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 #define MINOR_ALLOCED ((void *)-1)
112 * Bits for the md->flags field.
114 #define DMF_BLOCK_IO_FOR_SUSPEND 0
115 #define DMF_SUSPENDED 1
117 #define DMF_FREEING 3
118 #define DMF_DELETING 4
119 #define DMF_NOFLUSH_SUSPENDING 5
120 #define DMF_DEFERRED_REMOVE 6
121 #define DMF_SUSPENDED_INTERNALLY 7
124 * A dummy definition to make RCU happy.
125 * struct dm_table should never be dereferenced in this file.
132 * Work processed by per-device workqueue.
134 struct mapped_device {
135 struct srcu_struct io_barrier;
136 struct mutex suspend_lock;
141 * The current mapping.
142 * Use dm_get_live_table{_fast} or take suspend_lock for
145 struct dm_table __rcu *map;
147 struct list_head table_devices;
148 struct mutex table_devices_lock;
152 struct request_queue *queue;
154 /* Protect queue and type against concurrent access. */
155 struct mutex type_lock;
157 struct dm_target *immutable_target;
158 struct target_type *immutable_target_type;
160 struct gendisk *disk;
166 * A list of ios that arrived while we were suspended.
169 wait_queue_head_t wait;
170 struct work_struct work;
171 struct bio_list deferred;
172 spinlock_t deferred_lock;
175 * Processing queue (flush)
177 struct workqueue_struct *wq;
180 * io objects are allocated from here.
191 wait_queue_head_t eventq;
193 struct list_head uevent_list;
194 spinlock_t uevent_lock; /* Protect access to uevent_list */
197 * freeze/thaw support require holding onto a super block
199 struct super_block *frozen_sb;
200 struct block_device *bdev;
202 /* forced geometry settings */
203 struct hd_geometry geometry;
205 /* kobject and completion */
206 struct dm_kobject_holder kobj_holder;
208 /* zero-length flush that will be cloned and submitted to targets */
209 struct bio flush_bio;
211 /* the number of internal suspends */
212 unsigned internal_suspend_count;
214 struct dm_stats stats;
216 struct kthread_worker kworker;
217 struct task_struct *kworker_task;
219 /* for request-based merge heuristic in dm_request_fn() */
220 unsigned seq_rq_merge_deadline_usecs;
222 sector_t last_rq_pos;
223 ktime_t last_rq_start_time;
225 /* for blk-mq request-based DM support */
226 struct blk_mq_tag_set *tag_set;
231 #ifdef CONFIG_DM_MQ_DEFAULT
232 static bool use_blk_mq = true;
234 static bool use_blk_mq = false;
237 #define DM_MQ_NR_HW_QUEUES 1
238 #define DM_MQ_QUEUE_DEPTH 2048
240 static unsigned blk_mq_nr_hw_queues = DM_MQ_NR_HW_QUEUES;
241 static unsigned blk_mq_queue_depth = DM_MQ_QUEUE_DEPTH;
243 bool dm_use_blk_mq(struct mapped_device *md)
245 return md->use_blk_mq;
247 EXPORT_SYMBOL_GPL(dm_use_blk_mq);
250 * For mempools pre-allocation at the table loading time.
252 struct dm_md_mempools {
258 struct table_device {
259 struct list_head list;
261 struct dm_dev dm_dev;
264 #define RESERVED_BIO_BASED_IOS 16
265 #define RESERVED_REQUEST_BASED_IOS 256
266 #define RESERVED_MAX_IOS 1024
267 static struct kmem_cache *_io_cache;
268 static struct kmem_cache *_rq_tio_cache;
269 static struct kmem_cache *_rq_cache;
272 * Bio-based DM's mempools' reserved IOs set by the user.
274 static unsigned reserved_bio_based_ios = RESERVED_BIO_BASED_IOS;
277 * Request-based DM's mempools' reserved IOs set by the user.
279 static unsigned reserved_rq_based_ios = RESERVED_REQUEST_BASED_IOS;
281 static unsigned __dm_get_module_param(unsigned *module_param,
282 unsigned def, unsigned max)
284 unsigned param = ACCESS_ONCE(*module_param);
285 unsigned modified_param = 0;
288 modified_param = def;
289 else if (param > max)
290 modified_param = max;
292 if (modified_param) {
293 (void)cmpxchg(module_param, param, modified_param);
294 param = modified_param;
300 unsigned dm_get_reserved_bio_based_ios(void)
302 return __dm_get_module_param(&reserved_bio_based_ios,
303 RESERVED_BIO_BASED_IOS, RESERVED_MAX_IOS);
305 EXPORT_SYMBOL_GPL(dm_get_reserved_bio_based_ios);
307 unsigned dm_get_reserved_rq_based_ios(void)
309 return __dm_get_module_param(&reserved_rq_based_ios,
310 RESERVED_REQUEST_BASED_IOS, RESERVED_MAX_IOS);
312 EXPORT_SYMBOL_GPL(dm_get_reserved_rq_based_ios);
314 unsigned dm_get_blk_mq_nr_hw_queues(void)
316 return __dm_get_module_param(&blk_mq_nr_hw_queues, 1, 32);
319 unsigned dm_get_blk_mq_queue_depth(void)
321 return __dm_get_module_param(&blk_mq_queue_depth,
322 DM_MQ_QUEUE_DEPTH, BLK_MQ_MAX_DEPTH);
325 static int __init local_init(void)
329 /* allocate a slab for the dm_ios */
330 _io_cache = KMEM_CACHE(dm_io, 0);
334 _rq_tio_cache = KMEM_CACHE(dm_rq_target_io, 0);
336 goto out_free_io_cache;
338 _rq_cache = kmem_cache_create("dm_clone_request", sizeof(struct request),
339 __alignof__(struct request), 0, NULL);
341 goto out_free_rq_tio_cache;
343 r = dm_uevent_init();
345 goto out_free_rq_cache;
347 deferred_remove_workqueue = alloc_workqueue("kdmremove", WQ_UNBOUND, 1);
348 if (!deferred_remove_workqueue) {
350 goto out_uevent_exit;
354 r = register_blkdev(_major, _name);
356 goto out_free_workqueue;
364 destroy_workqueue(deferred_remove_workqueue);
368 kmem_cache_destroy(_rq_cache);
369 out_free_rq_tio_cache:
370 kmem_cache_destroy(_rq_tio_cache);
372 kmem_cache_destroy(_io_cache);
377 static void local_exit(void)
379 flush_scheduled_work();
380 destroy_workqueue(deferred_remove_workqueue);
382 kmem_cache_destroy(_rq_cache);
383 kmem_cache_destroy(_rq_tio_cache);
384 kmem_cache_destroy(_io_cache);
385 unregister_blkdev(_major, _name);
390 DMINFO("cleaned up");
393 static int (*_inits[])(void) __initdata = {
404 static void (*_exits[])(void) = {
415 static int __init dm_init(void)
417 const int count = ARRAY_SIZE(_inits);
421 for (i = 0; i < count; i++) {
436 static void __exit dm_exit(void)
438 int i = ARRAY_SIZE(_exits);
444 * Should be empty by this point.
446 idr_destroy(&_minor_idr);
450 * Block device functions
452 int dm_deleting_md(struct mapped_device *md)
454 return test_bit(DMF_DELETING, &md->flags);
457 static int dm_blk_open(struct block_device *bdev, fmode_t mode)
459 struct mapped_device *md;
461 spin_lock(&_minor_lock);
463 md = bdev->bd_disk->private_data;
467 if (test_bit(DMF_FREEING, &md->flags) ||
468 dm_deleting_md(md)) {
474 atomic_inc(&md->open_count);
476 spin_unlock(&_minor_lock);
478 return md ? 0 : -ENXIO;
481 static void dm_blk_close(struct gendisk *disk, fmode_t mode)
483 struct mapped_device *md;
485 spin_lock(&_minor_lock);
487 md = disk->private_data;
491 if (atomic_dec_and_test(&md->open_count) &&
492 (test_bit(DMF_DEFERRED_REMOVE, &md->flags)))
493 queue_work(deferred_remove_workqueue, &deferred_remove_work);
497 spin_unlock(&_minor_lock);
500 int dm_open_count(struct mapped_device *md)
502 return atomic_read(&md->open_count);
506 * Guarantees nothing is using the device before it's deleted.
508 int dm_lock_for_deletion(struct mapped_device *md, bool mark_deferred, bool only_deferred)
512 spin_lock(&_minor_lock);
514 if (dm_open_count(md)) {
517 set_bit(DMF_DEFERRED_REMOVE, &md->flags);
518 } else if (only_deferred && !test_bit(DMF_DEFERRED_REMOVE, &md->flags))
521 set_bit(DMF_DELETING, &md->flags);
523 spin_unlock(&_minor_lock);
528 int dm_cancel_deferred_remove(struct mapped_device *md)
532 spin_lock(&_minor_lock);
534 if (test_bit(DMF_DELETING, &md->flags))
537 clear_bit(DMF_DEFERRED_REMOVE, &md->flags);
539 spin_unlock(&_minor_lock);
544 static void do_deferred_remove(struct work_struct *w)
546 dm_deferred_remove();
549 sector_t dm_get_size(struct mapped_device *md)
551 return get_capacity(md->disk);
554 struct request_queue *dm_get_md_queue(struct mapped_device *md)
559 struct dm_stats *dm_get_stats(struct mapped_device *md)
564 static int dm_blk_getgeo(struct block_device *bdev, struct hd_geometry *geo)
566 struct mapped_device *md = bdev->bd_disk->private_data;
568 return dm_get_geometry(md, geo);
571 static int dm_get_live_table_for_ioctl(struct mapped_device *md,
572 struct dm_target **tgt, struct block_device **bdev,
573 fmode_t *mode, int *srcu_idx)
575 struct dm_table *map;
580 map = dm_get_live_table(md, srcu_idx);
581 if (!map || !dm_table_get_size(map))
584 /* We only support devices that have a single target */
585 if (dm_table_get_num_targets(map) != 1)
588 *tgt = dm_table_get_target(map, 0);
590 if (!(*tgt)->type->prepare_ioctl)
593 if (dm_suspended_md(md)) {
598 r = (*tgt)->type->prepare_ioctl(*tgt, bdev, mode);
605 dm_put_live_table(md, *srcu_idx);
606 if (r == -ENOTCONN && !fatal_signal_pending(current)) {
613 static int dm_blk_ioctl(struct block_device *bdev, fmode_t mode,
614 unsigned int cmd, unsigned long arg)
616 struct mapped_device *md = bdev->bd_disk->private_data;
617 struct dm_target *tgt;
618 struct block_device *tgt_bdev = NULL;
621 r = dm_get_live_table_for_ioctl(md, &tgt, &tgt_bdev, &mode, &srcu_idx);
627 * Target determined this ioctl is being issued against
628 * a logical partition of the parent bdev; so extra
629 * validation is needed.
