2 * Copyright (C) 2001, 2002 Sistina Software (UK) Limited.
3 * Copyright (C) 2004-2008 Red Hat, Inc. All rights reserved.
5 * This file is released under the GPL.
11 #include <linux/init.h>
12 #include <linux/module.h>
13 #include <linux/mutex.h>
14 #include <linux/moduleparam.h>
15 #include <linux/blkpg.h>
16 #include <linux/bio.h>
17 #include <linux/mempool.h>
18 #include <linux/slab.h>
19 #include <linux/idr.h>
20 #include <linux/hdreg.h>
21 #include <linux/delay.h>
22 #include <linux/wait.h>
23 #include <linux/kthread.h>
24 #include <linux/ktime.h>
25 #include <linux/elevator.h> /* for rq_end_sector() */
26 #include <linux/blk-mq.h>
29 #include <trace/events/block.h>
31 #define DM_MSG_PREFIX "core"
35 * ratelimit state to be used in DMXXX_LIMIT().
37 DEFINE_RATELIMIT_STATE(dm_ratelimit_state,
38 DEFAULT_RATELIMIT_INTERVAL,
39 DEFAULT_RATELIMIT_BURST);
40 EXPORT_SYMBOL(dm_ratelimit_state);
44 * Cookies are numeric values sent with CHANGE and REMOVE
45 * uevents while resuming, removing or renaming the device.
47 #define DM_COOKIE_ENV_VAR_NAME "DM_COOKIE"
48 #define DM_COOKIE_LENGTH 24
50 static const char *_name = DM_NAME;
52 static unsigned int major = 0;
53 static unsigned int _major = 0;
55 static DEFINE_IDR(_minor_idr);
57 static DEFINE_SPINLOCK(_minor_lock);
59 static void do_deferred_remove(struct work_struct *w);
61 static DECLARE_WORK(deferred_remove_work, do_deferred_remove);
63 static struct workqueue_struct *deferred_remove_workqueue;
67 * One of these is allocated per bio.
70 struct mapped_device *md;
74 unsigned long start_time;
75 spinlock_t endio_lock;
76 struct dm_stats_aux stats_aux;
80 * For request-based dm.
81 * One of these is allocated per request.
83 struct dm_rq_target_io {
84 struct mapped_device *md;
86 struct request *orig, *clone;
87 struct kthread_work work;
90 struct dm_stats_aux stats_aux;
91 unsigned long duration_jiffies;
96 * For request-based dm - the bio clones we allocate are embedded in these
99 * We allocate these with bio_alloc_bioset, using the front_pad parameter when
100 * the bioset is created - this means the bio has to come at the end of the
103 struct dm_rq_clone_bio_info {
105 struct dm_rq_target_io *tio;
109 #define MINOR_ALLOCED ((void *)-1)
112 * Bits for the md->flags field.
114 #define DMF_BLOCK_IO_FOR_SUSPEND 0
115 #define DMF_SUSPENDED 1
117 #define DMF_FREEING 3
118 #define DMF_DELETING 4
119 #define DMF_NOFLUSH_SUSPENDING 5
120 #define DMF_DEFERRED_REMOVE 6
121 #define DMF_SUSPENDED_INTERNALLY 7
124 * A dummy definition to make RCU happy.
125 * struct dm_table should never be dereferenced in this file.
132 * Work processed by per-device workqueue.
134 struct mapped_device {
135 struct srcu_struct io_barrier;
136 struct mutex suspend_lock;
141 * The current mapping.
142 * Use dm_get_live_table{_fast} or take suspend_lock for
145 struct dm_table __rcu *map;
147 struct list_head table_devices;
148 struct mutex table_devices_lock;
152 struct request_queue *queue;
154 /* Protect queue and type against concurrent access. */
155 struct mutex type_lock;
157 struct dm_target *immutable_target;
158 struct target_type *immutable_target_type;
160 struct gendisk *disk;
166 * A list of ios that arrived while we were suspended.
169 wait_queue_head_t wait;
170 struct work_struct work;
171 struct bio_list deferred;
172 spinlock_t deferred_lock;
175 * Processing queue (flush)
177 struct workqueue_struct *wq;
180 * io objects are allocated from here.
191 wait_queue_head_t eventq;
193 struct list_head uevent_list;
194 spinlock_t uevent_lock; /* Protect access to uevent_list */
197 * freeze/thaw support require holding onto a super block
199 struct super_block *frozen_sb;
200 struct block_device *bdev;
202 /* forced geometry settings */
203 struct hd_geometry geometry;
205 /* kobject and completion */
206 struct dm_kobject_holder kobj_holder;
208 /* zero-length flush that will be cloned and submitted to targets */
209 struct bio flush_bio;
211 /* the number of internal suspends */
212 unsigned internal_suspend_count;
214 struct dm_stats stats;
216 struct kthread_worker kworker;
217 struct task_struct *kworker_task;
219 /* for request-based merge heuristic in dm_request_fn() */
220 unsigned seq_rq_merge_deadline_usecs;
222 sector_t last_rq_pos;
223 ktime_t last_rq_start_time;
225 /* for blk-mq request-based DM support */
226 struct blk_mq_tag_set tag_set;
230 #ifdef CONFIG_DM_MQ_DEFAULT
231 static bool use_blk_mq = true;
233 static bool use_blk_mq = false;
236 bool dm_use_blk_mq(struct mapped_device *md)
238 return md->use_blk_mq;
242 * For mempools pre-allocation at the table loading time.
244 struct dm_md_mempools {
250 struct table_device {
251 struct list_head list;
253 struct dm_dev dm_dev;
256 #define RESERVED_BIO_BASED_IOS 16
257 #define RESERVED_REQUEST_BASED_IOS 256
258 #define RESERVED_MAX_IOS 1024
259 static struct kmem_cache *_io_cache;
260 static struct kmem_cache *_rq_tio_cache;
261 static struct kmem_cache *_rq_cache;
264 * Bio-based DM's mempools' reserved IOs set by the user.
266 static unsigned reserved_bio_based_ios = RESERVED_BIO_BASED_IOS;
269 * Request-based DM's mempools' reserved IOs set by the user.
271 static unsigned reserved_rq_based_ios = RESERVED_REQUEST_BASED_IOS;
273 static unsigned __dm_get_module_param(unsigned *module_param,
274 unsigned def, unsigned max)
276 unsigned param = ACCESS_ONCE(*module_param);
277 unsigned modified_param = 0;
280 modified_param = def;
281 else if (param > max)
282 modified_param = max;
284 if (modified_param) {
285 (void)cmpxchg(module_param, param, modified_param);
286 param = modified_param;
292 unsigned dm_get_reserved_bio_based_ios(void)
294 return __dm_get_module_param(&reserved_bio_based_ios,
295 RESERVED_BIO_BASED_IOS, RESERVED_MAX_IOS);
297 EXPORT_SYMBOL_GPL(dm_get_reserved_bio_based_ios);
299 unsigned dm_get_reserved_rq_based_ios(void)
301 return __dm_get_module_param(&reserved_rq_based_ios,
302 RESERVED_REQUEST_BASED_IOS, RESERVED_MAX_IOS);
304 EXPORT_SYMBOL_GPL(dm_get_reserved_rq_based_ios);
306 static int __init local_init(void)
310 /* allocate a slab for the dm_ios */
311 _io_cache = KMEM_CACHE(dm_io, 0);
315 _rq_tio_cache = KMEM_CACHE(dm_rq_target_io, 0);
317 goto out_free_io_cache;
319 _rq_cache = kmem_cache_create("dm_clone_request", sizeof(struct request),
320 __alignof__(struct request), 0, NULL);
322 goto out_free_rq_tio_cache;
324 r = dm_uevent_init();
326 goto out_free_rq_cache;
328 deferred_remove_workqueue = alloc_workqueue("kdmremove", WQ_UNBOUND, 1);
329 if (!deferred_remove_workqueue) {
331 goto out_uevent_exit;
335 r = register_blkdev(_major, _name);
337 goto out_free_workqueue;
345 destroy_workqueue(deferred_remove_workqueue);
349 kmem_cache_destroy(_rq_cache);
350 out_free_rq_tio_cache:
351 kmem_cache_destroy(_rq_tio_cache);
353 kmem_cache_destroy(_io_cache);
358 static void local_exit(void)
360 flush_scheduled_work();
361 destroy_workqueue(deferred_remove_workqueue);
363 kmem_cache_destroy(_rq_cache);
364 kmem_cache_destroy(_rq_tio_cache);
365 kmem_cache_destroy(_io_cache);
366 unregister_blkdev(_major, _name);
371 DMINFO("cleaned up");
374 static int (*_inits[])(void) __initdata = {
385 static void (*_exits[])(void) = {
396 static int __init dm_init(void)
398 const int count = ARRAY_SIZE(_inits);
402 for (i = 0; i < count; i++) {
417 static void __exit dm_exit(void)
419 int i = ARRAY_SIZE(_exits);
425 * Should be empty by this point.
427 idr_destroy(&_minor_idr);
431 * Block device functions
433 int dm_deleting_md(struct mapped_device *md)
435 return test_bit(DMF_DELETING, &md->flags);
438 static int dm_blk_open(struct block_device *bdev, fmode_t mode)
440 struct mapped_device *md;
442 spin_lock(&_minor_lock);
444 md = bdev->bd_disk->private_data;
448 if (test_bit(DMF_FREEING, &md->flags) ||
449 dm_deleting_md(md)) {
455 atomic_inc(&md->open_count);
457 spin_unlock(&_minor_lock);
459 return md ? 0 : -ENXIO;
462 static void dm_blk_close(struct gendisk *disk, fmode_t mode)
464 struct mapped_device *md;
466 spin_lock(&_minor_lock);
468 md = disk->private_data;
472 if (atomic_dec_and_test(&md->open_count) &&
473 (test_bit(DMF_DEFERRED_REMOVE, &md->flags)))
474 queue_work(deferred_remove_workqueue, &deferred_remove_work);
478 spin_unlock(&_minor_lock);
481 int dm_open_count(struct mapped_device *md)
483 return atomic_read(&md->open_count);
487 * Guarantees nothing is using the device before it's deleted.
489 int dm_lock_for_deletion(struct mapped_device *md, bool mark_deferred, bool only_deferred)
493 spin_lock(&_minor_lock);
495 if (dm_open_count(md)) {
498 set_bit(DMF_DEFERRED_REMOVE, &md->flags);
499 } else if (only_deferred && !test_bit(DMF_DEFERRED_REMOVE, &md->flags))
502 set_bit(DMF_DELETING, &md->flags);
504 spin_unlock(&_minor_lock);
509 int dm_cancel_deferred_remove(struct mapped_device *md)
513 spin_lock(&_minor_lock);
515 if (test_bit(DMF_DELETING, &md->flags))
518 clear_bit(DMF_DEFERRED_REMOVE, &md->flags);
520 spin_unlock(&_minor_lock);
525 static void do_deferred_remove(struct work_struct *w)
527 dm_deferred_remove();
530 sector_t dm_get_size(struct mapped_device *md)
532 return get_capacity(md->disk);
535 struct request_queue *dm_get_md_queue(struct mapped_device *md)
540 struct dm_stats *dm_get_stats(struct mapped_device *md)
545 static int dm_blk_getgeo(struct block_device *bdev, struct hd_geometry *geo)
547 struct mapped_device *md = bdev->bd_disk->private_data;
549 return dm_get_geometry(md, geo);
552 static int dm_grab_bdev_for_ioctl(struct mapped_device *md,
553 struct block_device **bdev,
556 struct dm_target *tgt;
557 struct dm_table *map;
562 map = dm_get_live_table(md, &srcu_idx);
563 if (!map || !dm_table_get_size(map))
566 /* We only support devices that have a single target */
567 if (dm_table_get_num_targets(map) != 1)
570 tgt = dm_table_get_target(map, 0);
571 if (!tgt->type->prepare_ioctl)
574 if (dm_suspended_md(md)) {
579 r = tgt->type->prepare_ioctl(tgt, bdev, mode);
584 dm_put_live_table(md, srcu_idx);
588 dm_put_live_table(md, srcu_idx);
589 if (r == -ENOTCONN && !fatal_signal_pending(current)) {
596 static int dm_blk_ioctl(struct block_device *bdev, fmode_t mode,
597 unsigned int cmd, unsigned long arg)
599 struct mapped_device *md = bdev->bd_disk->private_data;
602 r = dm_grab_bdev_for_ioctl(md, &bdev, &mode);
608 * Target determined this ioctl is being issued against
609 * a logical partition of the parent bdev; so extra
610 * validation is needed.
