1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 2001, 2002 Sistina Software (UK) Limited.
4 * Copyright (C) 2004-2008 Red Hat, Inc. All rights reserved.
6 * This file is released under the GPL.
11 #include "dm-uevent.h"
14 #include <linux/init.h>
15 #include <linux/module.h>
16 #include <linux/mutex.h>
17 #include <linux/sched/mm.h>
18 #include <linux/sched/signal.h>
19 #include <linux/blkpg.h>
20 #include <linux/bio.h>
21 #include <linux/mempool.h>
22 #include <linux/dax.h>
23 #include <linux/slab.h>
24 #include <linux/idr.h>
25 #include <linux/uio.h>
26 #include <linux/hdreg.h>
27 #include <linux/delay.h>
28 #include <linux/wait.h>
30 #include <linux/refcount.h>
31 #include <linux/part_stat.h>
32 #include <linux/blk-crypto.h>
33 #include <linux/blk-crypto-profile.h>
35 #define DM_MSG_PREFIX "core"
38 * Cookies are numeric values sent with CHANGE and REMOVE
39 * uevents while resuming, removing or renaming the device.
41 #define DM_COOKIE_ENV_VAR_NAME "DM_COOKIE"
42 #define DM_COOKIE_LENGTH 24
45 * For REQ_POLLED fs bio, this flag is set if we link mapped underlying
46 * dm_io into one list, and reuse bio->bi_private as the list head. Before
47 * ending this fs bio, we will recover its ->bi_private.
49 #define REQ_DM_POLL_LIST REQ_DRV
51 static const char *_name = DM_NAME;
53 static unsigned int major;
54 static unsigned int _major;
56 static DEFINE_IDR(_minor_idr);
58 static DEFINE_SPINLOCK(_minor_lock);
60 static void do_deferred_remove(struct work_struct *w);
62 static DECLARE_WORK(deferred_remove_work, do_deferred_remove);
64 static struct workqueue_struct *deferred_remove_workqueue;
66 atomic_t dm_global_event_nr = ATOMIC_INIT(0);
67 DECLARE_WAIT_QUEUE_HEAD(dm_global_eventq);
69 void dm_issue_global_event(void)
71 atomic_inc(&dm_global_event_nr);
72 wake_up(&dm_global_eventq);
75 DEFINE_STATIC_KEY_FALSE(stats_enabled);
76 DEFINE_STATIC_KEY_FALSE(swap_bios_enabled);
77 DEFINE_STATIC_KEY_FALSE(zoned_enabled);
80 * One of these is allocated (on-stack) per original bio.
87 unsigned int sector_count;
88 bool is_abnormal_io:1;
89 bool submit_as_polled:1;
92 static inline struct dm_target_io *clone_to_tio(struct bio *clone)
94 return container_of(clone, struct dm_target_io, clone);
97 void *dm_per_bio_data(struct bio *bio, size_t data_size)
99 if (!dm_tio_flagged(clone_to_tio(bio), DM_TIO_INSIDE_DM_IO))
100 return (char *)bio - DM_TARGET_IO_BIO_OFFSET - data_size;
101 return (char *)bio - DM_IO_BIO_OFFSET - data_size;
103 EXPORT_SYMBOL_GPL(dm_per_bio_data);
105 struct bio *dm_bio_from_per_bio_data(void *data, size_t data_size)
107 struct dm_io *io = (struct dm_io *)((char *)data + data_size);
109 if (io->magic == DM_IO_MAGIC)
110 return (struct bio *)((char *)io + DM_IO_BIO_OFFSET);
111 BUG_ON(io->magic != DM_TIO_MAGIC);
112 return (struct bio *)((char *)io + DM_TARGET_IO_BIO_OFFSET);
114 EXPORT_SYMBOL_GPL(dm_bio_from_per_bio_data);
116 unsigned int dm_bio_get_target_bio_nr(const struct bio *bio)
118 return container_of(bio, struct dm_target_io, clone)->target_bio_nr;
120 EXPORT_SYMBOL_GPL(dm_bio_get_target_bio_nr);
122 #define MINOR_ALLOCED ((void *)-1)
124 #define DM_NUMA_NODE NUMA_NO_NODE
125 static int dm_numa_node = DM_NUMA_NODE;
127 #define DEFAULT_SWAP_BIOS (8 * 1048576 / PAGE_SIZE)
128 static int swap_bios = DEFAULT_SWAP_BIOS;
129 static int get_swap_bios(void)
131 int latch = READ_ONCE(swap_bios);
133 if (unlikely(latch <= 0))
134 latch = DEFAULT_SWAP_BIOS;
138 struct table_device {
139 struct list_head list;
141 struct dm_dev dm_dev;
145 * Bio-based DM's mempools' reserved IOs set by the user.
147 #define RESERVED_BIO_BASED_IOS 16
148 static unsigned int reserved_bio_based_ios = RESERVED_BIO_BASED_IOS;
150 static int __dm_get_module_param_int(int *module_param, int min, int max)
152 int param = READ_ONCE(*module_param);
153 int modified_param = 0;
154 bool modified = true;
157 modified_param = min;
158 else if (param > max)
159 modified_param = max;
164 (void)cmpxchg(module_param, param, modified_param);
165 param = modified_param;
171 unsigned int __dm_get_module_param(unsigned int *module_param, unsigned int def, unsigned int max)
173 unsigned int param = READ_ONCE(*module_param);
174 unsigned int modified_param = 0;
177 modified_param = def;
178 else if (param > max)
179 modified_param = max;
181 if (modified_param) {
182 (void)cmpxchg(module_param, param, modified_param);
183 param = modified_param;
189 unsigned int dm_get_reserved_bio_based_ios(void)
191 return __dm_get_module_param(&reserved_bio_based_ios,
192 RESERVED_BIO_BASED_IOS, DM_RESERVED_MAX_IOS);
194 EXPORT_SYMBOL_GPL(dm_get_reserved_bio_based_ios);
196 static unsigned int dm_get_numa_node(void)
198 return __dm_get_module_param_int(&dm_numa_node,
199 DM_NUMA_NODE, num_online_nodes() - 1);
202 static int __init local_init(void)
206 r = dm_uevent_init();
210 deferred_remove_workqueue = alloc_ordered_workqueue("kdmremove", 0);
211 if (!deferred_remove_workqueue) {
213 goto out_uevent_exit;
217 r = register_blkdev(_major, _name);
219 goto out_free_workqueue;
227 destroy_workqueue(deferred_remove_workqueue);
234 static void local_exit(void)
236 destroy_workqueue(deferred_remove_workqueue);
238 unregister_blkdev(_major, _name);
243 DMINFO("cleaned up");
246 static int (*_inits[])(void) __initdata = {
257 static void (*_exits[])(void) = {
268 static int __init dm_init(void)
270 const int count = ARRAY_SIZE(_inits);
273 #if (IS_ENABLED(CONFIG_IMA) && !IS_ENABLED(CONFIG_IMA_DISABLE_HTABLE))
274 DMWARN("CONFIG_IMA_DISABLE_HTABLE is disabled."
275 " Duplicate IMA measurements will not be recorded in the IMA log.");
278 for (i = 0; i < count; i++) {
292 static void __exit dm_exit(void)
294 int i = ARRAY_SIZE(_exits);
300 * Should be empty by this point.
302 idr_destroy(&_minor_idr);
306 * Block device functions
308 int dm_deleting_md(struct mapped_device *md)
310 return test_bit(DMF_DELETING, &md->flags);
313 static int dm_blk_open(struct gendisk *disk, blk_mode_t mode)
315 struct mapped_device *md;
317 spin_lock(&_minor_lock);
319 md = disk->private_data;
323 if (test_bit(DMF_FREEING, &md->flags) ||
324 dm_deleting_md(md)) {
330 atomic_inc(&md->open_count);
332 spin_unlock(&_minor_lock);
334 return md ? 0 : -ENXIO;
337 static void dm_blk_close(struct gendisk *disk)
339 struct mapped_device *md;
341 spin_lock(&_minor_lock);
343 md = disk->private_data;
347 if (atomic_dec_and_test(&md->open_count) &&
348 (test_bit(DMF_DEFERRED_REMOVE, &md->flags)))
349 queue_work(deferred_remove_workqueue, &deferred_remove_work);
353 spin_unlock(&_minor_lock);
356 int dm_open_count(struct mapped_device *md)
358 return atomic_read(&md->open_count);
362 * Guarantees nothing is using the device before it's deleted.
364 int dm_lock_for_deletion(struct mapped_device *md, bool mark_deferred, bool only_deferred)
368 spin_lock(&_minor_lock);
370 if (dm_open_count(md)) {
373 set_bit(DMF_DEFERRED_REMOVE, &md->flags);
374 } else if (only_deferred && !test_bit(DMF_DEFERRED_REMOVE, &md->flags))
377 set_bit(DMF_DELETING, &md->flags);
379 spin_unlock(&_minor_lock);
384 int dm_cancel_deferred_remove(struct mapped_device *md)
388 spin_lock(&_minor_lock);
390 if (test_bit(DMF_DELETING, &md->flags))
393 clear_bit(DMF_DEFERRED_REMOVE, &md->flags);
395 spin_unlock(&_minor_lock);
400 static void do_deferred_remove(struct work_struct *w)
402 dm_deferred_remove();
405 static int dm_blk_getgeo(struct block_device *bdev, struct hd_geometry *geo)
407 struct mapped_device *md = bdev->bd_disk->private_data;
409 return dm_get_geometry(md, geo);
412 static int dm_prepare_ioctl(struct mapped_device *md, int *srcu_idx,
413 struct block_device **bdev)
415 struct dm_target *ti;
416 struct dm_table *map;
421 map = dm_get_live_table(md, srcu_idx);
422 if (!map || !dm_table_get_size(map))
425 /* We only support devices that have a single target */
426 if (map->num_targets != 1)
429 ti = dm_table_get_target(map, 0);
430 if (!ti->type->prepare_ioctl)
433 if (dm_suspended_md(md))
436 r = ti->type->prepare_ioctl(ti, bdev);
437 if (r == -ENOTCONN && !fatal_signal_pending(current)) {
438 dm_put_live_table(md, *srcu_idx);
446 static void dm_unprepare_ioctl(struct mapped_device *md, int srcu_idx)
448 dm_put_live_table(md, srcu_idx);
451 static int dm_blk_ioctl(struct block_device *bdev, blk_mode_t mode,
452 unsigned int cmd, unsigned long arg)
454 struct mapped_device *md = bdev->bd_disk->private_data;
457 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
463 * Target determined this ioctl is being issued against a
464 * subset of the parent bdev; require extra privileges.
466 if (!capable(CAP_SYS_RAWIO)) {
468 "%s: sending ioctl %x to DM device without required privilege.",
475 if (!bdev->bd_disk->fops->ioctl)
478 r = bdev->bd_disk->fops->ioctl(bdev, mode, cmd, arg);
480 dm_unprepare_ioctl(md, srcu_idx);
484 u64 dm_start_time_ns_from_clone(struct bio *bio)
486 return jiffies_to_nsecs(clone_to_tio(bio)->io->start_time);
488 EXPORT_SYMBOL_GPL(dm_start_time_ns_from_clone);
490 static inline bool bio_is_flush_with_data(struct bio *bio)
492 return ((bio->bi_opf & REQ_PREFLUSH) && bio->bi_iter.bi_size);
495 static inline unsigned int dm_io_sectors(struct dm_io *io, struct bio *bio)
498 * If REQ_PREFLUSH set, don't account payload, it will be
499 * submitted (and accounted) after this flush completes.
