Root/drivers/md/dm-snap.c

1/*
2 * dm-snapshot.c
3 *
4 * Copyright (C) 2001-2002 Sistina Software (UK) Limited.
5 *
6 * This file is released under the GPL.
7 */
8
9#include <linux/blkdev.h>
10#include <linux/device-mapper.h>
11#include <linux/delay.h>
12#include <linux/fs.h>
13#include <linux/init.h>
14#include <linux/kdev_t.h>
15#include <linux/list.h>
16#include <linux/mempool.h>
17#include <linux/module.h>
18#include <linux/slab.h>
19#include <linux/vmalloc.h>
20#include <linux/log2.h>
21#include <linux/dm-kcopyd.h>
22
23#include "dm-exception-store.h"
24
25#define DM_MSG_PREFIX "snapshots"
26
27static const char dm_snapshot_merge_target_name[] = "snapshot-merge";
28
29#define dm_target_is_snapshot_merge(ti) \
30    ((ti)->type->name == dm_snapshot_merge_target_name)
31
32/*
33 * The size of the mempool used to track chunks in use.
34 */
35#define MIN_IOS 256
36
37#define DM_TRACKED_CHUNK_HASH_SIZE 16
38#define DM_TRACKED_CHUNK_HASH(x) ((unsigned long)(x) & \
39                     (DM_TRACKED_CHUNK_HASH_SIZE - 1))
40
41struct dm_exception_table {
42    uint32_t hash_mask;
43    unsigned hash_shift;
44    struct list_head *table;
45};
46
47struct dm_snapshot {
48    struct rw_semaphore lock;
49
50    struct dm_dev *origin;
51    struct dm_dev *cow;
52
53    struct dm_target *ti;
54
55    /* List of snapshots per Origin */
56    struct list_head list;
57
58    /*
59     * You can't use a snapshot if this is 0 (e.g. if full).
60     * A snapshot-merge target never clears this.
61     */
62    int valid;
63
64    /* Origin writes don't trigger exceptions until this is set */
65    int active;
66
67    atomic_t pending_exceptions_count;
68
69    mempool_t *pending_pool;
70
71    struct dm_exception_table pending;
72    struct dm_exception_table complete;
73
74    /*
75     * pe_lock protects all pending_exception operations and access
76     * as well as the snapshot_bios list.
77     */
78    spinlock_t pe_lock;
79
80    /* Chunks with outstanding reads */
81    spinlock_t tracked_chunk_lock;
82    struct hlist_head tracked_chunk_hash[DM_TRACKED_CHUNK_HASH_SIZE];
83
84    /* The on disk metadata handler */
85    struct dm_exception_store *store;
86
87    struct dm_kcopyd_client *kcopyd_client;
88
89    /* Wait for events based on state_bits */
90    unsigned long state_bits;
91
92    /* Range of chunks currently being merged. */
93    chunk_t first_merging_chunk;
94    int num_merging_chunks;
95
96    /*
97     * The merge operation failed if this flag is set.
98     * Failure modes are handled as follows:
99     * - I/O error reading the header
100     * => don't load the target; abort.
101     * - Header does not have "valid" flag set
102     * => use the origin; forget about the snapshot.
103     * - I/O error when reading exceptions
104     * => don't load the target; abort.
105     * (We can't use the intermediate origin state.)
106     * - I/O error while merging
107     * => stop merging; set merge_failed; process I/O normally.
108     */
109    int merge_failed;
110
111    /*
112     * Incoming bios that overlap with chunks being merged must wait
113     * for them to be committed.
114     */
115    struct bio_list bios_queued_during_merge;
116};
117
118/*
119 * state_bits:
120 * RUNNING_MERGE - Merge operation is in progress.
121 * SHUTDOWN_MERGE - Set to signal that merge needs to be stopped;
122 * cleared afterwards.
123 */
124#define RUNNING_MERGE 0
125#define SHUTDOWN_MERGE 1
126
127DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(snapshot_copy_throttle,
128        "A percentage of time allocated for copy on write");
129
130struct dm_dev *dm_snap_origin(struct dm_snapshot *s)
131{
132    return s->origin;
133}
134EXPORT_SYMBOL(dm_snap_origin);
135
136struct dm_dev *dm_snap_cow(struct dm_snapshot *s)
137{
138    return s->cow;
139}
140EXPORT_SYMBOL(dm_snap_cow);
141
142static sector_t chunk_to_sector(struct dm_exception_store *store,
143                chunk_t chunk)
144{
145    return chunk << store->chunk_shift;
146}
147
148static int bdev_equal(struct block_device *lhs, struct block_device *rhs)
149{
150    /*
151     * There is only ever one instance of a particular block
152     * device so we can compare pointers safely.
153     */
154    return lhs == rhs;
155}
156
157struct dm_snap_pending_exception {
158    struct dm_exception e;
159
160    /*
161     * Origin buffers waiting for this to complete are held
162     * in a bio list
163     */
164    struct bio_list origin_bios;
165    struct bio_list snapshot_bios;
166
167    /* Pointer back to snapshot context */
168    struct dm_snapshot *snap;
169
170    /*
171     * 1 indicates the exception has already been sent to
172     * kcopyd.
173     */
174    int started;
175
176    /*
177     * For writing a complete chunk, bypassing the copy.
178     */
179    struct bio *full_bio;
180    bio_end_io_t *full_bio_end_io;
181    void *full_bio_private;
182};
183
184/*
185 * Hash table mapping origin volumes to lists of snapshots and
186 * a lock to protect it
187 */
188static struct kmem_cache *exception_cache;
189static struct kmem_cache *pending_cache;
190
191struct dm_snap_tracked_chunk {
192    struct hlist_node node;
193    chunk_t chunk;
194};
195
196static void init_tracked_chunk(struct bio *bio)
197{
198    struct dm_snap_tracked_chunk *c = dm_per_bio_data(bio, sizeof(struct dm_snap_tracked_chunk));
199    INIT_HLIST_NODE(&c->node);
200}
201
202static bool is_bio_tracked(struct bio *bio)
203{
204    struct dm_snap_tracked_chunk *c = dm_per_bio_data(bio, sizeof(struct dm_snap_tracked_chunk));
205    return !hlist_unhashed(&c->node);
206}
207
208static void track_chunk(struct dm_snapshot *s, struct bio *bio, chunk_t chunk)
209{
210    struct dm_snap_tracked_chunk *c = dm_per_bio_data(bio, sizeof(struct dm_snap_tracked_chunk));
211
212    c->chunk = chunk;
213
214    spin_lock_irq(&s->tracked_chunk_lock);
215    hlist_add_head(&c->node,
216               &s->tracked_chunk_hash[DM_TRACKED_CHUNK_HASH(chunk)]);
217    spin_unlock_irq(&s->tracked_chunk_lock);
218}
219
220static void stop_tracking_chunk(struct dm_snapshot *s, struct bio *bio)
221{
222    struct dm_snap_tracked_chunk *c = dm_per_bio_data(bio, sizeof(struct dm_snap_tracked_chunk));
223    unsigned long flags;
224
225    spin_lock_irqsave(&s->tracked_chunk_lock, flags);
226    hlist_del(&c->node);
227    spin_unlock_irqrestore(&s->tracked_chunk_lock, flags);
228}
229
230static int __chunk_is_tracked(struct dm_snapshot *s, chunk_t chunk)
231{
232    struct dm_snap_tracked_chunk *c;
233    int found = 0;
234
235    spin_lock_irq(&s->tracked_chunk_lock);
236
237    hlist_for_each_entry(c,
238        &s->tracked_chunk_hash[DM_TRACKED_CHUNK_HASH(chunk)], node) {
239        if (c->chunk == chunk) {
240            found = 1;
241            break;
242        }
243    }
244
245    spin_unlock_irq(&s->tracked_chunk_lock);
246
247    return found;
248}
249
250/*
251 * This conflicting I/O is extremely improbable in the caller,
252 * so msleep(1) is sufficient and there is no need for a wait queue.
253 */
254static void __check_for_conflicting_io(struct dm_snapshot *s, chunk_t chunk)
255{
256    while (__chunk_is_tracked(s, chunk))
257        msleep(1);
258}
259
260/*
261 * One of these per registered origin, held in the snapshot_origins hash
262 */
263struct origin {
264    /* The origin device */
265    struct block_device *bdev;
266
267    struct list_head hash_list;
268
269    /* List of snapshots for this origin */
270    struct list_head snapshots;
271};
272
273/*
274 * Size of the hash table for origin volumes. If we make this
275 * the size of the minors list then it should be nearly perfect
276 */
277#define ORIGIN_HASH_SIZE 256
278#define ORIGIN_MASK 0xFF
279static struct list_head *_origins;
280static struct rw_semaphore _origins_lock;
281
282static DECLARE_WAIT_QUEUE_HEAD(_pending_exceptions_done);
283static DEFINE_SPINLOCK(_pending_exceptions_done_spinlock);
284static uint64_t _pending_exceptions_done_count;
285
286static int init_origin_hash(void)
287{
288    int i;
289
290    _origins = kmalloc(ORIGIN_HASH_SIZE * sizeof(struct list_head),
291               GFP_KERNEL);
292    if (!_origins) {
293        DMERR("unable to allocate memory");
294        return -ENOMEM;
295    }
296
297    for (i = 0; i < ORIGIN_HASH_SIZE; i++)
298        INIT_LIST_HEAD(_origins + i);
299    init_rwsem(&_origins_lock);
300
301    return 0;
302}
303
304static void exit_origin_hash(void)
305{
306    kfree(_origins);
307}
308
309static unsigned origin_hash(struct block_device *bdev)
310{
311    return bdev->bd_dev & ORIGIN_MASK;
312}
313
314static struct origin *__lookup_origin(struct block_device *origin)
315{
316    struct list_head *ol;
317    struct origin *o;
318
319    ol = &_origins[origin_hash(origin)];
320    list_for_each_entry (o, ol, hash_list)
321        if (bdev_equal(o->bdev, origin))
322            return o;
323
324    return NULL;
325}
326
327static void __insert_origin(struct origin *o)
328{
329    struct list_head *sl = &_origins[origin_hash(o->bdev)];
330    list_add_tail(&o->hash_list, sl);
331}
332
333/*
334 * _origins_lock must be held when calling this function.