612 r = scsi_verify_blk_ioctl(NULL, cmd);
617 r = __blkdev_driver_ioctl(bdev, mode, cmd, arg);
623 static struct dm_io *alloc_io(struct mapped_device *md)
625 return mempool_alloc(md->io_pool, GFP_NOIO);
628 static void free_io(struct mapped_device *md, struct dm_io *io)
630 mempool_free(io, md->io_pool);
633 static void free_tio(struct mapped_device *md, struct dm_target_io *tio)
635 bio_put(&tio->clone);
638 static struct dm_rq_target_io *alloc_rq_tio(struct mapped_device *md,
641 return mempool_alloc(md->io_pool, gfp_mask);
644 static void free_rq_tio(struct dm_rq_target_io *tio)
646 mempool_free(tio, tio->md->io_pool);
649 static struct request *alloc_clone_request(struct mapped_device *md,
652 return mempool_alloc(md->rq_pool, gfp_mask);
655 static void free_clone_request(struct mapped_device *md, struct request *rq)
657 mempool_free(rq, md->rq_pool);
660 static int md_in_flight(struct mapped_device *md)
662 return atomic_read(&md->pending[READ]) +
663 atomic_read(&md->pending[WRITE]);
666 static void start_io_acct(struct dm_io *io)
668 struct mapped_device *md = io->md;
669 struct bio *bio = io->bio;
671 int rw = bio_data_dir(bio);
673 io->start_time = jiffies;
675 cpu = part_stat_lock();
676 part_round_stats(cpu, &dm_disk(md)->part0);
678 atomic_set(&dm_disk(md)->part0.in_flight[rw],
679 atomic_inc_return(&md->pending[rw]));
681 if (unlikely(dm_stats_used(&md->stats)))
682 dm_stats_account_io(&md->stats, bio->bi_rw, bio->bi_iter.bi_sector,
683 bio_sectors(bio), false, 0, &io->stats_aux);
686 static void end_io_acct(struct dm_io *io)
688 struct mapped_device *md = io->md;
689 struct bio *bio = io->bio;
690 unsigned long duration = jiffies - io->start_time;
692 int rw = bio_data_dir(bio);
694 generic_end_io_acct(rw, &dm_disk(md)->part0, io->start_time);
696 if (unlikely(dm_stats_used(&md->stats)))
697 dm_stats_account_io(&md->stats, bio->bi_rw, bio->bi_iter.bi_sector,
698 bio_sectors(bio), true, duration, &io->stats_aux);
701 * After this is decremented the bio must not be touched if it is
704 pending = atomic_dec_return(&md->pending[rw]);
705 atomic_set(&dm_disk(md)->part0.in_flight[rw], pending);
706 pending += atomic_read(&md->pending[rw^0x1]);
708 /* nudge anyone waiting on suspend queue */
714 * Add the bio to the list of deferred io.
716 static void queue_io(struct mapped_device *md, struct bio *bio)
720 spin_lock_irqsave(&md->deferred_lock, flags);
721 bio_list_add(&md->deferred, bio);
722 spin_unlock_irqrestore(&md->deferred_lock, flags);
723 queue_work(md->wq, &md->work);
727 * Everyone (including functions in this file), should use this
728 * function to access the md->map field, and make sure they call
729 * dm_put_live_table() when finished.
731 struct dm_table *dm_get_live_table(struct mapped_device *md, int *srcu_idx) __acquires(md->io_barrier)
733 *srcu_idx = srcu_read_lock(&md->io_barrier);
735 return srcu_dereference(md->map, &md->io_barrier);
738 void dm_put_live_table(struct mapped_device *md, int srcu_idx) __releases(md->io_barrier)
740 srcu_read_unlock(&md->io_barrier, srcu_idx);
743 void dm_sync_table(struct mapped_device *md)
745 synchronize_srcu(&md->io_barrier);
746 synchronize_rcu_expedited();
750 * A fast alternative to dm_get_live_table/dm_put_live_table.
751 * The caller must not block between these two functions.
753 static struct dm_table *dm_get_live_table_fast(struct mapped_device *md) __acquires(RCU)
756 return rcu_dereference(md->map);
759 static void dm_put_live_table_fast(struct mapped_device *md) __releases(RCU)
765 * Open a table device so we can use it as a map destination.
767 static int open_table_device(struct table_device *td, dev_t dev,
768 struct mapped_device *md)
770 static char *_claim_ptr = "I belong to device-mapper";
771 struct block_device *bdev;
775 BUG_ON(td->dm_dev.bdev);
777 bdev = blkdev_get_by_dev(dev, td->dm_dev.mode | FMODE_EXCL, _claim_ptr);
779 return PTR_ERR(bdev);
781 r = bd_link_disk_holder(bdev, dm_disk(md));
783 blkdev_put(bdev, td->dm_dev.mode | FMODE_EXCL);
787 td->dm_dev.bdev = bdev;
792 * Close a table device that we've been using.
794 static void close_table_device(struct table_device *td, struct mapped_device *md)
796 if (!td->dm_dev.bdev)
799 bd_unlink_disk_holder(td->dm_dev.bdev, dm_disk(md));
800 blkdev_put(td->dm_dev.bdev, td->dm_dev.mode | FMODE_EXCL);
801 td->dm_dev.bdev = NULL;
804 static struct table_device *find_table_device(struct list_head *l, dev_t dev,
806 struct table_device *td;
808 list_for_each_entry(td, l, list)
809 if (td->dm_dev.bdev->bd_dev == dev && td->dm_dev.mode == mode)
815 int dm_get_table_device(struct mapped_device *md, dev_t dev, fmode_t mode,
816 struct dm_dev **result) {
818 struct table_device *td;
820 mutex_lock(&md->table_devices_lock);
821 td = find_table_device(&md->table_devices, dev, mode);
823 td = kmalloc(sizeof(*td), GFP_KERNEL);
825 mutex_unlock(&md->table_devices_lock);
829 td->dm_dev.mode = mode;
830 td->dm_dev.bdev = NULL;
832 if ((r = open_table_device(td, dev, md))) {
833 mutex_unlock(&md->table_devices_lock);
838 format_dev_t(td->dm_dev.name, dev);
840 atomic_set(&td->count, 0);
841 list_add(&td->list, &md->table_devices);
843 atomic_inc(&td->count);
844 mutex_unlock(&md->table_devices_lock);
846 *result = &td->dm_dev;
849 EXPORT_SYMBOL_GPL(dm_get_table_device);
851 void dm_put_table_device(struct mapped_device *md, struct dm_dev *d)
853 struct table_device *td = container_of(d, struct table_device, dm_dev);
855 mutex_lock(&md->table_devices_lock);
856 if (atomic_dec_and_test(&td->count)) {
857 close_table_device(td, md);
861 mutex_unlock(&md->table_devices_lock);
863 EXPORT_SYMBOL(dm_put_table_device);
865 static void free_table_devices(struct list_head *devices)
867 struct list_head *tmp, *next;
869 list_for_each_safe(tmp, next, devices) {
870 struct table_device *td = list_entry(tmp, struct table_device, list);
872 DMWARN("dm_destroy: %s still exists with %d references",
873 td->dm_dev.name, atomic_read(&td->count));
879 * Get the geometry associated with a dm device
881 int dm_get_geometry(struct mapped_device *md, struct hd_geometry *geo)
889 * Set the geometry of a device.
891 int dm_set_geometry(struct mapped_device *md, struct hd_geometry *geo)
893 sector_t sz = (sector_t)geo->cylinders * geo->heads * geo->sectors;
895 if (geo->start > sz) {
896 DMWARN("Start sector is beyond the geometry limits.");
905 /*-----------------------------------------------------------------
907 * A more elegant soln is in the works that uses the queue
908 * merge fn, unfortunately there are a couple of changes to
909 * the block layer that I want to make for this. So in the
910 * interests of getting something for people to use I give
911 * you this clearly demarcated crap.
912 *---------------------------------------------------------------*/
914 static int __noflush_suspending(struct mapped_device *md)
916 return test_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
920 * Decrements the number of outstanding ios that a bio has been
921 * cloned into, completing the original io if necc.
923 static void dec_pending(struct dm_io *io, int error)
928 struct mapped_device *md = io->md;
930 /* Push-back supersedes any I/O errors */
931 if (unlikely(error)) {
932 spin_lock_irqsave(&io->endio_lock, flags);
933 if (!(io->error > 0 && __noflush_suspending(md)))
935 spin_unlock_irqrestore(&io->endio_lock, flags);
938 if (atomic_dec_and_test(&io->io_count)) {
939 if (io->error == DM_ENDIO_REQUEUE) {
941 * Target requested pushing back the I/O.
943 spin_lock_irqsave(&md->deferred_lock, flags);
944 if (__noflush_suspending(md))
945 bio_list_add_head(&md->deferred, io->bio);
947 /* noflush suspend was interrupted. */
949 spin_unlock_irqrestore(&md->deferred_lock, flags);
952 io_error = io->error;
957 if (io_error == DM_ENDIO_REQUEUE)
960 if ((bio->bi_rw & REQ_FLUSH) && bio->bi_iter.bi_size) {
962 * Preflush done for flush with data, reissue
965 bio->bi_rw &= ~REQ_FLUSH;
968 /* done with normal IO or empty flush */
969 trace_block_bio_complete(md->queue, bio, io_error);
970 bio->bi_error = io_error;
976 static void disable_write_same(struct mapped_device *md)
978 struct queue_limits *limits = dm_get_queue_limits(md);
980 /* device doesn't really support WRITE SAME, disable it */
981 limits->max_write_same_sectors = 0;
984 static void clone_endio(struct bio *bio)
986 int error = bio->bi_error;
988 struct dm_target_io *tio = container_of(bio, struct dm_target_io, clone);
989 struct dm_io *io = tio->io;
990 struct mapped_device *md = tio->io->md;
991 dm_endio_fn endio = tio->ti->type->end_io;
994 r = endio(tio->ti, bio, error);
995 if (r < 0 || r == DM_ENDIO_REQUEUE)
997 * error and requeue request are handled
1001 else if (r == DM_ENDIO_INCOMPLETE)
1002 /* The target will handle the io */
1005 DMWARN("unimplemented target endio return value: %d", r);
1010 if (unlikely(r == -EREMOTEIO && (bio->bi_rw & REQ_WRITE_SAME) &&
1011 !bdev_get_queue(bio->bi_bdev)->limits.max_write_same_sectors))
1012 disable_write_same(md);
1015 dec_pending(io, error);
1019 * Partial completion handling for request-based dm
1021 static void end_clone_bio(struct bio *clone)
1023 struct dm_rq_clone_bio_info *info =
1024 container_of(clone, struct dm_rq_clone_bio_info, clone);
1025 struct dm_rq_target_io *tio = info->tio;
1026 struct bio *bio = info->orig;
1027 unsigned int nr_bytes = info->orig->bi_iter.bi_size;
1028 int error = clone->bi_error;
1034 * An error has already been detected on the request.
