1 /* 2 * Block driver for the QCOW version 2 format 3 * 4 * Copyright (c) 2004-2006 Fabrice Bellard 5 * 6 * Permission is hereby granted, free of charge, to any person obtaining a copy 7 * of this software and associated documentation files (the "Software"), to deal 8 * in the Software without restriction, including without limitation the rights 9 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 10 * copies of the Software, and to permit persons to whom the Software is 11 * furnished to do so, subject to the following conditions: 12 * 13 * The above copyright notice and this permission notice shall be included in 14 * all copies or substantial portions of the Software. 15 * 16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN 22 * THE SOFTWARE. 23 */ 24 25 #include "qemu-common.h" 26 #include "block/block_int.h" 27 #include "block/qcow2.h" 28 #include "qemu/range.h" 29 30 static int64_t alloc_clusters_noref(BlockDriverState *bs, uint64_t size); 31 static int QEMU_WARN_UNUSED_RESULT update_refcount(BlockDriverState *bs, 32 int64_t offset, int64_t length, uint64_t addend, 33 bool decrease, enum qcow2_discard_type type); 34 35 static uint64_t get_refcount_ro0(const void *refcount_array, uint64_t index); 36 static uint64_t get_refcount_ro1(const void *refcount_array, uint64_t index); 37 static uint64_t get_refcount_ro2(const void *refcount_array, uint64_t index); 38 static uint64_t get_refcount_ro3(const void *refcount_array, uint64_t index); 39 static uint64_t get_refcount_ro4(const void *refcount_array, uint64_t index); 40 static uint64_t get_refcount_ro5(const void *refcount_array, uint64_t index); 41 static uint64_t get_refcount_ro6(const void *refcount_array, uint64_t index); 42 43 static void set_refcount_ro0(void *refcount_array, uint64_t index, 44 uint64_t value); 45 static void set_refcount_ro1(void *refcount_array, uint64_t index, 46 uint64_t value); 47 static void set_refcount_ro2(void *refcount_array, uint64_t index, 48 uint64_t value); 49 static void set_refcount_ro3(void *refcount_array, uint64_t index, 50 uint64_t value); 51 static void set_refcount_ro4(void *refcount_array, uint64_t index, 52 uint64_t value); 53 static void set_refcount_ro5(void *refcount_array, uint64_t index, 54 uint64_t value); 55 static void set_refcount_ro6(void *refcount_array, uint64_t index, 56 uint64_t value); 57 58 59 static Qcow2GetRefcountFunc *const get_refcount_funcs[] = { 60 &get_refcount_ro0, 61 &get_refcount_ro1, 62 &get_refcount_ro2, 63 &get_refcount_ro3, 64 &get_refcount_ro4, 65 &get_refcount_ro5, 66 &get_refcount_ro6 67 }; 68 69 static Qcow2SetRefcountFunc *const set_refcount_funcs[] = { 70 &set_refcount_ro0, 71 &set_refcount_ro1, 72 &set_refcount_ro2, 73 &set_refcount_ro3, 74 &set_refcount_ro4, 75 &set_refcount_ro5, 76 &set_refcount_ro6 77 }; 78 79 80 /*********************************************************/ 81 /* refcount handling */ 82 83 int qcow2_refcount_init(BlockDriverState *bs) 84 { 85 BDRVQcow2State *s = bs->opaque; 86 unsigned int refcount_table_size2, i; 87 int ret; 88 89 assert(s->refcount_order >= 0 && s->refcount_order <= 6); 90 91 s->get_refcount = get_refcount_funcs[s->refcount_order]; 92 s->set_refcount = set_refcount_funcs[s->refcount_order]; 93 94 assert(s->refcount_table_size <= INT_MAX / sizeof(uint64_t)); 95 refcount_table_size2 = s->refcount_table_size * sizeof(uint64_t); 96 s->refcount_table = g_try_malloc(refcount_table_size2); 97 98 if (s->refcount_table_size > 0) { 99 if (s->refcount_table == NULL) { 100 ret = -ENOMEM; 101 goto fail; 102 } 103 BLKDBG_EVENT(bs->file, BLKDBG_REFTABLE_LOAD); 104 ret = bdrv_pread(bs->file, s->refcount_table_offset, 105 s->refcount_table, refcount_table_size2); 106 if (ret < 0) { 107 goto fail; 108 } 109 for(i = 0; i < s->refcount_table_size; i++) 110 be64_to_cpus(&s->refcount_table[i]); 111 } 112 return 0; 113 fail: 114 return ret; 115 } 116 117 void qcow2_refcount_close(BlockDriverState *bs) 118 { 119 BDRVQcow2State *s = bs->opaque; 120 g_free(s->refcount_table); 121 } 122 123 124 static uint64_t get_refcount_ro0(const void *refcount_array, uint64_t index) 125 { 126 return (((const uint8_t *)refcount_array)[index / 8] >> (index % 8)) & 0x1; 127 } 128 129 static void set_refcount_ro0(void *refcount_array, uint64_t index, 130 uint64_t value) 131 { 132 assert(!(value >> 1)); 133 ((uint8_t *)refcount_array)[index / 8] &= ~(0x1 << (index % 8)); 134 ((uint8_t *)refcount_array)[index / 8] |= value << (index % 8); 135 } 136 137 static uint64_t get_refcount_ro1(const void *refcount_array, uint64_t index) 138 { 139 return (((const uint8_t *)refcount_array)[index / 4] >> (2 * (index % 4))) 140 & 0x3; 141 } 142 143 static void set_refcount_ro1(void *refcount_array, uint64_t index, 144 uint64_t value) 145 { 146 assert(!(value >> 2)); 147 ((uint8_t *)refcount_array)[index / 4] &= ~(0x3 << (2 * (index % 4))); 148 ((uint8_t *)refcount_array)[index / 4] |= value << (2 * (index % 4)); 149 } 150 151 static uint64_t get_refcount_ro2(const void *refcount_array, uint64_t index) 152 { 153 return (((const uint8_t *)refcount_array)[index / 2] >> (4 * (index % 2))) 154 & 0xf; 155 } 156 157 static void set_refcount_ro2(void *refcount_array, uint64_t index, 158 uint64_t value) 159 { 160 assert(!(value >> 4)); 161 ((uint8_t *)refcount_array)[index / 2] &= ~(0xf << (4 * (index % 2))); 162 ((uint8_t *)refcount_array)[index / 2] |= value << (4 * (index % 2)); 163 } 164 165 static uint64_t get_refcount_ro3(const void *refcount_array, uint64_t index) 166 { 167 return ((const uint8_t *)refcount_array)[index]; 168 } 169 170 static void set_refcount_ro3(void *refcount_array, uint64_t index, 171 uint64_t value) 172 { 173 assert(!(value >> 8)); 174 ((uint8_t *)refcount_array)[index] = value; 175 } 176 177 static uint64_t get_refcount_ro4(const void *refcount_array, uint64_t index) 178 { 179 return be16_to_cpu(((const uint16_t *)refcount_array)[index]); 180 } 181 182 static void set_refcount_ro4(void *refcount_array, uint64_t index, 183 uint64_t value) 184 { 185 assert(!(value >> 16)); 186 ((uint16_t *)refcount_array)[index] = cpu_to_be16(value); 187 } 188 189 static uint64_t get_refcount_ro5(const void *refcount_array, uint64_t index) 190 { 191 return be32_to_cpu(((const uint32_t *)refcount_array)[index]); 192 } 193 194 static void set_refcount_ro5(void *refcount_array, uint64_t index, 195 uint64_t value) 196 { 197 assert(!(value >> 32)); 198 ((uint32_t *)refcount_array)[index] = cpu_to_be32(value); 199 } 200 201 static uint64_t get_refcount_ro6(const void *refcount_array, uint64_t index) 202 { 203 return be64_to_cpu(((const uint64_t *)refcount_array)[index]); 204 } 205 206 static void set_refcount_ro6(void *refcount_array, uint64_t index, 207 uint64_t value) 208 { 209 ((uint64_t *)refcount_array)[index] = cpu_to_be64(value); 210 } 211 212 213 static int load_refcount_block(BlockDriverState *bs, 214 int64_t refcount_block_offset, 215 void **refcount_block) 216 { 217 BDRVQcow2State *s = bs->opaque; 218 int ret; 219 220 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_LOAD); 221 ret = qcow2_cache_get(bs, s->refcount_block_cache, refcount_block_offset, 222 refcount_block); 223 224 return ret; 225 } 226 227 /* 228 * Retrieves the refcount of the cluster given by its index and stores it in 229 * *refcount. Returns 0 on success and -errno on failure. 