335 * Returns number of snapshots registered using the supplied cow device, plus:
336 * snap_src - a snapshot suitable for use as a source of exception handover
337 * snap_dest - a snapshot capable of receiving exception handover.
338 * snap_merge - an existing snapshot-merge target linked to the same origin.
339 * There can be at most one snapshot-merge target. The parameter is optional.
340 *
341 * Possible return values and states of snap_src and snap_dest.
342 * 0: NULL, NULL - first new snapshot
343 * 1: snap_src, NULL - normal snapshot
344 * 2: snap_src, snap_dest - waiting for handover
345 * 2: snap_src, NULL - handed over, waiting for old to be deleted
346 * 1: NULL, snap_dest - source got destroyed without handover
347 */
348static int __find_snapshots_sharing_cow(struct dm_snapshot *snap,
349                    struct dm_snapshot **snap_src,
350                    struct dm_snapshot **snap_dest,
351                    struct dm_snapshot **snap_merge)
352{
353    struct dm_snapshot *s;
354    struct origin *o;
355    int count = 0;
356    int active;
357
358    o = __lookup_origin(snap->origin->bdev);
359    if (!o)
360        goto out;
361
362    list_for_each_entry(s, &o->snapshots, list) {
363        if (dm_target_is_snapshot_merge(s->ti) && snap_merge)
364            *snap_merge = s;
365        if (!bdev_equal(s->cow->bdev, snap->cow->bdev))
366            continue;
367
368        down_read(&s->lock);
369        active = s->active;
370        up_read(&s->lock);
371
372        if (active) {
373            if (snap_src)
374                *snap_src = s;
375        } else if (snap_dest)
376            *snap_dest = s;
377
378        count++;
379    }
380
381out:
382    return count;
383}
384
385/*
386 * On success, returns 1 if this snapshot is a handover destination,
387 * otherwise returns 0.
388 */
389static int __validate_exception_handover(struct dm_snapshot *snap)
390{
391    struct dm_snapshot *snap_src = NULL, *snap_dest = NULL;
392    struct dm_snapshot *snap_merge = NULL;
393
394    /* Does snapshot need exceptions handed over to it? */
395    if ((__find_snapshots_sharing_cow(snap, &snap_src, &snap_dest,
396                      &snap_merge) == 2) ||
397        snap_dest) {
398        snap->ti->error = "Snapshot cow pairing for exception "
399                  "table handover failed";
400        return -EINVAL;
401    }
402
403    /*
404     * If no snap_src was found, snap cannot become a handover
405     * destination.
406     */
407    if (!snap_src)
408        return 0;
409
410    /*
411     * Non-snapshot-merge handover?
412     */
413    if (!dm_target_is_snapshot_merge(snap->ti))
414        return 1;
415
416    /*
417     * Do not allow more than one merging snapshot.
418     */
419    if (snap_merge) {
420        snap->ti->error = "A snapshot is already merging.";
421        return -EINVAL;
422    }
423
424    if (!snap_src->store->type->prepare_merge ||
425        !snap_src->store->type->commit_merge) {
426        snap->ti->error = "Snapshot exception store does not "
427                  "support snapshot-merge.";
428        return -EINVAL;
429    }
430
431    return 1;
432}
433
434static void __insert_snapshot(struct origin *o, struct dm_snapshot *s)
435{
436    struct dm_snapshot *l;
437
438    /* Sort the list according to chunk size, largest-first smallest-last */
439    list_for_each_entry(l, &o->snapshots, list)
440        if (l->store->chunk_size < s->store->chunk_size)
441            break;
442    list_add_tail(&s->list, &l->list);
443}
444
445/*
446 * Make a note of the snapshot and its origin so we can look it
447 * up when the origin has a write on it.
448 *
449 * Also validate snapshot exception store handovers.
450 * On success, returns 1 if this registration is a handover destination,
451 * otherwise returns 0.
452 */
453static int register_snapshot(struct dm_snapshot *snap)
454{
455    struct origin *o, *new_o = NULL;
456    struct block_device *bdev = snap->origin->bdev;
457    int r = 0;
458
459    new_o = kmalloc(sizeof(*new_o), GFP_KERNEL);
460    if (!new_o)
461        return -ENOMEM;
462
463    down_write(&_origins_lock);
464
465    r = __validate_exception_handover(snap);
466    if (r < 0) {
467        kfree(new_o);
468        goto out;
469    }
470
471    o = __lookup_origin(bdev);
472    if (o)
473        kfree(new_o);
474    else {
475        /* New origin */
476        o = new_o;
477
478        /* Initialise the struct */
479        INIT_LIST_HEAD(&o->snapshots);
480        o->bdev = bdev;
481
482        __insert_origin(o);
483    }
484
485    __insert_snapshot(o, snap);
486
487out:
488    up_write(&_origins_lock);
489
490    return r;
491}
492
493/*
494 * Move snapshot to correct place in list according to chunk size.
495 */
496static void reregister_snapshot(struct dm_snapshot *s)
497{
498    struct block_device *bdev = s->origin->bdev;
499
500    down_write(&_origins_lock);
501
502    list_del(&s->list);
503    __insert_snapshot(__lookup_origin(bdev), s);
504
505    up_write(&_origins_lock);
506}
507
508static void unregister_snapshot(struct dm_snapshot *s)
509{
510    struct origin *o;
511
512    down_write(&_origins_lock);
513    o = __lookup_origin(s->origin->bdev);
514
515    list_del(&s->list);
516    if (o && list_empty(&o->snapshots)) {
517        list_del(&o->hash_list);
518        kfree(o);
519    }
520
521    up_write(&_origins_lock);
522}
523
524/*
525 * Implementation of the exception hash tables.
526 * The lowest hash_shift bits of the chunk number are ignored, allowing
527 * some consecutive chunks to be grouped together.
528 */
529static int dm_exception_table_init(struct dm_exception_table *et,
530                   uint32_t size, unsigned hash_shift)
531{
532    unsigned int i;
533
534    et->hash_shift = hash_shift;
535    et->hash_mask = size - 1;
536    et->table = dm_vcalloc(size, sizeof(struct list_head));
537    if (!et->table)
538        return -ENOMEM;
539
540    for (i = 0; i < size; i++)
541        INIT_LIST_HEAD(et->table + i);
542
543    return 0;
544}
545
546static void dm_exception_table_exit(struct dm_exception_table *et,
547                    struct kmem_cache *mem)
548{
549    struct list_head *slot;
550    struct dm_exception *ex, *next;
551    int i, size;
552
553    size = et->hash_mask + 1;
554    for (i = 0; i < size; i++) {
555        slot = et->table + i;
556
557        list_for_each_entry_safe (ex, next, slot, hash_list)
558            kmem_cache_free(mem, ex);
559    }
560
561    vfree(et->table);
562}
563
564static uint32_t exception_hash(struct dm_exception_table *et, chunk_t chunk)
565{
566    return (chunk >> et->hash_shift) & et->hash_mask;
567}
568
569static void dm_remove_exception(struct dm_exception *e)
570{
571    list_del(&e->hash_list);
572}
573
574/*
575 * Return the exception data for a sector, or NULL if not
576 * remapped.