631 r = scsi_verify_blk_ioctl(NULL, cmd);
636 r = __blkdev_driver_ioctl(tgt_bdev, mode, cmd, arg);
638 dm_put_live_table(md, srcu_idx);
642 static struct dm_io *alloc_io(struct mapped_device *md)
644 return mempool_alloc(md->io_pool, GFP_NOIO);
647 static void free_io(struct mapped_device *md, struct dm_io *io)
649 mempool_free(io, md->io_pool);
652 static void free_tio(struct mapped_device *md, struct dm_target_io *tio)
654 bio_put(&tio->clone);
657 static struct dm_rq_target_io *alloc_rq_tio(struct mapped_device *md,
660 return mempool_alloc(md->io_pool, gfp_mask);
663 static void free_rq_tio(struct dm_rq_target_io *tio)
665 mempool_free(tio, tio->md->io_pool);
668 static struct request *alloc_clone_request(struct mapped_device *md,
671 return mempool_alloc(md->rq_pool, gfp_mask);
674 static void free_clone_request(struct mapped_device *md, struct request *rq)
676 mempool_free(rq, md->rq_pool);
679 static int md_in_flight(struct mapped_device *md)
681 return atomic_read(&md->pending[READ]) +
682 atomic_read(&md->pending[WRITE]);
685 static void start_io_acct(struct dm_io *io)
687 struct mapped_device *md = io->md;
688 struct bio *bio = io->bio;
690 int rw = bio_data_dir(bio);
692 io->start_time = jiffies;
694 cpu = part_stat_lock();
695 part_round_stats(cpu, &dm_disk(md)->part0);
697 atomic_set(&dm_disk(md)->part0.in_flight[rw],
698 atomic_inc_return(&md->pending[rw]));
700 if (unlikely(dm_stats_used(&md->stats)))
701 dm_stats_account_io(&md->stats, bio->bi_rw, bio->bi_iter.bi_sector,
702 bio_sectors(bio), false, 0, &io->stats_aux);
705 static void end_io_acct(struct dm_io *io)
707 struct mapped_device *md = io->md;
708 struct bio *bio = io->bio;
709 unsigned long duration = jiffies - io->start_time;
711 int rw = bio_data_dir(bio);
713 generic_end_io_acct(rw, &dm_disk(md)->part0, io->start_time);
715 if (unlikely(dm_stats_used(&md->stats)))
716 dm_stats_account_io(&md->stats, bio->bi_rw, bio->bi_iter.bi_sector,
717 bio_sectors(bio), true, duration, &io->stats_aux);
720 * After this is decremented the bio must not be touched if it is
723 pending = atomic_dec_return(&md->pending[rw]);
724 atomic_set(&dm_disk(md)->part0.in_flight[rw], pending);
725 pending += atomic_read(&md->pending[rw^0x1]);
727 /* nudge anyone waiting on suspend queue */
733 * Add the bio to the list of deferred io.
735 static void queue_io(struct mapped_device *md, struct bio *bio)
739 spin_lock_irqsave(&md->deferred_lock, flags);
740 bio_list_add(&md->deferred, bio);
741 spin_unlock_irqrestore(&md->deferred_lock, flags);
742 queue_work(md->wq, &md->work);
746 * Everyone (including functions in this file), should use this
747 * function to access the md->map field, and make sure they call
748 * dm_put_live_table() when finished.
750 struct dm_table *dm_get_live_table(struct mapped_device *md, int *srcu_idx) __acquires(md->io_barrier)
752 *srcu_idx = srcu_read_lock(&md->io_barrier);
754 return srcu_dereference(md->map, &md->io_barrier);
757 void dm_put_live_table(struct mapped_device *md, int srcu_idx) __releases(md->io_barrier)
759 srcu_read_unlock(&md->io_barrier, srcu_idx);
762 void dm_sync_table(struct mapped_device *md)
764 synchronize_srcu(&md->io_barrier);
765 synchronize_rcu_expedited();
769 * A fast alternative to dm_get_live_table/dm_put_live_table.
770 * The caller must not block between these two functions.
772 static struct dm_table *dm_get_live_table_fast(struct mapped_device *md) __acquires(RCU)
775 return rcu_dereference(md->map);
778 static void dm_put_live_table_fast(struct mapped_device *md) __releases(RCU)
784 * Open a table device so we can use it as a map destination.
786 static int open_table_device(struct table_device *td, dev_t dev,
787 struct mapped_device *md)
789 static char *_claim_ptr = "I belong to device-mapper";
790 struct block_device *bdev;
794 BUG_ON(td->dm_dev.bdev);
796 bdev = blkdev_get_by_dev(dev, td->dm_dev.mode | FMODE_EXCL, _claim_ptr);
798 return PTR_ERR(bdev);
800 r = bd_link_disk_holder(bdev, dm_disk(md));
802 blkdev_put(bdev, td->dm_dev.mode | FMODE_EXCL);
806 td->dm_dev.bdev = bdev;
811 * Close a table device that we've been using.
813 static void close_table_device(struct table_device *td, struct mapped_device *md)
815 if (!td->dm_dev.bdev)
818 bd_unlink_disk_holder(td->dm_dev.bdev, dm_disk(md));
819 blkdev_put(td->dm_dev.bdev, td->dm_dev.mode | FMODE_EXCL);
820 td->dm_dev.bdev = NULL;
823 static struct table_device *find_table_device(struct list_head *l, dev_t dev,
825 struct table_device *td;
827 list_for_each_entry(td, l, list)
828 if (td->dm_dev.bdev->bd_dev == dev && td->dm_dev.mode == mode)
834 int dm_get_table_device(struct mapped_device *md, dev_t dev, fmode_t mode,
835 struct dm_dev **result) {
837 struct table_device *td;
839 mutex_lock(&md->table_devices_lock);
840 td = find_table_device(&md->table_devices, dev, mode);
842 td = kmalloc(sizeof(*td), GFP_KERNEL);
844 mutex_unlock(&md->table_devices_lock);
848 td->dm_dev.mode = mode;
849 td->dm_dev.bdev = NULL;
851 if ((r = open_table_device(td, dev, md))) {
852 mutex_unlock(&md->table_devices_lock);
857 format_dev_t(td->dm_dev.name, dev);
859 atomic_set(&td->count, 0);
860 list_add(&td->list, &md->table_devices);
862 atomic_inc(&td->count);
863 mutex_unlock(&md->table_devices_lock);
865 *result = &td->dm_dev;
868 EXPORT_SYMBOL_GPL(dm_get_table_device);
870 void dm_put_table_device(struct mapped_device *md, struct dm_dev *d)
872 struct table_device *td = container_of(d, struct table_device, dm_dev);
874 mutex_lock(&md->table_devices_lock);
875 if (atomic_dec_and_test(&td->count)) {
876 close_table_device(td, md);
880 mutex_unlock(&md->table_devices_lock);
882 EXPORT_SYMBOL(dm_put_table_device);
884 static void free_table_devices(struct list_head *devices)
886 struct list_head *tmp, *next;
888 list_for_each_safe(tmp, next, devices) {
889 struct table_device *td = list_entry(tmp, struct table_device, list);
891 DMWARN("dm_destroy: %s still exists with %d references",
892 td->dm_dev.name, atomic_read(&td->count));
898 * Get the geometry associated with a dm device
900 int dm_get_geometry(struct mapped_device *md, struct hd_geometry *geo)
908 * Set the geometry of a device.
910 int dm_set_geometry(struct mapped_device *md, struct hd_geometry *geo)
912 sector_t sz = (sector_t)geo->cylinders * geo->heads * geo->sectors;
914 if (geo->start > sz) {
915 DMWARN("Start sector is beyond the geometry limits.");
924 /*-----------------------------------------------------------------
926 * A more elegant soln is in the works that uses the queue
927 * merge fn, unfortunately there are a couple of changes to
928 * the block layer that I want to make for this. So in the
929 * interests of getting something for people to use I give
930 * you this clearly demarcated crap.
931 *---------------------------------------------------------------*/
933 static int __noflush_suspending(struct mapped_device *md)
935 return test_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
939 * Decrements the number of outstanding ios that a bio has been
940 * cloned into, completing the original io if necc.
942 static void dec_pending(struct dm_io *io, int error)
947 struct mapped_device *md = io->md;
949 /* Push-back supersedes any I/O errors */
950 if (unlikely(error)) {
951 spin_lock_irqsave(&io->endio_lock, flags);
952 if (!(io->error > 0 && __noflush_suspending(md)))
954 spin_unlock_irqrestore(&io->endio_lock, flags);
957 if (atomic_dec_and_test(&io->io_count)) {
958 if (io->error == DM_ENDIO_REQUEUE) {
960 * Target requested pushing back the I/O.
962 spin_lock_irqsave(&md->deferred_lock, flags);
963 if (__noflush_suspending(md))
964 bio_list_add_head(&md->deferred, io->bio);
966 /* noflush suspend was interrupted. */
968 spin_unlock_irqrestore(&md->deferred_lock, flags);
971 io_error = io->error;
976 if (io_error == DM_ENDIO_REQUEUE)
979 if ((bio->bi_rw & REQ_FLUSH) && bio->bi_iter.bi_size) {
981 * Preflush done for flush with data, reissue
984 bio->bi_rw &= ~REQ_FLUSH;
987 /* done with normal IO or empty flush */
988 trace_block_bio_complete(md->queue, bio, io_error);
989 bio->bi_error = io_error;
995 static void disable_write_same(struct mapped_device *md)
997 struct queue_limits *limits = dm_get_queue_limits(md);
999 /* device doesn't really support WRITE SAME, disable it */
1000 limits->max_write_same_sectors = 0;
1003 static void clone_endio(struct bio *bio)
1005 int error = bio->bi_error;
1007 struct dm_target_io *tio = container_of(bio, struct dm_target_io, clone);
1008 struct dm_io *io = tio->io;
1009 struct mapped_device *md = tio->io->md;
1010 dm_endio_fn endio = tio->ti->type->end_io;
1013 r = endio(tio->ti, bio, error);
1014 if (r < 0 || r == DM_ENDIO_REQUEUE)
1016 * error and requeue request are handled
1020 else if (r == DM_ENDIO_INCOMPLETE)
1021 /* The target will handle the io */
1024 DMWARN("unimplemented target endio return value: %d", r);
1029 if (unlikely(r == -EREMOTEIO && (bio->bi_rw & REQ_WRITE_SAME) &&
1030 !bdev_get_queue(bio->bi_bdev)->limits.max_write_same_sectors))
1031 disable_write_same(md);
1034 dec_pending(io, error);
1038 * Partial completion handling for request-based dm
1040 static void end_clone_bio(struct bio *clone)
1042 struct dm_rq_clone_bio_info *info =
1043 container_of(clone, struct dm_rq_clone_bio_info, clone);
1044 struct dm_rq_target_io *tio = info->tio;
1045 struct bio *bio = info->orig;
1046 unsigned int nr_bytes = info->orig->bi_iter.bi_size;
1047 int error = clone->bi_error;
1053 * An error has already been detected on the request.
1054 * Once error occurred, just let clone->end_io() handle
1060 * Don't notice the error to the upper layer yet.
1061 * The error handling decision is made by the target driver,
1062 * when the request is completed.
1069 * I/O for the bio successfully completed.
1070 * Notice the data completion to the upper layer.
1074 * bios are processed from the head of the list.
1075 * So the completing bio should always be rq->bio.
1076 * If it's not, something wrong is happening.
1078 if (tio->orig->bio != bio)
1079 DMERR("bio completion is going in the middle of the request");
1082 * Update the original request.
1083 * Do not use blk_end_request() here, because it may complete
1084 * the original request before the clone, and break the ordering.
1086 blk_update_request(tio->orig, 0, nr_bytes);
1089 static struct dm_rq_target_io *tio_from_request(struct request *rq)
1091 return (rq->q->mq_ops ? blk_mq_rq_to_pdu(rq) : rq->special);
1094 static void rq_end_stats(struct mapped_device *md, struct request *orig)
1096 if (unlikely(dm_stats_used(&md->stats))) {
1097 struct dm_rq_target_io *tio = tio_from_request(orig);
1098 tio->duration_jiffies = jiffies - tio->duration_jiffies;
1099 dm_stats_account_io(&md->stats, orig->cmd_flags, blk_rq_pos(orig),
1100 tio->n_sectors, true, tio->duration_jiffies,
1106 * Don't touch any member of the md after calling this function because
1107 * the md may be freed in dm_put() at the end of this function.
1108 * Or do dm_get() before calling this function and dm_put() later.
