Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 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 "dm.h" |
| 10 | #include "dm-snap.h" |
| 11 | #include "dm-io.h" |
| 12 | #include "kcopyd.h" |
| 13 | |
| 14 | #include <linux/mm.h> |
| 15 | #include <linux/pagemap.h> |
| 16 | #include <linux/vmalloc.h> |
| 17 | #include <linux/slab.h> |
| 18 | |
| 19 | /*----------------------------------------------------------------- |
| 20 | * Persistent snapshots, by persistent we mean that the snapshot |
| 21 | * will survive a reboot. |
| 22 | *---------------------------------------------------------------*/ |
| 23 | |
| 24 | /* |
| 25 | * We need to store a record of which parts of the origin have |
| 26 | * been copied to the snapshot device. The snapshot code |
| 27 | * requires that we copy exception chunks to chunk aligned areas |
| 28 | * of the COW store. It makes sense therefore, to store the |
| 29 | * metadata in chunk size blocks. |
| 30 | * |
| 31 | * There is no backward or forward compatibility implemented, |
| 32 | * snapshots with different disk versions than the kernel will |
| 33 | * not be usable. It is expected that "lvcreate" will blank out |
| 34 | * the start of a fresh COW device before calling the snapshot |
| 35 | * constructor. |
| 36 | * |
| 37 | * The first chunk of the COW device just contains the header. |
| 38 | * After this there is a chunk filled with exception metadata, |
| 39 | * followed by as many exception chunks as can fit in the |
| 40 | * metadata areas. |
| 41 | * |
| 42 | * All on disk structures are in little-endian format. The end |
| 43 | * of the exceptions info is indicated by an exception with a |
| 44 | * new_chunk of 0, which is invalid since it would point to the |
| 45 | * header chunk. |
| 46 | */ |
| 47 | |
| 48 | /* |
| 49 | * Magic for persistent snapshots: "SnAp" - Feeble isn't it. |
| 50 | */ |
| 51 | #define SNAP_MAGIC 0x70416e53 |
| 52 | |
| 53 | /* |
| 54 | * The on-disk version of the metadata. |
| 55 | */ |
| 56 | #define SNAPSHOT_DISK_VERSION 1 |
| 57 | |
| 58 | struct disk_header { |
| 59 | uint32_t magic; |
| 60 | |
| 61 | /* |
| 62 | * Is this snapshot valid. There is no way of recovering |
| 63 | * an invalid snapshot. |
| 64 | */ |
| 65 | uint32_t valid; |
| 66 | |
| 67 | /* |
| 68 | * Simple, incrementing version. no backward |
| 69 | * compatibility. |
| 70 | */ |
| 71 | uint32_t version; |
| 72 | |
| 73 | /* In sectors */ |
| 74 | uint32_t chunk_size; |
| 75 | }; |
| 76 | |
| 77 | struct disk_exception { |
| 78 | uint64_t old_chunk; |
| 79 | uint64_t new_chunk; |
| 80 | }; |
| 81 | |
| 82 | struct commit_callback { |
| 83 | void (*callback)(void *, int success); |
| 84 | void *context; |
| 85 | }; |
| 86 | |
| 87 | /* |
| 88 | * The top level structure for a persistent exception store. |
| 89 | */ |
| 90 | struct pstore { |
| 91 | struct dm_snapshot *snap; /* up pointer to my snapshot */ |
| 92 | int version; |
| 93 | int valid; |
| 94 | uint32_t chunk_size; |
| 95 | uint32_t exceptions_per_area; |
| 96 | |
| 97 | /* |
| 98 | * Now that we have an asynchronous kcopyd there is no |
| 99 | * need for large chunk sizes, so it wont hurt to have a |
| 100 | * whole chunks worth of metadata in memory at once. |
| 101 | */ |
| 102 | void *area; |
| 103 | |
