_devt(md->disk));
2449 cleanup_mapped_device(md);
2451 blk_mq_free_tag_set(&md->tag_set);
2453 free_table_devices(&md->table_devices);
2454 dm_stats_cleanup(&md->stats);
2457 module_put(THIS_MODULE);
2461 static void __bind_mempools(struct mapped_device *md, struct dm_table *t)
2463 struct dm_md_mempools *p = dm_table_get_md_mempools(t);
2466 /* The md already has necessary mempools. */
2467 if (dm_table_get_type(t) == DM_TYPE_BIO_BASED) {
2469 * Reload bioset because front_pad may have changed
2470 * because a different table was loaded.
2472 bioset_free(md->bs);
2477 * There's no need to reload with request-based dm
2478 * because the size of front_pad doesn't change.
2479 * Note for future: If you are to reload bioset,
2480 * prep-ed requests in the queue may refer
2481 * to bio from the old bioset, so you must walk
2482 * through the queue to unprep.
2487 BUG_ON(!p || md->io_pool || md->rq_pool || md->bs);
2489 md->io_pool = p->io_pool;
2491 md->rq_pool = p->rq_pool;
2497 /* mempool bind completed, no longer need any mempools in the table */
2498 dm_table_free_md_mempools(t);
2502 * Bind a table to the device.
2504 static void event_callback(void *context)
2506 unsigned long flags;
2508 struct mapped_device *md = (struct mapped_device *) context;
2510 spin_lock_irqsave(&md->uevent_lock, flags);
2511 list_splice_init(&md->uevent_list, &uevents);
2512 spin_unlock_irqrestore(&md->uevent_lock, flags);
2514 dm_send_uevents(&uevents, &disk_to_dev(md->disk)->kobj);
2516 atomic_inc(&md->event_nr);
2517 wake_up(&md->eventq);
2521 * Protected by md->suspend_lock obtained by dm_swap_table().
2523 static void __set_size(struct mapped_device *md, sector_t size)
2525 set_capacity(md->disk, size);
2527 i_size_write(md->bdev->bd_inode, (loff_t)size << SECTOR_SHIFT);
2531 * Returns old map, which caller must destroy.
2533 static struct dm_table *__bind(struct mapped_device *md, struct dm_table *t,
2534 struct queue_limits *limits)
2536 struct dm_table *old_map;
2537 struct request_queue *q = md->queue;
2540 size = dm_table_get_size(t);
2543 * Wipe any geometry if the size of the table changed.
2545 if (size != dm_get_size(md))
2546 memset(&md->geometry, 0, sizeof(md->geometry));
2548 __set_size(md, size);
2550 dm_table_event_callback(t, event_callback, md);
2553 * The queue hasn't been stopped yet, if the old table type wasn't
2554 * for request-based during suspension. So stop it to prevent
2555 * I/O mapping before resume.
2556 * This must be done before setting the queue restrictions,
2557 * because request-based dm may be run just after the setting.
2559 if (dm_table_request_based(t))
2562 __bind_mempools(md, t);
2564 old_map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2565 rcu_assign_pointer(md->map, t);
2566 md->immutable_target_type = dm_table_get_immutable_target_type(t);
2568 dm_table_set_restrictions(t, q, limits);
2576 * Returns unbound table for the caller to free.
2578 static struct dm_table *__unbind(struct mapped_device *md)
2580 struct dm_table *map = rcu_dereference_protected(md->map, 1);
2585 dm_table_event_callback(map, NULL, NULL);
2586 RCU_INIT_POINTER(md->map, NULL);
2593 * Constructor for a new device.
2595 int dm_create(int minor, struct mapped_device **result)
2597 struct mapped_device *md;
2599 md = alloc_dev(minor);
2610 * Functions to manage md->type.
2611 * All are required to hold md->type_lock.