1035 * Once error occurred, just let clone->end_io() handle
1041 * Don't notice the error to the upper layer yet.
1042 * The error handling decision is made by the target driver,
1043 * when the request is completed.
1050 * I/O for the bio successfully completed.
1051 * Notice the data completion to the upper layer.
1055 * bios are processed from the head of the list.
1056 * So the completing bio should always be rq->bio.
1057 * If it's not, something wrong is happening.
1059 if (tio->orig->bio != bio)
1060 DMERR("bio completion is going in the middle of the request");
1063 * Update the original request.
1064 * Do not use blk_end_request() here, because it may complete
1065 * the original request before the clone, and break the ordering.
1067 blk_update_request(tio->orig, 0, nr_bytes);
1070 static struct dm_rq_target_io *tio_from_request(struct request *rq)
1072 return (rq->q->mq_ops ? blk_mq_rq_to_pdu(rq) : rq->special);
1075 static void rq_end_stats(struct mapped_device *md, struct request *orig)
1077 if (unlikely(dm_stats_used(&md->stats))) {
1078 struct dm_rq_target_io *tio = tio_from_request(orig);
1079 tio->duration_jiffies = jiffies - tio->duration_jiffies;
1080 dm_stats_account_io(&md->stats, orig->cmd_flags, blk_rq_pos(orig),
1081 tio->n_sectors, true, tio->duration_jiffies,
1087 * Don't touch any member of the md after calling this function because
1088 * the md may be freed in dm_put() at the end of this function.
1089 * Or do dm_get() before calling this function and dm_put() later.
1091 static void rq_completed(struct mapped_device *md, int rw, bool run_queue)
1093 atomic_dec(&md->pending[rw]);
1095 /* nudge anyone waiting on suspend queue */
1096 if (!md_in_flight(md))
1100 * Run this off this callpath, as drivers could invoke end_io while
1101 * inside their request_fn (and holding the queue lock). Calling
1102 * back into ->request_fn() could deadlock attempting to grab the
1105 if (!md->queue->mq_ops && run_queue)
1106 blk_run_queue_async(md->queue);
1109 * dm_put() must be at the end of this function. See the comment above
1114 static void free_rq_clone(struct request *clone)
1116 struct dm_rq_target_io *tio = clone->end_io_data;
1117 struct mapped_device *md = tio->md;
1119 blk_rq_unprep_clone(clone);
1121 if (md->type == DM_TYPE_MQ_REQUEST_BASED)
1122 /* stacked on blk-mq queue(s) */
1123 tio->ti->type->release_clone_rq(clone);
1124 else if (!md->queue->mq_ops)
1125 /* request_fn queue stacked on request_fn queue(s) */
1126 free_clone_request(md, clone);
1128 * NOTE: for the blk-mq queue stacked on request_fn queue(s) case:
1129 * no need to call free_clone_request() because we leverage blk-mq by
1130 * allocating the clone at the end of the blk-mq pdu (see: clone_rq)
1133 if (!md->queue->mq_ops)
1138 * Complete the clone and the original request.
1139 * Must be called without clone's queue lock held,
1140 * see end_clone_request() for more details.
1142 static void dm_end_request(struct request *clone, int error)
1144 int rw = rq_data_dir(clone);
1145 struct dm_rq_target_io *tio = clone->end_io_data;
1146 struct mapped_device *md = tio->md;
1147 struct request *rq = tio->orig;
1149 if (rq->cmd_type == REQ_TYPE_BLOCK_PC) {
1150 rq->errors = clone->errors;
1151 rq->resid_len = clone->resid_len;
1155 * We are using the sense buffer of the original
1157 * So setting the length of the sense data is enough.
1159 rq->sense_len = clone->sense_len;
1162 free_rq_clone(clone);
1163 rq_end_stats(md, rq);
1165 blk_end_request_all(rq, error);
1167 blk_mq_end_request(rq, error);
1168 rq_completed(md, rw, true);
1171 static void dm_unprep_request(struct request *rq)
1173 struct dm_rq_target_io *tio = tio_from_request(rq);
1174 struct request *clone = tio->clone;
1176 if (!rq->q->mq_ops) {
1178 rq->cmd_flags &= ~REQ_DONTPREP;
1182 free_rq_clone(clone);
1183 else if (!tio->md->queue->mq_ops)
1188 * Requeue the original request of a clone.
1190 static void old_requeue_request(struct request *rq)
1192 struct request_queue *q = rq->q;
1193 unsigned long flags;
1195 spin_lock_irqsave(q->queue_lock, flags);
1196 blk_requeue_request(q, rq);
1197 blk_run_queue_async(q);
1198 spin_unlock_irqrestore(q->queue_lock, flags);
1201 static void dm_requeue_original_request(struct mapped_device *md,
1204 int rw = rq_data_dir(rq);
1206 dm_unprep_request(rq);
1208 rq_end_stats(md, rq);
1210 old_requeue_request(rq);
1212 blk_mq_requeue_request(rq);
1213 blk_mq_kick_requeue_list(rq->q);
1216 rq_completed(md, rw, false);
1219 static void old_stop_queue(struct request_queue *q)
1221 unsigned long flags;
1223 if (blk_queue_stopped(q))
1226 spin_lock_irqsave(q->queue_lock, flags);
1228 spin_unlock_irqrestore(q->queue_lock, flags);
1231 static void stop_queue(struct request_queue *q)
1236 blk_mq_stop_hw_queues(q);
1239 static void old_start_queue(struct request_queue *q)
1241 unsigned long flags;
1243 spin_lock_irqsave(q->queue_lock, flags);
1244 if (blk_queue_stopped(q))
1246 spin_unlock_irqrestore(q->queue_lock, flags);
1249 static void start_queue(struct request_queue *q)
1254 blk_mq_start_stopped_hw_queues(q, true);
1257 static void dm_done(struct request *clone, int error, bool mapped)
1260 struct dm_rq_target_io *tio = clone->end_io_data;
1261 dm_request_endio_fn rq_end_io = NULL;
1264 rq_end_io = tio->ti->type->rq_end_io;
1266 if (mapped && rq_end_io)
1267 r = rq_end_io(tio->ti, clone, error, &tio->info);
1270 if (unlikely(r == -EREMOTEIO && (clone->cmd_flags & REQ_WRITE_SAME) &&
1271 !clone->q->limits.max_write_same_sectors))
1272 disable_write_same(tio->md);
1275 /* The target wants to complete the I/O */
1276 dm_end_request(clone, r);
1277 else if (r == DM_ENDIO_INCOMPLETE)
1278 /* The target will handle the I/O */
1280 else if (r == DM_ENDIO_REQUEUE)
1281 /* The target wants to requeue the I/O */
1282 dm_requeue_original_request(tio->md, tio->orig);
1284 DMWARN("unimplemented target endio return value: %d", r);
1290 * Request completion handler for request-based dm
1292 static void dm_softirq_done(struct request *rq)
1295 struct dm_rq_target_io *tio = tio_from_request(rq);
1296 struct request *clone = tio->clone;
1300 rq_end_stats(tio->md, rq);
1301 rw = rq_data_dir(rq);
1302 if (!rq->q->mq_ops) {
1303 blk_end_request_all(rq, tio->error);
1304 rq_completed(tio->md, rw, false);
1307 blk_mq_end_request(rq, tio->error);
1308 rq_completed(tio->md, rw, false);
1313 if (rq->cmd_flags & REQ_FAILED)
1316 dm_done(clone, tio->error, mapped);
1320 * Complete the clone and the original request with the error status
1321 * through softirq context.
1323 static void dm_complete_request(struct request *rq, int error)
1325 struct dm_rq_target_io *tio = tio_from_request(rq);
1329 blk_complete_request(rq);
1331 blk_mq_complete_request(rq, error);
1335 * Complete the not-mapped clone and the original request with the error status
1336 * through softirq context.
1337 * Target's rq_end_io() function isn't called.
1338 * This may be used when the target's map_rq() or clone_and_map_rq() functions fail.
1340 static void dm_kill_unmapped_request(struct request *rq, int error)
1342 rq->cmd_flags |= REQ_FAILED;
1343 dm_complete_request(rq, error);
1347 * Called with the clone's queue lock held (for non-blk-mq)
1349 static void end_clone_request(struct request *clone, int error)
1351 struct dm_rq_target_io *tio = clone->end_io_data;
1353 if (!clone->q->mq_ops) {
1355 * For just cleaning up the information of the queue in which
1356 * the clone was dispatched.
1357 * The clone is *NOT* freed actually here because it is alloced
1358 * from dm own mempool (REQ_ALLOCED isn't set).
1360 __blk_put_request(clone->q, clone);
1364 * Actual request completion is done in a softirq context which doesn't
1365 * hold the clone's queue lock. Otherwise, deadlock could occur because:
1366 * - another request may be submitted by the upper level driver
1367 * of the stacking during the completion
1368 * - the submission which requires queue lock may be done
1369 * against this clone's queue
1371 dm_complete_request(tio->orig, error);
1375 * Return maximum size of I/O possible at the supplied sector up to the current
1378 static sector_t max_io_len_target_boundary(sector_t sector, struct dm_target *ti)
1380 sector_t target_offset = dm_target_offset(ti, sector);
1382 return ti->len - target_offset;
1385 static sector_t max_io_len(sector_t sector, struct dm_target *ti)
1387 sector_t len = max_io_len_target_boundary(sector, ti);
1388 sector_t offset, max_len;
1391 * Does the target need to split even further?
1393 if (ti->max_io_len) {
1394 offset = dm_target_offset(ti, sector);
1395 if (unlikely(ti->max_io_len & (ti->max_io_len - 1)))
1396 max_len = sector_div(offset, ti->max_io_len);
1398 max_len = offset & (ti->max_io_len - 1);
1399 max_len = ti->max_io_len - max_len;
1408 int dm_set_target_max_io_len(struct dm_target *ti, sector_t len)
1410 if (len > UINT_MAX) {
1411 DMERR("Specified maximum size of target IO (%llu) exceeds limit (%u)",
1412 (unsigned long long)len, UINT_MAX);
1413 ti->error = "Maximum size of target IO is too large";
1417 ti->max_io_len = (uint32_t) len;
1421 EXPORT_SYMBOL_GPL(dm_set_target_max_io_len);
1424 * A target may call dm_accept_partial_bio only from the map routine. It is
1425 * allowed for all bio types except REQ_FLUSH.
1427 * dm_accept_partial_bio informs the dm that the target only wants to process
1428 * additional n_sectors sectors of the bio and the rest of the data should be
1429 * sent in a next bio.
1431 * A diagram that explains the arithmetics:
1432 * +--------------------+---------------+-------+
1434 * +--------------------+---------------+-------+
1436 * <-------------- *tio->len_ptr --------------->
1437 * <------- bi_size ------->
1440 * Region 1 was already iterated over with bio_advance or similar function.
1441 * (it may be empty if the target doesn't use bio_advance)
1442 * Region 2 is the remaining bio size that the target wants to process.
1443 * (it may be empty if region 1 is non-empty, although there is no reason
1445 * The target requires that region 3 is to be sent in the next bio.
1447 * If the target wants to receive multiple copies of the bio (via num_*bios, etc),
1448 * the partially processed part (the sum of regions 1+2) must be the same for all
1449 * copies of the bio.