| 104 | /* |
| 105 | * Used to keep track of which metadata area the data in |
| 106 | * 'chunk' refers to. |
| 107 | */ |
| 108 | uint32_t current_area; |
| 109 | |
| 110 | /* |
| 111 | * The next free chunk for an exception. |
| 112 | */ |
| 113 | uint32_t next_free; |
| 114 | |
| 115 | /* |
| 116 | * The index of next free exception in the current |
| 117 | * metadata area. |
| 118 | */ |
| 119 | uint32_t current_committed; |
| 120 | |
| 121 | atomic_t pending_count; |
| 122 | uint32_t callback_count; |
| 123 | struct commit_callback *callbacks; |
| 124 | }; |
| 125 | |
| 126 | static inline unsigned int sectors_to_pages(unsigned int sectors) |
| 127 | { |
| 128 | return sectors / (PAGE_SIZE >> 9); |
| 129 | } |
| 130 | |
| 131 | static int alloc_area(struct pstore *ps) |
| 132 | { |
| 133 | int r = -ENOMEM; |
| 134 | size_t len; |
| 135 | |
| 136 | len = ps->chunk_size << SECTOR_SHIFT; |
| 137 | |
| 138 | /* |
| 139 | * Allocate the chunk_size block of memory that will hold |
| 140 | * a single metadata area. |
| 141 | */ |
| 142 | ps->area = vmalloc(len); |
| 143 | if (!ps->area) |
| 144 | return r; |
| 145 | |
| 146 | return 0; |
| 147 | } |
| 148 | |
| 149 | static void free_area(struct pstore *ps) |
| 150 | { |
| 151 | vfree(ps->area); |
| 152 | } |
| 153 | |
| 154 | /* |
| 155 | * Read or write a chunk aligned and sized block of data from a device. |
| 156 | */ |
| 157 | static int chunk_io(struct pstore *ps, uint32_t chunk, int rw) |
| 158 | { |
| 159 | struct io_region where; |
| 160 | unsigned long bits; |
| 161 | |
| 162 | where.bdev = ps->snap->cow->bdev; |
| 163 | where.sector = ps->chunk_size * chunk; |
| 164 | where.count = ps->chunk_size; |
| 165 | |
| 166 | return dm_io_sync_vm(1, &where, rw, ps->area, &bits); |
| 167 | } |
| 168 | |
| 169 | /* |
| 170 | * Read or write a metadata area. Remembering to skip the first |
| 171 | * chunk which holds the header. |
| 172 | */ |
| 173 | static int area_io(struct pstore *ps, uint32_t area, int rw) |
| 174 | { |
| 175 | int r; |
| 176 | uint32_t chunk; |
| 177 | |
| 178 | /* convert a metadata area index to a chunk index */ |
| 179 | chunk = 1 + ((ps->exceptions_per_area + 1) * area); |
| 180 | |
| 181 | r = chunk_io(ps, chunk, rw); |
| 182 | if (r) |
| 183 | return r; |
| 184 | |
| 185 | ps->current_area = area; |
| 186 | return 0; |
| 187 | } |
| 188 | |
| 189 | static int zero_area(struct pstore *ps, uint32_t area) |
| 190 | { |
| 191 | memset(ps->area, 0, ps->chunk_size << SECTOR_SHIFT); |
| 192 | return area_io(ps, area, WRITE); |
| 193 | } |
| 194 | |
| 195 | static int read_header(struct pstore *ps, int *new_snapshot) |
| 196 | { |
| 197 | int r; |
| 198 | struct disk_header *dh; |
| 199 | |
| 200 | r = chunk_io(ps, 0, READ); |
| 201 | if (r) |
| 202 | return r; |
| 203 | |
| 204 | dh = (struct disk_header *) ps->area; |
| 205 | |
| 206 | if (le32_to_cpu(dh->magic) == 0) { |
| 207 | *new_snapshot = 1; |
| 208 | |
| 209 | } else if (le32_to_cpu(dh->magic) == SNAP_MAGIC) { |
| 210 | *new_snapshot = 0; |
| 211 | ps->valid = le32_to_cpu(dh->valid); |
| 212 | ps->version = le32_to_cpu(dh->version); |
| 213 | ps->chunk_size = le32_to_cpu(dh->chunk_size); |
| 214 | |
| 215 | } else { |
| 216 | DMWARN("Invalid/corrupt snapshot"); |
| 217 | r = -ENXIO; |
| 218 | } |