2613 void dm_lock_md_type(struct mapped_device *md)
2615 mutex_lock(&md->type_lock);
2618 void dm_unlock_md_type(struct mapped_device *md)
2620 mutex_unlock(&md->type_lock);
2623 void dm_set_md_type(struct mapped_device *md, unsigned type)
2625 BUG_ON(!mutex_is_locked(&md->type_lock));
2629 unsigned dm_get_md_type(struct mapped_device *md)
2631 BUG_ON(!mutex_is_locked(&md->type_lock));
2635 struct target_type *dm_get_immutable_target_type(struct mapped_device *md)
2637 return md->immutable_target_type;
2641 * The queue_limits are only valid as long as you have a reference
2644 struct queue_limits *dm_get_queue_limits(struct mapped_device *md)
2646 BUG_ON(!atomic_read(&md->holders));
2647 return &md->queue->limits;
2649 EXPORT_SYMBOL_GPL(dm_get_queue_limits);
2651 static void init_rq_based_worker_thread(struct mapped_device *md)
2653 /* Initialize the request-based DM worker thread */
2654 init_kthread_worker(&md->kworker);
2655 md->kworker_task = kthread_run(kthread_worker_fn, &md->kworker,
2656 "kdmwork-%s", dm_device_name(md));
2660 * Fully initialize a request-based queue (->elevator, ->request_fn, etc).
2662 static int dm_init_request_based_queue(struct mapped_device *md)
2664 struct request_queue *q = NULL;
2666 /* Fully initialize the queue */
2667 q = blk_init_allocated_queue(md->queue, dm_request_fn, NULL);
2671 /* disable dm_request_fn's merge heuristic by default */
2672 md->seq_rq_merge_deadline_usecs = 0;
2675 dm_init_old_md_queue(md);
2676 blk_queue_softirq_done(md->queue, dm_softirq_done);
2677 blk_queue_prep_rq(md->queue, dm_prep_fn);
2679 init_rq_based_worker_thread(md);
2681 elv_register_queue(md->queue);
2686 static int dm_mq_init_request(void *data, struct request *rq,
2687 unsigned int hctx_idx, unsigned int request_idx,
2688 unsigned int numa_node)
2690 struct mapped_device *md = data;
2691 struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
2694 * Must initialize md member of tio, otherwise it won't
2695 * be available in dm_mq_queue_rq.
2702 static int dm_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
2703 const struct blk_mq_queue_data *bd)
2705 struct request *rq = bd->rq;
2706 struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
2707 struct mapped_device *md = tio->md;
2709 struct dm_table *map = dm_get_live_table(md, &srcu_idx);
2710 struct dm_target *ti;
2713 /* always use block 0 to find the target for flushes for now */
2715 if (!(rq->cmd_flags & REQ_FLUSH))
2716 pos = blk_rq_pos(rq);
2718 ti = dm_table_find_target(map, pos);
2719 if (!dm_target_is_valid(ti)) {
2720 dm_put_live_table(md, srcu_idx);
2721 DMERR_LIMIT("request attempted access beyond the end of device");
2723 * Must perform setup, that rq_completed() requires,
2724 * before returning BLK_MQ_RQ_QUEUE_ERROR
2726 dm_start_request(md, rq);
2727 return BLK_MQ_RQ_QUEUE_ERROR;
2729 dm_put_live_table(md, srcu_idx);
2731 if (ti->type->busy && ti->type->busy(ti))
2732 return BLK_MQ_RQ_QUEUE_BUSY;
2734 dm_start_request(md, rq);
2736 /* Init tio using md established in .init_request */
2737 init_tio(tio, rq, md);
2740 * Establish tio->ti before queuing work (map_tio_request)
2741 * or making direct call to map_request().
2745 /* Clone the request if underlying devices aren't blk-mq */
2746 if (dm_table_get_type(map) == DM_TYPE_REQUEST_BASED) {
2747 /* clone request is allocated at the end of the pdu */
2748 tio->clone = (void *)blk_mq_rq_to_pdu(rq) + sizeof(struct dm_rq_target_io);
2749 (void) clone_rq(rq, md, tio, GFP_ATOMIC);
2750 queue_kthread_work(&md->kworker, &tio->work);
2752 /* Direct call is fine since .queue_rq allows allocations */
2753 if (map_request(tio, rq, md) == DM_MAPIO_REQUEUE) {
2754 /* Undo dm_start_request() before requeuing */
2755 rq_end_stats(md, rq);
2756 rq_completed(md, rq_data_dir(rq), false);
2757 return BLK_MQ_RQ_QUEUE_BUSY;
2761 return BLK_MQ_RQ_QUEUE_OK;
2764 static struct blk_mq_ops dm_mq_ops = {
2765 .queue_rq = dm_mq_queue_rq,
