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/bio-integrity.h>
15 #include <linux/init.h>
16 #include <linux/module.h>
17 #include <linux/mutex.h>
18 #include <linux/sched/mm.h>
19 #include <linux/sched/signal.h>
20 #include <linux/blkpg.h>
21 #include <linux/bio.h>
22 #include <linux/mempool.h>
23 #include <linux/dax.h>
24 #include <linux/slab.h>
25 #include <linux/idr.h>
26 #include <linux/uio.h>
27 #include <linux/hdreg.h>
28 #include <linux/delay.h>
29 #include <linux/wait.h>
31 #include <linux/refcount.h>
32 #include <linux/part_stat.h>
33 #include <linux/blk-crypto.h>
34 #include <linux/blk-crypto-profile.h>
36 #define DM_MSG_PREFIX "core"
39 * Cookies are numeric values sent with CHANGE and REMOVE
40 * uevents while resuming, removing or renaming the device.
42 #define DM_COOKIE_ENV_VAR_NAME "DM_COOKIE"
43 #define DM_COOKIE_LENGTH 24
46 * For REQ_POLLED fs bio, this flag is set if we link mapped underlying
47 * dm_io into one list, and reuse bio->bi_private as the list head. Before
48 * ending this fs bio, we will recover its ->bi_private.
50 #define REQ_DM_POLL_LIST REQ_DRV
52 static const char *_name = DM_NAME;
54 static unsigned int major;
55 static unsigned int _major;
57 static DEFINE_IDR(_minor_idr);
59 static DEFINE_SPINLOCK(_minor_lock);
61 static void do_deferred_remove(struct work_struct *w);
63 static DECLARE_WORK(deferred_remove_work, do_deferred_remove);
65 static struct workqueue_struct *deferred_remove_workqueue;
67 atomic_t dm_global_event_nr = ATOMIC_INIT(0);
68 DECLARE_WAIT_QUEUE_HEAD(dm_global_eventq);
70 void dm_issue_global_event(void)
72 atomic_inc(&dm_global_event_nr);
73 wake_up(&dm_global_eventq);
76 DEFINE_STATIC_KEY_FALSE(stats_enabled);
77 DEFINE_STATIC_KEY_FALSE(swap_bios_enabled);
78 DEFINE_STATIC_KEY_FALSE(zoned_enabled);
81 * One of these is allocated (on-stack) per original bio.
88 unsigned int sector_count;
89 bool is_abnormal_io:1;
90 bool submit_as_polled:1;
93 static inline struct dm_target_io *clone_to_tio(struct bio *clone)
95 return container_of(clone, struct dm_target_io, clone);
98 void *dm_per_bio_data(struct bio *bio, size_t data_size)
100 if (!dm_tio_flagged(clone_to_tio(bio), DM_TIO_INSIDE_DM_IO))
101 return (char *)bio - DM_TARGET_IO_BIO_OFFSET - data_size;
102 return (char *)bio - DM_IO_BIO_OFFSET - data_size;
104 EXPORT_SYMBOL_GPL(dm_per_bio_data);
106 struct bio *dm_bio_from_per_bio_data(void *data, size_t data_size)
108 struct dm_io *io = (struct dm_io *)((char *)data + data_size);
110 if (io->magic == DM_IO_MAGIC)
111 return (struct bio *)((char *)io + DM_IO_BIO_OFFSET);
112 BUG_ON(io->magic != DM_TIO_MAGIC);
113 return (struct bio *)((char *)io + DM_TARGET_IO_BIO_OFFSET);
115 EXPORT_SYMBOL_GPL(dm_bio_from_per_bio_data);
117 unsigned int dm_bio_get_target_bio_nr(const struct bio *bio)
119 return container_of(bio, struct dm_target_io, clone)->target_bio_nr;
121 EXPORT_SYMBOL_GPL(dm_bio_get_target_bio_nr);
123 #define MINOR_ALLOCED ((void *)-1)
125 #define DM_NUMA_NODE NUMA_NO_NODE
126 static int dm_numa_node = DM_NUMA_NODE;
128 #define DEFAULT_SWAP_BIOS (8 * 1048576 / PAGE_SIZE)
129 static int swap_bios = DEFAULT_SWAP_BIOS;
130 static int get_swap_bios(void)
132 int latch = READ_ONCE(swap_bios);
134 if (unlikely(latch <= 0))
135 latch = DEFAULT_SWAP_BIOS;
139 struct table_device {
140 struct list_head list;
142 struct dm_dev dm_dev;
146 * Bio-based DM's mempools' reserved IOs set by the user.
148 #define RESERVED_BIO_BASED_IOS 16
149 static unsigned int reserved_bio_based_ios = RESERVED_BIO_BASED_IOS;
151 static int __dm_get_module_param_int(int *module_param, int min, int max)
153 int param = READ_ONCE(*module_param);
154 int modified_param = 0;
155 bool modified = true;
158 modified_param = min;
159 else if (param > max)
160 modified_param = max;
165 (void)cmpxchg(module_param, param, modified_param);
166 param = modified_param;
172 unsigned int __dm_get_module_param(unsigned int *module_param, unsigned int def, unsigned int max)
174 unsigned int param = READ_ONCE(*module_param);
175 unsigned int modified_param = 0;
178 modified_param = def;
179 else if (param > max)
180 modified_param = max;
182 if (modified_param) {
183 (void)cmpxchg(module_param, param, modified_param);
184 param = modified_param;
190 unsigned int dm_get_reserved_bio_based_ios(void)
192 return __dm_get_module_param(&reserved_bio_based_ios,
193 RESERVED_BIO_BASED_IOS, DM_RESERVED_MAX_IOS);
195 EXPORT_SYMBOL_GPL(dm_get_reserved_bio_based_ios);
197 static unsigned int dm_get_numa_node(void)
199 return __dm_get_module_param_int(&dm_numa_node,
200 DM_NUMA_NODE, num_online_nodes() - 1);
203 static int __init local_init(void)
207 r = dm_uevent_init();
211 deferred_remove_workqueue = alloc_ordered_workqueue("kdmremove", 0);
212 if (!deferred_remove_workqueue) {
214 goto out_uevent_exit;
218 r = register_blkdev(_major, _name);
220 goto out_free_workqueue;
228 destroy_workqueue(deferred_remove_workqueue);
235 static void local_exit(void)
237 destroy_workqueue(deferred_remove_workqueue);
239 unregister_blkdev(_major, _name);
244 DMINFO("cleaned up");
247 static int (*_inits[])(void) __initdata = {
258 static void (*_exits[])(void) = {
269 static int __init dm_init(void)
271 const int count = ARRAY_SIZE(_inits);
274 #if (IS_ENABLED(CONFIG_IMA) && !IS_ENABLED(CONFIG_IMA_DISABLE_HTABLE))
275 DMWARN("CONFIG_IMA_DISABLE_HTABLE is disabled."
276 " Duplicate IMA measurements will not be recorded in the IMA log.");
279 for (i = 0; i < count; i++) {
293 static void __exit dm_exit(void)
295 int i = ARRAY_SIZE(_exits);
301 * Should be empty by this point.
303 idr_destroy(&_minor_idr);
307 * Block device functions
309 int dm_deleting_md(struct mapped_device *md)
311 return test_bit(DMF_DELETING, &md->flags);
314 static int dm_blk_open(struct gendisk *disk, blk_mode_t mode)
316 struct mapped_device *md;
318 spin_lock(&_minor_lock);
320 md = disk->private_data;
324 if (test_bit(DMF_FREEING, &md->flags) ||
325 dm_deleting_md(md)) {
331 atomic_inc(&md->open_count);
333 spin_unlock(&_minor_lock);
335 return md ? 0 : -ENXIO;
338 static void dm_blk_close(struct gendisk *disk)
340 struct mapped_device *md;
342 spin_lock(&_minor_lock);
344 md = disk->private_data;
348 if (atomic_dec_and_test(&md->open_count) &&
349 (test_bit(DMF_DEFERRED_REMOVE, &md->flags)))
350 queue_work(deferred_remove_workqueue, &deferred_remove_work);
354 spin_unlock(&_minor_lock);
357 int dm_open_count(struct mapped_device *md)
359 return atomic_read(&md->open_count);
363 * Guarantees nothing is using the device before it's deleted.
365 int dm_lock_for_deletion(struct mapped_device *md, bool mark_deferred, bool only_deferred)
369 spin_lock(&_minor_lock);
371 if (dm_open_count(md)) {
374 set_bit(DMF_DEFERRED_REMOVE, &md->flags);
375 } else if (only_deferred && !test_bit(DMF_DEFERRED_REMOVE, &md->flags))
378 set_bit(DMF_DELETING, &md->flags);
380 spin_unlock(&_minor_lock);
385 int dm_cancel_deferred_remove(struct mapped_device *md)
389 spin_lock(&_minor_lock);
391 if (test_bit(DMF_DELETING, &md->flags))
394 clear_bit(DMF_DEFERRED_REMOVE, &md->flags);
396 spin_unlock(&_minor_lock);
401 static void do_deferred_remove(struct work_struct *w)
403 dm_deferred_remove();
406 static int dm_blk_getgeo(struct block_device *bdev, struct hd_geometry *geo)
408 struct mapped_device *md = bdev->bd_disk->private_data;
410 return dm_get_geometry(md, geo);
413 static int dm_prepare_ioctl(struct mapped_device *md, int *srcu_idx,
414 struct block_device **bdev)
416 struct dm_target *ti;
417 struct dm_table *map;
422 map = dm_get_live_table(md, srcu_idx);
423 if (!map || !dm_table_get_size(map))
426 /* We only support devices that have a single target */
427 if (map->num_targets != 1)
430 ti = dm_table_get_target(map, 0);
431 if (!ti->type->prepare_ioctl)
434 if (dm_suspended_md(md))
437 r = ti->type->prepare_ioctl(ti, bdev);
438 if (r == -ENOTCONN && !fatal_signal_pending(current)) {
439 dm_put_live_table(md, *srcu_idx);
447 static void dm_unprepare_ioctl(struct mapped_device *md, int srcu_idx)
449 dm_put_live_table(md, srcu_idx);
452 static int dm_blk_ioctl(struct block_device *bdev, blk_mode_t mode,
453 unsigned int cmd, unsigned long arg)
455 struct mapped_device *md = bdev->bd_disk->private_data;
458 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
464 * Target determined this ioctl is being issued against a
465 * subset of the parent bdev; require extra privileges.
467 if (!capable(CAP_SYS_RAWIO)) {
469 "%s: sending ioctl %x to DM device without required privilege.",
476 if (!bdev->bd_disk->fops->ioctl)
479 r = bdev->bd_disk->fops->ioctl(bdev, mode, cmd, arg);
481 dm_unprepare_ioctl(md, srcu_idx);
485 u64 dm_start_time_ns_from_clone(struct bio *bio)
487 return jiffies_to_nsecs(clone_to_tio(bio)->io->start_time);
489 EXPORT_SYMBOL_GPL(dm_start_time_ns_from_clone);
491 static inline bool bio_is_flush_with_data(struct bio *bio)
493 return ((bio->bi_opf & REQ_PREFLUSH) && bio->bi_iter.bi_size);
496 static inline unsigned int dm_io_sectors(struct dm_io *io, struct bio *bio)
499 * If REQ_PREFLUSH set, don't account payload, it will be
500 * submitted (and accounted) after this flush completes.
502 if (bio_is_flush_with_data(bio))
504 if (unlikely(dm_io_flagged(io, DM_IO_WAS_SPLIT)))
506 return bio_sectors(bio);
509 static void dm_io_acct(struct dm_io *io, bool end)
511 struct bio *bio = io->orig_bio;
513 if (dm_io_flagged(io, DM_IO_BLK_STAT)) {
515 bdev_start_io_acct(bio->bi_bdev, bio_op(bio),
518 bdev_end_io_acct(bio->bi_bdev, bio_op(bio),
519 dm_io_sectors(io, bio),
523 if (static_branch_unlikely(&stats_enabled) &&
524 unlikely(dm_stats_used(&io->md->stats))) {
527 if (unlikely(dm_io_flagged(io, DM_IO_WAS_SPLIT)))
528 sector = bio_end_sector(bio) - io->sector_offset;
530 sector = bio->bi_iter.bi_sector;
532 dm_stats_account_io(&io->md->stats, bio_data_dir(bio),
533 sector, dm_io_sectors(io, bio),
534 end, io->start_time, &io->stats_aux);
538 static void __dm_start_io_acct(struct dm_io *io)
540 dm_io_acct(io, false);
543 static void dm_start_io_acct(struct dm_io *io, struct bio *clone)
546 * Ensure IO accounting is only ever started once.