230 */ 231 int qcow2_get_refcount(BlockDriverState *bs, int64_t cluster_index, 232 uint64_t *refcount) 233 { 234 BDRVQcow2State *s = bs->opaque; 235 uint64_t refcount_table_index, block_index; 236 int64_t refcount_block_offset; 237 int ret; 238 void *refcount_block; 239 240 refcount_table_index = cluster_index >> s->refcount_block_bits; 241 if (refcount_table_index >= s->refcount_table_size) { 242 *refcount = 0; 243 return 0; 244 } 245 refcount_block_offset = 246 s->refcount_table[refcount_table_index] & REFT_OFFSET_MASK; 247 if (!refcount_block_offset) { 248 *refcount = 0; 249 return 0; 250 } 251 252 if (offset_into_cluster(s, refcount_block_offset)) { 253 qcow2_signal_corruption(bs, true, -1, -1, "Refblock offset %#" PRIx64 254 " unaligned (reftable index: %#" PRIx64 ")", 255 refcount_block_offset, refcount_table_index); 256 return -EIO; 257 } 258 259 ret = qcow2_cache_get(bs, s->refcount_block_cache, refcount_block_offset, 260 &refcount_block); 261 if (ret < 0) { 262 return ret; 263 } 264 265 block_index = cluster_index & (s->refcount_block_size - 1); 266 *refcount = s->get_refcount(refcount_block, block_index); 267 268 qcow2_cache_put(bs, s->refcount_block_cache, &refcount_block); 269 270 return 0; 271 } 272 273 /* 274 * Rounds the refcount table size up to avoid growing the table for each single 275 * refcount block that is allocated. 276 */ 277 static unsigned int next_refcount_table_size(BDRVQcow2State *s, 278 unsigned int min_size) 279 { 280 unsigned int min_clusters = (min_size >> (s->cluster_bits - 3)) + 1; 281 unsigned int refcount_table_clusters = 282 MAX(1, s->refcount_table_size >> (s->cluster_bits - 3)); 283 284 while (min_clusters > refcount_table_clusters) { 285 refcount_table_clusters = (refcount_table_clusters * 3 + 1) / 2; 286 } 287 288 return refcount_table_clusters << (s->cluster_bits - 3); 289 } 290 291 292 /* Checks if two offsets are described by the same refcount block */ 293 static int in_same_refcount_block(BDRVQcow2State *s, uint64_t offset_a, 294 uint64_t offset_b) 295 { 296 uint64_t block_a = offset_a >> (s->cluster_bits + s->refcount_block_bits); 297 uint64_t block_b = offset_b >> (s->cluster_bits + s->refcount_block_bits); 298 299 return (block_a == block_b); 300 } 301 302 /* 303 * Loads a refcount block. If it doesn't exist yet, it is allocated first 304 * (including growing the refcount table if needed). 305 * 306 * Returns 0 on success or -errno in error case 307 */ 308 static int alloc_refcount_block(BlockDriverState *bs, 309 int64_t cluster_index, void **refcount_block) 310 { 311 BDRVQcow2State *s = bs->opaque; 312 unsigned int refcount_table_index; 313 int ret; 314 315 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC); 316 317 /* Find the refcount block for the given cluster */ 318 refcount_table_index = cluster_index >> s->refcount_block_bits; 319 320 if (refcount_table_index < s->refcount_table_size) { 321 322 uint64_t refcount_block_offset = 323 s->refcount_table[refcount_table_index] & REFT_OFFSET_MASK; 324 325 /* If it's