577 */
578static struct dm_exception *dm_lookup_exception(struct dm_exception_table *et,
579                        chunk_t chunk)
580{
581    struct list_head *slot;
582    struct dm_exception *e;
583
584    slot = &et->table[exception_hash(et, chunk)];
585    list_for_each_entry (e, slot, hash_list)
586        if (chunk >= e->old_chunk &&
587            chunk <= e->old_chunk + dm_consecutive_chunk_count(e))
588            return e;
589
590    return NULL;
591}
592
593static struct dm_exception *alloc_completed_exception(void)
594{
595    struct dm_exception *e;
596
597    e = kmem_cache_alloc(exception_cache, GFP_NOIO);
598    if (!e)
599        e = kmem_cache_alloc(exception_cache, GFP_ATOMIC);
600
601    return e;
602}
603
604static void free_completed_exception(struct dm_exception *e)
605{
606    kmem_cache_free(exception_cache, e);
607}
608
609static struct dm_snap_pending_exception *alloc_pending_exception(struct dm_snapshot *s)
610{
611    struct dm_snap_pending_exception *pe = mempool_alloc(s->pending_pool,
612                                 GFP_NOIO);
613
614    atomic_inc(&s->pending_exceptions_count);
615    pe->snap = s;
616
617    return pe;
618}
619
620static void free_pending_exception(struct dm_snap_pending_exception *pe)
621{
622    struct dm_snapshot *s = pe->snap;
623
624    mempool_free(pe, s->pending_pool);
625    smp_mb__before_atomic_dec();
626    atomic_dec(&s->pending_exceptions_count);
627}
628
629static void dm_insert_exception(struct dm_exception_table *eh,
630                struct dm_exception *new_e)
631{
632    struct list_head *l;
633    struct dm_exception *e = NULL;
634
635    l = &eh->table[exception_hash(eh, new_e->old_chunk)];
636
637    /* Add immediately if this table doesn't support consecutive chunks */
638    if (!eh->hash_shift)
639        goto out;
640
641    /* List is ordered by old_chunk */
642    list_for_each_entry_reverse(e, l, hash_list) {
643        /* Insert after an existing chunk? */
644        if (new_e->old_chunk == (e->old_chunk +
645                     dm_consecutive_chunk_count(e) + 1) &&
646            new_e->new_chunk == (dm_chunk_number(e->new_chunk) +
647                     dm_consecutive_chunk_count(e) + 1)) {
648            dm_consecutive_chunk_count_inc(e);
649            free_completed_exception(new_e);
650            return;
651        }
652
653        /* Insert before an existing chunk? */
654        if (new_e->old_chunk == (e->old_chunk - 1) &&
655            new_e->new_chunk == (dm_chunk_number(e->new_chunk) - 1)) {
656            dm_consecutive_chunk_count_inc(e);
657            e->old_chunk--;
658            e->new_chunk--;
659            free_completed_exception(new_e);
660            return;
661        }
662
663        if (new_e->old_chunk > e->old_chunk)
664            break;
665    }
666
667out:
668    list_add(&new_e->hash_list, e ? &e->hash_list : l);
669}
670
671/*
672 * Callback used by the exception stores to load exceptions when
673 * initialising.
674 */
675static int dm_add_exception(void *context, chunk_t old, chunk_t new)
676{
677    struct dm_snapshot *s = context;
678    struct dm_exception *e;
679
680    e = alloc_completed_exception();
681    if (!e)
682        return -ENOMEM;
683
684    e->old_chunk = old;
685
686    /* Consecutive_count is implicitly initialised to zero */
687    e->new_chunk = new;
688
689    dm_insert_exception(&s->complete, e);
690
691    return 0;
692}
693
694/*
695 * Return a minimum chunk size of all snapshots that have the specified origin.
696 * Return zero if the origin has no snapshots.
697 */
698static uint32_t __minimum_chunk_size(struct origin *o)
699{
700    struct dm_snapshot *snap;
701    unsigned chunk_size = 0;
702
703    if (o)
704        list_for_each_entry(snap, &o->snapshots, list)
705            chunk_size = min_not_zero(chunk_size,
706                          snap->store->chunk_size);
707
708    return (uint32_t) chunk_size;
709}
710
711/*
712 * Hard coded magic.
713 */
714static int calc_max_buckets(void)
715{
716    /* use a fixed size of 2MB */
717    unsigned long mem = 2 * 1024 * 1024;
718    mem /= sizeof(struct list_head);
719
720    return mem;
721}
722
723/*
724 * Allocate room for a suitable hash table.
725 */
726static int init_hash_tables(struct dm_snapshot *s)
727{
728    sector_t hash_size, cow_dev_size, origin_dev_size, max_buckets;
729
730    /*
731     * Calculate based on the size of the original volume or
732     * the COW volume...
733     */
734    cow_dev_size = get_dev_size(s->cow->bdev);
735    origin_dev_size = get_dev_size(s->origin->bdev);
736    max_buckets = calc_max_buckets();
737
738    hash_size = min(origin_dev_size, cow_dev_size) >> s->store->chunk_shift;
739    hash_size = min(hash_size, max_buckets);
740
741    if (hash_size < 64)
742        hash_size = 64;
743    hash_size = rounddown_pow_of_two(hash_size);
744    if (dm_exception_table_init(&s->complete, hash_size,
745                    DM_CHUNK_CONSECUTIVE_BITS))
746        return -ENOMEM;
747
748    /*
749     * Allocate hash table for in-flight exceptions
750     * Make this smaller than the real hash table
751     */
752    hash_size >>= 3;
753    if (hash_size < 64)
754        hash_size = 64;
755
756    if (dm_exception_table_init(&s->pending, hash_size, 0)) {
757        dm_exception_table_exit(&s->complete, exception_cache);
758        return -ENOMEM;
759    }
760
761    return 0;
762}
763
764static void merge_shutdown(struct dm_snapshot *s)
765{
766    clear_bit_unlock(RUNNING_MERGE, &s->state_bits);
767    smp_mb__after_clear_bit();
768    wake_up_bit(&s->state_bits, RUNNING_MERGE);
769}
770
771static struct bio *__release_queued_bios_after_merge(struct dm_snapshot *s)
772{
773    s->first_merging_chunk = 0;
774    s->num_merging_chunks = 0;
775
776    return bio_list_get(&s->bios_queued_during_merge);
777}
778
779/*
780 * Remove one chunk from the index of completed exceptions.
781 */
782static int __remove_single_exception_chunk(struct dm_snapshot *s,
783                       chunk_t old_chunk)
784{
785    struct dm_exception *e;
786
787    e = dm_lookup_exception(&s->complete, old_chunk);
788    if (!e) {
789        DMERR("Corruption detected: exception for block %llu is "
790              "on disk but not in memory",
791              (unsigned long long)old_chunk);
792        return -EINVAL;
793    }
794
795    /*
796     * If this is the only chunk using this exception, remove exception.
797     */
798    if (!dm_consecutive_chunk_count(e)) {
799        dm_remove_exception(e);
800        free_completed_exception(e);
801        return 0;
802    }
803
804    /*
805     * The chunk may be either at the beginning or the end of a
806     * group of consecutive chunks - never in the middle. We are
807     * removing chunks in the opposite order to that in which they
808     * were added, so this should always be true.
809     * Decrement the consecutive chunk counter and adjust the
810     * starting point if necessary.
811     */
812    if (old_chunk == e->old_chunk) {
813        e->old_chunk++;
814        e->new_chunk++;
815    } else if (old_chunk != e->old_chunk +
816           dm_consecutive_chunk_count(e)) {
817        DMERR("Attempt to merge block %llu from the "
818              "middle of a chunk range [%llu - %llu]",
819              (unsigned long long)old_chunk,
820              (unsigned long long)e->old_chunk,
821              (unsigned long long)
822              e->old_chunk + dm_consecutive_chunk_count(e));
823        return -EINVAL;
824    }
825
826    dm_consecutive_chunk_count_dec(e);
827
828    return 0;
829}
830
831static void flush_bios(struct bio *bio);
832
833static int remove_single_exception_chunk(struct dm_snapshot *s)
834{
835    struct bio *b = NULL;
836    int r;
837    chunk_t old_chunk = s->first_merging_chunk + s->num_merging_chunks - 1;
838
839    down_write(&s->lock);
840
841    /*
842     * Process chunks (and associated exceptions) in reverse order
843     * so that dm_consecutive_chunk_count_dec() accounting works.
844     */
845    do {
846        r = __remove_single_exception_chunk(s, old_chunk);
847        if (r)
848            goto out;
849    } while (old_chunk-- > s->first_merging_chunk);
850
851    b = __release_queued_bios_after_merge(s);
852
853out:
854    up_write(&s->lock);
855    if (b)
856        flush_bios(b);
857
858    return r;
859}
860
861static int origin_write_extent(struct dm_snapshot *merging_snap,
862                   sector_t sector, unsigned chunk_size);
863
864static void merge_callback(int read_err, unsigned long write_err,
865               void *context);
866
867static uint64_t read_pending_exceptions_done_count(void)
868{
869    uint64_t pending_exceptions_done;
870
871    spin_lock(&_pending_exceptions_done_spinlock);
872    pending_exceptions_done = _pending_exceptions_done_count;
873    spin_unlock(&_pending_exceptions_done_spinlock);
874
875    return pending_exceptions_done;
876}
877
878static void increment_pending_exceptions_done_count(void)
879{
880    spin_lock(&_pending_exceptions_done_spinlock);
881    _pending_exceptions_done_count++;
882    spin_unlock(&_pending_exceptions_done_spinlock);
883
884    wake_up_all(&_pending_exceptions_done);
885}
886
887static void snapshot_merge_next_chunks(struct dm_snapshot *s)
888{
889    int i, linear_chunks;
890    chunk_t old_chunk, new_chunk;
891    struct dm_io_region src, dest;
892    sector_t io_size;
893    uint64_t previous_count;
894
895    BUG_ON(!test_bit(RUNNING_MERGE, &s->state_bits));
896    if (unlikely(test_bit(SHUTDOWN_MERGE, &s->state_bits)))
897        goto shut;
898
899    /*
900     * valid flag never changes during merge, so no lock required.