1110 static void rq_completed(struct mapped_device *md, int rw, bool run_queue)
1112 atomic_dec(&md->pending[rw]);
1114 /* nudge anyone waiting on suspend queue */
1115 if (!md_in_flight(md))
1119 * Run this off this callpath, as drivers could invoke end_io while
1120 * inside their request_fn (and holding the queue lock). Calling
1121 * back into ->request_fn() could deadlock attempting to grab the
1124 if (!md->queue->mq_ops && run_queue)
1125 blk_run_queue_async(md->queue);
1128 * dm_put() must be at the end of this function. See the comment above
1133 static void free_rq_clone(struct request *clone)
1135 struct dm_rq_target_io *tio = clone->end_io_data;
1136 struct mapped_device *md = tio->md;
1138 blk_rq_unprep_clone(clone);
1140 if (md->type == DM_TYPE_MQ_REQUEST_BASED)
1141 /* stacked on blk-mq queue(s) */
1142 tio->ti->type->release_clone_rq(clone);
1143 else if (!md->queue->mq_ops)
1144 /* request_fn queue stacked on request_fn queue(s) */
1145 free_clone_request(md, clone);
1147 * NOTE: for the blk-mq queue stacked on request_fn queue(s) case:
1148 * no need to call free_clone_request() because we leverage blk-mq by
1149 * allocating the clone at the end of the blk-mq pdu (see: clone_rq)
1152 if (!md->queue->mq_ops)
1157 * Complete the clone and the original request.
1158 * Must be called without clone's queue lock held,
1159 * see end_clone_request() for more details.
1161 static void dm_end_request(struct request *clone, int error)
1163 int rw = rq_data_dir(clone);
1164 struct dm_rq_target_io *tio = clone->end_io_data;
1165 struct mapped_device *md = tio->md;
1166 struct request *rq = tio->orig;
1168 if (rq->cmd_type == REQ_TYPE_BLOCK_PC) {
1169 rq->errors = clone->errors;
1170 rq->resid_len = clone->resid_len;
1174 * We are using the sense buffer of the original
1176 * So setting the length of the sense data is enough.
1178 rq->sense_len = clone->sense_len;
1181 free_rq_clone(clone);
1182 rq_end_stats(md, rq);
1184 blk_end_request_all(rq, error);
1186 blk_mq_end_request(rq, error);
1187 rq_completed(md, rw, true);
1190 static void dm_unprep_request(struct request *rq)
1192 struct dm_rq_target_io *tio = tio_from_request(rq);
1193 struct request *clone = tio->clone;
1195 if (!rq->q->mq_ops) {
1197 rq->cmd_flags &= ~REQ_DONTPREP;
1201 free_rq_clone(clone);
1205 * Requeue the original request of a clone.
1207 static void old_requeue_request(struct request *rq)
1209 struct request_queue *q = rq->q;
1210 unsigned long flags;
1212 spin_lock_irqsave(q->queue_lock, flags);
1213 blk_requeue_request(q, rq);
1214 blk_run_queue_async(q);
1215 spin_unlock_irqrestore(q->queue_lock, flags);
1218 static void dm_requeue_original_request(struct mapped_device *md,
1221 int rw = rq_data_dir(rq);
1223 dm_unprep_request(rq);
1225 rq_end_stats(md, rq);
1227 old_requeue_request(rq);
1229 blk_mq_requeue_request(rq);
1230 blk_mq_kick_requeue_list(rq->q);
1233 rq_completed(md, rw, false);
1236 static void old_stop_queue(struct request_queue *q)
1238 unsigned long flags;
1240 if (blk_queue_stopped(q))
1243 spin_lock_irqsave(q->queue_lock, flags);
1245 spin_unlock_irqrestore(q->queue_lock, flags);
1248 static void stop_queue(struct request_queue *q)
1253 blk_mq_stop_hw_queues(q);
1256 static void old_start_queue(struct request_queue *q)
1258 unsigned long flags;
1260 spin_lock_irqsave(q->queue_lock, flags);
1261 if (blk_queue_stopped(q))
1263 spin_unlock_irqrestore(q->queue_lock, flags);
1266 static void start_queue(struct request_queue *q)
1271 blk_mq_start_stopped_hw_queues(q, true);
1274 static void dm_done(struct request *clone, int error, bool mapped)
1277 struct dm_rq_target_io *tio = clone->end_io_data;
1278 dm_request_endio_fn rq_end_io = NULL;
1281 rq_end_io = tio->ti->type->rq_end_io;
1283 if (mapped && rq_end_io)
1284 r = rq_end_io(tio->ti, clone, error, &tio->info);
1287 if (unlikely(r == -EREMOTEIO && (clone->cmd_flags & REQ_WRITE_SAME) &&
1288 !clone->q->limits.max_write_same_sectors))
1289 disable_write_same(tio->md);
1292 /* The target wants to complete the I/O */
1293 dm_end_request(clone, r);
1294 else if (r == DM_ENDIO_INCOMPLETE)
1295 /* The target will handle the I/O */
1297 else if (r == DM_ENDIO_REQUEUE)
1298 /* The target wants to requeue the I/O */
1299 dm_requeue_original_request(tio->md, tio->orig);
1301 DMWARN("unimplemented target endio return value: %d", r);
1307 * Request completion handler for request-based dm
1309 static void dm_softirq_done(struct request *rq)
1312 struct dm_rq_target_io *tio = tio_from_request(rq);
1313 struct request *clone = tio->clone;
1317 rq_end_stats(tio->md, rq);
1318 rw = rq_data_dir(rq);
1319 if (!rq->q->mq_ops) {
1320 blk_end_request_all(rq, tio->error);
1321 rq_completed(tio->md, rw, false);
1324 blk_mq_end_request(rq, tio->error);
1325 rq_completed(tio->md, rw, false);
1330 if (rq->cmd_flags & REQ_FAILED)
1333 dm_done(clone, tio->error, mapped);
1337 * Complete the clone and the original request with the error status
1338 * through softirq context.
1340 static void dm_complete_request(struct request *rq, int error)
1342 struct dm_rq_target_io *tio = tio_from_request(rq);
1346 blk_complete_request(rq);
1348 blk_mq_complete_request(rq, rq->errors);
1352 * Complete the not-mapped clone and the original request with the error status
1353 * through softirq context.
1354 * Target's rq_end_io() function isn't called.
1355 * This may be used when the target's map_rq() or clone_and_map_rq() functions fail.
1357 static void dm_kill_unmapped_request(struct request *rq, int error)
1359 rq->cmd_flags |= REQ_FAILED;
1360 dm_complete_request(rq, error);
1364 * Called with the clone's queue lock held (for non-blk-mq)
1366 static void end_clone_request(struct request *clone, int error)
1368 struct dm_rq_target_io *tio = clone->end_io_data;
1370 if (!clone->q->mq_ops) {
1372 * For just cleaning up the information of the queue in which
1373 * the clone was dispatched.
1374 * The clone is *NOT* freed actually here because it is alloced
1375 * from dm own mempool (REQ_ALLOCED isn't set).
1377 __blk_put_request(clone->q, clone);
1381 * Actual request completion is done in a softirq context which doesn't
1382 * hold the clone's queue lock. Otherwise, deadlock could occur because:
1383 * - another request may be submitted by the upper level driver
1384 * of the stacking during the completion
1385 * - the submission which requires queue lock may be done
1386 * against this clone's queue
1388 dm_complete_request(tio->orig, error);
1392 * Return maximum size of I/O possible at the supplied sector up to the current
1395 static sector_t max_io_len_target_boundary(sector_t sector, struct dm_target *ti)
1397 sector_t target_offset = dm_target_offset(ti, sector);
1399 return ti->len - target_offset;
1402 static sector_t max_io_len(sector_t sector, struct dm_target *ti)
1404 sector_t len = max_io_len_target_boundary(sector, ti);
1405 sector_t offset, max_len;
1408 * Does the target need to split even further?
1410 if (ti->max_io_len) {
1411 offset = dm_target_offset(ti, sector);
1412 if (unlikely(ti->max_io_len & (ti->max_io_len - 1)))
1413 max_len = sector_div(offset, ti->max_io_len);
1415 max_len = offset & (ti->max_io_len - 1);
1416 max_len = ti->max_io_len - max_len;
1425 int dm_set_target_max_io_len(struct dm_target *ti, sector_t len)
1427 if (len > UINT_MAX) {
1428 DMERR("Specified maximum size of target IO (%llu) exceeds limit (%u)",
1429 (unsigned long long)len, UINT_MAX);
1430 ti->error = "Maximum size of target IO is too large";
1434 ti->max_io_len = (uint32_t) len;
1438 EXPORT_SYMBOL_GPL(dm_set_target_max_io_len);
1441 * A target may call dm_accept_partial_bio only from the map routine. It is
1442 * allowed for all bio types except REQ_FLUSH.
1444 * dm_accept_partial_bio informs the dm that the target only wants to process
1445 * additional n_sectors sectors of the bio and the rest of the data should be
1446 * sent in a next bio.
1448 * A diagram that explains the arithmetics:
1449 * +--------------------+---------------+-------+
1451 * +--------------------+---------------+-------+
1453 * <-------------- *tio->len_ptr --------------->
1454 * <------- bi_size ------->
1457 * Region 1 was already iterated over with bio_advance or similar function.
1458 * (it may be empty if the target doesn't use bio_advance)
1459 * Region 2 is the remaining bio size that the target wants to process.
1460 * (it may be empty if region 1 is non-empty, although there is no reason
1462 * The target requires that region 3 is to be sent in the next bio.
1464 * If the target wants to receive multiple copies of the bio (via num_*bios, etc),
1465 * the partially processed part (the sum of regions 1+2) must be the same for all
1466 * copies of the bio.
1468 void dm_accept_partial_bio(struct bio *bio, unsigned n_sectors)
1470 struct dm_target_io *tio = container_of(bio, struct dm_target_io, clone);
1471 unsigned bi_size = bio->bi_iter.bi_size >> SECTOR_SHIFT;
1472 BUG_ON(bio->bi_rw & REQ_FLUSH);
1473 BUG_ON(bi_size > *tio->len_ptr);
1474 BUG_ON(n_sectors > bi_size);
1475 *tio->len_ptr -= bi_size - n_sectors;
1476 bio->bi_iter.bi_size = n_sectors << SECTOR_SHIFT;
1478 EXPORT_SYMBOL_GPL(dm_accept_partial_bio);
1480 static void __map_bio(struct dm_target_io *tio)
1484 struct mapped_device *md;
1485 struct bio *clone = &tio->clone;
1486 struct dm_target *ti = tio->ti;
1488 clone->bi_end_io = clone_endio;
1491 * Map the clone. If r == 0 we don't need to do
1492 * anything, the target has assumed ownership of
1495 atomic_inc(&tio->io->io_count);
1496 sector = clone->bi_iter.bi_sector;
1497 r = ti->type->map(ti, clone);
1498 if (r == DM_MAPIO_REMAPPED) {
1499 /* the bio has been remapped so dispatch it */
1501 trace_block_bio_remap(bdev_get_queue(clone->bi_bdev), clone,
1502 tio->io->bio->bi_bdev->bd_dev, sector);
1504 generic_make_request(clone);
1505 } else if (r < 0 || r == DM_MAPIO_REQUEUE) {
1506 /* error the io and bail out, or requeue it if needed */
1508 dec_pending(tio->io, r);
1510 } else if (r != DM_MAPIO_SUBMITTED) {
1511 DMWARN("unimplemented target map return value: %d", r);
1517 struct mapped_device *md;
1518 struct dm_table *map;
1522 unsigned sector_count;
1525 static void bio_setup_sector(struct bio *bio, sector_t sector, unsigned len)
1527 bio->bi_iter.bi_sector = sector;
1528 bio->bi_iter.bi_size = to_bytes(len);
1532 * Creates a bio that consists of range of complete bvecs.