already there, we're done */ 326 if (refcount_block_offset) { 327 if (offset_into_cluster(s, refcount_block_offset)) { 328 qcow2_signal_corruption(bs, true, -1, -1, "Refblock offset %#" 329 PRIx64 " unaligned (reftable index: " 330 "%#x)", refcount_block_offset, 331 refcount_table_index); 332 return -EIO; 333 } 334 335 return load_refcount_block(bs, refcount_block_offset, 336 refcount_block); 337 } 338 } 339 340 /* 341 * If we came here, we need to allocate something. Something is at least 342 * a cluster for the new refcount block. It may also include a new refcount 343 * table if the old refcount table is too small. 344 * 345 * Note that allocating clusters here needs some special care: 346 * 347 * - We can't use the normal qcow2_alloc_clusters(), it would try to 348 * increase the refcount and very likely we would end up with an endless 349 * recursion. Instead we must place the refcount blocks in a way that 350 * they can describe them themselves. 351 * 352 * - We need to consider that at this point we are inside update_refcounts 353 * and potentially doing an initial refcount increase. This means that 354 * some clusters have already been allocated by the caller, but their 355 * refcount isn't accurate yet. If we allocate clusters for metadata, we 356 * need to return -EAGAIN to signal the caller that it needs to restart 357 * the search for free clusters. 358 * 359 * - alloc_clusters_noref and qcow2_free_clusters may load a different 360 * refcount block into the cache 361 */ 362 363 *refcount_block = NULL; 364 365 /* We write to the refcount table, so we might depend on L2 tables */ 366 ret = qcow2_cache_flush(bs, s->l2_table_cache); 367 if (ret < 0) { 368 return ret; 369 } 370 371 /* Allocate the refcount block itself and mark it as used */ 372 int64_t new_block = alloc_clusters_noref(bs, s->cluster_size); 373 if (new_block < 0) { 374 return new_block; 375 } 376 377 #ifdef DEBUG_ALLOC2 378 fprintf(stderr, "qcow2: Allocate refcount block %d for %" PRIx64 379 " at %" PRIx64 "\n", 380 refcount_table_index, cluster_index << s->cluster_bits, new_block); 381 #endif 382 383 if (in_same_refcount_block(s, new_block, cluster_index << s->cluster_bits)) { 384 /* Zero the new refcount block before updating it */ 385 ret = qcow2_cache_get_empty(bs, s->refcount_block_cache, new_block, 386 refcount_block); 387 if (ret < 0) { 388 goto fail_block; 389 } 390 391 memset(*refcount_block, 0, s->cluster_size); 392 393 /* The block describes itself, need to update the cache */ 394 int block_index = (new_block >> s->cluster_bits) & 395 (s->refcount_block_size - 1); 396 s->set_refcount(*refcount_block, block_index, 1); 397 } else { 398 /* Described somewhere else. This can recurse at most twice before we 399 * arrive at a block that describes itself. */ 400 ret = update_refcount(bs, new_block, s->cluster_size, 1, false, 401 QCOW2_DISCARD_NEVER); 402 if (ret < 0) { 403 goto fail_block; 404 } 405 406 ret = qcow2_cache_flush(bs, s->refcount_block_cache); 407 if (ret < 0) { 408 goto fail_block; 409 } 410 411 /* Initialize the new refcount block only after updating its refcount, 412 * update_refcount uses the refcount cache itself */ 413 ret = qcow2_cache_get_empty(bs, s->refcount_block_cache, new_block, 414 refcount_block); 415 if (ret < 0) { 416 goto fail_block; 417 } 418 419 memset(*refcount_block, 0, s->cluster_size); 420 } 421 422 /* Now the new refcount block needs to be written to disk */ 423 