901     */
902    if (!s->valid) {
903        DMERR("Snapshot is invalid: can't merge");
904        goto shut;
905    }
906
907    linear_chunks = s->store->type->prepare_merge(s->store, &old_chunk,
908                              &new_chunk);
909    if (linear_chunks <= 0) {
910        if (linear_chunks < 0) {
911            DMERR("Read error in exception store: "
912                  "shutting down merge");
913            down_write(&s->lock);
914            s->merge_failed = 1;
915            up_write(&s->lock);
916        }
917        goto shut;
918    }
919
920    /* Adjust old_chunk and new_chunk to reflect start of linear region */
921    old_chunk = old_chunk + 1 - linear_chunks;
922    new_chunk = new_chunk + 1 - linear_chunks;
923
924    /*
925     * Use one (potentially large) I/O to copy all 'linear_chunks'
926     * from the exception store to the origin
927     */
928    io_size = linear_chunks * s->store->chunk_size;
929
930    dest.bdev = s->origin->bdev;
931    dest.sector = chunk_to_sector(s->store, old_chunk);
932    dest.count = min(io_size, get_dev_size(dest.bdev) - dest.sector);
933
934    src.bdev = s->cow->bdev;
935    src.sector = chunk_to_sector(s->store, new_chunk);
936    src.count = dest.count;
937
938    /*
939     * Reallocate any exceptions needed in other snapshots then
940     * wait for the pending exceptions to complete.
941     * Each time any pending exception (globally on the system)
942     * completes we are woken and repeat the process to find out
943     * if we can proceed. While this may not seem a particularly
944     * efficient algorithm, it is not expected to have any
945     * significant impact on performance.
946     */
947    previous_count = read_pending_exceptions_done_count();
948    while (origin_write_extent(s, dest.sector, io_size)) {
949        wait_event(_pending_exceptions_done,
950               (read_pending_exceptions_done_count() !=
951                previous_count));
952        /* Retry after the wait, until all exceptions are done. */
953        previous_count = read_pending_exceptions_done_count();
954    }
955
956    down_write(&s->lock);
957    s->first_merging_chunk = old_chunk;
958    s->num_merging_chunks = linear_chunks;
959    up_write(&s->lock);
960
961    /* Wait until writes to all 'linear_chunks' drain */
962    for (i = 0; i < linear_chunks; i++)
963        __check_for_conflicting_io(s, old_chunk + i);
964
965    dm_kcopyd_copy(s->kcopyd_client, &src, 1, &dest, 0, merge_callback, s);
966    return;
967
968shut:
969    merge_shutdown(s);
970}
971
972static void error_bios(struct bio *bio);
973
974static void merge_callback(int read_err, unsigned long write_err, void *context)
975{
976    struct dm_snapshot *s = context;
977    struct bio *b = NULL;
978
979    if (read_err || write_err) {
980        if (read_err)
981            DMERR("Read error: shutting down merge.");
982        else
983            DMERR("Write error: shutting down merge.");
984        goto shut;
985    }
986
987    if (s->store->type->commit_merge(s->store,
988                     s->num_merging_chunks) < 0) {
989        DMERR("Write error in exception store: shutting down merge");
990        goto shut;
991    }
992
993    if (remove_single_exception_chunk(s) < 0)
994        goto shut;
995
996    snapshot_merge_next_chunks(s);
997
998    return;
999
1000shut:
1001    down_write(&s->lock);
1002    s->merge_failed = 1;
1003    b = __release_queued_bios_after_merge(s);
1004    up_write(&s->lock);
1005    error_bios(b);
1006
1007    merge_shutdown(s);
1008}
1009
1010static void start_merge(struct dm_snapshot *s)
1011{
1012    if (!test_and_set_bit(RUNNING_MERGE, &s->state_bits))
1013        snapshot_merge_next_chunks(s);
1014}
1015
1016static int wait_schedule(void *ptr)
1017{
1018    schedule();
1019
1020    return 0;
1021}
1022
1023/*
1024 * Stop the merging process and wait until it finishes.
1025 */
1026static void stop_merge(struct dm_snapshot *s)
1027{
1028    set_bit(SHUTDOWN_MERGE, &s->state_bits);
1029    wait_on_bit(&s->state_bits, RUNNING_MERGE, wait_schedule,
1030            TASK_UNINTERRUPTIBLE);
1031    clear_bit(SHUTDOWN_MERGE, &s->state_bits);
1032}
1033
1034/*
1035 * Construct a snapshot mapping: <origin_dev> <COW-dev> <p/n> <chunk-size>
1036 */
1037static int snapshot_ctr(struct dm_target *ti, unsigned int argc, char **argv)
1038{
1039    struct dm_snapshot *s;
1040    int i;
1041    int r = -EINVAL;
1042    char *origin_path, *cow_path;
1043    unsigned args_used, num_flush_bios = 1;
1044    fmode_t origin_mode = FMODE_READ;
1045
1046    if (argc != 4) {
1047        ti->error = "requires exactly 4 arguments";
1048        r = -EINVAL;
1049        goto bad;
1050    }
1051
1052    if (dm_target_is_snapshot_merge(ti)) {
1053        num_flush_bios = 2;
1054        origin_mode = FMODE_WRITE;
1055    }
1056
1057    s = kmalloc(sizeof(*s), GFP_KERNEL);
1058    if (!s) {
1059        ti->error = "Cannot allocate private snapshot structure";
1060        r = -ENOMEM;
1061        goto bad;
1062    }
1063
1064    origin_path = argv[0];
1065    argv++;
1066    argc--;
1067
1068    r = dm_get_device(ti, origin_path, origin_mode, &s->origin);
1069    if (r) {
1070        ti->error = "Cannot get origin device";
1071        goto bad_origin;
1072    }
1073
1074    cow_path = argv[0];
1075    argv++;
1076    argc--;
1077
1078    r = dm_get_device(ti, cow_path, dm_table_get_mode(ti->table), &s->cow);
1079    if (r) {
1080        ti->error = "Cannot get COW device";
1081        goto bad_cow;
1082    }
1083
1084    r = dm_exception_store_create(ti, argc, argv, s, &args_used, &s->store);
1085    if (r) {
1086        ti->error = "Couldn't create exception store";
1087        r = -EINVAL;
1088        goto bad_store;
1089    }
1090
1091    argv += args_used;
1092    argc -= args_used;
1093
1094    s->ti = ti;
1095    s->valid = 1;
1096    s->active = 0;
1097    atomic_set(&s->pending_exceptions_count, 0);
1098    init_rwsem(&s->lock);
1099    INIT_LIST_HEAD(&s->list);
1100    spin_lock_init(&s->pe_lock);
1101    s->state_bits = 0;
1102    s->merge_failed = 0;
1103    s->first_merging_chunk = 0;
1104    s->num_merging_chunks = 0;
1105    bio_list_init(&s->bios_queued_during_merge);
1106
1107    /* Allocate hash table for COW data */
1108    if (init_hash_tables(s)) {
1109        ti->error = "Unable to allocate hash table space";
1110        r = -ENOMEM;
1111        goto bad_hash_tables;
1112    }
1113
1114    s->kcopyd_client = dm_kcopyd_client_create(&dm_kcopyd_throttle);
1115    if (IS_ERR(s->kcopyd_client)) {
1116        r = PTR_ERR(s->kcopyd_client);
1117        ti->error = "Could not create kcopyd client";
1118        goto bad_kcopyd;
1119    }
1120
1121    s->pending_pool = mempool_create_slab_pool(MIN_IOS, pending_cache);
1122    if (!s->pending_pool) {
1123        ti->error = "Could not allocate mempool for pending exceptions";
1124        goto bad_pending_pool;
1125    }
1126
1127    for (i = 0; i < DM_TRACKED_CHUNK_HASH_SIZE; i++)
1128        INIT_HLIST_HEAD(&s->tracked_chunk_hash[i]);
1129
1130    spin_lock_init(&s->tracked_chunk_lock);
1131
1132    ti->private = s;
1133    ti->num_flush_bios = num_flush_bios;
1134    ti->per_bio_data_size = sizeof(struct dm_snap_tracked_chunk);
1135
1136    /* Add snapshot to the list of snapshots for this origin */
1137    /* Exceptions aren't triggered till snapshot_resume() is called */
1138    r = register_snapshot(s);
1139    if (r == -ENOMEM) {
1140        ti->error = "Snapshot origin struct allocation failed";
1141        goto bad_load_and_register;
1142    } else if (r < 0) {
1143        /* invalid handover, register_snapshot has set ti->error */
1144        goto bad_load_and_register;
1145    }
1146
1147    /*
1148     * Metadata must only be loaded into one table at once, so skip this
1149     * if metadata will be handed over during resume.
1150     * Chunk size will be set during the handover - set it to zero to
1151     * ensure it's ignored.