| 219 | |
| 220 | return r; |
| 221 | } |
| 222 | |
| 223 | static int write_header(struct pstore *ps) |
| 224 | { |
| 225 | struct disk_header *dh; |
| 226 | |
| 227 | memset(ps->area, 0, ps->chunk_size << SECTOR_SHIFT); |
| 228 | |
| 229 | dh = (struct disk_header *) ps->area; |
| 230 | dh->magic = cpu_to_le32(SNAP_MAGIC); |
| 231 | dh->valid = cpu_to_le32(ps->valid); |
| 232 | dh->version = cpu_to_le32(ps->version); |
| 233 | dh->chunk_size = cpu_to_le32(ps->chunk_size); |
| 234 | |
| 235 | return chunk_io(ps, 0, WRITE); |
| 236 | } |
| 237 | |
| 238 | /* |
| 239 | * Access functions for the disk exceptions, these do the endian conversions. |
| 240 | */ |
| 241 | static struct disk_exception *get_exception(struct pstore *ps, uint32_t index) |
| 242 | { |
| 243 | if (index >= ps->exceptions_per_area) |
| 244 | return NULL; |
| 245 | |
| 246 | return ((struct disk_exception *) ps->area) + index; |
| 247 | } |
| 248 | |
| 249 | static int read_exception(struct pstore *ps, |
| 250 | uint32_t index, struct disk_exception *result) |
| 251 | { |
| 252 | struct disk_exception *e; |
| 253 | |
| 254 | e = get_exception(ps, index); |
| 255 | if (!e) |
| 256 | return -EINVAL; |
| 257 | |
| 258 | /* copy it */ |
| 259 | result->old_chunk = le64_to_cpu(e->old_chunk); |
| 260 | result->new_chunk = le64_to_cpu(e->new_chunk); |
| 261 | |
| 262 | return 0; |
| 263 | } |
| 264 | |
| 265 | static int write_exception(struct pstore *ps, |
| 266 | uint32_t index, struct disk_exception *de) |
| 267 | { |
| 268 | struct disk_exception *e; |
| 269 | |
| 270 | e = get_exception(ps, index); |
| 271 | if (!e) |
| 272 | return -EINVAL; |
| 273 | |
| 274 | /* copy it */ |
| 275 | e->old_chunk = cpu_to_le64(de->old_chunk); |
| 276 | e->new_chunk = cpu_to_le64(de->new_chunk); |
| 277 | |
| 278 | return 0; |
| 279 | } |
| 280 | |
| 281 | /* |
| 282 | * Registers the exceptions that are present in the current area. |
| 283 | * 'full' is filled in to indicate if the area has been |
| 284 | * filled. |
| 285 | */ |
| 286 | static int insert_exceptions(struct pstore *ps, int *full) |
| 287 | { |
| 288 | int r; |
| 289 | unsigned int i; |
| 290 | struct disk_exception de; |
| 291 | |
| 292 | /* presume the area is full */ |
| 293 | *full = 1; |
| 294 | |
| 295 | for (i = 0; i < ps->exceptions_per_area; i++) { |
| 296 | r = read_exception(ps, i, &de); |
| 297 | |
| 298 | if (r) |
| 299 | return r; |
| 300 | |
| 301 | /* |
| 302 | * If the new_chunk is pointing at the start of |
| 303 | * the COW device, where the first metadata area |
| 304 | * is we know that we've hit the end of the |
| 305 | * exceptions. Therefore the area is not full. |
| 306 | */ |
| 307 | if (de.new_chunk == 0LL) { |
| 308 | ps->current_committed = i; |
| 309 | *full = 0; |
| 310 | break; |
| 311 | } |
| 312 | |
| 313 | /* |
| 314 | * Keep track of the start of the free chunks. |
| 315 | */ |
| 316 | if (ps->next_free <= de.new_chunk) |
| 317 | ps->next_free = de.new_chunk + 1; |
| 318 | |
| 319 | /* |
| 320 | * Otherwise we add the exception to the snapshot. |
| 321 | */ |
| 322 | r = dm_add_exception(ps->snap, de.old_chunk, de.new_chunk); |
| 323 | if (r) |
| 324 | return r; |
| 325 | } |
| 326 | |
| 327 | return 0; |
| 328 | } |
| 329 | |
| 330 | static int read_exceptions(struct pstore *ps) |