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE); 424 qcow2_cache_entry_mark_dirty(bs, s->refcount_block_cache, *refcount_block); 425 ret = qcow2_cache_flush(bs, s->refcount_block_cache); 426 if (ret < 0) { 427 goto fail_block; 428 } 429 430 /* If the refcount table is big enough, just hook the block up there */ 431 if (refcount_table_index < s->refcount_table_size) { 432 uint64_t data64 = cpu_to_be64(new_block); 433 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_HOOKUP); 434 ret = bdrv_pwrite_sync(bs->file, 435 s->refcount_table_offset + refcount_table_index * sizeof(uint64_t), 436 &data64, sizeof(data64)); 437 if (ret < 0) { 438 goto fail_block; 439 } 440 441 s->refcount_table[refcount_table_index] = new_block; 442 443 /* The new refcount block may be where the caller intended to put its 444 * data, so let it restart the search. */ 445 return -EAGAIN; 446 } 447 448 qcow2_cache_put(bs, s->refcount_block_cache, refcount_block); 449 450 /* 451 * If we come here, we need to grow the refcount table. Again, a new 452 * refcount table needs some space and we can't simply allocate to avoid 453 * endless recursion. 454 * 455 * Therefore let's grab new refcount blocks at the end of the image, which 456 * will describe themselves and the new refcount table. This way we can 457 * reference them only in the new table and do the switch to the new 458 * refcount table at once without producing an inconsistent state in 459 * between. 460 */ 461 BLKDBG_EVENT(bs->file, BLKDBG_REFTABLE_GROW); 462 463 /* Calculate the number of refcount blocks needed so far; this will be the 464 * basis for calculating the index of the first cluster used for the 465 * self-describing refcount structures which we are about to create. 466 * 467 * Because we reached this point, there cannot be any refcount entries for 468 * cluster_index or higher indices yet. However, because new_block has been 469 * allocated to describe that cluster (and it will assume this role later 470 * on), we cannot use that index; also, new_block may actually have a higher 471 * cluster index than cluster_index, so it needs to be taken into account 472 * here (and 1 needs to be added to its value because that cluster is used). 473 */ 474 uint64_t blocks_used = DIV_ROUND_UP(MAX(cluster_index + 1, 475 (new_block >> s->cluster_bits) + 1), 476 s->refcount_block_size); 477 478 if (blocks_used > QCOW_MAX_REFTABLE_SIZE / sizeof(uint64_t)) { 479 return -EFBIG; 480 } 481 482 /* And now we need at least one block more for the new metadata */ 483 uint64_t table_size = next_refcount_table_size(s, blocks_used + 1); 484 uint64_t last_table_size; 485 uint64_t blocks_clusters; 486 do { 487 uint64_t table_clusters = 488 size_to_clusters(s, table_size * sizeof(uint64_t)); 489 blocks_clusters = 1 + 490 ((table_clusters + s->refcount_block_size - 1) 491 / s->refcount_block_size); 492 uint64_t meta_clusters = table_clusters + blocks_clusters; 493 494 last_table_size = table_size; 495 table_size = next_refcount_table_size(s, blocks_used + 496 ((meta_clusters + s->refcount_block_size - 1) 497 / s->refcount_block_size)); 498 499 } while (last_table_size != table_size); 500 501 #ifdef DEBUG_ALLOC2 502 fprintf(stderr, "qcow2: Grow refcount table %" PRId32 " => %" PRId64 "\n", 503 s->refcount_table_size, table_size); 504 #endif 505 506 /* Create the new refcount table and blocks */ 507 uint64_t meta_offset = (blocks_used * s->refcount_block_size) * 508 s->cluster_size; 