548 if (dm_io_flagged(io, DM_IO_ACCOUNTED))
551 /* Expect no possibility for race unless DM_TIO_IS_DUPLICATE_BIO. */
552 if (!clone || likely(dm_tio_is_normal(clone_to_tio(clone)))) {
553 dm_io_set_flag(io, DM_IO_ACCOUNTED);
556 /* Can afford locking given DM_TIO_IS_DUPLICATE_BIO */
557 spin_lock_irqsave(&io->lock, flags);
558 if (dm_io_flagged(io, DM_IO_ACCOUNTED)) {
559 spin_unlock_irqrestore(&io->lock, flags);
562 dm_io_set_flag(io, DM_IO_ACCOUNTED);
563 spin_unlock_irqrestore(&io->lock, flags);
566 __dm_start_io_acct(io);
569 static void dm_end_io_acct(struct dm_io *io)
571 dm_io_acct(io, true);
574 static struct dm_io *alloc_io(struct mapped_device *md, struct bio *bio, gfp_t gfp_mask)
577 struct dm_target_io *tio;
580 clone = bio_alloc_clone(NULL, bio, gfp_mask, &md->mempools->io_bs);
581 if (unlikely(!clone))
583 tio = clone_to_tio(clone);
585 dm_tio_set_flag(tio, DM_TIO_INSIDE_DM_IO);
588 io = container_of(tio, struct dm_io, tio);
589 io->magic = DM_IO_MAGIC;
590 io->status = BLK_STS_OK;
592 /* one ref is for submission, the other is for completion */
593 atomic_set(&io->io_count, 2);
594 this_cpu_inc(*md->pending_io);
597 spin_lock_init(&io->lock);
598 io->start_time = jiffies;
600 if (blk_queue_io_stat(md->queue))
601 dm_io_set_flag(io, DM_IO_BLK_STAT);
603 if (static_branch_unlikely(&stats_enabled) &&
604 unlikely(dm_stats_used(&md->stats)))
605 dm_stats_record_start(&md->stats, &io->stats_aux);
610 static void free_io(struct dm_io *io)
612 bio_put(&io->tio.clone);
615 static struct bio *alloc_tio(struct clone_info *ci, struct dm_target *ti,
616 unsigned int target_bio_nr, unsigned int *len, gfp_t gfp_mask)
618 struct mapped_device *md = ci->io->md;
619 struct dm_target_io *tio;
622 if (!ci->io->tio.io) {
623 /* the dm_target_io embedded in ci->io is available */
625 /* alloc_io() already initialized embedded clone */
628 clone = bio_alloc_clone(NULL, ci->bio, gfp_mask,
633 /* REQ_DM_POLL_LIST shouldn't be inherited */
634 clone->bi_opf &= ~REQ_DM_POLL_LIST;
636 tio = clone_to_tio(clone);
637 tio->flags = 0; /* also clears DM_TIO_INSIDE_DM_IO */
640 tio->magic = DM_TIO_MAGIC;
643 tio->target_bio_nr = target_bio_nr;
647 /* Set default bdev, but target must bio_set_dev() before issuing IO */
648 clone->bi_bdev = md->disk->part0;
649 if (likely(ti != NULL) && unlikely(ti->needs_bio_set_dev))
650 bio_set_dev(clone, md->disk->part0);
653 clone->bi_iter.bi_size = to_bytes(*len);
654 if (bio_integrity(clone))
655 bio_integrity_trim(clone);
661 static void free_tio(struct bio *clone)
663 if (dm_tio_flagged(clone_to_tio(clone), DM_TIO_INSIDE_DM_IO))
669 * Add the bio to the list of deferred io.
671 static void queue_io(struct mapped_device *md, struct bio *bio)
675 spin_lock_irqsave(&md->deferred_lock, flags);
676 bio_list_add(&md->deferred, bio);
677 spin_unlock_irqrestore(&md->deferred_lock, flags);
678 queue_work(md->wq, &md->work);
682 * Everyone (including functions in this file), should use this
683 * function to access the md->map field, and make sure they call
684 * dm_put_live_table() when finished.
686 struct dm_table *dm_get_live_table(struct mapped_device *md,
687 int *srcu_idx) __acquires(md->io_barrier)
689 *srcu_idx = srcu_read_lock(&md->io_barrier);
691 return srcu_dereference(md->map, &md->io_barrier);
694 void dm_put_live_table(struct mapped_device *md,
695 int srcu_idx) __releases(md->io_barrier)
697 srcu_read_unlock(&md->io_barrier, srcu_idx);
700 void dm_sync_table(struct mapped_device *md)
702 synchronize_srcu(&md->io_barrier);
703 synchronize_rcu_expedited();
707 * A fast alternative to dm_get_live_table/dm_put_live_table.
708 * The caller must not block between these two functions.
710 static struct dm_table *dm_get_live_table_fast(struct mapped_device *md) __acquires(RCU)
713 return rcu_dereference(md->map);
716 static void dm_put_live_table_fast(struct mapped_device *md) __releases(RCU)
721 static char *_dm_claim_ptr = "I belong to device-mapper";
724 * Open a table device so we can use it as a map destination.
726 static struct table_device *open_table_device(struct mapped_device *md,
727 dev_t dev, blk_mode_t mode)
729 struct table_device *td;
730 struct file *bdev_file;
731 struct block_device *bdev;
735 td = kmalloc_node(sizeof(*td), GFP_KERNEL, md->numa_node_id);
737 return ERR_PTR(-ENOMEM);
738 refcount_set(&td->count, 1);
740 bdev_file = bdev_file_open_by_dev(dev, mode, _dm_claim_ptr, NULL);
741 if (IS_ERR(bdev_file)) {
742 r = PTR_ERR(bdev_file);
746 bdev = file_bdev(bdev_file);
749 * We can be called before the dm disk is added. In that case we can't
750 * register the holder relation here. It will be done once add_disk was
753 if (md->disk->slave_dir) {
754 r = bd_link_disk_holder(bdev, md->disk);
759 td->dm_dev.mode = mode;
760 td->dm_dev.bdev = bdev;
761 td->dm_dev.bdev_file = bdev_file;
762 td->dm_dev.dax_dev = fs_dax_get_by_bdev(bdev, &part_off,
764 format_dev_t(td->dm_dev.name, dev);
765 list_add(&td->list, &md->table_devices);
769 __fput_sync(bdev_file);
776 * Close a table device that we've been using.
778 static void close_table_device(struct table_device *td, struct mapped_device *md)
780 if (md->disk->slave_dir)
781 bd_unlink_disk_holder(td->dm_dev.bdev, md->disk);
783 /* Leverage async fput() if DMF_DEFERRED_REMOVE set */
784 if (unlikely(test_bit(DMF_DEFERRED_REMOVE, &md->flags)))
785 fput(td->dm_dev.bdev_file);
787 __fput_sync(td->dm_dev.bdev_file);
789 put_dax(td->dm_dev.dax_dev);
794 static struct table_device *find_table_device(struct list_head *l, dev_t dev,
797 struct table_device *td;
799 list_for_each_entry(td, l, list)
800 if (td->dm_dev.bdev->bd_dev == dev && td->dm_dev.mode == mode)
806 int dm_get_table_device(struct mapped_device *md, dev_t dev, blk_mode_t mode,
807 struct dm_dev **result)
809 struct table_device *td;
811 mutex_lock(&md->table_devices_lock);
812 td = find_table_device(&md->table_devices, dev, mode);
814 td = open_table_device(md, dev, mode);
816 mutex_unlock(&md->table_devices_lock);
820 refcount_inc(&td->count);
822 mutex_unlock(&md->table_devices_lock);
824 *result = &td->dm_dev;
828 void dm_put_table_device(struct mapped_device *md, struct dm_dev *d)
830 struct table_device *td = container_of(d, struct table_device, dm_dev);
832 mutex_lock(&md->table_devices_lock);
833 if (refcount_dec_and_test(&td->count))
834 close_table_device(td, md);
835 mutex_unlock(&md->table_devices_lock);
839 * Get the geometry associated with a dm device
841 int dm_get_geometry(struct mapped_device *md, struct hd_geometry *geo)
849 * Set the geometry of a device.
851 int dm_set_geometry(struct mapped_device *md, struct hd_geometry *geo)
853 sector_t sz = (sector_t)geo->cylinders * geo->heads * geo->sectors;
855 if (geo->start > sz) {
856 DMERR("Start sector is beyond the geometry limits.");
865 static int __noflush_suspending(struct mapped_device *md)
867 return test_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
870 static void dm_requeue_add_io(struct dm_io *io, bool first_stage)
872 struct mapped_device *md = io->md;
875 struct dm_io *next = md->requeue_list;
877 md->requeue_list = io;
880 bio_list_add_head(&md->deferred, io->orig_bio);
884 static void dm_kick_requeue(struct mapped_device *md, bool first_stage)
887 queue_work(md->wq, &md->requeue_work);
889 queue_work(md->wq, &md->work);
893 * Return true if the dm_io's original bio is requeued.
894 * io->status is updated with error if requeue disallowed.
896 static bool dm_handle_requeue(struct dm_io *io, bool first_stage)
898 struct bio *bio = io->orig_bio;
899 bool handle_requeue = (io->status == BLK_STS_DM_REQUEUE);
900 bool handle_polled_eagain = ((io->status == BLK_STS_AGAIN) &&
901 (bio->bi_opf & REQ_POLLED));
902 struct mapped_device *md = io->md;
903 bool requeued = false;
905 if (handle_requeue || handle_polled_eagain) {
908 if (bio->bi_opf & REQ_POLLED) {
910 * Upper layer won't help us poll split bio
911 * (io->orig_bio may only reflect a subset of the
912 * pre-split original) so clear REQ_POLLED.
914 bio_clear_polled(bio);
918 * Target requested pushing back the I/O or
919 * polled IO hit BLK_STS_AGAIN.
921 spin_lock_irqsave(&md->deferred_lock, flags);
922 if ((__noflush_suspending(md) &&
923 !WARN_ON_ONCE(dm_is_zone_write(md, bio))) ||
924 handle_polled_eagain || first_stage) {
925 dm_requeue_add_io(io, first_stage);
929 * noflush suspend was interrupted or this is
930 * a write to a zoned target.
932 io->status = BLK_STS_IOERR;
934 spin_unlock_irqrestore(&md->deferred_lock, flags);
938 dm_kick_requeue(md, first_stage);
943 static void __dm_io_complete(struct dm_io *io, bool first_stage)
945 struct bio *bio = io->orig_bio;
946 struct mapped_device *md = io->md;
947 blk_status_t io_error;
950 requeued = dm_handle_requeue(io, first_stage);
951 if (requeued && first_stage)
954 io_error = io->status;
955 if (dm_io_flagged(io, DM_IO_ACCOUNTED))
957 else if (!io_error) {
959 * Must handle target that DM_MAPIO_SUBMITTED only to
960 * then bio_endio() rather than dm_submit_bio_remap()
962 __dm_start_io_acct(io);
967 this_cpu_dec(*md->pending_io);
969 /* nudge anyone waiting on suspend queue */
970 if (unlikely(wq_has_sleeper(&md->wait)))
973 /* Return early if the original bio was requeued */
977 if (bio_is_flush_with_data(bio)) {
979 * Preflush done for flush with data, reissue
980 * without REQ_PREFLUSH.
