1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Copyright (C) 2001-2002 Sistina Software (UK) Limited. 4 * 5 * This file is released under the GPL. 6 */ 7 8 #include <linux/blkdev.h> 9 #include <linux/device-mapper.h> 10 #include <linux/delay.h> 11 #include <linux/fs.h> 12 #include <linux/init.h> 13 #include <linux/kdev_t.h> 14 #include <linux/list.h> 15 #include <linux/list_bl.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.h" 24 25 #include "dm-exception-store.h" 26 27 #define DM_MSG_PREFIX "snapshots" 28 29 static const char dm_snapshot_merge_target_name[] = "snapshot-merge"; 30 31 #define dm_target_is_snapshot_merge(ti) \ 32 ((ti)->type->name == dm_snapshot_merge_target_name) 33 34 /* 35 * The size of the mempool used to track chunks in use. 36 */ 37 #define MIN_IOS 256 38 39 #define DM_TRACKED_CHUNK_HASH_SIZE 16 40 #define DM_TRACKED_CHUNK_HASH(x) ((unsigned long)(x) & \ 41 (DM_TRACKED_CHUNK_HASH_SIZE - 1)) 42 43 struct dm_exception_table { 44 uint32_t hash_mask; 45 unsigned int hash_shift; 46 struct hlist_bl_head *table; 47 }; 48 49 struct dm_snapshot { 50 struct rw_semaphore lock; 51 52 struct dm_dev *origin; 53 struct dm_dev *cow; 54 55 struct dm_target *ti; 56 57 /* List of snapshots per Origin */ 58 struct list_head list; 59 60 /* 61 * You can't use a snapshot if this is 0 (e.g. if full). 62 * A snapshot-merge target never clears this. 63 */ 64 int valid; 65 66 /* 67 * The snapshot overflowed because of a write to the snapshot device. 68 * We don't have to invalidate the snapshot in this case, but we need 69 * to prevent further writes. 70 */ 71 int snapshot_overflowed; 72 73 /* Origin writes don't trigger exceptions until this is set */ 74 int active; 75 76 atomic_t pending_exceptions_count; 77 78 spinlock_t pe_allocation_lock; 79 80 /* Protected by "pe_allocation_lock" */ 81 sector_t exception_start_sequence; 82 83 /* Protected by kcopyd single-threaded callback */ 84 sector_t exception_complete_sequence; 85 86 /* 87 * A list of pending exceptions that completed out of order. 88 * Protected by kcopyd single-threaded callback. 89 */ 90 struct rb_root out_of_order_tree; 91 92 mempool_t pending_pool; 93 94 struct dm_exception_table pending; 95 struct dm_exception_table complete; 96 97 /* 98 * pe_lock protects all pending_exception operations and access 99 * as well as the snapshot_bios list. 100 */ 101 spinlock_t pe_lock; 102 103 /* Chunks with outstanding reads */ 104 spinlock_t tracked_chunk_lock; 105 struct hlist_head tracked_chunk_hash[DM_TRACKED_CHUNK_HASH_SIZE]; 106 107 /* The on disk metadata handler */ 108 struct dm_exception_store *store; 109 110 unsigned int in_progress; 111 struct wait_queue_head in_progress_wait; 112 113 struct dm_kcopyd_client *kcopyd_client; 114 115 /* Wait for events based on state_bits */ 116 unsigned long state_bits; 117 118 /* Range of chunks currently being merged. */ 119 chunk_t first_merging_chunk; 120 int num_merging_chunks; 121 122 /* 123 * The merge operation failed if this flag is set. 124 * Failure modes are handled as follows: 125 * - I/O error reading the header 126 * => don't load the target; abort. 127 * - Header does not have "valid" flag set 128 * => use the origin; forget about the snapshot. 129 * - I/O error when reading exceptions 130 * => don't load the target; abort. 131 * (We can't use the intermediate origin state.) 132 * - I/O error while merging 133 * => stop merging; set merge_failed; process I/O normally. 134 */ 135 bool merge_failed:1; 136 137 bool discard_zeroes_cow:1; 138 bool discard_passdown_origin:1; 139 140 /* 141 * Incoming bios that overlap with chunks being merged must wait 142 * for them to be committed. 143 */ 144 struct bio_list bios_queued_during_merge; 145 }; 146 147 /* 148 * state_bits: 149 * RUNNING_MERGE - Merge operation is in progress. 150 * SHUTDOWN_MERGE - Set to signal that merge needs to be stopped; 151 * cleared afterwards. 152 */ 153 #define RUNNING_MERGE 0 154 #define SHUTDOWN_MERGE 1 155 156 /* 157 * Maximum number of chunks being copied on write. 158 * 159 * The value was decided experimentally as a trade-off between memory 160 * consumption, stalling the kernel's workqueues and maintaining a high enough 161 * throughput. 162 */ 163 #define DEFAULT_COW_THRESHOLD 2048 164 165 static unsigned int cow_threshold = DEFAULT_COW_THRESHOLD; 166 module_param_named(snapshot_cow_threshold, cow_threshold, uint, 0644); 167 MODULE_PARM_DESC(snapshot_cow_threshold, "Maximum number of chunks being copied on write"); 168 169 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(snapshot_copy_throttle, 170 "A percentage of time allocated for copy on write"); 171 172 struct dm_dev *dm_snap_origin(struct dm_snapshot *s) 173 { 174 return s->origin; 175 } 176 EXPORT_SYMBOL(dm_snap_origin); 177 178 struct dm_dev *dm_snap_cow(struct dm_snapshot *s) 179 { 180 return s->cow; 181 } 182 EXPORT_SYMBOL(dm_snap_cow); 183 184 static sector_t chunk_to_sector(struct dm_exception_store *store, 185 chunk_t chunk) 186 { 187 return chunk << store->chunk_shift; 188 } 189 190 static int bdev_equal(struct block_device *lhs, struct block_device *rhs) 191 { 192 /* 193 * There is only ever one instance of a particular block 194 * device so we can compare pointers safely. 195 */ 196 return lhs == rhs; 197 } 198 199 struct dm_snap_pending_exception { 200 struct dm_exception e; 201 202 /* 203 * Origin buffers waiting for this to complete are held 204 * in a bio list 205 */ 206 struct bio_list origin_bios; 207 struct bio_list snapshot_bios; 208 209 /* Pointer back to snapshot context */ 210 struct dm_snapshot *snap; 211 212 /* 213 * 1 indicates the exception has already been sent to 214 * kcopyd. 215 */ 216 int started; 217 218 /* There was copying error. */ 219 int copy_error; 220 221 /* A sequence number, it is used for in-order completion. */ 222 sector_t exception_sequence; 223 224 struct rb_node out_of_order_node; 225 226 /* 227 * For writing a complete chunk, bypassing the copy. 228 */ 229 struct bio *full_bio; 230 bio_end_io_t *full_bio_end_io; 231 }; 232 233 /* 234 * Hash table mapping origin volumes to lists of snapshots and 235 * a lock to protect it 236 */ 237 static struct kmem_cache *exception_cache; 238 static struct kmem_cache *pending_cache; 239 240 struct dm_snap_tracked_chunk { 241 struct hlist_node node; 242 chunk_t chunk; 243 }; 244 245 static void init_tracked_chunk(struct bio *bio) 246 { 247 struct dm_snap_tracked_chunk *c = dm_per_bio_data(bio, sizeof(struct dm_snap_tracked_chunk)); 248 249 INIT_HLIST_NODE(&c->node); 250 } 251 252 static bool is_bio_tracked(struct bio *bio) 253 { 254 struct dm_snap_tracked_chunk *c = dm_per_bio_data(bio, sizeof(struct dm_snap_tracked_chunk)); 255 256 return !hlist_unhashed(&c->node); 257 } 258 259 static void track_chunk(struct dm_snapshot *s, struct bio *bio, chunk_t chunk) 260 { 261 struct dm_snap_tracked_chunk *c = dm_per_bio_data(bio, sizeof(struct dm_snap_tracked_chunk)); 262 263 c->chunk = chunk; 264 265 spin_lock_irq(&s->tracked_chunk_lock); 266 hlist_add_head(&c->node, 267 &s->tracked_chunk_hash[DM_TRACKED_CHUNK_HASH(chunk)]); 268 spin_unlock_irq(&s->tracked_chunk_lock); 269 } 270 271 static void stop_tracking_chunk(struct dm_snapshot *s, struct bio *bio) 272 { 273 struct dm_snap_tracked_chunk *c = dm_per_bio_data(bio, sizeof(struct dm_snap_tracked_chunk)); 274 unsigned long flags; 275 276 spin_lock_irqsave(&s->tracked_chunk_lock, flags); 277 hlist_del(&c->node); 278 spin_unlock_irqrestore(&s->tracked_chunk_lock, flags); 279 } 280 281 static int __chunk_is_tracked(struct dm_snapshot *s, chunk_t chunk) 282 { 283 struct dm_snap_tracked_chunk *c; 284 int found = 0; 285 286 spin_lock_irq(&s->tracked_chunk_lock); 287 288 hlist_for_each_entry(c, 289 &s->tracked_chunk_hash[DM_TRACKED_CHUNK_HASH(chunk)], node) { 290 if (c->chunk == chunk) { 291 found = 1; 292 break; 293 } 294 } 295 296 spin_unlock_irq(&s->tracked_chunk_lock); 297 298 return found; 299 } 300 301 /* 302 * This conflicting I/O is extremely improbable in the caller, 303 * so fsleep(1000) is sufficient and there is no need for a wait queue. 304 */ 305 static void __check_for_conflicting_io(struct dm_snapshot *s, chunk_t chunk) 306 { 307 while (__chunk_is_tracked(s, chunk)) 308 fsleep(1000); 309 } 310 311 /* 312 * One of these per registered origin, held in the snapshot_origins hash 313 */ 314 struct origin { 315 /* The origin device */ 316 struct block_device *bdev; 317 318 struct list_head hash_list; 319 320 /* List of snapshots for this origin */ 321 struct list_head snapshots; 322 }; 323 324 /* 325 * This structure is allocated for each origin target 326 */ 327 struct dm_origin { 328 struct dm_dev *dev; 329 struct dm_target *ti; 330 unsigned int split_boundary; 331 struct list_head hash_list; 332 }; 333 334 /* 335 * Size of the hash table for origin volumes. If we make this 336 * the size of the minors list then it should be nearly perfect 337 */ 338 #define ORIGIN_HASH_SIZE 256 339 #define ORIGIN_MASK 0xFF 340 static struct list_head *_origins; 341 static struct list_head *_dm_origins; 342 static struct rw_semaphore _origins_lock; 343 344 static DECLARE_WAIT_QUEUE_HEAD(_pending_exceptions_done); 345 static DEFINE_SPINLOCK(_pending_exceptions_done_spinlock); 346 static uint64_t _pending_exceptions_done_count; 347 348 static int init_origin_hash(void) 349 { 350 int i; 351 352 _origins = kmalloc_array(ORIGIN_HASH_SIZE, sizeof(struct list_head), 353 GFP_KERNEL); 354 if (!_origins) { 355 DMERR("unable to allocate memory for _origins"); 356 return -ENOMEM; 357 } 358 for (i = 0; i < ORIGIN_HASH_SIZE; i++) 359 INIT_LIST_HEAD(_origins + i); 360 361 _dm_origins = kmalloc_array(ORIGIN_HASH_SIZE, 362 sizeof(struct list_head), 363 GFP_KERNEL); 364 if (!