1534 static void clone_bio(struct dm_target_io *tio, struct bio *bio,
1535 sector_t sector, unsigned len)
1537 struct bio *clone = &tio->clone;
1539 __bio_clone_fast(clone, bio);
1541 if (bio_integrity(bio))
1542 bio_integrity_clone(clone, bio, GFP_NOIO);
1544 bio_advance(clone, to_bytes(sector - clone->bi_iter.bi_sector));
1545 clone->bi_iter.bi_size = to_bytes(len);
1547 if (bio_integrity(bio))
1548 bio_integrity_trim(clone, 0, len);
1551 static struct dm_target_io *alloc_tio(struct clone_info *ci,
1552 struct dm_target *ti,
1553 unsigned target_bio_nr)
1555 struct dm_target_io *tio;
1558 clone = bio_alloc_bioset(GFP_NOIO, 0, ci->md->bs);
1559 tio = container_of(clone, struct dm_target_io, clone);
1563 tio->target_bio_nr = target_bio_nr;
1568 static void __clone_and_map_simple_bio(struct clone_info *ci,
1569 struct dm_target *ti,
1570 unsigned target_bio_nr, unsigned *len)
1572 struct dm_target_io *tio = alloc_tio(ci, ti, target_bio_nr);
1573 struct bio *clone = &tio->clone;
1577 __bio_clone_fast(clone, ci->bio);
1579 bio_setup_sector(clone, ci->sector, *len);
1584 static void __send_duplicate_bios(struct clone_info *ci, struct dm_target *ti,
1585 unsigned num_bios, unsigned *len)
1587 unsigned target_bio_nr;
1589 for (target_bio_nr = 0; target_bio_nr < num_bios; target_bio_nr++)
1590 __clone_and_map_simple_bio(ci, ti, target_bio_nr, len);
1593 static int __send_empty_flush(struct clone_info *ci)
1595 unsigned target_nr = 0;
1596 struct dm_target *ti;
1598 BUG_ON(bio_has_data(ci->bio));
1599 while ((ti = dm_table_get_target(ci->map, target_nr++)))
1600 __send_duplicate_bios(ci, ti, ti->num_flush_bios, NULL);
1605 static void __clone_and_map_data_bio(struct clone_info *ci, struct dm_target *ti,
1606 sector_t sector, unsigned *len)
1608 struct bio *bio = ci->bio;
1609 struct dm_target_io *tio;
1610 unsigned target_bio_nr;
1611 unsigned num_target_bios = 1;
1614 * Does the target want to receive duplicate copies of the bio?
1616 if (bio_data_dir(bio) == WRITE && ti->num_write_bios)
1617 num_target_bios = ti->num_write_bios(ti, bio);
1619 for (target_bio_nr = 0; target_bio_nr < num_target_bios; target_bio_nr++) {
1620 tio = alloc_tio(ci, ti, target_bio_nr);
1622 clone_bio(tio, bio, sector, *len);
1627 typedef unsigned (*get_num_bios_fn)(struct dm_target *ti);
1629 static unsigned get_num_discard_bios(struct dm_target *ti)
1631 return ti->num_discard_bios;
1634 static unsigned get_num_write_same_bios(struct dm_target *ti)
1636 return ti->num_write_same_bios;
1639 typedef bool (*is_split_required_fn)(struct dm_target *ti);
1641 static bool is_split_required_for_discard(struct dm_target *ti)
1643 return ti->split_discard_bios;
1646 static int __send_changing_extent_only(struct clone_info *ci,
1647 get_num_bios_fn get_num_bios,
1648 is_split_required_fn is_split_required)
1650 struct dm_target *ti;
1655 ti = dm_table_find_target(ci->map, ci->sector);
1656 if (!dm_target_is_valid(ti))
1660 * Even though the device advertised support for this type of
1661 * request, that does not mean every target supports it, and
1662 * reconfiguration might also have changed that since the
1663 * check was performed.
1665 num_bios = get_num_bios ? get_num_bios(ti) : 0;
1669 if (is_split_required && !is_split_required(ti))
1670 len = min((sector_t)ci->sector_count, max_io_len_target_boundary(ci->sector, ti));
1672 len = min((sector_t)ci->sector_count, max_io_len(ci->sector, ti));
1674 __send_duplicate_bios(ci, ti, num_bios, &len);
1677 } while (ci->sector_count -= len);
1682 static int __send_discard(struct clone_info *ci)
1684 return __send_changing_extent_only(ci, get_num_discard_bios,
1685 is_split_required_for_discard);
1688 static int __send_write_same(struct clone_info *ci)
1690 return __send_changing_extent_only(ci, get_num_write_same_bios, NULL);
1694 * Select the correct strategy for processing a non-flush bio.
1696 static int __split_and_process_non_flush(struct clone_info *ci)
1698 struct bio *bio = ci->bio;
1699 struct dm_target *ti;
1702 if (unlikely(bio->bi_rw & REQ_DISCARD))
1703 return __send_discard(ci);
1704 else if (unlikely(bio->bi_rw & REQ_WRITE_SAME))
1705 return __send_write_same(ci);
1707 ti = dm_table_find_target(ci->map, ci->sector);
1708 if (!dm_target_is_valid(ti))
1711 len = min_t(sector_t, max_io_len(ci->sector, ti), ci->sector_count);
1713 __clone_and_map_data_bio(ci, ti, ci->sector, &len);
1716 ci->sector_count -= len;
1722 * Entry point to split a bio into clones and submit them to the targets.
1724 static void __split_and_process_bio(struct mapped_device *md,
1725 struct dm_table *map, struct bio *bio)
1727 struct clone_info ci;
1730 if (unlikely(!map)) {
1737 ci.io = alloc_io(md);
1739 atomic_set(&ci.io->io_count, 1);
1742 spin_lock_init(&ci.io->endio_lock);
1743 ci.sector = bio->bi_iter.bi_sector;
1745 start_io_acct(ci.io);
1747 if (bio->bi_rw & REQ_FLUSH) {
1748 ci.bio = &ci.md->flush_bio;
1749 ci.sector_count = 0;
1750 error = __send_empty_flush(&ci);
1751 /* dec_pending submits any data associated with flush */
1754 ci.sector_count = bio_sectors(bio);
1755 while (ci.sector_count && !error)
1756 error = __split_and_process_non_flush(&ci);
1759 /* drop the extra reference count */
1760 dec_pending(ci.io, error);
1762 /*-----------------------------------------------------------------
1764 *---------------------------------------------------------------*/
1767 * The request function that just remaps the bio built up by
1770 static blk_qc_t dm_make_request(struct request_queue *q, struct bio *bio)
1772 int rw = bio_data_dir(bio);
1773 struct mapped_device *md = q->queuedata;
1775 struct dm_table *map;
1777 map = dm_get_live_table(md, &srcu_idx);
1779 generic_start_io_acct(rw, bio_sectors(bio), &dm_disk(md)->part0);
1781 /* if we're suspended, we have to queue this io for later */
1782 if (unlikely(test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags))) {
1783 dm_put_live_table(md, srcu_idx);
1785 if (bio_rw(bio) != READA)
1789 return BLK_QC_T_NONE;
1792 __split_and_process_bio(md, map, bio);
1793 dm_put_live_table(md, srcu_idx);
1794 return BLK_QC_T_NONE;
1797 int dm_request_based(struct mapped_device *md)
1799 return blk_queue_stackable(md->queue);
1802 static void dm_dispatch_clone_request(struct request *clone, struct request *rq)
1806 if (blk_queue_io_stat(clone->q))
1807 clone->cmd_flags |= REQ_IO_STAT;
1809 clone->start_time = jiffies;
1810 r = blk_insert_cloned_request(clone->q, clone);
1812 /* must complete clone in terms of original request */
1813 dm_complete_request(rq, r);
1816 static int dm_rq_bio_constructor(struct bio *bio, struct bio *bio_orig,
1819 struct dm_rq_target_io *tio = data;
1820 struct dm_rq_clone_bio_info *info =
1821 container_of(bio, struct dm_rq_clone_bio_info, clone);
1823 info->orig = bio_orig;
1825 bio->bi_end_io = end_clone_bio;
1830 static int setup_clone(struct request *clone, struct request *rq,
1831 struct dm_rq_target_io *tio, gfp_t gfp_mask)
1835 r = blk_rq_prep_clone(clone, rq, tio->md->bs, gfp_mask,
1836 dm_rq_bio_constructor, tio);
1840 clone->cmd = rq->cmd;
1841 clone->cmd_len = rq->cmd_len;
1842 clone->sense = rq->sense;
1843 clone->end_io = end_clone_request;
1844 clone->end_io_data = tio;
1851 static struct request *clone_rq(struct request *rq, struct mapped_device *md,
1852 struct dm_rq_target_io *tio, gfp_t gfp_mask)
1855 * Create clone for use with .request_fn request_queue
1857 * Do not allocate a clone if tio->clone was already set
1858 * (see: dm_mq_queue_rq).
1860 bool alloc_clone = !tio->clone;
1861 struct request *clone;
1864 clone = alloc_clone_request(md, gfp_mask);
1870 blk_rq_init(NULL, clone);
1871 if (setup_clone(clone, rq, tio, gfp_mask)) {
1874 free_clone_request(md, clone);
1881 static void map_tio_request(struct kthread_work *work);
1883 static void init_tio(struct dm_rq_target_io *tio, struct request *rq,
1884 struct mapped_device *md)
1892 * Avoid initializing info for blk-mq; it passes
1893 * target-specific data through info.ptr
1894 * (see: dm_mq_init_request)
1896 if (!md->init_tio_pdu)
1897 memset(&tio->info, 0, sizeof(tio->info));
1898 if (md->kworker_task)
1899 init_kthread_work(&tio->work, map_tio_request);
1902 static struct dm_rq_target_io *prep_tio(struct request *rq,
1903 struct mapped_device *md, gfp_t gfp_mask)
1905 struct dm_rq_target_io *tio;
1907 struct dm_table *table;
1909 tio = alloc_rq_tio(md, gfp_mask);
1913 init_tio(tio, rq, md);
1915 table = dm_get_live_table(md, &srcu_idx);
1916 if (!dm_table_mq_request_based(table)) {
1917 if (!clone_rq(rq, md, tio, gfp_mask)) {
1918 dm_put_live_table(md, srcu_idx);
1923 dm_put_live_table(md, srcu_idx);
1929 * Called with the queue lock held.
