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) {
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;
608 r = dm_get_live_table_for_ioctl(md, &tgt, &bdev, &mode, &srcu_idx);
614 * Target determined this ioctl is being issued against
615 * a logical partition of the parent bdev; so extra
616 * validation is needed.
618 r = scsi_verify_blk_ioctl(NULL, cmd);
623 r = __blkdev_driver_ioctl(bdev, mode, cmd, arg);
625 dm_put_live_table(md, srcu_idx);
629 static struct dm_io *alloc_io(struct mapped_device *md)
631 return mempool_alloc(md->io_pool, GFP_NOIO);
634 static void free_io(struct mapped_device *md, struct dm_io *io)
636 mempool_free(io, md->io_pool);
639 static void free_tio(struct mapped_device *md, struct dm_target_io *tio)
641 bio_put(&tio->clone);
644 static struct dm_rq_target_io *alloc_rq_tio(struct mapped_device *md,
647 return mempool_alloc(md->io_pool, gfp_mask);
650 static void free_rq_tio(struct dm_rq_target_io *tio)
652 mempool_free(tio, tio->md->io_pool);
655 static struct request *alloc_clone_request(struct mapped_device *md,
658 return mempool_alloc(md->rq_pool, gfp_mask);
661 static void free_clone_request(struct mapped_device *md, struct request *rq)
663 mempool_free(rq, md->rq_pool);
666 static int md_in_flight(struct mapped_device *md)
668 return atomic_read(&md->pending[READ]) +
669 atomic_read(&md->pending[WRITE]);
672 static void start_io_acct(struct dm_io *io)
674 struct mapped_device *md = io->md;
675 struct bio *bio = io->bio;
677 int rw = bio_data_dir(bio);
679 io->start_time = jiffies;
681 cpu = part_stat_lock();
682 part_round_stats(cpu, &dm_disk(md)->part0);
684 atomic_set(&dm_disk(md)->part0.in_flight[rw],
685 atomic_inc_return(&md->pending[rw]));
687 if (unlikely(dm_stats_used(&md->stats)))
688 dm_stats_account_io(&md->stats, bio->bi_rw, bio->bi_iter.bi_sector,
689 bio_sectors(bio), false, 0, &io->stats_aux);
692 static void end_io_acct(struct dm_io *io)
694 struct mapped_device *md = io->md;
695 struct bio *bio = io->bio;
696 unsigned long duration = jiffies - io->start_time;
698 int rw = bio_data_dir(bio);
700 generic_end_io_acct(rw, &dm_disk(md)->part0, io->start_time);
702 if (unlikely(dm_stats_used(&md->stats)))
703 dm_stats_account_io(&md->stats, bio->bi_rw, bio->bi_iter.bi_sector,
704 bio_sectors(bio), true, duration, &io->stats_aux);
707 * After this is decremented the bio must not be touched if it is
710 pending = atomic_dec_return(&md->pending[rw]);
711 atomic_set(&dm_disk(md)->part0.in_flight[rw], pending);
712 pending += atomic_read(&md->pending[rw^0x1]);
714 /* nudge anyone waiting on suspend queue */
720 * Add the bio to the list of deferred io.
722 static void queue_io(struct mapped_device *md, struct bio *bio)
726 spin_lock_irqsave(&md->deferred_lock, flags);
727 bio_list_add(&md->deferred, bio);
728 spin_unlock_irqrestore(&md->deferred_lock, flags);
729 queue_work(md->wq, &md->work);
733 * Everyone (including functions in this file), should use this
734 * function to access the md->map field, and make sure they call
735 * dm_put_live_table() when finished.
737 struct dm_table *dm_get_live_table(struct mapped_device *md, int *srcu_idx) __acquires(md->io_barrier)
739 *srcu_idx = srcu_read_lock(&md->io_barrier);
741 return srcu_dereference(md->map, &md->io_barrier);
744 void dm_put_live_table(struct mapped_device *md, int srcu_idx) __releases(md->io_barrier)
746 srcu_read_unlock(&md->io_barrier, srcu_idx);
749 void dm_sync_table(struct mapped_device *md)
751 synchronize_srcu(&md->io_barrier);
752 synchronize_rcu_expedited();
756 * A fast alternative to dm_get_live_table/dm_put_live_table.
757 * The caller must not block between these two functions.
759 static struct dm_table *dm_get_live_table_fast(struct mapped_device *md) __acquires(RCU)
762 return rcu_dereference(md->map);
765 static void dm_put_live_table_fast(struct mapped_device *md) __releases(RCU)
771 * Open a table device so we can use it as a map destination.
773 static int open_table_device(struct table_device *td, dev_t dev,
774 struct mapped_device *md)
776 static char *_claim_ptr = "I belong to device-mapper";
777 struct block_device *bdev;
781 BUG_ON(td->dm_dev.bdev);
783 bdev = blkdev_get_by_dev(dev, td->dm_dev.mode | FMODE_EXCL, _claim_ptr);
785 return PTR_ERR(bdev);
787 r = bd_link_disk_holder(bdev, dm_disk(md));
789 blkdev_put(bdev, td->dm_dev.mode | FMODE_EXCL);
793 td->dm_dev.bdev = bdev;
798 * Close a table device that we've been using.
800 static void close_table_device(struct table_device *td, struct mapped_device *md)
802 if (!td->dm_dev.bdev)
805 bd_unlink_disk_holder(td->dm_dev.bdev, dm_disk(md));
806 blkdev_put(td->dm_dev.bdev, td->dm_dev.mode | FMODE_EXCL);
807 td->dm_dev.bdev = NULL;
810 static struct table_device *find_table_device(struct list_head *l, dev_t dev,
812 struct table_device *td;
814 list_for_each_entry(td, l, list)
815 if (td->dm_dev.bdev->bd_dev == dev && td->dm_dev.mode == mode)
821 int dm_get_table_device(struct mapped_device *md, dev_t dev, fmode_t mode,
822 struct dm_dev **result) {
824 struct table_device *td;
826 mutex_lock(&md->table_devices_lock);
827 td = find_table_device(&md->table_devices, dev, mode);
829 td = kmalloc(sizeof(*td), GFP_KERNEL);
831 mutex_unlock(&md->table_devices_lock);
835 td->dm_dev.mode = mode;
836 td->dm_dev.bdev = NULL;
838 if ((r = open_table_device(td, dev, md))) {
839 mutex_unlock(&md->table_devices_lock);
844 format_dev_t(td->dm_dev.name, dev);
846 atomic_set(&td->count, 0);
847 list_add(&td->list, &md->table_devices);
849 atomic_inc(&td->count);
850 mutex_unlock(&md->table_devices_lock);
852 *result = &td->dm_dev;
855 EXPORT_SYMBOL_GPL(dm_get_table_device);
857 void dm_put_table_device(struct mapped_device *md, struct dm_dev *d)
859 struct table_device *td = container_of(d, struct table_device, dm_dev);
861 mutex_lock(&md->table_devices_lock);
862 if (atomic_dec_and_test(&td->count)) {
863 close_table_device(td, md);
867 mutex_unlock(&md->table_devices_lock);
869 EXPORT_SYMBOL(dm_put_table_device);
871 static void free_table_devices(struct list_head *devices)
873 struct list_head *tmp, *next;
875 list_for_each_safe(tmp, next, devices) {
876 struct table_device *td = list_entry(tmp, struct table_device, list);
878 DMWARN("dm_destroy: %s still exists with %d references",
879 td->dm_dev.name, atomic_read(&td->count));
885 * Get the geometry associated with a dm device
887 int dm_get_geometry(struct mapped_device *md, struct hd_geometry *geo)
895 * Set the geometry of a device.
897 int dm_set_geometry(struct mapped_device *md, struct hd_geometry *geo)
899 sector_t sz = (sector_t)geo->cylinders * geo->heads * geo->sectors;
901 if (geo->start > sz) {
902 DMWARN("Start sector is beyond the geometry limits.");
911 /*-----------------------------------------------------------------
913 * A more elegant soln is in the works that uses the queue
914 * merge fn, unfortunately there are a couple of changes to
915 * the block layer that I want to make for this. So in the
916 * interests of getting something for people to use I give
917 * you this clearly demarcated crap.
918 *---------------------------------------------------------------*/
920 static int __noflush_suspending(struct mapped_device *md)
922 return test_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
926 * Decrements the number of outstanding ios that a bio has been
927 * cloned into, completing the original io if necc.
929 static void dec_pending(struct dm_io *io, int error)
934 struct mapped_device *md = io->md;
936 /* Push-back supersedes any I/O errors */
937 if (unlikely(error)) {
938 spin_lock_irqsave(&io->endio_lock, flags);
939 if (!(io->error > 0 && __noflush_suspending(md)))
941 spin_unlock_irqrestore(&io->endio_lock, flags);
944 if (atomic_dec_and_test(&io->io_count)) {
945 if (io->error == DM_ENDIO_REQUEUE) {
947 * Target requested pushing back the I/O.