_dm_origins) { 365 DMERR("unable to allocate memory for _dm_origins"); 366 kfree(_origins); 367 return -ENOMEM; 368 } 369 for (i = 0; i < ORIGIN_HASH_SIZE; i++) 370 INIT_LIST_HEAD(_dm_origins + i); 371 372 init_rwsem(&_origins_lock); 373 374 return 0; 375 } 376 377 static void exit_origin_hash(void) 378 { 379 kfree(_origins); 380 kfree(_dm_origins); 381 } 382 383 static unsigned int origin_hash(struct block_device *bdev) 384 { 385 return bdev->bd_dev & ORIGIN_MASK; 386 } 387 388 static struct origin *__lookup_origin(struct block_device *origin) 389 { 390 struct list_head *ol; 391 struct origin *o; 392 393 ol = &_origins[origin_hash(origin)]; 394 list_for_each_entry(o, ol, hash_list) 395 if (bdev_equal(o->bdev, origin)) 396 return o; 397 398 return NULL; 399 } 400 401 static void __insert_origin(struct origin *o) 402 { 403 struct list_head *sl = &_origins[origin_hash(o->bdev)]; 404 405 list_add_tail(&o->hash_list, sl); 406 } 407 408 static struct dm_origin *__lookup_dm_origin(struct block_device *origin) 409 { 410 struct list_head *ol; 411 struct dm_origin *o; 412 413 ol = &_dm_origins[origin_hash(origin)]; 414 list_for_each_entry(o, ol, hash_list) 415 if (bdev_equal(o->dev->bdev, origin)) 416 return o; 417 418 return NULL; 419 } 420 421 static void __insert_dm_origin(struct dm_origin *o) 422 { 423 struct list_head *sl = &_dm_origins[origin_hash(o->dev->bdev)]; 424 425 list_add_tail(&o->hash_list, sl); 426 } 427 428 static void __remove_dm_origin(struct dm_origin *o) 429 { 430 list_del(&o->hash_list); 431 } 432 433 /* 434 * _origins_lock must be held when calling this function. 435 * Returns number of snapshots registered using the supplied cow device, plus: 436 * snap_src - a snapshot suitable for use as a source of exception handover 437 * snap_dest - a snapshot capable of receiving exception handover. 438 * snap_merge - an existing snapshot-merge target linked to the same origin. 439 * There can be at most one snapshot-merge target. The parameter is optional. 440 * 441 * Possible return values and states of snap_src and snap_dest. 442 * 0: NULL, NULL - first new snapshot 443 * 1: snap_src, NULL - normal snapshot 444 * 2: snap_src, snap_dest - waiting for handover 445 * 2: snap_src, NULL - handed over, waiting for old to be deleted 446 * 1: NULL, snap_dest - source got destroyed without handover 447 */ 448 static int __find_snapshots_sharing_cow(struct dm_snapshot *snap, 449 struct dm_snapshot **snap_src, 450 struct dm_snapshot **snap_dest, 451 struct dm_snapshot **snap_merge) 452 { 453 struct dm_snapshot *s; 454 struct origin *o; 455 int count = 0; 456 int active; 457 458 o = __lookup_origin(snap->origin->bdev); 459 if (!o) 460 goto out; 461 462 list_for_each_entry(s, &o->snapshots, list) { 463 if (dm_target_is_snapshot_merge(s->ti) && snap_merge) 464 *snap_merge = s; 465 if (!bdev_equal(s->cow->bdev, snap->cow->bdev)) 466 continue; 467 468 down_read(&s->lock); 469 active = s->active; 470 up_read(&s->lock); 471 472 if (active) { 473 if (snap_src) 474 *snap_src = s; 475 } else if (snap_dest) 476 *snap_dest = s; 477 478 count++; 479 } 480 481 out: 482 return count; 483 } 484 485 /* 486 * On success, returns 1 if this snapshot is a handover destination, 487 * otherwise returns 0. 488 */ 489 static int __validate_exception_handover(struct dm_snapshot *snap) 490 { 491 struct dm_snapshot *snap_src = NULL, *snap_dest = NULL; 492 struct dm_snapshot *snap_merge = NULL; 493 494 /* Does snapshot need exceptions handed over to it? */ 495 if ((__find_snapshots_sharing_cow(snap, &snap_src, &snap_dest, 496 &snap_merge) == 2) || 497 snap_dest) { 498 snap->ti->error = "Snapshot cow pairing for exception " 499 "table handover failed"; 500 return -EINVAL; 501 } 502 503 /* 504 * If no snap_src was found, snap cannot become a handover 505 * destination. 506 */ 507 if (!snap_src) 508 return 0; 509 510 /* 511 * Non-snapshot-merge handover? 512 */ 513 if (!dm_target_is_snapshot_merge(snap->ti)) 514 return 1; 515 516 /* 517 * Do not allow more than one merging snapshot. 518 */ 519 if (snap_merge) { 520 snap->ti->error = "A snapshot is already merging."; 521 return -EINVAL; 522 } 523 524 if (!snap_src->store->type->prepare_merge || 525 !snap_src->store->type->commit_merge) { 526 snap->ti->error = "Snapshot exception store does not " 527 "support snapshot-merge."; 528 return -EINVAL; 529 } 530 531 return 1; 532 } 533 534 static void __insert_snapshot(struct origin *o, struct dm_snapshot *s) 535 { 536 struct dm_snapshot *l; 537 538 /* Sort the list according to chunk size, largest-first smallest-last */ 539 list_for_each_entry(l, &o->snapshots, list) 540 if (l->store->chunk_size < s->store->chunk_size) 541 break; 542 list_add_tail(&s->list, &l->list); 543 } 544 545 /* 546 * Make a note of the snapshot and its origin so we can look it 547 * up when the origin has a write on it. 548 * 549 * Also validate snapshot exception store handovers. 550 * On success, returns 1 if this registration is a handover destination, 551 * otherwise returns 0. 552 */ 553 static int register_snapshot(struct dm_snapshot *snap) 554 { 555 struct origin *o, *new_o = NULL; 556 struct block_device *bdev = snap->origin->bdev; 557 int r = 0; 558 559 new_o = kmalloc(sizeof(*new_o), GFP_KERNEL); 560 if (!new_o) 561 return -ENOMEM; 562 563 down_write(&_origins_lock); 564 565 r = __validate_exception_handover(snap); 566 if (r < 0) { 567 kfree(new_o); 568 goto out; 569 } 570 571 o = __lookup_origin(bdev); 572 if (o) 573 kfree(new_o); 574 else { 575 /* New origin */ 576 o = new_o; 577 578 /* Initialise the struct */ 579 INIT_LIST_HEAD(&o->snapshots); 580 o->bdev = bdev; 581 582 __insert_origin(o); 583 } 584 585 __insert_snapshot(o, snap); 586 587 out: 588 up_write(&_origins_lock); 589 590 return r; 591 } 592 593 /* 594 * Move snapshot to correct place in list according to chunk size. 595 */ 596 static void reregister_snapshot(struct dm_snapshot *s) 597 { 598 struct block_device *bdev = s->origin->bdev; 599 600 down_write(&_origins_lock); 601 602 list_del(&s->list); 603 __insert_snapshot(__lookup_origin(bdev), s); 604 605 up_write(&_origins_lock); 606 } 607 608 static void unregister_snapshot(struct dm_snapshot *s) 609 { 610 struct origin *o; 611 612 down_write(&_origins_lock); 613 o = __lookup_origin(s->origin->bdev); 614 615 list_del(&s->list); 616 if (o && list_empty(&o->snapshots)) { 617 list_del(&o->hash_list); 618 kfree(o); 619 } 620 621 up_write(&_origins_lock); 622 } 623 624 /* 625 * Implementation of the exception hash tables. 626 * The lowest hash_shift bits of the chunk number are ignored, allowing 627 * some consecutive chunks to be grouped together. 628 */ 629 static uint32_t exception_hash(struct dm_exception_table *et, chunk_t chunk); 630 631 /* Lock to protect access to the completed and pending exception hash tables. */ 632 struct dm_exception_table_lock { 633 struct hlist_bl_head *complete_slot; 634 struct hlist_bl_head *pending_slot; 635 }; 636 637 static void dm_exception_table_lock_init(struct dm_snapshot *s, chunk_t chunk, 638 struct dm_exception_table_lock *lock) 639 { 640 struct dm_exception_table *complete = &s->complete; 641 struct dm_exception_table *pending = &s->pending; 642 643 lock->complete_slot = &complete->table[exception_hash(complete, chunk)]; 644 lock->pending_slot = &pending->table[exception_hash(pending, chunk)]; 645 } 646 647 static void dm_exception_table_lock(struct dm_exception_table_lock *lock) 648 { 649 hlist_bl_lock(lock->complete_slot); 650 hlist_bl_lock(lock->pending_slot); 651 } 652 653 static void dm_exception_table_unlock(struct dm_exception_table_lock *lock) 654 { 655 hlist_bl_unlock(lock->pending_slot); 656 hlist_bl_unlock(lock->complete_slot); 657 } 658 659 static int dm_exception_table_init(struct dm_exception_table *et, 660 uint32_t size, unsigned int hash_shift) 661 { 662 unsigned int i; 663 664 et->hash_shift = hash_shift; 665 et->hash_mask = size - 1; 666 et->table = kvmalloc_array(size, sizeof(struct hlist_bl_head), 667 GFP_KERNEL); 668 if (!et->table) 669 return -ENOMEM; 670 671 for (i = 0; i < size; i++) 672 INIT_HLIST_BL_HEAD(et->table + i); 673 674 return 0; 675 } 676 677 static void dm_exception_table_exit(struct dm_exception_table *et, 678 struct kmem_cache *mem) 679 { 680 struct hlist_bl_head *slot; 681 struct dm_exception *ex; 682 struct hlist_bl_node *pos, *n; 683 int i, size; 684 685 size = et->hash_mask + 1; 686 for (i = 0; i < size; i++) { 687 slot = et->table + i; 688 689 hlist_bl_for_each_entry_safe(ex, pos, n, slot, hash_list) 690 kmem_cache_free(mem, ex); 691 } 692 693 kvfree(et->table); 694 } 695 696 static uint32_t exception_hash(struct dm_exception_table *et, chunk_t chunk) 697 { 698 return (chunk >> et->hash_shift) & et->hash_mask; 699 } 700 701 static void dm_remove_exception(struct dm_exception *e) 702 { 703 hlist_bl_del(&e->hash_list); 704 } 705 706 /* 707 * Return the exception data for a sector, or NULL if not 708 * remapped. 