1931 static int dm_prep_fn(struct request_queue *q, struct request *rq)
1933 struct mapped_device *md = q->queuedata;
1934 struct dm_rq_target_io *tio;
1936 if (unlikely(rq->special)) {
1937 DMWARN("Already has something in rq->special.");
1938 return BLKPREP_KILL;
1941 tio = prep_tio(rq, md, GFP_ATOMIC);
1943 return BLKPREP_DEFER;
1946 rq->cmd_flags |= REQ_DONTPREP;
1953 * 0 : the request has been processed
1954 * DM_MAPIO_REQUEUE : the original request needs to be requeued
1955 * < 0 : the request was completed due to failure
1957 static int map_request(struct dm_rq_target_io *tio, struct request *rq,
1958 struct mapped_device *md)
1961 struct dm_target *ti = tio->ti;
1962 struct request *clone = NULL;
1966 r = ti->type->map_rq(ti, clone, &tio->info);
1968 r = ti->type->clone_and_map_rq(ti, rq, &tio->info, &clone);
1970 /* The target wants to complete the I/O */
1971 dm_kill_unmapped_request(rq, r);
1974 if (r != DM_MAPIO_REMAPPED)
1976 if (setup_clone(clone, rq, tio, GFP_ATOMIC)) {
1978 ti->type->release_clone_rq(clone);
1979 return DM_MAPIO_REQUEUE;
1984 case DM_MAPIO_SUBMITTED:
1985 /* The target has taken the I/O to submit by itself later */
1987 case DM_MAPIO_REMAPPED:
1988 /* The target has remapped the I/O so dispatch it */
1989 trace_block_rq_remap(clone->q, clone, disk_devt(dm_disk(md)),
1991 dm_dispatch_clone_request(clone, rq);
1993 case DM_MAPIO_REQUEUE:
1994 /* The target wants to requeue the I/O */
1995 dm_requeue_original_request(md, tio->orig);
1999 DMWARN("unimplemented target map return value: %d", r);
2003 /* The target wants to complete the I/O */
2004 dm_kill_unmapped_request(rq, r);
2011 static void map_tio_request(struct kthread_work *work)
2013 struct dm_rq_target_io *tio = container_of(work, struct dm_rq_target_io, work);
2014 struct request *rq = tio->orig;
2015 struct mapped_device *md = tio->md;
2017 if (map_request(tio, rq, md) == DM_MAPIO_REQUEUE)
2018 dm_requeue_original_request(md, rq);
2021 static void dm_start_request(struct mapped_device *md, struct request *orig)
2023 if (!orig->q->mq_ops)
2024 blk_start_request(orig);
2026 blk_mq_start_request(orig);
2027 atomic_inc(&md->pending[rq_data_dir(orig)]);
2029 if (md->seq_rq_merge_deadline_usecs) {
2030 md->last_rq_pos = rq_end_sector(orig);
2031 md->last_rq_rw = rq_data_dir(orig);
2032 md->last_rq_start_time = ktime_get();
2035 if (unlikely(dm_stats_used(&md->stats))) {
2036 struct dm_rq_target_io *tio = tio_from_request(orig);
2037 tio->duration_jiffies = jiffies;
2038 tio->n_sectors = blk_rq_sectors(orig);
2039 dm_stats_account_io(&md->stats, orig->cmd_flags, blk_rq_pos(orig),
2040 tio->n_sectors, false, 0, &tio->stats_aux);
2044 * Hold the md reference here for the in-flight I/O.
2045 * We can't rely on the reference count by device opener,
2046 * because the device may be closed during the request completion
2047 * when all bios are completed.
2048 * See the comment in rq_completed() too.
2053 #define MAX_SEQ_RQ_MERGE_DEADLINE_USECS 100000
2055 ssize_t dm_attr_rq_based_seq_io_merge_deadline_show(struct mapped_device *md, char *buf)
2057 return sprintf(buf, "%u\n", md->seq_rq_merge_deadline_usecs);
2060 ssize_t dm_attr_rq_based_seq_io_merge_deadline_store(struct mapped_device *md,
2061 const char *buf, size_t count)
2065 if (!dm_request_based(md) || md->use_blk_mq)
2068 if (kstrtouint(buf, 10, &deadline))
2071 if (deadline > MAX_SEQ_RQ_MERGE_DEADLINE_USECS)
2072 deadline = MAX_SEQ_RQ_MERGE_DEADLINE_USECS;
2074 md->seq_rq_merge_deadline_usecs = deadline;
2079 static bool dm_request_peeked_before_merge_deadline(struct mapped_device *md)
2081 ktime_t kt_deadline;
2083 if (!md->seq_rq_merge_deadline_usecs)
2086 kt_deadline = ns_to_ktime((u64)md->seq_rq_merge_deadline_usecs * NSEC_PER_USEC);
2087 kt_deadline = ktime_add_safe(md->last_rq_start_time, kt_deadline);
2089 return !ktime_after(ktime_get(), kt_deadline);
2093 * q->request_fn for request-based dm.
2094 * Called with the queue lock held.
2096 static void dm_request_fn(struct request_queue *q)
2098 struct mapped_device *md = q->queuedata;
2099 struct dm_target *ti = md->immutable_target;
2101 struct dm_rq_target_io *tio;
2104 if (unlikely(!ti)) {
2106 struct dm_table *map = dm_get_live_table(md, &srcu_idx);
2108 ti = dm_table_find_target(map, pos);
2109 dm_put_live_table(md, srcu_idx);
2113 * For suspend, check blk_queue_stopped() and increment
2114 * ->pending within a single queue_lock not to increment the
2115 * number of in-flight I/Os after the queue is stopped in
2118 while (!blk_queue_stopped(q)) {
2119 rq = blk_peek_request(q);
2123 /* always use block 0 to find the target for flushes for now */
2125 if (!(rq->cmd_flags & REQ_FLUSH))
2126 pos = blk_rq_pos(rq);
2128 if ((dm_request_peeked_before_merge_deadline(md) &&
2129 md_in_flight(md) && rq->bio && rq->bio->bi_vcnt == 1 &&
2130 md->last_rq_pos == pos && md->last_rq_rw == rq_data_dir(rq)) ||
2131 (ti->type->busy && ti->type->busy(ti))) {
2132 blk_delay_queue(q, HZ / 100);
2136 dm_start_request(md, rq);
2138 tio = tio_from_request(rq);
2139 /* Establish tio->ti before queuing work (map_tio_request) */
2141 queue_kthread_work(&md->kworker, &tio->work);
2142 BUG_ON(!irqs_disabled());
2146 static int dm_any_congested(void *congested_data, int bdi_bits)
2149 struct mapped_device *md = congested_data;
2150 struct dm_table *map;
2152 if (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
2153 if (dm_request_based(md)) {
2155 * With request-based DM we only need to check the
2156 * top-level queue for congestion.
2158 r = md->queue->backing_dev_info.wb.state & bdi_bits;
2160 map = dm_get_live_table_fast(md);
2162 r = dm_table_any_congested(map, bdi_bits);
2163 dm_put_live_table_fast(md);
2170 /*-----------------------------------------------------------------
2171 * An IDR is used to keep track of allocated minor numbers.
2172 *---------------------------------------------------------------*/
2173 static void free_minor(int minor)
2175 spin_lock(&_minor_lock);
2176 idr_remove(&_minor_idr, minor);
2177 spin_unlock(&_minor_lock);
2181 * See if the device with a specific minor # is free.
2183 static int specific_minor(int minor)
2187 if (minor >= (1 << MINORBITS))
2190 idr_preload(GFP_KERNEL);
2191 spin_lock(&_minor_lock);
2193 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, minor, minor + 1, GFP_NOWAIT);
2195 spin_unlock(&_minor_lock);
2198 return r == -ENOSPC ? -EBUSY : r;
2202 static int next_free_minor(int *minor)
2206 idr_preload(GFP_KERNEL);
2207 spin_lock(&_minor_lock);
2209 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, 0, 1 << MINORBITS, GFP_NOWAIT);
2211 spin_unlock(&_minor_lock);
2219 static const struct block_device_operations dm_blk_dops;
2221 static void dm_wq_work(struct work_struct *work);
2223 static void dm_init_md_queue(struct mapped_device *md)
2226 * Request-based dm devices cannot be stacked on top of bio-based dm
2227 * devices. The type of this dm device may not have been decided yet.
2228 * The type is decided at the first table loading time.
2229 * To prevent problematic device stacking, clear the queue flag
2230 * for request stacking support until then.
2232 * This queue is new, so no concurrency on the queue_flags.
2234 queue_flag_clear_unlocked(QUEUE_FLAG_STACKABLE, md->queue);
2237 * Initialize data that will only be used by a non-blk-mq DM queue
2238 * - must do so here (in alloc_dev callchain) before queue is used
2240 md->queue->queuedata = md;
2241 md->queue->backing_dev_info.congested_data = md;
2244 static void dm_init_old_md_queue(struct mapped_device *md)
2246 md->use_blk_mq = false;
2247 dm_init_md_queue(md);
2250 * Initialize aspects of queue that aren't relevant for blk-mq
2252 md->queue->backing_dev_info.congested_fn = dm_any_congested;
2253 blk_queue_bounce_limit(md->queue, BLK_BOUNCE_ANY);
2256 static void cleanup_mapped_device(struct mapped_device *md)
2259 destroy_workqueue(md->wq);
2260 if (md->kworker_task)
2261 kthread_stop(md->kworker_task);
2262 mempool_destroy(md->io_pool);
2263 mempool_destroy(md->rq_pool);
2265 bioset_free(md->bs);
2267 cleanup_srcu_struct(&md->io_barrier);
2270 spin_lock(&_minor_lock);
2271 md->disk->private_data = NULL;
2272 spin_unlock(&_minor_lock);
2273 del_gendisk(md->disk);
2278 blk_cleanup_queue(md->queue);
2287 * Allocate and initialise a blank device with a given minor.
2289 static struct mapped_device *alloc_dev(int minor)
2292 struct mapped_device *md = kzalloc(sizeof(*md), GFP_KERNEL);
2296 DMWARN("unable to allocate device, out of memory.");
2300 if (!try_module_get(THIS_MODULE))
2301 goto bad_module_get;
2303 /* get a minor number for the dev */
2304 if (minor == DM_ANY_MINOR)
2305 r = next_free_minor(&minor);
2307 r = specific_minor(minor);
2311 r = init_srcu_struct(&md->io_barrier);
2313 goto bad_io_barrier;
2315 md->use_blk_mq = use_blk_mq;
2316 md->init_tio_pdu = false;
2317 md->type = DM_TYPE_NONE;
2318 mutex_init(&md->suspend_lock);
2319 mutex_init(&md->type_lock);
2320 mutex_init(&md->table_devices_lock);
2321 spin_lock_init(&md->deferred_lock);
2322 atomic_set(&md->holders, 1);
2323 atomic_set(&md->open_count, 0);
2324 atomic_set(&md->event_nr, 0);
2325 atomic_set(&md->uevent_seq, 0);
2326 INIT_LIST_HEAD(&md->uevent_list);
2327 INIT_LIST_HEAD(&md->table_devices);
2328 spin_lock_init(&md->uevent_lock);
2330 md->queue = blk_alloc_queue(GFP_KERNEL);
2334 dm_init_md_queue(md);
2336 md->disk = alloc_disk(1);
2340 atomic_set(&md->pending[0], 0);
2341 atomic_set(&md->pending[1], 0);
2342 init_waitqueue_head(&md->wait);
2343 INIT_WORK(&md->work, dm_wq_work);
2344 init_waitqueue_head(&md->eventq);
2345 init_completion(&md->kobj_holder.completion);
2346 md->kworker_task = NULL;
2348 md->disk->major = _major;
2349 md->disk->first_minor = minor;
2350 md->disk->fops = &dm_blk_dops;
2351 md->disk->queue = md->queue;
2352 md->disk->private_data = md;
2353 sprintf(md->disk->disk_name, "dm-%d", minor);
2355 format_dev_t(md->name, MKDEV(_major, minor));
2357 md->wq = alloc_workqueue("kdmflush", WQ_MEM_RECLAIM, 0);
2361 md->bdev = bdget_disk(md->disk, 0);
2365 bio_init(&md->flush_bio);
2366 md->flush_bio.bi_bdev = md->bdev;
2367 md->flush_bio.bi_rw = WRITE_FLUSH;
2369 dm_stats_init(&md->stats);
2371 /* Populate the mapping, nobody knows we exist yet */
2372 spin_lock(&_minor_lock);
2373 old_md = idr_replace(&_minor_idr, md, minor);
2374 spin_unlock(&_minor_lock);
2376 BUG_ON(old_md != MINOR_ALLOCED);
2381 cleanup_mapped_device(md);
2385 module_put(THIS_MODULE);
2391 static void unlock_fs(struct mapped_device *md);
2393 static void free_dev(struct mapped_device *md)
2395 int minor = MINOR(disk_devt(md->disk));
2399 cleanup_mapped_device(md);
2401 blk_mq_free_tag_set(md->tag_set);
2405 free_table_devices(&md->table_devices);
2406 dm_stats_cleanup(&md->stats);
2409 module_put(THIS_MODULE);
2413 static void __bind_mempools(struct mapped_device *md, struct dm_table *t)
2415 struct dm_md_mempools *p = dm_table_get_md_mempools(t);
2418 /* The md already has necessary mempools. */
2419 if (dm_table_get_type(t) == DM_TYPE_BIO_BASED) {
2421 * Reload bioset because front_pad may have changed
2422 * because a different table was loaded.