949 spin_lock_irqsave(&md->deferred_lock, flags);
950 if (__noflush_suspending(md))
951 bio_list_add_head(&md->deferred, io->bio);
953 /* noflush suspend was interrupted. */
955 spin_unlock_irqrestore(&md->deferred_lock, flags);
958 io_error = io->error;
963 if (io_error == DM_ENDIO_REQUEUE)
966 if ((bio->bi_rw & REQ_FLUSH) && bio->bi_iter.bi_size) {
968 * Preflush done for flush with data, reissue
971 bio->bi_rw &= ~REQ_FLUSH;
974 /* done with normal IO or empty flush */
975 trace_block_bio_complete(md->queue, bio, io_error);
976 bio->bi_error = io_error;
982 static void disable_write_same(struct mapped_device *md)
984 struct queue_limits *limits = dm_get_queue_limits(md);
986 /* device doesn't really support WRITE SAME, disable it */
987 limits->max_write_same_sectors = 0;
990 static void clone_endio(struct bio *bio)
992 int error = bio->bi_error;
994 struct dm_target_io *tio = container_of(bio, struct dm_target_io, clone);
995 struct dm_io *io = tio->io;
996 struct mapped_device *md = tio->io->md;
997 dm_endio_fn endio = tio->ti->type->end_io;
1000 r = endio(tio->ti, bio, error);
1001 if (r < 0 || r == DM_ENDIO_REQUEUE)
1003 * error and requeue request are handled
1007 else if (r == DM_ENDIO_INCOMPLETE)
1008 /* The target will handle the io */
1011 DMWARN("unimplemented target endio return value: %d", r);
1016 if (unlikely(r == -EREMOTEIO && (bio->bi_rw & REQ_WRITE_SAME) &&
1017 !bdev_get_queue(bio->bi_bdev)->limits.max_write_same_sectors))
1018 disable_write_same(md);
1021 dec_pending(io, error);
1025 * Partial completion handling for request-based dm
1027 static void end_clone_bio(struct bio *clone)
1029 struct dm_rq_clone_bio_info *info =
1030 container_of(clone, struct dm_rq_clone_bio_info, clone);
1031 struct dm_rq_target_io *tio = info->tio;
1032 struct bio *bio = info->orig;
1033 unsigned int nr_bytes = info->orig->bi_iter.bi_size;
1039 * An error has already been detected on the request.
1040 * Once error occurred, just let clone->end_io() handle
1044 else if (bio->bi_error) {
1046 * Don't notice the error to the upper layer yet.
1047 * The error handling decision is made by the target driver,
1048 * when the request is completed.
1050 tio->error = bio->bi_error;
1055 * I/O for the bio successfully completed.
1056 * Notice the data completion to the upper layer.
1060 * bios are processed from the head of the list.
1061 * So the completing bio should always be rq->bio.
1062 * If it's not, something wrong is happening.
1064 if (tio->orig->bio != bio)
1065 DMERR("bio completion is going in the middle of the request");
1068 * Update the original request.
1069 * Do not use blk_end_request() here, because it may complete
1070 * the original request before the clone, and break the ordering.
1072 blk_update_request(tio->orig, 0, nr_bytes);
1075 static struct dm_rq_target_io *tio_from_request(struct request *rq)
1077 return (rq->q->mq_ops ? blk_mq_rq_to_pdu(rq) : rq->special);
1080 static void rq_end_stats(struct mapped_device *md, struct request *orig)
1082 if (unlikely(dm_stats_used(&md->stats))) {
1083 struct dm_rq_target_io *tio = tio_from_request(orig);
1084 tio->duration_jiffies = jiffies - tio->duration_jiffies;
1085 dm_stats_account_io(&md->stats, orig->cmd_flags, blk_rq_pos(orig),
1086 tio->n_sectors, true, tio->duration_jiffies,
1092 * Don't touch any member of the md after calling this function because
1093 * the md may be freed in dm_put() at the end of this function.
1094 * Or do dm_get() before calling this function and dm_put() later.
1096 static void rq_completed(struct mapped_device *md, int rw, bool run_queue)
1098 atomic_dec(&md->pending[rw]);
1100 /* nudge anyone waiting on suspend queue */
1101 if (!md_in_flight(md))
1105 * Run this off this callpath, as drivers could invoke end_io while
1106 * inside their request_fn (and holding the queue lock). Calling
1107 * back into ->request_fn() could deadlock attempting to grab the
1111 if (md->queue->mq_ops)
1112 blk_mq_run_hw_queues(md->queue, true);
1114 blk_run_queue_async(md->queue);
1118 * dm_put() must be at the end of this function. See the comment above
1123 static void free_rq_clone(struct request *clone)
1125 struct dm_rq_target_io *tio = clone->end_io_data;
1126 struct mapped_device *md = tio->md;
1128 blk_rq_unprep_clone(clone);
1130 if (md->type == DM_TYPE_MQ_REQUEST_BASED)
1131 /* stacked on blk-mq queue(s) */
1132 tio->ti->type->release_clone_rq(clone);
1133 else if (!md->queue->mq_ops)
1134 /* request_fn queue stacked on request_fn queue(s) */
1135 free_clone_request(md, clone);
1137 * NOTE: for the blk-mq queue stacked on request_fn queue(s) case:
1138 * no need to call free_clone_request() because we leverage blk-mq by
1139 * allocating the clone at the end of the blk-mq pdu (see: clone_rq)
1142 if (!md->queue->mq_ops)
1147 * Complete the clone and the original request.
1148 * Must be called without clone's queue lock held,
1149 * see end_clone_request() for more details.
1151 static void dm_end_request(struct request *clone, int error)
1153 int rw = rq_data_dir(clone);
1154 struct dm_rq_target_io *tio = clone->end_io_data;
1155 struct mapped_device *md = tio->md;
1156 struct request *rq = tio->orig;
1158 if (rq->cmd_type == REQ_TYPE_BLOCK_PC) {
1159 rq->errors = clone->errors;
1160 rq->resid_len = clone->resid_len;
1164 * We are using the sense buffer of the original
1166 * So setting the length of the sense data is enough.
1168 rq->sense_len = clone->sense_len;
1171 free_rq_clone(clone);
1172 rq_end_stats(md, rq);
1174 blk_end_request_all(rq, error);
1176 blk_mq_end_request(rq, error);
1177 rq_completed(md, rw, true);
1180 static void dm_unprep_request(struct request *rq)
1182 struct dm_rq_target_io *tio = tio_from_request(rq);
1183 struct request *clone = tio->clone;
1185 if (!rq->q->mq_ops) {
1187 rq->cmd_flags &= ~REQ_DONTPREP;
1191 free_rq_clone(clone);
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 dm_unprep_request(rq);
1215 rq_end_stats(md, 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);
1335 blk_complete_request(rq);
1339 * Complete the not-mapped clone and the original request with the error status
1340 * through softirq context.
1341 * Target's rq_end_io() function isn't called.
1342 * This may be used when the target's map_rq() or clone_and_map_rq() functions fail.
1344 static void dm_kill_unmapped_request(struct request *rq, int error)
1346 rq->cmd_flags |= REQ_FAILED;
1347 dm_complete_request(rq, error);
1351 * Called with the clone's queue lock held (for non-blk-mq)
1353 static void end_clone_request(struct request *clone, int error)
1355 struct dm_rq_target_io *tio = clone->end_io_data;
1357 if (!clone->q->mq_ops) {
1359 * For just cleaning up the information of the queue in which
1360 * the clone was dispatched.
1361 * The clone is *NOT* freed actually here because it is alloced
1362 * from dm own mempool (REQ_ALLOCED isn't set).
1364 __blk_put_request(clone->q, clone);
1368 * Actual request completion is done in a softirq context which doesn't
1369 * hold the clone's queue lock. Otherwise, deadlock could occur because:
1370 * - another request may be submitted by the upper level driver
1371 * of the stacking during the completion
1372 * - the submission which requires queue lock may be done
1373 * against this clone's queue
1375 dm_complete_request(tio->orig, error);
1379 * Return maximum size of I/O possible at the supplied sector up to the current
1382 static sector_t max_io_len_target_boundary(sector_t sector, struct dm_target *ti)
1384 sector_t target_offset = dm_target_offset(ti, sector);
1386 return ti->len - target_offset;
1389 static sector_t max_io_len(sector_t sector, struct dm_target *ti)
1391 sector_t len = max_io_len_target_boundary(sector, ti);
1392 sector_t offset, max_len;
1395 * Does the target need to split even further?
1397 if (ti->max_io_len) {
1398 offset = dm_target_offset(ti, sector);
1399 if (unlikely(ti->max_io_len & (ti->max_io_len - 1)))
1400 max_len = sector_div(offset, ti->max_io_len);
1402 max_len = offset & (ti->max_io_len - 1);
1403 max_len = ti->max_io_len - max_len;
1412 int dm_set_target_max_io_len(struct dm_target *ti, sector_t len)
1414 if (len > UINT_MAX) {
1415 DMERR("Specified maximum size of target IO (%llu) exceeds limit (%u)",
1416 (unsigned long long)len, UINT_MAX);
1417 ti->error = "Maximum size of target IO is too large";
1421 ti->max_io_len = (uint32_t) len;
1425 EXPORT_SYMBOL_GPL(dm_set_target_max_io_len);
1428 * A target may call dm_accept_partial_bio only from the map routine. It is
1429 * allowed for all bio types except REQ_FLUSH.
1431 * dm_accept_partial_bio informs the dm that the target only wants to process
1432 * additional n_sectors sectors of the bio and the rest of the data should be
1433 * sent in a next bio.
1435 * A diagram that explains the arithmetics:
1436 * +--------------------+---------------+-------+
1438 * +--------------------+---------------+-------+
1440 * <-------------- *tio->len_ptr --------------->
1441 * <------- bi_size ------->
1444 * Region 1 was already iterated over with bio_advance or similar function.
1445 * (it may be empty if the target doesn't use bio_advance)
1446 * Region 2 is the remaining bio size that the target wants to process.
1447 * (it may be empty if region 1 is non-empty, although there is no reason
1449 * The target requires that region 3 is to be sent in the next bio.
1451 * If the target wants to receive multiple copies of the bio (via num_*bios, etc),
1452 * the partially processed part (the sum of regions 1+2) must be the same for all
1453 * copies of the bio.