1451 void dm_accept_partial_bio(struct bio *bio, unsigned n_sectors)
1453 struct dm_target_io *tio = container_of(bio, struct dm_target_io, clone);
1454 unsigned bi_size = bio->bi_iter.bi_size >> SECTOR_SHIFT;
1455 BUG_ON(bio->bi_rw & REQ_FLUSH);
1456 BUG_ON(bi_size > *tio->len_ptr);
1457 BUG_ON(n_sectors > bi_size);
1458 *tio->len_ptr -= bi_size - n_sectors;
1459 bio->bi_iter.bi_size = n_sectors << SECTOR_SHIFT;
1461 EXPORT_SYMBOL_GPL(dm_accept_partial_bio);
1463 static void __map_bio(struct dm_target_io *tio)
1467 struct mapped_device *md;
1468 struct bio *clone = &tio->clone;
1469 struct dm_target *ti = tio->ti;
1471 clone->bi_end_io = clone_endio;
1474 * Map the clone. If r == 0 we don't need to do
1475 * anything, the target has assumed ownership of
1478 atomic_inc(&tio->io->io_count);
1479 sector = clone->bi_iter.bi_sector;
1480 r = ti->type->map(ti, clone);
1481 if (r == DM_MAPIO_REMAPPED) {
1482 /* the bio has been remapped so dispatch it */
1484 trace_block_bio_remap(bdev_get_queue(clone->bi_bdev), clone,
1485 tio->io->bio->bi_bdev->bd_dev, sector);
1487 generic_make_request(clone);
1488 } else if (r < 0 || r == DM_MAPIO_REQUEUE) {
1489 /* error the io and bail out, or requeue it if needed */
1491 dec_pending(tio->io, r);
1493 } else if (r != DM_MAPIO_SUBMITTED) {
1494 DMWARN("unimplemented target map return value: %d", r);
1500 struct mapped_device *md;
1501 struct dm_table *map;
1505 unsigned sector_count;
1508 static void bio_setup_sector(struct bio *bio, sector_t sector, unsigned len)
1510 bio->bi_iter.bi_sector = sector;
1511 bio->bi_iter.bi_size = to_bytes(len);
1515 * Creates a bio that consists of range of complete bvecs.
1517 static void clone_bio(struct dm_target_io *tio, struct bio *bio,
1518 sector_t sector, unsigned len)
1520 struct bio *clone = &tio->clone;
1522 __bio_clone_fast(clone, bio);
1524 if (bio_integrity(bio))
1525 bio_integrity_clone(clone, bio, GFP_NOIO);
1527 bio_advance(clone, to_bytes(sector - clone->bi_iter.bi_sector));
1528 clone->bi_iter.bi_size = to_bytes(len);
1530 if (bio_integrity(bio))
1531 bio_integrity_trim(clone, 0, len);
1534 static struct dm_target_io *alloc_tio(struct clone_info *ci,
1535 struct dm_target *ti,
1536 unsigned target_bio_nr)
1538 struct dm_target_io *tio;
1541 clone = bio_alloc_bioset(GFP_NOIO, 0, ci->md->bs);
1542 tio = container_of(clone, struct dm_target_io, clone);
1546 tio->target_bio_nr = target_bio_nr;
1551 static void __clone_and_map_simple_bio(struct clone_info *ci,
1552 struct dm_target *ti,
1553 unsigned target_bio_nr, unsigned *len)
1555 struct dm_target_io *tio = alloc_tio(ci, ti, target_bio_nr);
1556 struct bio *clone = &tio->clone;
1560 __bio_clone_fast(clone, ci->bio);
1562 bio_setup_sector(clone, ci->sector, *len);
1567 static void __send_duplicate_bios(struct clone_info *ci, struct dm_target *ti,
1568 unsigned num_bios, unsigned *len)
1570 unsigned target_bio_nr;
1572 for (target_bio_nr = 0; target_bio_nr < num_bios; target_bio_nr++)
1573 __clone_and_map_simple_bio(ci, ti, target_bio_nr, len);
1576 static int __send_empty_flush(struct clone_info *ci)
1578 unsigned target_nr = 0;
1579 struct dm_target *ti;
1581 BUG_ON(bio_has_data(ci->bio));
1582 while ((ti = dm_table_get_target(ci->map, target_nr++)))
1583 __send_duplicate_bios(ci, ti, ti->num_flush_bios, NULL);
1588 static void __clone_and_map_data_bio(struct clone_info *ci, struct dm_target *ti,
1589 sector_t sector, unsigned *len)
1591 struct bio *bio = ci->bio;
1592 struct dm_target_io *tio;
1593 unsigned target_bio_nr;
1594 unsigned num_target_bios = 1;
1597 * Does the target want to receive duplicate copies of the bio?
1599 if (bio_data_dir(bio) == WRITE && ti->num_write_bios)
1600 num_target_bios = ti->num_write_bios(ti, bio);
1602 for (target_bio_nr = 0; target_bio_nr < num_target_bios; target_bio_nr++) {
1603 tio = alloc_tio(ci, ti, target_bio_nr);
1605 clone_bio(tio, bio, sector, *len);
1610 typedef unsigned (*get_num_bios_fn)(struct dm_target *ti);
1612 static unsigned get_num_discard_bios(struct dm_target *ti)
1614 return ti->num_discard_bios;
1617 static unsigned get_num_write_same_bios(struct dm_target *ti)
1619 return ti->num_write_same_bios;
1622 typedef bool (*is_split_required_fn)(struct dm_target *ti);
1624 static bool is_split_required_for_discard(struct dm_target *ti)
1626 return ti->split_discard_bios;
1629 static int __send_changing_extent_only(struct clone_info *ci,
1630 get_num_bios_fn get_num_bios,
1631 is_split_required_fn is_split_required)
1633 struct dm_target *ti;
1638 ti = dm_table_find_target(ci->map, ci->sector);
1639 if (!dm_target_is_valid(ti))
1643 * Even though the device advertised support for this type of
1644 * request, that does not mean every target supports it, and
1645 * reconfiguration might also have changed that since the
1646 * check was performed.
1648 num_bios = get_num_bios ? get_num_bios(ti) : 0;
1652 if (is_split_required && !is_split_required(ti))
1653 len = min((sector_t)ci->sector_count, max_io_len_target_boundary(ci->sector, ti));
1655 len = min((sector_t)ci->sector_count, max_io_len(ci->sector, ti));
1657 __send_duplicate_bios(ci, ti, num_bios, &len);
1660 } while (ci->sector_count -= len);
1665 static int __send_discard(struct clone_info *ci)
1667 return __send_changing_extent_only(ci, get_num_discard_bios,
1668 is_split_required_for_discard);
1671 static int __send_write_same(struct clone_info *ci)
1673 return __send_changing_extent_only(ci, get_num_write_same_bios, NULL);
1677 * Select the correct strategy for processing a non-flush bio.
1679 static int __split_and_process_non_flush(struct clone_info *ci)
1681 struct bio *bio = ci->bio;
1682 struct dm_target *ti;
1685 if (unlikely(bio->bi_rw & REQ_DISCARD))
1686 return __send_discard(ci);
1687 else if (unlikely(bio->bi_rw & REQ_WRITE_SAME))
1688 return __send_write_same(ci);
1690 ti = dm_table_find_target(ci->map, ci->sector);
1691 if (!dm_target_is_valid(ti))
1694 len = min_t(sector_t, max_io_len(ci->sector, ti), ci->sector_count);
1696 __clone_and_map_data_bio(ci, ti, ci->sector, &len);
1699 ci->sector_count -= len;
1705 * Entry point to split a bio into clones and submit them to the targets.
1707 static void __split_and_process_bio(struct mapped_device *md,
1708 struct dm_table *map, struct bio *bio)
1710 struct clone_info ci;
1713 if (unlikely(!map)) {
1720 ci.io = alloc_io(md);
1722 atomic_set(&ci.io->io_count, 1);
1725 spin_lock_init(&ci.io->endio_lock);
1726 ci.sector = bio->bi_iter.bi_sector;
1728 start_io_acct(ci.io);
1730 if (bio->bi_rw & REQ_FLUSH) {
1731 ci.bio = &ci.md->flush_bio;
1732 ci.sector_count = 0;
1733 error = __send_empty_flush(&ci);
1734 /* dec_pending submits any data associated with flush */
1737 ci.sector_count = bio_sectors(bio);
1738 while (ci.sector_count && !error)
1739 error = __split_and_process_non_flush(&ci);
1742 /* drop the extra reference count */
1743 dec_pending(ci.io, error);
1745 /*-----------------------------------------------------------------
1747 *---------------------------------------------------------------*/
1750 * The request function that just remaps the bio built up by
1753 static blk_qc_t dm_make_request(struct request_queue *q, struct bio *bio)
1755 int rw = bio_data_dir(bio);
1756 struct mapped_device *md = q->queuedata;
1758 struct dm_table *map;
1760 map = dm_get_live_table(md, &srcu_idx);
1762 generic_start_io_acct(rw, bio_sectors(bio), &dm_disk(md)->part0);
1764 /* if we're suspended, we have to queue this io for later */
1765 if (unlikely(test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags))) {
1766 dm_put_live_table(md, srcu_idx);
1768 if (bio_rw(bio) != READA)
1772 return BLK_QC_T_NONE;
1775 __split_and_process_bio(md, map, bio);
1776 dm_put_live_table(md, srcu_idx);
1777 return BLK_QC_T_NONE;
1780 int dm_request_based(struct mapped_device *md)
1782 return blk_queue_stackable(md->queue);
1785 static void dm_dispatch_clone_request(struct request *clone, struct request *rq)
1789 if (blk_queue_io_stat(clone->q))
1790 clone->cmd_flags |= REQ_IO_STAT;
1792 clone->start_time = jiffies;
1793 r = blk_insert_cloned_request(clone->q, clone);
1795 /* must complete clone in terms of original request */
1796 dm_complete_request(rq, r);
1799 static int dm_rq_bio_constructor(struct bio *bio, struct bio *bio_orig,
1802 struct dm_rq_target_io *tio = data;
1803 struct dm_rq_clone_bio_info *info =
1804 container_of(bio, struct dm_rq_clone_bio_info, clone);
1806 info->orig = bio_orig;
1808 bio->bi_end_io = end_clone_bio;
1813 static int setup_clone(struct request *clone, struct request *rq,
1814 struct dm_rq_target_io *tio, gfp_t gfp_mask)
1818 r = blk_rq_prep_clone(clone, rq, tio->md->bs, gfp_mask,
1819 dm_rq_bio_constructor, tio);
1823 clone->cmd = rq->cmd;
1824 clone->cmd_len = rq->cmd_len;
1825 clone->sense = rq->sense;
1826 clone->end_io = end_clone_request;
1827 clone->end_io_data = tio;
1834 static struct request *clone_rq(struct request *rq, struct mapped_device *md,
1835 struct dm_rq_target_io *tio, gfp_t gfp_mask)
1838 * Do not allocate a clone if tio->clone was already set
1839 * (see: dm_mq_queue_rq).
1841 bool alloc_clone = !tio->clone;
1842 struct request *clone;
1845 clone = alloc_clone_request(md, gfp_mask);
1851 blk_rq_init(NULL, clone);
1852 if (setup_clone(clone, rq, tio, gfp_mask)) {
1855 free_clone_request(md, clone);
1862 static void map_tio_request(struct kthread_work *work);
1864 static void init_tio(struct dm_rq_target_io *tio, struct request *rq,
1865 struct mapped_device *md)
1872 memset(&tio->info, 0, sizeof(tio->info));
1873 if (md->kworker_task)
1874 init_kthread_work(&tio->work, map_tio_request);
1877 static struct dm_rq_target_io *prep_tio(struct request *rq,
1878 struct mapped_device *md, gfp_t gfp_mask)
1880 struct dm_rq_target_io *tio;
1882 struct dm_table *table;
1884 tio = alloc_rq_tio(md, gfp_mask);
1888 init_tio(tio, rq, md);
1890 table = dm_get_live_table(md, &srcu_idx);
1891 if (!dm_table_mq_request_based(table)) {
1892 if (!clone_rq(rq, md, tio, gfp_mask)) {
1893 dm_put_live_table(md, srcu_idx);
1898 dm_put_live_table(md, srcu_idx);
1904 * Called with the queue lock held.