709 */ 710 static struct dm_exception *dm_lookup_exception(struct dm_exception_table *et, 711 chunk_t chunk) 712 { 713 struct hlist_bl_head *slot; 714 struct hlist_bl_node *pos; 715 struct dm_exception *e; 716 717 slot = &et->table[exception_hash(et, chunk)]; 718 hlist_bl_for_each_entry(e, pos, slot, hash_list) 719 if (chunk >= e->old_chunk && 720 chunk <= e->old_chunk + dm_consecutive_chunk_count(e)) 721 return e; 722 723 return NULL; 724 } 725 726 static struct dm_exception *alloc_completed_exception(gfp_t gfp) 727 { 728 struct dm_exception *e; 729 730 e = kmem_cache_alloc(exception_cache, gfp); 731 if (!e && gfp == GFP_NOIO) 732 e = kmem_cache_alloc(exception_cache, GFP_ATOMIC); 733 734 return e; 735 } 736 737 static void free_completed_exception(struct dm_exception *e) 738 { 739 kmem_cache_free(exception_cache, e); 740 } 741 742 static struct dm_snap_pending_exception *alloc_pending_exception(struct dm_snapshot *s) 743 { 744 struct dm_snap_pending_exception *pe = mempool_alloc(&s->pending_pool, 745 GFP_NOIO); 746 747 atomic_inc(&s->pending_exceptions_count); 748 pe->snap = s; 749 750 return pe; 751 } 752 753 static void free_pending_exception(struct dm_snap_pending_exception *pe) 754 { 755 struct dm_snapshot *s = pe->snap; 756 757 mempool_free(pe, &s->pending_pool); 758 smp_mb__before_atomic(); 759 atomic_dec(&s->pending_exceptions_count); 760 } 761 762 static void dm_insert_exception(struct dm_exception_table *eh, 763 struct dm_exception *new_e) 764 { 765 struct hlist_bl_head *l; 766 struct hlist_bl_node *pos; 767 struct dm_exception *e = NULL; 768 769 l = &eh->table[exception_hash(eh, new_e->old_chunk)]; 770 771 /* Add immediately if this table doesn't support consecutive chunks */ 772 if (!eh->hash_shift) 773 goto out; 774 775 /* List is ordered by old_chunk */ 776 hlist_bl_for_each_entry(e, pos, l, hash_list) { 777 /* Insert after an existing chunk? */ 778 if (new_e->old_chunk == (e->old_chunk + 779 dm_consecutive_chunk_count(e) + 1) && 780 new_e->new_chunk == (dm_chunk_number(e->new_chunk) + 781 dm_consecutive_chunk_count(e) + 1)) { 782 dm_consecutive_chunk_count_inc(e); 783 free_completed_exception(new_e); 784 return; 785 } 786 787 /* Insert before an existing chunk? */ 788 if (new_e->old_chunk == (e->old_chunk - 1) && 789 new_e->new_chunk == (dm_chunk_number(e->new_chunk) - 1)) { 790 dm_consecutive_chunk_count_inc(e); 791 e->old_chunk--; 792 e->new_chunk--; 793 free_completed_exception(new_e); 794 return; 795 } 796 797 if (new_e->old_chunk < e->old_chunk) 798 break; 799 } 800 801 out: 802 if (!e) { 803 /* 804 * Either the table doesn't support consecutive chunks or slot 805 * l is empty. 806 */ 807 hlist_bl_add_head(&new_e->hash_list, l); 808 } else if (new_e->old_chunk < e->old_chunk) { 809 /* Add before an existing exception */ 810 hlist_bl_add_before(&new_e->hash_list, &e->hash_list); 811 } else { 812 /* Add to l's tail: e is the last exception in this slot */ 813 hlist_bl_add_behind(&new_e->hash_list, &e->hash_list); 814 } 815 } 816 817 /* 818 * Callback used by the exception stores to load exceptions when 819 * initialising. 820 */ 821 static int dm_add_exception(void *context, chunk_t old, chunk_t new) 822 { 823 struct dm_exception_table_lock lock; 824 struct dm_snapshot *s = context; 825 struct dm_exception *e; 826 827 e = alloc_completed_exception(GFP_KERNEL); 828 if (!e) 829 return -ENOMEM; 830 831 e->old_chunk = old; 832 833 /* Consecutive_count is implicitly initialised to zero */ 834 e->new_chunk = new; 835 836 /* 837 * Although there is no need to lock access to the exception tables 838 * here, if we don't then hlist_bl_add_head(), called by 839 * dm_insert_exception(), will complain about accessing the 840 * corresponding list without locking it first. 841 */ 842 dm_exception_table_lock_init(s, old, &lock); 843 844 dm_exception_table_lock(&lock); 845 dm_insert_exception(&s->complete, e); 846 dm_exception_table_unlock(&lock); 847 848 return 0; 849 } 850 851 /* 852 * Return a minimum chunk size of all snapshots that have the specified origin. 853 * Return zero if the origin has no snapshots. 854 */ 855 static uint32_t __minimum_chunk_size(struct origin *o) 856 { 857 struct dm_snapshot *snap; 858 unsigned int chunk_size = rounddown_pow_of_two(UINT_MAX); 859 860 if (o) 861 list_for_each_entry(snap, &o->snapshots, list) 862 chunk_size = min_not_zero(chunk_size, 863 snap->store->chunk_size); 864 865 return (uint32_t) chunk_size; 866 } 867 868 /* 869 * Hard coded magic. 870 */ 871 static int calc_max_buckets(void) 872 { 873 /* use a fixed size of 2MB */ 874 unsigned long mem = 2 * 1024 * 1024; 875 876 mem /= sizeof(struct hlist_bl_head); 877 878 return mem; 879 } 880 881 /* 882 * Allocate room for a suitable hash table. 883 */ 884 static int init_hash_tables(struct dm_snapshot *s) 885 { 886 sector_t hash_size, cow_dev_size, max_buckets; 887 888 /* 889 * Calculate based on the size of the original volume or 890 * the COW volume... 891 */ 892 cow_dev_size = get_dev_size(s->cow->bdev); 893 max_buckets = calc_max_buckets(); 894 895 hash_size = cow_dev_size >> s->store->chunk_shift; 896 hash_size = min(hash_size, max_buckets); 897 898 if (hash_size < 64) 899 hash_size = 64; 900 hash_size = rounddown_pow_of_two(hash_size); 901 if (dm_exception_table_init(&s->complete, hash_size, 902 DM_CHUNK_CONSECUTIVE_BITS)) 903 return -ENOMEM; 904 905 /* 906 * Allocate hash table for in-flight exceptions 907 * Make this smaller than the real hash table 908 */ 909 hash_size >>= 3; 910 if (hash_size < 64) 911 hash_size = 64; 912 913 if (dm_exception_table_init(&s->pending, hash_size, 0)) { 914 dm_exception_table_exit(&s->complete, exception_cache); 915 return -ENOMEM; 916 } 917 918 return 0; 919 } 920 921 static void merge_shutdown(struct dm_snapshot *s) 922 { 923 clear_bit_unlock(RUNNING_MERGE, &s->state_bits); 924 smp_mb__after_atomic(); 925 wake_up_bit(&s->state_bits, RUNNING_MERGE); 926 } 927 928 static struct bio *__release_queued_bios_after_merge(struct dm_snapshot *s) 929 { 930 s->first_merging_chunk = 0; 931 s->num_merging_chunks = 0; 932 933 return bio_list_get(&s->bios_queued_during_merge); 934 } 935 936 /* 937 * Remove one chunk from the index of completed exceptions. 938 */ 939 static int __remove_single_exception_chunk(struct dm_snapshot *s, 940 chunk_t old_chunk) 941 { 942 struct dm_exception *e; 943 944 e = dm_lookup_exception(&s->complete, old_chunk); 945 if (!e) { 946 DMERR("Corruption detected: exception for block %llu is " 947 "on disk but not in memory", 948 (unsigned long long)old_chunk); 949 return -EINVAL; 950 } 951 952 /* 953 * If this is the only chunk using this exception, remove exception. 954 */ 955 if (!dm_consecutive_chunk_count(e)) { 956 dm_remove_exception(e); 957 free_completed_exception(e); 958 return 0; 959 } 960 961 /* 962 * The chunk may be either at the beginning or the end of a 963 * group of consecutive chunks - never in the middle. We are 964 * removing chunks in the opposite order to that in which they 965 * were added, so this should always be true. 966 * Decrement the consecutive chunk counter and adjust the 967 * starting point if necessary. 968 */ 969 if (old_chunk == e->old_chunk) { 970 e->old_chunk++; 971 e->new_chunk++; 972 } else if (old_chunk != e->old_chunk + 973 dm_consecutive_chunk_count(e)) { 974 DMERR("Attempt to merge block %llu from the " 975 "middle of a chunk range [%llu - %llu]", 976 (unsigned long long)old_chunk, 977 (unsigned long long)e->old_chunk, 978 (unsigned long long) 979 e->old_chunk + dm_consecutive_chunk_count(e)); 980 return -EINVAL; 981 } 982 983 dm_consecutive_chunk_count_dec(e); 984 985 return 0; 986 } 987 988 static void flush_bios(struct bio *bio); 989 990 static int remove_single_exception_chunk(struct dm_snapshot *s) 991 { 992 struct bio *b = NULL; 993 int r; 994 chunk_t old_chunk = s->first_merging_chunk + s->num_merging_chunks - 1; 995 996 down_write(&s->lock); 997 998 /* 999 * Process chunks (and associated exceptions) in reverse order 1000 * so that dm_consecutive_chunk_count_dec() accounting works. 1001 */ 1002 do { 1003 r = __remove_single_exception_chunk(s, old_chunk); 1004 if (r) 1005 goto out; 1006 } while (old_chunk-- > s->first_merging_chunk); 1007 1008 b = __release_queued_bios_after_merge(s); 1009 1010 out: 1011 up_write(&s->lock); 1012 if (b) 1013 flush_bios(b); 1014 1015 return r; 1016 } 1017 1018 static int origin_write_extent(struct dm_snapshot *merging_snap, 1019 sector_t sector, unsigned int chunk_size); 1020 1021 static void merge_callback(int read_err, unsigned long write_err, 1022 void *context); 1023 1024 static uint64_t read_pending_exceptions_done_count(void) 1025 { 1026 uint64_t pending_exceptions_done; 1027 1028 spin_lock(&_pending_exceptions_done_spinlock); 1029 pending_exceptions_done = _pending_exceptions_done_count; 1030 spin_unlock(&_pending_exceptions_done_spinlock); 1031 1032 return pending_exceptions_done; 1033 } 1034 1035 static void increment_pending_exceptions_done_count(void) 1036 { 1037 spin_lock(&_pending_exceptions_done_spinlock); 1038 _pending_exceptions_done_count++; 1039 spin_unlock(&_pending_exceptions_done_spinlock); 1040 1041 wake_up_all(&_pending_exceptions_done); 1042 } 1043 1044 static void snapshot_merge_next_chunks(struct dm_snapshot *s) 1045 { 1046 int i, linear_chunks; 1047 chunk_t old_chunk, new_chunk; 1048 struct dm_io_region src, dest; 1049 sector_t io_size; 1050 uint64_t previous_count; 1051 1052 BUG_ON(!test_bit(RUNNING_MERGE, &s->state_bits)); 1053 if (unlikely(test_bit(SHUTDOWN_MERGE, &s->state_bits))) 1054 goto shut; 1055 1056 /* 1057 * valid flag never changes during merge, so no lock required. 1058 */ 1059 if (!s->valid) { 1060 DMERR("Snapshot is invalid: can't merge"); 1061 goto shut; 1062 } 1063 1064 linear_chunks = s->store->type->prepare_merge(s->store, &old_chunk, 1065 &new_chunk); 1066 if (linear_chunks <= 0) { 1067 if (linear_chunks < 0) { 1068 DMERR("Read error in exception store: " 1069 "shutting down merge"); 1070 down_write(&s->lock); 1071 s->merge_failed = true; 1072 up_write(&s->lock); 1073 } 1074 goto shut; 1075 } 1076 1077 /* Adjust old_chunk and new_chunk to reflect start of linear region */ 1078 old_chunk = old_chunk + 1 - linear_chunks; 1079 new_chunk = new_chunk + 1 - linear_chunks; 1080 1081 /* 1082 * Use one (potentially large) I/O to copy all 'linear_chunks' 1083 * from the exception store to the origin 1084 */ 1085 io_size = linear_chunks * s->store->chunk_size; 1086 1087 dest.bdev = s->origin->bdev; 1088 dest.sector = chunk_to_sector(s->store, old_chunk); 1089 dest.count = min(io_size, get_dev_size(dest.bdev) - dest.sector); 1090 1091 src.bdev = s->cow->bdev; 1092 src.sector = chunk_to_sector(s->store, new_chunk); 1093 src.count = dest.count; 1094 1095 /* 1096 * Reallocate any exceptions needed in other snapshots then 1097 * wait for the pending exceptions to complete. 1098 * Each time any pending exception (globally on the system) 1099 * completes we are woken and repeat the process to find out 1100 * if we can proceed. While this may not seem a particularly 1101 * efficient algorithm, it is not expected to have any 1102 * significant impact on performance. 