1455 void dm_accept_partial_bio(struct bio *bio, unsigned n_sectors)
1457 struct dm_target_io *tio = container_of(bio, struct dm_target_io, clone);
1458 unsigned bi_size = bio->bi_iter.bi_size >> SECTOR_SHIFT;
1459 BUG_ON(bio->bi_rw & REQ_FLUSH);
1460 BUG_ON(bi_size > *tio->len_ptr);
1461 BUG_ON(n_sectors > bi_size);
1462 *tio->len_ptr -= bi_size - n_sectors;
1463 bio->bi_iter.bi_size = n_sectors << SECTOR_SHIFT;
1465 EXPORT_SYMBOL_GPL(dm_accept_partial_bio);
1467 static void __map_bio(struct dm_target_io *tio)
1471 struct mapped_device *md;
1472 struct bio *clone = &tio->clone;
1473 struct dm_target *ti = tio->ti;
1475 clone->bi_end_io = clone_endio;
1478 * Map the clone. If r == 0 we don't need to do
1479 * anything, the target has assumed ownership of
1482 atomic_inc(&tio->io->io_count);
1483 sector = clone->bi_iter.bi_sector;
1484 r = ti->type->map(ti, clone);
1485 if (r == DM_MAPIO_REMAPPED) {
1486 /* the bio has been remapped so dispatch it */
1488 trace_block_bio_remap(bdev_get_queue(clone->bi_bdev), clone,
1489 tio->io->bio->bi_bdev->bd_dev, sector);
1491 generic_make_request(clone);
1492 } else if (r < 0 || r == DM_MAPIO_REQUEUE) {
1493 /* error the io and bail out, or requeue it if needed */
1495 dec_pending(tio->io, r);
1497 } else if (r != DM_MAPIO_SUBMITTED) {
1498 DMWARN("unimplemented target map return value: %d", r);
1504 struct mapped_device *md;
1505 struct dm_table *map;
1509 unsigned sector_count;
1512 static void bio_setup_sector(struct bio *bio, sector_t sector, unsigned len)
1514 bio->bi_iter.bi_sector = sector;
1515 bio->bi_iter.bi_size = to_bytes(len);
1519 * Creates a bio that consists of range of complete bvecs.
1521 static void clone_bio(struct dm_target_io *tio, struct bio *bio,
1522 sector_t sector, unsigned len)
1524 struct bio *clone = &tio->clone;
1526 __bio_clone_fast(clone, bio);
1528 if (bio_integrity(bio))
1529 bio_integrity_clone(clone, bio, GFP_NOIO);
1531 bio_advance(clone, to_bytes(sector - clone->bi_iter.bi_sector));
1532 clone->bi_iter.bi_size = to_bytes(len);
1534 if (bio_integrity(bio))
1535 bio_integrity_trim(clone, 0, len);
1538 static struct dm_target_io *alloc_tio(struct clone_info *ci,
1539 struct dm_target *ti,
1540 unsigned target_bio_nr)
1542 struct dm_target_io *tio;
1545 clone = bio_alloc_bioset(GFP_NOIO, 0, ci->md->bs);
1546 tio = container_of(clone, struct dm_target_io, clone);
1550 tio->target_bio_nr = target_bio_nr;
1555 static void __clone_and_map_simple_bio(struct clone_info *ci,
1556 struct dm_target *ti,
1557 unsigned target_bio_nr, unsigned *len)
1559 struct dm_target_io *tio = alloc_tio(ci, ti, target_bio_nr);
1560 struct bio *clone = &tio->clone;
1564 __bio_clone_fast(clone, ci->bio);
1566 bio_setup_sector(clone, ci->sector, *len);
1571 static void __send_duplicate_bios(struct clone_info *ci, struct dm_target *ti,
1572 unsigned num_bios, unsigned *len)
1574 unsigned target_bio_nr;
1576 for (target_bio_nr = 0; target_bio_nr < num_bios; target_bio_nr++)
1577 __clone_and_map_simple_bio(ci, ti, target_bio_nr, len);
1580 static int __send_empty_flush(struct clone_info *ci)
1582 unsigned target_nr = 0;
1583 struct dm_target *ti;
1585 BUG_ON(bio_has_data(ci->bio));
1586 while ((ti = dm_table_get_target(ci->map, target_nr++)))
1587 __send_duplicate_bios(ci, ti, ti->num_flush_bios, NULL);
1592 static void __clone_and_map_data_bio(struct clone_info *ci, struct dm_target *ti,
1593 sector_t sector, unsigned *len)
1595 struct bio *bio = ci->bio;
1596 struct dm_target_io *tio;
1597 unsigned target_bio_nr;
1598 unsigned num_target_bios = 1;
1601 * Does the target want to receive duplicate copies of the bio?
1603 if (bio_data_dir(bio) == WRITE && ti->num_write_bios)
1604 num_target_bios = ti->num_write_bios(ti, bio);
1606 for (target_bio_nr = 0; target_bio_nr < num_target_bios; target_bio_nr++) {
1607 tio = alloc_tio(ci, ti, target_bio_nr);
1609 clone_bio(tio, bio, sector, *len);
1614 typedef unsigned (*get_num_bios_fn)(struct dm_target *ti);
1616 static unsigned get_num_discard_bios(struct dm_target *ti)
1618 return ti->num_discard_bios;
1621 static unsigned get_num_write_same_bios(struct dm_target *ti)
1623 return ti->num_write_same_bios;
1626 typedef bool (*is_split_required_fn)(struct dm_target *ti);
1628 static bool is_split_required_for_discard(struct dm_target *ti)
1630 return ti->split_discard_bios;
1633 static int __send_changing_extent_only(struct clone_info *ci,
1634 get_num_bios_fn get_num_bios,
1635 is_split_required_fn is_split_required)
1637 struct dm_target *ti;
1642 ti = dm_table_find_target(ci->map, ci->sector);
1643 if (!dm_target_is_valid(ti))
1647 * Even though the device advertised support for this type of
1648 * request, that does not mean every target supports it, and
1649 * reconfiguration might also have changed that since the
1650 * check was performed.
1652 num_bios = get_num_bios ? get_num_bios(ti) : 0;
1656 if (is_split_required && !is_split_required(ti))
1657 len = min((sector_t)ci->sector_count, max_io_len_target_boundary(ci->sector, ti));
1659 len = min((sector_t)ci->sector_count, max_io_len(ci->sector, ti));
1661 __send_duplicate_bios(ci, ti, num_bios, &len);
1664 } while (ci->sector_count -= len);
1669 static int __send_discard(struct clone_info *ci)
1671 return __send_changing_extent_only(ci, get_num_discard_bios,
1672 is_split_required_for_discard);
1675 static int __send_write_same(struct clone_info *ci)
1677 return __send_changing_extent_only(ci, get_num_write_same_bios, NULL);
1681 * Select the correct strategy for processing a non-flush bio.
1683 static int __split_and_process_non_flush(struct clone_info *ci)
1685 struct bio *bio = ci->bio;
1686 struct dm_target *ti;
1689 if (unlikely(bio->bi_rw & REQ_DISCARD))
1690 return __send_discard(ci);
1691 else if (unlikely(bio->bi_rw & REQ_WRITE_SAME))
1692 return __send_write_same(ci);
1694 ti = dm_table_find_target(ci->map, ci->sector);
1695 if (!dm_target_is_valid(ti))
1698 len = min_t(sector_t, max_io_len(ci->sector, ti), ci->sector_count);
1700 __clone_and_map_data_bio(ci, ti, ci->sector, &len);
1703 ci->sector_count -= len;
1709 * Entry point to split a bio into clones and submit them to the targets.
1711 static void __split_and_process_bio(struct mapped_device *md,
1712 struct dm_table *map, struct bio *bio)
1714 struct clone_info ci;
1717 if (unlikely(!map)) {
1724 ci.io = alloc_io(md);
1726 atomic_set(&ci.io->io_count, 1);
1729 spin_lock_init(&ci.io->endio_lock);
1730 ci.sector = bio->bi_iter.bi_sector;
1732 start_io_acct(ci.io);
1734 if (bio->bi_rw & REQ_FLUSH) {
1735 ci.bio = &ci.md->flush_bio;
1736 ci.sector_count = 0;
1737 error = __send_empty_flush(&ci);
1738 /* dec_pending submits any data associated with flush */
1741 ci.sector_count = bio_sectors(bio);
1742 while (ci.sector_count && !error)
1743 error = __split_and_process_non_flush(&ci);
1746 /* drop the extra reference count */
1747 dec_pending(ci.io, error);
1749 /*-----------------------------------------------------------------
1751 *---------------------------------------------------------------*/
1754 * The request function that just remaps the bio built up by
1757 static void dm_make_request(struct request_queue *q, struct bio *bio)
1759 int rw = bio_data_dir(bio);
1760 struct mapped_device *md = q->queuedata;
1762 struct dm_table *map;
1764 map = dm_get_live_table(md, &srcu_idx);
1766 generic_start_io_acct(rw, bio_sectors(bio), &dm_disk(md)->part0);
1768 /* if we're suspended, we have to queue this io for later */
1769 if (unlikely(test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags))) {
1770 dm_put_live_table(md, srcu_idx);
1772 if (bio_rw(bio) != READA)
1779 __split_and_process_bio(md, map, bio);
1780 dm_put_live_table(md, srcu_idx);
1784 int dm_request_based(struct mapped_device *md)
1786 return blk_queue_stackable(md->queue);
1789 static void dm_dispatch_clone_request(struct request *clone, struct request *rq)
1793 if (blk_queue_io_stat(clone->q))
1794 clone->cmd_flags |= REQ_IO_STAT;
1796 clone->start_time = jiffies;
1797 r = blk_insert_cloned_request(clone->q, clone);
1799 /* must complete clone in terms of original request */
1800 dm_complete_request(rq, r);
1803 static int dm_rq_bio_constructor(struct bio *bio, struct bio *bio_orig,
1806 struct dm_rq_target_io *tio = data;
1807 struct dm_rq_clone_bio_info *info =
1808 container_of(bio, struct dm_rq_clone_bio_info, clone);
1810 info->orig = bio_orig;
1812 bio->bi_end_io = end_clone_bio;
1817 static int setup_clone(struct request *clone, struct request *rq,
1818 struct dm_rq_target_io *tio, gfp_t gfp_mask)
1822 r = blk_rq_prep_clone(clone, rq, tio->md->bs, gfp_mask,
1823 dm_rq_bio_constructor, tio);
1827 clone->cmd = rq->cmd;
1828 clone->cmd_len = rq->cmd_len;
1829 clone->sense = rq->sense;
1830 clone->end_io = end_clone_request;
1831 clone->end_io_data = tio;
1838 static struct request *clone_rq(struct request *rq, struct mapped_device *md,
1839 struct dm_rq_target_io *tio, gfp_t gfp_mask)
1842 * Do not allocate a clone if tio->clone was already set
1843 * (see: dm_mq_queue_rq).