982 bio->bi_opf &= ~REQ_PREFLUSH;
985 /* done with normal IO or empty flush */
987 bio->bi_status = io_error;
992 static void dm_wq_requeue_work(struct work_struct *work)
994 struct mapped_device *md = container_of(work, struct mapped_device,
999 /* reuse deferred lock to simplify dm_handle_requeue */
1000 spin_lock_irqsave(&md->deferred_lock, flags);
1001 io = md->requeue_list;
1002 md->requeue_list = NULL;
1003 spin_unlock_irqrestore(&md->deferred_lock, flags);
1006 struct dm_io *next = io->next;
1008 dm_io_rewind(io, &md->disk->bio_split);
1011 __dm_io_complete(io, false);
1018 * Two staged requeue:
1020 * 1) io->orig_bio points to the real original bio, and the part mapped to
1021 * this io must be requeued, instead of other parts of the original bio.
1023 * 2) io->orig_bio points to new cloned bio which matches the requeued dm_io.
1025 static void dm_io_complete(struct dm_io *io)
1030 * Only dm_io that has been split needs two stage requeue, otherwise
1031 * we may run into long bio clone chain during suspend and OOM could
1034 * Also flush data dm_io won't be marked as DM_IO_WAS_SPLIT, so they
1035 * also aren't handled via the first stage requeue.
1037 if (dm_io_flagged(io, DM_IO_WAS_SPLIT))
1038 first_requeue = true;
1040 first_requeue = false;
1042 __dm_io_complete(io, first_requeue);
1046 * Decrements the number of outstanding ios that a bio has been
1047 * cloned into, completing the original io if necc.
1049 static inline void __dm_io_dec_pending(struct dm_io *io)
1051 if (atomic_dec_and_test(&io->io_count))
1055 static void dm_io_set_error(struct dm_io *io, blk_status_t error)
1057 unsigned long flags;
1059 /* Push-back supersedes any I/O errors */
1060 spin_lock_irqsave(&io->lock, flags);
1061 if (!(io->status == BLK_STS_DM_REQUEUE &&
1062 __noflush_suspending(io->md))) {
1065 spin_unlock_irqrestore(&io->lock, flags);
1068 static void dm_io_dec_pending(struct dm_io *io, blk_status_t error)
1070 if (unlikely(error))
1071 dm_io_set_error(io, error);
1073 __dm_io_dec_pending(io);
1077 * The queue_limits are only valid as long as you have a reference
1078 * count on 'md'. But _not_ imposing verification to avoid atomic_read(),
1080 static inline struct queue_limits *dm_get_queue_limits(struct mapped_device *md)
1082 return &md->queue->limits;
1085 void disable_discard(struct mapped_device *md)
1087 struct queue_limits *limits = dm_get_queue_limits(md);
1089 /* device doesn't really support DISCARD, disable it */
1090 limits->max_hw_discard_sectors = 0;
1093 void disable_write_zeroes(struct mapped_device *md)
1095 struct queue_limits *limits = dm_get_queue_limits(md);
1097 /* device doesn't really support WRITE ZEROES, disable it */
1098 limits->max_write_zeroes_sectors = 0;
1101 static bool swap_bios_limit(struct dm_target *ti, struct bio *bio)
1103 return unlikely((bio->bi_opf & REQ_SWAP) != 0) && unlikely(ti->limit_swap_bios);
1106 static void clone_endio(struct bio *bio)
1108 blk_status_t error = bio->bi_status;
1109 struct dm_target_io *tio = clone_to_tio(bio);
1110 struct dm_target *ti = tio->ti;
1111 dm_endio_fn endio = likely(ti != NULL) ? ti->type->end_io : NULL;
1112 struct dm_io *io = tio->io;
1113 struct mapped_device *md = io->md;
1115 if (unlikely(error == BLK_STS_TARGET)) {
1116 if (bio_op(bio) == REQ_OP_DISCARD &&
1117 !bdev_max_discard_sectors(bio->bi_bdev))
1118 disable_discard(md);
1119 else if (bio_op(bio) == REQ_OP_WRITE_ZEROES &&
1120 !bdev_write_zeroes_sectors(bio->bi_bdev))
1121 disable_write_zeroes(md);
1124 if (static_branch_unlikely(&zoned_enabled) &&
1125 unlikely(bdev_is_zoned(bio->bi_bdev)))
1126 dm_zone_endio(io, bio);
1129 int r = endio(ti, bio, &error);
1132 case DM_ENDIO_REQUEUE:
1133 if (static_branch_unlikely(&zoned_enabled)) {
1135 * Requeuing writes to a sequential zone of a zoned
1136 * target will break the sequential write pattern:
1139 if (WARN_ON_ONCE(dm_is_zone_write(md, bio)))
1140 error = BLK_STS_IOERR;
1142 error = BLK_STS_DM_REQUEUE;
1144 error = BLK_STS_DM_REQUEUE;
1148 case DM_ENDIO_INCOMPLETE:
1149 /* The target will handle the io */
1152 DMCRIT("unimplemented target endio return value: %d", r);
1157 if (static_branch_unlikely(&swap_bios_enabled) &&
1158 likely(ti != NULL) && unlikely(swap_bios_limit(ti, bio)))
1159 up(&md->swap_bios_semaphore);
1162 dm_io_dec_pending(io, error);
1166 * Return maximum size of I/O possible at the supplied sector up to the current
1169 static inline sector_t max_io_len_target_boundary(struct dm_target *ti,
1170 sector_t target_offset)
1172 return ti->len - target_offset;
1175 static sector_t __max_io_len(struct dm_target *ti, sector_t sector,
1176 unsigned int max_granularity,
1177 unsigned int max_sectors)
1179 sector_t target_offset = dm_target_offset(ti, sector);
1180 sector_t len = max_io_len_target_boundary(ti, target_offset);
1183 * Does the target need to split IO even further?
1184 * - varied (per target) IO splitting is a tenet of DM; this
1185 * explains why stacked chunk_sectors based splitting via
1186 * bio_split_to_limits() isn't possible here.
1188 if (!max_granularity)
1190 return min_t(sector_t, len,
1191 min(max_sectors ? : queue_max_sectors(ti->table->md->queue),
1192 blk_boundary_sectors_left(target_offset, max_granularity)));
1195 static inline sector_t max_io_len(struct dm_target *ti, sector_t sector)
1197 return __max_io_len(ti, sector, ti->max_io_len, 0);
1200 int dm_set_target_max_io_len(struct dm_target *ti, sector_t len)
1202 if (len > UINT_MAX) {
1203 DMERR("Specified maximum size of target IO (%llu) exceeds limit (%u)",
1204 (unsigned long long)len, UINT_MAX);
1205 ti->error = "Maximum size of target IO is too large";
1209 ti->max_io_len = (uint32_t) len;
1213 EXPORT_SYMBOL_GPL(dm_set_target_max_io_len);
1215 static struct dm_target *dm_dax_get_live_target(struct mapped_device *md,
1216 sector_t sector, int *srcu_idx)
1217 __acquires(md->io_barrier)
1219 struct dm_table *map;
1220 struct dm_target *ti;
1222 map = dm_get_live_table(md, srcu_idx);
1226 ti = dm_table_find_target(map, sector);
1233 static long dm_dax_direct_access(struct dax_device *dax_dev, pgoff_t pgoff,
1234 long nr_pages, enum dax_access_mode mode, void **kaddr,
1237 struct mapped_device *md = dax_get_private(dax_dev);
1238 sector_t sector = pgoff * PAGE_SECTORS;
1239 struct dm_target *ti;
1240 long len, ret = -EIO;
1243 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1247 if (!ti->type->direct_access)
1249 len = max_io_len(ti, sector) / PAGE_SECTORS;
1252 nr_pages = min(len, nr_pages);
1253 ret = ti->type->direct_access(ti, pgoff, nr_pages, mode, kaddr, pfn);
1256 dm_put_live_table(md, srcu_idx);
1261 static int dm_dax_zero_page_range(struct dax_device *dax_dev, pgoff_t pgoff,
1264 struct mapped_device *md = dax_get_private(dax_dev);
1265 sector_t sector = pgoff * PAGE_SECTORS;
1266 struct dm_target *ti;
1270 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1274 if (WARN_ON(!ti->type->dax_zero_page_range)) {
1276 * ->zero_page_range() is mandatory dax operation. If we are
1277 * here, something is wrong.
1281 ret = ti->type->dax_zero_page_range(ti, pgoff, nr_pages);
1283 dm_put_live_table(md, srcu_idx);
1288 static size_t dm_dax_recovery_write(struct dax_device *dax_dev, pgoff_t pgoff,
1289 void *addr, size_t bytes, struct iov_iter *i)
1291 struct mapped_device *md = dax_get_private(dax_dev);
1292 sector_t sector = pgoff * PAGE_SECTORS;
1293 struct dm_target *ti;
1297 ti = dm_dax_get_live_target(md, sector, &srcu_idx);
1298 if (!ti || !ti->type->dax_recovery_write)
1301 ret = ti->type->dax_recovery_write(ti, pgoff, addr, bytes, i);
1303 dm_put_live_table(md, srcu_idx);
1308 * A target may call dm_accept_partial_bio only from the map routine. It is
1309 * allowed for all bio types except REQ_PREFLUSH, REQ_OP_ZONE_* zone management
1310 * operations, REQ_OP_ZONE_APPEND (zone append writes) and any bio serviced by
1311 * __send_duplicate_bios().
1313 * dm_accept_partial_bio informs the dm that the target only wants to process
1314 * additional n_sectors sectors of the bio and the rest of the data should be
1315 * sent in a next bio.
1317 * A diagram that explains the arithmetics:
1318 * +--------------------+---------------+-------+
1320 * +--------------------+---------------+-------+
1322 * <-------------- *tio->len_ptr --------------->
1323 * <----- bio_sectors ----->
1326 * Region 1 was already iterated over with bio_advance or similar function.
1327 * (it may be empty if the target doesn't use bio_advance)
1328 * Region 2 is the remaining bio size that the target wants to process.
1329 * (it may be empty if region 1 is non-empty, although there is no reason
1331 * The target requires that region 3 is to be sent in the next bio.
1333 * If the target wants to receive multiple copies of the bio (via num_*bios, etc),
1334 * the partially processed part (the sum of regions 1+2) must be the same for all
1335 * copies of the bio.
1337 void dm_accept_partial_bio(struct bio *bio, unsigned int n_sectors)
1339 struct dm_target_io *tio = clone_to_tio(bio);
1340 struct dm_io *io = tio->io;
1341 unsigned int bio_sectors = bio_sectors(bio);
1343 BUG_ON(dm_tio_flagged(tio, DM_TIO_IS_DUPLICATE_BIO));
1344 BUG_ON(op_is_zone_mgmt(bio_op(bio)));
1345 BUG_ON(bio_op(bio) == REQ_OP_ZONE_APPEND);
1346 BUG_ON(bio_sectors > *tio->len_ptr);
1347 BUG_ON(n_sectors > bio_sectors);
1349 *tio->len_ptr -= bio_sectors - n_sectors;
1350 bio->bi_iter.bi_size = n_sectors << SECTOR_SHIFT;
1353 * __split_and_process_bio() may have already saved mapped part
1354 * for accounting but it is being reduced so update accordingly.
1356 dm_io_set_flag(io, DM_IO_WAS_SPLIT);
1357 io->sectors = n_sectors;
1358 io->sector_offset = bio_sectors(io->orig_bio);
1360 EXPORT_SYMBOL_GPL(dm_accept_partial_bio);
1363 * @clone: clone bio that DM core passed to target's .map function
1364 * @tgt_clone: clone of @clone bio that target needs submitted
1366 * Targets should use this interface to submit bios they take
1367 * ownership of when returning DM_MAPIO_SUBMITTED.
1369 * Target should also enable ti->accounts_remapped_io
1371 void dm_submit_bio_remap(struct bio *clone, struct bio *tgt_clone)
1373 struct dm_target_io *tio = clone_to_tio(clone);
1374 struct dm_io *io = tio->io;
1376 /* establish bio that will get submitted */
1381 * Account io->origin_bio to DM dev on behalf of target
1382 * that took ownership of IO with DM_MAPIO_SUBMITTED.