2766 .map_queue = blk_mq_map_queue,
2767 .complete = dm_softirq_done,
2768 .init_request = dm_mq_init_request,
2771 static int dm_init_request_based_blk_mq_queue(struct mapped_device *md)
2773 unsigned md_type = dm_get_md_type(md);
2774 struct request_queue *q;
2777 memset(&md->tag_set, 0, sizeof(md->tag_set));
2778 md->tag_set.ops = &dm_mq_ops;
2779 md->tag_set.queue_depth = BLKDEV_MAX_RQ;
2780 md->tag_set.numa_node = NUMA_NO_NODE;
2781 md->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
2782 md->tag_set.nr_hw_queues = 1;
2783 if (md_type == DM_TYPE_REQUEST_BASED) {
2784 /* make the memory for non-blk-mq clone part of the pdu */
2785 md->tag_set.cmd_size = sizeof(struct dm_rq_target_io) + sizeof(struct request);
2787 md->tag_set.cmd_size = sizeof(struct dm_rq_target_io);
2788 md->tag_set.driver_data = md;
2790 err = blk_mq_alloc_tag_set(&md->tag_set);
2794 q = blk_mq_init_allocated_queue(&md->tag_set, md->queue);
2800 dm_init_md_queue(md);
2802 /* backfill 'mq' sysfs registration normally done in blk_register_queue */
2803 blk_mq_register_disk(md->disk);
2805 if (md_type == DM_TYPE_REQUEST_BASED)
2806 init_rq_based_worker_thread(md);
2811 blk_mq_free_tag_set(&md->tag_set);
2815 static unsigned filter_md_type(unsigned type, struct mapped_device *md)
2817 if (type == DM_TYPE_BIO_BASED)
2820 return !md->use_blk_mq ? DM_TYPE_REQUEST_BASED : DM_TYPE_MQ_REQUEST_BASED;
2824 * Setup the DM device's queue based on md's type
2826 int dm_setup_md_queue(struct mapped_device *md)
2829 unsigned md_type = filter_md_type(dm_get_md_type(md), md);
2832 case DM_TYPE_REQUEST_BASED:
2833 r = dm_init_request_based_queue(md);
2835 DMWARN("Cannot initialize queue for request-based mapped device");
2839 case DM_TYPE_MQ_REQUEST_BASED:
2840 r = dm_init_request_based_blk_mq_queue(md);
2842 DMWARN("Cannot initialize queue for request-based blk-mq mapped device");
2846 case DM_TYPE_BIO_BASED:
2847 dm_init_old_md_queue(md);
2848 blk_queue_make_request(md->queue, dm_make_request);
2850 * DM handles splitting bios as needed. Free the bio_split bioset
2851 * since it won't be used (saves 1 process per bio-based DM device).
2853 bioset_free(md->queue->bio_split);
2854 md->queue->bio_split = NULL;
2861 struct mapped_device *dm_get_md(dev_t dev)
2863 struct mapped_device *md;
2864 unsigned minor = MINOR(dev);
2866 if (MAJOR(dev) != _major || minor >= (1 << MINORBITS))
2869 spin_lock(&_minor_lock);
2871 md = idr_find(&_minor_idr, minor);
2873 if ((md == MINOR_ALLOCED ||
2874 (MINOR(disk_devt(dm_disk(md))) != minor) ||
2875 dm_deleting_md(md) ||
2876 test_bit(DMF_FREEING, &md->flags))) {
2884 spin_unlock(&_minor_lock);
2888 EXPORT_SYMBOL_GPL(dm_get_md);
2890 void *dm_get_mdptr(struct mapped_device *md)
2892 return md->interface_ptr;
2895 void dm_set_mdptr(struct mapped_device *md, void *ptr)
2897 md->interface_ptr = ptr;
2900 void dm_get(struct mapped_device *md)
2902 atomic_inc(&md->holders);
2903 BUG_ON(test_bit(DMF_FREEING, &md->flags));
2906 int dm_hold(struct mapped_device *md)
2908 spin_lock(&_minor_lock);
2909 if (test_bit(DMF_FREEING, &md->flags)) {
2910 spin_unlock(&_minor_lock);
2914 spin_unlock(&_minor_lock);
2917 EXPORT_SYMBOL_GPL(dm_hold);
2919 const char *dm_device_name(struct mapped_device *md)
2923 EXPORT_SYMBOL_GPL(dm_device_name);
2925 static void __dm_destroy(struct mapped_device *md, bool wait)
2927 struct request_queue *q = dm_get_md_queue(md);
2928 struct dm_table *map;
2933 spin_lock(&_minor_lock);
2934 idr_replace(&_minor_idr, MINOR_ALLOCED, MINOR(disk_devt(dm_disk(md))));
2935 set_bit(DMF_FREEING, &md->flags);
2936 spin_unlock(&_minor_lock);
2938 spin_lock_irq(q->queue_lock);
2939 queue_flag_set(QUEUE_FLAG_DYING, q);
2940 spin_unlock_irq(q->queue_lock);
2942 if (dm_request_based(md) && md->kworker_task)
2943 flush_kthread_worker(&md->kworker);
2946 * Take suspend_lock so that presuspend and postsuspend methods
2947 * do not race with internal suspend.