1845 bool alloc_clone = !tio->clone;
1846 struct request *clone;
1849 clone = alloc_clone_request(md, gfp_mask);
1855 blk_rq_init(NULL, clone);
1856 if (setup_clone(clone, rq, tio, gfp_mask)) {
1859 free_clone_request(md, clone);
1866 static void map_tio_request(struct kthread_work *work);
1868 static void init_tio(struct dm_rq_target_io *tio, struct request *rq,
1869 struct mapped_device *md)
1876 memset(&tio->info, 0, sizeof(tio->info));
1877 if (md->kworker_task)
1878 init_kthread_work(&tio->work, map_tio_request);
1881 static struct dm_rq_target_io *prep_tio(struct request *rq,
1882 struct mapped_device *md, gfp_t gfp_mask)
1884 struct dm_rq_target_io *tio;
1886 struct dm_table *table;
1888 tio = alloc_rq_tio(md, gfp_mask);
1892 init_tio(tio, rq, md);
1894 table = dm_get_live_table(md, &srcu_idx);
1895 if (!dm_table_mq_request_based(table)) {
1896 if (!clone_rq(rq, md, tio, gfp_mask)) {
1897 dm_put_live_table(md, srcu_idx);
1902 dm_put_live_table(md, srcu_idx);
1908 * Called with the queue lock held.
1910 static int dm_prep_fn(struct request_queue *q, struct request *rq)
1912 struct mapped_device *md = q->queuedata;
1913 struct dm_rq_target_io *tio;
1915 if (unlikely(rq->special)) {
1916 DMWARN("Already has something in rq->special.");
1917 return BLKPREP_KILL;
1920 tio = prep_tio(rq, md, GFP_ATOMIC);
1922 return BLKPREP_DEFER;
1925 rq->cmd_flags |= REQ_DONTPREP;
1932 * 0 : the request has been processed
1933 * DM_MAPIO_REQUEUE : the original request needs to be requeued
1934 * < 0 : the request was completed due to failure
1936 static int map_request(struct dm_rq_target_io *tio, struct request *rq,
1937 struct mapped_device *md)
1940 struct dm_target *ti = tio->ti;
1941 struct request *clone = NULL;
1945 r = ti->type->map_rq(ti, clone, &tio->info);
1947 r = ti->type->clone_and_map_rq(ti, rq, &tio->info, &clone);
1949 /* The target wants to complete the I/O */
1950 dm_kill_unmapped_request(rq, r);
1953 if (r != DM_MAPIO_REMAPPED)
1955 if (setup_clone(clone, rq, tio, GFP_ATOMIC)) {
1957 ti->type->release_clone_rq(clone);
1958 return DM_MAPIO_REQUEUE;
1963 case DM_MAPIO_SUBMITTED:
1964 /* The target has taken the I/O to submit by itself later */
1966 case DM_MAPIO_REMAPPED:
1967 /* The target has remapped the I/O so dispatch it */
1968 trace_block_rq_remap(clone->q, clone, disk_devt(dm_disk(md)),
1970 dm_dispatch_clone_request(clone, rq);
1972 case DM_MAPIO_REQUEUE:
1973 /* The target wants to requeue the I/O */
1974 dm_requeue_original_request(md, tio->orig);
1978 DMWARN("unimplemented target map return value: %d", r);
1982 /* The target wants to complete the I/O */
1983 dm_kill_unmapped_request(rq, r);
1990 static void map_tio_request(struct kthread_work *work)
1992 struct dm_rq_target_io *tio = container_of(work, struct dm_rq_target_io, work);
1993 struct request *rq = tio->orig;
1994 struct mapped_device *md = tio->md;
1996 if (map_request(tio, rq, md) == DM_MAPIO_REQUEUE)
1997 dm_requeue_original_request(md, rq);
2000 static void dm_start_request(struct mapped_device *md, struct request *orig)
2002 if (!orig->q->mq_ops)
2003 blk_start_request(orig);
2005 blk_mq_start_request(orig);
2006 atomic_inc(&md->pending[rq_data_dir(orig)]);
2008 if (md->seq_rq_merge_deadline_usecs) {
2009 md->last_rq_pos = rq_end_sector(orig);
2010 md->last_rq_rw = rq_data_dir(orig);
2011 md->last_rq_start_time = ktime_get();
2014 if (unlikely(dm_stats_used(&md->stats))) {
2015 struct dm_rq_target_io *tio = tio_from_request(orig);
2016 tio->duration_jiffies = jiffies;
2017 tio->n_sectors = blk_rq_sectors(orig);
2018 dm_stats_account_io(&md->stats, orig->cmd_flags, blk_rq_pos(orig),
2019 tio->n_sectors, false, 0, &tio->stats_aux);
2023 * Hold the md reference here for the in-flight I/O.
2024 * We can't rely on the reference count by device opener,
2025 * because the device may be closed during the request completion
2026 * when all bios are completed.
2027 * See the comment in rq_completed() too.
2032 #define MAX_SEQ_RQ_MERGE_DEADLINE_USECS 100000
2034 ssize_t dm_attr_rq_based_seq_io_merge_deadline_show(struct mapped_device *md, char *buf)
2036 return sprintf(buf, "%u\n", md->seq_rq_merge_deadline_usecs);
2039 ssize_t dm_attr_rq_based_seq_io_merge_deadline_store(struct mapped_device *md,
2040 const char *buf, size_t count)
2044 if (!dm_request_based(md) || md->use_blk_mq)
2047 if (kstrtouint(buf, 10, &deadline))
2050 if (deadline > MAX_SEQ_RQ_MERGE_DEADLINE_USECS)
2051 deadline = MAX_SEQ_RQ_MERGE_DEADLINE_USECS;
2053 md->seq_rq_merge_deadline_usecs = deadline;
2058 static bool dm_request_peeked_before_merge_deadline(struct mapped_device *md)
2060 ktime_t kt_deadline;
2062 if (!md->seq_rq_merge_deadline_usecs)
2065 kt_deadline = ns_to_ktime((u64)md->seq_rq_merge_deadline_usecs * NSEC_PER_USEC);
2066 kt_deadline = ktime_add_safe(md->last_rq_start_time, kt_deadline);
2068 return !ktime_after(ktime_get(), kt_deadline);
2072 * q->request_fn for request-based dm.
2073 * Called with the queue lock held.
2075 static void dm_request_fn(struct request_queue *q)
2077 struct mapped_device *md = q->queuedata;
2079 struct dm_table *map = dm_get_live_table(md, &srcu_idx);
2080 struct dm_target *ti;
2082 struct dm_rq_target_io *tio;
2086 * For suspend, check blk_queue_stopped() and increment
2087 * ->pending within a single queue_lock not to increment the
2088 * number of in-flight I/Os after the queue is stopped in
2091 while (!blk_queue_stopped(q)) {
2092 rq = blk_peek_request(q);
2096 /* always use block 0 to find the target for flushes for now */
2098 if (!(rq->cmd_flags & REQ_FLUSH))
2099 pos = blk_rq_pos(rq);
2101 ti = dm_table_find_target(map, pos);
2102 if (!dm_target_is_valid(ti)) {
2104 * Must perform setup, that rq_completed() requires,
2105 * before calling dm_kill_unmapped_request
2107 DMERR_LIMIT("request attempted access beyond the end of device");
2108 dm_start_request(md, rq);
2109 dm_kill_unmapped_request(rq, -EIO);
2113 if (dm_request_peeked_before_merge_deadline(md) &&
2114 md_in_flight(md) && rq->bio && rq->bio->bi_vcnt == 1 &&
2115 md->last_rq_pos == pos && md->last_rq_rw == rq_data_dir(rq))
2118 if (ti->type->busy && ti->type->busy(ti))
2121 dm_start_request(md, rq);
2123 tio = tio_from_request(rq);
2124 /* Establish tio->ti before queuing work (map_tio_request) */
2126 queue_kthread_work(&md->kworker, &tio->work);
2127 BUG_ON(!irqs_disabled());
2133 blk_delay_queue(q, HZ / 100);
2135 dm_put_live_table(md, srcu_idx);
2138 static int dm_any_congested(void *congested_data, int bdi_bits)
2141 struct mapped_device *md = congested_data;
2142 struct dm_table *map;
2144 if (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
2145 map = dm_get_live_table_fast(md);
2148 * Request-based dm cares about only own queue for
2149 * the query about congestion status of request_queue
2151 if (dm_request_based(md))
2152 r = md->queue->backing_dev_info.wb.state &
2155 r = dm_table_any_congested(map, bdi_bits);
2157 dm_put_live_table_fast(md);
2163 /*-----------------------------------------------------------------
2164 * An IDR is used to keep track of allocated minor numbers.
2165 *---------------------------------------------------------------*/
2166 static void free_minor(int minor)
2168 spin_lock(&_minor_lock);
2169 idr_remove(&_minor_idr, minor);
2170 spin_unlock(&_minor_lock);
2174 * See if the device with a specific minor # is free.
2176 static int specific_minor(int minor)
2180 if (minor >= (1 << MINORBITS))
2183 idr_preload(GFP_KERNEL);
2184 spin_lock(&_minor_lock);
2186 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, minor, minor + 1, GFP_NOWAIT);
2188 spin_unlock(&_minor_lock);
2191 return r == -ENOSPC ? -EBUSY : r;
2195 static int next_free_minor(int *minor)
2199 idr_preload(GFP_KERNEL);
2200 spin_lock(&_minor_lock);
2202 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, 0, 1 << MINORBITS, GFP_NOWAIT);
2204 spin_unlock(&_minor_lock);
2212 static const struct block_device_operations dm_blk_dops;
2214 static void dm_wq_work(struct work_struct *work);
2216 static void dm_init_md_queue(struct mapped_device *md)
2219 * Request-based dm devices cannot be stacked on top of bio-based dm
2220 * devices. The type of this dm device may not have been decided yet.