501 if (bio_is_flush_with_data(bio))
503 if (unlikely(dm_io_flagged(io, DM_IO_WAS_SPLIT)))
505 return bio_sectors(bio);
508 static void dm_io_acct(struct dm_io *io, bool end)
510 struct bio *bio = io->orig_bio;
512 if (dm_io_flagged(io, DM_IO_BLK_STAT)) {
514 bdev_start_io_acct(bio->bi_bdev, bio_op(bio),
517 bdev_end_io_acct(bio->bi_bdev, bio_op(bio),
518 dm_io_sectors(io, bio),
522 if (static_branch_unlikely(&stats_enabled) &&
523 unlikely(dm_stats_used(&io->md->stats))) {
526 if (unlikely(dm_io_flagged(io, DM_IO_WAS_SPLIT)))
527 sector = bio_end_sector(bio) - io->sector_offset;
529 sector = bio->bi_iter.bi_sector;
531 dm_stats_account_io(&io->md->stats, bio_data_dir(bio),
532 sector, dm_io_sectors(io, bio),
533 end, io->start_time, &io->stats_aux);
537 static void __dm_start_io_acct(struct dm_io *io)
539 dm_io_acct(io, false);
542 static void dm_start_io_acct(struct dm_io *io, struct bio *clone)
545 * Ensure IO accounting is only ever started once.
547 if (dm_io_flagged(io, DM_IO_ACCOUNTED))
550 /* Expect no possibility for race unless DM_TIO_IS_DUPLICATE_BIO. */
551 if (!clone || likely(dm_tio_is_normal(clone_to_tio(clone)))) {
552 dm_io_set_flag(io, DM_IO_ACCOUNTED);
555 /* Can afford locking given DM_TIO_IS_DUPLICATE_BIO */
556 spin_lock_irqsave(&io->lock, flags);
557 if (dm_io_flagged(io, DM_IO_ACCOUNTED)) {
558 spin_unlock_irqrestore(&io->lock, flags);
561 dm_io_set_flag(io, DM_IO_ACCOUNTED);
562 spin_unlock_irqrestore(&io->lock, flags);
565 __dm_start_io_acct(io);
568 static void dm_end_io_acct(struct dm_io *io)
570 dm_io_acct(io, true);
573 static struct dm_io *alloc_io(struct mapped_device *md, struct bio *bio, gfp_t gfp_mask)
576 struct dm_target_io *tio;
579 clone = bio_alloc_clone(NULL, bio, gfp_mask, &md->mempools->io_bs);
580 if (unlikely(!clone))
582 tio = clone_to_tio(clone);
584 dm_tio_set_flag(tio, DM_TIO_INSIDE_DM_IO);
587 io = container_of(tio, struct dm_io, tio);
588 io->magic = DM_IO_MAGIC;
589 io->status = BLK_STS_OK;
591 /* one ref is for submission, the other is for completion */
592 atomic_set(&io->io_count, 2);
593 this_cpu_inc(*md->pending_io);
596 spin_lock_init(&io->lock);
597 io->start_time = jiffies;
599 if (blk_queue_io_stat(md->queue))
600 dm_io_set_flag(io, DM_IO_BLK_STAT);
602 if (static_branch_unlikely(&stats_enabled) &&
603 unlikely(dm_stats_used(&md->stats)))
604 dm_stats_record_start(&md->stats, &io->stats_aux);
609 static void free_io(struct dm_io *io)
611 bio_put(&io->tio.clone);
614 static struct bio *alloc_tio(struct clone_info *ci, struct dm_target *ti,
615 unsigned int target_bio_nr, unsigned int *len, gfp_t gfp_mask)
617 struct mapped_device *md = ci->io->md;
618 struct dm_target_io *tio;
621 if (!ci->io->tio.io) {
622 /* the dm_target_io embedded in ci->io is available */
624 /* alloc_io() already initialized embedded clone */
627 clone = bio_alloc_clone(NULL, ci->bio, gfp_mask,
632 /* REQ_DM_POLL_LIST shouldn't be inherited */
633 clone->bi_opf &= ~REQ_DM_POLL_LIST;
635 tio = clone_to_tio(clone);
636 tio->flags = 0; /* also clears DM_TIO_INSIDE_DM_IO */
639 tio->magic = DM_TIO_MAGIC;
642 tio->target_bio_nr = target_bio_nr;
646 /* Set default bdev, but target must bio_set_dev() before issuing IO */
647 clone->bi_bdev = md->disk->part0;
648 if (unlikely(ti->needs_bio_set_dev))
649 bio_set_dev(clone, md->disk->part0);
652 clone->bi_iter.bi_size = to_bytes(*len);
653 if (bio_integrity(clone))
654 bio_integrity_trim(clone);
660 static void free_tio(struct bio *clone)
662 if (dm_tio_flagged(clone_to_tio(clone), DM_TIO_INSIDE_DM_IO))
668 * Add the bio to the list of deferred io.
670 static void queue_io(struct mapped_device *md, struct bio *bio)
674 spin_lock_irqsave(&md->deferred_lock, flags);
675 bio_list_add(&md->deferred, bio);
676 spin_unlock_irqrestore(&md->deferred_lock, flags);
677 queue_work(md->wq, &md->work);
681 * Everyone (including functions in this file), should use this
682 * function to access the md->map field, and make sure they call
683 * dm_put_live_table() when finished.
685 struct dm_table *dm_get_live_table(struct mapped_device *md,
686 int *srcu_idx) __acquires(md->io_barrier)
688 *srcu_idx = srcu_read_lock(&md->io_barrier);
690 return srcu_dereference(md->map, &md->io_barrier);
693 void dm_put_live_table(struct mapped_device *md,
694 int srcu_idx) __releases(md->io_barrier)
696 srcu_read_unlock(&md->io_barrier, srcu_idx);
699 void dm_sync_table(struct mapped_device *md)
701 synchronize_srcu(&md->io_barrier);
702 synchronize_rcu_expedited();
706 * A fast alternative to dm_get_live_table/dm_put_live_table.
707 * The caller must not block between these two functions.
709 static struct dm_table *dm_get_live_table_fast(struct mapped_device *md) __acquires(RCU)
712 return rcu_dereference(md->map);
715 static void dm_put_live_table_fast(struct mapped_device *md) __releases(RCU)
720 static char *_dm_claim_ptr = "I belong to device-mapper";
723 * Open a table device so we can use it as a map destination.
725 static struct table_device *open_table_device(struct mapped_device *md,
726 dev_t dev, blk_mode_t mode)
728 struct table_device *td;
729 struct file *bdev_file;
730 struct block_device *bdev;
734 td = kmalloc_node(sizeof(*td), GFP_KERNEL, md->numa_node_id);
736 return ERR_PTR(-ENOMEM);
737 refcount_set(&td->count, 1);
739 bdev_file = bdev_file_open_by_dev(dev, mode, _dm_claim_ptr, NULL);
740 if (IS_ERR(bdev_file)) {
741 r = PTR_ERR(bdev_file);
745 bdev = file_bdev(bdev_file);
748 * We can be called before the dm disk is added. In that case we can't
749 * register the holder relation here. It will be done once add_disk was
752 if (md->disk->slave_dir) {
753 r = bd_link_disk_holder(bdev, md->disk);
758 td->dm_dev.mode = mode;
759 td->dm_dev.bdev = bdev;
760 td->dm_dev.bdev_file = bdev_file;
761 td->dm_dev.dax_dev = fs_dax_get_by_bdev(bdev, &part_off,
763 format_dev_t(td->dm_dev.name, dev);
764 list_add(&td->list, &md->table_devices);
775 * Close a table device that we've been using.
777 static void close_table_device(struct table_device *td, struct mapped_device *md)
779 if (md->disk->slave_dir)
780 bd_unlink_disk_holder(td->dm_dev.bdev, md->disk);
781 fput(td->dm_dev.bdev_file);
782 put_dax(td->dm_dev.dax_dev);
787 static struct table_device *find_table_device(struct list_head *l, dev_t dev,
790 struct table_device *td;
792 list_for_each_entry(td, l, list)
793 if (td->dm_dev.bdev->bd_dev == dev && td->dm_dev.mode == mode)
799 int dm_get_table_device(struct mapped_device *md, dev_t dev, blk_mode_t mode,
800 struct dm_dev **result)
802 struct table_device *td;
804 mutex_lock(&md->table_devices_lock);
805 td = find_table_device(&md->table_devices, dev, mode);
807 td = open_table_device(md, dev, mode);
809 mutex_unlock(&md->table_devices_lock);
813 refcount_inc(&td->count);
815 mutex_unlock(&md->table_devices_lock);
817 *result = &td->dm_dev;
821 void dm_put_table_device(struct mapped_device *md, struct dm_dev *d)
823 struct table_device *td = container_of(d, struct table_device, dm_dev);
825 mutex_lock(&md->table_devices_lock);
826 if (refcount_dec_and_test(&td->count))
827 close_table_device(td, md);
828 mutex_unlock(&md->table_devices_lock);
832 * Get the geometry associated with a dm device
834 int dm_get_geometry(struct mapped_device *md, struct hd_geometry *geo)
842 * Set the geometry of a device.
844 int dm_set_geometry(struct mapped_device *md, struct hd_geometry *geo)
846 sector_t sz = (sector_t)geo->cylinders * geo->heads * geo->sectors;
848 if (geo->start > sz) {
849 DMERR("Start sector is beyond the geometry limits.");
858 static int __noflush_suspending(struct mapped_device *md)
860 return test_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
863 static void dm_requeue_add_io(struct dm_io *io, bool first_stage)
865 struct mapped_device *md = io->md;
868 struct dm_io *next = md->requeue_list;
870 md->requeue_list = io;
873 bio_list_add_head(&md->deferred, io->orig_bio);
877 static void dm_kick_requeue(struct mapped_device *md, bool first_stage)
880 queue_work(md->wq, &md->requeue_work);
882 queue_work(md->wq, &md->work);
886 * Return true if the dm_io's original bio is requeued.
887 * io->status is updated with error if requeue disallowed.
889 static bool dm_handle_requeue(struct dm_io *io, bool first_stage)
891 struct bio *bio = io->orig_bio;
892 bool handle_requeue = (io->status == BLK_STS_DM_REQUEUE);
893 bool handle_polled_eagain = ((io->status == BLK_STS_AGAIN) &&
894 (bio->bi_opf & REQ_POLLED));
895 struct mapped_device *md = io->md;
896 bool requeued = false;
898 if (handle_requeue || handle_polled_eagain) {
901 if (bio->bi_opf & REQ_POLLED) {
903 * Upper layer won't help us poll split bio
904 * (io->orig_bio may only reflect a subset of the
905 * pre-split original) so clear REQ_POLLED.
907 bio_clear_polled(bio);
911 * Target requested pushing back the I/O or
912 * polled IO hit BLK_STS_AGAIN.
914 spin_lock_irqsave(&md->deferred_lock, flags);
915 if ((__noflush_suspending(md) &&
916 !WARN_ON_ONCE(dm_is_zone_write(md, bio))) ||
917 handle_polled_eagain || first_stage) {
918 dm_requeue_add_io(io, first_stage);
922 * noflush suspend was interrupted or this is
923 * a write to a zoned target.