1103 */ 1104 previous_count = read_pending_exceptions_done_count(); 1105 while (origin_write_extent(s, dest.sector, io_size)) { 1106 wait_event(_pending_exceptions_done, 1107 (read_pending_exceptions_done_count() != 1108 previous_count)); 1109 /* Retry after the wait, until all exceptions are done. */ 1110 previous_count = read_pending_exceptions_done_count(); 1111 } 1112 1113 down_write(&s->lock); 1114 s->first_merging_chunk = old_chunk; 1115 s->num_merging_chunks = linear_chunks; 1116 up_write(&s->lock); 1117 1118 /* Wait until writes to all 'linear_chunks' drain */ 1119 for (i = 0; i < linear_chunks; i++) 1120 __check_for_conflicting_io(s, old_chunk + i); 1121 1122 dm_kcopyd_copy(s->kcopyd_client, &src, 1, &dest, 0, merge_callback, s); 1123 return; 1124 1125 shut: 1126 merge_shutdown(s); 1127 } 1128 1129 static void error_bios(struct bio *bio); 1130 1131 static void merge_callback(int read_err, unsigned long write_err, void *context) 1132 { 1133 struct dm_snapshot *s = context; 1134 struct bio *b = NULL; 1135 1136 if (read_err || write_err) { 1137 if (read_err) 1138 DMERR("Read error: shutting down merge."); 1139 else 1140 DMERR("Write error: shutting down merge."); 1141 goto shut; 1142 } 1143 1144 if (blkdev_issue_flush(s->origin->bdev) < 0) { 1145 DMERR("Flush after merge failed: shutting down merge"); 1146 goto shut; 1147 } 1148 1149 if (s->store->type->commit_merge(s->store, 1150 s->num_merging_chunks) < 0) { 1151 DMERR("Write error in exception store: shutting down merge"); 1152 goto shut; 1153 } 1154 1155 if (remove_single_exception_chunk(s) < 0) 1156 goto shut; 1157 1158 snapshot_merge_next_chunks(s); 1159 1160 return; 1161 1162 shut: 1163 down_write(&s->lock); 1164 s->merge_failed = true; 1165 b = __release_queued_bios_after_merge(s); 1166 up_write(&s->lock); 1167 error_bios(b); 1168 1169 merge_shutdown(s); 1170 } 1171 1172 static void start_merge(struct dm_snapshot *s) 1173 { 1174 if (!test_and_set_bit(RUNNING_MERGE, &s->state_bits)) 1175 snapshot_merge_next_chunks(s); 1176 } 1177 1178 /* 1179 * Stop the merging process and wait until it finishes. 1180 */ 1181 static void stop_merge(struct dm_snapshot *s) 1182 { 1183 set_bit(SHUTDOWN_MERGE, &s->state_bits); 1184 wait_on_bit(&s->state_bits, RUNNING_MERGE, TASK_UNINTERRUPTIBLE); 1185 clear_bit(SHUTDOWN_MERGE, &s->state_bits); 1186 } 1187 1188 static int parse_snapshot_features(struct dm_arg_set *as, struct dm_snapshot *s, 1189 struct dm_target *ti) 1190 { 1191 int r; 1192 unsigned int argc; 1193 const char *arg_name; 1194 1195 static const struct dm_arg _args[] = { 1196 {0, 2, "Invalid number of feature arguments"}, 1197 }; 1198 1199 /* 1200 * No feature arguments supplied. 1201 */ 1202 if (!as->argc) 1203 return 0; 1204 1205 r = dm_read_arg_group(_args, as, &argc, &ti->error); 1206 if (r) 1207 return -EINVAL; 1208 1209 while (argc && !r) { 1210 arg_name = dm_shift_arg(as); 1211 argc--; 1212 1213 if (!strcasecmp(arg_name, "discard_zeroes_cow")) 1214 s->discard_zeroes_cow = true; 1215 1216 else if (!strcasecmp(arg_name, "discard_passdown_origin")) 1217 s->discard_passdown_origin = true; 1218 1219 else { 1220 ti->error = "Unrecognised feature requested"; 1221 r = -EINVAL; 1222 break; 1223 } 1224 } 1225 1226 if (!s->discard_zeroes_cow && s->discard_passdown_origin) { 1227 /* 1228 * TODO: really these are disjoint.. but ti->num_discard_bios 1229 * and dm_bio_get_target_bio_nr() require rigid constraints. 1230 */ 1231 ti->error = "discard_passdown_origin feature depends on discard_zeroes_cow"; 1232 r = -EINVAL; 1233 } 1234 1235 return r; 1236 } 1237 1238 /* 1239 * Construct a snapshot mapping: 1240 * <origin_dev> <COW-dev> <p|po|n> <chunk-size> [<# feature args> [<arg>]*] 1241 */ 1242 static int snapshot_ctr(struct dm_target *ti, unsigned int argc, char **argv) 1243 { 1244 struct dm_snapshot *s; 1245 struct dm_arg_set as; 1246 int i; 1247 int r = -EINVAL; 1248 char *origin_path, *cow_path; 1249 dev_t origin_dev, cow_dev; 1250 unsigned int args_used, num_flush_bios = 1; 1251 fmode_t origin_mode = FMODE_READ; 1252 1253 if (argc < 4) { 1254 ti->error = "requires 4 or more arguments"; 1255 r = -EINVAL; 1256 goto bad; 1257 } 1258 1259 if (dm_target_is_snapshot_merge(ti)) { 1260 num_flush_bios = 2; 1261 origin_mode = FMODE_WRITE; 1262 } 1263 1264 s = kzalloc(sizeof(*s), GFP_KERNEL); 1265 if (!s) { 1266 ti->error = "Cannot allocate private snapshot structure"; 1267 r = -ENOMEM; 1268 goto bad; 1269 } 1270 1271 as.argc = argc; 1272 as.argv = argv; 1273 dm_consume_args(&as, 4); 1274 r = parse_snapshot_features(&as, s, ti); 1275 if (r) 1276 goto bad_features; 1277 1278 origin_path = argv[0]; 1279 argv++; 1280 argc--; 1281 1282 r = dm_get_device(ti, origin_path, origin_mode, &s->origin); 1283 if (r) { 1284 ti->error = "Cannot get origin device"; 1285 goto bad_origin; 1286 } 1287 origin_dev = s->origin->bdev->bd_dev; 1288 1289 cow_path = argv[0]; 1290 argv++; 1291 argc--; 1292 1293 cow_dev = dm_get_dev_t(cow_path); 1294 if (cow_dev && cow_dev == origin_dev) { 1295 ti->error = "COW device cannot be the same as origin device"; 1296 r = -EINVAL; 1297 goto bad_cow; 1298 } 1299 1300 r = dm_get_device(ti, cow_path, dm_table_get_mode(ti->table), &s->cow); 1301 if (r) { 1302 ti->error = "Cannot get COW device"; 1303 goto bad_cow; 1304 } 1305 1306 r = dm_exception_store_create(ti, argc, argv, s, &args_used, &s->store); 1307 if (r) { 1308 ti->error = "Couldn't create exception store"; 1309 r = -EINVAL; 1310 goto bad_store; 1311 } 1312 1313 argv += args_used; 1314 argc -= args_used; 1315 1316 s->ti = ti; 1317 s->valid = 1; 1318 s->snapshot_overflowed = 0; 1319 s->active = 0; 1320 atomic_set(&s->pending_exceptions_count, 0); 1321 spin_lock_init(&s->pe_allocation_lock); 1322 s->exception_start_sequence = 0; 1323 s->exception_complete_sequence = 0; 1324 s->out_of_order_tree = RB_ROOT; 1325 init_rwsem(&s->lock); 1326 INIT_LIST_HEAD(&s->list); 1327 spin_lock_init(&s->pe_lock); 1328 s->state_bits = 0; 1329 s->merge_failed = false; 1330 s->first_merging_chunk = 0; 1331 s->num_merging_chunks = 0; 1332 bio_list_init(&s->bios_queued_during_merge); 1333 1334 /* Allocate hash table for COW data */ 1335 if (init_hash_tables(s)) { 1336 ti->error = "Unable to allocate hash table space"; 1337 r = -ENOMEM; 1338 goto bad_hash_tables; 1339 } 1340 1341 init_waitqueue_head(&s->in_progress_wait); 1342 1343 s->kcopyd_client = dm_kcopyd_client_create(&dm_kcopyd_throttle); 1344 if (IS_ERR(s->kcopyd_client)) { 1345 r = PTR_ERR(s->kcopyd_client); 1346 ti->error = "Could not create kcopyd client"; 1347 goto bad_kcopyd; 1348 } 1349 1350 r = mempool_init_slab_pool(&s->pending_pool, MIN_IOS, pending_cache); 1351 if (r) { 1352 ti->error = "Could not allocate mempool for pending exceptions"; 1353 goto bad_pending_pool; 1354 } 1355 1356 for (i = 0; i < DM_TRACKED_CHUNK_HASH_SIZE; i++) 1357 INIT_HLIST_HEAD(&s->tracked_chunk_hash[i]); 1358 1359 spin_lock_init(&s->tracked_chunk_lock); 1360 1361 ti->private = s; 1362 ti->num_flush_bios = num_flush_bios; 1363 if (s->discard_zeroes_cow) 1364 ti->num_discard_bios = (s->discard_passdown_origin ? 2 : 1); 1365 ti->per_io_data_size = sizeof(struct dm_snap_tracked_chunk); 1366 1367 /* Add snapshot to the list of snapshots for this origin */ 1368 /* Exceptions aren't triggered till snapshot_resume() is called */ 1369 r = register_snapshot(s); 1370 if (r == -ENOMEM) { 1371 ti->error = "Snapshot origin struct allocation failed"; 1372 goto bad_load_and_register; 1373 } else if (r < 0) { 1374 /* invalid handover, register_snapshot has set ti->error */ 1375 goto bad_load_and_register; 1376 } 1377 1378 /* 1379 * Metadata must only be loaded into one table at once, so skip this 1380 * if metadata will be handed over during resume. 1381 * Chunk size will be set during the handover - set it to zero to 1382 * ensure it's ignored. 1383 */ 1384 if (r > 0) { 1385 s->store->chunk_size = 0; 1386 return 0; 1387 } 1388 1389 r = s->store->type->read_metadata(s->store, dm_add_exception, 1390 (void *)s); 1391 if (r < 0) { 1392 ti->error = "Failed to read snapshot metadata"; 1393 goto bad_read_metadata; 1394 } else if (r > 0) { 1395 s->valid = 0; 1396 DMWARN("Snapshot is marked invalid."); 1397 } 1398 1399 if (!s->store->chunk_size) { 1400 ti->error = "Chunk size not set"; 1401 r = -EINVAL; 1402 goto bad_read_metadata; 1403 } 1404 1405 r = dm_set_target_max_io_len(ti, s->store->chunk_size); 1406 if (r) 1407 goto bad_read_metadata; 1408 1409 return 0; 1410 1411 bad_read_metadata: 1412 unregister_snapshot(s); 1413 bad_load_and_register: 1414 mempool_exit(&s->pending_pool); 1415 bad_pending_pool: 1416 dm_kcopyd_client_destroy(s->kcopyd_client); 1417 bad_kcopyd: 1418 dm_exception_table_exit(&s->pending, pending_cache); 1419 dm_exception_table_exit(&s->complete, exception_cache); 1420 bad_hash_tables: 1421 dm_exception_store_destroy(s->store); 1422 bad_store: 1423 dm_put_device(ti, s->cow); 1424 bad_cow: 1425 dm_put_device(ti, s->origin); 1426 bad_origin: 1427 bad_features: 1428 kfree(s); 1429 bad: 1430 return r; 1431 } 1432 1433 static void __free_exceptions(struct dm_snapshot *s) 1434 { 1435 dm_kcopyd_client_destroy(s->kcopyd_client); 1436 s->kcopyd_client = NULL; 1437 1438 dm_exception_table_exit(&s->pending, pending_cache); 1439 dm_exception_table_exit(&s->complete, exception_cache); 1440 } 1441 1442 static void __handover_exceptions(struct dm_snapshot *snap_src, 1443 struct dm_snapshot *snap_dest) 1444 { 1445 union { 1446 struct dm_exception_table table_swap; 1447 struct dm_exception_store *store_swap; 1448 } u; 1449 1450 /* 1451 * Swap all snapshot context information between the two instances. 