2949 mutex_lock(&md->suspend_lock);
2950 map = dm_get_live_table(md, &srcu_idx);
2951 if (!dm_suspended_md(md)) {
2952 dm_table_presuspend_targets(map);
2953 dm_table_postsuspend_targets(map);
2955 /* dm_put_live_table must be before msleep, otherwise deadlock is possible */
2956 dm_put_live_table(md, srcu_idx);
2957 mutex_unlock(&md->suspend_lock);
2960 * Rare, but there may be I/O requests still going to complete,
2961 * for example. Wait for all references to disappear.
2962 * No one should increment the reference count of the mapped_device,
2963 * after the mapped_device state becomes DMF_FREEING.
2966 while (atomic_read(&md->holders))
2968 else if (atomic_read(&md->holders))
2969 DMWARN("%s: Forcibly removing mapped_device still in use! (%d users)",
2970 dm_device_name(md), atomic_read(&md->holders));
2973 dm_table_destroy(__unbind(md));
2977 void dm_destroy(struct mapped_device *md)
2979 __dm_destroy(md, true);
2982 void dm_destroy_immediate(struct mapped_device *md)
2984 __dm_destroy(md, false);
2987 void dm_put(struct mapped_device *md)
2989 atomic_dec(&md->holders);
2991 EXPORT_SYMBOL_GPL(dm_put);
2993 static int dm_wait_for_completion(struct mapped_device *md, int interruptible)
2996 DECLARE_WAITQUEUE(wait, current);
2998 add_wait_queue(&md->wait, &wait);
3001 set_current_state(interruptible);
3003 if (!md_in_flight(md))
3006 if (interruptible == TASK_INTERRUPTIBLE &&
3007 signal_pending(current)) {
3014 set_current_state(TASK_RUNNING);
3016 remove_wait_queue(&md->wait, &wait);
3022 * Process the deferred bios
3024 static void dm_wq_work(struct work_struct *work)
3026 struct mapped_device *md = container_of(work, struct mapped_device,
3030 struct dm_table *map;
3032 map = dm_get_live_table(md, &srcu_idx);
3034 while (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
3035 spin_lock_irq(&md->deferred_lock);
3036 c = bio_list_pop(&md->deferred);
3037 spin_unlock_irq(&md->deferred_lock);
3042 if (dm_request_based(md))
3043 generic_make_request(c);
3045 __split_and_process_bio(md, map, c);
3048 dm_put_live_table(md, srcu_idx);
3051 static void dm_queue_flush(struct mapped_device *md)
3053 clear_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
3054 smp_mb__after_atomic();
3055 queue_work(md->wq, &md->work);
3059 * Swap in a new table, returning the old one for the caller to destroy.
3061 struct dm_table *dm_swap_table(struct mapped_device *md, struct dm_table *table)
3063 struct dm_table *live_map = NULL, *map = ERR_PTR(-EINVAL);
3064 struct queue_limits limits;
3067 mutex_lock(&md->suspend_lock);
3069 /* device must be suspended */
3070 if (!dm_suspended_md(md))
3074 * If the new table has no data devices, retain the existing limits.
3075 * This helps multipath with queue_if_no_path if all paths disappear,
3076 * then new I/O is queued based on these limits, and then some paths
3079 if (dm_table_has_no_data_devices(table)) {
3080 live_map = dm_get_live_table_fast(md);
3082 limits = md->queue->limits;
3083 dm_put_live_table_fast(md);
3087 r = dm_calculate_queue_limits(table, &limits);
3094 map = __bind(md, table, &limits);
3097 mutex_unlock(&md->suspend_lock);
3102 * Functions to lock and unlock any filesystem running on the
3105 static int lock_fs(struct mapped_device *md)
3109 WARN_ON(md->frozen_sb);
3111 md->frozen_sb = freeze_bdev(md->bdev);
3112 if (IS_ERR(md->frozen_sb)) {
3113 r = PTR_ERR(md->frozen_sb);
3114 md->frozen_sb = NULL;
3118 set_bit(DMF_FROZEN, &md->flags);
3123 static void unlock_fs(struct mapped_device *md)
3125 if (!test_bit(DMF_FROZEN, &md->flags))
3128 thaw_bdev(md->bdev, md->frozen_sb);
3129 md->frozen_sb = NULL;
3130 clear_bit(DMF_FROZEN, &md->flags);
3134 * If __dm_suspend returns 0, the device is completely quiescent
3135 * now. There is no request-processing activity. All new requests
3136 * are being added to md->deferred list.