1152     */
1153    if (r > 0) {
1154        s->store->chunk_size = 0;
1155        return 0;
1156    }
1157
1158    r = s->store->type->read_metadata(s->store, dm_add_exception,
1159                      (void *)s);
1160    if (r < 0) {
1161        ti->error = "Failed to read snapshot metadata";
1162        goto bad_read_metadata;
1163    } else if (r > 0) {
1164        s->valid = 0;
1165        DMWARN("Snapshot is marked invalid.");
1166    }
1167
1168    if (!s->store->chunk_size) {
1169        ti->error = "Chunk size not set";
1170        goto bad_read_metadata;
1171    }
1172
1173    r = dm_set_target_max_io_len(ti, s->store->chunk_size);
1174    if (r)
1175        goto bad_read_metadata;
1176
1177    return 0;
1178
1179bad_read_metadata:
1180    unregister_snapshot(s);
1181
1182bad_load_and_register:
1183    mempool_destroy(s->pending_pool);
1184
1185bad_pending_pool:
1186    dm_kcopyd_client_destroy(s->kcopyd_client);
1187
1188bad_kcopyd:
1189    dm_exception_table_exit(&s->pending, pending_cache);
1190    dm_exception_table_exit(&s->complete, exception_cache);
1191
1192bad_hash_tables:
1193    dm_exception_store_destroy(s->store);
1194
1195bad_store:
1196    dm_put_device(ti, s->cow);
1197
1198bad_cow:
1199    dm_put_device(ti, s->origin);
1200
1201bad_origin:
1202    kfree(s);
1203
1204bad:
1205    return r;
1206}
1207
1208static void __free_exceptions(struct dm_snapshot *s)
1209{
1210    dm_kcopyd_client_destroy(s->kcopyd_client);
1211    s->kcopyd_client = NULL;
1212
1213    dm_exception_table_exit(&s->pending, pending_cache);
1214    dm_exception_table_exit(&s->complete, exception_cache);
1215}
1216
1217static void __handover_exceptions(struct dm_snapshot *snap_src,
1218                  struct dm_snapshot *snap_dest)
1219{
1220    union {
1221        struct dm_exception_table table_swap;
1222        struct dm_exception_store *store_swap;
1223    } u;
1224
1225    /*
1226     * Swap all snapshot context information between the two instances.
1227     */
1228    u.table_swap = snap_dest->complete;
1229    snap_dest->complete = snap_src->complete;
1230    snap_src->complete = u.table_swap;
1231
1232    u.store_swap = snap_dest->store;
1233    snap_dest->store = snap_src->store;
1234    snap_src->store = u.store_swap;
1235
1236    snap_dest->store->snap = snap_dest;
1237    snap_src->store->snap = snap_src;
1238
1239    snap_dest->ti->max_io_len = snap_dest->store->chunk_size;
1240    snap_dest->valid = snap_src->valid;
1241
1242    /*
1243     * Set source invalid to ensure it receives no further I/O.
1244     */
1245    snap_src->valid = 0;
1246}
1247
1248static void snapshot_dtr(struct dm_target *ti)
1249{
1250#ifdef CONFIG_DM_DEBUG
1251    int i;
1252#endif
1253    struct dm_snapshot *s = ti->private;
1254    struct dm_snapshot *snap_src = NULL, *snap_dest = NULL;
1255
1256    down_read(&_origins_lock);
1257    /* Check whether exception handover must be cancelled */
1258    (void) __find_snapshots_sharing_cow(s, &snap_src, &snap_dest, NULL);
1259    if (snap_src && snap_dest && (s == snap_src)) {
1260        down_write(&snap_dest->lock);
1261        snap_dest->valid = 0;
1262        up_write(&snap_dest->lock);
1263        DMERR("Cancelling snapshot handover.");
1264    }
1265    up_read(&_origins_lock);
1266
1267    if (dm_target_is_snapshot_merge(ti))
1268        stop_merge(s);
1269
1270    /* Prevent further origin writes from using this snapshot. */
1271    /* After this returns there can be no new kcopyd jobs. */
1272    unregister_snapshot(s);
1273
1274    while (atomic_read(&s->pending_exceptions_count))
1275        msleep(1);
1276    /*
1277     * Ensure instructions in mempool_destroy aren't reordered
1278     * before atomic_read.
1279     */
1280    smp_mb();
1281
1282#ifdef CONFIG_DM_DEBUG
1283    for (i = 0; i < DM_TRACKED_CHUNK_HASH_SIZE; i++)
1284        BUG_ON(!hlist_empty(&s->tracked_chunk_hash[i]));
1285#endif
1286
1287    __free_exceptions(s);
1288
1289    mempool_destroy(s->pending_pool);
1290
1291    dm_exception_store_destroy(s->store);
1292
1293    dm_put_device(ti, s->cow);
1294
1295    dm_put_device(ti, s->origin);
1296
1297    kfree(s);
1298}
1299
1300/*
1301 * Flush a list of buffers.
1302 */
1303static void flush_bios(struct bio *bio)
1304{
1305    struct bio *n;
1306
1307    while (bio) {
1308        n = bio->bi_next;
1309        bio->bi_next = NULL;
1310        generic_make_request(bio);
1311        bio = n;
1312    }
1313}
1314
1315static int do_origin(struct dm_dev *origin, struct bio *bio);
1316
1317/*
1318 * Flush a list of buffers.
1319 */
1320static void retry_origin_bios(struct dm_snapshot *s, struct bio *bio)
1321{
1322    struct bio *n;
1323    int r;
1324
1325    while (bio) {
1326        n = bio->bi_next;
1327        bio->bi_next = NULL;
1328        r = do_origin(s->origin, bio);
1329        if (r == DM_MAPIO_REMAPPED)
1330            generic_make_request(bio);
1331        bio = n;
1332    }
1333}
1334
1335/*
1336 * Error a list of buffers.
1337 */
1338static void error_bios(struct bio *bio)
1339{
1340    struct bio *n;
1341
1342    while (bio) {
1343        n = bio->bi_next;
1344        bio->bi_next = NULL;
1345        bio_io_error(bio);
1346        bio = n;
1347    }
1348}
1349
1350static void __invalidate_snapshot(struct dm_snapshot *s, int err)
1351{
1352    if (!s->valid)
1353        return;
1354
1355    if (err == -EIO)
1356        DMERR("Invalidating snapshot: Error reading/writing.");
1357    else if (err == -ENOMEM)
1358        DMERR("Invalidating snapshot: Unable to allocate exception.");
1359
1360    if (s->store->type->drop_snapshot)
1361        s->store->type->drop_snapshot(s->store);
1362
1363    s->valid = 0;
1364
1365    dm_table_event(s->ti->table);
1366}
1367
1368static void pending_complete(struct dm_snap_pending_exception *pe, int success)
1369{
1370    struct dm_exception *e;
1371    struct dm_snapshot *s = pe->snap;
1372    struct bio *origin_bios = NULL;
1373    struct bio *snapshot_bios = NULL;
1374    struct bio *full_bio = NULL;
1375    int error = 0;
1376
1377    if (!success) {
1378        /* Read/write error - snapshot is unusable */
1379        down_write(&s->lock);
1380        __invalidate_snapshot(s, -EIO);
1381        error = 1;
1382        goto out;
1383    }
1384
1385    e = alloc_completed_exception();
1386    if (!e) {
1387        down_write(&s->lock);
1388        __invalidate_snapshot(s, -ENOMEM);
1389        error = 1;
1390        goto out;
1391    }
1392    *e = pe->e;
1393
1394    down_write(&s->lock);
1395    if (!s->valid) {
1396        free_completed_exception(e);
1397        error = 1;
1398        goto out;
1399    }
1400
1401    /* Check for conflicting reads */
1402    __check_for_conflicting_io(s, pe->e.old_chunk);
1403
1404    /*
1405     * Add a proper exception, and remove the
1406     * in-flight exception from the list.
1407     */
1408    dm_insert_exception(&s->complete, e);
1409
1410out:
1411    dm_remove_exception(&pe->e);
1412    snapshot_bios = bio_list_get(&pe->snapshot_bios);
1413    origin_bios = bio_list_get(&pe->origin_bios);
1414    full_bio = pe->full_bio;
1415    if (full_bio) {
1416        full_bio->bi_end_io = pe->full_bio_end_io;
1417        full_bio->bi_private = pe->full_bio_private;
1418    }
1419    free_pending_exception(pe);
1420
1421    increment_pending_exceptions_done_count();
1422
1423    up_write(&s->lock);
1424
1425    /* Submit any pending write bios */
1426    if (error) {
1427        if (full_bio)
1428            bio_io_error(full_bio);
1429        error_bios(snapshot_bios);
1430    } else {
1431        if (full_bio)
1432            bio_endio(full_bio, 0);
1433        flush_bios(snapshot_bios);
1434    }
1435
1436    retry_origin_bios(s, origin_bios);
1437}
1438
1439static void commit_callback(void *context, int success)
1440{
1441    struct dm_snap_pending_exception *pe = context;
1442
1443    pending_complete(pe, success);
1444}
1445
1446/*
1447 * Called when the copy I/O has finished. kcopyd actually runs
1448 * this code so don't block.
1449 */
1450static void copy_callback(int read_err, unsigned long write_err, void *context)
1451{
1452    struct dm_snap_pending_exception *pe = context;
1453    struct dm_snapshot *s = pe->snap;
1454
1455    if (read_err || write_err)
1456        pending_complete(pe, 0);
1457
1458    else
1459        /* Update the metadata if we are persistent */
1460        s->store->type->commit_exception(s->store, &pe->e,
1461                         commit_callback, pe);
1462}
1463
1464/*
1465 * Dispatches the copy operation to kcopyd.