2424 bioset_free(md->bs);
2429 * There's no need to reload with request-based dm
2430 * because the size of front_pad doesn't change.
2431 * Note for future: If you are to reload bioset,
2432 * prep-ed requests in the queue may refer
2433 * to bio from the old bioset, so you must walk
2434 * through the queue to unprep.
2439 BUG_ON(!p || md->io_pool || md->rq_pool || md->bs);
2441 md->io_pool = p->io_pool;
2443 md->rq_pool = p->rq_pool;
2449 /* mempool bind completed, no longer need any mempools in the table */
2450 dm_table_free_md_mempools(t);
2454 * Bind a table to the device.
2456 static void event_callback(void *context)
2458 unsigned long flags;
2460 struct mapped_device *md = (struct mapped_device *) context;
2462 spin_lock_irqsave(&md->uevent_lock, flags);
2463 list_splice_init(&md->uevent_list, &uevents);
2464 spin_unlock_irqrestore(&md->uevent_lock, flags);
2466 dm_send_uevents(&uevents, &disk_to_dev(md->disk)->kobj);
2468 atomic_inc(&md->event_nr);
2469 wake_up(&md->eventq);
2473 * Protected by md->suspend_lock obtained by dm_swap_table().
2475 static void __set_size(struct mapped_device *md, sector_t size)
2477 set_capacity(md->disk, size);
2479 i_size_write(md->bdev->bd_inode, (loff_t)size << SECTOR_SHIFT);
2483 * Returns old map, which caller must destroy.
2485 static struct dm_table *__bind(struct mapped_device *md, struct dm_table *t,
2486 struct queue_limits *limits)
2488 struct dm_table *old_map;
2489 struct request_queue *q = md->queue;
2492 size = dm_table_get_size(t);
2495 * Wipe any geometry if the size of the table changed.
2497 if (size != dm_get_size(md))
2498 memset(&md->geometry, 0, sizeof(md->geometry));
2500 __set_size(md, size);
2502 dm_table_event_callback(t, event_callback, md);
2505 * The queue hasn't been stopped yet, if the old table type wasn't
2506 * for request-based during suspension. So stop it to prevent
2507 * I/O mapping before resume.
2508 * This must be done before setting the queue restrictions,
2509 * because request-based dm may be run just after the setting.
2511 if (dm_table_request_based(t)) {
2514 * Leverage the fact that request-based DM targets are
2515 * immutable singletons and establish md->immutable_target
2516 * - used to optimize both dm_request_fn and dm_mq_queue_rq
2518 md->immutable_target = dm_table_get_immutable_target(t);
2521 __bind_mempools(md, t);
2523 old_map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2524 rcu_assign_pointer(md->map, t);
2525 md->immutable_target_type = dm_table_get_immutable_target_type(t);
2527 dm_table_set_restrictions(t, q, limits);
2535 * Returns unbound table for the caller to free.
2537 static struct dm_table *__unbind(struct mapped_device *md)
2539 struct dm_table *map = rcu_dereference_protected(md->map, 1);
2544 dm_table_event_callback(map, NULL, NULL);
2545 RCU_INIT_POINTER(md->map, NULL);
2552 * Constructor for a new device.
2554 int dm_create(int minor, struct mapped_device **result)
2556 struct mapped_device *md;
2558 md = alloc_dev(minor);
2569 * Functions to manage md->type.
2570 * All are required to hold md->type_lock.
2572 void dm_lock_md_type(struct mapped_device *md)
2574 mutex_lock(&md->type_lock);
2577 void dm_unlock_md_type(struct mapped_device *md)
2579 mutex_unlock(&md->type_lock);
2582 void dm_set_md_type(struct mapped_device *md, unsigned type)
2584 BUG_ON(!mutex_is_locked(&md->type_lock));
2588 unsigned dm_get_md_type(struct mapped_device *md)
2593 struct target_type *dm_get_immutable_target_type(struct mapped_device *md)
2595 return md->immutable_target_type;
2599 * The queue_limits are only valid as long as you have a reference
2602 struct queue_limits *dm_get_queue_limits(struct mapped_device *md)
2604 BUG_ON(!atomic_read(&md->holders));
2605 return &md->queue->limits;
2607 EXPORT_SYMBOL_GPL(dm_get_queue_limits);
2609 static void init_rq_based_worker_thread(struct mapped_device *md)
2611 /* Initialize the request-based DM worker thread */
2612 init_kthread_worker(&md->kworker);
2613 md->kworker_task = kthread_run(kthread_worker_fn, &md->kworker,
2614 "kdmwork-%s", dm_device_name(md));
2618 * Fully initialize a request-based queue (->elevator, ->request_fn, etc).
2620 static int dm_init_request_based_queue(struct mapped_device *md)
2622 struct request_queue *q = NULL;
2624 /* Fully initialize the queue */
2625 q = blk_init_allocated_queue(md->queue, dm_request_fn, NULL);
2629 /* disable dm_request_fn's merge heuristic by default */
2630 md->seq_rq_merge_deadline_usecs = 0;
2633 dm_init_old_md_queue(md);
2634 blk_queue_softirq_done(md->queue, dm_softirq_done);
2635 blk_queue_prep_rq(md->queue, dm_prep_fn);
2637 init_rq_based_worker_thread(md);
2639 elv_register_queue(md->queue);
2644 static int dm_mq_init_request(void *data, struct request *rq,
2645 unsigned int hctx_idx, unsigned int request_idx,
2646 unsigned int numa_node)
2648 struct mapped_device *md = data;
2649 struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
2652 * Must initialize md member of tio, otherwise it won't
2653 * be available in dm_mq_queue_rq.
2657 if (md->init_tio_pdu) {
2658 /* target-specific per-io data is immediately after the tio */
2659 tio->info.ptr = tio + 1;
2665 static int dm_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
2666 const struct blk_mq_queue_data *bd)
2668 struct request *rq = bd->rq;
2669 struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
2670 struct mapped_device *md = tio->md;
2671 struct dm_target *ti = md->immutable_target;
2673 if (unlikely(!ti)) {
2675 struct dm_table *map = dm_get_live_table(md, &srcu_idx);
2677 ti = dm_table_find_target(map, 0);
2678 dm_put_live_table(md, srcu_idx);
2681 if (ti->type->busy && ti->type->busy(ti))
2682 return BLK_MQ_RQ_QUEUE_BUSY;
2684 dm_start_request(md, rq);
2686 /* Init tio using md established in .init_request */
2687 init_tio(tio, rq, md);
2690 * Establish tio->ti before queuing work (map_tio_request)
2691 * or making direct call to map_request().
2696 * Both the table and md type cannot change after initial table load
2698 if (dm_get_md_type(md) == DM_TYPE_REQUEST_BASED) {
2700 * Clone the request if underlying devices aren't blk-mq
2701 * - clone request is allocated at the end of the pdu
2703 tio->clone = blk_mq_rq_to_pdu(rq) + sizeof(*tio) + ti->per_io_data_size;
2704 (void) clone_rq(rq, md, tio, GFP_ATOMIC);
2705 queue_kthread_work(&md->kworker, &tio->work);
2707 /* Direct call is fine since .queue_rq allows allocations */
2708 if (map_request(tio, rq, md) == DM_MAPIO_REQUEUE) {
2709 /* Undo dm_start_request() before requeuing */
2710 rq_end_stats(md, rq);
2711 rq_completed(md, rq_data_dir(rq), false);
2712 return BLK_MQ_RQ_QUEUE_BUSY;
2716 return BLK_MQ_RQ_QUEUE_OK;
2719 static struct blk_mq_ops dm_mq_ops = {
2720 .queue_rq = dm_mq_queue_rq,
2721 .map_queue = blk_mq_map_queue,
2722 .complete = dm_softirq_done,
2723 .init_request = dm_mq_init_request,
2726 static int dm_init_request_based_blk_mq_queue(struct mapped_device *md,
2727 struct dm_target *immutable_tgt)
2729 unsigned md_type = dm_get_md_type(md);
2730 struct request_queue *q;
2733 md->tag_set = kzalloc(sizeof(struct blk_mq_tag_set), GFP_KERNEL);
2737 md->tag_set->ops = &dm_mq_ops;
2738 md->tag_set->queue_depth = dm_get_blk_mq_queue_depth();
2739 md->tag_set->numa_node = NUMA_NO_NODE;
2740 md->tag_set->flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
2741 md->tag_set->nr_hw_queues = dm_get_blk_mq_nr_hw_queues();
2742 md->tag_set->driver_data = md;
2744 md->tag_set->cmd_size = sizeof(struct dm_rq_target_io);
2745 if (immutable_tgt && immutable_tgt->per_io_data_size) {
2746 /* any target-specific per-io data is immediately after the tio */
2747 md->tag_set->cmd_size += immutable_tgt->per_io_data_size;
2748 md->init_tio_pdu = true;
2750 if (md_type == DM_TYPE_REQUEST_BASED) {
2751 /* put the memory for non-blk-mq clone at the end of the pdu */
2752 md->tag_set->cmd_size += sizeof(struct request);
2755 err = blk_mq_alloc_tag_set(md->tag_set);
2757 goto out_kfree_tag_set;
2759 q = blk_mq_init_allocated_queue(md->tag_set, md->queue);
2765 dm_init_md_queue(md);
2767 /* backfill 'mq' sysfs registration normally done in blk_register_queue */
2768 blk_mq_register_disk(md->disk);
2770 if (md_type == DM_TYPE_REQUEST_BASED)
2771 init_rq_based_worker_thread(md);
2776 blk_mq_free_tag_set(md->tag_set);
2783 static unsigned filter_md_type(unsigned type, struct mapped_device *md)
2785 if (type == DM_TYPE_BIO_BASED)
2788 return !md->use_blk_mq ? DM_TYPE_REQUEST_BASED : DM_TYPE_MQ_REQUEST_BASED;
2792 * Setup the DM device's queue based on md's type
2794 int dm_setup_md_queue(struct mapped_device *md, struct dm_table *t)
2797 unsigned md_type = filter_md_type(dm_get_md_type(md), md);
2800 case DM_TYPE_REQUEST_BASED:
2801 r = dm_init_request_based_queue(md);
2803 DMWARN("Cannot initialize queue for request-based mapped device");
2807 case DM_TYPE_MQ_REQUEST_BASED:
2808 r = dm_init_request_based_blk_mq_queue(md, dm_table_get_immutable_target(t));
2810 DMWARN("Cannot initialize queue for request-based blk-mq mapped device");
2814 case DM_TYPE_BIO_BASED:
2815 dm_init_old_md_queue(md);
2816 blk_queue_make_request(md->queue, dm_make_request);
2818 * DM handles splitting bios as needed. Free the bio_split bioset
2819 * since it won't be used (saves 1 process per bio-based DM device).