2221 * The type is decided at the first table loading time.
2222 * To prevent problematic device stacking, clear the queue flag
2223 * for request stacking support until then.
2225 * This queue is new, so no concurrency on the queue_flags.
2227 queue_flag_clear_unlocked(QUEUE_FLAG_STACKABLE, md->queue);
2230 * Initialize data that will only be used by a non-blk-mq DM queue
2231 * - must do so here (in alloc_dev callchain) before queue is used
2233 md->queue->queuedata = md;
2234 md->queue->backing_dev_info.congested_data = md;
2237 static void dm_init_old_md_queue(struct mapped_device *md)
2239 md->use_blk_mq = false;
2240 dm_init_md_queue(md);
2243 * Initialize aspects of queue that aren't relevant for blk-mq
2245 md->queue->backing_dev_info.congested_fn = dm_any_congested;
2246 blk_queue_bounce_limit(md->queue, BLK_BOUNCE_ANY);
2249 static void cleanup_mapped_device(struct mapped_device *md)
2252 destroy_workqueue(md->wq);
2253 if (md->kworker_task)
2254 kthread_stop(md->kworker_task);
2255 mempool_destroy(md->io_pool);
2256 mempool_destroy(md->rq_pool);
2258 bioset_free(md->bs);
2260 cleanup_srcu_struct(&md->io_barrier);
2263 spin_lock(&_minor_lock);
2264 md->disk->private_data = NULL;
2265 spin_unlock(&_minor_lock);
2266 if (blk_get_integrity(md->disk))
2267 blk_integrity_unregister(md->disk);
2268 del_gendisk(md->disk);
2273 blk_cleanup_queue(md->queue);
2282 * Allocate and initialise a blank device with a given minor.
2284 static struct mapped_device *alloc_dev(int minor)
2287 struct mapped_device *md = kzalloc(sizeof(*md), GFP_KERNEL);
2291 DMWARN("unable to allocate device, out of memory.");
2295 if (!try_module_get(THIS_MODULE))
2296 goto bad_module_get;
2298 /* get a minor number for the dev */
2299 if (minor == DM_ANY_MINOR)
2300 r = next_free_minor(&minor);
2302 r = specific_minor(minor);
2306 r = init_srcu_struct(&md->io_barrier);
2308 goto bad_io_barrier;
2310 md->use_blk_mq = use_blk_mq;
2311 md->type = DM_TYPE_NONE;
2312 mutex_init(&md->suspend_lock);
2313 mutex_init(&md->type_lock);
2314 mutex_init(&md->table_devices_lock);
2315 spin_lock_init(&md->deferred_lock);
2316 atomic_set(&md->holders, 1);
2317 atomic_set(&md->open_count, 0);
2318 atomic_set(&md->event_nr, 0);
2319 atomic_set(&md->uevent_seq, 0);
2320 INIT_LIST_HEAD(&md->uevent_list);
2321 INIT_LIST_HEAD(&md->table_devices);
2322 spin_lock_init(&md->uevent_lock);
2324 md->queue = blk_alloc_queue(GFP_KERNEL);
2328 dm_init_md_queue(md);
2330 md->disk = alloc_disk(1);
2334 atomic_set(&md->pending[0], 0);
2335 atomic_set(&md->pending[1], 0);
2336 init_waitqueue_head(&md->wait);
2337 INIT_WORK(&md->work, dm_wq_work);
2338 init_waitqueue_head(&md->eventq);
2339 init_completion(&md->kobj_holder.completion);
2340 md->kworker_task = NULL;
2342 md->disk->major = _major;
2343 md->disk->first_minor = minor;
2344 md->disk->fops = &dm_blk_dops;
2345 md->disk->queue = md->queue;
2346 md->disk->private_data = md;
2347 sprintf(md->disk->disk_name, "dm-%d", minor);
2349 format_dev_t(md->name, MKDEV(_major, minor));
2351 md->wq = alloc_workqueue("kdmflush", WQ_MEM_RECLAIM, 0);
2355 md->bdev = bdget_disk(md->disk, 0);
2359 bio_init(&md->flush_bio);
2360 md->flush_bio.bi_bdev = md->bdev;
2361 md->flush_bio.bi_rw = WRITE_FLUSH;
2363 dm_stats_init(&md->stats);
2365 /* Populate the mapping, nobody knows we exist yet */
2366 spin_lock(&_minor_lock);
2367 old_md = idr_replace(&_minor_idr, md, minor);
2368 spin_unlock(&_minor_lock);
2370 BUG_ON(old_md != MINOR_ALLOCED);
2375 cleanup_mapped_device(md);
2379 module_put(THIS_MODULE);
2385 static void unlock_fs(struct mapped_device *md);
2387 static void free_dev(struct mapped_device *md)
2389 int minor = MINOR(disk_devt(md->disk));
2393 cleanup_mapped_device(md);
2395 blk_mq_free_tag_set(&md->tag_set);
2397 free_table_devices(&md->table_devices);
2398 dm_stats_cleanup(&md->stats);
2401 module_put(THIS_MODULE);
2405 static void __bind_mempools(struct mapped_device *md, struct dm_table *t)
2407 struct dm_md_mempools *p = dm_table_get_md_mempools(t);
2410 /* The md already has necessary mempools. */
2411 if (dm_table_get_type(t) == DM_TYPE_BIO_BASED) {
2413 * Reload bioset because front_pad may have changed
2414 * because a different table was loaded.
2416 bioset_free(md->bs);
2421 * There's no need to reload with request-based dm
2422 * because the size of front_pad doesn't change.
2423 * Note for future: If you are to reload bioset,
2424 * prep-ed requests in the queue may refer
2425 * to bio from the old bioset, so you must walk
2426 * through the queue to unprep.
2431 BUG_ON(!p || md->io_pool || md->rq_pool || md->bs);
2433 md->io_pool = p->io_pool;
2435 md->rq_pool = p->rq_pool;
2441 /* mempool bind completed, no longer need any mempools in the table */
2442 dm_table_free_md_mempools(t);
2446 * Bind a table to the device.
2448 static void event_callback(void *context)
2450 unsigned long flags;
2452 struct mapped_device *md = (struct mapped_device *) context;
2454 spin_lock_irqsave(&md->uevent_lock, flags);
2455 list_splice_init(&md->uevent_list, &uevents);
2456 spin_unlock_irqrestore(&md->uevent_lock, flags);
2458 dm_send_uevents(&uevents, &disk_to_dev(md->disk)->kobj);
2460 atomic_inc(&md->event_nr);
2461 wake_up(&md->eventq);
2465 * Protected by md->suspend_lock obtained by dm_swap_table().
2467 static void __set_size(struct mapped_device *md, sector_t size)
2469 set_capacity(md->disk, size);
2471 i_size_write(md->bdev->bd_inode, (loff_t)size << SECTOR_SHIFT);
2475 * Returns old map, which caller must destroy.
2477 static struct dm_table *__bind(struct mapped_device *md, struct dm_table *t,
2478 struct queue_limits *limits)
2480 struct dm_table *old_map;
2481 struct request_queue *q = md->queue;
2484 size = dm_table_get_size(t);
2487 * Wipe any geometry if the size of the table changed.
2489 if (size != dm_get_size(md))
2490 memset(&md->geometry, 0, sizeof(md->geometry));
2492 __set_size(md, size);
2494 dm_table_event_callback(t, event_callback, md);
2497 * The queue hasn't been stopped yet, if the old table type wasn't
2498 * for request-based during suspension. So stop it to prevent
2499 * I/O mapping before resume.
2500 * This must be done before setting the queue restrictions,
2501 * because request-based dm may be run just after the setting.
2503 if (dm_table_request_based(t))
2506 __bind_mempools(md, t);
2508 old_map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2509 rcu_assign_pointer(md->map, t);
2510 md->immutable_target_type = dm_table_get_immutable_target_type(t);
2512 dm_table_set_restrictions(t, q, limits);
2520 * Returns unbound table for the caller to free.
2522 static struct dm_table *__unbind(struct mapped_device *md)
2524 struct dm_table *map = rcu_dereference_protected(md->map, 1);
2529 dm_table_event_callback(map, NULL, NULL);
2530 RCU_INIT_POINTER(md->map, NULL);
2537 * Constructor for a new device.
2539 int dm_create(int minor, struct mapped_device **result)
2541 struct mapped_device *md;
2543 md = alloc_dev(minor);
2554 * Functions to manage md->type.
2555 * All are required to hold md->type_lock.
2557 void dm_lock_md_type(struct mapped_device *md)
2559 mutex_lock(&md->type_lock);
2562 void dm_unlock_md_type(struct mapped_device *md)
2564 mutex_unlock(&md->type_lock);
2567 void dm_set_md_type(struct mapped_device *md, unsigned type)
2569 BUG_ON(!mutex_is_locked(&md->type_lock));
2573 unsigned dm_get_md_type(struct mapped_device *md)
2575 BUG_ON(!mutex_is_locked(&md->type_lock));
2579 struct target_type *dm_get_immutable_target_type(struct mapped_device *md)
2581 return md->immutable_target_type;
2585 * The queue_limits are only valid as long as you have a reference
2588 struct queue_limits *dm_get_queue_limits(struct mapped_device *md)
2590 BUG_ON(!atomic_read(&md->holders));
2591 return &md->queue->limits;
2593 EXPORT_SYMBOL_GPL(dm_get_queue_limits);
2595 static void init_rq_based_worker_thread(struct mapped_device *md)
2597 /* Initialize the request-based DM worker thread */
2598 init_kthread_worker(&md->kworker);
2599 md->kworker_task = kthread_run(kthread_worker_fn, &md->kworker,
2600 "kdmwork-%s", dm_device_name(md));
2604 * Fully initialize a request-based queue (->elevator, ->request_fn, etc).