1906 static int dm_prep_fn(struct request_queue *q, struct request *rq)
1908 struct mapped_device *md = q->queuedata;
1909 struct dm_rq_target_io *tio;
1911 if (unlikely(rq->special)) {
1912 DMWARN("Already has something in rq->special.");
1913 return BLKPREP_KILL;
1916 tio = prep_tio(rq, md, GFP_ATOMIC);
1918 return BLKPREP_DEFER;
1921 rq->cmd_flags |= REQ_DONTPREP;
1928 * 0 : the request has been processed
1929 * DM_MAPIO_REQUEUE : the original request needs to be requeued
1930 * < 0 : the request was completed due to failure
1932 static int map_request(struct dm_rq_target_io *tio, struct request *rq,
1933 struct mapped_device *md)
1936 struct dm_target *ti = tio->ti;
1937 struct request *clone = NULL;
1941 r = ti->type->map_rq(ti, clone, &tio->info);
1943 r = ti->type->clone_and_map_rq(ti, rq, &tio->info, &clone);
1945 /* The target wants to complete the I/O */
1946 dm_kill_unmapped_request(rq, r);
1949 if (r != DM_MAPIO_REMAPPED)
1951 if (setup_clone(clone, rq, tio, GFP_ATOMIC)) {
1953 ti->type->release_clone_rq(clone);
1954 return DM_MAPIO_REQUEUE;
1959 case DM_MAPIO_SUBMITTED:
1960 /* The target has taken the I/O to submit by itself later */
1962 case DM_MAPIO_REMAPPED:
1963 /* The target has remapped the I/O so dispatch it */
1964 trace_block_rq_remap(clone->q, clone, disk_devt(dm_disk(md)),
1966 dm_dispatch_clone_request(clone, rq);
1968 case DM_MAPIO_REQUEUE:
1969 /* The target wants to requeue the I/O */
1970 dm_requeue_original_request(md, tio->orig);
1974 DMWARN("unimplemented target map return value: %d", r);
1978 /* The target wants to complete the I/O */
1979 dm_kill_unmapped_request(rq, r);
1986 static void map_tio_request(struct kthread_work *work)
1988 struct dm_rq_target_io *tio = container_of(work, struct dm_rq_target_io, work);
1989 struct request *rq = tio->orig;
1990 struct mapped_device *md = tio->md;
1992 if (map_request(tio, rq, md) == DM_MAPIO_REQUEUE)
1993 dm_requeue_original_request(md, rq);
1996 static void dm_start_request(struct mapped_device *md, struct request *orig)
1998 if (!orig->q->mq_ops)
1999 blk_start_request(orig);
2001 blk_mq_start_request(orig);
2002 atomic_inc(&md->pending[rq_data_dir(orig)]);
2004 if (md->seq_rq_merge_deadline_usecs) {
2005 md->last_rq_pos = rq_end_sector(orig);
2006 md->last_rq_rw = rq_data_dir(orig);
2007 md->last_rq_start_time = ktime_get();
2010 if (unlikely(dm_stats_used(&md->stats))) {
2011 struct dm_rq_target_io *tio = tio_from_request(orig);
2012 tio->duration_jiffies = jiffies;
2013 tio->n_sectors = blk_rq_sectors(orig);
2014 dm_stats_account_io(&md->stats, orig->cmd_flags, blk_rq_pos(orig),
2015 tio->n_sectors, false, 0, &tio->stats_aux);
2019 * Hold the md reference here for the in-flight I/O.
2020 * We can't rely on the reference count by device opener,
2021 * because the device may be closed during the request completion
2022 * when all bios are completed.
2023 * See the comment in rq_completed() too.
2028 #define MAX_SEQ_RQ_MERGE_DEADLINE_USECS 100000
2030 ssize_t dm_attr_rq_based_seq_io_merge_deadline_show(struct mapped_device *md, char *buf)
2032 return sprintf(buf, "%u\n", md->seq_rq_merge_deadline_usecs);
2035 ssize_t dm_attr_rq_based_seq_io_merge_deadline_store(struct mapped_device *md,
2036 const char *buf, size_t count)
2040 if (!dm_request_based(md) || md->use_blk_mq)
2043 if (kstrtouint(buf, 10, &deadline))
2046 if (deadline > MAX_SEQ_RQ_MERGE_DEADLINE_USECS)
2047 deadline = MAX_SEQ_RQ_MERGE_DEADLINE_USECS;
2049 md->seq_rq_merge_deadline_usecs = deadline;
2054 static bool dm_request_peeked_before_merge_deadline(struct mapped_device *md)
2056 ktime_t kt_deadline;
2058 if (!md->seq_rq_merge_deadline_usecs)
2061 kt_deadline = ns_to_ktime((u64)md->seq_rq_merge_deadline_usecs * NSEC_PER_USEC);
2062 kt_deadline = ktime_add_safe(md->last_rq_start_time, kt_deadline);
2064 return !ktime_after(ktime_get(), kt_deadline);
2068 * q->request_fn for request-based dm.
2069 * Called with the queue lock held.
2071 static void dm_request_fn(struct request_queue *q)
2073 struct mapped_device *md = q->queuedata;
2074 struct dm_target *ti = md->immutable_target;
2076 struct dm_rq_target_io *tio;
2079 if (unlikely(!ti)) {
2081 struct dm_table *map = dm_get_live_table(md, &srcu_idx);
2083 ti = dm_table_find_target(map, pos);
2084 dm_put_live_table(md, srcu_idx);
2088 * For suspend, check blk_queue_stopped() and increment
2089 * ->pending within a single queue_lock not to increment the
2090 * number of in-flight I/Os after the queue is stopped in
2093 while (!blk_queue_stopped(q)) {
2094 rq = blk_peek_request(q);
2098 /* always use block 0 to find the target for flushes for now */
2100 if (!(rq->cmd_flags & REQ_FLUSH))
2101 pos = blk_rq_pos(rq);
2103 if ((dm_request_peeked_before_merge_deadline(md) &&
2104 md_in_flight(md) && rq->bio && rq->bio->bi_vcnt == 1 &&
2105 md->last_rq_pos == pos && md->last_rq_rw == rq_data_dir(rq)) ||
2106 (ti->type->busy && ti->type->busy(ti))) {
2107 blk_delay_queue(q, HZ / 100);
2111 dm_start_request(md, rq);
2113 tio = tio_from_request(rq);
2114 /* Establish tio->ti before queuing work (map_tio_request) */
2116 queue_kthread_work(&md->kworker, &tio->work);
2117 BUG_ON(!irqs_disabled());
2121 static int dm_any_congested(void *congested_data, int bdi_bits)
2124 struct mapped_device *md = congested_data;
2125 struct dm_table *map;
2127 if (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
2128 if (dm_request_based(md)) {
2130 * With request-based DM we only need to check the
2131 * top-level queue for congestion.
2133 r = md->queue->backing_dev_info.wb.state & bdi_bits;
2135 map = dm_get_live_table_fast(md);
2137 r = dm_table_any_congested(map, bdi_bits);
2138 dm_put_live_table_fast(md);
2145 /*-----------------------------------------------------------------
2146 * An IDR is used to keep track of allocated minor numbers.
2147 *---------------------------------------------------------------*/
2148 static void free_minor(int minor)
2150 spin_lock(&_minor_lock);
2151 idr_remove(&_minor_idr, minor);
2152 spin_unlock(&_minor_lock);
2156 * See if the device with a specific minor # is free.
2158 static int specific_minor(int minor)
2162 if (minor >= (1 << MINORBITS))
2165 idr_preload(GFP_KERNEL);
2166 spin_lock(&_minor_lock);
2168 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, minor, minor + 1, GFP_NOWAIT);
2170 spin_unlock(&_minor_lock);
2173 return r == -ENOSPC ? -EBUSY : r;
2177 static int next_free_minor(int *minor)
2181 idr_preload(GFP_KERNEL);
2182 spin_lock(&_minor_lock);
2184 r = idr_alloc(&_minor_idr, MINOR_ALLOCED, 0, 1 << MINORBITS, GFP_NOWAIT);
2186 spin_unlock(&_minor_lock);
2194 static const struct block_device_operations dm_blk_dops;
2196 static void dm_wq_work(struct work_struct *work);
2198 static void dm_init_md_queue(struct mapped_device *md)
2201 * Request-based dm devices cannot be stacked on top of bio-based dm
2202 * devices. The type of this dm device may not have been decided yet.
2203 * The type is decided at the first table loading time.
2204 * To prevent problematic device stacking, clear the queue flag
2205 * for request stacking support until then.
2207 * This queue is new, so no concurrency on the queue_flags.
2209 queue_flag_clear_unlocked(QUEUE_FLAG_STACKABLE, md->queue);
2212 * Initialize data that will only be used by a non-blk-mq DM queue
2213 * - must do so here (in alloc_dev callchain) before queue is used
2215 md->queue->queuedata = md;
2216 md->queue->backing_dev_info.congested_data = md;
2219 static void dm_init_old_md_queue(struct mapped_device *md)
2221 md->use_blk_mq = false;
2222 dm_init_md_queue(md);
2225 * Initialize aspects of queue that aren't relevant for blk-mq
2227 md->queue->backing_dev_info.congested_fn = dm_any_congested;
2228 blk_queue_bounce_limit(md->queue, BLK_BOUNCE_ANY);
2231 static void cleanup_mapped_device(struct mapped_device *md)
2234 destroy_workqueue(md->wq);
2235 if (md->kworker_task)
2236 kthread_stop(md->kworker_task);
2237 mempool_destroy(md->io_pool);
2238 mempool_destroy(md->rq_pool);
2240 bioset_free(md->bs);
2242 cleanup_srcu_struct(&md->io_barrier);
2245 spin_lock(&_minor_lock);
2246 md->disk->private_data = NULL;
2247 spin_unlock(&_minor_lock);
2248 del_gendisk(md->disk);
2253 blk_cleanup_queue(md->queue);
2262 * Allocate and initialise a blank device with a given minor.