2821 bioset_free(md->queue->bio_split);
2822 md->queue->bio_split = NULL;
2829 struct mapped_device *dm_get_md(dev_t dev)
2831 struct mapped_device *md;
2832 unsigned minor = MINOR(dev);
2834 if (MAJOR(dev) != _major || minor >= (1 << MINORBITS))
2837 spin_lock(&_minor_lock);
2839 md = idr_find(&_minor_idr, minor);
2841 if ((md == MINOR_ALLOCED ||
2842 (MINOR(disk_devt(dm_disk(md))) != minor) ||
2843 dm_deleting_md(md) ||
2844 test_bit(DMF_FREEING, &md->flags))) {
2852 spin_unlock(&_minor_lock);
2856 EXPORT_SYMBOL_GPL(dm_get_md);
2858 void *dm_get_mdptr(struct mapped_device *md)
2860 return md->interface_ptr;
2863 void dm_set_mdptr(struct mapped_device *md, void *ptr)
2865 md->interface_ptr = ptr;
2868 void dm_get(struct mapped_device *md)
2870 atomic_inc(&md->holders);
2871 BUG_ON(test_bit(DMF_FREEING, &md->flags));
2874 int dm_hold(struct mapped_device *md)
2876 spin_lock(&_minor_lock);
2877 if (test_bit(DMF_FREEING, &md->flags)) {
2878 spin_unlock(&_minor_lock);
2882 spin_unlock(&_minor_lock);
2885 EXPORT_SYMBOL_GPL(dm_hold);
2887 const char *dm_device_name(struct mapped_device *md)
2891 EXPORT_SYMBOL_GPL(dm_device_name);
2893 static void __dm_destroy(struct mapped_device *md, bool wait)
2895 struct dm_table *map;
2900 spin_lock(&_minor_lock);
2901 idr_replace(&_minor_idr, MINOR_ALLOCED, MINOR(disk_devt(dm_disk(md))));
2902 set_bit(DMF_FREEING, &md->flags);
2903 spin_unlock(&_minor_lock);
2905 if (dm_request_based(md) && md->kworker_task)
2906 flush_kthread_worker(&md->kworker);
2909 * Take suspend_lock so that presuspend and postsuspend methods
2910 * do not race with internal suspend.
2912 mutex_lock(&md->suspend_lock);
2913 map = dm_get_live_table(md, &srcu_idx);
2914 if (!dm_suspended_md(md)) {
2915 dm_table_presuspend_targets(map);
2916 dm_table_postsuspend_targets(map);
2918 /* dm_put_live_table must be before msleep, otherwise deadlock is possible */
2919 dm_put_live_table(md, srcu_idx);
2920 mutex_unlock(&md->suspend_lock);
2923 * Rare, but there may be I/O requests still going to complete,
2924 * for example. Wait for all references to disappear.
2925 * No one should increment the reference count of the mapped_device,
2926 * after the mapped_device state becomes DMF_FREEING.
2929 while (atomic_read(&md->holders))
2931 else if (atomic_read(&md->holders))
2932 DMWARN("%s: Forcibly removing mapped_device still in use! (%d users)",
2933 dm_device_name(md), atomic_read(&md->holders));
2936 dm_table_destroy(__unbind(md));
2940 void dm_destroy(struct mapped_device *md)
2942 __dm_destroy(md, true);
2945 void dm_destroy_immediate(struct mapped_device *md)
2947 __dm_destroy(md, false);
2950 void dm_put(struct mapped_device *md)
2952 atomic_dec(&md->holders);
2954 EXPORT_SYMBOL_GPL(dm_put);
2956 static int dm_wait_for_completion(struct mapped_device *md, int interruptible)
2959 DECLARE_WAITQUEUE(wait, current);
2961 add_wait_queue(&md->wait, &wait);
2964 set_current_state(interruptible);
2966 if (!md_in_flight(md))
2969 if (interruptible == TASK_INTERRUPTIBLE &&
2970 signal_pending(current)) {
2977 set_current_state(TASK_RUNNING);
2979 remove_wait_queue(&md->wait, &wait);
2985 * Process the deferred bios
2987 static void dm_wq_work(struct work_struct *work)
2989 struct mapped_device *md = container_of(work, struct mapped_device,
2993 struct dm_table *map;
2995 map = dm_get_live_table(md, &srcu_idx);
2997 while (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
2998 spin_lock_irq(&md->deferred_lock);
2999 c = bio_list_pop(&md->deferred);
3000 spin_unlock_irq(&md->deferred_lock);
3005 if (dm_request_based(md))
3006 generic_make_request(c);
3008 __split_and_process_bio(md, map, c);
3011 dm_put_live_table(md, srcu_idx);
3014 static void dm_queue_flush(struct mapped_device *md)
3016 clear_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
3017 smp_mb__after_atomic();
3018 queue_work(md->wq, &md->work);
3022 * Swap in a new table, returning the old one for the caller to destroy.
3024 struct dm_table *dm_swap_table(struct mapped_device *md, struct dm_table *table)
3026 struct dm_table *live_map = NULL, *map = ERR_PTR(-EINVAL);
3027 struct queue_limits limits;
3030 mutex_lock(&md->suspend_lock);
3032 /* device must be suspended */
3033 if (!dm_suspended_md(md))
3037 * If the new table has no data devices, retain the existing limits.
3038 * This helps multipath with queue_if_no_path if all paths disappear,
3039 * then new I/O is queued based on these limits, and then some paths
3042 if (dm_table_has_no_data_devices(table)) {
3043 live_map = dm_get_live_table_fast(md);
3045 limits = md->queue->limits;
3046 dm_put_live_table_fast(md);
3050 r = dm_calculate_queue_limits(table, &limits);
3057 map = __bind(md, table, &limits);
3060 mutex_unlock(&md->suspend_lock);
3065 * Functions to lock and unlock any filesystem running on the
3068 static int lock_fs(struct mapped_device *md)
3072 WARN_ON(md->frozen_sb);
3074 md->frozen_sb = freeze_bdev(md->bdev);
3075 if (IS_ERR(md->frozen_sb)) {
3076 r = PTR_ERR(md->frozen_sb);
3077 md->frozen_sb = NULL;
3081 set_bit(DMF_FROZEN, &md->flags);
3086 static void unlock_fs(struct mapped_device *md)
3088 if (!test_bit(DMF_FROZEN, &md->flags))
3091 thaw_bdev(md->bdev, md->frozen_sb);
3092 md->frozen_sb = NULL;
3093 clear_bit(DMF_FROZEN, &md->flags);
3097 * If __dm_suspend returns 0, the device is completely quiescent
3098 * now. There is no request-processing activity. All new requests
3099 * are being added to md->deferred list.
3101 * Caller must hold md->suspend_lock
3103 static int __dm_suspend(struct mapped_device *md, struct dm_table *map,
3104 unsigned suspend_flags, int interruptible)
3106 bool do_lockfs = suspend_flags & DM_SUSPEND_LOCKFS_FLAG;
3107 bool noflush = suspend_flags & DM_SUSPEND_NOFLUSH_FLAG;
3111 * DMF_NOFLUSH_SUSPENDING must be set before presuspend.
3112 * This flag is cleared before dm_suspend returns.
3115 set_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
3118 * This gets reverted if there's an error later and the targets
3119 * provide the .presuspend_undo hook.
3121 dm_table_presuspend_targets(map);
3124 * Flush I/O to the device.
3125 * Any I/O submitted after lock_fs() may not be flushed.
3126 * noflush takes precedence over do_lockfs.
3127 * (lock_fs() flushes I/Os and waits for them to complete.)
3129 if (!noflush && do_lockfs) {
3132 dm_table_presuspend_undo_targets(map);
3138 * Here we must make sure that no processes are submitting requests
3139 * to target drivers i.e. no one may be executing
3140 * __split_and_process_bio. This is called from dm_request and
3143 * To get all processes out of __split_and_process_bio in dm_request,
3144 * we take the write lock. To prevent any process from reentering
3145 * __split_and_process_bio from dm_request and quiesce the thread
3146 * (dm_wq_work), we set BMF_BLOCK_IO_FOR_SUSPEND and call
3147 * flush_workqueue(md->wq).
3149 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
3151 synchronize_srcu(&md->io_barrier);
3154 * Stop md->queue before flushing md->wq in case request-based
3155 * dm defers requests to md->wq from md->queue.
3157 if (dm_request_based(md)) {
3158 stop_queue(md->queue);
3159 if (md->kworker_task)
3160 flush_kthread_worker(&md->kworker);
3163 flush_workqueue(md->wq);
3166 * At this point no more requests are entering target request routines.
3167 * We call dm_wait_for_completion to wait for all existing requests
3170 r = dm_wait_for_completion(md, interruptible);
3173 clear_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
3175 synchronize_srcu(&md->io_barrier);
3177 /* were we interrupted ? */
3181 if (dm_request_based(md))
3182 start_queue(md->queue);
3185 dm_table_presuspend_undo_targets(map);
3186 /* pushback list is already flushed, so skip flush */
3193 * We need to be able to change a mapping table under a mounted
3194 * filesystem. For example we might want to move some data in
3195 * the background. Before the table can be swapped with
3196 * dm_bind_table, dm_suspend must be called to flush any in
3197 * flight bios and ensure that any further io gets deferred.
3200 * Suspend mechanism in request-based dm.
3202 * 1. Flush all I/Os by lock_fs() if needed.
3203 * 2. Stop dispatching any I/O by stopping the request_queue.
3204 * 3. Wait for all in-flight I/Os to be completed or requeued.
3206 * To abort suspend, start the request_queue.
3208 int dm_suspend(struct mapped_device *md, unsigned suspend_flags)
3210 struct dm_table *map = NULL;
3214 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
3216 if (dm_suspended_md(md)) {
3221 if (dm_suspended_internally_md(md)) {
3222 /* already internally suspended, wait for internal resume */
3223 mutex_unlock(&md->suspend_lock);
3224 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
3230 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3232 r = __dm_suspend(md, map, suspend_flags, TASK_INTERRUPTIBLE);
3236 set_bit(DMF_SUSPENDED, &md->flags);
3238 dm_table_postsuspend_targets(map);
3241 mutex_unlock(&md->suspend_lock);
3245 static int __dm_resume(struct mapped_device *md, struct dm_table *map)
3248 int r = dm_table_resume_targets(map);
3256 * Flushing deferred I/Os must be done after targets are resumed
3257 * so that mapping of targets can work correctly.
3258 * Request-based dm is queueing the deferred I/Os in its request_queue.
3260 if (dm_request_based(md))
3261 start_queue(md->queue);
3268 int dm_resume(struct mapped_device *md)
3271 struct dm_table *map = NULL;
3274 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
3276 if (!dm_suspended_md(md))
3279 if (dm_suspended_internally_md(md)) {
3280 /* already internally suspended, wait for internal resume */
3281 mutex_unlock(&md->suspend_lock);
3282 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
3288 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3289 if (!map || !dm_table_get_size(map))
3292 r = __dm_resume(md, map);
3296 clear_bit(DMF_SUSPENDED, &md->flags);
3300 mutex_unlock(&md->suspend_lock);
3306 * Internal suspend/resume works like userspace-driven suspend. It waits
3307 * until all bios finish and prevents issuing new bios to the target drivers.
3308 * It may be used only from the kernel.
3311 static void __dm_internal_suspend(struct mapped_device *md, unsigned suspend_flags)
3313 struct dm_table *map = NULL;
3315 if (md->internal_suspend_count++)
3316 return; /* nested internal suspend */
3318 if (dm_suspended_md(md)) {
3319 set_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3320 return; /* nest suspend */
3323 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3326 * Using TASK_UNINTERRUPTIBLE because only NOFLUSH internal suspend is
3327 * supported. Properly supporting a TASK_INTERRUPTIBLE internal suspend
3328 * would require changing .presuspend to return an error -- avoid this
3329 * until there is a need for more elaborate variants of internal suspend.