1384 dm_start_io_acct(io, clone);
1386 trace_block_bio_remap(tgt_clone, disk_devt(io->md->disk),
1388 submit_bio_noacct(tgt_clone);
1390 EXPORT_SYMBOL_GPL(dm_submit_bio_remap);
1392 static noinline void __set_swap_bios_limit(struct mapped_device *md, int latch)
1394 mutex_lock(&md->swap_bios_lock);
1395 while (latch < md->swap_bios) {
1397 down(&md->swap_bios_semaphore);
1400 while (latch > md->swap_bios) {
1402 up(&md->swap_bios_semaphore);
1405 mutex_unlock(&md->swap_bios_lock);
1408 static void __map_bio(struct bio *clone)
1410 struct dm_target_io *tio = clone_to_tio(clone);
1411 struct dm_target *ti = tio->ti;
1412 struct dm_io *io = tio->io;
1413 struct mapped_device *md = io->md;
1416 clone->bi_end_io = clone_endio;
1421 tio->old_sector = clone->bi_iter.bi_sector;
1423 if (static_branch_unlikely(&swap_bios_enabled) &&
1424 unlikely(swap_bios_limit(ti, clone))) {
1425 int latch = get_swap_bios();
1427 if (unlikely(latch != md->swap_bios))
1428 __set_swap_bios_limit(md, latch);
1429 down(&md->swap_bios_semaphore);
1432 if (likely(ti->type->map == linear_map))
1433 r = linear_map(ti, clone);
1434 else if (ti->type->map == stripe_map)
1435 r = stripe_map(ti, clone);
1437 r = ti->type->map(ti, clone);
1440 case DM_MAPIO_SUBMITTED:
1441 /* target has assumed ownership of this io */
1442 if (!ti->accounts_remapped_io)
1443 dm_start_io_acct(io, clone);
1445 case DM_MAPIO_REMAPPED:
1446 dm_submit_bio_remap(clone, NULL);
1449 case DM_MAPIO_REQUEUE:
1450 if (static_branch_unlikely(&swap_bios_enabled) &&
1451 unlikely(swap_bios_limit(ti, clone)))
1452 up(&md->swap_bios_semaphore);
1454 if (r == DM_MAPIO_KILL)
1455 dm_io_dec_pending(io, BLK_STS_IOERR);
1457 dm_io_dec_pending(io, BLK_STS_DM_REQUEUE);
1460 DMCRIT("unimplemented target map return value: %d", r);
1465 static void setup_split_accounting(struct clone_info *ci, unsigned int len)
1467 struct dm_io *io = ci->io;
1469 if (ci->sector_count > len) {
1471 * Split needed, save the mapped part for accounting.
1472 * NOTE: dm_accept_partial_bio() will update accordingly.
1474 dm_io_set_flag(io, DM_IO_WAS_SPLIT);
1476 io->sector_offset = bio_sectors(ci->bio);
1480 static void alloc_multiple_bios(struct bio_list *blist, struct clone_info *ci,
1481 struct dm_target *ti, unsigned int num_bios,
1482 unsigned *len, gfp_t gfp_flag)
1485 int try = (gfp_flag & GFP_NOWAIT) ? 0 : 1;
1487 for (; try < 2; try++) {
1490 if (try && num_bios > 1)
1491 mutex_lock(&ci->io->md->table_devices_lock);
1492 for (bio_nr = 0; bio_nr < num_bios; bio_nr++) {
1493 bio = alloc_tio(ci, ti, bio_nr, len,
1494 try ? GFP_NOIO : GFP_NOWAIT);
1498 bio_list_add(blist, bio);
1500 if (try && num_bios > 1)
1501 mutex_unlock(&ci->io->md->table_devices_lock);
1502 if (bio_nr == num_bios)
1505 while ((bio = bio_list_pop(blist)))
1510 static unsigned int __send_duplicate_bios(struct clone_info *ci, struct dm_target *ti,
1511 unsigned int num_bios, unsigned int *len,
1514 struct bio_list blist = BIO_EMPTY_LIST;
1516 unsigned int ret = 0;
1518 if (WARN_ON_ONCE(num_bios == 0)) /* num_bios = 0 is a bug in caller */
1521 /* dm_accept_partial_bio() is not supported with shared tio->len_ptr */
1523 setup_split_accounting(ci, *len);
1526 * Using alloc_multiple_bios(), even if num_bios is 1, to consistently
1527 * support allocating using GFP_NOWAIT with GFP_NOIO fallback.
1529 alloc_multiple_bios(&blist, ci, ti, num_bios, len, gfp_flag);
1530 while ((clone = bio_list_pop(&blist))) {
1532 dm_tio_set_flag(clone_to_tio(clone), DM_TIO_IS_DUPLICATE_BIO);
1540 static void __send_empty_flush(struct clone_info *ci)
1542 struct dm_table *t = ci->map;
1543 struct bio flush_bio;
1546 * Use an on-stack bio for this, it's safe since we don't
1547 * need to reference it after submit. It's just used as
1548 * the basis for the clone(s).
1550 bio_init(&flush_bio, ci->io->md->disk->part0, NULL, 0,
1551 REQ_OP_WRITE | REQ_PREFLUSH | REQ_SYNC);
1553 ci->bio = &flush_bio;
1554 ci->sector_count = 0;
1555 ci->io->tio.clone.bi_iter.bi_size = 0;
1557 if (!t->flush_bypasses_map) {
1558 for (unsigned int i = 0; i < t->num_targets; i++) {
1560 struct dm_target *ti = dm_table_get_target(t, i);
1562 if (unlikely(ti->num_flush_bios == 0))
1565 atomic_add(ti->num_flush_bios, &ci->io->io_count);
1566 bios = __send_duplicate_bios(ci, ti, ti->num_flush_bios,
1568 atomic_sub(ti->num_flush_bios - bios, &ci->io->io_count);
1572 * Note that there's no need to grab t->devices_lock here
1573 * because the targets that support flush optimization don't
1574 * modify the list of devices.
1576 struct list_head *devices = dm_table_get_devices(t);
1577 unsigned int len = 0;
1578 struct dm_dev_internal *dd;
1579 list_for_each_entry(dd, devices, list) {
1582 * Note that the structure dm_target_io is not
1583 * associated with any target (because the device may be
1584 * used by multiple targets), so we set tio->ti = NULL.
1585 * We must check for NULL in the I/O processing path, to
1586 * avoid NULL pointer dereference.
1588 clone = alloc_tio(ci, NULL, 0, &len, GFP_NOIO);
1589 atomic_add(1, &ci->io->io_count);
1590 bio_set_dev(clone, dd->dm_dev->bdev);
1591 clone->bi_end_io = clone_endio;
1592 dm_submit_bio_remap(clone, NULL);
1597 * alloc_io() takes one extra reference for submission, so the
1598 * reference won't reach 0 without the following subtraction
1600 atomic_sub(1, &ci->io->io_count);
1602 bio_uninit(ci->bio);
1605 static void __send_abnormal_io(struct clone_info *ci, struct dm_target *ti,
1606 unsigned int num_bios, unsigned int max_granularity,
1607 unsigned int max_sectors)
1609 unsigned int len, bios;
1611 len = min_t(sector_t, ci->sector_count,
1612 __max_io_len(ti, ci->sector, max_granularity, max_sectors));
1614 atomic_add(num_bios, &ci->io->io_count);
1615 bios = __send_duplicate_bios(ci, ti, num_bios, &len, GFP_NOIO);
1617 * alloc_io() takes one extra reference for submission, so the
1618 * reference won't reach 0 without the following (+1) subtraction
1620 atomic_sub(num_bios - bios + 1, &ci->io->io_count);
1623 ci->sector_count -= len;
1626 static bool is_abnormal_io(struct bio *bio)
1628 switch (bio_op(bio)) {
1633 case REQ_OP_DISCARD:
1634 case REQ_OP_SECURE_ERASE:
1635 case REQ_OP_WRITE_ZEROES:
1636 case REQ_OP_ZONE_RESET_ALL:
1643 static blk_status_t __process_abnormal_io(struct clone_info *ci,
1644 struct dm_target *ti)
1646 unsigned int num_bios = 0;
1647 unsigned int max_granularity = 0;
1648 unsigned int max_sectors = 0;
1649 struct queue_limits *limits = dm_get_queue_limits(ti->table->md);
1651 switch (bio_op(ci->bio)) {
1652 case REQ_OP_DISCARD:
1653 num_bios = ti->num_discard_bios;
1654 max_sectors = limits->max_discard_sectors;
1655 if (ti->max_discard_granularity)
1656 max_granularity = max_sectors;
1658 case REQ_OP_SECURE_ERASE:
1659 num_bios = ti->num_secure_erase_bios;
1660 max_sectors = limits->max_secure_erase_sectors;
1662 case REQ_OP_WRITE_ZEROES:
1663 num_bios = ti->num_write_zeroes_bios;
1664 max_sectors = limits->max_write_zeroes_sectors;
1671 * Even though the device advertised support for this type of
1672 * request, that does not mean every target supports it, and
1673 * reconfiguration might also have changed that since the
1674 * check was performed.
1676 if (unlikely(!num_bios))
1677 return BLK_STS_NOTSUPP;
1679 __send_abnormal_io(ci, ti, num_bios, max_granularity, max_sectors);
1685 * Reuse ->bi_private as dm_io list head for storing all dm_io instances
1686 * associated with this bio, and this bio's bi_private needs to be
1687 * stored in dm_io->data before the reuse.
1689 * bio->bi_private is owned by fs or upper layer, so block layer won't
1690 * touch it after splitting. Meantime it won't be changed by anyone after
1691 * bio is submitted. So this reuse is safe.
1693 static inline struct dm_io **dm_poll_list_head(struct bio *bio)
1695 return (struct dm_io **)&bio->bi_private;
1698 static void dm_queue_poll_io(struct bio *bio, struct dm_io *io)
1700 struct dm_io **head = dm_poll_list_head(bio);
1702 if (!(bio->bi_opf & REQ_DM_POLL_LIST)) {
1703 bio->bi_opf |= REQ_DM_POLL_LIST;
1705 * Save .bi_private into dm_io, so that we can reuse
1706 * .bi_private as dm_io list head for storing dm_io list
1708 io->data = bio->bi_private;
1710 /* tell block layer to poll for completion */
1711 bio->bi_cookie = ~BLK_QC_T_NONE;
1716 * bio recursed due to split, reuse original poll list,
1717 * and save bio->bi_private too.
1719 io->data = (*head)->data;
1727 * Select the correct strategy for processing a non-flush bio.
1729 static blk_status_t __split_and_process_bio(struct clone_info *ci)
1732 struct dm_target *ti;
1735 ti = dm_table_find_target(ci->map, ci->sector);
1737 return BLK_STS_IOERR;
1739 if (unlikely(ci->is_abnormal_io))
1740 return __process_abnormal_io(ci, ti);
1743 * Only support bio polling for normal IO, and the target io is
1744 * exactly inside the dm_io instance (verified in dm_poll_dm_io)
1746 ci->submit_as_polled = !!(ci->bio->bi_opf & REQ_POLLED);
1748 len = min_t(sector_t, max_io_len(ti, ci->sector), ci->sector_count);
1749 setup_split_accounting(ci, len);
1751 if (unlikely(ci->bio->bi_opf & REQ_NOWAIT)) {
1752 if (unlikely(!dm_target_supports_nowait(ti->type)))
1753 return BLK_STS_NOTSUPP;
1755 clone = alloc_tio(ci, ti, 0, &len, GFP_NOWAIT);
1756 if (unlikely(!clone))
1757 return BLK_STS_AGAIN;
1759 clone = alloc_tio(ci, ti, 0, &len, GFP_NOIO);
1764 ci->sector_count -= len;
1769 static void init_clone_info(struct clone_info *ci, struct dm_io *io,
1770 struct dm_table *map, struct bio *bio, bool is_abnormal)
1775 ci->is_abnormal_io = is_abnormal;
1776 ci->submit_as_polled = false;
1777 ci->sector = bio->bi_iter.bi_sector;
1778 ci->sector_count = bio_sectors(bio);
1780 /* Shouldn't happen but sector_count was being set to 0 so... */
1781 if (static_branch_unlikely(&zoned_enabled) &&
1782 WARN_ON_ONCE(op_is_zone_mgmt(bio_op(bio)) && ci->sector_count))
1783 ci->sector_count = 0;
1786 #ifdef CONFIG_BLK_DEV_ZONED
1787 static inline bool dm_zone_bio_needs_split(struct mapped_device *md,
1791 * For mapped device that need zone append emulation, we must
1792 * split any large BIO that straddles zone boundaries.