2606 static int dm_init_request_based_queue(struct mapped_device *md)
2608 struct request_queue *q = NULL;
2610 /* Fully initialize the queue */
2611 q = blk_init_allocated_queue(md->queue, dm_request_fn, NULL);
2615 /* disable dm_request_fn's merge heuristic by default */
2616 md->seq_rq_merge_deadline_usecs = 0;
2619 dm_init_old_md_queue(md);
2620 blk_queue_softirq_done(md->queue, dm_softirq_done);
2621 blk_queue_prep_rq(md->queue, dm_prep_fn);
2623 init_rq_based_worker_thread(md);
2625 elv_register_queue(md->queue);
2630 static int dm_mq_init_request(void *data, struct request *rq,
2631 unsigned int hctx_idx, unsigned int request_idx,
2632 unsigned int numa_node)
2634 struct mapped_device *md = data;
2635 struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
2638 * Must initialize md member of tio, otherwise it won't
2639 * be available in dm_mq_queue_rq.
2646 static int dm_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
2647 const struct blk_mq_queue_data *bd)
2649 struct request *rq = bd->rq;
2650 struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
2651 struct mapped_device *md = tio->md;
2653 struct dm_table *map = dm_get_live_table(md, &srcu_idx);
2654 struct dm_target *ti;
2657 /* always use block 0 to find the target for flushes for now */
2659 if (!(rq->cmd_flags & REQ_FLUSH))
2660 pos = blk_rq_pos(rq);
2662 ti = dm_table_find_target(map, pos);
2663 if (!dm_target_is_valid(ti)) {
2664 dm_put_live_table(md, srcu_idx);
2665 DMERR_LIMIT("request attempted access beyond the end of device");
2667 * Must perform setup, that rq_completed() requires,
2668 * before returning BLK_MQ_RQ_QUEUE_ERROR
2670 dm_start_request(md, rq);
2671 return BLK_MQ_RQ_QUEUE_ERROR;
2673 dm_put_live_table(md, srcu_idx);
2675 if (ti->type->busy && ti->type->busy(ti))
2676 return BLK_MQ_RQ_QUEUE_BUSY;
2678 dm_start_request(md, rq);
2680 /* Init tio using md established in .init_request */
2681 init_tio(tio, rq, md);
2684 * Establish tio->ti before queuing work (map_tio_request)
2685 * or making direct call to map_request().
2689 /* Clone the request if underlying devices aren't blk-mq */
2690 if (dm_table_get_type(map) == DM_TYPE_REQUEST_BASED) {
2691 /* clone request is allocated at the end of the pdu */
2692 tio->clone = (void *)blk_mq_rq_to_pdu(rq) + sizeof(struct dm_rq_target_io);
2693 (void) clone_rq(rq, md, tio, GFP_ATOMIC);
2694 queue_kthread_work(&md->kworker, &tio->work);
2696 /* Direct call is fine since .queue_rq allows allocations */
2697 if (map_request(tio, rq, md) == DM_MAPIO_REQUEUE) {
2698 /* Undo dm_start_request() before requeuing */
2699 rq_end_stats(md, rq);
2700 rq_completed(md, rq_data_dir(rq), false);
2701 return BLK_MQ_RQ_QUEUE_BUSY;
2705 return BLK_MQ_RQ_QUEUE_OK;
2708 static struct blk_mq_ops dm_mq_ops = {
2709 .queue_rq = dm_mq_queue_rq,
2710 .map_queue = blk_mq_map_queue,
2711 .complete = dm_softirq_done,
2712 .init_request = dm_mq_init_request,
2715 static int dm_init_request_based_blk_mq_queue(struct mapped_device *md)
2717 unsigned md_type = dm_get_md_type(md);
2718 struct request_queue *q;
2721 memset(&md->tag_set, 0, sizeof(md->tag_set));
2722 md->tag_set.ops = &dm_mq_ops;
2723 md->tag_set.queue_depth = BLKDEV_MAX_RQ;
2724 md->tag_set.numa_node = NUMA_NO_NODE;
2725 md->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
2726 md->tag_set.nr_hw_queues = 1;
2727 if (md_type == DM_TYPE_REQUEST_BASED) {
2728 /* make the memory for non-blk-mq clone part of the pdu */
2729 md->tag_set.cmd_size = sizeof(struct dm_rq_target_io) + sizeof(struct request);
2731 md->tag_set.cmd_size = sizeof(struct dm_rq_target_io);
2732 md->tag_set.driver_data = md;
2734 err = blk_mq_alloc_tag_set(&md->tag_set);
2738 q = blk_mq_init_allocated_queue(&md->tag_set, md->queue);
2744 dm_init_md_queue(md);
2746 /* backfill 'mq' sysfs registration normally done in blk_register_queue */
2747 blk_mq_register_disk(md->disk);
2749 if (md_type == DM_TYPE_REQUEST_BASED)
2750 init_rq_based_worker_thread(md);
2755 blk_mq_free_tag_set(&md->tag_set);
2759 static unsigned filter_md_type(unsigned type, struct mapped_device *md)
2761 if (type == DM_TYPE_BIO_BASED)
2764 return !md->use_blk_mq ? DM_TYPE_REQUEST_BASED : DM_TYPE_MQ_REQUEST_BASED;
2768 * Setup the DM device's queue based on md's type
2770 int dm_setup_md_queue(struct mapped_device *md)
2773 unsigned md_type = filter_md_type(dm_get_md_type(md), md);
2776 case DM_TYPE_REQUEST_BASED:
2777 r = dm_init_request_based_queue(md);
2779 DMWARN("Cannot initialize queue for request-based mapped device");
2783 case DM_TYPE_MQ_REQUEST_BASED:
2784 r = dm_init_request_based_blk_mq_queue(md);
2786 DMWARN("Cannot initialize queue for request-based blk-mq mapped device");
2790 case DM_TYPE_BIO_BASED:
2791 dm_init_old_md_queue(md);
2792 blk_queue_make_request(md->queue, dm_make_request);
2794 * DM handles splitting bios as needed. Free the bio_split bioset
2795 * since it won't be used (saves 1 process per bio-based DM device).
2797 bioset_free(md->queue->bio_split);
2798 md->queue->bio_split = NULL;
2805 struct mapped_device *dm_get_md(dev_t dev)
2807 struct mapped_device *md;
2808 unsigned minor = MINOR(dev);
2810 if (MAJOR(dev) != _major || minor >= (1 << MINORBITS))
2813 spin_lock(&_minor_lock);
2815 md = idr_find(&_minor_idr, minor);
2817 if ((md == MINOR_ALLOCED ||
2818 (MINOR(disk_devt(dm_disk(md))) != minor) ||
2819 dm_deleting_md(md) ||
2820 test_bit(DMF_FREEING, &md->flags))) {
2828 spin_unlock(&_minor_lock);
2832 EXPORT_SYMBOL_GPL(dm_get_md);
2834 void *dm_get_mdptr(struct mapped_device *md)
2836 return md->interface_ptr;
2839 void dm_set_mdptr(struct mapped_device *md, void *ptr)
2841 md->interface_ptr = ptr;
2844 void dm_get(struct mapped_device *md)
2846 atomic_inc(&md->holders);
2847 BUG_ON(test_bit(DMF_FREEING, &md->flags));
2850 int dm_hold(struct mapped_device *md)
2852 spin_lock(&_minor_lock);
2853 if (test_bit(DMF_FREEING, &md->flags)) {
2854 spin_unlock(&_minor_lock);
2858 spin_unlock(&_minor_lock);
2861 EXPORT_SYMBOL_GPL(dm_hold);
2863 const char *dm_device_name(struct mapped_device *md)
2867 EXPORT_SYMBOL_GPL(dm_device_name);
2869 static void __dm_destroy(struct mapped_device *md, bool wait)
2871 struct dm_table *map;
2876 map = dm_get_live_table(md, &srcu_idx);
2878 spin_lock(&_minor_lock);
2879 idr_replace(&_minor_idr, MINOR_ALLOCED, MINOR(disk_devt(dm_disk(md))));
2880 set_bit(DMF_FREEING, &md->flags);
2881 spin_unlock(&_minor_lock);
2883 if (dm_request_based(md) && md->kworker_task)
2884 flush_kthread_worker(&md->kworker);
2887 * Take suspend_lock so that presuspend and postsuspend methods
2888 * do not race with internal suspend.
2890 mutex_lock(&md->suspend_lock);
2891 if (!dm_suspended_md(md)) {
2892 dm_table_presuspend_targets(map);
2893 dm_table_postsuspend_targets(map);
2895 mutex_unlock(&md->suspend_lock);
2897 /* dm_put_live_table must be before msleep, otherwise deadlock is possible */
2898 dm_put_live_table(md, srcu_idx);
2901 * Rare, but there may be I/O requests still going to complete,
2902 * for example. Wait for all references to disappear.
2903 * No one should increment the reference count of the mapped_device,
2904 * after the mapped_device state becomes DMF_FREEING.