1466 */
1467static void start_copy(struct dm_snap_pending_exception *pe)
1468{
1469    struct dm_snapshot *s = pe->snap;
1470    struct dm_io_region src, dest;
1471    struct block_device *bdev = s->origin->bdev;
1472    sector_t dev_size;
1473
1474    dev_size = get_dev_size(bdev);
1475
1476    src.bdev = bdev;
1477    src.sector = chunk_to_sector(s->store, pe->e.old_chunk);
1478    src.count = min((sector_t)s->store->chunk_size, dev_size - src.sector);
1479
1480    dest.bdev = s->cow->bdev;
1481    dest.sector = chunk_to_sector(s->store, pe->e.new_chunk);
1482    dest.count = src.count;
1483
1484    /* Hand over to kcopyd */
1485    dm_kcopyd_copy(s->kcopyd_client, &src, 1, &dest, 0, copy_callback, pe);
1486}
1487
1488static void full_bio_end_io(struct bio *bio, int error)
1489{
1490    void *callback_data = bio->bi_private;
1491
1492    dm_kcopyd_do_callback(callback_data, 0, error ? 1 : 0);
1493}
1494
1495static void start_full_bio(struct dm_snap_pending_exception *pe,
1496               struct bio *bio)
1497{
1498    struct dm_snapshot *s = pe->snap;
1499    void *callback_data;
1500
1501    pe->full_bio = bio;
1502    pe->full_bio_end_io = bio->bi_end_io;
1503    pe->full_bio_private = bio->bi_private;
1504
1505    callback_data = dm_kcopyd_prepare_callback(s->kcopyd_client,
1506                           copy_callback, pe);
1507
1508    bio->bi_end_io = full_bio_end_io;
1509    bio->bi_private = callback_data;
1510
1511    generic_make_request(bio);
1512}
1513
1514static struct dm_snap_pending_exception *
1515__lookup_pending_exception(struct dm_snapshot *s, chunk_t chunk)
1516{
1517    struct dm_exception *e = dm_lookup_exception(&s->pending, chunk);
1518
1519    if (!e)
1520        return NULL;
1521
1522    return container_of(e, struct dm_snap_pending_exception, e);
1523}
1524
1525/*
1526 * Looks to see if this snapshot already has a pending exception
1527 * for this chunk, otherwise it allocates a new one and inserts
1528 * it into the pending table.
1529 *
1530 * NOTE: a write lock must be held on snap->lock before calling
1531 * this.
1532 */
1533static struct dm_snap_pending_exception *
1534__find_pending_exception(struct dm_snapshot *s,
1535             struct dm_snap_pending_exception *pe, chunk_t chunk)
1536{
1537    struct dm_snap_pending_exception *pe2;
1538
1539    pe2 = __lookup_pending_exception(s, chunk);
1540    if (pe2) {
1541        free_pending_exception(pe);
1542        return pe2;
1543    }
1544
1545    pe->e.old_chunk = chunk;
1546    bio_list_init(&pe->origin_bios);
1547    bio_list_init(&pe->snapshot_bios);
1548    pe->started = 0;
1549    pe->full_bio = NULL;
1550
1551    if (s->store->type->prepare_exception(s->store, &pe->e)) {
1552        free_pending_exception(pe);
1553        return NULL;
1554    }
1555
1556    dm_insert_exception(&s->pending, &pe->e);
1557
1558    return pe;
1559}
1560
1561static void remap_exception(struct dm_snapshot *s, struct dm_exception *e,
1562                struct bio *bio, chunk_t chunk)
1563{
1564    bio->bi_bdev = s->cow->bdev;
1565    bio->bi_sector = chunk_to_sector(s->store,
1566                     dm_chunk_number(e->new_chunk) +
1567                     (chunk - e->old_chunk)) +
1568                     (bio->bi_sector &
1569                      s->store->chunk_mask);
1570}
1571
1572static int snapshot_map(struct dm_target *ti, struct bio *bio)
1573{
1574    struct dm_exception *e;
1575    struct dm_snapshot *s = ti->private;
1576    int r = DM_MAPIO_REMAPPED;
1577    chunk_t chunk;
1578    struct dm_snap_pending_exception *pe = NULL;
1579
1580    init_tracked_chunk(bio);
1581
1582    if (bio->bi_rw & REQ_FLUSH) {
1583        bio->bi_bdev = s->cow->bdev;
1584        return DM_MAPIO_REMAPPED;
1585    }
1586
1587    chunk = sector_to_chunk(s->store, bio->bi_sector);
1588
1589    /* Full snapshots are not usable */
1590    /* To get here the table must be live so s->active is always set. */
1591    if (!s->valid)
1592        return -EIO;
1593
1594    /* FIXME: should only take write lock if we need
1595     * to copy an exception */
1596    down_write(&s->lock);
1597
1598    if (!s->valid) {
1599        r = -EIO;
1600        goto out_unlock;
1601    }
1602
1603    /* If the block is already remapped - use that, else remap it */
1604    e = dm_lookup_exception(&s->complete, chunk);
1605    if (e) {
1606        remap_exception(s, e, bio, chunk);
1607        goto out_unlock;
1608    }
1609
1610    /*
1611     * Write to snapshot - higher level takes care of RW/RO
1612     * flags so we should only get this if we are
1613     * writeable.
1614     */
1615    if (bio_rw(bio) == WRITE) {
1616        pe = __lookup_pending_exception(s, chunk);
1617        if (!pe) {
1618            up_write(&s->lock);
1619            pe = alloc_pending_exception(s);
1620            down_write(&s->lock);
1621
1622            if (!s->valid) {
1623                free_pending_exception(pe);
1624                r = -EIO;
1625                goto out_unlock;
1626            }
1627
1628            e = dm_lookup_exception(&s->complete, chunk);
1629            if (e) {
1630                free_pending_exception(pe);
1631                remap_exception(s, e, bio, chunk);
1632                goto out_unlock;
1633            }
1634
1635            pe = __find_pending_exception(s, pe, chunk);
1636            if (!pe) {
1637                __invalidate_snapshot(s, -ENOMEM);
1638                r = -EIO;
1639                goto out_unlock;
1640            }
1641        }
1642
1643        remap_exception(s, &pe->e, bio, chunk);
1644
1645        r = DM_MAPIO_SUBMITTED;
1646
1647        if (!pe->started &&
1648            bio->bi_size == (s->store->chunk_size << SECTOR_SHIFT)) {
1649            pe->started = 1;
1650            up_write(&s->lock);
1651            start_full_bio(pe, bio);
1652            goto out;
1653        }
1654
1655        bio_list_add(&pe->snapshot_bios, bio);
1656
1657        if (!pe->started) {
1658            /* this is protected by snap->lock */
1659            pe->started = 1;
1660            up_write(&s->lock);
1661            start_copy(pe);
1662            goto out;
1663        }
1664    } else {
1665        bio->bi_bdev = s->origin->bdev;
1666        track_chunk(s, bio, chunk);
1667    }
1668
1669out_unlock:
1670    up_write(&s->lock);
1671out:
1672    return r;
1673}
1674
1675/*
1676 * A snapshot-merge target behaves like a combination of a snapshot
1677 * target and a snapshot-origin target. It only generates new
1678 * exceptions in other snapshots and not in the one that is being
1679 * merged.
1680 *
1681 * For each chunk, if there is an existing exception, it is used to
1682 * redirect I/O to the cow device. Otherwise I/O is sent to the origin,
1683 * which in turn might generate exceptions in other snapshots.
1684 * If merging is currently taking place on the chunk in question, the
1685 * I/O is deferred by adding it to s->bios_queued_during_merge.