3138 * Caller must hold md->suspend_lock
3140 static int __dm_suspend(struct mapped_device *md, struct dm_table *map,
3141 unsigned suspend_flags, int interruptible,
3142 int dmf_suspended_flag)
3144 bool do_lockfs = suspend_flags & DM_SUSPEND_LOCKFS_FLAG;
3145 bool noflush = suspend_flags & DM_SUSPEND_NOFLUSH_FLAG;
3149 * DMF_NOFLUSH_SUSPENDING must be set before presuspend.
3150 * This flag is cleared before dm_suspend returns.
3153 set_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
3156 * This gets reverted if there's an error later and the targets
3157 * provide the .presuspend_undo hook.
3159 dm_table_presuspend_targets(map);
3162 * Flush I/O to the device.
3163 * Any I/O submitted after lock_fs() may not be flushed.
3164 * noflush takes precedence over do_lockfs.
3165 * (lock_fs() flushes I/Os and waits for them to complete.)
3167 if (!noflush && do_lockfs) {
3170 dm_table_presuspend_undo_targets(map);
3176 * Here we must make sure that no processes are submitting requests
3177 * to target drivers i.e. no one may be executing
3178 * __split_and_process_bio. This is called from dm_request and
3181 * To get all processes out of __split_and_process_bio in dm_request,
3182 * we take the write lock. To prevent any process from reentering
3183 * __split_and_process_bio from dm_request and quiesce the thread
3184 * (dm_wq_work), we set BMF_BLOCK_IO_FOR_SUSPEND and call
3185 * flush_workqueue(md->wq).
3187 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
3189 synchronize_srcu(&md->io_barrier);
3192 * Stop md->queue before flushing md->wq in case request-based
3193 * dm defers requests to md->wq from md->queue.
3195 if (dm_request_based(md)) {
3196 stop_queue(md->queue);
3197 if (md->kworker_task)
3198 flush_kthread_worker(&md->kworker);
3201 flush_workqueue(md->wq);
3204 * At this point no more requests are entering target request routines.
3205 * We call dm_wait_for_completion to wait for all existing requests
3208 r = dm_wait_for_completion(md, interruptible);
3210 set_bit(dmf_suspended_flag, &md->flags);
3213 clear_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
3215 synchronize_srcu(&md->io_barrier);
3217 /* were we interrupted ? */
3221 if (dm_request_based(md))
3222 start_queue(md->queue);
3225 dm_table_presuspend_undo_targets(map);
3226 /* pushback list is already flushed, so skip flush */
3233 * We need to be able to change a mapping table under a mounted
3234 * filesystem. For example we might want to move some data in
3235 * the background. Before the table can be swapped with
3236 * dm_bind_table, dm_suspend must be called to flush any in
3237 * flight bios and ensure that any further io gets deferred.
3240 * Suspend mechanism in request-based dm.
3242 * 1. Flush all I/Os by lock_fs() if needed.
3243 * 2. Stop dispatching any I/O by stopping the request_queue.
3244 * 3. Wait for all in-flight I/Os to be completed or requeued.
3246 * To abort suspend, start the request_queue.
3248 int dm_suspend(struct mapped_device *md, unsigned suspend_flags)
3250 struct dm_table *map = NULL;
3254 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
3256 if (dm_suspended_md(md)) {
3261 if (dm_suspended_internally_md(md)) {
3262 /* already internally suspended, wait for internal resume */
3263 mutex_unlock(&md->suspend_lock);
3264 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
3270 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3272 r = __dm_suspend(md, map, suspend_flags, TASK_INTERRUPTIBLE, DMF_SUSPENDED);
3276 dm_table_postsuspend_targets(map);
3279 mutex_unlock(&md->suspend_lock);
3283 static int __dm_resume(struct mapped_device *md, struct dm_table *map)
3286 int r = dm_table_resume_targets(map);
3294 * Flushing deferred I/Os must be done after targets are resumed
3295 * so that mapping of targets can work correctly.