1794 return dm_emulate_zone_append(md) && bio_straddles_zones(bio) &&
1795 !bio_flagged(bio, BIO_ZONE_WRITE_PLUGGING);
1797 static inline bool dm_zone_plug_bio(struct mapped_device *md, struct bio *bio)
1799 return dm_emulate_zone_append(md) && blk_zone_plug_bio(bio, 0);
1802 static blk_status_t __send_zone_reset_all_emulated(struct clone_info *ci,
1803 struct dm_target *ti)
1805 struct bio_list blist = BIO_EMPTY_LIST;
1806 struct mapped_device *md = ci->io->md;
1807 unsigned int zone_sectors = md->disk->queue->limits.chunk_sectors;
1808 unsigned long *need_reset;
1809 unsigned int i, nr_zones, nr_reset;
1810 unsigned int num_bios = 0;
1811 blk_status_t sts = BLK_STS_OK;
1812 sector_t sector = ti->begin;
1816 nr_zones = ti->len >> ilog2(zone_sectors);
1817 need_reset = bitmap_zalloc(nr_zones, GFP_NOIO);
1819 return BLK_STS_RESOURCE;
1821 ret = dm_zone_get_reset_bitmap(md, ci->map, ti->begin,
1822 nr_zones, need_reset);
1824 sts = BLK_STS_IOERR;
1828 /* If we have no zone to reset, we are done. */
1829 nr_reset = bitmap_weight(need_reset, nr_zones);
1833 atomic_add(nr_zones, &ci->io->io_count);
1835 for (i = 0; i < nr_zones; i++) {
1837 if (!test_bit(i, need_reset)) {
1838 sector += zone_sectors;
1842 if (bio_list_empty(&blist)) {
1843 /* This may take a while, so be nice to others */
1848 * We may need to reset thousands of zones, so let's
1849 * not go crazy with the clone allocation.
1851 alloc_multiple_bios(&blist, ci, ti, min(nr_reset, 32),
1855 /* Get a clone and change it to a regular reset operation. */
1856 clone = bio_list_pop(&blist);
1857 clone->bi_opf &= ~REQ_OP_MASK;
1858 clone->bi_opf |= REQ_OP_ZONE_RESET | REQ_SYNC;
1859 clone->bi_iter.bi_sector = sector;
1860 clone->bi_iter.bi_size = 0;
1863 sector += zone_sectors;
1868 WARN_ON_ONCE(!bio_list_empty(&blist));
1869 atomic_sub(nr_zones - num_bios, &ci->io->io_count);
1870 ci->sector_count = 0;
1873 bitmap_free(need_reset);
1878 static void __send_zone_reset_all_native(struct clone_info *ci,
1879 struct dm_target *ti)
1883 atomic_add(1, &ci->io->io_count);
1884 bios = __send_duplicate_bios(ci, ti, 1, NULL, GFP_NOIO);
1885 atomic_sub(1 - bios, &ci->io->io_count);
1887 ci->sector_count = 0;
1890 static blk_status_t __send_zone_reset_all(struct clone_info *ci)
1892 struct dm_table *t = ci->map;
1893 blk_status_t sts = BLK_STS_OK;
1895 for (unsigned int i = 0; i < t->num_targets; i++) {
1896 struct dm_target *ti = dm_table_get_target(t, i);
1898 if (ti->zone_reset_all_supported) {
1899 __send_zone_reset_all_native(ci, ti);
1903 sts = __send_zone_reset_all_emulated(ci, ti);
1904 if (sts != BLK_STS_OK)
1908 /* Release the reference that alloc_io() took for submission. */
1909 atomic_sub(1, &ci->io->io_count);
1915 static inline bool dm_zone_bio_needs_split(struct mapped_device *md,
1920 static inline bool dm_zone_plug_bio(struct mapped_device *md, struct bio *bio)
1924 static blk_status_t __send_zone_reset_all(struct clone_info *ci)
1926 return BLK_STS_NOTSUPP;
1931 * Entry point to split a bio into clones and submit them to the targets.
1933 static void dm_split_and_process_bio(struct mapped_device *md,
1934 struct dm_table *map, struct bio *bio)
1936 struct clone_info ci;
1938 blk_status_t error = BLK_STS_OK;
1939 bool is_abnormal, need_split;
1941 is_abnormal = is_abnormal_io(bio);
1942 if (static_branch_unlikely(&zoned_enabled)) {
1943 /* Special case REQ_OP_ZONE_RESET_ALL as it cannot be split. */
1944 need_split = (bio_op(bio) != REQ_OP_ZONE_RESET_ALL) &&
1945 (is_abnormal || dm_zone_bio_needs_split(md, bio));
1947 need_split = is_abnormal;
1950 if (unlikely(need_split)) {
1952 * Use bio_split_to_limits() for abnormal IO (e.g. discard, etc)
1953 * otherwise associated queue_limits won't be imposed.
1954 * Also split the BIO for mapped devices needing zone append
1955 * emulation to ensure that the BIO does not cross zone
1958 bio = bio_split_to_limits(bio);
1964 * Use the block layer zone write plugging for mapped devices that
1965 * need zone append emulation (e.g. dm-crypt).
1967 if (static_branch_unlikely(&zoned_enabled) && dm_zone_plug_bio(md, bio))
1970 /* Only support nowait for normal IO */
1971 if (unlikely(bio->bi_opf & REQ_NOWAIT) && !is_abnormal) {
1972 io = alloc_io(md, bio, GFP_NOWAIT);
1973 if (unlikely(!io)) {
1974 /* Unable to do anything without dm_io. */
1975 bio_wouldblock_error(bio);
1979 io = alloc_io(md, bio, GFP_NOIO);
1981 init_clone_info(&ci, io, map, bio, is_abnormal);
1983 if (bio->bi_opf & REQ_PREFLUSH) {
1984 __send_empty_flush(&ci);
1985 /* dm_io_complete submits any data associated with flush */
1989 if (static_branch_unlikely(&zoned_enabled) &&
1990 (bio_op(bio) == REQ_OP_ZONE_RESET_ALL)) {
1991 error = __send_zone_reset_all(&ci);
1995 error = __split_and_process_bio(&ci);
1996 if (error || !ci.sector_count)
1999 * Remainder must be passed to submit_bio_noacct() so it gets handled
2000 * *after* bios already submitted have been completely processed.
2002 bio_trim(bio, io->sectors, ci.sector_count);
2003 trace_block_split(bio, bio->bi_iter.bi_sector);
2004 bio_inc_remaining(bio);
2005 submit_bio_noacct(bio);
2008 * Drop the extra reference count for non-POLLED bio, and hold one
2009 * reference for POLLED bio, which will be released in dm_poll_bio
2011 * Add every dm_io instance into the dm_io list head which is stored
2012 * in bio->bi_private, so that dm_poll_bio can poll them all.
2014 if (error || !ci.submit_as_polled) {
2016 * In case of submission failure, the extra reference for
2017 * submitting io isn't consumed yet
2020 atomic_dec(&io->io_count);
2021 dm_io_dec_pending(io, error);
2023 dm_queue_poll_io(bio, io);
2026 static void dm_submit_bio(struct bio *bio)
2028 struct mapped_device *md = bio->bi_bdev->bd_disk->private_data;
2030 struct dm_table *map;
2032 map = dm_get_live_table(md, &srcu_idx);
2034 /* If suspended, or map not yet available, queue this IO for later */
2035 if (unlikely(test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) ||
2037 if (bio->bi_opf & REQ_NOWAIT)
2038 bio_wouldblock_error(bio);
2039 else if (bio->bi_opf & REQ_RAHEAD)
2046 dm_split_and_process_bio(md, map, bio);
2048 dm_put_live_table(md, srcu_idx);
2051 static bool dm_poll_dm_io(struct dm_io *io, struct io_comp_batch *iob,
2054 WARN_ON_ONCE(!dm_tio_is_normal(&io->tio));
2056 /* don't poll if the mapped io is done */
2057 if (atomic_read(&io->io_count) > 1)
2058 bio_poll(&io->tio.clone, iob, flags);
2060 /* bio_poll holds the last reference */
2061 return atomic_read(&io->io_count) == 1;
2064 static int dm_poll_bio(struct bio *bio, struct io_comp_batch *iob,
2067 struct dm_io **head = dm_poll_list_head(bio);
2068 struct dm_io *list = *head;
2069 struct dm_io *tmp = NULL;
2070 struct dm_io *curr, *next;
2072 /* Only poll normal bio which was marked as REQ_DM_POLL_LIST */
2073 if (!(bio->bi_opf & REQ_DM_POLL_LIST))
2076 WARN_ON_ONCE(!list);
2079 * Restore .bi_private before possibly completing dm_io.
2081 * bio_poll() is only possible once @bio has been completely
2082 * submitted via submit_bio_noacct()'s depth-first submission.
2083 * So there is no dm_queue_poll_io() race associated with
2084 * clearing REQ_DM_POLL_LIST here.
2086 bio->bi_opf &= ~REQ_DM_POLL_LIST;
2087 bio->bi_private = list->data;
2089 for (curr = list, next = curr->next; curr; curr = next, next =
2090 curr ? curr->next : NULL) {
2091 if (dm_poll_dm_io(curr, iob, flags)) {
2093 * clone_endio() has already occurred, so no
2094 * error handling is needed here.
2096 __dm_io_dec_pending(curr);
2105 bio->bi_opf |= REQ_DM_POLL_LIST;
2106 /* Reset bio->bi_private to dm_io list head */
2114 *---------------------------------------------------------------
2115 * An IDR is used to keep track of allocated minor numbers.
2116 *---------------------------------------------------------------
2118 static void free_minor(int minor)
2120 spin_lock(&_minor_lock);
2121 idr_remove(&_minor_idr, minor);
2122 spin_unlock(&_minor_lock);
2126 * See if the device with a specific minor # is free.
2128 static int specific_minor(int minor)
2132 if (minor >= (1 << MINORBITS))
2135 idr_preload(GFP_KERNEL);
2136 spin_lock(&_minor_lock);
2138 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, minor, minor + 1, GFP_NOWAIT);
2140 spin_unlock(&_minor_lock);
2143 return r == -ENOSPC ? -EBUSY : r;
2147 static int next_free_minor(int *minor)
2151 idr_preload(GFP_KERNEL);
2152 spin_lock(&_minor_lock);
2154 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, 0, 1 << MINORBITS, GFP_NOWAIT);
2156 spin_unlock(&_minor_lock);
2164 static const struct block_device_operations dm_blk_dops;
2165 static const struct block_device_operations dm_rq_blk_dops;
2166 static const struct dax_operations dm_dax_ops;
2168 static void dm_wq_work(struct work_struct *work);
2170 #ifdef CONFIG_BLK_INLINE_ENCRYPTION
2171 static void dm_queue_destroy_crypto_profile(struct request_queue *q)
2173 dm_destroy_crypto_profile(q->crypto_profile);
2176 #else /* CONFIG_BLK_INLINE_ENCRYPTION */
2178 static inline void dm_queue_destroy_crypto_profile(struct request_queue *q)
2181 #endif /* !CONFIG_BLK_INLINE_ENCRYPTION */
2183 static void cleanup_mapped_device(struct mapped_device *md)
2186 destroy_workqueue(md->wq);
2187 dm_free_md_mempools(md->mempools);
2190 dax_remove_host(md->disk);
2191 kill_dax(md->dax_dev);
2192 put_dax(md->dax_dev);
2197 spin_lock(&_minor_lock);
2198 md->disk->private_data = NULL;
2199 spin_unlock(&_minor_lock);
2200 if (dm_get_md_type(md) != DM_TYPE_NONE) {
2201 struct table_device *td;
2204 list_for_each_entry(td, &md->table_devices, list) {
2205 bd_unlink_disk_holder(td->dm_dev.bdev,
2210 * Hold lock to make sure del_gendisk() won't concurrent
2211 * with open/close_table_device().