2264 static struct mapped_device *alloc_dev(int minor)
2267 struct mapped_device *md = kzalloc(sizeof(*md), GFP_KERNEL);
2271 DMWARN("unable to allocate device, out of memory.");
2275 if (!try_module_get(THIS_MODULE))
2276 goto bad_module_get;
2278 /* get a minor number for the dev */
2279 if (minor == DM_ANY_MINOR)
2280 r = next_free_minor(&minor);
2282 r = specific_minor(minor);
2286 r = init_srcu_struct(&md->io_barrier);
2288 goto bad_io_barrier;
2290 md->use_blk_mq = use_blk_mq;
2291 md->type = DM_TYPE_NONE;
2292 mutex_init(&md->suspend_lock);
2293 mutex_init(&md->type_lock);
2294 mutex_init(&md->table_devices_lock);
2295 spin_lock_init(&md->deferred_lock);
2296 atomic_set(&md->holders, 1);
2297 atomic_set(&md->open_count, 0);
2298 atomic_set(&md->event_nr, 0);
2299 atomic_set(&md->uevent_seq, 0);
2300 INIT_LIST_HEAD(&md->uevent_list);
2301 INIT_LIST_HEAD(&md->table_devices);
2302 spin_lock_init(&md->uevent_lock);
2304 md->queue = blk_alloc_queue(GFP_KERNEL);
2308 dm_init_md_queue(md);
2310 md->disk = alloc_disk(1);
2314 atomic_set(&md->pending[0], 0);
2315 atomic_set(&md->pending[1], 0);
2316 init_waitqueue_head(&md->wait);
2317 INIT_WORK(&md->work, dm_wq_work);
2318 init_waitqueue_head(&md->eventq);
2319 init_completion(&md->kobj_holder.completion);
2320 md->kworker_task = NULL;
2322 md->disk->major = _major;
2323 md->disk->first_minor = minor;
2324 md->disk->fops = &dm_blk_dops;
2325 md->disk->queue = md->queue;
2326 md->disk->private_data = md;
2327 sprintf(md->disk->disk_name, "dm-%d", minor);
2329 format_dev_t(md->name, MKDEV(_major, minor));
2331 md->wq = alloc_workqueue("kdmflush", WQ_MEM_RECLAIM, 0);
2335 md->bdev = bdget_disk(md->disk, 0);
2339 bio_init(&md->flush_bio);
2340 md->flush_bio.bi_bdev = md->bdev;
2341 md->flush_bio.bi_rw = WRITE_FLUSH;
2343 dm_stats_init(&md->stats);
2345 /* Populate the mapping, nobody knows we exist yet */
2346 spin_lock(&_minor_lock);
2347 old_md = idr_replace(&_minor_idr, md, minor);
2348 spin_unlock(&_minor_lock);
2350 BUG_ON(old_md != MINOR_ALLOCED);
2355 cleanup_mapped_device(md);
2359 module_put(THIS_MODULE);
2365 static void unlock_fs(struct mapped_device *md);
2367 static void free_dev(struct mapped_device *md)
2369 int minor = MINOR(disk_devt(md->disk));
2373 cleanup_mapped_device(md);
2375 blk_mq_free_tag_set(&md->tag_set);
2377 free_table_devices(&md->table_devices);
2378 dm_stats_cleanup(&md->stats);
2381 module_put(THIS_MODULE);
2385 static void __bind_mempools(struct mapped_device *md, struct dm_table *t)
2387 struct dm_md_mempools *p = dm_table_get_md_mempools(t);
2390 /* The md already has necessary mempools. */
2391 if (dm_table_get_type(t) == DM_TYPE_BIO_BASED) {
2393 * Reload bioset because front_pad may have changed
2394 * because a different table was loaded.
2396 bioset_free(md->bs);
2401 * There's no need to reload with request-based dm
2402 * because the size of front_pad doesn't change.
2403 * Note for future: If you are to reload bioset,
2404 * prep-ed requests in the queue may refer
2405 * to bio from the old bioset, so you must walk
2406 * through the queue to unprep.
2411 BUG_ON(!p || md->io_pool || md->rq_pool || md->bs);
2413 md->io_pool = p->io_pool;
2415 md->rq_pool = p->rq_pool;
2421 /* mempool bind completed, no longer need any mempools in the table */
2422 dm_table_free_md_mempools(t);
2426 * Bind a table to the device.
2428 static void event_callback(void *context)
2430 unsigned long flags;
2432 struct mapped_device *md = (struct mapped_device *) context;
2434 spin_lock_irqsave(&md->uevent_lock, flags);
2435 list_splice_init(&md->uevent_list, &uevents);
2436 spin_unlock_irqrestore(&md->uevent_lock, flags);
2438 dm_send_uevents(&uevents, &disk_to_dev(md->disk)->kobj);
2440 atomic_inc(&md->event_nr);
2441 wake_up(&md->eventq);
2445 * Protected by md->suspend_lock obtained by dm_swap_table().
2447 static void __set_size(struct mapped_device *md, sector_t size)
2449 set_capacity(md->disk, size);
2451 i_size_write(md->bdev->bd_inode, (loff_t)size << SECTOR_SHIFT);
2455 * Returns old map, which caller must destroy.
2457 static struct dm_table *__bind(struct mapped_device *md, struct dm_table *t,
2458 struct queue_limits *limits)
2460 struct dm_table *old_map;
2461 struct request_queue *q = md->queue;
2464 size = dm_table_get_size(t);
2467 * Wipe any geometry if the size of the table changed.
2469 if (size != dm_get_size(md))
2470 memset(&md->geometry, 0, sizeof(md->geometry));
2472 __set_size(md, size);
2474 dm_table_event_callback(t, event_callback, md);
2477 * The queue hasn't been stopped yet, if the old table type wasn't
2478 * for request-based during suspension. So stop it to prevent
2479 * I/O mapping before resume.
2480 * This must be done before setting the queue restrictions,
2481 * because request-based dm may be run just after the setting.
2483 if (dm_table_request_based(t)) {
2486 * Leverage the fact that request-based DM targets are
2487 * immutable singletons and establish md->immutable_target
2488 * - used to optimize both dm_request_fn and dm_mq_queue_rq
2490 md->immutable_target = dm_table_get_immutable_target(t);
2493 __bind_mempools(md, t);
2495 old_map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
2496 rcu_assign_pointer(md->map, t);
2497 md->immutable_target_type = dm_table_get_immutable_target_type(t);
2499 dm_table_set_restrictions(t, q, limits);
2507 * Returns unbound table for the caller to free.
2509 static struct dm_table *__unbind(struct mapped_device *md)
2511 struct dm_table *map = rcu_dereference_protected(md->map, 1);
2516 dm_table_event_callback(map, NULL, NULL);
2517 RCU_INIT_POINTER(md->map, NULL);
2524 * Constructor for a new device.
2526 int dm_create(int minor, struct mapped_device **result)
2528 struct mapped_device *md;
2530 md = alloc_dev(minor);
2541 * Functions to manage md->type.
2542 * All are required to hold md->type_lock.
2544 void dm_lock_md_type(struct mapped_device *md)
2546 mutex_lock(&md->type_lock);
2549 void dm_unlock_md_type(struct mapped_device *md)
2551 mutex_unlock(&md->type_lock);
2554 void dm_set_md_type(struct mapped_device *md, unsigned type)
2556 BUG_ON(!mutex_is_locked(&md->type_lock));
2560 unsigned dm_get_md_type(struct mapped_device *md)
2565 struct target_type *dm_get_immutable_target_type(struct mapped_device *md)
2567 return md->immutable_target_type;
2571 * The queue_limits are only valid as long as you have a reference
2574 struct queue_limits *dm_get_queue_limits(struct mapped_device *md)
2576 BUG_ON(!atomic_read(&md->holders));
2577 return &md->queue->limits;
2579 EXPORT_SYMBOL_GPL(dm_get_queue_limits);
2581 static void init_rq_based_worker_thread(struct mapped_device *md)
2583 /* Initialize the request-based DM worker thread */
2584 init_kthread_worker(&md->kworker);
2585 md->kworker_task = kthread_run(kthread_worker_fn, &md->kworker,
2586 "kdmwork-%s", dm_device_name(md));
2590 * Fully initialize a request-based queue (->elevator, ->request_fn, etc).
2592 static int dm_init_request_based_queue(struct mapped_device *md)
2594 struct request_queue *q = NULL;
2596 /* Fully initialize the queue */
2597 q = blk_init_allocated_queue(md->queue, dm_request_fn, NULL);
2601 /* disable dm_request_fn's merge heuristic by default */
2602 md->seq_rq_merge_deadline_usecs = 0;
2605 dm_init_old_md_queue(md);
2606 blk_queue_softirq_done(md->queue, dm_softirq_done);
2607 blk_queue_prep_rq(md->queue, dm_prep_fn);
2609 init_rq_based_worker_thread(md);
2611 elv_register_queue(md->queue);
2616 static int dm_mq_init_request(void *data, struct request *rq,
2617 unsigned int hctx_idx, unsigned int request_idx,
2618 unsigned int numa_node)
2620 struct mapped_device *md = data;
2621 struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
2624 * Must initialize md member of tio, otherwise it won't
2625 * be available in dm_mq_queue_rq.
2632 static int dm_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
2633 const struct blk_mq_queue_data *bd)
2635 struct request *rq = bd->rq;
2636 struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
2637 struct mapped_device *md = tio->md;
2638 struct dm_target *ti = md->immutable_target;
2640 if (unlikely(!ti)) {
2642 struct dm_table *map = dm_get_live_table(md, &srcu_idx);
2644 ti = dm_table_find_target(map, 0);
2645 dm_put_live_table(md, srcu_idx);
2648 if (ti->type->busy && ti->type->busy(ti))
2649 return BLK_MQ_RQ_QUEUE_BUSY;
2651 dm_start_request(md, rq);
2653 /* Init tio using md established in .init_request */
2654 init_tio(tio, rq, md);
2657 * Establish tio->ti before queuing work (map_tio_request)
2658 * or making direct call to map_request().
2663 * Both the table and md type cannot change after initial table load
2665 if (dm_get_md_type(md) == DM_TYPE_REQUEST_BASED) {
2666 /* clone request is allocated at the end of the pdu */
2667 tio->clone = (void *)blk_mq_rq_to_pdu(rq) + sizeof(struct dm_rq_target_io);
2668 (void) clone_rq(rq, md, tio, GFP_ATOMIC);
2669 queue_kthread_work(&md->kworker, &tio->work);
2671 /* Direct call is fine since .queue_rq allows allocations */
2672 if (map_request(tio, rq, md) == DM_MAPIO_REQUEUE) {
2673 /* Undo dm_start_request() before requeuing */
2674 rq_end_stats(md, rq);
2675 rq_completed(md, rq_data_dir(rq), false);
2676 return BLK_MQ_RQ_QUEUE_BUSY;
2680 return BLK_MQ_RQ_QUEUE_OK;
2683 static struct blk_mq_ops dm_mq_ops = {
2684 .queue_rq = dm_mq_queue_rq,
2685 .map_queue = blk_mq_map_queue,
2686 .complete = dm_softirq_done,
2687 .init_request = dm_mq_init_request,
2690 static int dm_init_request_based_blk_mq_queue(struct mapped_device *md)
2692 unsigned md_type = dm_get_md_type(md);
2693 struct request_queue *q;
2696 memset(&md->tag_set, 0, sizeof(md->tag_set));
2697 md->tag_set.ops = &dm_mq_ops;
2698 md->tag_set.queue_depth = BLKDEV_MAX_RQ;
2699 md->tag_set.numa_node = NUMA_NO_NODE;
2700 md->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
2701 md->tag_set.nr_hw_queues = 1;
2702 if (md_type == DM_TYPE_REQUEST_BASED) {
2703 /* make the memory for non-blk-mq clone part of the pdu */
2704 md->tag_set.cmd_size = sizeof(struct dm_rq_target_io) + sizeof(struct request);
2706 md->tag_set.cmd_size = sizeof(struct dm_rq_target_io);
2707 md->tag_set.driver_data = md;
2709 err = blk_mq_alloc_tag_set(&md->tag_set);
2713 q = blk_mq_init_allocated_queue(&md->tag_set, md->queue);
2719 dm_init_md_queue(md);
2721 /* backfill 'mq' sysfs registration normally done in blk_register_queue */
2722 blk_mq_register_disk(md->disk);
2724 if (md_type == DM_TYPE_REQUEST_BASED)
2725 init_rq_based_worker_thread(md);
2730 blk_mq_free_tag_set(&md->tag_set);
2734 static unsigned filter_md_type(unsigned type, struct mapped_device *md)
2736 if (type == DM_TYPE_BIO_BASED)
2739 return !md->use_blk_mq ? DM_TYPE_REQUEST_BASED : DM_TYPE_MQ_REQUEST_BASED;
2743 * Setup the DM device's queue based on md's type
2745 int dm_setup_md_queue(struct mapped_device *md)
2748 unsigned md_type = filter_md_type(dm_get_md_type(md), md);
2751 case DM_TYPE_REQUEST_BASED:
2752 r = dm_init_request_based_queue(md);
2754 DMWARN("Cannot initialize queue for request-based mapped device");
2758 case DM_TYPE_MQ_REQUEST_BASED:
2759 r = dm_init_request_based_blk_mq_queue(md);
2761 DMWARN("Cannot initialize queue for request-based blk-mq mapped device");
2765 case DM_TYPE_BIO_BASED:
2766 dm_init_old_md_queue(md);
2767 blk_queue_make_request(md->queue, dm_make_request);
2769 * DM handles splitting bios as needed. Free the bio_split bioset
2770 * since it won't be used (saves 1 process per bio-based DM device).