925 io->status = BLK_STS_IOERR;
927 spin_unlock_irqrestore(&md->deferred_lock, flags);
931 dm_kick_requeue(md, first_stage);
936 static void __dm_io_complete(struct dm_io *io, bool first_stage)
938 struct bio *bio = io->orig_bio;
939 struct mapped_device *md = io->md;
940 blk_status_t io_error;
943 requeued = dm_handle_requeue(io, first_stage);
944 if (requeued && first_stage)
947 io_error = io->status;
948 if (dm_io_flagged(io, DM_IO_ACCOUNTED))
950 else if (!io_error) {
952 * Must handle target that DM_MAPIO_SUBMITTED only to
953 * then bio_endio() rather than dm_submit_bio_remap()
955 __dm_start_io_acct(io);
960 this_cpu_dec(*md->pending_io);
962 /* nudge anyone waiting on suspend queue */
963 if (unlikely(wq_has_sleeper(&md->wait)))
966 /* Return early if the original bio was requeued */
970 if (bio_is_flush_with_data(bio)) {
972 * Preflush done for flush with data, reissue
973 * without REQ_PREFLUSH.
975 bio->bi_opf &= ~REQ_PREFLUSH;
978 /* done with normal IO or empty flush */
980 bio->bi_status = io_error;
985 static void dm_wq_requeue_work(struct work_struct *work)
987 struct mapped_device *md = container_of(work, struct mapped_device,
992 /* reuse deferred lock to simplify dm_handle_requeue */
993 spin_lock_irqsave(&md->deferred_lock, flags);
994 io = md->requeue_list;
995 md->requeue_list = NULL;
996 spin_unlock_irqrestore(&md->deferred_lock, flags);
999 struct dm_io *next = io->next;
1001 dm_io_rewind(io, &md->disk->bio_split);
1004 __dm_io_complete(io, false);
1011 * Two staged requeue:
1013 * 1) io->orig_bio points to the real original bio, and the part mapped to
1014 * this io must be requeued, instead of other parts of the original bio.
1016 * 2) io->orig_bio points to new cloned bio which matches the requeued dm_io.
1018 static void dm_io_complete(struct dm_io *io)
1023 * Only dm_io that has been split needs two stage requeue, otherwise
1024 * we may run into long bio clone chain during suspend and OOM could
1027 * Also flush data dm_io won't be marked as DM_IO_WAS_SPLIT, so they
1028 * also aren't handled via the first stage requeue.
1030 if (dm_io_flagged(io, DM_IO_WAS_SPLIT))
1031 first_requeue = true;
1033 first_requeue = false;
1035 __dm_io_complete(io, first_requeue);
1039 * Decrements the number of outstanding ios that a bio has been
1040 * cloned into, completing the original io if necc.
1042 static inline void __dm_io_dec_pending(struct dm_io *io)
1044 if (atomic_dec_and_test(&io->io_count))
1048 static void dm_io_set_error(struct dm_io *io, blk_status_t error)
1050 unsigned long flags;
1052 /* Push-back supersedes any I/O errors */
1053 spin_lock_irqsave(&io->lock, flags);
1054 if (!(io->status == BLK_STS_DM_REQUEUE &&
1055 __noflush_suspending(io->md))) {
1058 spin_unlock_irqrestore(&io->lock, flags);
1061 static void dm_io_dec_pending(struct dm_io *io, blk_status_t error)
1063 if (unlikely(error))
1064 dm_io_set_error(io, error);
1066 __dm_io_dec_pending(io);
1070 * The queue_limits are only valid as long as you have a reference
1071 * count on 'md'. But _not_ imposing verification to avoid atomic_read(),
1073 static inline struct queue_limits *dm_get_queue_limits(struct mapped_device *md)
1075 return &md->queue->limits;
1078 void disable_discard(struct mapped_device *md)
1080 struct queue_limits *limits = dm_get_queue_limits(md);
1082 /* device doesn't really support DISCARD, disable it */
1083 limits->max_discard_sectors = 0;
1086 void disable_write_zeroes(struct mapped_device *md)
1088 struct queue_limits *limits = dm_get_queue_limits(md);
1090 /* device doesn't really support WRITE ZEROES, disable it */
1091 limits->max_write_zeroes_sectors = 0;
1094 static bool swap_bios_limit(struct dm_target *ti, struct bio *bio)
1096 return unlikely((bio->bi_opf & REQ_SWAP) != 0) && unlikely(ti->limit_swap_bios);
1099 static void clone_endio(struct bio *bio)
1101 blk_status_t error = bio->bi_status;
1102 struct dm_target_io *tio = clone_to_tio(bio);
1103 struct dm_target *ti = tio->ti;
1104 dm_endio_fn endio = ti->type->end_io;
1105 struct dm_io *io = tio->io;
1106 struct mapped_device *md = io->md;
1108 if (unlikely(error == BLK_STS_TARGET)) {
1109 if (bio_op(bio) == REQ_OP_DISCARD &&
1110 !bdev_max_discard_sectors(bio->bi_bdev))
1111 disable_discard(md);
1112 else if (bio_op(bio) == REQ_OP_WRITE_ZEROES &&
1113 !bdev_write_zeroes_sectors(bio->bi_bdev))
1114 disable_write_zeroes(md);
1117 if (static_branch_unlikely(&zoned_enabled) &&
1118 unlikely(bdev_is_zoned(bio->bi_bdev)))
1119 dm_zone_endio(io, bio);
1122 int r = endio(ti, bio, &error);
1125 case DM_ENDIO_REQUEUE:
1126 if (static_branch_unlikely(&zoned_enabled)) {
1128 * Requeuing writes to a sequential zone of a zoned
1129 * target will break the sequential write pattern:
1132 if (WARN_ON_ONCE(dm_is_zone_write(md, bio)))
1133 error = BLK_STS_IOERR;
1135 error = BLK_STS_DM_REQUEUE;
1137 error = BLK_STS_DM_REQUEUE;
1141 case DM_ENDIO_INCOMPLETE:
1142 /* The target will handle the io */
1145 DMCRIT("unimplemented target endio return value: %d", r);
1150 if (static_branch_unlikely(&swap_bios_enabled) &&
1151 unlikely(swap_bios_limit(ti, bio)))
1152 up(&md->swap_bios_semaphore);
1155 dm_io_dec_pending(io, error);
1159 * Return maximum size of I/O possible at the supplied sector up to the current
1162 static inline sector_t max_io_len_target_boundary(struct dm_target *ti,
1163 sector_t target_offset)
1165 return ti->len - target_offset;
1168 static sector_t __max_io_len(struct dm_target *ti, sector_t sector,
1169 unsigned int max_granularity,
1170 unsigned int max_sectors)
1172 sector_t target_offset = dm_target_offset(ti, sector);
1173 sector_t len = max_io_len_target_boundary(ti, target_offset);
1176 * Does the target need to split IO even further?
1177 * - varied (per target) IO splitting is a tenet of DM; this
1178 * explains why stacked chunk_sectors based splitting via
1179 * bio_split_to_limits() isn't possible here.
1181 if (!max_granularity)
1183 return min_t(sector_t, len,
1184 min(max_sectors ? : queue_max_sectors(ti->table->md->queue),
1185 blk_chunk_sectors_left(target_offset, max_granularity)));
1188 static inline sector_t max_io_len(struct dm_target *ti, sector_t sector)
1190 return __max_io_len(ti, sector, ti->max_io_len, 0);
1193 int dm_set_target_max_io_len(struct dm_target *ti, sector_t len)
1195 if (len > UINT_MAX) {
1196 DMERR("Specified maximum size of target IO (%llu) exceeds limit (%u)",
1197 (unsigned long long)len, UINT_MAX);
1198 ti->error = "Maximum size of target IO is too large";
1202 ti->max_io_len = (uint32_t) len;
1206 EXPORT_SYMBOL_GPL(dm_set_target_max_io_len);
1208 static struct dm_target *dm_dax_get_live_target(struct mapped_device *md,
1209 sector_t sector, int *srcu_idx)
1210 __acquires(md->io_barrier)
1212 struct dm_table *map;
1213 struct dm_target *ti;
1215 map = dm_get_live_table(md, srcu_idx);
1219 ti = dm_table_find_target(map, sector);
1226 static long dm_dax_direct_access(struct dax_device *dax_dev, pgoff_t pgoff,
1227 long nr_pages, enum dax_access_mode mode, void **kaddr,
1230 struct mapped_device *md = dax_get_private(dax_dev);
1231 sector_t sector = pgoff * PAGE_SECTORS;
1232 struct dm_target *ti;
1233 long len, ret = -EIO;
1236 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1240 if (!ti->type->direct_access)
1242 len = max_io_len(ti, sector) / PAGE_SECTORS;
1245 nr_pages = min(len, nr_pages);
1246 ret = ti->type->direct_access(ti, pgoff, nr_pages, mode, kaddr, pfn);
1249 dm_put_live_table(md, srcu_idx);
1254 static int dm_dax_zero_page_range(struct dax_device *dax_dev, pgoff_t pgoff,
1257 struct mapped_device *md = dax_get_private(dax_dev);
1258 sector_t sector = pgoff * PAGE_SECTORS;
1259 struct dm_target *ti;
1263 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1267 if (WARN_ON(!ti->type->dax_zero_page_range)) {
1269 * ->zero_page_range() is mandatory dax operation. If we are
1270 * here, something is wrong.
1274 ret = ti->type->dax_zero_page_range(ti, pgoff, nr_pages);
1276 dm_put_live_table(md, srcu_idx);
1281 static size_t dm_dax_recovery_write(struct dax_device *dax_dev, pgoff_t pgoff,
1282 void *addr, size_t bytes, struct iov_iter *i)
1284 struct mapped_device *md = dax_get_private(dax_dev);
1285 sector_t sector = pgoff * PAGE_SECTORS;
1286 struct dm_target *ti;
1290 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1291 if (!ti || !ti->type->dax_recovery_write)
1294 ret = ti->type->dax_recovery_write(ti, pgoff, addr, bytes, i);
1296 dm_put_live_table(md, srcu_idx);
1301 * A target may call dm_accept_partial_bio only from the map routine. It is
1302 * allowed for all bio types except REQ_PREFLUSH, REQ_OP_ZONE_* zone management
1303 * operations, REQ_OP_ZONE_APPEND (zone append writes) and any bio serviced by
1304 * __send_duplicate_bios().
1306 * dm_accept_partial_bio informs the dm that the target only wants to process
1307 * additional n_sectors sectors of the bio and the rest of the data should be
1308 * sent in a next bio.
1310 * A diagram that explains the arithmetics:
1311 * +--------------------+---------------+-------+
1313 * +--------------------+---------------+-------+
1315 * <-------------- *tio->len_ptr --------------->
1316 * <----- bio_sectors ----->
1319 * Region 1 was already iterated over with bio_advance or similar function.
1320 * (it may be empty if the target doesn't use bio_advance)
1321 * Region 2 is the remaining bio size that the target wants to process.
1322 * (it may be empty if region 1 is non-empty, although there is no reason
1324 * The target requires that region 3 is to be sent in the next bio.
1326 * If the target wants to receive multiple copies of the bio (via num_*bios, etc),
1327 * the partially processed part (the sum of regions 1+2) must be the same for all
1328 * copies of the bio.