2213 mutex_lock(&md->table_devices_lock);
2214 del_gendisk(md->disk);
2215 mutex_unlock(&md->table_devices_lock);
2217 dm_queue_destroy_crypto_profile(md->queue);
2221 if (md->pending_io) {
2222 free_percpu(md->pending_io);
2223 md->pending_io = NULL;
2226 cleanup_srcu_struct(&md->io_barrier);
2228 mutex_destroy(&md->suspend_lock);
2229 mutex_destroy(&md->type_lock);
2230 mutex_destroy(&md->table_devices_lock);
2231 mutex_destroy(&md->swap_bios_lock);
2233 dm_mq_cleanup_mapped_device(md);
2237 * Allocate and initialise a blank device with a given minor.
2239 static struct mapped_device *alloc_dev(int minor)
2241 int r, numa_node_id = dm_get_numa_node();
2242 struct dax_device *dax_dev;
2243 struct mapped_device *md;
2246 md = kvzalloc_node(sizeof(*md), GFP_KERNEL, numa_node_id);
2248 DMERR("unable to allocate device, out of memory.");
2252 if (!try_module_get(THIS_MODULE))
2253 goto bad_module_get;
2255 /* get a minor number for the dev */
2256 if (minor == DM_ANY_MINOR)
2257 r = next_free_minor(&minor);
2259 r = specific_minor(minor);
2263 r = init_srcu_struct(&md->io_barrier);
2265 goto bad_io_barrier;
2267 md->numa_node_id = numa_node_id;
2268 md->init_tio_pdu = false;
2269 md->type = DM_TYPE_NONE;
2270 mutex_init(&md->suspend_lock);
2271 mutex_init(&md->type_lock);
2272 mutex_init(&md->table_devices_lock);
2273 spin_lock_init(&md->deferred_lock);
2274 atomic_set(&md->holders, 1);
2275 atomic_set(&md->open_count, 0);
2276 atomic_set(&md->event_nr, 0);
2277 atomic_set(&md->uevent_seq, 0);
2278 INIT_LIST_HEAD(&md->uevent_list);
2279 INIT_LIST_HEAD(&md->table_devices);
2280 spin_lock_init(&md->uevent_lock);
2283 * default to bio-based until DM table is loaded and md->type
2284 * established. If request-based table is loaded: blk-mq will
2285 * override accordingly.
2287 md->disk = blk_alloc_disk(NULL, md->numa_node_id);
2288 if (IS_ERR(md->disk))
2290 md->queue = md->disk->queue;
2292 init_waitqueue_head(&md->wait);
2293 INIT_WORK(&md->work, dm_wq_work);
2294 INIT_WORK(&md->requeue_work, dm_wq_requeue_work);
2295 init_waitqueue_head(&md->eventq);
2296 init_completion(&md->kobj_holder.completion);
2298 md->requeue_list = NULL;
2299 md->swap_bios = get_swap_bios();
2300 sema_init(&md->swap_bios_semaphore, md->swap_bios);
2301 mutex_init(&md->swap_bios_lock);
2303 md->disk->major = _major;
2304 md->disk->first_minor = minor;
2305 md->disk->minors = 1;
2306 md->disk->flags |= GENHD_FL_NO_PART;
2307 md->disk->fops = &dm_blk_dops;
2308 md->disk->private_data = md;
2309 sprintf(md->disk->disk_name, "dm-%d", minor);
2311 dax_dev = alloc_dax(md, &dm_dax_ops);
2312 if (IS_ERR(dax_dev)) {
2313 if (PTR_ERR(dax_dev) != -EOPNOTSUPP)
2316 set_dax_nocache(dax_dev);
2317 set_dax_nomc(dax_dev);
2318 md->dax_dev = dax_dev;
2319 if (dax_add_host(dax_dev, md->disk))
2323 format_dev_t(md->name, MKDEV(_major, minor));
2325 md->wq = alloc_workqueue("kdmflush/%s", WQ_MEM_RECLAIM, 0, md->name);
2329 md->pending_io = alloc_percpu(unsigned long);
2330 if (!md->pending_io)
2333 r = dm_stats_init(&md->stats);
2337 /* Populate the mapping, nobody knows we exist yet */
2338 spin_lock(&_minor_lock);
2339 old_md = idr_replace(&_minor_idr, md, minor);
2340 spin_unlock(&_minor_lock);
2342 BUG_ON(old_md != MINOR_ALLOCED);
2347 cleanup_mapped_device(md);
2351 module_put(THIS_MODULE);
2357 static void unlock_fs(struct mapped_device *md);
2359 static void free_dev(struct mapped_device *md)
2361 int minor = MINOR(disk_devt(md->disk));
2365 cleanup_mapped_device(md);
2367 WARN_ON_ONCE(!list_empty(&md->table_devices));
2368 dm_stats_cleanup(&md->stats);
2371 module_put(THIS_MODULE);
2376 * Bind a table to the device.
2378 static void event_callback(void *context)
2380 unsigned long flags;
2382 struct mapped_device *md = context;
2384 spin_lock_irqsave(&md->uevent_lock, flags);
2385 list_splice_init(&md->uevent_list, &uevents);
2386 spin_unlock_irqrestore(&md->uevent_lock, flags);
2388 dm_send_uevents(&uevents, &disk_to_dev(md->disk)->kobj);
2390 atomic_inc(&md->event_nr);
2391 wake_up(&md->eventq);
2392 dm_issue_global_event();
2396 * Returns old map, which caller must destroy.
2398 static struct dm_table *__bind(struct mapped_device *md, struct dm_table *t,
2399 struct queue_limits *limits)
2401 struct dm_table *old_map;
2405 lockdep_assert_held(&md->suspend_lock);
2407 size = dm_table_get_size(t);
2410 * Wipe any geometry if the size of the table changed.
2412 if (size != dm_get_size(md))
2413 memset(&md->geometry, 0, sizeof(md->geometry));
2415 set_capacity(md->disk, size);
2417 dm_table_event_callback(t, event_callback, md);
2419 if (dm_table_request_based(t)) {
2421 * Leverage the fact that request-based DM targets are
2422 * immutable singletons - used to optimize dm_mq_queue_rq.
2424 md->immutable_target = dm_table_get_immutable_target(t);
2427 * There is no need to reload with request-based dm because the
2428 * size of front_pad doesn't change.
2430 * Note for future: If you are to reload bioset, prep-ed
2431 * requests in the queue may refer to bio from the old bioset,
2432 * so you must walk through the queue to unprep.
2434 if (!md->mempools) {
2435 md->mempools = t->mempools;
2440 * The md may already have mempools that need changing.
2441 * If so, reload bioset because front_pad may have changed
2442 * because a different table was loaded.
2444 dm_free_md_mempools(md->mempools);
2445 md->mempools = t->mempools;
2449 ret = dm_table_set_restrictions(t, md->queue, limits);
2451 old_map = ERR_PTR(ret);
2455 old_map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2456 rcu_assign_pointer(md->map, (void *)t);
2457 md->immutable_target_type = dm_table_get_immutable_target_type(t);
2466 * Returns unbound table for the caller to free.
2468 static struct dm_table *__unbind(struct mapped_device *md)
2470 struct dm_table *map = rcu_dereference_protected(md->map, 1);
2475 dm_table_event_callback(map, NULL, NULL);
2476 RCU_INIT_POINTER(md->map, NULL);
2483 * Constructor for a new device.
2485 int dm_create(int minor, struct mapped_device **result)
2487 struct mapped_device *md;
2489 md = alloc_dev(minor);
2493 dm_ima_reset_data(md);
2500 * Functions to manage md->type.
2501 * All are required to hold md->type_lock.
2503 void dm_lock_md_type(struct mapped_device *md)
2505 mutex_lock(&md->type_lock);
2508 void dm_unlock_md_type(struct mapped_device *md)
2510 mutex_unlock(&md->type_lock);
2513 void dm_set_md_type(struct mapped_device *md, enum dm_queue_mode type)
2515 BUG_ON(!mutex_is_locked(&md->type_lock));
2519 enum dm_queue_mode dm_get_md_type(struct mapped_device *md)
2524 struct target_type *dm_get_immutable_target_type(struct mapped_device *md)
2526 return md->immutable_target_type;
2530 * Setup the DM device's queue based on md's type
2532 int dm_setup_md_queue(struct mapped_device *md, struct dm_table *t)
2534 enum dm_queue_mode type = dm_table_get_type(t);
2535 struct queue_limits limits;
2536 struct table_device *td;
2539 WARN_ON_ONCE(type == DM_TYPE_NONE);
2541 if (type == DM_TYPE_REQUEST_BASED) {
2542 md->disk->fops = &dm_rq_blk_dops;
2543 r = dm_mq_init_request_queue(md, t);
2545 DMERR("Cannot initialize queue for request-based dm mapped device");
2550 r = dm_calculate_queue_limits(t, &limits);
2552 DMERR("Cannot calculate initial queue limits");
2555 r = dm_table_set_restrictions(t, md->queue, &limits);
2560 * Hold lock to make sure add_disk() and del_gendisk() won't concurrent
2561 * with open_table_device() and close_table_device().
2563 mutex_lock(&md->table_devices_lock);
2564 r = add_disk(md->disk);
2565 mutex_unlock(&md->table_devices_lock);
2570 * Register the holder relationship for devices added before the disk
2573 list_for_each_entry(td, &md->table_devices, list) {
2574 r = bd_link_disk_holder(td->dm_dev.bdev, md->disk);
2576 goto out_undo_holders;
2579 r = dm_sysfs_init(md);
2581 goto out_undo_holders;
2587 list_for_each_entry_continue_reverse(td, &md->table_devices, list)
2588 bd_unlink_disk_holder(td->dm_dev.bdev, md->disk);
2589 mutex_lock(&md->table_devices_lock);
2590 del_gendisk(md->disk);
2591 mutex_unlock(&md->table_devices_lock);
2595 struct mapped_device *dm_get_md(dev_t dev)
2597 struct mapped_device *md;
2598 unsigned int minor = MINOR(dev);
2600 if (MAJOR(dev) != _major || minor >= (1 << MINORBITS))
2603 spin_lock(&_minor_lock);
2605 md = idr_find(&_minor_idr, minor);
2606 if (!md || md == MINOR_ALLOCED || (MINOR(disk_devt(dm_disk(md))) != minor) ||
2607 test_bit(DMF_FREEING, &md->flags) || dm_deleting_md(md)) {
2613 spin_unlock(&_minor_lock);
2617 EXPORT_SYMBOL_GPL(dm_get_md);
2619 void *dm_get_mdptr(struct mapped_device *md)
2621 return md->interface_ptr;
2624 void dm_set_mdptr(struct mapped_device *md, void *ptr)
2626 md->interface_ptr = ptr;
2629 void dm_get(struct mapped_device *md)
2631 atomic_inc(&md->holders);
2632 BUG_ON(test_bit(DMF_FREEING, &md->flags));
2635 int dm_hold(struct mapped_device *md)
2637 spin_lock(&_minor_lock);
2638 if (test_bit(DMF_FREEING, &md->flags)) {
2639 spin_unlock(&_minor_lock);
2643 spin_unlock(&_minor_lock);
2646 EXPORT_SYMBOL_GPL(dm_hold);
2648 const char *dm_device_name(struct mapped_device *md)
2652 EXPORT_SYMBOL_GPL(dm_device_name);
2654 static void __dm_destroy(struct mapped_device *md, bool wait)
2656 struct dm_table *map;
2661 spin_lock(&_minor_lock);
2662 idr_replace(&_minor_idr, MINOR_ALLOCED, MINOR(disk_devt(dm_disk(md))));
2663 set_bit(DMF_FREEING, &md->flags);
2664 spin_unlock(&_minor_lock);
2666 blk_mark_disk_dead(md->disk);
2669 * Take suspend_lock so that presuspend and postsuspend methods
2670 * do not race with internal suspend.
2672 mutex_lock(&md->suspend_lock);
2673 map = dm_get_live_table(md, &srcu_idx);
2674 if (!dm_suspended_md(md)) {
2675 dm_table_presuspend_targets(map);
2676 set_bit(DMF_SUSPENDED, &md->flags);
2677 set_bit(DMF_POST_SUSPENDING, &md->flags);
2678 dm_table_postsuspend_targets(map);
2680 /* dm_put_live_table must be before fsleep, otherwise deadlock is possible */
2681 dm_put_live_table(md, srcu_idx);
2682 mutex_unlock(&md->suspend_lock);
2685 * Rare, but there may be I/O requests still going to complete,
2686 * for example. Wait for all references to disappear.