2907 while (atomic_read(&md->holders))
2909 else if (atomic_read(&md->holders))
2910 DMWARN("%s: Forcibly removing mapped_device still in use! (%d users)",
2911 dm_device_name(md), atomic_read(&md->holders));
2914 dm_table_destroy(__unbind(md));
2918 void dm_destroy(struct mapped_device *md)
2920 __dm_destroy(md, true);
2923 void dm_destroy_immediate(struct mapped_device *md)
2925 __dm_destroy(md, false);
2928 void dm_put(struct mapped_device *md)
2930 atomic_dec(&md->holders);
2932 EXPORT_SYMBOL_GPL(dm_put);
2934 static int dm_wait_for_completion(struct mapped_device *md, int interruptible)
2937 DECLARE_WAITQUEUE(wait, current);
2939 add_wait_queue(&md->wait, &wait);
2942 set_current_state(interruptible);
2944 if (!md_in_flight(md))
2947 if (interruptible == TASK_INTERRUPTIBLE &&
2948 signal_pending(current)) {
2955 set_current_state(TASK_RUNNING);
2957 remove_wait_queue(&md->wait, &wait);
2963 * Process the deferred bios
2965 static void dm_wq_work(struct work_struct *work)
2967 struct mapped_device *md = container_of(work, struct mapped_device,
2971 struct dm_table *map;
2973 map = dm_get_live_table(md, &srcu_idx);
2975 while (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
2976 spin_lock_irq(&md->deferred_lock);
2977 c = bio_list_pop(&md->deferred);
2978 spin_unlock_irq(&md->deferred_lock);
2983 if (dm_request_based(md))
2984 generic_make_request(c);
2986 __split_and_process_bio(md, map, c);
2989 dm_put_live_table(md, srcu_idx);
2992 static void dm_queue_flush(struct mapped_device *md)
2994 clear_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2995 smp_mb__after_atomic();
2996 queue_work(md->wq, &md->work);
3000 * Swap in a new table, returning the old one for the caller to destroy.
3002 struct dm_table *dm_swap_table(struct mapped_device *md, struct dm_table *table)
3004 struct dm_table *live_map = NULL, *map = ERR_PTR(-EINVAL);
3005 struct queue_limits limits;
3008 mutex_lock(&md->suspend_lock);
3010 /* device must be suspended */
3011 if (!dm_suspended_md(md))
3015 * If the new table has no data devices, retain the existing limits.
3016 * This helps multipath with queue_if_no_path if all paths disappear,
3017 * then new I/O is queued based on these limits, and then some paths
3020 if (dm_table_has_no_data_devices(table)) {
3021 live_map = dm_get_live_table_fast(md);
3023 limits = md->queue->limits;
3024 dm_put_live_table_fast(md);
3028 r = dm_calculate_queue_limits(table, &limits);
3035 map = __bind(md, table, &limits);
3038 mutex_unlock(&md->suspend_lock);
3043 * Functions to lock and unlock any filesystem running on the
3046 static int lock_fs(struct mapped_device *md)
3050 WARN_ON(md->frozen_sb);
3052 md->frozen_sb = freeze_bdev(md->bdev);
3053 if (IS_ERR(md->frozen_sb)) {
3054 r = PTR_ERR(md->frozen_sb);
3055 md->frozen_sb = NULL;
3059 set_bit(DMF_FROZEN, &md->flags);
3064 static void unlock_fs(struct mapped_device *md)
3066 if (!test_bit(DMF_FROZEN, &md->flags))
3069 thaw_bdev(md->bdev, md->frozen_sb);
3070 md->frozen_sb = NULL;
3071 clear_bit(DMF_FROZEN, &md->flags);
3075 * If __dm_suspend returns 0, the device is completely quiescent
3076 * now. There is no request-processing activity. All new requests
3077 * are being added to md->deferred list.
3079 * Caller must hold md->suspend_lock
3081 static int __dm_suspend(struct mapped_device *md, struct dm_table *map,
3082 unsigned suspend_flags, int interruptible)
3084 bool do_lockfs = suspend_flags & DM_SUSPEND_LOCKFS_FLAG;
3085 bool noflush = suspend_flags & DM_SUSPEND_NOFLUSH_FLAG;
3089 * DMF_NOFLUSH_SUSPENDING must be set before presuspend.
3090 * This flag is cleared before dm_suspend returns.
3093 set_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
3096 * This gets reverted if there's an error later and the targets
3097 * provide the .presuspend_undo hook.
3099 dm_table_presuspend_targets(map);
3102 * Flush I/O to the device.
3103 * Any I/O submitted after lock_fs() may not be flushed.
3104 * noflush takes precedence over do_lockfs.
3105 * (lock_fs() flushes I/Os and waits for them to complete.)
3107 if (!noflush && do_lockfs) {
3110 dm_table_presuspend_undo_targets(map);
3116 * Here we must make sure that no processes are submitting requests
3117 * to target drivers i.e. no one may be executing
3118 * __split_and_process_bio. This is called from dm_request and
3121 * To get all processes out of __split_and_process_bio in dm_request,
3122 * we take the write lock. To prevent any process from reentering
3123 * __split_and_process_bio from dm_request and quiesce the thread
3124 * (dm_wq_work), we set BMF_BLOCK_IO_FOR_SUSPEND and call
3125 * flush_workqueue(md->wq).
3127 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
3129 synchronize_srcu(&md->io_barrier);
3132 * Stop md->queue before flushing md->wq in case request-based
3133 * dm defers requests to md->wq from md->queue.
3135 if (dm_request_based(md)) {
3136 stop_queue(md->queue);
3137 if (md->kworker_task)
3138 flush_kthread_worker(&md->kworker);
3141 flush_workqueue(md->wq);
3144 * At this point no more requests are entering target request routines.
3145 * We call dm_wait_for_completion to wait for all existing requests
3148 r = dm_wait_for_completion(md, interruptible);
3151 clear_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
3153 synchronize_srcu(&md->io_barrier);
3155 /* were we interrupted ? */
3159 if (dm_request_based(md))
3160 start_queue(md->queue);
3163 dm_table_presuspend_undo_targets(map);
3164 /* pushback list is already flushed, so skip flush */
3171 * We need to be able to change a mapping table under a mounted
3172 * filesystem. For example we might want to move some data in
3173 * the background. Before the table can be swapped with
3174 * dm_bind_table, dm_suspend must be called to flush any in
3175 * flight bios and ensure that any further io gets deferred.
3178 * Suspend mechanism in request-based dm.
3180 * 1. Flush all I/Os by lock_fs() if needed.
3181 * 2. Stop dispatching any I/O by stopping the request_queue.
3182 * 3. Wait for all in-flight I/Os to be completed or requeued.
3184 * To abort suspend, start the request_queue.
3186 int dm_suspend(struct mapped_device *md, unsigned suspend_flags)
3188 struct dm_table *map = NULL;
3192 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
3194 if (dm_suspended_md(md)) {
3199 if (dm_suspended_internally_md(md)) {
3200 /* already internally suspended, wait for internal resume */
3201 mutex_unlock(&md->suspend_lock);
3202 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
3208 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3210 r = __dm_suspend(md, map, suspend_flags, TASK_INTERRUPTIBLE);
3214 set_bit(DMF_SUSPENDED, &md->flags);
3216 dm_table_postsuspend_targets(map);
3219 mutex_unlock(&md->suspend_lock);
3223 static int __dm_resume(struct mapped_device *md, struct dm_table *map)
3226 int r = dm_table_resume_targets(map);
3234 * Flushing deferred I/Os must be done after targets are resumed
3235 * so that mapping of targets can work correctly.
3236 * Request-based dm is queueing the deferred I/Os in its request_queue.
3238 if (dm_request_based(md))
3239 start_queue(md->queue);
3246 int dm_resume(struct mapped_device *md)
3249 struct dm_table *map = NULL;
3252 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
3254 if (!dm_suspended_md(md))
3257 if (dm_suspended_internally_md(md)) {
3258 /* already internally suspended, wait for internal resume */
3259 mutex_unlock(&md->suspend_lock);
3260 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
3266 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3267 if (!map || !dm_table_get_size(map))
3270 r = __dm_resume(md, map);
3274 clear_bit(DMF_SUSPENDED, &md->flags);
3278 mutex_unlock(&md->suspend_lock);
3284 * Internal suspend/resume works like userspace-driven suspend. It waits
3285 * until all bios finish and prevents issuing new bios to the target drivers.
3286 * It may be used only from the kernel.
3289 static void __dm_internal_suspend(struct mapped_device *md, unsigned suspend_flags)
3291 struct dm_table *map = NULL;
3293 if (md->internal_suspend_count++)
3294 return; /* nested internal suspend */
3296 if (dm_suspended_md(md)) {
3297 set_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3298 return; /* nest suspend */
3301 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3304 * Using TASK_UNINTERRUPTIBLE because only NOFLUSH internal suspend is
3305 * supported. Properly supporting a TASK_INTERRUPTIBLE internal suspend
3306 * would require changing .presuspend to return an error -- avoid this
3307 * until there is a need for more elaborate variants of internal suspend.
3309 (void) __dm_suspend(md, map, suspend_flags, TASK_UNINTERRUPTIBLE);
3311 set_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3313 dm_table_postsuspend_targets(map);
3316 static void __dm_internal_resume(struct mapped_device *md)
3318 BUG_ON(!md->internal_suspend_count);
3320 if (--md->internal_suspend_count)
3321 return; /* resume from nested internal suspend */
3323 if (dm_suspended_md(md))
3324 goto done; /* resume from nested suspend */
3327 * NOTE: existing callers don't need to call dm_table_resume_targets
3328 * (which may fail -- so best to avoid it for now by passing NULL map)
3330 (void) __dm_resume(md, NULL);
3333 clear_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3334 smp_mb__after_atomic();
3335 wake_up_bit(&md->flags, DMF_SUSPENDED_INTERNALLY);
3338 void dm_internal_suspend_noflush(struct mapped_device *md)
3340 mutex_lock(&md->suspend_lock);
3341 __dm_internal_suspend(md, DM_SUSPEND_NOFLUSH_FLAG);
3342 mutex_unlock(&md->suspend_lock);
3344 EXPORT_SYMBOL_GPL(dm_internal_suspend_noflush);
3346 void dm_internal_resume(struct mapped_device *md)
3348 mutex_lock(&md->suspend_lock);
3349 __dm_internal_resume(md);
3350 mutex_unlock(&md->suspend_lock);
3352 EXPORT_SYMBOL_GPL(dm_internal_resume);
3355 * Fast variants of internal suspend/resume hold md->suspend_lock,
3356 * which prevents interaction with userspace-driven suspend.