2772 bioset_free(md->queue->bio_split);
2773 md->queue->bio_split = NULL;
2780 struct mapped_device *dm_get_md(dev_t dev)
2782 struct mapped_device *md;
2783 unsigned minor = MINOR(dev);
2785 if (MAJOR(dev) != _major || minor >= (1 << MINORBITS))
2788 spin_lock(&_minor_lock);
2790 md = idr_find(&_minor_idr, minor);
2792 if ((md == MINOR_ALLOCED ||
2793 (MINOR(disk_devt(dm_disk(md))) != minor) ||
2794 dm_deleting_md(md) ||
2795 test_bit(DMF_FREEING, &md->flags))) {
2803 spin_unlock(&_minor_lock);
2807 EXPORT_SYMBOL_GPL(dm_get_md);
2809 void *dm_get_mdptr(struct mapped_device *md)
2811 return md->interface_ptr;
2814 void dm_set_mdptr(struct mapped_device *md, void *ptr)
2816 md->interface_ptr = ptr;
2819 void dm_get(struct mapped_device *md)
2821 atomic_inc(&md->holders);
2822 BUG_ON(test_bit(DMF_FREEING, &md->flags));
2825 int dm_hold(struct mapped_device *md)
2827 spin_lock(&_minor_lock);
2828 if (test_bit(DMF_FREEING, &md->flags)) {
2829 spin_unlock(&_minor_lock);
2833 spin_unlock(&_minor_lock);
2836 EXPORT_SYMBOL_GPL(dm_hold);
2838 const char *dm_device_name(struct mapped_device *md)
2842 EXPORT_SYMBOL_GPL(dm_device_name);
2844 static void __dm_destroy(struct mapped_device *md, bool wait)
2846 struct dm_table *map;
2851 spin_lock(&_minor_lock);
2852 idr_replace(&_minor_idr, MINOR_ALLOCED, MINOR(disk_devt(dm_disk(md))));
2853 set_bit(DMF_FREEING, &md->flags);
2854 spin_unlock(&_minor_lock);
2856 if (dm_request_based(md) && md->kworker_task)
2857 flush_kthread_worker(&md->kworker);
2860 * Take suspend_lock so that presuspend and postsuspend methods
2861 * do not race with internal suspend.
2863 mutex_lock(&md->suspend_lock);
2864 map = dm_get_live_table(md, &srcu_idx);
2865 if (!dm_suspended_md(md)) {
2866 dm_table_presuspend_targets(map);
2867 dm_table_postsuspend_targets(map);
2869 /* dm_put_live_table must be before msleep, otherwise deadlock is possible */
2870 dm_put_live_table(md, srcu_idx);
2871 mutex_unlock(&md->suspend_lock);
2874 * Rare, but there may be I/O requests still going to complete,
2875 * for example. Wait for all references to disappear.
2876 * No one should increment the reference count of the mapped_device,
2877 * after the mapped_device state becomes DMF_FREEING.
2880 while (atomic_read(&md->holders))
2882 else if (atomic_read(&md->holders))
2883 DMWARN("%s: Forcibly removing mapped_device still in use! (%d users)",
2884 dm_device_name(md), atomic_read(&md->holders));
2887 dm_table_destroy(__unbind(md));
2891 void dm_destroy(struct mapped_device *md)
2893 __dm_destroy(md, true);
2896 void dm_destroy_immediate(struct mapped_device *md)
2898 __dm_destroy(md, false);
2901 void dm_put(struct mapped_device *md)
2903 atomic_dec(&md->holders);
2905 EXPORT_SYMBOL_GPL(dm_put);
2907 static int dm_wait_for_completion(struct mapped_device *md, int interruptible)
2910 DECLARE_WAITQUEUE(wait, current);
2912 add_wait_queue(&md->wait, &wait);
2915 set_current_state(interruptible);
2917 if (!md_in_flight(md))
2920 if (interruptible == TASK_INTERRUPTIBLE &&
2921 signal_pending(current)) {
2928 set_current_state(TASK_RUNNING);
2930 remove_wait_queue(&md->wait, &wait);
2936 * Process the deferred bios
2938 static void dm_wq_work(struct work_struct *work)
2940 struct mapped_device *md = container_of(work, struct mapped_device,
2944 struct dm_table *map;
2946 map = dm_get_live_table(md, &srcu_idx);
2948 while (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
2949 spin_lock_irq(&md->deferred_lock);
2950 c = bio_list_pop(&md->deferred);
2951 spin_unlock_irq(&md->deferred_lock);
2956 if (dm_request_based(md))
2957 generic_make_request(c);
2959 __split_and_process_bio(md, map, c);
2962 dm_put_live_table(md, srcu_idx);
2965 static void dm_queue_flush(struct mapped_device *md)
2967 clear_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
2968 smp_mb__after_atomic();
2969 queue_work(md->wq, &md->work);
2973 * Swap in a new table, returning the old one for the caller to destroy.
2975 struct dm_table *dm_swap_table(struct mapped_device *md, struct dm_table *table)
2977 struct dm_table *live_map = NULL, *map = ERR_PTR(-EINVAL);
2978 struct queue_limits limits;
2981 mutex_lock(&md->suspend_lock);
2983 /* device must be suspended */
2984 if (!dm_suspended_md(md))
2988 * If the new table has no data devices, retain the existing limits.
2989 * This helps multipath with queue_if_no_path if all paths disappear,
2990 * then new I/O is queued based on these limits, and then some paths
2993 if (dm_table_has_no_data_devices(table)) {
2994 live_map = dm_get_live_table_fast(md);
2996 limits = md->queue->limits;
2997 dm_put_live_table_fast(md);
3001 r = dm_calculate_queue_limits(table, &limits);
3008 map = __bind(md, table, &limits);
3011 mutex_unlock(&md->suspend_lock);
3016 * Functions to lock and unlock any filesystem running on the
3019 static int lock_fs(struct mapped_device *md)
3023 WARN_ON(md->frozen_sb);
3025 md->frozen_sb = freeze_bdev(md->bdev);
3026 if (IS_ERR(md->frozen_sb)) {
3027 r = PTR_ERR(md->frozen_sb);
3028 md->frozen_sb = NULL;
3032 set_bit(DMF_FROZEN, &md->flags);
3037 static void unlock_fs(struct mapped_device *md)
3039 if (!test_bit(DMF_FROZEN, &md->flags))
3042 thaw_bdev(md->bdev, md->frozen_sb);
3043 md->frozen_sb = NULL;
3044 clear_bit(DMF_FROZEN, &md->flags);
3048 * If __dm_suspend returns 0, the device is completely quiescent
3049 * now. There is no request-processing activity. All new requests
3050 * are being added to md->deferred list.
3052 * Caller must hold md->suspend_lock
3054 static int __dm_suspend(struct mapped_device *md, struct dm_table *map,
3055 unsigned suspend_flags, int interruptible)
3057 bool do_lockfs = suspend_flags & DM_SUSPEND_LOCKFS_FLAG;
3058 bool noflush = suspend_flags & DM_SUSPEND_NOFLUSH_FLAG;
3062 * DMF_NOFLUSH_SUSPENDING must be set before presuspend.
3063 * This flag is cleared before dm_suspend returns.
3066 set_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
3069 * This gets reverted if there's an error later and the targets
3070 * provide the .presuspend_undo hook.
3072 dm_table_presuspend_targets(map);
3075 * Flush I/O to the device.
3076 * Any I/O submitted after lock_fs() may not be flushed.
3077 * noflush takes precedence over do_lockfs.
3078 * (lock_fs() flushes I/Os and waits for them to complete.)
3080 if (!noflush && do_lockfs) {
3083 dm_table_presuspend_undo_targets(map);
3089 * Here we must make sure that no processes are submitting requests
3090 * to target drivers i.e. no one may be executing
3091 * __split_and_process_bio. This is called from dm_request and
3094 * To get all processes out of __split_and_process_bio in dm_request,
3095 * we take the write lock. To prevent any process from reentering
3096 * __split_and_process_bio from dm_request and quiesce the thread
3097 * (dm_wq_work), we set BMF_BLOCK_IO_FOR_SUSPEND and call
3098 * flush_workqueue(md->wq).
3100 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
3102 synchronize_srcu(&md->io_barrier);
3105 * Stop md->queue before flushing md->wq in case request-based
3106 * dm defers requests to md->wq from md->queue.
3108 if (dm_request_based(md)) {
3109 stop_queue(md->queue);
3110 if (md->kworker_task)
3111 flush_kthread_worker(&md->kworker);
3114 flush_workqueue(md->wq);
3117 * At this point no more requests are entering target request routines.
3118 * We call dm_wait_for_completion to wait for all existing requests
3121 r = dm_wait_for_completion(md, interruptible);
3124 clear_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
3126 synchronize_srcu(&md->io_barrier);
3128 /* were we interrupted ? */
3132 if (dm_request_based(md))
3133 start_queue(md->queue);
3136 dm_table_presuspend_undo_targets(map);
3137 /* pushback list is already flushed, so skip flush */
3144 * We need to be able to change a mapping table under a mounted
3145 * filesystem. For example we might want to move some data in
3146 * the background. Before the table can be swapped with
3147 * dm_bind_table, dm_suspend must be called to flush any in
3148 * flight bios and ensure that any further io gets deferred.
3151 * Suspend mechanism in request-based dm.
3153 * 1. Flush all I/Os by lock_fs() if needed.
3154 * 2. Stop dispatching any I/O by stopping the request_queue.
3155 * 3. Wait for all in-flight I/Os to be completed or requeued.
3157 * To abort suspend, start the request_queue.
3159 int dm_suspend(struct mapped_device *md, unsigned suspend_flags)
3161 struct dm_table *map = NULL;
3165 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
3167 if (dm_suspended_md(md)) {
3172 if (dm_suspended_internally_md(md)) {
3173 /* already internally suspended, wait for internal resume */
3174 mutex_unlock(&md->suspend_lock);
3175 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
3181 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3183 r = __dm_suspend(md, map, suspend_flags, TASK_INTERRUPTIBLE);
3187 set_bit(DMF_SUSPENDED, &md->flags);
3189 dm_table_postsuspend_targets(map);
3192 mutex_unlock(&md->suspend_lock);
3196 static int __dm_resume(struct mapped_device *md, struct dm_table *map)
3199 int r = dm_table_resume_targets(map);
3207 * Flushing deferred I/Os must be done after targets are resumed
3208 * so that mapping of targets can work correctly.
3209 * Request-based dm is queueing the deferred I/Os in its request_queue.
3211 if (dm_request_based(md))
3212 start_queue(md->queue);
3219 int dm_resume(struct mapped_device *md)
3222 struct dm_table *map = NULL;
3225 mutex_lock_nested(&md->suspend_lock, SINGLE_DEPTH_NESTING);
3227 if (!dm_suspended_md(md))
3230 if (dm_suspended_internally_md(md)) {
3231 /* already internally suspended, wait for internal resume */
3232 mutex_unlock(&md->suspend_lock);
3233 r = wait_on_bit(&md->flags, DMF_SUSPENDED_INTERNALLY, TASK_INTERRUPTIBLE);
3239 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3240 if (!map || !dm_table_get_size(map))
3243 r = __dm_resume(md, map);
3247 clear_bit(DMF_SUSPENDED, &md->flags);
3251 mutex_unlock(&md->suspend_lock);
3257 * Internal suspend/resume works like userspace-driven suspend. It waits
3258 * until all bios finish and prevents issuing new bios to the target drivers.