2687 * No one should increment the reference count of the mapped_device,
2688 * after the mapped_device state becomes DMF_FREEING.
2691 while (atomic_read(&md->holders))
2693 else if (atomic_read(&md->holders))
2694 DMWARN("%s: Forcibly removing mapped_device still in use! (%d users)",
2695 dm_device_name(md), atomic_read(&md->holders));
2697 dm_table_destroy(__unbind(md));
2701 void dm_destroy(struct mapped_device *md)
2703 __dm_destroy(md, true);
2706 void dm_destroy_immediate(struct mapped_device *md)
2708 __dm_destroy(md, false);
2711 void dm_put(struct mapped_device *md)
2713 atomic_dec(&md->holders);
2715 EXPORT_SYMBOL_GPL(dm_put);
2717 static bool dm_in_flight_bios(struct mapped_device *md)
2720 unsigned long sum = 0;
2722 for_each_possible_cpu(cpu)
2723 sum += *per_cpu_ptr(md->pending_io, cpu);
2728 static int dm_wait_for_bios_completion(struct mapped_device *md, unsigned int task_state)
2734 prepare_to_wait(&md->wait, &wait, task_state);
2736 if (!dm_in_flight_bios(md))
2739 if (signal_pending_state(task_state, current)) {
2746 finish_wait(&md->wait, &wait);
2753 static int dm_wait_for_completion(struct mapped_device *md, unsigned int task_state)
2757 if (!queue_is_mq(md->queue))
2758 return dm_wait_for_bios_completion(md, task_state);
2761 if (!blk_mq_queue_inflight(md->queue))
2764 if (signal_pending_state(task_state, current)) {
2776 * Process the deferred bios
2778 static void dm_wq_work(struct work_struct *work)
2780 struct mapped_device *md = container_of(work, struct mapped_device, work);
2783 while (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
2784 spin_lock_irq(&md->deferred_lock);
2785 bio = bio_list_pop(&md->deferred);
2786 spin_unlock_irq(&md->deferred_lock);
2791 submit_bio_noacct(bio);
2796 static void dm_queue_flush(struct mapped_device *md)
2798 clear_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2799 smp_mb__after_atomic();
2800 queue_work(md->wq, &md->work);
2804 * Swap in a new table, returning the old one for the caller to destroy.
2806 struct dm_table *dm_swap_table(struct mapped_device *md, struct dm_table *table)
2808 struct dm_table *live_map = NULL, *map = ERR_PTR(-EINVAL);
2809 struct queue_limits limits;
2812 mutex_lock(&md->suspend_lock);
2814 /* device must be suspended */
2815 if (!dm_suspended_md(md))
2819 * If the new table has no data devices, retain the existing limits.
2820 * This helps multipath with queue_if_no_path if all paths disappear,
2821 * then new I/O is queued based on these limits, and then some paths
2824 if (dm_table_has_no_data_devices(table)) {
2825 live_map = dm_get_live_table_fast(md);
2827 limits = md->queue->limits;
2828 dm_put_live_table_fast(md);
2832 r = dm_calculate_queue_limits(table, &limits);
2839 map = __bind(md, table, &limits);
2840 dm_issue_global_event();
2843 mutex_unlock(&md->suspend_lock);
2848 * Functions to lock and unlock any filesystem running on the
2851 static int lock_fs(struct mapped_device *md)
2855 WARN_ON(test_bit(DMF_FROZEN, &md->flags));
2857 r = bdev_freeze(md->disk->part0);
2859 set_bit(DMF_FROZEN, &md->flags);
2863 static void unlock_fs(struct mapped_device *md)
2865 if (!test_bit(DMF_FROZEN, &md->flags))
2867 bdev_thaw(md->disk->part0);
2868 clear_bit(DMF_FROZEN, &md->flags);
2872 * @suspend_flags: DM_SUSPEND_LOCKFS_FLAG and/or DM_SUSPEND_NOFLUSH_FLAG
2873 * @task_state: e.g. TASK_INTERRUPTIBLE or TASK_UNINTERRUPTIBLE
2874 * @dmf_suspended_flag: DMF_SUSPENDED or DMF_SUSPENDED_INTERNALLY
2876 * If __dm_suspend returns 0, the device is completely quiescent
2877 * now. There is no request-processing activity. All new requests
2878 * are being added to md->deferred list.
2880 static int __dm_suspend(struct mapped_device *md, struct dm_table *map,
2881 unsigned int suspend_flags, unsigned int task_state,
2882 int dmf_suspended_flag)
2884 bool do_lockfs = suspend_flags & DM_SUSPEND_LOCKFS_FLAG;
2885 bool noflush = suspend_flags & DM_SUSPEND_NOFLUSH_FLAG;
2888 lockdep_assert_held(&md->suspend_lock);
2891 * DMF_NOFLUSH_SUSPENDING must be set before presuspend.
2892 * This flag is cleared before dm_suspend returns.
2895 set_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
2897 DMDEBUG("%s: suspending with flush", dm_device_name(md));
2900 * This gets reverted if there's an error later and the targets
2901 * provide the .presuspend_undo hook.
2903 dm_table_presuspend_targets(map);
2906 * Flush I/O to the device.
2907 * Any I/O submitted after lock_fs() may not be flushed.
2908 * noflush takes precedence over do_lockfs.
2909 * (lock_fs() flushes I/Os and waits for them to complete.)
2911 if (!noflush && do_lockfs) {
2914 dm_table_presuspend_undo_targets(map);
2920 * Here we must make sure that no processes are submitting requests
2921 * to target drivers i.e. no one may be executing
2922 * dm_split_and_process_bio from dm_submit_bio.
2924 * To get all processes out of dm_split_and_process_bio in dm_submit_bio,
2925 * we take the write lock. To prevent any process from reentering
2926 * dm_split_and_process_bio from dm_submit_bio and quiesce the thread
2927 * (dm_wq_work), we set DMF_BLOCK_IO_FOR_SUSPEND and call
2928 * flush_workqueue(md->wq).
2930 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2932 synchronize_srcu(&md->io_barrier);
2935 * Stop md->queue before flushing md->wq in case request-based
2936 * dm defers requests to md->wq from md->queue.
2938 if (dm_request_based(md))
2939 dm_stop_queue(md->queue);
2941 flush_workqueue(md->wq);
2944 * At this point no more requests are entering target request routines.
2945 * We call dm_wait_for_completion to wait for all existing requests
2948 r = dm_wait_for_completion(md, task_state);
2950 set_bit(dmf_suspended_flag, &md->flags);
2953 clear_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
2955 synchronize_srcu(&md->io_barrier);
2957 /* were we interrupted ? */
2961 if (dm_request_based(md))
2962 dm_start_queue(md->queue);
2965 dm_table_presuspend_undo_targets(map);
2966 /* pushback list is already flushed, so skip flush */
2973 * We need to be able to change a mapping table under a mounted
2974 * filesystem. For example we might want to move some data in
2975 * the background. Before the table can be swapped with
2976 * dm_bind_table, dm_suspend must be called to flush any in
2977 * flight bios and ensure that any further io gets deferred.
2980 * Suspend mechanism in request-based dm.
2982 * 1. Flush all I/Os by lock_fs() if needed.
2983 * 2. Stop dispatching any I/O by stopping the request_queue.
2984 * 3. Wait for all in-flight I/Os to be completed or requeued.
2986 * To abort suspend, start the request_queue.
2988 int dm_suspend(struct mapped_device *md, unsigned int suspend_flags)
2990 struct dm_table *map = NULL;
2994 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
2996 if (dm_suspended_md(md)) {
3001 if (dm_suspended_internally_md(md)) {
3002 /* already internally suspended, wait for internal resume */
3003 mutex_unlock(&md->suspend_lock);
3004 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
3010 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3012 /* avoid deadlock with fs/namespace.c:do_mount() */
3013 suspend_flags &= ~DM_SUSPEND_LOCKFS_FLAG;
3016 r = __dm_suspend(md, map, suspend_flags, TASK_INTERRUPTIBLE, DMF_SUSPENDED);
3020 set_bit(DMF_POST_SUSPENDING, &md->flags);
3021 dm_table_postsuspend_targets(map);
3022 clear_bit(DMF_POST_SUSPENDING, &md->flags);
3025 mutex_unlock(&md->suspend_lock);
3029 static int __dm_resume(struct mapped_device *md, struct dm_table *map)
3032 int r = dm_table_resume_targets(map);
3041 * Flushing deferred I/Os must be done after targets are resumed
3042 * so that mapping of targets can work correctly.
3043 * Request-based dm is queueing the deferred I/Os in its request_queue.
3045 if (dm_request_based(md))
3046 dm_start_queue(md->queue);
3053 int dm_resume(struct mapped_device *md)
3056 struct dm_table *map = NULL;
3060 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
3062 if (!dm_suspended_md(md))
3065 if (dm_suspended_internally_md(md)) {
3066 /* already internally suspended, wait for internal resume */
3067 mutex_unlock(&md->suspend_lock);
3068 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
3074 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3075 if (!map || !dm_table_get_size(map))
3078 r = __dm_resume(md, map);
3082 clear_bit(DMF_SUSPENDED, &md->flags);
3084 mutex_unlock(&md->suspend_lock);
3090 * Internal suspend/resume works like userspace-driven suspend. It waits
3091 * until all bios finish and prevents issuing new bios to the target drivers.
3092 * It may be used only from the kernel.
3095 static void __dm_internal_suspend(struct mapped_device *md, unsigned int suspend_flags)
3097 struct dm_table *map = NULL;
3099 lockdep_assert_held(&md->suspend_lock);
3101 if (md->internal_suspend_count++)
3102 return; /* nested internal suspend */
3104 if (dm_suspended_md(md)) {
3105 set_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3106 return; /* nest suspend */
3109 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3112 * Using TASK_UNINTERRUPTIBLE because only NOFLUSH internal suspend is
3113 * supported. Properly supporting a TASK_INTERRUPTIBLE internal suspend
3114 * would require changing .presuspend to return an error -- avoid this
3115 * until there is a need for more elaborate variants of internal suspend.
3117 (void) __dm_suspend(md, map, suspend_flags, TASK_UNINTERRUPTIBLE,
3118 DMF_SUSPENDED_INTERNALLY);
3120 set_bit(DMF_POST_SUSPENDING, &md->flags);
3121 dm_table_postsuspend_targets(map);
3122 clear_bit(DMF_POST_SUSPENDING, &md->flags);
3125 static void __dm_internal_resume(struct mapped_device *md)
3128 struct dm_table *map;
3130 BUG_ON(!md->internal_suspend_count);
3132 if (--md->internal_suspend_count)
3133 return; /* resume from nested internal suspend */
3135 if (dm_suspended_md(md))
3136 goto done; /* resume from nested suspend */
3138 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3139 r = __dm_resume(md, map);
3142 * If a preresume method of some target failed, we are in a
3143 * tricky situation. We can't return an error to the caller. We
3144 * can't fake success because then the "resume" and
3145 * "postsuspend" methods would not be paired correctly, and it
3146 * would break various targets, for example it would cause list
3147 * corruption in the "origin" target.
3149 * So, we fake normal suspend here, to make sure that the
3150 * "resume" and "postsuspend" methods will be paired correctly.
3152 DMERR("Preresume method failed: %d", r);
3153 set_bit(DMF_SUSPENDED, &md->flags);
3156 clear_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3157 smp_mb__after_atomic();
3158 wake_up_bit(&md->flags, DMF_SUSPENDED_INTERNALLY);
3161 void dm_internal_suspend_noflush(struct mapped_device *md)
3163 mutex_lock(&md->suspend_lock);
3164 __dm_internal_suspend(md, DM_SUSPEND_NOFLUSH_FLAG);
3165 mutex_unlock(&md->suspend_lock);
3167 EXPORT_SYMBOL_GPL(dm_internal_suspend_noflush);
3169 void dm_internal_resume(struct mapped_device *md)
3171 mutex_lock(&md->suspend_lock);
3172 __dm_internal_resume(md);
3173 mutex_unlock(&md->suspend_lock);
3175 EXPORT_SYMBOL_GPL(dm_internal_resume);
3178 * Fast variants of internal suspend/resume hold md->suspend_lock,
3179 * which prevents interaction with userspace-driven suspend.