| 331 | { |
| 332 | uint32_t area; |
| 333 | int r, full = 1; |
| 334 | |
| 335 | /* |
| 336 | * Keeping reading chunks and inserting exceptions until |
| 337 | * we find a partially full area. |
| 338 | */ |
| 339 | for (area = 0; full; area++) { |
| 340 | r = area_io(ps, area, READ); |
| 341 | if (r) |
| 342 | return r; |
| 343 | |
| 344 | r = insert_exceptions(ps, &full); |
| 345 | if (r) |
| 346 | return r; |
| 347 | } |
| 348 | |
| 349 | return 0; |
| 350 | } |
| 351 | |
| 352 | static inline struct pstore *get_info(struct exception_store *store) |
| 353 | { |
| 354 | return (struct pstore *) store->context; |
| 355 | } |
| 356 | |
| 357 | static void persistent_fraction_full(struct exception_store *store, |
| 358 | sector_t *numerator, sector_t *denominator) |
| 359 | { |
| 360 | *numerator = get_info(store)->next_free * store->snap->chunk_size; |
| 361 | *denominator = get_dev_size(store->snap->cow->bdev); |
| 362 | } |
| 363 | |
| 364 | static void persistent_destroy(struct exception_store *store) |
| 365 | { |
| 366 | struct pstore *ps = get_info(store); |
| 367 | |
| 368 | dm_io_put(sectors_to_pages(ps->chunk_size)); |
| 369 | vfree(ps->callbacks); |
| 370 | free_area(ps); |
| 371 | kfree(ps); |
| 372 | } |
| 373 | |
| 374 | static int persistent_read_metadata(struct exception_store *store) |
| 375 | { |
| 376 | int r, new_snapshot; |
| 377 | struct pstore *ps = get_info(store); |
| 378 | |
| 379 | /* |
| 380 | * Read the snapshot header. |
| 381 | */ |
| 382 | r = read_header(ps, &new_snapshot); |
| 383 | if (r) |
| 384 | return r; |
| 385 | |
| 386 | /* |
| 387 | * Do we need to setup a new snapshot ? |
| 388 | */ |
| 389 | if (new_snapshot) { |
| 390 | r = write_header(ps); |
| 391 | if (r) { |
| 392 | DMWARN("write_header failed"); |
| 393 | return r; |
| 394 | } |
| 395 | |
| 396 | r = zero_area(ps, 0); |
| 397 | if (r) { |
| 398 | DMWARN("zero_area(0) failed"); |
| 399 | return r; |
| 400 | } |
| 401 | |
| 402 | } else { |
| 403 | /* |
| 404 | * Sanity checks. |
| 405 | */ |
| 406 | if (!ps->valid) { |
| 407 | DMWARN("snapshot is marked invalid"); |
| 408 | return -EINVAL; |
| 409 | } |
| 410 | |
| 411 | if (ps->version != SNAPSHOT_DISK_VERSION) { |
| 412 | DMWARN("unable to handle snapshot disk version %d", |
| 413 | ps->version); |
| 414 | return -EINVAL; |
| 415 | } |
| 416 | |
| 417 | /* |
| 418 | * Read the metadata. |
| 419 | */ |
| 420 | r = read_exceptions(ps); |
| 421 | if (r) |
| 422 | return r; |
| 423 | } |
| 424 | |
| 425 | return 0; |
| 426 | } |
| 427 | |
| 428 | static int persistent_prepare(struct exception_store *store, |
| 429 | struct exception *e) |
| 430 | { |
| 431 | struct pstore *ps = get_info(store); |
| 432 | uint32_t stride; |
| 433 | sector_t size = get_dev_size(store->snap->cow->bdev); |
| 434 | |
| 435 | /* Is there enough room ? */ |
| 436 | if (size < ((ps->next_free + 1) * store->snap->chunk_size)) |
| 437 | return -ENOSPC; |
| 438 | |
| 439 | e->new_chunk = ps->next_free; |
| 440 | |
| 441 | /* |
| 442 | * Move onto the next free pending, making sure to take |
| 443 | * into account the location of the metadata chunks. |
| 444 | */ |
| 445 | stride = (ps->exceptions_per_area + 1); |
| 446 | if ((++ps->next_free % stride) == 1) |
| 447 | ps->next_free++; |