1452 */ 1453 u.table_swap = snap_dest->complete; 1454 snap_dest->complete = snap_src->complete; 1455 snap_src->complete = u.table_swap; 1456 1457 u.store_swap = snap_dest->store; 1458 snap_dest->store = snap_src->store; 1459 snap_dest->store->userspace_supports_overflow = u.store_swap->userspace_supports_overflow; 1460 snap_src->store = u.store_swap; 1461 1462 snap_dest->store->snap = snap_dest; 1463 snap_src->store->snap = snap_src; 1464 1465 snap_dest->ti->max_io_len = snap_dest->store->chunk_size; 1466 snap_dest->valid = snap_src->valid; 1467 snap_dest->snapshot_overflowed = snap_src->snapshot_overflowed; 1468 1469 /* 1470 * Set source invalid to ensure it receives no further I/O. 1471 */ 1472 snap_src->valid = 0; 1473 } 1474 1475 static void snapshot_dtr(struct dm_target *ti) 1476 { 1477 #ifdef CONFIG_DM_DEBUG 1478 int i; 1479 #endif 1480 struct dm_snapshot *s = ti->private; 1481 struct dm_snapshot *snap_src = NULL, *snap_dest = NULL; 1482 1483 down_read(&_origins_lock); 1484 /* Check whether exception handover must be cancelled */ 1485 (void) __find_snapshots_sharing_cow(s, &snap_src, &snap_dest, NULL); 1486 if (snap_src && snap_dest && (s == snap_src)) { 1487 down_write(&snap_dest->lock); 1488 snap_dest->valid = 0; 1489 up_write(&snap_dest->lock); 1490 DMERR("Cancelling snapshot handover."); 1491 } 1492 up_read(&_origins_lock); 1493 1494 if (dm_target_is_snapshot_merge(ti)) 1495 stop_merge(s); 1496 1497 /* Prevent further origin writes from using this snapshot. */ 1498 /* After this returns there can be no new kcopyd jobs. */ 1499 unregister_snapshot(s); 1500 1501 while (atomic_read(&s->pending_exceptions_count)) 1502 fsleep(1000); 1503 /* 1504 * Ensure instructions in mempool_exit aren't reordered 1505 * before atomic_read. 1506 */ 1507 smp_mb(); 1508 1509 #ifdef CONFIG_DM_DEBUG 1510 for (i = 0; i < DM_TRACKED_CHUNK_HASH_SIZE; i++) 1511 BUG_ON(!hlist_empty(&s->tracked_chunk_hash[i])); 1512 #endif 1513 1514 __free_exceptions(s); 1515 1516 mempool_exit(&s->pending_pool); 1517 1518 dm_exception_store_destroy(s->store); 1519 1520 dm_put_device(ti, s->cow); 1521 1522 dm_put_device(ti, s->origin); 1523 1524 WARN_ON(s->in_progress); 1525 1526 kfree(s); 1527 } 1528 1529 static void account_start_copy(struct dm_snapshot *s) 1530 { 1531 spin_lock(&s->in_progress_wait.lock); 1532 s->in_progress++; 1533 spin_unlock(&s->in_progress_wait.lock); 1534 } 1535 1536 static void account_end_copy(struct dm_snapshot *s) 1537 { 1538 spin_lock(&s->in_progress_wait.lock); 1539 BUG_ON(!s->in_progress); 1540 s->in_progress--; 1541 if (likely(s->in_progress <= cow_threshold) && 1542 unlikely(waitqueue_active(&s->in_progress_wait))) 1543 wake_up_locked(&s->in_progress_wait); 1544 spin_unlock(&s->in_progress_wait.lock); 1545 } 1546 1547 static bool wait_for_in_progress(struct dm_snapshot *s, bool unlock_origins) 1548 { 1549 if (unlikely(s->in_progress > cow_threshold)) { 1550 spin_lock(&s->in_progress_wait.lock); 1551 if (likely(s->in_progress > cow_threshold)) { 1552 /* 1553 * NOTE: this throttle doesn't account for whether 1554 * the caller is servicing an IO that will trigger a COW 1555 * so excess throttling may result for chunks not required 1556 * to be COW'd. But if cow_threshold was reached, extra 1557 * throttling is unlikely to negatively impact performance. 1558 */ 1559 DECLARE_WAITQUEUE(wait, current); 1560 1561 __add_wait_queue(&s->in_progress_wait, &wait); 1562 __set_current_state(TASK_UNINTERRUPTIBLE); 1563 spin_unlock(&s->in_progress_wait.lock); 1564 if (unlock_origins) 1565 up_read(&_origins_lock); 1566 io_schedule(); 1567 remove_wait_queue(&s->in_progress_wait, &wait); 1568 return false; 1569 } 1570 spin_unlock(&s->in_progress_wait.lock); 1571 } 1572 return true; 1573 } 1574 1575 /* 1576 * Flush a list of buffers. 1577 */ 1578 static void flush_bios(struct bio *bio) 1579 { 1580 struct bio *n; 1581 1582 while (bio) { 1583 n = bio->bi_next; 1584 bio->bi_next = NULL; 1585 submit_bio_noacct(bio); 1586 bio = n; 1587 } 1588 } 1589 1590 static int do_origin(struct dm_dev *origin, struct bio *bio, bool limit); 1591 1592 /* 1593 * Flush a list of buffers. 1594 */ 1595 static void retry_origin_bios(struct dm_snapshot *s, struct bio *bio) 1596 { 1597 struct bio *n; 1598 int r; 1599 1600 while (bio) { 1601 n = bio->bi_next; 1602 bio->bi_next = NULL; 1603 r = do_origin(s->origin, bio, false); 1604 if (r == DM_MAPIO_REMAPPED) 1605 submit_bio_noacct(bio); 1606 bio = n; 1607 } 1608 } 1609 1610 /* 1611 * Error a list of buffers. 1612 */ 1613 static void error_bios(struct bio *bio) 1614 { 1615 struct bio *n; 1616 1617 while (bio) { 1618 n = bio->bi_next; 1619 bio->bi_next = NULL; 1620 bio_io_error(bio); 1621 bio = n; 1622 } 1623 } 1624 1625 static void __invalidate_snapshot(struct dm_snapshot *s, int err) 1626 { 1627 if (!s->valid) 1628 return; 1629 1630 if (err == -EIO) 1631 DMERR("Invalidating snapshot: Error reading/writing."); 1632 else if (err == -ENOMEM) 1633 DMERR("Invalidating snapshot: Unable to allocate exception."); 1634 1635 if (s->store->type->drop_snapshot) 1636 s->store->type->drop_snapshot(s->store); 1637 1638 s->valid = 0; 1639 1640 dm_table_event(s->ti->table); 1641 } 1642 1643 static void invalidate_snapshot(struct dm_snapshot *s, int err) 1644 { 1645 down_write(&s->lock); 1646 __invalidate_snapshot(s, err); 1647 up_write(&s->lock); 1648 } 1649 1650 static void pending_complete(void *context, int success) 1651 { 1652 struct dm_snap_pending_exception *pe = context; 1653 struct dm_exception *e; 1654 struct dm_snapshot *s = pe->snap; 1655 struct bio *origin_bios = NULL; 1656 struct bio *snapshot_bios = NULL; 1657 struct bio *full_bio = NULL; 1658 struct dm_exception_table_lock lock; 1659 int error = 0; 1660 1661 dm_exception_table_lock_init(s, pe->e.old_chunk, &lock); 1662 1663 if (!success) { 1664 /* Read/write error - snapshot is unusable */ 1665 invalidate_snapshot(s, -EIO); 1666 error = 1; 1667 1668 dm_exception_table_lock(&lock); 1669 goto out; 1670 } 1671 1672 e = alloc_completed_exception(GFP_NOIO); 1673 if (!e) { 1674 invalidate_snapshot(s, -ENOMEM); 1675 error = 1; 1676 1677 dm_exception_table_lock(&lock); 1678 goto out; 1679 } 1680 *e = pe->e; 1681 1682 down_read(&s->lock); 1683 dm_exception_table_lock(&lock); 1684 if (!s->valid) { 1685 up_read(&s->lock); 1686 free_completed_exception(e); 1687 error = 1; 1688 1689 goto out; 1690 } 1691 1692 /* 1693 * Add a proper exception. After inserting the completed exception all 1694 * subsequent snapshot reads to this chunk will be redirected to the 1695 * COW device. This ensures that we do not starve. Moreover, as long 1696 * as the pending exception exists, neither origin writes nor snapshot 1697 * merging can overwrite the chunk in origin. 1698 */ 1699 dm_insert_exception(&s->complete, e); 1700 up_read(&s->lock); 1701 1702 /* Wait for conflicting reads to drain */ 1703 if (__chunk_is_tracked(s, pe->e.old_chunk)) { 1704 dm_exception_table_unlock(&lock); 1705 __check_for_conflicting_io(s, pe->e.old_chunk); 1706 dm_exception_table_lock(&lock); 1707 } 1708 1709 out: 1710 /* Remove the in-flight exception from the list */ 1711 dm_remove_exception(&pe->e); 1712 1713 dm_exception_table_unlock(&lock); 1714 1715 snapshot_bios = bio_list_get(&pe->snapshot_bios); 1716 origin_bios = bio_list_get(&pe->origin_bios); 1717 full_bio = pe->full_bio; 1718 if (full_bio) 1719 full_bio->bi_end_io = pe->full_bio_end_io; 1720 increment_pending_exceptions_done_count(); 1721 1722 /* Submit any pending write bios */ 1723 if (error) { 1724 if (full_bio) 1725 bio_io_error(full_bio); 1726 error_bios(snapshot_bios); 1727 } else { 1728 if (full_bio) 1729 bio_endio(full_bio); 1730 flush_bios(snapshot_bios); 1731 } 1732 1733 retry_origin_bios(s, origin_bios); 1734 1735 free_pending_exception(pe); 1736 } 1737 1738 static void complete_exception(struct dm_snap_pending_exception *pe) 1739 { 1740 struct dm_snapshot *s = pe->snap; 1741 1742 /* Update the metadata if we are persistent */ 1743 s->store->type->commit_exception(s->store, &pe->e, !pe->copy_error, 1744 pending_complete, pe); 1745 } 1746 1747 /* 1748 * Called when the copy I/O has finished. kcopyd actually runs 1749 * this code so don't block. 1750 */ 1751 static void copy_callback(int read_err, unsigned long write_err, void *context) 1752 { 1753 struct dm_snap_pending_exception *pe = context; 1754 struct dm_snapshot *s = pe->snap; 1755 1756 pe->copy_error = read_err || write_err; 1757 1758 if (pe->exception_sequence == s->exception_complete_sequence) { 1759 struct rb_node *next; 1760 1761 s->exception_complete_sequence++; 1762 complete_exception(pe); 1763 1764 next = rb_first(&s->out_of_order_tree); 1765 while (next) { 1766 pe = rb_entry(next, struct dm_snap_pending_exception, 1767 out_of_order_node); 1768 if (pe->exception_sequence != s->exception_complete_sequence) 1769 break; 1770 next = rb_next(next); 1771 s->exception_complete_sequence++; 1772 rb_erase(&pe->out_of_order_node, &s->out_of_order_tree); 1773 complete_exception(pe); 1774 cond_resched(); 1775 } 1776 } else { 1777 struct rb_node *parent = NULL; 1778 struct rb_node **p = &s->out_of_order_tree.rb_node; 1779 struct dm_snap_pending_exception *pe2; 1780 1781 while (*p) { 1782 pe2 = rb_entry(*p, struct dm_snap_pending_exception, out_of_order_node); 1783 parent = *p; 1784 1785 BUG_ON(pe->exception_sequence == pe2->exception_sequence); 1786 if (pe->exception_sequence < pe2->exception_sequence) 1787 p = &((*p)->rb_left); 1788 else 1789 p = &((*p)->rb_right); 1790 } 1791 1792 rb_link_node(&pe->out_of_order_node, parent, p); 1793 rb_insert_color(&pe->out_of_order_node, &s->out_of_order_tree); 1794 } 1795 account_end_copy(s); 1796 } 1797 1798 /* 1799 * Dispatches the copy operation to kcopyd. 