3182 void dm_internal_suspend_fast(struct mapped_device *md)
3184 mutex_lock(&md->suspend_lock);
3185 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
3188 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
3189 synchronize_srcu(&md->io_barrier);
3190 flush_workqueue(md->wq);
3191 dm_wait_for_completion(md, TASK_UNINTERRUPTIBLE);
3193 EXPORT_SYMBOL_GPL(dm_internal_suspend_fast);
3195 void dm_internal_resume_fast(struct mapped_device *md)
3197 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
3203 mutex_unlock(&md->suspend_lock);
3205 EXPORT_SYMBOL_GPL(dm_internal_resume_fast);
3208 *---------------------------------------------------------------
3209 * Event notification.
3210 *---------------------------------------------------------------
3212 int dm_kobject_uevent(struct mapped_device *md, enum kobject_action action,
3213 unsigned int cookie, bool need_resize_uevent)
3216 unsigned int noio_flag;
3217 char udev_cookie[DM_COOKIE_LENGTH];
3218 char *envp[3] = { NULL, NULL, NULL };
3219 char **envpp = envp;
3221 snprintf(udev_cookie, DM_COOKIE_LENGTH, "%s=%u",
3222 DM_COOKIE_ENV_VAR_NAME, cookie);
3223 *envpp++ = udev_cookie;
3225 if (need_resize_uevent) {
3226 *envpp++ = "RESIZE=1";
3229 noio_flag = memalloc_noio_save();
3231 r = kobject_uevent_env(&disk_to_dev(md->disk)->kobj, action, envp);
3233 memalloc_noio_restore(noio_flag);
3238 uint32_t dm_next_uevent_seq(struct mapped_device *md)
3240 return atomic_add_return(1, &md->uevent_seq);
3243 uint32_t dm_get_event_nr(struct mapped_device *md)
3245 return atomic_read(&md->event_nr);
3248 int dm_wait_event(struct mapped_device *md, int event_nr)
3250 return wait_event_interruptible(md->eventq,
3251 (event_nr != atomic_read(&md->event_nr)));
3254 void dm_uevent_add(struct mapped_device *md, struct list_head *elist)
3256 unsigned long flags;
3258 spin_lock_irqsave(&md->uevent_lock, flags);
3259 list_add(elist, &md->uevent_list);
3260 spin_unlock_irqrestore(&md->uevent_lock, flags);
3264 * The gendisk is only valid as long as you have a reference
3267 struct gendisk *dm_disk(struct mapped_device *md)
3271 EXPORT_SYMBOL_GPL(dm_disk);
3273 struct kobject *dm_kobject(struct mapped_device *md)
3275 return &md->kobj_holder.kobj;
3278 struct mapped_device *dm_get_from_kobject(struct kobject *kobj)
3280 struct mapped_device *md;
3282 md = container_of(kobj, struct mapped_device, kobj_holder.kobj);
3284 spin_lock(&_minor_lock);
3285 if (test_bit(DMF_FREEING, &md->flags) || dm_deleting_md(md)) {
3291 spin_unlock(&_minor_lock);
3296 int dm_suspended_md(struct mapped_device *md)
3298 return test_bit(DMF_SUSPENDED, &md->flags);
3301 static int dm_post_suspending_md(struct mapped_device *md)
3303 return test_bit(DMF_POST_SUSPENDING, &md->flags);
3306 int dm_suspended_internally_md(struct mapped_device *md)
3308 return test_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3311 int dm_test_deferred_remove_flag(struct mapped_device *md)
3313 return test_bit(DMF_DEFERRED_REMOVE, &md->flags);
3316 int dm_suspended(struct dm_target *ti)
3318 return dm_suspended_md(ti->table->md);
3320 EXPORT_SYMBOL_GPL(dm_suspended);
3322 int dm_post_suspending(struct dm_target *ti)
3324 return dm_post_suspending_md(ti->table->md);
3326 EXPORT_SYMBOL_GPL(dm_post_suspending);
3328 int dm_noflush_suspending(struct dm_target *ti)
3330 return __noflush_suspending(ti->table->md);
3332 EXPORT_SYMBOL_GPL(dm_noflush_suspending);
3334 void dm_free_md_mempools(struct dm_md_mempools *pools)
3339 bioset_exit(&pools->bs);
3340 bioset_exit(&pools->io_bs);
3353 struct pr_keys *read_keys;
3354 struct pr_held_reservation *rsv;
3357 static int dm_call_pr(struct block_device *bdev, iterate_devices_callout_fn fn,
3360 struct mapped_device *md = bdev->bd_disk->private_data;
3361 struct dm_table *table;
3362 struct dm_target *ti;
3363 int ret = -ENOTTY, srcu_idx;
3365 table = dm_get_live_table(md, &srcu_idx);
3366 if (!table || !dm_table_get_size(table))
3369 /* We only support devices that have a single target */
3370 if (table->num_targets != 1)
3372 ti = dm_table_get_target(table, 0);
3374 if (dm_suspended_md(md)) {
3380 if (!ti->type->iterate_devices)
3383 ti->type->iterate_devices(ti, fn, pr);
3386 dm_put_live_table(md, srcu_idx);
3391 * For register / unregister we need to manually call out to every path.
3393 static int __dm_pr_register(struct dm_target *ti, struct dm_dev *dev,
3394 sector_t start, sector_t len, void *data)
3396 struct dm_pr *pr = data;
3397 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3400 if (!ops || !ops->pr_register) {
3401 pr->ret = -EOPNOTSUPP;
3405 ret = ops->pr_register(dev->bdev, pr->old_key, pr->new_key, pr->flags);
3418 static int dm_pr_register(struct block_device *bdev, u64 old_key, u64 new_key,
3430 ret = dm_call_pr(bdev, __dm_pr_register, &pr);
3432 /* Didn't even get to register a path */
3443 /* unregister all paths if we failed to register any path */
3444 pr.old_key = new_key;
3447 pr.fail_early = false;
3448 (void) dm_call_pr(bdev, __dm_pr_register, &pr);
3453 static int __dm_pr_reserve(struct dm_target *ti, struct dm_dev *dev,
3454 sector_t start, sector_t len, void *data)
3456 struct dm_pr *pr = data;
3457 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3459 if (!ops || !ops->pr_reserve) {
3460 pr->ret = -EOPNOTSUPP;
3464 pr->ret = ops->pr_reserve(dev->bdev, pr->old_key, pr->type, pr->flags);
3471 static int dm_pr_reserve(struct block_device *bdev, u64 key, enum pr_type type,
3478 .fail_early = false,
3483 ret = dm_call_pr(bdev, __dm_pr_reserve, &pr);
3491 * If there is a non-All Registrants type of reservation, the release must be
3492 * sent down the holding path. For the cases where there is no reservation or
3493 * the path is not the holder the device will also return success, so we must
3494 * try each path to make sure we got the correct path.
3496 static int __dm_pr_release(struct dm_target *ti, struct dm_dev *dev,
3497 sector_t start, sector_t len, void *data)
3499 struct dm_pr *pr = data;
3500 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3502 if (!ops || !ops->pr_release) {
3503 pr->ret = -EOPNOTSUPP;
3507 pr->ret = ops->pr_release(dev->bdev, pr->old_key, pr->type);
3514 static int dm_pr_release(struct block_device *bdev, u64 key, enum pr_type type)
3519 .fail_early = false,
3523 ret = dm_call_pr(bdev, __dm_pr_release, &pr);
3530 static int __dm_pr_preempt(struct dm_target *ti, struct dm_dev *dev,
3531 sector_t start, sector_t len, void *data)
3533 struct dm_pr *pr = data;
3534 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3536 if (!ops || !ops->pr_preempt) {
3537 pr->ret = -EOPNOTSUPP;
3541 pr->ret = ops->pr_preempt(dev->bdev, pr->old_key, pr->new_key, pr->type,
3549 static int dm_pr_preempt(struct block_device *bdev, u64 old_key, u64 new_key,
3550 enum pr_type type, bool abort)
3556 .fail_early = false,
3560 ret = dm_call_pr(bdev, __dm_pr_preempt, &pr);
3567 static int dm_pr_clear(struct block_device *bdev, u64 key)
3569 struct mapped_device *md = bdev->bd_disk->private_data;
3570 const struct pr_ops *ops;
3573 r = dm_prepare_ioctl(md, &srcu_idx, &bdev);
3577 ops = bdev->bd_disk->fops->pr_ops;
3578 if (ops && ops->pr_clear)
3579 r = ops->pr_clear(bdev, key);
3583 dm_unprepare_ioctl(md, srcu_idx);
3587 static int __dm_pr_read_keys(struct dm_target *ti, struct dm_dev *dev,
3588 sector_t start, sector_t len, void *data)
3590 struct dm_pr *pr = data;
3591 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3593 if (!ops || !ops->pr_read_keys) {
3594 pr->ret = -EOPNOTSUPP;
3598 pr->ret = ops->pr_read_keys(dev->bdev, pr->read_keys);
3605 static int dm_pr_read_keys(struct block_device *bdev, struct pr_keys *keys)
3612 ret = dm_call_pr(bdev, __dm_pr_read_keys, &pr);
3619 static int __dm_pr_read_reservation(struct dm_target *ti, struct dm_dev *dev,
3620 sector_t start, sector_t len, void *data)
3622 struct dm_pr *pr = data;
3623 const struct pr_ops *ops = dev->bdev->bd_disk->fops->pr_ops;
3625 if (!ops || !ops->pr_read_reservation) {
3626 pr->ret = -EOPNOTSUPP;
3630 pr->ret = ops->pr_read_reservation(dev->bdev, pr->rsv);
3637 static int dm_pr_read_reservation(struct block_device *bdev,
3638 struct pr_held_reservation *rsv)
3645 ret = dm_call_pr(bdev, __dm_pr_read_reservation, &pr);
3652 static const struct pr_ops dm_pr_ops = {
3653 .pr_register = dm_pr_register,
3654 .pr_reserve = dm_pr_reserve,
3655 .pr_release = dm_pr_release,
3656 .pr_preempt = dm_pr_preempt,
3657 .pr_clear = dm_pr_clear,
3658 .pr_read_keys = dm_pr_read_keys,
3659 .pr_read_reservation = dm_pr_read_reservation,
3662 static const struct block_device_operations dm_blk_dops = {
3663 .submit_bio = dm_submit_bio,
3664 .poll_bio = dm_poll_bio,
3665 .open = dm_blk_open,
3666 .release = dm_blk_close,
3667 .ioctl = dm_blk_ioctl,
3668 .getgeo = dm_blk_getgeo,
3669 .report_zones = dm_blk_report_zones,
3670 .pr_ops = &dm_pr_ops,
3671 .owner = THIS_MODULE
3674 static const struct block_device_operations dm_rq_blk_dops = {
3675 .open = dm_blk_open,
3676 .release = dm_blk_close,
3677 .ioctl = dm_blk_ioctl,
3678 .getgeo = dm_blk_getgeo,
3679 .pr_ops = &dm_pr_ops,
3680 .owner = THIS_MODULE
3683 static const struct dax_operations dm_dax_ops = {
3684 .direct_access = dm_dax_direct_access,
3685 .zero_page_range = dm_dax_zero_page_range,
3686 .recovery_write = dm_dax_recovery_write,
3692 module_init(dm_init);
3693 module_exit(dm_exit);
3695 module_param(major, uint, 0);
3696 MODULE_PARM_DESC(major, "The major number of the device mapper");
3698 module_param(reserved_bio_based_ios, uint, 0644);
3699 MODULE_PARM_DESC(reserved_bio_based_ios, "Reserved IOs in bio-based mempools");
3701 module_param(dm_numa_node, int, 0644);
3702 MODULE_PARM_DESC(dm_numa_node, "NUMA node for DM device memory allocations");
3704 module_param(swap_bios, int, 0644);
3705 MODULE_PARM_DESC(swap_bios, "Maximum allowed inflight swap IOs");
3707 MODULE_DESCRIPTION(DM_NAME " driver");
3709 MODULE_LICENSE("GPL");