509 uint64_t table_offset = meta_offset + blocks_clusters * s->cluster_size; 510 uint64_t *new_table = g_try_new0(uint64_t, table_size); 511 void *new_blocks = g_try_malloc0(blocks_clusters * s->cluster_size); 512 513 assert(table_size > 0 && blocks_clusters > 0); 514 if (new_table == NULL || new_blocks == NULL) { 515 ret = -ENOMEM; 516 goto fail_table; 517 } 518 519 /* Fill the new refcount table */ 520 memcpy(new_table, s->refcount_table, 521 s->refcount_table_size * sizeof(uint64_t)); 522 new_table[refcount_table_index] = new_block; 523 524 int i; 525 for (i = 0; i < blocks_clusters; i++) { 526 new_table[blocks_used + i] = meta_offset + (i * s->cluster_size); 527 } 528 529 /* Fill the refcount blocks */ 530 uint64_t table_clusters = size_to_clusters(s, table_size * sizeof(uint64_t)); 531 int block = 0; 532 for (i = 0; i < table_clusters + blocks_clusters; i++) { 533 s->set_refcount(new_blocks, block++, 1); 534 } 535 536 /* Write refcount blocks to disk */ 537 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE_BLOCKS); 538 ret = bdrv_pwrite_sync(bs->file, meta_offset, new_blocks, 539 blocks_clusters * s->cluster_size); 540 g_free(new_blocks); 541 new_blocks = NULL; 542 if (ret < 0) { 543 goto fail_table; 544 } 545 546 /* Write refcount table to disk */ 547 for(i = 0; i < table_size; i++) { 548 cpu_to_be64s(&new_table[i]); 549 } 550 551 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE_TABLE); 552 ret = bdrv_pwrite_sync(bs->file, table_offset, new_table, 553 table_size * sizeof(uint64_t)); 554 if (ret < 0) { 555 goto fail_table; 556 } 557 558 for(i = 0; i < table_size; i++) { 559 be64_to_cpus(&new_table[i]); 560 } 561 562 /* Hook up the new refcount table in the qcow2 header */ 563 uint8_t data[12]; 564 cpu_to_be64w((uint64_t*)data, table_offset); 565 cpu_to_be32w((uint32_t*)(data + 8), table_clusters); 566 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_SWITCH_TABLE); 567 ret = bdrv_pwrite_sync(bs->file, offsetof(QCowHeader, refcount_table_offset), 568 data, sizeof(data)); 569 if (ret < 0) { 570 goto fail_table; 571 } 572 573 /* And switch it in memory */ 574 uint64_t old_table_offset = s->refcount_table_offset; 575 uint64_t old_table_size = s->refcount_table_size; 576 577 g_free(s->refcount_table); 578 s->refcount_table = new_table; 579 s->refcount_table_size = table_size; 580 s->refcount_table_offset = table_offset; 581 582 /* Free old table. */ 583 qcow2_free_clusters(bs, old_table_offset, old_table_size * sizeof(uint64_t), 584 QCOW2_DISCARD_OTHER); 585 586 ret = load_refcount_block(bs, new_block, refcount_block); 587 if (ret < 0) { 588 return ret; 589 } 590 591 /* If we were trying to do the initial refcount update for some cluster 592 * allocation, we might have used the same clusters to store newly 593 * allocated metadata. Make the caller search some new space. */ 594 return -EAGAIN; 595 596 fail_table: 597 g_free(new_blocks); 598 g_free(new_table); 599 fail_block: 600 if (*refcount_block != NULL) { 601 qcow2_cache_put(bs, s->refcount_block_cache, refcount_block); 602 } 603 return ret; 604 } 605 606 void qcow2_process_discards(BlockDriverState *bs, int ret) 607 { 608 BDRVQcow2State *s = bs->opaque; 609 Qcow2DiscardRegion *d, *next; 610 611 QTAILQ_FOREACH_SAFE(d, &s->discards, next, next) { 612 QTAILQ_REMOVE(&s->discards, d, next); 613 614 /* Discard is optional, ignore the return value */ 615 if (ret >= 0) { 616 bdrv_discard(bs->file, 617 d->offset >> BDRV_SECTOR_BITS, 618 d->bytes >> BDRV_SECTOR_BITS); 619 } 620 621 g_free(d); 622 } 623 } 624 625 static void update_refcount_discard(BlockDriverState *bs, 626 uint64_t offset, uint64_t length) 627 { 628 BDRVQcow2State *s = bs->opaque; 629 Qcow2DiscardRegion *d, *p, *next; 630 631 QTAILQ_FOREACH(d, &s->discards, next) { 632 uint64_t new_start = MIN(offset, d->offset); 633 uint64_t new_end = MAX(offset + length, d->offset + d->bytes); 634 635 if (new_end - new_start <= length + d->bytes) { 636 /* There can't be any overlap, areas ending up here have no 637 * references any more and therefore shouldn't get freed another 638 * time. */ 639 assert(d->bytes + length == new_end - new_start); 640 d->offset = new_start; 641 d->bytes = new_end - new_start; 642 goto found; 643 } 644 } 645 646 d = g_malloc(sizeof(*d)); 647 *d = (Qcow2DiscardRegion) { 648 .bs = bs, 649 .offset = offset, 650 .bytes = length, 651 }; 652 QTAILQ_INSERT_TAIL(&s->discards, d, next); 653 654 found: 655 /* Merge discard requests if they are adjacent now */ 656 QTAILQ_FOREACH_SAFE(p, &s->discards, next, next) { 657 if (p == d 658 || p->offset > d->offset + d->bytes 659 || d->offset > p->offset + p->bytes) 660 { 661 continue; 662 } 663 664 /* Still no overlap possible */ 665 assert(p->offset == d->offset + d->bytes 666 || d->offset == p->offset + p->bytes); 667 668 QTAILQ_REMOVE(&s->discards, p, next); 669 d->offset = MIN(d->offset, p->offset); 670 d->bytes += p->bytes; 671 g_free(p); 672 } 673 } 674 675 /* XXX: cache several refcount block clusters ? */ 676 /* @addend is the absolute value of the addend; if @decrease is set, @addend 677 * will be subtracted from the current refcount, otherwise it will be added */ 678 static int QEMU_WARN_UNUSED_RESULT update_refcount(BlockDriverState *bs, 679 int64_t offset, 680 int64_t length, 681 uint64_t addend, 682 bool decrease, 683 enum qcow2_discard_type type) 684 { 685 BDRVQcow2State *s = bs->opaque; 686 int64_t start, last, cluster_offset; 687 void *refcount_block = NULL; 688 int64_t old_table_index = -1; 689 int ret; 690 691 #ifdef DEBUG_ALLOC2 692 fprintf(stderr, "update_refcount: offset=%" PRId64 " size=%" PRId64 693 " addend=%s%" PRIu64 "\n", offset, length, decrease ? "-" : "", 694 addend); 695 #endif 696 if (length < 0) { 697 return -EINVAL; 698 } else if (length == 0) { 699 return 0; 700 } 701 702 if (decrease) { 703 qcow2_cache_set_dependency(bs, s->refcount_block_cache, 704 s->l2_table_cache); 705 } 706 707 start = start_of_cluster(s, offset); 708 last = start_of_cluster(s, offset + length - 1); 709 for(cluster_offset = start; cluster_offset <= last; 710 cluster_offset += s->cluster_size) 711 { 712 int block_index; 713 uint64_t refcount; 714 int64_t cluster_index = cluster_offset >> s->cluster_bits; 715 int64_t table_index = cluster_index >> s->refcount_block_bits; 716 717 /* Load the refcount block and allocate it if needed */ 718 if (table_index != old_table_index) { 719 if (refcount_block) { 720 qcow2_cache_put(bs, s->refcount_block_cache, &refcount_block); 721 } 722 ret = alloc_refcount_block(bs, cluster_index, &refcount_block); 723 if (ret < 0) { 724 goto fail; 725 } 726 } 727 old_table_index = table_index; 728 729 qcow2_cache_entry_mark_dirty(bs, s->refcount_block_cache, 730 refcount_block); 731 732 /* we can update the count and save it */ 733 block_index = cluster_index & (s->refcount_block_size - 1); 734 735 refcount = s->get_refcount(refcount_block, block_index); 