3296 * Request-based dm is queueing the deferred I/Os in its request_queue.
3298 if (dm_request_based(md))
3299 start_queue(md->queue);
3306 int dm_resume(struct mapped_device *md)
3309 struct dm_table *map = NULL;
3313 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
3315 if (!dm_suspended_md(md))
3318 if (dm_suspended_internally_md(md)) {
3319 /* already internally suspended, wait for internal resume */
3320 mutex_unlock(&md->suspend_lock);
3321 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
3327 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3328 if (!map || !dm_table_get_size(map))
3331 r = __dm_resume(md, map);
3335 clear_bit(DMF_SUSPENDED, &md->flags);
3337 mutex_unlock(&md->suspend_lock);
3343 * Internal suspend/resume works like userspace-driven suspend. It waits
3344 * until all bios finish and prevents issuing new bios to the target drivers.
3345 * It may be used only from the kernel.
3348 static void __dm_internal_suspend(struct mapped_device *md, unsigned suspend_flags)
3350 struct dm_table *map = NULL;
3352 if (md->internal_suspend_count++)
3353 return; /* nested internal suspend */
3355 if (dm_suspended_md(md)) {
3356 set_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3357 return; /* nest suspend */
3360 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3363 * Using TASK_UNINTERRUPTIBLE because only NOFLUSH internal suspend is
3364 * supported. Properly supporting a TASK_INTERRUPTIBLE internal suspend
3365 * would require changing .presuspend to return an error -- avoid this
3366 * until there is a need for more elaborate variants of internal suspend.
3368 (void) __dm_suspend(md, map, suspend_flags, TASK_UNINTERRUPTIBLE,
3369 DMF_SUSPENDED_INTERNALLY);
3371 dm_table_postsuspend_targets(map);
3374 static void __dm_internal_resume(struct mapped_device *md)
3376 BUG_ON(!md->internal_suspend_count);
3378 if (--md->internal_suspend_count)
3379 return; /* resume from nested internal suspend */
3381 if (dm_suspended_md(md))
3382 goto done; /* resume from nested suspend */
3385 * NOTE: existing callers don't need to call dm_table_resume_targets
3386 * (which may fail -- so best to avoid it for now by passing NULL map)
3388 (void) __dm_resume(md, NULL);
3391 clear_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3392 smp_mb__after_atomic();
3393 wake_up_bit(&md->flags, DMF_SUSPENDED_INTERNALLY);
3396 void dm_internal_suspend_noflush(struct mapped_device *md)
3398 mutex_lock(&md->suspend_lock);
3399 __dm_internal_suspend(md, DM_SUSPEND_NOFLUSH_FLAG);
3400 mutex_unlock(&md->suspend_lock);
3402 EXPORT_SYMBOL_GPL(dm_internal_suspend_noflush);
3404 void dm_internal_resume(struct mapped_device *md)
3406 mutex_lock(&md->suspend_lock);
3407 __dm_internal_resume(md);
3408 mutex_unlock(&md->suspend_lock);
3410 EXPORT_SYMBOL_GPL(dm_internal_resume);
3413 * Fast variants of internal suspend/resume hold md->suspend_lock,
3414 * which prevents interaction with userspace-driven suspend.
3417 void dm_internal_suspend_fast(struct mapped_device *md)
3419 mutex_lock(&md->suspend_lock);
3420 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
3423 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
3424 synchronize_srcu(&md->io_barrier);
3425 flush_workqueue(md->wq);
3426 dm_wait_for_completion(md, TASK_UNINTERRUPTIBLE);
3428 EXPORT_SYMBOL_GPL(dm_internal_suspend_fast);
3430 void dm_internal_resume_fast(struct mapped_device *md)
3432 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
3438 mutex_unlock(&md->suspend_lock);
3440 EXPORT_SYMBOL_GPL(dm_internal_resume_fast);
3442 /*-----------------------------------------------------------------
3443 * Event notification.