3331 (void) __dm_suspend(md, map, suspend_flags, TASK_UNINTERRUPTIBLE);
3333 set_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3335 dm_table_postsuspend_targets(map);
3338 static void __dm_internal_resume(struct mapped_device *md)
3340 BUG_ON(!md->internal_suspend_count);
3342 if (--md->internal_suspend_count)
3343 return; /* resume from nested internal suspend */
3345 if (dm_suspended_md(md))
3346 goto done; /* resume from nested suspend */
3349 * NOTE: existing callers don't need to call dm_table_resume_targets
3350 * (which may fail -- so best to avoid it for now by passing NULL map)
3352 (void) __dm_resume(md, NULL);
3355 clear_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3356 smp_mb__after_atomic();
3357 wake_up_bit(&md->flags, DMF_SUSPENDED_INTERNALLY);
3360 void dm_internal_suspend_noflush(struct mapped_device *md)
3362 mutex_lock(&md->suspend_lock);
3363 __dm_internal_suspend(md, DM_SUSPEND_NOFLUSH_FLAG);
3364 mutex_unlock(&md->suspend_lock);
3366 EXPORT_SYMBOL_GPL(dm_internal_suspend_noflush);
3368 void dm_internal_resume(struct mapped_device *md)
3370 mutex_lock(&md->suspend_lock);
3371 __dm_internal_resume(md);
3372 mutex_unlock(&md->suspend_lock);
3374 EXPORT_SYMBOL_GPL(dm_internal_resume);
3377 * Fast variants of internal suspend/resume hold md->suspend_lock,
3378 * which prevents interaction with userspace-driven suspend.
3381 void dm_internal_suspend_fast(struct mapped_device *md)
3383 mutex_lock(&md->suspend_lock);
3384 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
3387 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
3388 synchronize_srcu(&md->io_barrier);
3389 flush_workqueue(md->wq);
3390 dm_wait_for_completion(md, TASK_UNINTERRUPTIBLE);
3392 EXPORT_SYMBOL_GPL(dm_internal_suspend_fast);
3394 void dm_internal_resume_fast(struct mapped_device *md)
3396 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
3402 mutex_unlock(&md->suspend_lock);
3404 EXPORT_SYMBOL_GPL(dm_internal_resume_fast);
3406 /*-----------------------------------------------------------------
3407 * Event notification.
3408 *---------------------------------------------------------------*/
3409 int dm_kobject_uevent(struct mapped_device *md, enum kobject_action action,
3412 char udev_cookie[DM_COOKIE_LENGTH];
3413 char *envp[] = { udev_cookie, NULL };
3416 return kobject_uevent(&disk_to_dev(md->disk)->kobj, action);
3418 snprintf(udev_cookie, DM_COOKIE_LENGTH, "%s=%u",
3419 DM_COOKIE_ENV_VAR_NAME, cookie);
3420 return kobject_uevent_env(&disk_to_dev(md->disk)->kobj,
3425 uint32_t dm_next_uevent_seq(struct mapped_device *md)
3427 return atomic_add_return(1, &md->uevent_seq);
3430 uint32_t dm_get_event_nr(struct mapped_device *md)
3432 return atomic_read(&md->event_nr);
3435 int dm_wait_event(struct mapped_device *md, int event_nr)
3437 return wait_event_interruptible(md->eventq,
3438 (event_nr != atomic_read(&md->event_nr)));
3441 void dm_uevent_add(struct mapped_device *md, struct list_head *elist)
3443 unsigned long flags;
3445 spin_lock_irqsave(&md->uevent_lock, flags);
3446 list_add(elist, &md->uevent_list);
3447 spin_unlock_irqrestore(&md->uevent_lock, flags);
3451 * The gendisk is only valid as long as you have a reference
3454 struct gendisk *dm_disk(struct mapped_device *md)
3458 EXPORT_SYMBOL_GPL(dm_disk);
3460 struct kobject *dm_kobject(struct mapped_device *md)
3462 return &md->kobj_holder.kobj;
3465 struct mapped_device *dm_get_from_kobject(struct kobject *kobj)
3467 struct mapped_device *md;
3469 md = container_of(kobj, struct mapped_device, kobj_holder.kobj);
3471 if (test_bit(DMF_FREEING, &md->flags) ||
3479 int dm_suspended_md(struct mapped_device *md)
3481 return test_bit(DMF_SUSPENDED, &md->flags);
3484 int dm_suspended_internally_md(struct mapped_device *md)
3486 return test_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3489 int dm_test_deferred_remove_flag(struct mapped_device *md)
3491 return test_bit(DMF_DEFERRED_REMOVE, &md->flags);
3494 int dm_suspended(struct dm_target *ti)
3496 return dm_suspended_md(dm_table_get_md(ti->table));
3498 EXPORT_SYMBOL_GPL(dm_suspended);
3500 int dm_noflush_suspending(struct dm_target *ti)
3502 return __noflush_suspending(dm_table_get_md(ti->table));
3504 EXPORT_SYMBOL_GPL(dm_noflush_suspending);
3506 struct dm_md_mempools *dm_alloc_md_mempools(struct mapped_device *md, unsigned type,
3507 unsigned integrity, unsigned per_io_data_size)
3509 struct dm_md_mempools *pools = kzalloc(sizeof(*pools), GFP_KERNEL);
3510 struct kmem_cache *cachep = NULL;
3511 unsigned int pool_size = 0;
3512 unsigned int front_pad;
3517 type = filter_md_type(type, md);
3520 case DM_TYPE_BIO_BASED:
3522 pool_size = dm_get_reserved_bio_based_ios();
3523 front_pad = roundup(per_io_data_size, __alignof__(struct dm_target_io)) + offsetof(struct dm_target_io, clone);
3525 case DM_TYPE_REQUEST_BASED:
3526 cachep = _rq_tio_cache;
3527 pool_size = dm_get_reserved_rq_based_ios();
3528 pools->rq_pool = mempool_create_slab_pool(pool_size, _rq_cache);
3529 if (!pools->rq_pool)
3531 /* fall through to setup remaining rq-based pools */
3532 case DM_TYPE_MQ_REQUEST_BASED:
3534 pool_size = dm_get_reserved_rq_based_ios();
3535 front_pad = offsetof(struct dm_rq_clone_bio_info, clone);
3536 /* per_io_data_size is used for blk-mq pdu at queue allocation */
3543 pools->io_pool = mempool_create_slab_pool(pool_size, cachep);
3544 if (!pools->io_pool)
3548 pools->bs = bioset_create_nobvec(pool_size, front_pad);
3552 if (integrity && bioset_integrity_create(pools->bs, pool_size))
3558 dm_free_md_mempools(pools);
3563 void dm_free_md_mempools(struct dm_md_mempools *pools)
3568 mempool_destroy(pools->io_pool);
3569 mempool_destroy(pools->rq_pool);
3572 bioset_free(pools->bs);
3577 static int dm_pr_register(struct block_device *bdev, u64 old_key, u64 new_key,
3580 struct mapped_device *md = bdev->bd_disk->private_data;
3581 const struct pr_ops *ops;
3582 struct dm_target *tgt;
3586 r = dm_get_live_table_for_ioctl(md, &tgt, &bdev, &mode, &srcu_idx);
3590 ops = bdev->bd_disk->fops->pr_ops;
3591 if (ops && ops->pr_register)
3592 r = ops->pr_register(bdev, old_key, new_key, flags);
3596 dm_put_live_table(md, srcu_idx);
3600 static int dm_pr_reserve(struct block_device *bdev, u64 key, enum pr_type type,
3603 struct mapped_device *md = bdev->bd_disk->private_data;
3604 const struct pr_ops *ops;
3605 struct dm_target *tgt;
3609 r = dm_get_live_table_for_ioctl(md, &tgt, &bdev, &mode, &srcu_idx);
3613 ops = bdev->bd_disk->fops->pr_ops;
3614 if (ops && ops->pr_reserve)
3615 r = ops->pr_reserve(bdev, key, type, flags);
3619 dm_put_live_table(md, srcu_idx);
3623 static int dm_pr_release(struct block_device *bdev, u64 key, enum pr_type type)
3625 struct mapped_device *md = bdev->bd_disk->private_data;
3626 const struct pr_ops *ops;
3627 struct dm_target *tgt;
3631 r = dm_get_live_table_for_ioctl(md, &tgt, &bdev, &mode, &srcu_idx);
3635 ops = bdev->bd_disk->fops->pr_ops;
3636 if (ops && ops->pr_release)
3637 r = ops->pr_release(bdev, key, type);
3641 dm_put_live_table(md, srcu_idx);
3645 static int dm_pr_preempt(struct block_device *bdev, u64 old_key, u64 new_key,
3646 enum pr_type type, bool abort)
3648 struct mapped_device *md = bdev->bd_disk->private_data;
3649 const struct pr_ops *ops;
3650 struct dm_target *tgt;
3654 r = dm_get_live_table_for_ioctl(md, &tgt, &bdev, &mode, &srcu_idx);
3658 ops = bdev->bd_disk->fops->pr_ops;
3659 if (ops && ops->pr_preempt)
3660 r = ops->pr_preempt(bdev, old_key, new_key, type, abort);
3664 dm_put_live_table(md, srcu_idx);
3668 static int dm_pr_clear(struct block_device *bdev, u64 key)
3670 struct mapped_device *md = bdev->bd_disk->private_data;
3671 const struct pr_ops *ops;
3672 struct dm_target *tgt;
3676 r = dm_get_live_table_for_ioctl(md, &tgt, &bdev, &mode, &srcu_idx);
3680 ops = bdev->bd_disk->fops->pr_ops;
3681 if (ops && ops->pr_clear)
3682 r = ops->pr_clear(bdev, key);
3686 dm_put_live_table(md, srcu_idx);
3690 static const struct pr_ops dm_pr_ops = {
3691 .pr_register = dm_pr_register,
3692 .pr_reserve = dm_pr_reserve,
3693 .pr_release = dm_pr_release,
3694 .pr_preempt = dm_pr_preempt,
3695 .pr_clear = dm_pr_clear,
3698 static const struct block_device_operations dm_blk_dops = {
3699 .open = dm_blk_open,
3700 .release = dm_blk_close,
3701 .ioctl = dm_blk_ioctl,
3702 .getgeo = dm_blk_getgeo,
3703 .pr_ops = &dm_pr_ops,
3704 .owner = THIS_MODULE
3710 module_init(dm_init);
3711 module_exit(dm_exit);
3713 module_param(major, uint, 0);
3714 MODULE_PARM_DESC(major, "The major number of the device mapper");
3716 module_param(reserved_bio_based_ios, uint, S_IRUGO | S_IWUSR);
3717 MODULE_PARM_DESC(reserved_bio_based_ios, "Reserved IOs in bio-based mempools");
3719 module_param(reserved_rq_based_ios, uint, S_IRUGO | S_IWUSR);
3720 MODULE_PARM_DESC(reserved_rq_based_ios, "Reserved IOs in request-based mempools");
3722 module_param(use_blk_mq, bool, S_IRUGO | S_IWUSR);
3723 MODULE_PARM_DESC(use_blk_mq, "Use block multiqueue for request-based DM devices");
3725 module_param(blk_mq_nr_hw_queues, uint, S_IRUGO | S_IWUSR);
3726 MODULE_PARM_DESC(blk_mq_nr_hw_queues, "Number of hardware queues for blk-mq request-based DM devices");
3728 module_param(blk_mq_queue_depth, uint, S_IRUGO | S_IWUSR);
3729 MODULE_PARM_DESC(blk_mq_queue_depth, "Queue depth for blk-mq request-based DM devices");
3731 MODULE_DESCRIPTION(DM_NAME " driver");
3732 MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
3733 MODULE_LICENSE("GPL");