1330 void dm_accept_partial_bio(struct bio *bio, unsigned int n_sectors)
1332 struct dm_target_io *tio = clone_to_tio(bio);
1333 struct dm_io *io = tio->io;
1334 unsigned int bio_sectors = bio_sectors(bio);
1336 BUG_ON(dm_tio_flagged(tio, DM_TIO_IS_DUPLICATE_BIO));
1337 BUG_ON(op_is_zone_mgmt(bio_op(bio)));
1338 BUG_ON(bio_op(bio) == REQ_OP_ZONE_APPEND);
1339 BUG_ON(bio_sectors > *tio->len_ptr);
1340 BUG_ON(n_sectors > bio_sectors);
1342 *tio->len_ptr -= bio_sectors - n_sectors;
1343 bio->bi_iter.bi_size = n_sectors << SECTOR_SHIFT;
1346 * __split_and_process_bio() may have already saved mapped part
1347 * for accounting but it is being reduced so update accordingly.
1349 dm_io_set_flag(io, DM_IO_WAS_SPLIT);
1350 io->sectors = n_sectors;
1351 io->sector_offset = bio_sectors(io->orig_bio);
1353 EXPORT_SYMBOL_GPL(dm_accept_partial_bio);
1356 * @clone: clone bio that DM core passed to target's .map function
1357 * @tgt_clone: clone of @clone bio that target needs submitted
1359 * Targets should use this interface to submit bios they take
1360 * ownership of when returning DM_MAPIO_SUBMITTED.
1362 * Target should also enable ti->accounts_remapped_io
1364 void dm_submit_bio_remap(struct bio *clone, struct bio *tgt_clone)
1366 struct dm_target_io *tio = clone_to_tio(clone);
1367 struct dm_io *io = tio->io;
1369 /* establish bio that will get submitted */
1374 * Account io->origin_bio to DM dev on behalf of target
1375 * that took ownership of IO with DM_MAPIO_SUBMITTED.
1377 dm_start_io_acct(io, clone);
1379 trace_block_bio_remap(tgt_clone, disk_devt(io->md->disk),
1381 submit_bio_noacct(tgt_clone);
1383 EXPORT_SYMBOL_GPL(dm_submit_bio_remap);
1385 static noinline void __set_swap_bios_limit(struct mapped_device *md, int latch)
1387 mutex_lock(&md->swap_bios_lock);
1388 while (latch < md->swap_bios) {
1390 down(&md->swap_bios_semaphore);
1393 while (latch > md->swap_bios) {
1395 up(&md->swap_bios_semaphore);
1398 mutex_unlock(&md->swap_bios_lock);
1401 static void __map_bio(struct bio *clone)
1403 struct dm_target_io *tio = clone_to_tio(clone);
1404 struct dm_target *ti = tio->ti;
1405 struct dm_io *io = tio->io;
1406 struct mapped_device *md = io->md;
1409 clone->bi_end_io = clone_endio;
1414 tio->old_sector = clone->bi_iter.bi_sector;
1416 if (static_branch_unlikely(&swap_bios_enabled) &&
1417 unlikely(swap_bios_limit(ti, clone))) {
1418 int latch = get_swap_bios();
1420 if (unlikely(latch != md->swap_bios))
1421 __set_swap_bios_limit(md, latch);
1422 down(&md->swap_bios_semaphore);
1425 if (static_branch_unlikely(&zoned_enabled)) {
1427 * Check if the IO needs a special mapping due to zone append
1428 * emulation on zoned target. In this case, dm_zone_map_bio()
1429 * calls the target map operation.
1431 if (unlikely(dm_emulate_zone_append(md)))
1432 r = dm_zone_map_bio(tio);
1437 if (likely(ti->type->map == linear_map))
1438 r = linear_map(ti, clone);
1439 else if (ti->type->map == stripe_map)
1440 r = stripe_map(ti, clone);
1442 r = ti->type->map(ti, clone);
1446 case DM_MAPIO_SUBMITTED:
1447 /* target has assumed ownership of this io */
1448 if (!ti->accounts_remapped_io)
1449 dm_start_io_acct(io, clone);
1451 case DM_MAPIO_REMAPPED:
1452 dm_submit_bio_remap(clone, NULL);
1455 case DM_MAPIO_REQUEUE:
1456 if (static_branch_unlikely(&swap_bios_enabled) &&
1457 unlikely(swap_bios_limit(ti, clone)))
1458 up(&md->swap_bios_semaphore);
1460 if (r == DM_MAPIO_KILL)
1461 dm_io_dec_pending(io, BLK_STS_IOERR);
1463 dm_io_dec_pending(io, BLK_STS_DM_REQUEUE);
1466 DMCRIT("unimplemented target map return value: %d", r);
1471 static void setup_split_accounting(struct clone_info *ci, unsigned int len)
1473 struct dm_io *io = ci->io;
1475 if (ci->sector_count > len) {
1477 * Split needed, save the mapped part for accounting.
1478 * NOTE: dm_accept_partial_bio() will update accordingly.
1480 dm_io_set_flag(io, DM_IO_WAS_SPLIT);
1482 io->sector_offset = bio_sectors(ci->bio);
1486 static void alloc_multiple_bios(struct bio_list *blist, struct clone_info *ci,
1487 struct dm_target *ti, unsigned int num_bios,
1488 unsigned *len, gfp_t gfp_flag)
1491 int try = (gfp_flag & GFP_NOWAIT) ? 0 : 1;
1493 for (; try < 2; try++) {
1496 if (try && num_bios > 1)
1497 mutex_lock(&ci->io->md->table_devices_lock);
1498 for (bio_nr = 0; bio_nr < num_bios; bio_nr++) {
1499 bio = alloc_tio(ci, ti, bio_nr, len,
1500 try ? GFP_NOIO : GFP_NOWAIT);
1504 bio_list_add(blist, bio);
1506 if (try && num_bios > 1)
1507 mutex_unlock(&ci->io->md->table_devices_lock);
1508 if (bio_nr == num_bios)
1511 while ((bio = bio_list_pop(blist)))
1516 static unsigned int __send_duplicate_bios(struct clone_info *ci, struct dm_target *ti,
1517 unsigned int num_bios, unsigned int *len,
1520 struct bio_list blist = BIO_EMPTY_LIST;
1522 unsigned int ret = 0;
1524 if (WARN_ON_ONCE(num_bios == 0)) /* num_bios = 0 is a bug in caller */
1527 /* dm_accept_partial_bio() is not supported with shared tio->len_ptr */
1529 setup_split_accounting(ci, *len);
1532 * Using alloc_multiple_bios(), even if num_bios is 1, to consistently
1533 * support allocating using GFP_NOWAIT with GFP_NOIO fallback.
1535 alloc_multiple_bios(&blist, ci, ti, num_bios, len, gfp_flag);
1536 while ((clone = bio_list_pop(&blist))) {
1538 dm_tio_set_flag(clone_to_tio(clone), DM_TIO_IS_DUPLICATE_BIO);
1546 static void __send_empty_flush(struct clone_info *ci)
1548 struct dm_table *t = ci->map;
1549 struct bio flush_bio;
1552 * Use an on-stack bio for this, it's safe since we don't
1553 * need to reference it after submit. It's just used as
1554 * the basis for the clone(s).
1556 bio_init(&flush_bio, ci->io->md->disk->part0, NULL, 0,
1557 REQ_OP_WRITE | REQ_PREFLUSH | REQ_SYNC);
1559 ci->bio = &flush_bio;
1560 ci->sector_count = 0;
1561 ci->io->tio.clone.bi_iter.bi_size = 0;
1563 for (unsigned int i = 0; i < t->num_targets; i++) {
1565 struct dm_target *ti = dm_table_get_target(t, i);
1567 if (unlikely(ti->num_flush_bios == 0))
1570 atomic_add(ti->num_flush_bios, &ci->io->io_count);
1571 bios = __send_duplicate_bios(ci, ti, ti->num_flush_bios,
1573 atomic_sub(ti->num_flush_bios - bios, &ci->io->io_count);
1577 * alloc_io() takes one extra reference for submission, so the
1578 * reference won't reach 0 without the following subtraction
1580 atomic_sub(1, &ci->io->io_count);
1582 bio_uninit(ci->bio);
1585 static void __send_abnormal_io(struct clone_info *ci, struct dm_target *ti,
1586 unsigned int num_bios, unsigned int max_granularity,
1587 unsigned int max_sectors)
1589 unsigned int len, bios;
1591 len = min_t(sector_t, ci->sector_count,
1592 __max_io_len(ti, ci->sector, max_granularity, max_sectors));
1594 atomic_add(num_bios, &ci->io->io_count);
1595 bios = __send_duplicate_bios(ci, ti, num_bios, &len, GFP_NOIO);
1597 * alloc_io() takes one extra reference for submission, so the
1598 * reference won't reach 0 without the following (+1) subtraction
1600 atomic_sub(num_bios - bios + 1, &ci->io->io_count);
1603 ci->sector_count -= len;
1606 static bool is_abnormal_io(struct bio *bio)
1608 enum req_op op = bio_op(bio);
1610 if (op != REQ_OP_READ && op != REQ_OP_WRITE && op != REQ_OP_FLUSH) {
1612 case REQ_OP_DISCARD:
1613 case REQ_OP_SECURE_ERASE:
1614 case REQ_OP_WRITE_ZEROES:
1624 static blk_status_t __process_abnormal_io(struct clone_info *ci,
1625 struct dm_target *ti)
1627 unsigned int num_bios = 0;
1628 unsigned int max_granularity = 0;
1629 unsigned int max_sectors = 0;
1630 struct queue_limits *limits = dm_get_queue_limits(ti->table->md);
1632 switch (bio_op(ci->bio)) {
1633 case REQ_OP_DISCARD:
1634 num_bios = ti->num_discard_bios;
1635 max_sectors = limits->max_discard_sectors;
1636 if (ti->max_discard_granularity)
1637 max_granularity = max_sectors;
1639 case REQ_OP_SECURE_ERASE:
1640 num_bios = ti->num_secure_erase_bios;
1641 max_sectors = limits->max_secure_erase_sectors;
1642 if (ti->max_secure_erase_granularity)
1643 max_granularity = max_sectors;
1645 case REQ_OP_WRITE_ZEROES:
1646 num_bios = ti->num_write_zeroes_bios;
1647 max_sectors = limits->max_write_zeroes_sectors;
1648 if (ti->max_write_zeroes_granularity)
1649 max_granularity = max_sectors;
1656 * Even though the device advertised support for this type of
1657 * request, that does not mean every target supports it, and
1658 * reconfiguration might also have changed that since the
1659 * check was performed.
1661 if (unlikely(!num_bios))
1662 return BLK_STS_NOTSUPP;
1664 __send_abnormal_io(ci, ti, num_bios, max_granularity, max_sectors);
1670 * Reuse ->bi_private as dm_io list head for storing all dm_io instances
1671 * associated with this bio, and this bio's bi_private needs to be
1672 * stored in dm_io->data before the reuse.
1674 * bio->bi_private is owned by fs or upper layer, so block layer won't
1675 * touch it after splitting. Meantime it won't be changed by anyone after
1676 * bio is submitted. So this reuse is safe.
1678 static inline struct dm_io **dm_poll_list_head(struct bio *bio)
1680 return (struct dm_io **)&bio->bi_private;
1683 static void dm_queue_poll_io(struct bio *bio, struct dm_io *io)
1685 struct dm_io **head = dm_poll_list_head(bio);
1687 if (!(bio->bi_opf & REQ_DM_POLL_LIST)) {
1688 bio->bi_opf |= REQ_DM_POLL_LIST;
1690 * Save .bi_private into dm_io, so that we can reuse
1691 * .bi_private as dm_io list head for storing dm_io list
1693 io->data = bio->bi_private;
1695 /* tell block layer to poll for completion */
1696 bio->bi_cookie = ~BLK_QC_T_NONE;
1701 * bio recursed due to split, reuse original poll list,
1702 * and save bio->bi_private too.