1686 */
1687static int snapshot_merge_map(struct dm_target *ti, struct bio *bio)
1688{
1689    struct dm_exception *e;
1690    struct dm_snapshot *s = ti->private;
1691    int r = DM_MAPIO_REMAPPED;
1692    chunk_t chunk;
1693
1694    init_tracked_chunk(bio);
1695
1696    if (bio->bi_rw & REQ_FLUSH) {
1697        if (!dm_bio_get_target_bio_nr(bio))
1698            bio->bi_bdev = s->origin->bdev;
1699        else
1700            bio->bi_bdev = s->cow->bdev;
1701        return DM_MAPIO_REMAPPED;
1702    }
1703
1704    chunk = sector_to_chunk(s->store, bio->bi_sector);
1705
1706    down_write(&s->lock);
1707
1708    /* Full merging snapshots are redirected to the origin */
1709    if (!s->valid)
1710        goto redirect_to_origin;
1711
1712    /* If the block is already remapped - use that */
1713    e = dm_lookup_exception(&s->complete, chunk);
1714    if (e) {
1715        /* Queue writes overlapping with chunks being merged */
1716        if (bio_rw(bio) == WRITE &&
1717            chunk >= s->first_merging_chunk &&
1718            chunk < (s->first_merging_chunk +
1719                 s->num_merging_chunks)) {
1720            bio->bi_bdev = s->origin->bdev;
1721            bio_list_add(&s->bios_queued_during_merge, bio);
1722            r = DM_MAPIO_SUBMITTED;
1723            goto out_unlock;
1724        }
1725
1726        remap_exception(s, e, bio, chunk);
1727
1728        if (bio_rw(bio) == WRITE)
1729            track_chunk(s, bio, chunk);
1730        goto out_unlock;
1731    }
1732
1733redirect_to_origin:
1734    bio->bi_bdev = s->origin->bdev;
1735
1736    if (bio_rw(bio) == WRITE) {
1737        up_write(&s->lock);
1738        return do_origin(s->origin, bio);
1739    }
1740
1741out_unlock:
1742    up_write(&s->lock);
1743
1744    return r;
1745}
1746
1747static int snapshot_end_io(struct dm_target *ti, struct bio *bio, int error)
1748{
1749    struct dm_snapshot *s = ti->private;
1750
1751    if (is_bio_tracked(bio))
1752        stop_tracking_chunk(s, bio);
1753
1754    return 0;
1755}
1756
1757static void snapshot_merge_presuspend(struct dm_target *ti)
1758{
1759    struct dm_snapshot *s = ti->private;
1760
1761    stop_merge(s);
1762}
1763
1764static int snapshot_preresume(struct dm_target *ti)
1765{
1766    int r = 0;
1767    struct dm_snapshot *s = ti->private;
1768    struct dm_snapshot *snap_src = NULL, *snap_dest = NULL;
1769
1770    down_read(&_origins_lock);
1771    (void) __find_snapshots_sharing_cow(s, &snap_src, &snap_dest, NULL);
1772    if (snap_src && snap_dest) {
1773        down_read(&snap_src->lock);
1774        if (s == snap_src) {
1775            DMERR("Unable to resume snapshot source until "
1776                  "handover completes.");
1777            r = -EINVAL;
1778        } else if (!dm_suspended(snap_src->ti)) {
1779            DMERR("Unable to perform snapshot handover until "
1780                  "source is suspended.");
1781            r = -EINVAL;
1782        }
1783        up_read(&snap_src->lock);
1784    }
1785    up_read(&_origins_lock);
1786
1787    return r;
1788}
1789
1790static void snapshot_resume(struct dm_target *ti)
1791{
1792    struct dm_snapshot *s = ti->private;
1793    struct dm_snapshot *snap_src = NULL, *snap_dest = NULL;
1794
1795    down_read(&_origins_lock);
1796    (void) __find_snapshots_sharing_cow(s, &snap_src, &snap_dest, NULL);
1797    if (snap_src && snap_dest) {
1798        down_write(&snap_src->lock);
1799        down_write_nested(&snap_dest->lock, SINGLE_DEPTH_NESTING);
1800        __handover_exceptions(snap_src, snap_dest);
1801        up_write(&snap_dest->lock);
1802        up_write(&snap_src->lock);
1803    }
1804    up_read(&_origins_lock);
1805
1806    /* Now we have correct chunk size, reregister */
1807    reregister_snapshot(s);
1808
1809    down_write(&s->lock);
1810    s->active = 1;
1811    up_write(&s->lock);
1812}
1813
1814static uint32_t get_origin_minimum_chunksize(struct block_device *bdev)
1815{
1816    uint32_t min_chunksize;
1817
1818    down_read(&_origins_lock);
1819    min_chunksize = __minimum_chunk_size(__lookup_origin(bdev));
1820    up_read(&_origins_lock);
1821
1822    return min_chunksize;
1823}
1824
1825static void snapshot_merge_resume(struct dm_target *ti)
1826{
1827    struct dm_snapshot *s = ti->private;
1828
1829    /*
1830     * Handover exceptions from existing snapshot.
1831     */
1832    snapshot_resume(ti);
1833
1834    /*
1835     * snapshot-merge acts as an origin, so set ti->max_io_len
1836     */
1837    ti->max_io_len = get_origin_minimum_chunksize(s->origin->bdev);
1838
1839    start_merge(s);
1840}
1841
1842static void snapshot_status(struct dm_target *ti, status_type_t type,
1843                unsigned status_flags, char *result, unsigned maxlen)
1844{
1845    unsigned sz = 0;
1846    struct dm_snapshot *snap = ti->private;
1847
1848    switch (type) {
1849    case STATUSTYPE_INFO:
1850
1851        down_write(&snap->lock);
1852
1853        if (!snap->valid)
1854            DMEMIT("Invalid");
1855        else if (snap->merge_failed)
1856            DMEMIT("Merge failed");
1857        else {
1858            if (snap->store->type->usage) {
1859                sector_t total_sectors, sectors_allocated,
1860                     metadata_sectors;
1861                snap->store->type->usage(snap->store,
1862                             &total_sectors,
1863                             &sectors_allocated,
1864                             &metadata_sectors);
1865                DMEMIT("%llu/%llu %llu",
1866                       (unsigned long long)sectors_allocated,
1867                       (unsigned long long)total_sectors,
1868                       (unsigned long long)metadata_sectors);
1869            }
1870            else
1871                DMEMIT("Unknown");
1872        }
1873
1874        up_write(&snap->lock);
1875
1876        break;
1877
1878    case STATUSTYPE_TABLE:
1879        /*
1880         * kdevname returns a static pointer so we need
1881         * to make private copies if the output is to
1882         * make sense.
1883         */
1884        DMEMIT("%s %s", snap->origin->name, snap->cow->name);
1885        snap->store->type->status(snap->store, type, result + sz,
1886                      maxlen - sz);
1887        break;
1888    }
1889}
1890
1891static int snapshot_iterate_devices(struct dm_target *ti,
1892                    iterate_devices_callout_fn fn, void *data)
1893{
1894    struct dm_snapshot *snap = ti->private;
1895    int r;
1896
1897    r = fn(ti, snap->origin, 0, ti->len, data);
1898
1899    if (!r)
1900        r = fn(ti, snap->cow, 0, get_dev_size(snap->cow->bdev), data);
1901
1902    return r;
1903}
1904
1905
1906/*-----------------------------------------------------------------
1907 * Origin methods
1908 *---------------------------------------------------------------*/
1909
1910/*
1911 * If no exceptions need creating, DM_MAPIO_REMAPPED is returned and any
1912 * supplied bio was ignored. The caller may submit it immediately.
1913 * (No remapping actually occurs as the origin is always a direct linear
1914 * map.)
1915 *
1916 * If further exceptions are required, DM_MAPIO_SUBMITTED is returned
1917 * and any supplied bio is added to a list to be submitted once all
1918 * the necessary exceptions exist.
1919 */
1920static int __origin_write(struct list_head *snapshots, sector_t sector,
1921              struct bio *bio)
1922{
1923    int r = DM_MAPIO_REMAPPED;
1924    struct dm_snapshot *snap;
1925    struct dm_exception *e;
1926    struct dm_snap_pending_exception *pe;
1927    struct dm_snap_pending_exception *pe_to_start_now = NULL;
1928    struct dm_snap_pending_exception *pe_to_start_last = NULL;
1929    chunk_t chunk;
1930
1931    /* Do all the snapshots on this origin */
1932    list_for_each_entry (snap, snapshots, list) {
1933        /*
1934         * Don't make new exceptions in a merging snapshot
1935         * because it has effectively been deleted
1936         */
1937        if (dm_target_is_snapshot_merge(snap->ti))
1938            continue;
1939
1940        down_write(&snap->lock);
1941
1942        /* Only deal with valid and active snapshots */
1943        if (!snap->valid || !snap->active)
1944            goto next_snapshot;
1945
1946        /* Nothing to do if writing beyond end of snapshot */
1947        if (sector >= dm_table_get_size(snap->ti->table))
1948            goto next_snapshot;
1949
1950        /*
1951         * Remember, different snapshots can have
1952         * different chunk sizes.
1953         */
1954        chunk = sector_to_chunk(snap->store, sector);
1955
1956        /*
1957         * Check exception table to see if block
1958         * is already remapped in this snapshot
1959         * and trigger an exception if not.
1960         */
1961        e = dm_lookup_exception(&snap->complete, chunk);
1962        if (e)
1963            goto next_snapshot;
1964
1965        pe = __lookup_pending_exception(snap, chunk);
1966        if (!pe) {
1967            up_write(&snap->lock);
1968            pe = alloc_pending_exception(snap);
1969            down_write(&snap->lock);
1970
1971            if (!snap->valid) {
1972                free_pending_exception(pe);
1973                goto next_snapshot;
1974            }
1975
1976            e = dm_lookup_exception(&snap->complete, chunk);
1977            if (e) {
1978                free_pending_exception(pe);
1979                goto next_snapshot;
1980            }
1981
1982            pe = __find_pending_exception(snap, pe, chunk);
1983            if (!pe) {
1984                __invalidate_snapshot(snap, -ENOMEM);
1985                goto next_snapshot;
1986            }
1987        }
1988
1989        r = DM_MAPIO_SUBMITTED;
1990
1991        /*
1992         * If an origin bio was supplied, queue it to wait for the
1993         * completion of this exception, and start this one last,
1994         * at the end of the function.
1995         */
1996        if (bio) {
1997            bio_list_add(&pe->origin_bios, bio);
1998            bio = NULL;
1999
2000            if (!pe->started) {
2001                pe->started = 1;
2002                pe_to_start_last = pe;
2003            }
2004        }
2005
2006        if (!pe->started) {
2007            pe->started = 1;
2008            pe_to_start_now = pe;
2009        }
2010
2011next_snapshot:
2012        up_write(&snap->lock);
2013
2014        if (pe_to_start_now) {
2015            start_copy(pe_to_start_now);
2016            pe_to_start_now = NULL;
2017        }
2018    }
2019
2020    /*
2021     * Submit the exception against which the bio is queued last,
2022     * to give the other exceptions a head start.