| 448 | |
| 449 | atomic_inc(&ps->pending_count); |
| 450 | return 0; |
| 451 | } |
| 452 | |
| 453 | static void persistent_commit(struct exception_store *store, |
| 454 | struct exception *e, |
| 455 | void (*callback) (void *, int success), |
| 456 | void *callback_context) |
| 457 | { |
| 458 | int r; |
| 459 | unsigned int i; |
| 460 | struct pstore *ps = get_info(store); |
| 461 | struct disk_exception de; |
| 462 | struct commit_callback *cb; |
| 463 | |
| 464 | de.old_chunk = e->old_chunk; |
| 465 | de.new_chunk = e->new_chunk; |
| 466 | write_exception(ps, ps->current_committed++, &de); |
| 467 | |
| 468 | /* |
| 469 | * Add the callback to the back of the array. This code |
| 470 | * is the only place where the callback array is |
| 471 | * manipulated, and we know that it will never be called |
| 472 | * multiple times concurrently. |
| 473 | */ |
| 474 | cb = ps->callbacks + ps->callback_count++; |
| 475 | cb->callback = callback; |
| 476 | cb->context = callback_context; |
| 477 | |
| 478 | /* |
| 479 | * If there are no more exceptions in flight, or we have |
| 480 | * filled this metadata area we commit the exceptions to |
| 481 | * disk. |
| 482 | */ |
| 483 | if (atomic_dec_and_test(&ps->pending_count) || |
| 484 | (ps->current_committed == ps->exceptions_per_area)) { |
| 485 | r = area_io(ps, ps->current_area, WRITE); |
| 486 | if (r) |
| 487 | ps->valid = 0; |
| 488 | |
| 489 | for (i = 0; i < ps->callback_count; i++) { |
| 490 | cb = ps->callbacks + i; |
| 491 | cb->callback(cb->context, r == 0 ? 1 : 0); |
| 492 | } |
| 493 | |
| 494 | ps->callback_count = 0; |
| 495 | } |
| 496 | |
| 497 | /* |
| 498 | * Have we completely filled the current area ? |
| 499 | */ |
| 500 | if (ps->current_committed == ps->exceptions_per_area) { |
| 501 | ps->current_committed = 0; |
| 502 | r = zero_area(ps, ps->current_area + 1); |
| 503 | if (r) |
| 504 | ps->valid = 0; |
| 505 | } |
| 506 | } |
| 507 | |
| 508 | static void persistent_drop(struct exception_store *store) |
| 509 | { |
| 510 | struct pstore *ps = get_info(store); |
| 511 | |
| 512 | ps->valid = 0; |
| 513 | if (write_header(ps)) |
| 514 | DMWARN("write header failed"); |
| 515 | } |
| 516 | |
| 517 | int dm_create_persistent(struct exception_store *store, uint32_t chunk_size) |
| 518 | { |
| 519 | int r; |
| 520 | struct pstore *ps; |
| 521 | |
| 522 | r = dm_io_get(sectors_to_pages(chunk_size)); |
| 523 | if (r) |
| 524 | return r; |
| 525 | |
| 526 | /* allocate the pstore */ |
| 527 | ps = kmalloc(sizeof(*ps), GFP_KERNEL); |
| 528 | if (!ps) { |
| 529 | r = -ENOMEM; |
| 530 | goto bad; |
| 531 | } |
| 532 | |
| 533 | ps->snap = store->snap; |
| 534 | ps->valid = 1; |
| 535 | ps->version = SNAPSHOT_DISK_VERSION; |
| 536 | ps->chunk_size = chunk_size; |
| 537 | ps->exceptions_per_area = (chunk_size << SECTOR_SHIFT) / |
| 538 | sizeof(struct disk_exception); |
| 539 | ps->next_free = 2; /* skipping the header and first area */ |
| 540 | ps->current_committed = 0; |
| 541 | |
| 542 | r = alloc_area(ps); |
| 543 | if (r) |
| 544 | goto bad; |
| 545 | |
| 546 | /* |
| 547 | * Allocate space for all the callbacks. |
| 548 | */ |
| 549 | ps->callback_count = 0; |
| 550 | atomic_set(&ps->pending_count, 0); |
| 551 | ps->callbacks = dm_vcalloc(ps->exceptions_per_area, |