1704 io->data = (*head)->data;
1712 * Select the correct strategy for processing a non-flush bio.
1714 static blk_status_t __split_and_process_bio(struct clone_info *ci)
1717 struct dm_target *ti;
1720 ti = dm_table_find_target(ci->map, ci->sector);
1722 return BLK_STS_IOERR;
1724 if (unlikely(ci->is_abnormal_io))
1725 return __process_abnormal_io(ci, ti);
1728 * Only support bio polling for normal IO, and the target io is
1729 * exactly inside the dm_io instance (verified in dm_poll_dm_io)
1731 ci->submit_as_polled = !!(ci->bio->bi_opf & REQ_POLLED);
1733 len = min_t(sector_t, max_io_len(ti, ci->sector), ci->sector_count);
1734 setup_split_accounting(ci, len);
1736 if (unlikely(ci->bio->bi_opf & REQ_NOWAIT)) {
1737 if (unlikely(!dm_target_supports_nowait(ti->type)))
1738 return BLK_STS_NOTSUPP;
1740 clone = alloc_tio(ci, ti, 0, &len, GFP_NOWAIT);
1741 if (unlikely(!clone))
1742 return BLK_STS_AGAIN;
1744 clone = alloc_tio(ci, ti, 0, &len, GFP_NOIO);
1749 ci->sector_count -= len;
1754 static void init_clone_info(struct clone_info *ci, struct dm_io *io,
1755 struct dm_table *map, struct bio *bio, bool is_abnormal)
1760 ci->is_abnormal_io = is_abnormal;
1761 ci->submit_as_polled = false;
1762 ci->sector = bio->bi_iter.bi_sector;
1763 ci->sector_count = bio_sectors(bio);
1765 /* Shouldn't happen but sector_count was being set to 0 so... */
1766 if (static_branch_unlikely(&zoned_enabled) &&
1767 WARN_ON_ONCE(op_is_zone_mgmt(bio_op(bio)) && ci->sector_count))
1768 ci->sector_count = 0;
1772 * Entry point to split a bio into clones and submit them to the targets.
1774 static void dm_split_and_process_bio(struct mapped_device *md,
1775 struct dm_table *map, struct bio *bio)
1777 struct clone_info ci;
1779 blk_status_t error = BLK_STS_OK;
1782 is_abnormal = is_abnormal_io(bio);
1783 if (unlikely(is_abnormal)) {
1785 * Use bio_split_to_limits() for abnormal IO (e.g. discard, etc)
1786 * otherwise associated queue_limits won't be imposed.
1788 bio = bio_split_to_limits(bio);
1793 /* Only support nowait for normal IO */
1794 if (unlikely(bio->bi_opf & REQ_NOWAIT) && !is_abnormal) {
1795 io = alloc_io(md, bio, GFP_NOWAIT);
1796 if (unlikely(!io)) {
1797 /* Unable to do anything without dm_io. */
1798 bio_wouldblock_error(bio);
1802 io = alloc_io(md, bio, GFP_NOIO);
1804 init_clone_info(&ci, io, map, bio, is_abnormal);
1806 if (bio->bi_opf & REQ_PREFLUSH) {
1807 __send_empty_flush(&ci);
1808 /* dm_io_complete submits any data associated with flush */
1812 error = __split_and_process_bio(&ci);
1813 if (error || !ci.sector_count)
1816 * Remainder must be passed to submit_bio_noacct() so it gets handled
1817 * *after* bios already submitted have been completely processed.
1819 bio_trim(bio, io->sectors, ci.sector_count);
1820 trace_block_split(bio, bio->bi_iter.bi_sector);
1821 bio_inc_remaining(bio);
1822 submit_bio_noacct(bio);
1825 * Drop the extra reference count for non-POLLED bio, and hold one
1826 * reference for POLLED bio, which will be released in dm_poll_bio
1828 * Add every dm_io instance into the dm_io list head which is stored
1829 * in bio->bi_private, so that dm_poll_bio can poll them all.
1831 if (error || !ci.submit_as_polled) {
1833 * In case of submission failure, the extra reference for
1834 * submitting io isn't consumed yet
1837 atomic_dec(&io->io_count);
1838 dm_io_dec_pending(io, error);
1840 dm_queue_poll_io(bio, io);
1843 static void dm_submit_bio(struct bio *bio)
1845 struct mapped_device *md = bio->bi_bdev->bd_disk->private_data;
1847 struct dm_table *map;
1849 map = dm_get_live_table(md, &srcu_idx);
1851 /* If suspended, or map not yet available, queue this IO for later */
1852 if (unlikely(test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) ||
1854 if (bio->bi_opf & REQ_NOWAIT)
1855 bio_wouldblock_error(bio);
1856 else if (bio->bi_opf & REQ_RAHEAD)
1863 dm_split_and_process_bio(md, map, bio);
1865 dm_put_live_table(md, srcu_idx);
1868 static bool dm_poll_dm_io(struct dm_io *io, struct io_comp_batch *iob,
1871 WARN_ON_ONCE(!dm_tio_is_normal(&io->tio));
1873 /* don't poll if the mapped io is done */
1874 if (atomic_read(&io->io_count) > 1)
1875 bio_poll(&io->tio.clone, iob, flags);
1877 /* bio_poll holds the last reference */
1878 return atomic_read(&io->io_count) == 1;
1881 static int dm_poll_bio(struct bio *bio, struct io_comp_batch *iob,
1884 struct dm_io **head = dm_poll_list_head(bio);
1885 struct dm_io *list = *head;
1886 struct dm_io *tmp = NULL;
1887 struct dm_io *curr, *next;
1889 /* Only poll normal bio which was marked as REQ_DM_POLL_LIST */
1890 if (!(bio->bi_opf & REQ_DM_POLL_LIST))
1893 WARN_ON_ONCE(!list);
1896 * Restore .bi_private before possibly completing dm_io.
1898 * bio_poll() is only possible once @bio has been completely
1899 * submitted via submit_bio_noacct()'s depth-first submission.
1900 * So there is no dm_queue_poll_io() race associated with
1901 * clearing REQ_DM_POLL_LIST here.
1903 bio->bi_opf &= ~REQ_DM_POLL_LIST;
1904 bio->bi_private = list->data;
1906 for (curr = list, next = curr->next; curr; curr = next, next =
1907 curr ? curr->next : NULL) {
1908 if (dm_poll_dm_io(curr, iob, flags)) {
1910 * clone_endio() has already occurred, so no
1911 * error handling is needed here.
1913 __dm_io_dec_pending(curr);
1922 bio->bi_opf |= REQ_DM_POLL_LIST;
1923 /* Reset bio->bi_private to dm_io list head */
1931 *---------------------------------------------------------------
1932 * An IDR is used to keep track of allocated minor numbers.
1933 *---------------------------------------------------------------
1935 static void free_minor(int minor)
1937 spin_lock(&_minor_lock);
1938 idr_remove(&_minor_idr, minor);
1939 spin_unlock(&_minor_lock);
1943 * See if the device with a specific minor # is free.
1945 static int specific_minor(int minor)
1949 if (minor >= (1 << MINORBITS))
1952 idr_preload(GFP_KERNEL);
1953 spin_lock(&_minor_lock);
1955 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, minor, minor + 1, GFP_NOWAIT);
1957 spin_unlock(&_minor_lock);
1960 return r == -ENOSPC ? -EBUSY : r;
1964 static int next_free_minor(int *minor)
1968 idr_preload(GFP_KERNEL);
1969 spin_lock(&_minor_lock);
1971 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, 0, 1 << MINORBITS, GFP_NOWAIT);
1973 spin_unlock(&_minor_lock);
1981 static const struct block_device_operations dm_blk_dops;
1982 static const struct block_device_operations dm_rq_blk_dops;
1983 static const struct dax_operations dm_dax_ops;
1985 static void dm_wq_work(struct work_struct *work);
1987 #ifdef CONFIG_BLK_INLINE_ENCRYPTION
1988 static void dm_queue_destroy_crypto_profile(struct request_queue *q)
1990 dm_destroy_crypto_profile(q->crypto_profile);
1993 #else /* CONFIG_BLK_INLINE_ENCRYPTION */
1995 static inline void dm_queue_destroy_crypto_profile(struct request_queue *q)
1998 #endif /* !CONFIG_BLK_INLINE_ENCRYPTION */
2000 static void cleanup_mapped_device(struct mapped_device *md)
2003 destroy_workqueue(md->wq);
2004 dm_free_md_mempools(md->mempools);
2007 dax_remove_host(md->disk);
2008 kill_dax(md->dax_dev);
2009 put_dax(md->dax_dev);
2013 dm_cleanup_zoned_dev(md);
2015 spin_lock(&_minor_lock);
2016 md->disk->private_data = NULL;
2017 spin_unlock(&_minor_lock);
2018 if (dm_get_md_type(md) != DM_TYPE_NONE) {
2019 struct table_device *td;
2022 list_for_each_entry(td, &md->table_devices, list) {
2023 bd_unlink_disk_holder(td->dm_dev.bdev,
2028 * Hold lock to make sure del_gendisk() won't concurrent
2029 * with open/close_table_device().
2031 mutex_lock(&md->table_devices_lock);
2032 del_gendisk(md->disk);
2033 mutex_unlock(&md->table_devices_lock);
2035 dm_queue_destroy_crypto_profile(md->queue);
2039 if (md->pending_io) {
2040 free_percpu(md->pending_io);
2041 md->pending_io = NULL;
2044 cleanup_srcu_struct(&md->io_barrier);
2046 mutex_destroy(&md->suspend_lock);
2047 mutex_destroy(&md->type_lock);
2048 mutex_destroy(&md->table_devices_lock);
2049 mutex_destroy(&md->swap_bios_lock);
2051 dm_mq_cleanup_mapped_device(md);
2055 * Allocate and initialise a blank device with a given minor.
2057 static struct mapped_device *alloc_dev(int minor)
2059 int r, numa_node_id = dm_get_numa_node();
2060 struct mapped_device *md;
2063 md = kvzalloc_node(sizeof(*md), GFP_KERNEL, numa_node_id);
2065 DMERR("unable to allocate device, out of memory.");
2069 if (!try_module_get(THIS_MODULE))
2070 goto bad_module_get;
2072 /* get a minor number for the dev */
2073 if (minor == DM_ANY_MINOR)
2074 r = next_free_minor(&minor);
2076 r = specific_minor(minor);
2080 r = init_srcu_struct(&md->io_barrier);
2082 goto bad_io_barrier;
2084 md->numa_node_id = numa_node_id;
2085 md->init_tio_pdu = false;
2086 md->type = DM_TYPE_NONE;
2087 mutex_init(&md->suspend_lock);
2088 mutex_init(&md->type_lock);
2089 mutex_init(&md->table_devices_lock);
2090 spin_lock_init(&md->deferred_lock);
2091 atomic_set(&md->holders, 1);
2092 atomic_set(&md->open_count, 0);
2093 atomic_set(&md->event_nr, 0);
2094 atomic_set(&md->uevent_seq, 0);
2095 INIT_LIST_HEAD(&md->uevent_list);
2096 INIT_LIST_HEAD(&md->table_devices);
2097 spin_lock_init(&md->uevent_lock);
2100 * default to bio-based until DM table is loaded and md->type
2101 * established. If request-based table is loaded: blk-mq will
2102 * override accordingly.