2023     */
2024    if (pe_to_start_last)
2025        start_copy(pe_to_start_last);
2026
2027    return r;
2028}
2029
2030/*
2031 * Called on a write from the origin driver.
2032 */
2033static int do_origin(struct dm_dev *origin, struct bio *bio)
2034{
2035    struct origin *o;
2036    int r = DM_MAPIO_REMAPPED;
2037
2038    down_read(&_origins_lock);
2039    o = __lookup_origin(origin->bdev);
2040    if (o)
2041        r = __origin_write(&o->snapshots, bio->bi_sector, bio);
2042    up_read(&_origins_lock);
2043
2044    return r;
2045}
2046
2047/*
2048 * Trigger exceptions in all non-merging snapshots.
2049 *
2050 * The chunk size of the merging snapshot may be larger than the chunk
2051 * size of some other snapshot so we may need to reallocate multiple
2052 * chunks in other snapshots.
2053 *
2054 * We scan all the overlapping exceptions in the other snapshots.
2055 * Returns 1 if anything was reallocated and must be waited for,
2056 * otherwise returns 0.
2057 *
2058 * size must be a multiple of merging_snap's chunk_size.
2059 */
2060static int origin_write_extent(struct dm_snapshot *merging_snap,
2061                   sector_t sector, unsigned size)
2062{
2063    int must_wait = 0;
2064    sector_t n;
2065    struct origin *o;
2066
2067    /*
2068     * The origin's __minimum_chunk_size() got stored in max_io_len
2069     * by snapshot_merge_resume().
2070     */
2071    down_read(&_origins_lock);
2072    o = __lookup_origin(merging_snap->origin->bdev);
2073    for (n = 0; n < size; n += merging_snap->ti->max_io_len)
2074        if (__origin_write(&o->snapshots, sector + n, NULL) ==
2075            DM_MAPIO_SUBMITTED)
2076            must_wait = 1;
2077    up_read(&_origins_lock);
2078
2079    return must_wait;
2080}
2081
2082/*
2083 * Origin: maps a linear range of a device, with hooks for snapshotting.
2084 */
2085
2086/*
2087 * Construct an origin mapping: <dev_path>
2088 * The context for an origin is merely a 'struct dm_dev *'
2089 * pointing to the real device.
2090 */
2091static int origin_ctr(struct dm_target *ti, unsigned int argc, char **argv)
2092{
2093    int r;
2094    struct dm_dev *dev;
2095
2096    if (argc != 1) {
2097        ti->error = "origin: incorrect number of arguments";
2098        return -EINVAL;
2099    }
2100
2101    r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &dev);
2102    if (r) {
2103        ti->error = "Cannot get target device";
2104        return r;
2105    }
2106
2107    ti->private = dev;
2108    ti->num_flush_bios = 1;
2109
2110    return 0;
2111}
2112
2113static void origin_dtr(struct dm_target *ti)
2114{
2115    struct dm_dev *dev = ti->private;
2116    dm_put_device(ti, dev);
2117}
2118
2119static int origin_map(struct dm_target *ti, struct bio *bio)
2120{
2121    struct dm_dev *dev = ti->private;
2122    bio->bi_bdev = dev->bdev;
2123
2124    if (bio->bi_rw & REQ_FLUSH)
2125        return DM_MAPIO_REMAPPED;
2126
2127    /* Only tell snapshots if this is a write */
2128    return (bio_rw(bio) == WRITE) ? do_origin(dev, bio) : DM_MAPIO_REMAPPED;
2129}
2130
2131/*
2132 * Set the target "max_io_len" field to the minimum of all the snapshots'
2133 * chunk sizes.
2134 */
2135static void origin_resume(struct dm_target *ti)
2136{
2137    struct dm_dev *dev = ti->private;
2138
2139    ti->max_io_len = get_origin_minimum_chunksize(dev->bdev);
2140}
2141
2142static void origin_status(struct dm_target *ti, status_type_t type,
2143              unsigned status_flags, char *result, unsigned maxlen)
2144{
2145    struct dm_dev *dev = ti->private;
2146
2147    switch (type) {
2148    case STATUSTYPE_INFO:
2149        result[0] = '\0';
2150        break;
2151
2152    case STATUSTYPE_TABLE:
2153        snprintf(result, maxlen, "%s", dev->name);
2154        break;
2155    }
2156}
2157
2158static int origin_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
2159            struct bio_vec *biovec, int max_size)
2160{
2161    struct dm_dev *dev = ti->private;
2162    struct request_queue *q = bdev_get_queue(dev->bdev);
2163
2164    if (!q->merge_bvec_fn)
2165        return max_size;
2166
2167    bvm->bi_bdev = dev->bdev;
2168
2169    return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
2170}
2171
2172static int origin_iterate_devices(struct dm_target *ti,
2173                  iterate_devices_callout_fn fn, void *data)
2174{
2175    struct dm_dev *dev = ti->private;
2176
2177    return fn(ti, dev, 0, ti->len, data);
2178}
2179
2180static struct target_type origin_target = {
2181    .name = "snapshot-origin",
2182    .version = {1, 8, 1},
2183    .module = THIS_MODULE,
2184    .ctr = origin_ctr,
2185    .dtr = origin_dtr,
2186    .map = origin_map,
2187    .resume = origin_resume,
2188    .status = origin_status,
2189    .merge = origin_merge,
2190    .iterate_devices = origin_iterate_devices,
2191};
2192
2193static struct target_type snapshot_target = {
2194    .name = "snapshot",
2195    .version = {1, 11, 1},
2196    .module = THIS_MODULE,
2197    .ctr = snapshot_ctr,
2198    .dtr = snapshot_dtr,
2199    .map = snapshot_map,
2200    .end_io = snapshot_end_io,
2201    .preresume = snapshot_preresume,
2202    .resume = snapshot_resume,
2203    .status = snapshot_status,
2204    .iterate_devices = snapshot_iterate_devices,
2205};
2206
2207static struct target_type merge_target = {
2208    .name = dm_snapshot_merge_target_name,
2209    .version = {1, 2, 0},
2210    .module = THIS_MODULE,
2211    .ctr = snapshot_ctr,
2212    .dtr = snapshot_dtr,
2213    .map = snapshot_merge_map,
2214    .end_io = snapshot_end_io,
2215    .presuspend = snapshot_merge_presuspend,
2216    .preresume = snapshot_preresume,
2217    .resume = snapshot_merge_resume,
2218    .status = snapshot_status,
2219    .iterate_devices = snapshot_iterate_devices,
2220};
2221
2222static int __init dm_snapshot_init(void)
2223{
2224    int r;
2225
2226    r = dm_exception_store_init();
2227    if (r) {
2228        DMERR("Failed to initialize exception stores");
2229        return r;
2230    }
2231
2232    r = dm_register_target(&snapshot_target);
2233    if (r < 0) {
2234        DMERR("snapshot target register failed %d", r);
2235        goto bad_register_snapshot_target;
2236    }
2237
2238    r = dm_register_target(&origin_target);
2239    if (r < 0) {
2240        DMERR("Origin target register failed %d", r);
2241        goto bad_register_origin_target;
2242    }
2243
2244    r = dm_register_target(&merge_target);
2245    if (r < 0) {
2246        DMERR("Merge target register failed %d", r);
2247        goto bad_register_merge_target;
2248    }
2249
2250    r = init_origin_hash();
2251    if (r) {
2252        DMERR("init_origin_hash failed.");
2253        goto bad_origin_hash;
2254    }
2255
2256    exception_cache = KMEM_CACHE(dm_exception, 0);
2257    if (!exception_cache) {
2258        DMERR("Couldn't create exception cache.");
2259        r = -ENOMEM;
2260        goto bad_exception_cache;
2261    }
2262
2263    pending_cache = KMEM_CACHE(dm_snap_pending_exception, 0);
2264    if (!pending_cache) {
2265        DMERR("Couldn't create pending cache.");
2266        r = -ENOMEM;
2267        goto bad_pending_cache;
2268    }
2269
2270    return 0;
2271
2272bad_pending_cache:
2273    kmem_cache_destroy(exception_cache);
2274bad_exception_cache:
2275    exit_origin_hash();
2276bad_origin_hash:
2277    dm_unregister_target(&merge_target);
2278bad_register_merge_target:
2279    dm_unregister_target(&origin_target);
2280bad_register_origin_target:
2281    dm_unregister_target(&snapshot_target);
2282bad_register_snapshot_target:
2283    dm_exception_store_exit();
2284
2285    return r;
2286}
2287
2288static void __exit dm_snapshot_exit(void)
2289{
2290    dm_unregister_target(&snapshot_target);
2291    dm_unregister_target(&origin_target);
2292    dm_unregister_target(&merge_target);
2293
2294    exit_origin_hash();
2295    kmem_cache_destroy(pending_cache);
2296    kmem_cache_destroy(exception_cache);
2297
2298    dm_exception_store_exit();
2299}
2300
2301/* Module hooks */
2302module_init(dm_snapshot_init);
2303module_exit(dm_snapshot_exit);
2304
2305MODULE_DESCRIPTION(DM_NAME " snapshot target");
2306MODULE_AUTHOR("Joe Thornber");
2307MODULE_LICENSE("GPL");
2308MODULE_ALIAS("dm-snapshot-origin");
2309MODULE_ALIAS("dm-snapshot-merge");
2310

Archive Download this file



interactive