3359 void dm_internal_suspend_fast(struct mapped_device *md)
3361 mutex_lock(&md->suspend_lock);
3362 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
3365 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
3366 synchronize_srcu(&md->io_barrier);
3367 flush_workqueue(md->wq);
3368 dm_wait_for_completion(md, TASK_UNINTERRUPTIBLE);
3370 EXPORT_SYMBOL_GPL(dm_internal_suspend_fast);
3372 void dm_internal_resume_fast(struct mapped_device *md)
3374 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
3380 mutex_unlock(&md->suspend_lock);
3382 EXPORT_SYMBOL_GPL(dm_internal_resume_fast);
3384 /*-----------------------------------------------------------------
3385 * Event notification.
3386 *---------------------------------------------------------------*/
3387 int dm_kobject_uevent(struct mapped_device *md, enum kobject_action action,
3390 char udev_cookie[DM_COOKIE_LENGTH];
3391 char *envp[] = { udev_cookie, NULL };
3394 return kobject_uevent(&disk_to_dev(md->disk)->kobj, action);
3396 snprintf(udev_cookie, DM_COOKIE_LENGTH, "%s=%u",
3397 DM_COOKIE_ENV_VAR_NAME, cookie);
3398 return kobject_uevent_env(&disk_to_dev(md->disk)->kobj,
3403 uint32_t dm_next_uevent_seq(struct mapped_device *md)
3405 return atomic_add_return(1, &md->uevent_seq);
3408 uint32_t dm_get_event_nr(struct mapped_device *md)
3410 return atomic_read(&md->event_nr);
3413 int dm_wait_event(struct mapped_device *md, int event_nr)
3415 return wait_event_interruptible(md->eventq,
3416 (event_nr != atomic_read(&md->event_nr)));
3419 void dm_uevent_add(struct mapped_device *md, struct list_head *elist)
3421 unsigned long flags;
3423 spin_lock_irqsave(&md->uevent_lock, flags);
3424 list_add(elist, &md->uevent_list);
3425 spin_unlock_irqrestore(&md->uevent_lock, flags);
3429 * The gendisk is only valid as long as you have a reference
3432 struct gendisk *dm_disk(struct mapped_device *md)
3436 EXPORT_SYMBOL_GPL(dm_disk);
3438 struct kobject *dm_kobject(struct mapped_device *md)
3440 return &md->kobj_holder.kobj;
3443 struct mapped_device *dm_get_from_kobject(struct kobject *kobj)
3445 struct mapped_device *md;
3447 md = container_of(kobj, struct mapped_device, kobj_holder.kobj);
3449 if (test_bit(DMF_FREEING, &md->flags) ||
3457 int dm_suspended_md(struct mapped_device *md)
3459 return test_bit(DMF_SUSPENDED, &md->flags);
3462 int dm_suspended_internally_md(struct mapped_device *md)
3464 return test_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3467 int dm_test_deferred_remove_flag(struct mapped_device *md)
3469 return test_bit(DMF_DEFERRED_REMOVE, &md->flags);
3472 int dm_suspended(struct dm_target *ti)
3474 return dm_suspended_md(dm_table_get_md(ti->table));
3476 EXPORT_SYMBOL_GPL(dm_suspended);
3478 int dm_noflush_suspending(struct dm_target *ti)
3480 return __noflush_suspending(dm_table_get_md(ti->table));
3482 EXPORT_SYMBOL_GPL(dm_noflush_suspending);
3484 struct dm_md_mempools *dm_alloc_md_mempools(struct mapped_device *md, unsigned type,
3485 unsigned integrity, unsigned per_bio_data_size)
3487 struct dm_md_mempools *pools = kzalloc(sizeof(*pools), GFP_KERNEL);
3488 struct kmem_cache *cachep = NULL;
3489 unsigned int pool_size = 0;
3490 unsigned int front_pad;
3495 type = filter_md_type(type, md);
3498 case DM_TYPE_BIO_BASED:
3500 pool_size = dm_get_reserved_bio_based_ios();
3501 front_pad = roundup(per_bio_data_size, __alignof__(struct dm_target_io)) + offsetof(struct dm_target_io, clone);
3503 case DM_TYPE_REQUEST_BASED:
3504 cachep = _rq_tio_cache;
3505 pool_size = dm_get_reserved_rq_based_ios();
3506 pools->rq_pool = mempool_create_slab_pool(pool_size, _rq_cache);
3507 if (!pools->rq_pool)
3509 /* fall through to setup remaining rq-based pools */
3510 case DM_TYPE_MQ_REQUEST_BASED:
3512 pool_size = dm_get_reserved_rq_based_ios();
3513 front_pad = offsetof(struct dm_rq_clone_bio_info, clone);
3514 /* per_bio_data_size is not used. See __bind_mempools(). */
3515 WARN_ON(per_bio_data_size != 0);
3522 pools->io_pool = mempool_create_slab_pool(pool_size, cachep);
3523 if (!pools->io_pool)
3527 pools->bs = bioset_create_nobvec(pool_size, front_pad);
3531 if (integrity && bioset_integrity_create(pools->bs, pool_size))
3537 dm_free_md_mempools(pools);
3542 void dm_free_md_mempools(struct dm_md_mempools *pools)
3547 mempool_destroy(pools->io_pool);
3548 mempool_destroy(pools->rq_pool);
3551 bioset_free(pools->bs);
3556 static int dm_pr_register(struct block_device *bdev, u64 old_key, u64 new_key,
3559 struct mapped_device *md = bdev->bd_disk->private_data;
3560 const struct pr_ops *ops;
3561 struct dm_target *tgt;
3565 r = dm_get_live_table_for_ioctl(md, &tgt, &bdev, &mode, &srcu_idx);
3569 ops = bdev->bd_disk->fops->pr_ops;
3570 if (ops && ops->pr_register)
3571 r = ops->pr_register(bdev, old_key, new_key, flags);
3575 dm_put_live_table(md, srcu_idx);
3579 static int dm_pr_reserve(struct block_device *bdev, u64 key, enum pr_type type,
3582 struct mapped_device *md = bdev->bd_disk->private_data;
3583 const struct pr_ops *ops;
3584 struct dm_target *tgt;
3588 r = dm_get_live_table_for_ioctl(md, &tgt, &bdev, &mode, &srcu_idx);
3592 ops = bdev->bd_disk->fops->pr_ops;
3593 if (ops && ops->pr_reserve)
3594 r = ops->pr_reserve(bdev, key, type, flags);
3598 dm_put_live_table(md, srcu_idx);
3602 static int dm_pr_release(struct block_device *bdev, u64 key, enum pr_type type)
3604 struct mapped_device *md = bdev->bd_disk->private_data;
3605 const struct pr_ops *ops;
3606 struct dm_target *tgt;
3610 r = dm_get_live_table_for_ioctl(md, &tgt, &bdev, &mode, &srcu_idx);
3614 ops = bdev->bd_disk->fops->pr_ops;
3615 if (ops && ops->pr_release)
3616 r = ops->pr_release(bdev, key, type);
3620 dm_put_live_table(md, srcu_idx);
3624 static int dm_pr_preempt(struct block_device *bdev, u64 old_key, u64 new_key,
3625 enum pr_type type, bool abort)
3627 struct mapped_device *md = bdev->bd_disk->private_data;
3628 const struct pr_ops *ops;
3629 struct dm_target *tgt;
3633 r = dm_get_live_table_for_ioctl(md, &tgt, &bdev, &mode, &srcu_idx);
3637 ops = bdev->bd_disk->fops->pr_ops;
3638 if (ops && ops->pr_preempt)
3639 r = ops->pr_preempt(bdev, old_key, new_key, type, abort);
3643 dm_put_live_table(md, srcu_idx);
3647 static int dm_pr_clear(struct block_device *bdev, u64 key)
3649 struct mapped_device *md = bdev->bd_disk->private_data;
3650 const struct pr_ops *ops;
3651 struct dm_target *tgt;
3655 r = dm_get_live_table_for_ioctl(md, &tgt, &bdev, &mode, &srcu_idx);
3659 ops = bdev->bd_disk->fops->pr_ops;
3660 if (ops && ops->pr_clear)
3661 r = ops->pr_clear(bdev, key);
3665 dm_put_live_table(md, srcu_idx);
3669 static const struct pr_ops dm_pr_ops = {
3670 .pr_register = dm_pr_register,
3671 .pr_reserve = dm_pr_reserve,
3672 .pr_release = dm_pr_release,
3673 .pr_preempt = dm_pr_preempt,
3674 .pr_clear = dm_pr_clear,
3677 static const struct block_device_operations dm_blk_dops = {
3678 .open = dm_blk_open,
3679 .release = dm_blk_close,
3680 .ioctl = dm_blk_ioctl,
3681 .getgeo = dm_blk_getgeo,
3682 .pr_ops = &dm_pr_ops,
3683 .owner = THIS_MODULE
3689 module_init(dm_init);
3690 module_exit(dm_exit);
3692 module_param(major, uint, 0);
3693 MODULE_PARM_DESC(major, "The major number of the device mapper");
3695 module_param(reserved_bio_based_ios, uint, S_IRUGO | S_IWUSR);
3696 MODULE_PARM_DESC(reserved_bio_based_ios, "Reserved IOs in bio-based mempools");
3698 module_param(reserved_rq_based_ios, uint, S_IRUGO | S_IWUSR);
3699 MODULE_PARM_DESC(reserved_rq_based_ios, "Reserved IOs in request-based mempools");
3701 module_param(use_blk_mq, bool, S_IRUGO | S_IWUSR);
3702 MODULE_PARM_DESC(use_blk_mq, "Use block multiqueue for request-based DM devices");
3704 MODULE_DESCRIPTION(DM_NAME " driver");
3705 MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
3706 MODULE_LICENSE("GPL");