2104 md->disk = blk_alloc_disk(NULL, md->numa_node_id);
2105 if (IS_ERR(md->disk))
2107 md->queue = md->disk->queue;
2109 init_waitqueue_head(&md->wait);
2110 INIT_WORK(&md->work, dm_wq_work);
2111 INIT_WORK(&md->requeue_work, dm_wq_requeue_work);
2112 init_waitqueue_head(&md->eventq);
2113 init_completion(&md->kobj_holder.completion);
2115 md->requeue_list = NULL;
2116 md->swap_bios = get_swap_bios();
2117 sema_init(&md->swap_bios_semaphore, md->swap_bios);
2118 mutex_init(&md->swap_bios_lock);
2120 md->disk->major = _major;
2121 md->disk->first_minor = minor;
2122 md->disk->minors = 1;
2123 md->disk->flags |= GENHD_FL_NO_PART;
2124 md->disk->fops = &dm_blk_dops;
2125 md->disk->private_data = md;
2126 sprintf(md->disk->disk_name, "dm-%d", minor);
2128 if (IS_ENABLED(CONFIG_FS_DAX)) {
2129 md->dax_dev = alloc_dax(md, &dm_dax_ops);
2130 if (IS_ERR(md->dax_dev)) {
2134 set_dax_nocache(md->dax_dev);
2135 set_dax_nomc(md->dax_dev);
2136 if (dax_add_host(md->dax_dev, md->disk))
2140 format_dev_t(md->name, MKDEV(_major, minor));
2142 md->wq = alloc_workqueue("kdmflush/%s", WQ_MEM_RECLAIM, 0, md->name);
2146 md->pending_io = alloc_percpu(unsigned long);
2147 if (!md->pending_io)
2150 r = dm_stats_init(&md->stats);
2154 /* Populate the mapping, nobody knows we exist yet */
2155 spin_lock(&_minor_lock);
2156 old_md = idr_replace(&_minor_idr, md, minor);
2157 spin_unlock(&_minor_lock);
2159 BUG_ON(old_md != MINOR_ALLOCED);
2164 cleanup_mapped_device(md);
2168 module_put(THIS_MODULE);
2174 static void unlock_fs(struct mapped_device *md);
2176 static void free_dev(struct mapped_device *md)
2178 int minor = MINOR(disk_devt(md->disk));
2182 cleanup_mapped_device(md);
2184 WARN_ON_ONCE(!list_empty(&md->table_devices));
2185 dm_stats_cleanup(&md->stats);
2188 module_put(THIS_MODULE);
2193 * Bind a table to the device.
2195 static void event_callback(void *context)
2197 unsigned long flags;
2199 struct mapped_device *md = context;
2201 spin_lock_irqsave(&md->uevent_lock, flags);
2202 list_splice_init(&md->uevent_list, &uevents);
2203 spin_unlock_irqrestore(&md->uevent_lock, flags);
2205 dm_send_uevents(&uevents, &disk_to_dev(md->disk)->kobj);
2207 atomic_inc(&md->event_nr);
2208 wake_up(&md->eventq);
2209 dm_issue_global_event();
2213 * Returns old map, which caller must destroy.
2215 static struct dm_table *__bind(struct mapped_device *md, struct dm_table *t,
2216 struct queue_limits *limits)
2218 struct dm_table *old_map;
2222 lockdep_assert_held(&md->suspend_lock);
2224 size = dm_table_get_size(t);
2227 * Wipe any geometry if the size of the table changed.
2229 if (size != dm_get_size(md))
2230 memset(&md->geometry, 0, sizeof(md->geometry));
2232 set_capacity(md->disk, size);
2234 dm_table_event_callback(t, event_callback, md);
2236 if (dm_table_request_based(t)) {
2238 * Leverage the fact that request-based DM targets are
2239 * immutable singletons - used to optimize dm_mq_queue_rq.
2241 md->immutable_target = dm_table_get_immutable_target(t);
2244 * There is no need to reload with request-based dm because the
2245 * size of front_pad doesn't change.
2247 * Note for future: If you are to reload bioset, prep-ed
2248 * requests in the queue may refer to bio from the old bioset,
2249 * so you must walk through the queue to unprep.
2251 if (!md->mempools) {
2252 md->mempools = t->mempools;
2257 * The md may already have mempools that need changing.
2258 * If so, reload bioset because front_pad may have changed
2259 * because a different table was loaded.
2261 dm_free_md_mempools(md->mempools);
2262 md->mempools = t->mempools;
2266 ret = dm_table_set_restrictions(t, md->queue, limits);
2268 old_map = ERR_PTR(ret);
2272 old_map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2273 rcu_assign_pointer(md->map, (void *)t);
2274 md->immutable_target_type = dm_table_get_immutable_target_type(t);
2283 * Returns unbound table for the caller to free.
2285 static struct dm_table *__unbind(struct mapped_device *md)
2287 struct dm_table *map = rcu_dereference_protected(md->map, 1);
2292 dm_table_event_callback(map, NULL, NULL);
2293 RCU_INIT_POINTER(md->map, NULL);
2300 * Constructor for a new device.
2302 int dm_create(int minor, struct mapped_device **result)
2304 struct mapped_device *md;
2306 md = alloc_dev(minor);
2310 dm_ima_reset_data(md);
2317 * Functions to manage md->type.
2318 * All are required to hold md->type_lock.
2320 void dm_lock_md_type(struct mapped_device *md)
2322 mutex_lock(&md->type_lock);
2325 void dm_unlock_md_type(struct mapped_device *md)
2327 mutex_unlock(&md->type_lock);
2330 void dm_set_md_type(struct mapped_device *md, enum dm_queue_mode type)
2332 BUG_ON(!mutex_is_locked(&md->type_lock));
2336 enum dm_queue_mode dm_get_md_type(struct mapped_device *md)
2341 struct target_type *dm_get_immutable_target_type(struct mapped_device *md)
2343 return md->immutable_target_type;
2347 * Setup the DM device's queue based on md's type
2349 int dm_setup_md_queue(struct mapped_device *md, struct dm_table *t)
2351 enum dm_queue_mode type = dm_table_get_type(t);
2352 struct queue_limits limits;
2353 struct table_device *td;
2357 case DM_TYPE_REQUEST_BASED:
2358 md->disk->fops = &dm_rq_blk_dops;
2359 r = dm_mq_init_request_queue(md, t);
2361 DMERR("Cannot initialize queue for request-based dm mapped device");
2365 case DM_TYPE_BIO_BASED:
2366 case DM_TYPE_DAX_BIO_BASED:
2367 blk_queue_flag_set(QUEUE_FLAG_IO_STAT, md->queue);
2374 r = dm_calculate_queue_limits(t, &limits);
2376 DMERR("Cannot calculate initial queue limits");
2379 r = dm_table_set_restrictions(t, md->queue, &limits);
2384 * Hold lock to make sure add_disk() and del_gendisk() won't concurrent
2385 * with open_table_device() and close_table_device().
2387 mutex_lock(&md->table_devices_lock);
2388 r = add_disk(md->disk);
2389 mutex_unlock(&md->table_devices_lock);
2394 * Register the holder relationship for devices added before the disk
2397 list_for_each_entry(td, &md->table_devices, list) {
2398 r = bd_link_disk_holder(td->dm_dev.bdev, md->disk);
2400 goto out_undo_holders;
2403 r = dm_sysfs_init(md);
2405 goto out_undo_holders;
2411 list_for_each_entry_continue_reverse(td, &md->table_devices, list)
2412 bd_unlink_disk_holder(td->dm_dev.bdev, md->disk);
2413 mutex_lock(&md->table_devices_lock);
2414 del_gendisk(md->disk);
2415 mutex_unlock(&md->table_devices_lock);
2419 struct mapped_device *dm_get_md(dev_t dev)
2421 struct mapped_device *md;
2422 unsigned int minor = MINOR(dev);
2424 if (MAJOR(dev) != _major || minor >= (1 << MINORBITS))
2427 spin_lock(&_minor_lock);
2429 md = idr_find(&_minor_idr, minor);
2430 if (!md || md == MINOR_ALLOCED || (MINOR(disk_devt(dm_disk(md))) != minor) ||
2431 test_bit(DMF_FREEING, &md->flags) || dm_deleting_md(md)) {
2437 spin_unlock(&_minor_lock);
2441 EXPORT_SYMBOL_GPL(dm_get_md);
2443 void *dm_get_mdptr(struct mapped_device *md)
2445 return md->interface_ptr;
2448 void dm_set_mdptr(struct mapped_device *md, void *ptr)
2450 md->interface_ptr = ptr;
2453 void dm_get(struct mapped_device *md)
2455 atomic_inc(&md->holders);
2456 BUG_ON(test_bit(DMF_FREEING, &md->flags));
2459 int dm_hold(struct mapped_device *md)
2461 spin_lock(&_minor_lock);
2462 if (test_bit(DMF_FREEING, &md->flags)) {
2463 spin_unlock(&_minor_lock);
2467 spin_unlock(&_minor_lock);
2470 EXPORT_SYMBOL_GPL(dm_hold);
2472 const char *dm_device_name(struct mapped_device *md)
2476 EXPORT_SYMBOL_GPL(dm_device_name);
2478 static void __dm_destroy(struct mapped_device *md, bool wait)
2480 struct dm_table *map;
2485 spin_lock(&_minor_lock);
2486 idr_replace(&_minor_idr, MINOR_ALLOCED, MINOR(disk_devt(dm_disk(md))));
2487 set_bit(DMF_FREEING, &md->flags);
2488 spin_unlock(&_minor_lock);
2490 blk_mark_disk_dead(md->disk);
2493 * Take suspend_lock so that presuspend and postsuspend methods
2494 * do not race with internal suspend.
2496 mutex_lock(&md->suspend_lock);
2497 map = dm_get_live_table(md, &srcu_idx);
2498 if (!dm_suspended_md(md)) {
2499 dm_table_presuspend_targets(map);
2500 set_bit(DMF_SUSPENDED, &md->flags);
2501 set_bit(DMF_POST_SUSPENDING, &md->flags);
2502 dm_table_postsuspend_targets(map);
2504 /* dm_put_live_table must be before fsleep, otherwise deadlock is possible */
2505 dm_put_live_table(md, srcu_idx);
2506 mutex_unlock(&md->suspend_lock);
2509 * Rare, but there may be I/O requests still going to complete,
2510 * for example. Wait for all references to disappear.
2511 * No one should increment the reference count of the mapped_device,
2512 * after the mapped_device state becomes DMF_FREEING.
2515 while (atomic_read(&md->holders))
2517 else if (atomic_read(&md->holders))
2518 DMWARN("%s: Forcibly removing mapped_device still in use! (%d users)",
2519 dm_device_name(md), atomic_read(&md->holders));
2521 dm_table_destroy(__unbind(md));
2525 void dm_destroy(struct mapped_device *md)
2527 __dm_destroy(md, true);
2530 void dm_destroy_immediate(struct mapped_device *md)
2532 __dm_destroy(md, false);
2535 void dm_put(struct mapped_device *md)
2537 atomic_dec(&md->holders);
2539 EXPORT_SYMBOL_GPL(dm_put);
2541 static bool dm_in_flight_bios(struct mapped_device *md)
2544 unsigned long sum = 0;
2546 for_each_possible_cpu(cpu)
2547 sum += *per_cpu_ptr(md->pending_io, cpu);
2552 static int dm_wait_for_bios_completion(struct mapped_device *md, unsigned int task_state)
2558 prepare_to_wait(&md->wait, &wait, task_state);
2560 if (!dm_in_flight_bios(md))
2563 if (signal_pending_state(task_state, current)) {
2570 finish_wait(&md->wait, &wait);
2577 static int dm_wait_for_completion(struct mapped_device *md, unsigned int task_state)
2581 if (!queue_is_mq(md->queue))
2582 return dm_wait_for_bios_completion(md, task_state);
2585 if (!blk_mq_queue_inflight(md->queue))
2588 if (signal_pending_state(task_state, current)) {
2600 * Process the deferred bios
2602 static void dm_wq_work(struct work_struct *work)
2604 struct mapped_device *md = container_of(work, struct mapped_device, work);
2607 while (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
2608 spin_lock_irq(&md->deferred_lock);
2609 bio = bio_list_pop(&md->deferred);
2610 spin_unlock_irq(&md->deferred_lock);
2615 submit_bio_noacct(bio);
2620 static void dm_queue_flush(struct mapped_device *md)
2622 clear_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2623 smp_mb__after_atomic();
2624 queue_work(md->wq, &md->work);
2628 * Swap in a new table, returning the old one for the caller to destroy.