3444 *---------------------------------------------------------------*/
3445 int dm_kobject_uevent(struct mapped_device *md, enum kobject_action action,
3448 char udev_cookie[DM_COOKIE_LENGTH];
3449 char *envp[] = { udev_cookie, NULL };
3452 return kobject_uevent(&disk_to_dev(md->disk)->kobj, action);
3454 snprintf(udev_cookie, DM_COOKIE_LENGTH, "%s=%u",
3455 DM_COOKIE_ENV_VAR_NAME, cookie);
3456 return kobject_uevent_env(&disk_to_dev(md->disk)->kobj,
3461 uint32_t dm_next_uevent_seq(struct mapped_device *md)
3463 return atomic_add_return(1, &md->uevent_seq);
3466 uint32_t dm_get_event_nr(struct mapped_device *md)
3468 return atomic_read(&md->event_nr);
3471 int dm_wait_event(struct mapped_device *md, int event_nr)
3473 return wait_event_interruptible(md->eventq,
3474 (event_nr != atomic_read(&md->event_nr)));
3477 void dm_uevent_add(struct mapped_device *md, struct list_head *elist)
3479 unsigned long flags;
3481 spin_lock_irqsave(&md->uevent_lock, flags);
3482 list_add(elist, &md->uevent_list);
3483 spin_unlock_irqrestore(&md->uevent_lock, flags);
3487 * The gendisk is only valid as long as you have a reference
3490 struct gendisk *dm_disk(struct mapped_device *md)
3494 EXPORT_SYMBOL_GPL(dm_disk);
3496 struct kobject *dm_kobject(struct mapped_device *md)
3498 return &md->kobj_holder.kobj;
3501 struct mapped_device *dm_get_from_kobject(struct kobject *kobj)
3503 struct mapped_device *md;
3505 md = container_of(kobj, struct mapped_device, kobj_holder.kobj);
3507 if (test_bit(DMF_FREEING, &md->flags) ||
3515 int dm_suspended_md(struct mapped_device *md)
3517 return test_bit(DMF_SUSPENDED, &md->flags);
3520 int dm_suspended_internally_md(struct mapped_device *md)
3522 return test_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3525 int dm_test_deferred_remove_flag(struct mapped_device *md)
3527 return test_bit(DMF_DEFERRED_REMOVE, &md->flags);
3530 int dm_suspended(struct dm_target *ti)
3532 return dm_suspended_md(dm_table_get_md(ti->table));
3534 EXPORT_SYMBOL_GPL(dm_suspended);
3536 int dm_noflush_suspending(struct dm_target *ti)
3538 return __noflush_suspending(dm_table_get_md(ti->table));
3540 EXPORT_SYMBOL_GPL(dm_noflush_suspending);
3542 struct dm_md_mempools *dm_alloc_md_mempools(struct mapped_device *md, unsigned type,
3543 unsigned integrity, unsigned per_bio_data_size)
3545 struct dm_md_mempools *pools = kzalloc(sizeof(*pools), GFP_KERNEL);
3546 struct kmem_cache *cachep = NULL;
3547 unsigned int pool_size = 0;
3548 unsigned int front_pad;
3553 type = filter_md_type(type, md);
3556 case DM_TYPE_BIO_BASED:
3558 pool_size = dm_get_reserved_bio_based_ios();
3559 front_pad = roundup(per_bio_data_size, __alignof__(struct dm_target_io)) + offsetof(struct dm_target_io, clone);
3561 case DM_TYPE_REQUEST_BASED:
3562 cachep = _rq_tio_cache;
3563 pool_size = dm_get_reserved_rq_based_ios();
3564 pools->rq_pool = mempool_create_slab_pool(pool_size, _rq_cache);
3565 if (!pools->rq_pool)
3567 /* fall through to setup remaining rq-based pools */
3568 case DM_TYPE_MQ_REQUEST_BASED:
3570 pool_size = dm_get_reserved_rq_based_ios();
3571 front_pad = offsetof(struct dm_rq_clone_bio_info, clone);
3572 /* per_bio_data_size is not used. See __bind_mempools(). */
3573 WARN_ON(per_bio_data_size != 0);
3580 pools->io_pool = mempool_create_slab_pool(pool_size, cachep);
3581 if (!pools->io_pool)
3585 pools->bs = bioset_create_nobvec(pool_size, front_pad);
3589 if (integrity && bioset_integrity_create(pools->bs, pool_size))
3595 dm_free_md_mempools(pools);