| 552 | sizeof(*ps->callbacks)); |
| 553 | |
| 554 | if (!ps->callbacks) { |
| 555 | r = -ENOMEM; |
| 556 | goto bad; |
| 557 | } |
| 558 | |
| 559 | store->destroy = persistent_destroy; |
| 560 | store->read_metadata = persistent_read_metadata; |
| 561 | store->prepare_exception = persistent_prepare; |
| 562 | store->commit_exception = persistent_commit; |
| 563 | store->drop_snapshot = persistent_drop; |
| 564 | store->fraction_full = persistent_fraction_full; |
| 565 | store->context = ps; |
| 566 | |
| 567 | return 0; |
| 568 | |
| 569 | bad: |
| 570 | dm_io_put(sectors_to_pages(chunk_size)); |
Jesper Juhl | f9101210 | 2005-09-10 00:26:54 -0700 | [diff] [blame] | 571 | if (ps && ps->area) |
| 572 | free_area(ps); |
| 573 | kfree(ps); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 574 | return r; |
| 575 | } |
| 576 | |
| 577 | /*----------------------------------------------------------------- |
| 578 | * Implementation of the store for non-persistent snapshots. |
| 579 | *---------------------------------------------------------------*/ |
| 580 | struct transient_c { |
| 581 | sector_t next_free; |
| 582 | }; |
| 583 | |
| 584 | static void transient_destroy(struct exception_store *store) |
| 585 | { |
| 586 | kfree(store->context); |
| 587 | } |
| 588 | |
| 589 | static int transient_read_metadata(struct exception_store *store) |
| 590 | { |
| 591 | return 0; |
| 592 | } |
| 593 | |
| 594 | static int transient_prepare(struct exception_store *store, struct exception *e) |
| 595 | { |
| 596 | struct transient_c *tc = (struct transient_c *) store->context; |
| 597 | sector_t size = get_dev_size(store->snap->cow->bdev); |
| 598 | |
| 599 | if (size < (tc->next_free + store->snap->chunk_size)) |
| 600 | return -1; |
| 601 | |
| 602 | e->new_chunk = sector_to_chunk(store->snap, tc->next_free); |
| 603 | tc->next_free += store->snap->chunk_size; |
| 604 | |
| 605 | return 0; |
| 606 | } |
| 607 | |
| 608 | static void transient_commit(struct exception_store *store, |
| 609 | struct exception *e, |
| 610 | void (*callback) (void *, int success), |
| 611 | void *callback_context) |
| 612 | { |
| 613 | /* Just succeed */ |
| 614 | callback(callback_context, 1); |
| 615 | } |
| 616 | |
| 617 | static void transient_fraction_full(struct exception_store *store, |
| 618 | sector_t *numerator, sector_t *denominator) |
| 619 | { |
| 620 | *numerator = ((struct transient_c *) store->context)->next_free; |
| 621 | *denominator = get_dev_size(store->snap->cow->bdev); |
| 622 | } |
| 623 | |
| 624 | int dm_create_transient(struct exception_store *store, |
| 625 | struct dm_snapshot *s, int blocksize) |
| 626 | { |
| 627 | struct transient_c *tc; |
| 628 | |
| 629 | memset(store, 0, sizeof(*store)); |
| 630 | store->destroy = transient_destroy; |
| 631 | store->read_metadata = transient_read_metadata; |
| 632 | store->prepare_exception = transient_prepare; |
| 633 | store->commit_exception = transient_commit; |
| 634 | store->fraction_full = transient_fraction_full; |
| 635 | store->snap = s; |
| 636 | |
| 637 | tc = kmalloc(sizeof(struct transient_c), GFP_KERNEL); |
| 638 | if (!tc) |
| 639 | return -ENOMEM; |
| 640 | |
| 641 | tc->next_free = 0; |
| 642 | store->context = tc; |
| 643 | |
| 644 | return 0; |
| 645 | } |