2630 struct dm_table *dm_swap_table(struct mapped_device *md, struct dm_table *table)
2632 struct dm_table *live_map = NULL, *map = ERR_PTR(-EINVAL);
2633 struct queue_limits limits;
2636 mutex_lock(&md->suspend_lock);
2638 /* device must be suspended */
2639 if (!dm_suspended_md(md))
2643 * If the new table has no data devices, retain the existing limits.
2644 * This helps multipath with queue_if_no_path if all paths disappear,
2645 * then new I/O is queued based on these limits, and then some paths
2648 if (dm_table_has_no_data_devices(table)) {
2649 live_map = dm_get_live_table_fast(md);
2651 limits = md->queue->limits;
2652 dm_put_live_table_fast(md);
2656 r = dm_calculate_queue_limits(table, &limits);
2663 map = __bind(md, table, &limits);
2664 dm_issue_global_event();
2667 mutex_unlock(&md->suspend_lock);
2672 * Functions to lock and unlock any filesystem running on the
2675 static int lock_fs(struct mapped_device *md)
2679 WARN_ON(test_bit(DMF_FROZEN, &md->flags));
2681 r = bdev_freeze(md->disk->part0);
2683 set_bit(DMF_FROZEN, &md->flags);
2687 static void unlock_fs(struct mapped_device *md)
2689 if (!test_bit(DMF_FROZEN, &md->flags))
2691 bdev_thaw(md->disk->part0);
2692 clear_bit(DMF_FROZEN, &md->flags);
2696 * @suspend_flags: DM_SUSPEND_LOCKFS_FLAG and/or DM_SUSPEND_NOFLUSH_FLAG
2697 * @task_state: e.g. TASK_INTERRUPTIBLE or TASK_UNINTERRUPTIBLE
2698 * @dmf_suspended_flag: DMF_SUSPENDED or DMF_SUSPENDED_INTERNALLY
2700 * If __dm_suspend returns 0, the device is completely quiescent
2701 * now. There is no request-processing activity. All new requests
2702 * are being added to md->deferred list.
2704 static int __dm_suspend(struct mapped_device *md, struct dm_table *map,
2705 unsigned int suspend_flags, unsigned int task_state,
2706 int dmf_suspended_flag)
2708 bool do_lockfs = suspend_flags & DM_SUSPEND_LOCKFS_FLAG;
2709 bool noflush = suspend_flags & DM_SUSPEND_NOFLUSH_FLAG;
2712 lockdep_assert_held(&md->suspend_lock);
2715 * DMF_NOFLUSH_SUSPENDING must be set before presuspend.
2716 * This flag is cleared before dm_suspend returns.
2719 set_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
2721 DMDEBUG("%s: suspending with flush", dm_device_name(md));
2724 * This gets reverted if there's an error later and the targets
2725 * provide the .presuspend_undo hook.
2727 dm_table_presuspend_targets(map);
2730 * Flush I/O to the device.
2731 * Any I/O submitted after lock_fs() may not be flushed.
2732 * noflush takes precedence over do_lockfs.
2733 * (lock_fs() flushes I/Os and waits for them to complete.)
2735 if (!noflush && do_lockfs) {
2738 dm_table_presuspend_undo_targets(map);
2744 * Here we must make sure that no processes are submitting requests
2745 * to target drivers i.e. no one may be executing
2746 * dm_split_and_process_bio from dm_submit_bio.
2748 * To get all processes out of dm_split_and_process_bio in dm_submit_bio,
2749 * we take the write lock. To prevent any process from reentering
2750 * dm_split_and_process_bio from dm_submit_bio and quiesce the thread
2751 * (dm_wq_work), we set DMF_BLOCK_IO_FOR_SUSPEND and call
2752 * flush_workqueue(md->wq).
2754 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2756 synchronize_srcu(&md->io_barrier);
2759 * Stop md->queue before flushing md->wq in case request-based
2760 * dm defers requests to md->wq from md->queue.
2762 if (dm_request_based(md))
2763 dm_stop_queue(md->queue);
2765 flush_workqueue(md->wq);
2768 * At this point no more requests are entering target request routines.
2769 * We call dm_wait_for_completion to wait for all existing requests
2772 r = dm_wait_for_completion(md, task_state);
2774 set_bit(dmf_suspended_flag, &md->flags);
2777 clear_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
2779 synchronize_srcu(&md->io_barrier);
2781 /* were we interrupted ? */
2785 if (dm_request_based(md))
2786 dm_start_queue(md->queue);
2789 dm_table_presuspend_undo_targets(map);
2790 /* pushback list is already flushed, so skip flush */
2797 * We need to be able to change a mapping table under a mounted
2798 * filesystem. For example we might want to move some data in
2799 * the background. Before the table can be swapped with
2800 * dm_bind_table, dm_suspend must be called to flush any in
2801 * flight bios and ensure that any further io gets deferred.
2804 * Suspend mechanism in request-based dm.
2806 * 1. Flush all I/Os by lock_fs() if needed.
2807 * 2. Stop dispatching any I/O by stopping the request_queue.
2808 * 3. Wait for all in-flight I/Os to be completed or requeued.
2810 * To abort suspend, start the request_queue.
2812 int dm_suspend(struct mapped_device *md, unsigned int suspend_flags)
2814 struct dm_table *map = NULL;
2818 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
2820 if (dm_suspended_md(md)) {
2825 if (dm_suspended_internally_md(md)) {
2826 /* already internally suspended, wait for internal resume */
2827 mutex_unlock(&md->suspend_lock);
2828 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
2834 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2836 /* avoid deadlock with fs/namespace.c:do_mount() */
2837 suspend_flags &= ~DM_SUSPEND_LOCKFS_FLAG;
2840 r = __dm_suspend(md, map, suspend_flags, TASK_INTERRUPTIBLE, DMF_SUSPENDED);
2844 set_bit(DMF_POST_SUSPENDING, &md->flags);
2845 dm_table_postsuspend_targets(map);
2846 clear_bit(DMF_POST_SUSPENDING, &md->flags);
2849 mutex_unlock(&md->suspend_lock);
2853 static int __dm_resume(struct mapped_device *md, struct dm_table *map)
2856 int r = dm_table_resume_targets(map);
2865 * Flushing deferred I/Os must be done after targets are resumed
2866 * so that mapping of targets can work correctly.
2867 * Request-based dm is queueing the deferred I/Os in its request_queue.
2869 if (dm_request_based(md))
2870 dm_start_queue(md->queue);
2877 int dm_resume(struct mapped_device *md)
2880 struct dm_table *map = NULL;
2884 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
2886 if (!dm_suspended_md(md))
2889 if (dm_suspended_internally_md(md)) {
2890 /* already internally suspended, wait for internal resume */
2891 mutex_unlock(&md->suspend_lock);
2892 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
2898 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2899 if (!map || !dm_table_get_size(map))
2902 r = __dm_resume(md, map);
2906 clear_bit(DMF_SUSPENDED, &md->flags);
2908 mutex_unlock(&md->suspend_lock);
2914 * Internal suspend/resume works like userspace-driven suspend. It waits
2915 * until all bios finish and prevents issuing new bios to the target drivers.
2916 * It may be used only from the kernel.
2919 static void __dm_internal_suspend(struct mapped_device *md, unsigned int suspend_flags)
2921 struct dm_table *map = NULL;
2923 lockdep_assert_held(&md->suspend_lock);
2925 if (md->internal_suspend_count++)
2926 return; /* nested internal suspend */
2928 if (dm_suspended_md(md)) {
2929 set_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
2930 return; /* nest suspend */
2933 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2936 * Using TASK_UNINTERRUPTIBLE because only NOFLUSH internal suspend is
2937 * supported. Properly supporting a TASK_INTERRUPTIBLE internal suspend
2938 * would require changing .presuspend to return an error -- avoid this
2939 * until there is a need for more elaborate variants of internal suspend.
2941 (void) __dm_suspend(md, map, suspend_flags, TASK_UNINTERRUPTIBLE,
2942 DMF_SUSPENDED_INTERNALLY);
2944 set_bit(DMF_POST_SUSPENDING, &md->flags);
2945 dm_table_postsuspend_targets(map);
2946 clear_bit(DMF_POST_SUSPENDING, &md->flags);
2949 static void __dm_internal_resume(struct mapped_device *md)
2952 struct dm_table *map;
2954 BUG_ON(!md->internal_suspend_count);
2956 if (--md->internal_suspend_count)
2957 return; /* resume from nested internal suspend */
2959 if (dm_suspended_md(md))
2960 goto done; /* resume from nested suspend */
2962 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2963 r = __dm_resume(md, map);
2966 * If a preresume method of some target failed, we are in a
2967 * tricky situation. We can't return an error to the caller. We
2968 * can't fake success because then the "resume" and
2969 * "postsuspend" methods would not be paired correctly, and it
2970 * would break various targets, for example it would cause list
2971 * corruption in the "origin" target.
2973 * So, we fake normal suspend here, to make sure that the
2974 * "resume" and "postsuspend" methods will be paired correctly.
2976 DMERR("Preresume method failed: %d", r);
2977 set_bit(DMF_SUSPENDED, &md->flags);
2980 clear_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
2981 smp_mb__after_atomic();
2982 wake_up_bit(&md->flags, DMF_SUSPENDED_INTERNALLY);
2985 void dm_internal_suspend_noflush(struct mapped_device *md)
2987 mutex_lock(&md->suspend_lock);
2988 __dm_internal_suspend(md, DM_SUSPEND_NOFLUSH_FLAG);
2989 mutex_unlock(&md->suspend_lock);
2991 EXPORT_SYMBOL_GPL(dm_internal_suspend_noflush);
2993 void dm_internal_resume(struct mapped_device *md)
2995 mutex_lock(&md->suspend_lock);
2996 __dm_internal_resume(md);
2997 mutex_unlock(&md->suspend_lock);
2999 EXPORT_SYMBOL_GPL(dm_internal_resume);
3002 * Fast variants of internal suspend/resume hold md->suspend_lock,
3003 * which prevents interaction with userspace-driven suspend.
3006 void dm_internal_suspend_fast(struct mapped_device *md)
3008 mutex_lock(&md->suspend_lock);
3009 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
3012 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
3013 synchronize_srcu(&md->io_barrier);
3014 flush_workqueue(md->wq);
3015 dm_wait_for_completion(md, TASK_UNINTERRUPTIBLE);
3017 EXPORT_SYMBOL_GPL(dm_internal_suspend_fast);
3019 void dm_internal_resume_fast(struct mapped_device *md)
3021 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
3027 mutex_unlock(&md->suspend_lock);
3029 EXPORT_SYMBOL_GPL(dm_internal_resume_fast);
3032 *---------------------------------------------------------------
3033 * Event notification.