1800 */ 1801 static void start_copy(struct dm_snap_pending_exception *pe) 1802 { 1803 struct dm_snapshot *s = pe->snap; 1804 struct dm_io_region src, dest; 1805 struct block_device *bdev = s->origin->bdev; 1806 sector_t dev_size; 1807 1808 dev_size = get_dev_size(bdev); 1809 1810 src.bdev = bdev; 1811 src.sector = chunk_to_sector(s->store, pe->e.old_chunk); 1812 src.count = min((sector_t)s->store->chunk_size, dev_size - src.sector); 1813 1814 dest.bdev = s->cow->bdev; 1815 dest.sector = chunk_to_sector(s->store, pe->e.new_chunk); 1816 dest.count = src.count; 1817 1818 /* Hand over to kcopyd */ 1819 account_start_copy(s); 1820 dm_kcopyd_copy(s->kcopyd_client, &src, 1, &dest, 0, copy_callback, pe); 1821 } 1822 1823 static void full_bio_end_io(struct bio *bio) 1824 { 1825 void *callback_data = bio->bi_private; 1826 1827 dm_kcopyd_do_callback(callback_data, 0, bio->bi_status ? 1 : 0); 1828 } 1829 1830 static void start_full_bio(struct dm_snap_pending_exception *pe, 1831 struct bio *bio) 1832 { 1833 struct dm_snapshot *s = pe->snap; 1834 void *callback_data; 1835 1836 pe->full_bio = bio; 1837 pe->full_bio_end_io = bio->bi_end_io; 1838 1839 account_start_copy(s); 1840 callback_data = dm_kcopyd_prepare_callback(s->kcopyd_client, 1841 copy_callback, pe); 1842 1843 bio->bi_end_io = full_bio_end_io; 1844 bio->bi_private = callback_data; 1845 1846 submit_bio_noacct(bio); 1847 } 1848 1849 static struct dm_snap_pending_exception * 1850 __lookup_pending_exception(struct dm_snapshot *s, chunk_t chunk) 1851 { 1852 struct dm_exception *e = dm_lookup_exception(&s->pending, chunk); 1853 1854 if (!e) 1855 return NULL; 1856 1857 return container_of(e, struct dm_snap_pending_exception, e); 1858 } 1859 1860 /* 1861 * Inserts a pending exception into the pending table. 1862 * 1863 * NOTE: a write lock must be held on the chunk's pending exception table slot 1864 * before calling this. 1865 */ 1866 static struct dm_snap_pending_exception * 1867 __insert_pending_exception(struct dm_snapshot *s, 1868 struct dm_snap_pending_exception *pe, chunk_t chunk) 1869 { 1870 pe->e.old_chunk = chunk; 1871 bio_list_init(&pe->origin_bios); 1872 bio_list_init(&pe->snapshot_bios); 1873 pe->started = 0; 1874 pe->full_bio = NULL; 1875 1876 spin_lock(&s->pe_allocation_lock); 1877 if (s->store->type->prepare_exception(s->store, &pe->e)) { 1878 spin_unlock(&s->pe_allocation_lock); 1879 free_pending_exception(pe); 1880 return NULL; 1881 } 1882 1883 pe->exception_sequence = s->exception_start_sequence++; 1884 spin_unlock(&s->pe_allocation_lock); 1885 1886 dm_insert_exception(&s->pending, &pe->e); 1887 1888 return pe; 1889 } 1890 1891 /* 1892 * Looks to see if this snapshot already has a pending exception 1893 * for this chunk, otherwise it allocates a new one and inserts 1894 * it into the pending table. 1895 * 1896 * NOTE: a write lock must be held on the chunk's pending exception table slot 1897 * before calling this. 1898 */ 1899 static struct dm_snap_pending_exception * 1900 __find_pending_exception(struct dm_snapshot *s, 1901 struct dm_snap_pending_exception *pe, chunk_t chunk) 1902 { 1903 struct dm_snap_pending_exception *pe2; 1904 1905 pe2 = __lookup_pending_exception(s, chunk); 1906 if (pe2) { 1907 free_pending_exception(pe); 1908 return pe2; 1909 } 1910 1911 return __insert_pending_exception(s, pe, chunk); 1912 } 1913 1914 static void remap_exception(struct dm_snapshot *s, struct dm_exception *e, 1915 struct bio *bio, chunk_t chunk) 1916 { 1917 bio_set_dev(bio, s->cow->bdev); 1918 bio->bi_iter.bi_sector = 1919 chunk_to_sector(s->store, dm_chunk_number(e->new_chunk) + 1920 (chunk - e->old_chunk)) + 1921 (bio->bi_iter.bi_sector & s->store->chunk_mask); 1922 } 1923 1924 static void zero_callback(int read_err, unsigned long write_err, void *context) 1925 { 1926 struct bio *bio = context; 1927 struct dm_snapshot *s = bio->bi_private; 1928 1929 account_end_copy(s); 1930 bio->bi_status = write_err ? BLK_STS_IOERR : 0; 1931 bio_endio(bio); 1932 } 1933 1934 static void zero_exception(struct dm_snapshot *s, struct dm_exception *e, 1935 struct bio *bio, chunk_t chunk) 1936 { 1937 struct dm_io_region dest; 1938 1939 dest.bdev = s->cow->bdev; 1940 dest.sector = bio->bi_iter.bi_sector; 1941 dest.count = s->store->chunk_size; 1942 1943 account_start_copy(s); 1944 WARN_ON_ONCE(bio->bi_private); 1945 bio->bi_private = s; 1946 dm_kcopyd_zero(s->kcopyd_client, 1, &dest, 0, zero_callback, bio); 1947 } 1948 1949 static bool io_overlaps_chunk(struct dm_snapshot *s, struct bio *bio) 1950 { 1951 return bio->bi_iter.bi_size == 1952 (s->store->chunk_size << SECTOR_SHIFT); 1953 } 1954 1955 static int snapshot_map(struct dm_target *ti, struct bio *bio) 1956 { 1957 struct dm_exception *e; 1958 struct dm_snapshot *s = ti->private; 1959 int r = DM_MAPIO_REMAPPED; 1960 chunk_t chunk; 1961 struct dm_snap_pending_exception *pe = NULL; 1962 struct dm_exception_table_lock lock; 1963 1964 init_tracked_chunk(bio); 1965 1966 if (bio->bi_opf & REQ_PREFLUSH) { 1967 bio_set_dev(bio, s->cow->bdev); 1968 return DM_MAPIO_REMAPPED; 1969 } 1970 1971 chunk = sector_to_chunk(s->store, bio->bi_iter.bi_sector); 1972 dm_exception_table_lock_init(s, chunk, &lock); 1973 1974 /* Full snapshots are not usable */ 1975 /* To get here the table must be live so s->active is always set. */ 1976 if (!s->valid) 1977 return DM_MAPIO_KILL; 1978 1979 if (bio_data_dir(bio) == WRITE) { 1980 while (unlikely(!wait_for_in_progress(s, false))) 1981 ; /* wait_for_in_progress() has slept */ 1982 } 1983 1984 down_read(&s->lock); 1985 dm_exception_table_lock(&lock); 1986 1987 if (!s->valid || (unlikely(s->snapshot_overflowed) && 1988 bio_data_dir(bio) == WRITE)) { 1989 r = DM_MAPIO_KILL; 1990 goto out_unlock; 1991 } 1992 1993 if (unlikely(bio_op(bio) == REQ_OP_DISCARD)) { 1994 if (s->discard_passdown_origin && dm_bio_get_target_bio_nr(bio)) { 1995 /* 1996 * passdown discard to origin (without triggering 1997 * snapshot exceptions via do_origin; doing so would 1998 * defeat the goal of freeing space in origin that is 1999 * implied by the "discard_passdown_origin" feature) 2000 */ 2001 bio_set_dev(bio, s->origin->bdev); 2002 track_chunk(s, bio, chunk); 2003 goto out_unlock; 2004 } 2005 /* discard to snapshot (target_bio_nr == 0) zeroes exceptions */ 2006 } 2007 2008 /* If the block is already remapped - use that, else remap it */ 2009 e = dm_lookup_exception(&s->complete, chunk); 2010 if (e) { 2011 remap_exception(s, e, bio, chunk); 2012 if (unlikely(bio_op(bio) == REQ_OP_DISCARD) && 2013 io_overlaps_chunk(s, bio)) { 2014 dm_exception_table_unlock(&lock); 2015 up_read(&s->lock); 2016 zero_exception(s, e, bio, chunk); 2017 r = DM_MAPIO_SUBMITTED; /* discard is not issued */ 2018 goto out; 2019 } 2020 goto out_unlock; 2021 } 2022 2023 if (unlikely(bio_op(bio) == REQ_OP_DISCARD)) { 2024 /* 2025 * If no exception exists, complete discard immediately 2026 * otherwise it'll trigger copy-out. 2027 */ 2028 bio_endio(bio); 2029 r = DM_MAPIO_SUBMITTED; 2030 goto out_unlock; 2031 } 2032 2033 /* 2034 * Write to snapshot - higher level takes care of RW/RO 2035 * flags so we should only get this if we are 2036 * writable. 2037 */ 2038 if (bio_data_dir(bio) == WRITE) { 2039 pe = __lookup_pending_exception(s, chunk); 2040 if (!pe) { 2041 dm_exception_table_unlock(&lock); 2042 pe = alloc_pending_exception(s); 2043 dm_exception_table_lock(&lock); 2044 2045 e = dm_lookup_exception(&s->complete, chunk); 2046 if (e) { 2047 free_pending_exception(pe); 2048 remap_exception(s, e, bio, chunk); 2049 goto out_unlock; 2050 } 2051 2052 pe = __find_pending_exception(s, pe, chunk); 2053 if (!pe) { 2054 dm_exception_table_unlock(&lock); 2055 up_read(&s->lock); 2056 2057 down_write(&s->lock); 2058 2059 if (s->store->userspace_supports_overflow) { 2060 if (s->valid && !s->snapshot_overflowed) { 2061 s->snapshot_overflowed = 1; 2062 DMERR("Snapshot overflowed: Unable to allocate exception."); 2063 } 2064 } else 2065 __invalidate_snapshot(s, -ENOMEM); 2066 up_write(&s->lock); 2067 2068 r = DM_MAPIO_KILL; 2069 goto out; 2070 } 2071 } 2072 2073 remap_exception(s, &pe->e, bio, chunk); 2074 2075 r = DM_MAPIO_SUBMITTED; 2076 2077 if (!pe->started && io_overlaps_chunk(s, bio)) { 2078 pe->started = 1; 2079 2080 dm_exception_table_unlock(&lock); 2081 up_read(&s->lock); 2082 2083 start_full_bio(pe, bio); 2084 goto out; 2085 } 2086 2087 bio_list_add(&pe->snapshot_bios, bio); 2088 2089 if (!pe->started) { 2090 /* this is protected by the exception table lock */ 2091 pe->started = 1; 2092 2093 dm_exception_table_unlock(&lock); 2094 up_read(&s->lock); 2095 2096 start_copy(pe); 2097 goto out; 2098 } 2099 } else { 2100 bio_set_dev(bio, s->origin->bdev); 2101 track_chunk(s, bio, chunk); 2102 } 2103 2104 out_unlock: 2105 dm_exception_table_unlock(&lock); 2106 up_read(&s->lock); 2107 out: 2108 return r; 2109 } 2110 2111 /* 2112 * A snapshot-merge target behaves like a combination of a snapshot 2113 * target and a snapshot-origin target. It only generates new 2114 * exceptions in other snapshots and not in the one that is being 2115 * merged. 2116 * 2117 * For each chunk, if there is an existing exception, it is used to 2118 * redirect I/O to the cow device. Otherwise I/O is sent to the origin, 2119 * which in turn might generate exceptions in other snapshots. 2120 * If merging is currently taking place on the chunk in question, the 2121 * I/O is deferred by adding it to s->bios_queued_during_merge. 