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