3034 *---------------------------------------------------------------
3036 int dm_kobject_uevent(struct mapped_device *md, enum kobject_action action,
3037 unsigned int cookie, bool need_resize_uevent)
3040 unsigned int noio_flag;
3041 char udev_cookie[DM_COOKIE_LENGTH];
3042 char *envp[3] = { NULL, NULL, NULL };
3043 char **envpp = envp;
3045 snprintf(udev_cookie, DM_COOKIE_LENGTH, "%s=%u",
3046 DM_COOKIE_ENV_VAR_NAME, cookie);
3047 *envpp++ = udev_cookie;
3049 if (need_resize_uevent) {
3050 *envpp++ = "RESIZE=1";
3053 noio_flag = memalloc_noio_save();
3055 r = kobject_uevent_env(&disk_to_dev(md->disk)->kobj, action, envp);
3057 memalloc_noio_restore(noio_flag);
3062 uint32_t dm_next_uevent_seq(struct mapped_device *md)
3064 return atomic_add_return(1, &md->uevent_seq);
3067 uint32_t dm_get_event_nr(struct mapped_device *md)
3069 return atomic_read(&md->event_nr);
3072 int dm_wait_event(struct mapped_device *md, int event_nr)
3074 return wait_event_interruptible(md->eventq,
3075 (event_nr != atomic_read(&md->event_nr)));
3078 void dm_uevent_add(struct mapped_device *md, struct list_head *elist)
3080 unsigned long flags;
3082 spin_lock_irqsave(&md->uevent_lock, flags);
3083 list_add(elist, &md->uevent_list);
3084 spin_unlock_irqrestore(&md->uevent_lock, flags);
3088 * The gendisk is only valid as long as you have a reference
3091 struct gendisk *dm_disk(struct mapped_device *md)
3095 EXPORT_SYMBOL_GPL(dm_disk);
3097 struct kobject *dm_kobject(struct mapped_device *md)
3099 return &md->kobj_holder.kobj;
3102 struct mapped_device *dm_get_from_kobject(struct kobject *kobj)
3104 struct mapped_device *md;
3106 md = container_of(kobj, struct mapped_device, kobj_holder.kobj);
3108 spin_lock(&_minor_lock);
3109 if (test_bit(DMF_FREEING, &md->flags) || dm_deleting_md(md)) {
3115 spin_unlock(&_minor_lock);
3120 int dm_suspended_md(struct mapped_device *md)
3122 return test_bit(DMF_SUSPENDED, &md->flags);
3125 static int dm_post_suspending_md(struct mapped_device *md)
3127 return test_bit(DMF_POST_SUSPENDING, &md->flags);
3130 int dm_suspended_internally_md(struct mapped_device *md)
3132 return test_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3135 int dm_test_deferred_remove_flag(struct mapped_device *md)
3137 return test_bit(DMF_DEFERRED_REMOVE, &md->flags);
3140 int dm_suspended(struct dm_target *ti)
3142 return dm_suspended_md(ti->table->md);
3144 EXPORT_SYMBOL_GPL(dm_suspended);
3146 int dm_post_suspending(struct dm_target *ti)
3148 return dm_post_suspending_md(ti->table->md);
3150 EXPORT_SYMBOL_GPL(dm_post_suspending);
3152 int dm_noflush_suspending(struct dm_target *ti)
3154 return __noflush_suspending(ti->table->md);
3156 EXPORT_SYMBOL_GPL(dm_noflush_suspending);
3158 void dm_free_md_mempools(struct dm_md_mempools *pools)
3163 bioset_exit(&pools->bs);
3164 bioset_exit(&pools->io_bs);
3177 struct pr_keys *read_keys;
3178 struct pr_held_reservation *rsv;
3181 static int dm_call_pr(struct block_device *bdev, iterate_devices_callout_fn fn,
3184 struct mapped_device *md = bdev->bd_disk->private_data;
3185 struct dm_table *table;
3186 struct dm_target *ti;
3187 int ret = -ENOTTY, srcu_idx;
3189 table = dm_get_live_table(md, &srcu_idx);
3190 if (!table || !dm_table_get_size(table))
3193 /* We only support devices that have a single target */
3194 if (table->num_targets != 1)
3196 ti = dm_table_get_target(table, 0);
3198 if (dm_suspended_md(md)) {
3204 if (!ti->type->iterate_devices)
3207 ti->type->iterate_devices(ti, fn, pr);
3210 dm_put_live_table(md, srcu_idx);
3215 * For register / unregister we need to manually call out to every path.
3217 static int __dm_pr_register(struct dm_target *ti, struct dm_dev *dev,
3218 sector_t start, sector_t len, void *data)
3220 struct dm_pr *pr = data;
3221 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3224 if (!ops || !ops->pr_register) {
3225 pr->ret = -EOPNOTSUPP;
3229 ret = ops->pr_register(dev->bdev, pr->old_key, pr->new_key, pr->flags);
3242 static int dm_pr_register(struct block_device *bdev, u64 old_key, u64 new_key,
3254 ret = dm_call_pr(bdev, __dm_pr_register, &pr);
3256 /* Didn't even get to register a path */
3267 /* unregister all paths if we failed to register any path */
3268 pr.old_key = new_key;
3271 pr.fail_early = false;
3272 (void) dm_call_pr(bdev, __dm_pr_register, &pr);
3277 static int __dm_pr_reserve(struct dm_target *ti, struct dm_dev *dev,
3278 sector_t start, sector_t len, void *data)
3280 struct dm_pr *pr = data;
3281 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3283 if (!ops || !ops->pr_reserve) {
3284 pr->ret = -EOPNOTSUPP;
3288 pr->ret = ops->pr_reserve(dev->bdev, pr->old_key, pr->type, pr->flags);
3295 static int dm_pr_reserve(struct block_device *bdev, u64 key, enum pr_type type,
3302 .fail_early = false,
3307 ret = dm_call_pr(bdev, __dm_pr_reserve, &pr);
3315 * If there is a non-All Registrants type of reservation, the release must be
3316 * sent down the holding path. For the cases where there is no reservation or
3317 * the path is not the holder the device will also return success, so we must
3318 * try each path to make sure we got the correct path.
3320 static int __dm_pr_release(struct dm_target *ti, struct dm_dev *dev,
3321 sector_t start, sector_t len, void *data)
3323 struct dm_pr *pr = data;
3324 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3326 if (!ops || !ops->pr_release) {
3327 pr->ret = -EOPNOTSUPP;
3331 pr->ret = ops->pr_release(dev->bdev, pr->old_key, pr->type);
3338 static int dm_pr_release(struct block_device *bdev, u64 key, enum pr_type type)
3343 .fail_early = false,
3347 ret = dm_call_pr(bdev, __dm_pr_release, &pr);
3354 static int __dm_pr_preempt(struct dm_target *ti, struct dm_dev *dev,
3355 sector_t start, sector_t len, void *data)
3357 struct dm_pr *pr = data;
3358 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3360 if (!ops || !ops->pr_preempt) {
3361 pr->ret = -EOPNOTSUPP;
3365 pr->ret = ops->pr_preempt(dev->bdev, pr->old_key, pr->new_key, pr->type,
3373 static int dm_pr_preempt(struct block_device *bdev, u64 old_key, u64 new_key,
3374 enum pr_type type, bool abort)
3380 .fail_early = false,
3384 ret = dm_call_pr(bdev, __dm_pr_preempt, &pr);
3391 static int dm_pr_clear(struct block_device *bdev, u64 key)
3393 struct mapped_device *md = bdev->bd_disk->private_data;
3394 const struct pr_ops *ops;
3397 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
3401 ops = bdev->bd_disk->fops->pr_ops;
3402 if (ops && ops->pr_clear)
3403 r = ops->pr_clear(bdev, key);
3407 dm_unprepare_ioctl(md, srcu_idx);
3411 static int __dm_pr_read_keys(struct dm_target *ti, struct dm_dev *dev,
3412 sector_t start, sector_t len, void *data)
3414 struct dm_pr *pr = data;
3415 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3417 if (!ops || !ops->pr_read_keys) {
3418 pr->ret = -EOPNOTSUPP;
3422 pr->ret = ops->pr_read_keys(dev->bdev, pr->read_keys);
3429 static int dm_pr_read_keys(struct block_device *bdev, struct pr_keys *keys)
3436 ret = dm_call_pr(bdev, __dm_pr_read_keys, &pr);
3443 static int __dm_pr_read_reservation(struct dm_target *ti, struct dm_dev *dev,
3444 sector_t start, sector_t len, void *data)
3446 struct dm_pr *pr = data;
3447 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3449 if (!ops || !ops->pr_read_reservation) {
3450 pr->ret = -EOPNOTSUPP;
3454 pr->ret = ops->pr_read_reservation(dev->bdev, pr->rsv);
3461 static int dm_pr_read_reservation(struct block_device *bdev,
3462 struct pr_held_reservation *rsv)
3469 ret = dm_call_pr(bdev, __dm_pr_read_reservation, &pr);
3476 static const struct pr_ops dm_pr_ops = {
3477 .pr_register = dm_pr_register,
3478 .pr_reserve = dm_pr_reserve,
3479 .pr_release = dm_pr_release,
3480 .pr_preempt = dm_pr_preempt,
3481 .pr_clear = dm_pr_clear,
3482 .pr_read_keys = dm_pr_read_keys,
3483 .pr_read_reservation = dm_pr_read_reservation,
3486 static const struct block_device_operations dm_blk_dops = {
3487 .submit_bio = dm_submit_bio,
3488 .poll_bio = dm_poll_bio,
3489 .open = dm_blk_open,
3490 .release = dm_blk_close,
3491 .ioctl = dm_blk_ioctl,
3492 .getgeo = dm_blk_getgeo,
3493 .report_zones = dm_blk_report_zones,
3494 .pr_ops = &dm_pr_ops,
3495 .owner = THIS_MODULE
3498 static const struct block_device_operations dm_rq_blk_dops = {
3499 .open = dm_blk_open,
3500 .release = dm_blk_close,
3501 .ioctl = dm_blk_ioctl,
3502 .getgeo = dm_blk_getgeo,
3503 .pr_ops = &dm_pr_ops,
3504 .owner = THIS_MODULE
3507 static const struct dax_operations dm_dax_ops = {
3508 .direct_access = dm_dax_direct_access,
3509 .zero_page_range = dm_dax_zero_page_range,
3510 .recovery_write = dm_dax_recovery_write,
3516 module_init(dm_init);
3517 module_exit(dm_exit);
3519 module_param(major, uint, 0);
3520 MODULE_PARM_DESC(major, "The major number of the device mapper");
3522 module_param(reserved_bio_based_ios, uint, 0644);
3523 MODULE_PARM_DESC(reserved_bio_based_ios, "Reserved IOs in bio-based mempools");
3525 module_param(dm_numa_node, int, 0644);
3526 MODULE_PARM_DESC(dm_numa_node, "NUMA node for DM device memory allocations");
3528 module_param(swap_bios, int, 0644);
3529 MODULE_PARM_DESC(swap_bios, "Maximum allowed inflight swap IOs");
3531 MODULE_DESCRIPTION(DM_NAME " driver");
3532 MODULE_AUTHOR("Joe Thornber <dm-devel@lists.linux.dev>");
3533 MODULE_LICENSE("GPL");