2122 */ 2123 static int snapshot_merge_map(struct dm_target *ti, struct bio *bio) 2124 { 2125 struct dm_exception *e; 2126 struct dm_snapshot *s = ti->private; 2127 int r = DM_MAPIO_REMAPPED; 2128 chunk_t chunk; 2129 2130 init_tracked_chunk(bio); 2131 2132 if (bio->bi_opf & REQ_PREFLUSH) { 2133 if (!dm_bio_get_target_bio_nr(bio)) 2134 bio_set_dev(bio, s->origin->bdev); 2135 else 2136 bio_set_dev(bio, s->cow->bdev); 2137 return DM_MAPIO_REMAPPED; 2138 } 2139 2140 if (unlikely(bio_op(bio) == REQ_OP_DISCARD)) { 2141 /* Once merging, discards no longer effect change */ 2142 bio_endio(bio); 2143 return DM_MAPIO_SUBMITTED; 2144 } 2145 2146 chunk = sector_to_chunk(s->store, bio->bi_iter.bi_sector); 2147 2148 down_write(&s->lock); 2149 2150 /* Full merging snapshots are redirected to the origin */ 2151 if (!s->valid) 2152 goto redirect_to_origin; 2153 2154 /* If the block is already remapped - use that */ 2155 e = dm_lookup_exception(&s->complete, chunk); 2156 if (e) { 2157 /* Queue writes overlapping with chunks being merged */ 2158 if (bio_data_dir(bio) == WRITE && 2159 chunk >= s->first_merging_chunk && 2160 chunk < (s->first_merging_chunk + 2161 s->num_merging_chunks)) { 2162 bio_set_dev(bio, s->origin->bdev); 2163 bio_list_add(&s->bios_queued_during_merge, bio); 2164 r = DM_MAPIO_SUBMITTED; 2165 goto out_unlock; 2166 } 2167 2168 remap_exception(s, e, bio, chunk); 2169 2170 if (bio_data_dir(bio) == WRITE) 2171 track_chunk(s, bio, chunk); 2172 goto out_unlock; 2173 } 2174 2175 redirect_to_origin: 2176 bio_set_dev(bio, s->origin->bdev); 2177 2178 if (bio_data_dir(bio) == WRITE) { 2179 up_write(&s->lock); 2180 return do_origin(s->origin, bio, false); 2181 } 2182 2183 out_unlock: 2184 up_write(&s->lock); 2185 2186 return r; 2187 } 2188 2189 static int snapshot_end_io(struct dm_target *ti, struct bio *bio, 2190 blk_status_t *error) 2191 { 2192 struct dm_snapshot *s = ti->private; 2193 2194 if (is_bio_tracked(bio)) 2195 stop_tracking_chunk(s, bio); 2196 2197 return DM_ENDIO_DONE; 2198 } 2199 2200 static void snapshot_merge_presuspend(struct dm_target *ti) 2201 { 2202 struct dm_snapshot *s = ti->private; 2203 2204 stop_merge(s); 2205 } 2206 2207 static int snapshot_preresume(struct dm_target *ti) 2208 { 2209 int r = 0; 2210 struct dm_snapshot *s = ti->private; 2211 struct dm_snapshot *snap_src = NULL, *snap_dest = NULL; 2212 2213 down_read(&_origins_lock); 2214 (void) __find_snapshots_sharing_cow(s, &snap_src, &snap_dest, NULL); 2215 if (snap_src && snap_dest) { 2216 down_read(&snap_src->lock); 2217 if (s == snap_src) { 2218 DMERR("Unable to resume snapshot source until " 2219 "handover completes."); 2220 r = -EINVAL; 2221 } else if (!dm_suspended(snap_src->ti)) { 2222 DMERR("Unable to perform snapshot handover until " 2223 "source is suspended."); 2224 r = -EINVAL; 2225 } 2226 up_read(&snap_src->lock); 2227 } 2228 up_read(&_origins_lock); 2229 2230 return r; 2231 } 2232 2233 static void snapshot_resume(struct dm_target *ti) 2234 { 2235 struct dm_snapshot *s = ti->private; 2236 struct dm_snapshot *snap_src = NULL, *snap_dest = NULL, *snap_merging = NULL; 2237 struct dm_origin *o; 2238 struct mapped_device *origin_md = NULL; 2239 bool must_restart_merging = false; 2240 2241 down_read(&_origins_lock); 2242 2243 o = __lookup_dm_origin(s->origin->bdev); 2244 if (o) 2245 origin_md = dm_table_get_md(o->ti->table); 2246 if (!origin_md) { 2247 (void) __find_snapshots_sharing_cow(s, NULL, NULL, &snap_merging); 2248 if (snap_merging) 2249 origin_md = dm_table_get_md(snap_merging->ti->table); 2250 } 2251 if (origin_md == dm_table_get_md(ti->table)) 2252 origin_md = NULL; 2253 if (origin_md) { 2254 if (dm_hold(origin_md)) 2255 origin_md = NULL; 2256 } 2257 2258 up_read(&_origins_lock); 2259 2260 if (origin_md) { 2261 dm_internal_suspend_fast(origin_md); 2262 if (snap_merging && test_bit(RUNNING_MERGE, &snap_merging->state_bits)) { 2263 must_restart_merging = true; 2264 stop_merge(snap_merging); 2265 } 2266 } 2267 2268 down_read(&_origins_lock); 2269 2270 (void) __find_snapshots_sharing_cow(s, &snap_src, &snap_dest, NULL); 2271 if (snap_src && snap_dest) { 2272 down_write(&snap_src->lock); 2273 down_write_nested(&snap_dest->lock, SINGLE_DEPTH_NESTING); 2274 __handover_exceptions(snap_src, snap_dest); 2275 up_write(&snap_dest->lock); 2276 up_write(&snap_src->lock); 2277 } 2278 2279 up_read(&_origins_lock); 2280 2281 if (origin_md) { 2282 if (must_restart_merging) 2283 start_merge(snap_merging); 2284 dm_internal_resume_fast(origin_md); 2285 dm_put(origin_md); 2286 } 2287 2288 /* Now we have correct chunk size, reregister */ 2289 reregister_snapshot(s); 2290 2291 down_write(&s->lock); 2292 s->active = 1; 2293 up_write(&s->lock); 2294 } 2295 2296 static uint32_t get_origin_minimum_chunksize(struct block_device *bdev) 2297 { 2298 uint32_t min_chunksize; 2299 2300 down_read(&_origins_lock); 2301 min_chunksize = __minimum_chunk_size(__lookup_origin(bdev)); 2302 up_read(&_origins_lock); 2303 2304 return min_chunksize; 2305 } 2306 2307 static void snapshot_merge_resume(struct dm_target *ti) 2308 { 2309 struct dm_snapshot *s = ti->private; 2310 2311 /* 2312 * Handover exceptions from existing snapshot. 2313 */ 2314 snapshot_resume(ti); 2315 2316 /* 2317 * snapshot-merge acts as an origin, so set ti->max_io_len 2318 */ 2319 ti->max_io_len = get_origin_minimum_chunksize(s->origin->bdev); 2320 2321 start_merge(s); 2322 } 2323 2324 static void snapshot_status(struct dm_target *ti, status_type_t type, 2325 unsigned int status_flags, char *result, unsigned int maxlen) 2326 { 2327 unsigned int sz = 0; 2328 struct dm_snapshot *snap = ti->private; 2329 unsigned int num_features; 2330 2331 switch (type) { 2332 case STATUSTYPE_INFO: 2333 2334 down_write(&snap->lock); 2335 2336 if (!snap->valid) 2337 DMEMIT("Invalid"); 2338 else if (snap->merge_failed) 2339 DMEMIT("Merge failed"); 2340 else if (snap->snapshot_overflowed) 2341 DMEMIT("Overflow"); 2342 else { 2343 if (snap->store->type->usage) { 2344 sector_t total_sectors, sectors_allocated, 2345 metadata_sectors; 2346 snap->store->type->usage(snap->store, 2347 &total_sectors, 2348 §ors_allocated, 2349 &metadata_sectors); 2350 DMEMIT("%llu/%llu %llu", 2351 (unsigned long long)sectors_allocated, 2352 (unsigned long long)total_sectors, 2353 (unsigned long long)metadata_sectors); 2354 } else 2355 DMEMIT("Unknown"); 2356 } 2357 2358 up_write(&snap->lock); 2359 2360 break; 2361 2362 case STATUSTYPE_TABLE: 2363 /* 2364 * kdevname returns a static pointer so we need 2365 * to make private copies if the output is to 2366 * make sense. 2367 */ 2368 DMEMIT("%s %s", snap->origin->name, snap->cow->name); 2369 sz += snap->store->type->status(snap->store, type, result + sz, 2370 maxlen - sz); 2371 num_features = snap->discard_zeroes_cow + snap->discard_passdown_origin; 2372 if (num_features) { 2373 DMEMIT(" %u", num_features); 2374 if (snap->discard_zeroes_cow) 2375 DMEMIT(" discard_zeroes_cow"); 2376 if (snap->discard_passdown_origin) 2377 DMEMIT(" discard_passdown_origin"); 2378 } 2379 break; 2380 2381 case STATUSTYPE_IMA: 2382 DMEMIT_TARGET_NAME_VERSION(ti->type); 2383 DMEMIT(",snap_origin_name=%s", snap->origin->name); 2384 DMEMIT(",snap_cow_name=%s", snap->cow->name); 2385 DMEMIT(",snap_valid=%c", snap->valid ? 'y' : 'n'); 2386 DMEMIT(",snap_merge_failed=%c", snap->merge_failed ? 'y' : 'n'); 2387 DMEMIT(",snapshot_overflowed=%c", snap->snapshot_overflowed ? 'y' : 'n'); 2388 DMEMIT(";"); 2389 break; 2390 } 2391 } 2392 2393 static int snapshot_iterate_devices(struct dm_target *ti, 2394 iterate_devices_callout_fn fn, void *data) 2395 { 2396 struct dm_snapshot *snap = ti->private; 2397 int r; 2398 2399 r = fn(ti, snap->origin, 0, ti->len, data); 2400 2401 if (!r) 2402 r = fn(ti, snap->cow, 0, get_dev_size(snap->cow->bdev), data); 2403 2404 return r; 2405 } 2406 2407 static void snapshot_io_hints(struct dm_target *ti, struct queue_limits *limits) 2408 { 2409 struct dm_snapshot *snap = ti->private; 2410 2411 if (snap->discard_zeroes_cow) { 2412 struct dm_snapshot *snap_src = NULL, *snap_dest = NULL; 2413 2414 down_read(&_origins_lock); 2415 2416 (void) __find_snapshots_sharing_cow(snap, &snap_src, &snap_dest, NULL); 2417 if (snap_src && snap_dest) 2418 snap = snap_src; 2419 2420 /* All discards are split on chunk_size boundary */ 2421 limits->discard_granularity = snap->store->chunk_size; 2422 limits->max_discard_sectors = snap->store->chunk_size; 2423 2424 up_read(&_origins_lock); 2425 } 2426 } 2427 2428 /* 2429 *--------------------------------------------------------------- 2430 * Origin methods 2431 *--------------------------------------------------------------- 2432 */ 2433 /* 2434 * If no exceptions need creating, DM_MAPIO_REMAPPED is returned and any 2435 * supplied bio was ignored. The caller may submit it immediately. 2436 * (No remapping actually occurs as the origin is always a direct linear 2437 * map.) 2438 * 2439 * If further exceptions are required, DM_MAPIO_SUBMITTED is returned 2440 * and any supplied bio is added to a list to be submitted once all 2441 * the necessary exceptions exist. 2442 */ 2443 static int __origin_write(struct list_head *snapshots, sector_t sector, 2444 struct bio *bio) 2445 { 2446 int r = DM_MAPIO_REMAPPED; 2447 struct dm_snapshot *snap; 2448 struct dm_exception *e; 2449 struct dm_snap_pending_exception *pe, *pe2; 2450 struct dm_snap_pending_exception *pe_to_start_now = NULL; 2451 struct dm_snap_pending_exception *pe_to_start_last = NULL; 2452 struct dm_exception_table_lock lock; 2453 chunk_t chunk; 2454 2455 /* Do all the snapshots on this origin */ 2456 list_for_each_entry(snap, snapshots, list) { 2457 /* 2458 * Don't make new exceptions in a merging snapshot 2459 * because it has effectively been deleted 2460 */ 2461 if (dm_target_is_snapshot_merge(snap->ti)) 2462 continue; 2463 2464 /* Nothing to do if writing beyond end of snapshot */ 2465 if (sector >= dm_table_get_size(snap->ti->table)) 2466 continue; 2467 2468 /* 2469 * Remember, different snapshots can have 2470 * different chunk sizes. 2471 */ 2472 chunk = sector_to_chunk(snap->store, sector); 2473 dm_exception_table_lock_init(snap, chunk, &lock); 2474 2475 down_read(&snap->lock); 2476 dm_exception_table_lock(&lock); 2477 2478 /* Only deal with valid and active snapshots */ 2479 if (!snap->valid || !snap->active) 2480 goto next_snapshot; 2481 2482 pe = __lookup_pending_exception(snap, chunk); 2483 if (!pe) { 2484 /* 2485 * Check exception table to see if block is already 2486 * remapped in this snapshot and trigger an exception 2487 * if not. 2488 */ 2489 e = dm_lookup_exception(&snap->complete, chunk); 2490 if (e) 2491 goto next_snapshot; 2492 2493 dm_exception_table_unlock(&lock); 2494 pe = alloc_pending_exception(snap); 2495 dm_exception_table_lock(&lock); 2496 2497 pe2 = __lookup_pending_exception(snap, chunk); 2498 2499 if (!pe2) { 2500 e = dm_lookup_exception(&snap->complete, chunk); 2501 if (e) { 2502 free_pending_exception(pe); 2503 goto next_snapshot; 2504 } 2505 2506 pe = __insert_pending_exception(snap, pe, chunk); 2507 if (!pe) { 2508 dm_exception_table_unlock(&lock); 2509 up_read(&snap->lock); 2510 2511 invalidate_snapshot(snap, -ENOMEM); 2512 continue; 2513 } 2514 } else { 2515 free_pending_exception(pe); 2516 pe = pe2; 2517 } 2518 } 2519 2520 r = DM_MAPIO_SUBMITTED; 2521 2522 /* 2523 * If an origin bio was supplied, queue it to wait for the 2524 * completion of this exception, and start this one last, 2525 * at the end of the function. 