3259 * It may be used only from the kernel.
3262 static void __dm_internal_suspend(struct mapped_device *md, unsigned suspend_flags)
3264 struct dm_table *map = NULL;
3266 if (md->internal_suspend_count++)
3267 return; /* nested internal suspend */
3269 if (dm_suspended_md(md)) {
3270 set_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3271 return; /* nest suspend */
3274 map = rcu_dereference_protected(md->map, lockdep_is_held(&md->suspend_lock));
3277 * Using TASK_UNINTERRUPTIBLE because only NOFLUSH internal suspend is
3278 * supported. Properly supporting a TASK_INTERRUPTIBLE internal suspend
3279 * would require changing .presuspend to return an error -- avoid this
3280 * until there is a need for more elaborate variants of internal suspend.
3282 (void) __dm_suspend(md, map, suspend_flags, TASK_UNINTERRUPTIBLE);
3284 set_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3286 dm_table_postsuspend_targets(map);
3289 static void __dm_internal_resume(struct mapped_device *md)
3291 BUG_ON(!md->internal_suspend_count);
3293 if (--md->internal_suspend_count)
3294 return; /* resume from nested internal suspend */
3296 if (dm_suspended_md(md))
3297 goto done; /* resume from nested suspend */
3300 * NOTE: existing callers don't need to call dm_table_resume_targets
3301 * (which may fail -- so best to avoid it for now by passing NULL map)
3303 (void) __dm_resume(md, NULL);
3306 clear_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3307 smp_mb__after_atomic();
3308 wake_up_bit(&md->flags, DMF_SUSPENDED_INTERNALLY);
3311 void dm_internal_suspend_noflush(struct mapped_device *md)
3313 mutex_lock(&md->suspend_lock);
3314 __dm_internal_suspend(md, DM_SUSPEND_NOFLUSH_FLAG);
3315 mutex_unlock(&md->suspend_lock);
3317 EXPORT_SYMBOL_GPL(dm_internal_suspend_noflush);
3319 void dm_internal_resume(struct mapped_device *md)
3321 mutex_lock(&md->suspend_lock);
3322 __dm_internal_resume(md);
3323 mutex_unlock(&md->suspend_lock);
3325 EXPORT_SYMBOL_GPL(dm_internal_resume);
3328 * Fast variants of internal suspend/resume hold md->suspend_lock,
3329 * which prevents interaction with userspace-driven suspend.
3332 void dm_internal_suspend_fast(struct mapped_device *md)
3334 mutex_lock(&md->suspend_lock);
3335 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
3338 set_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags);
3339 synchronize_srcu(&md->io_barrier);
3340 flush_workqueue(md->wq);
3341 dm_wait_for_completion(md, TASK_UNINTERRUPTIBLE);
3343 EXPORT_SYMBOL_GPL(dm_internal_suspend_fast);
3345 void dm_internal_resume_fast(struct mapped_device *md)
3347 if (dm_suspended_md(md) || dm_suspended_internally_md(md))
3353 mutex_unlock(&md->suspend_lock);
3355 EXPORT_SYMBOL_GPL(dm_internal_resume_fast);
3357 /*-----------------------------------------------------------------
3358 * Event notification.
3359 *---------------------------------------------------------------*/
3360 int dm_kobject_uevent(struct mapped_device *md, enum kobject_action action,
3363 char udev_cookie[DM_COOKIE_LENGTH];
3364 char *envp[] = { udev_cookie, NULL };
3367 return kobject_uevent(&disk_to_dev(md->disk)->kobj, action);
3369 snprintf(udev_cookie, DM_COOKIE_LENGTH, "%s=%u",
3370 DM_COOKIE_ENV_VAR_NAME, cookie);
3371 return kobject_uevent_env(&disk_to_dev(md->disk)->kobj,
3376 uint32_t dm_next_uevent_seq(struct mapped_device *md)
3378 return atomic_add_return(1, &md->uevent_seq);
3381 uint32_t dm_get_event_nr(struct mapped_device *md)
3383 return atomic_read(&md->event_nr);
3386 int dm_wait_event(struct mapped_device *md, int event_nr)
3388 return wait_event_interruptible(md->eventq,
3389 (event_nr != atomic_read(&md->event_nr)));
3392 void dm_uevent_add(struct mapped_device *md, struct list_head *elist)
3394 unsigned long flags;
3396 spin_lock_irqsave(&md->uevent_lock, flags);
3397 list_add(elist, &md->uevent_list);
3398 spin_unlock_irqrestore(&md->uevent_lock, flags);
3402 * The gendisk is only valid as long as you have a reference
3405 struct gendisk *dm_disk(struct mapped_device *md)
3409 EXPORT_SYMBOL_GPL(dm_disk);
3411 struct kobject *dm_kobject(struct mapped_device *md)
3413 return &md->kobj_holder.kobj;
3416 struct mapped_device *dm_get_from_kobject(struct kobject *kobj)
3418 struct mapped_device *md;
3420 md = container_of(kobj, struct mapped_device, kobj_holder.kobj);
3422 if (test_bit(DMF_FREEING, &md->flags) ||
3430 int dm_suspended_md(struct mapped_device *md)
3432 return test_bit(DMF_SUSPENDED, &md->flags);
3435 int dm_suspended_internally_md(struct mapped_device *md)
3437 return test_bit(DMF_SUSPENDED_INTERNALLY, &md->flags);
3440 int dm_test_deferred_remove_flag(struct mapped_device *md)
3442 return test_bit(DMF_DEFERRED_REMOVE, &md->flags);
3445 int dm_suspended(struct dm_target *ti)
3447 return dm_suspended_md(dm_table_get_md(ti->table));
3449 EXPORT_SYMBOL_GPL(dm_suspended);
3451 int dm_noflush_suspending(struct dm_target *ti)
3453 return __noflush_suspending(dm_table_get_md(ti->table));
3455 EXPORT_SYMBOL_GPL(dm_noflush_suspending);
3457 struct dm_md_mempools *dm_alloc_md_mempools(struct mapped_device *md, unsigned type,
3458 unsigned integrity, unsigned per_bio_data_size)
3460 struct dm_md_mempools *pools = kzalloc(sizeof(*pools), GFP_KERNEL);
3461 struct kmem_cache *cachep = NULL;
3462 unsigned int pool_size = 0;
3463 unsigned int front_pad;
3468 type = filter_md_type(type, md);
3471 case DM_TYPE_BIO_BASED:
3473 pool_size = dm_get_reserved_bio_based_ios();
3474 front_pad = roundup(per_bio_data_size, __alignof__(struct dm_target_io)) + offsetof(struct dm_target_io, clone);
3476 case DM_TYPE_REQUEST_BASED:
3477 cachep = _rq_tio_cache;
3478 pool_size = dm_get_reserved_rq_based_ios();
3479 pools->rq_pool = mempool_create_slab_pool(pool_size, _rq_cache);
3480 if (!pools->rq_pool)
3482 /* fall through to setup remaining rq-based pools */
3483 case DM_TYPE_MQ_REQUEST_BASED:
3485 pool_size = dm_get_reserved_rq_based_ios();
3486 front_pad = offsetof(struct dm_rq_clone_bio_info, clone);
3487 /* per_bio_data_size is not used. See __bind_mempools(). */
3488 WARN_ON(per_bio_data_size != 0);
3495 pools->io_pool = mempool_create_slab_pool(pool_size, cachep);
3496 if (!pools->io_pool)
3500 pools->bs = bioset_create_nobvec(pool_size, front_pad);
3504 if (integrity && bioset_integrity_create(pools->bs, pool_size))
3510 dm_free_md_mempools(pools);
3515 void dm_free_md_mempools(struct dm_md_mempools *pools)
3520 mempool_destroy(pools->io_pool);
3521 mempool_destroy(pools->rq_pool);
3524 bioset_free(pools->bs);
3529 static int dm_pr_register(struct block_device *bdev, u64 old_key, u64 new_key,
3532 struct mapped_device *md = bdev->bd_disk->private_data;
3533 const struct pr_ops *ops;
3537 r = dm_grab_bdev_for_ioctl(md, &bdev, &mode);
3541 ops = bdev->bd_disk->fops->pr_ops;
3542 if (ops && ops->pr_register)
3543 r = ops->pr_register(bdev, old_key, new_key, flags);
3551 static int dm_pr_reserve(struct block_device *bdev, u64 key, enum pr_type type,
3554 struct mapped_device *md = bdev->bd_disk->private_data;
3555 const struct pr_ops *ops;
3559 r = dm_grab_bdev_for_ioctl(md, &bdev, &mode);
3563 ops = bdev->bd_disk->fops->pr_ops;
3564 if (ops && ops->pr_reserve)
3565 r = ops->pr_reserve(bdev, key, type, flags);
3573 static int dm_pr_release(struct block_device *bdev, u64 key, enum pr_type type)
3575 struct mapped_device *md = bdev->bd_disk->private_data;
3576 const struct pr_ops *ops;
3580 r = dm_grab_bdev_for_ioctl(md, &bdev, &mode);
3584 ops = bdev->bd_disk->fops->pr_ops;
3585 if (ops && ops->pr_release)
3586 r = ops->pr_release(bdev, key, type);
3594 static int dm_pr_preempt(struct block_device *bdev, u64 old_key, u64 new_key,
3595 enum pr_type type, bool abort)
3597 struct mapped_device *md = bdev->bd_disk->private_data;
3598 const struct pr_ops *ops;
3602 r = dm_grab_bdev_for_ioctl(md, &bdev, &mode);
3606 ops = bdev->bd_disk->fops->pr_ops;
3607 if (ops && ops->pr_preempt)
3608 r = ops->pr_preempt(bdev, old_key, new_key, type, abort);
3616 static int dm_pr_clear(struct block_device *bdev, u64 key)
3618 struct mapped_device *md = bdev->bd_disk->private_data;
3619 const struct pr_ops *ops;
3623 r = dm_grab_bdev_for_ioctl(md, &bdev, &mode);
3627 ops = bdev->bd_disk->fops->pr_ops;
3628 if (ops && ops->pr_clear)
3629 r = ops->pr_clear(bdev, key);
3637 static const struct pr_ops dm_pr_ops = {
3638 .pr_register = dm_pr_register,
3639 .pr_reserve = dm_pr_reserve,
3640 .pr_release = dm_pr_release,
3641 .pr_preempt = dm_pr_preempt,
3642 .pr_clear = dm_pr_clear,
3645 static const struct block_device_operations dm_blk_dops = {
3646 .open = dm_blk_open,
3647 .release = dm_blk_close,
3648 .ioctl = dm_blk_ioctl,
3649 .getgeo = dm_blk_getgeo,
3650 .pr_ops = &dm_pr_ops,
3651 .owner = THIS_MODULE
3657 module_init(dm_init);
3658 module_exit(dm_exit);
3660 module_param(major, uint, 0);
3661 MODULE_PARM_DESC(major, "The major number of the device mapper");
3663 module_param(reserved_bio_based_ios, uint, S_IRUGO | S_IWUSR);
3664 MODULE_PARM_DESC(reserved_bio_based_ios, "Reserved IOs in bio-based mempools");
3666 module_param(reserved_rq_based_ios, uint, S_IRUGO | S_IWUSR);
3667 MODULE_PARM_DESC(reserved_rq_based_ios, "Reserved IOs in request-based mempools");
3669 module_param(use_blk_mq, bool, S_IRUGO | S_IWUSR);
3670 MODULE_PARM_DESC(use_blk_mq, "Use block multiqueue for request-based DM devices");
3672 MODULE_DESCRIPTION(DM_NAME " driver");
3673 MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
3674 MODULE_LICENSE("GPL");