3600 void dm_free_md_mempools(struct dm_md_mempools *pools)
3605 mempool_destroy(pools->io_pool);
3606 mempool_destroy(pools->rq_pool);
3609 bioset_free(pools->bs);
3614 static int dm_pr_register(struct block_device *bdev, u64 old_key, u64 new_key,
3617 struct mapped_device *md = bdev->bd_disk->private_data;
3618 const struct pr_ops *ops;
3619 struct dm_target *tgt;
3623 r = dm_get_live_table_for_ioctl(md, &tgt, &bdev, &mode, &srcu_idx);
3627 ops = bdev->bd_disk->fops->pr_ops;
3628 if (ops && ops->pr_register)
3629 r = ops->pr_register(bdev, old_key, new_key, flags);
3633 dm_put_live_table(md, srcu_idx);
3637 static int dm_pr_reserve(struct block_device *bdev, u64 key, enum pr_type type,
3640 struct mapped_device *md = bdev->bd_disk->private_data;
3641 const struct pr_ops *ops;
3642 struct dm_target *tgt;
3646 r = dm_get_live_table_for_ioctl(md, &tgt, &bdev, &mode, &srcu_idx);
3650 ops = bdev->bd_disk->fops->pr_ops;
3651 if (ops && ops->pr_reserve)
3652 r = ops->pr_reserve(bdev, key, type, flags);
3656 dm_put_live_table(md, srcu_idx);
3660 static int dm_pr_release(struct block_device *bdev, u64 key, enum pr_type type)
3662 struct mapped_device *md = bdev->bd_disk->private_data;
3663 const struct pr_ops *ops;
3664 struct dm_target *tgt;
3668 r = dm_get_live_table_for_ioctl(md, &tgt, &bdev, &mode, &srcu_idx);
3672 ops = bdev->bd_disk->fops->pr_ops;
3673 if (ops && ops->pr_release)
3674 r = ops->pr_release(bdev, key, type);
3678 dm_put_live_table(md, srcu_idx);
3682 static int dm_pr_preempt(struct block_device *bdev, u64 old_key, u64 new_key,
3683 enum pr_type type, bool abort)
3685 struct mapped_device *md = bdev->bd_disk->private_data;
3686 const struct pr_ops *ops;
3687 struct dm_target *tgt;
3691 r = dm_get_live_table_for_ioctl(md, &tgt, &bdev, &mode, &srcu_idx);
3695 ops = bdev->bd_disk->fops->pr_ops;
3696 if (ops && ops->pr_preempt)
3697 r = ops->pr_preempt(bdev, old_key, new_key, type, abort);
3701 dm_put_live_table(md, srcu_idx);
3705 static int dm_pr_clear(struct block_device *bdev, u64 key)
3707 struct mapped_device *md = bdev->bd_disk->private_data;
3708 const struct pr_ops *ops;
3709 struct dm_target *tgt;
3713 r = dm_get_live_table_for_ioctl(md, &tgt, &bdev, &mode, &srcu_idx);
3717 ops = bdev->bd_disk->fops->pr_ops;
3718 if (ops && ops->pr_clear)
3719 r = ops->pr_clear(bdev, key);
3723 dm_put_live_table(md, srcu_idx);
3727 static const struct pr_ops dm_pr_ops = {
3728 .pr_register = dm_pr_register,
3729 .pr_reserve = dm_pr_reserve,
3730 .pr_release = dm_pr_release,
3731 .pr_preempt = dm_pr_preempt,
3732 .pr_clear = dm_pr_clear,
3735 static const struct block_device_operations dm_blk_dops = {
3736 .open = dm_blk_open,
3737 .release = dm_blk_close,
3738 .ioctl = dm_blk_ioctl,
3739 .getgeo = dm_blk_getgeo,
3740 .pr_ops = &dm_pr_ops,
3741 .owner = THIS_MODULE
3747 module_init(dm_init);
3748 module_exit(dm_exit);
3750 module_param(major, uint, 0);
3751 MODULE_PARM_DESC(major, "The major number of the device mapper");
3753 module_param(reserved_bio_based_ios, uint, S_IRUGO | S_IWUSR);
3754 MODULE_PARM_DESC(reserved_bio_based_ios, "Reserved IOs in bio-based mempools");
3756 module_param(reserved_rq_based_ios, uint, S_IRUGO | S_IWUSR);
3757 MODULE_PARM_DESC(reserved_rq_based_ios, "Reserved IOs in request-based mempools");
3759 module_param(use_blk_mq, bool, S_IRUGO | S_IWUSR);
3760 MODULE_PARM_DESC(use_blk_mq, "Use block multiqueue for request-based DM devices");
3762 MODULE_DESCRIPTION(DM_NAME " driver");
3763 MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
3764 MODULE_LICENSE("GPL");