2526 */ 2527 if (bio) { 2528 bio_list_add(&pe->origin_bios, bio); 2529 bio = NULL; 2530 2531 if (!pe->started) { 2532 pe->started = 1; 2533 pe_to_start_last = pe; 2534 } 2535 } 2536 2537 if (!pe->started) { 2538 pe->started = 1; 2539 pe_to_start_now = pe; 2540 } 2541 2542 next_snapshot: 2543 dm_exception_table_unlock(&lock); 2544 up_read(&snap->lock); 2545 2546 if (pe_to_start_now) { 2547 start_copy(pe_to_start_now); 2548 pe_to_start_now = NULL; 2549 } 2550 } 2551 2552 /* 2553 * Submit the exception against which the bio is queued last, 2554 * to give the other exceptions a head start. 2555 */ 2556 if (pe_to_start_last) 2557 start_copy(pe_to_start_last); 2558 2559 return r; 2560 } 2561 2562 /* 2563 * Called on a write from the origin driver. 2564 */ 2565 static int do_origin(struct dm_dev *origin, struct bio *bio, bool limit) 2566 { 2567 struct origin *o; 2568 int r = DM_MAPIO_REMAPPED; 2569 2570 again: 2571 down_read(&_origins_lock); 2572 o = __lookup_origin(origin->bdev); 2573 if (o) { 2574 if (limit) { 2575 struct dm_snapshot *s; 2576 2577 list_for_each_entry(s, &o->snapshots, list) 2578 if (unlikely(!wait_for_in_progress(s, true))) 2579 goto again; 2580 } 2581 2582 r = __origin_write(&o->snapshots, bio->bi_iter.bi_sector, bio); 2583 } 2584 up_read(&_origins_lock); 2585 2586 return r; 2587 } 2588 2589 /* 2590 * Trigger exceptions in all non-merging snapshots. 2591 * 2592 * The chunk size of the merging snapshot may be larger than the chunk 2593 * size of some other snapshot so we may need to reallocate multiple 2594 * chunks in other snapshots. 2595 * 2596 * We scan all the overlapping exceptions in the other snapshots. 2597 * Returns 1 if anything was reallocated and must be waited for, 2598 * otherwise returns 0. 2599 * 2600 * size must be a multiple of merging_snap's chunk_size. 2601 */ 2602 static int origin_write_extent(struct dm_snapshot *merging_snap, 2603 sector_t sector, unsigned int size) 2604 { 2605 int must_wait = 0; 2606 sector_t n; 2607 struct origin *o; 2608 2609 /* 2610 * The origin's __minimum_chunk_size() got stored in max_io_len 2611 * by snapshot_merge_resume(). 2612 */ 2613 down_read(&_origins_lock); 2614 o = __lookup_origin(merging_snap->origin->bdev); 2615 for (n = 0; n < size; n += merging_snap->ti->max_io_len) 2616 if (__origin_write(&o->snapshots, sector + n, NULL) == 2617 DM_MAPIO_SUBMITTED) 2618 must_wait = 1; 2619 up_read(&_origins_lock); 2620 2621 return must_wait; 2622 } 2623 2624 /* 2625 * Origin: maps a linear range of a device, with hooks for snapshotting. 2626 */ 2627 2628 /* 2629 * Construct an origin mapping: <dev_path> 2630 * The context for an origin is merely a 'struct dm_dev *' 2631 * pointing to the real device. 2632 */ 2633 static int origin_ctr(struct dm_target *ti, unsigned int argc, char **argv) 2634 { 2635 int r; 2636 struct dm_origin *o; 2637 2638 if (argc != 1) { 2639 ti->error = "origin: incorrect number of arguments"; 2640 return -EINVAL; 2641 } 2642 2643 o = kmalloc(sizeof(struct dm_origin), GFP_KERNEL); 2644 if (!o) { 2645 ti->error = "Cannot allocate private origin structure"; 2646 r = -ENOMEM; 2647 goto bad_alloc; 2648 } 2649 2650 r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &o->dev); 2651 if (r) { 2652 ti->error = "Cannot get target device"; 2653 goto bad_open; 2654 } 2655 2656 o->ti = ti; 2657 ti->private = o; 2658 ti->num_flush_bios = 1; 2659 2660 return 0; 2661 2662 bad_open: 2663 kfree(o); 2664 bad_alloc: 2665 return r; 2666 } 2667 2668 static void origin_dtr(struct dm_target *ti) 2669 { 2670 struct dm_origin *o = ti->private; 2671 2672 dm_put_device(ti, o->dev); 2673 kfree(o); 2674 } 2675 2676 static int origin_map(struct dm_target *ti, struct bio *bio) 2677 { 2678 struct dm_origin *o = ti->private; 2679 unsigned int available_sectors; 2680 2681 bio_set_dev(bio, o->dev->bdev); 2682 2683 if (unlikely(bio->bi_opf & REQ_PREFLUSH)) 2684 return DM_MAPIO_REMAPPED; 2685 2686 if (bio_data_dir(bio) != WRITE) 2687 return DM_MAPIO_REMAPPED; 2688 2689 available_sectors = o->split_boundary - 2690 ((unsigned int)bio->bi_iter.bi_sector & (o->split_boundary - 1)); 2691 2692 if (bio_sectors(bio) > available_sectors) 2693 dm_accept_partial_bio(bio, available_sectors); 2694 2695 /* Only tell snapshots if this is a write */ 2696 return do_origin(o->dev, bio, true); 2697 } 2698 2699 /* 2700 * Set the target "max_io_len" field to the minimum of all the snapshots' 2701 * chunk sizes. 2702 */ 2703 static void origin_resume(struct dm_target *ti) 2704 { 2705 struct dm_origin *o = ti->private; 2706 2707 o->split_boundary = get_origin_minimum_chunksize(o->dev->bdev); 2708 2709 down_write(&_origins_lock); 2710 __insert_dm_origin(o); 2711 up_write(&_origins_lock); 2712 } 2713 2714 static void origin_postsuspend(struct dm_target *ti) 2715 { 2716 struct dm_origin *o = ti->private; 2717 2718 down_write(&_origins_lock); 2719 __remove_dm_origin(o); 2720 up_write(&_origins_lock); 2721 } 2722 2723 static void origin_status(struct dm_target *ti, status_type_t type, 2724 unsigned int status_flags, char *result, unsigned int maxlen) 2725 { 2726 struct dm_origin *o = ti->private; 2727 2728 switch (type) { 2729 case STATUSTYPE_INFO: 2730 result[0] = '\0'; 2731 break; 2732 2733 case STATUSTYPE_TABLE: 2734 snprintf(result, maxlen, "%s", o->dev->name); 2735 break; 2736 case STATUSTYPE_IMA: 2737 result[0] = '\0'; 2738 break; 2739 } 2740 } 2741 2742 static int origin_iterate_devices(struct dm_target *ti, 2743 iterate_devices_callout_fn fn, void *data) 2744 { 2745 struct dm_origin *o = ti->private; 2746 2747 return fn(ti, o->dev, 0, ti->len, data); 2748 } 2749 2750 static struct target_type origin_target = { 2751 .name = "snapshot-origin", 2752 .version = {1, 9, 0}, 2753 .module = THIS_MODULE, 2754 .ctr = origin_ctr, 2755 .dtr = origin_dtr, 2756 .map = origin_map, 2757 .resume = origin_resume, 2758 .postsuspend = origin_postsuspend, 2759 .status = origin_status, 2760 .iterate_devices = origin_iterate_devices, 2761 }; 2762 2763 static struct target_type snapshot_target = { 2764 .name = "snapshot", 2765 .version = {1, 16, 0}, 2766 .module = THIS_MODULE, 2767 .ctr = snapshot_ctr, 2768 .dtr = snapshot_dtr, 2769 .map = snapshot_map, 2770 .end_io = snapshot_end_io, 2771 .preresume = snapshot_preresume, 2772 .resume = snapshot_resume, 2773 .status = snapshot_status, 2774 .iterate_devices = snapshot_iterate_devices, 2775 .io_hints = snapshot_io_hints, 2776 }; 2777 2778 static struct target_type merge_target = { 2779 .name = dm_snapshot_merge_target_name, 2780 .version = {1, 5, 0}, 2781 .module = THIS_MODULE, 2782 .ctr = snapshot_ctr, 2783 .dtr = snapshot_dtr, 2784 .map = snapshot_merge_map, 2785 .end_io = snapshot_end_io, 2786 .presuspend = snapshot_merge_presuspend, 2787 .preresume = snapshot_preresume, 2788 .resume = snapshot_merge_resume, 2789 .status = snapshot_status, 2790 .iterate_devices = snapshot_iterate_devices, 2791 .io_hints = snapshot_io_hints, 2792 }; 2793 2794 static int __init dm_snapshot_init(void) 2795 { 2796 int r; 2797 2798 r = dm_exception_store_init(); 2799 if (r) { 2800 DMERR("Failed to initialize exception stores"); 2801 return r; 2802 } 2803 2804 r = init_origin_hash(); 2805 if (r) { 2806 DMERR("init_origin_hash failed."); 2807 goto bad_origin_hash; 2808 } 2809 2810 exception_cache = KMEM_CACHE(dm_exception, 0); 2811 if (!exception_cache) { 2812 DMERR("Couldn't create exception cache."); 2813 r = -ENOMEM; 2814 goto bad_exception_cache; 2815 } 2816 2817 pending_cache = KMEM_CACHE(dm_snap_pending_exception, 0); 2818 if (!pending_cache) { 2819 DMERR("Couldn't create pending cache."); 2820 r = -ENOMEM; 2821 goto bad_pending_cache; 2822 } 2823 2824 r = dm_register_target(&snapshot_target); 2825 if (r < 0) { 2826 DMERR("snapshot target register failed %d", r); 2827 goto bad_register_snapshot_target; 2828 } 2829 2830 r = dm_register_target(&origin_target); 2831 if (r < 0) { 2832 DMERR("Origin target register failed %d", r); 2833 goto bad_register_origin_target; 2834 } 2835 2836 r = dm_register_target(&merge_target); 2837 if (r < 0) { 2838 DMERR("Merge target register failed %d", r); 2839 goto bad_register_merge_target; 2840 } 2841 2842 return 0; 2843 2844 bad_register_merge_target: 2845 dm_unregister_target(&origin_target); 2846 bad_register_origin_target: 2847 dm_unregister_target(&snapshot_target); 2848 bad_register_snapshot_target: 2849 kmem_cache_destroy(pending_cache); 2850 bad_pending_cache: 2851 kmem_cache_destroy(exception_cache); 2852 bad_exception_cache: 2853 exit_origin_hash(); 2854 bad_origin_hash: 2855 dm_exception_store_exit(); 2856 2857 return r; 2858 } 2859 2860 static void __exit dm_snapshot_exit(void) 2861 { 2862 dm_unregister_target(&snapshot_target); 2863 dm_unregister_target(&origin_target); 2864 dm_unregister_target(&merge_target); 2865 2866 exit_origin_hash(); 2867 kmem_cache_destroy(pending_cache); 2868 kmem_cache_destroy(exception_cache); 2869 2870 dm_exception_store_exit(); 2871 } 2872 2873 /* Module hooks */ 2874 module_init(dm_snapshot_init); 2875 module_exit(dm_snapshot_exit); 2876 2877 MODULE_DESCRIPTION(DM_NAME " snapshot target"); 2878 MODULE_AUTHOR("Joe Thornber"); 2879 MODULE_LICENSE("GPL"); 2880 MODULE_ALIAS("dm-snapshot-origin"); 2881 MODULE_ALIAS("dm-snapshot-merge"); 2882