736 if (decrease ? (refcount - addend > refcount) 737 : (refcount + addend < refcount || 738 refcount + addend > s->refcount_max)) 739 { 740 ret = -EINVAL; 741 goto fail; 742 } 743 if (decrease) { 744 refcount -= addend; 745 } else { 746 refcount += addend; 747 } 748 if (refcount == 0 && cluster_index < s->free_cluster_index) { 749 s->free_cluster_index = cluster_index; 750 } 751 s->set_refcount(refcount_block, block_index, refcount); 752 753 if (refcount == 0 && s->discard_passthrough[type]) { 754 update_refcount_discard(bs, cluster_offset, s->cluster_size); 755 } 756 } 757 758 ret = 0; 759 fail: 760 if (!s->cache_discards) { 761 qcow2_process_discards(bs, ret); 762 } 763 764 /* Write last changed block to disk */ 765 if (refcount_block) { 766 qcow2_cache_put(bs, s->refcount_block_cache, &refcount_block); 767 } 768 769 /* 770 * Try do undo any updates if an error is returned (This may succeed in 771 * some cases like ENOSPC for allocating a new refcount block) 772 */ 773 if (ret < 0) { 774 int dummy; 775 dummy = update_refcount(bs, offset, cluster_offset - offset, addend, 776 !decrease, QCOW2_DISCARD_NEVER); 777 (void)dummy; 778 } 779 780 return ret; 781 } 782 783 /* 784 * Increases or decreases the refcount of a given cluster. 785 * 786 * @addend is the absolute value of the addend; if @decrease is set, @addend 787 * will be subtracted from the current refcount, otherwise it will be added. 788 * 789 * On success 0 is returned; on failure -errno is returned. 790 */ 791 int qcow2_update_cluster_refcount(BlockDriverState *bs, 792 int64_t cluster_index, 793 uint64_t addend, bool decrease, 794 enum qcow2_discard_type type) 795 { 796 BDRVQcow2State *s = bs->opaque; 797 int ret; 798 799 ret = update_refcount(bs, cluster_index << s->cluster_bits, 1, addend, 800 decrease, type); 801 if (ret < 0) { 802 return ret; 803 } 804 805 return 0; 806 } 807 808 809 810 /*********************************************************/ 811 /* cluster allocation functions */ 812 813 814 815 /* return < 0 if error */ 816 static int64_t alloc_clusters_noref(BlockDriverState *bs, uint64_t size) 817 { 818 BDRVQcow2State *s = bs->opaque; 819 uint64_t i, nb_clusters, refcount; 820 int ret; 821 822 /* We can't allocate clusters if they may still be queued for discard. */ 823 if (s->cache_discards) { 824 qcow2_process_discards(bs, 0); 825 } 826 827 nb_clusters = size_to_clusters(s, size); 828 retry: 829 for(i = 0; i < nb_clusters; i++) { 830 uint64_t next_cluster_index = s->free_cluster_index++; 831 ret = qcow2_get_refcount(bs, next_cluster_index, &refcount); 832 833 if (ret < 0) { 834 return ret; 835 } else if (refcount != 0) { 836 goto retry; 837 } 838 } 839 840 /* Make sure that all offsets in the "allocated" range are representable 841 * in an int64_t */ 842 if (s->free_cluster_index > 0 && 843 s->free_cluster_index - 1 > (INT64_MAX >> s->cluster_bits)) 844 { 845 return -EFBIG; 846 } 847 848 #ifdef DEBUG_ALLOC2 849 fprintf(stderr, "alloc_clusters: size=%" PRId64 " -> %" PRId64 "\n", 850 size, 851 (s->free_cluster_index - nb_clusters) << s->cluster_bits); 852 #endif 853 return (s->free_cluster_index - nb_clusters) << s->cluster_bits; 854 } 855 856 int64_t qcow2_alloc_clusters(BlockDriverState *bs, uint64_t size) 857 { 858 int64_t offset; 859 int ret; 860 861 BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_ALLOC); 862 do { 863 offset = alloc_clusters_noref(bs, size); 864 if (offset < 0) { 865 return offset; 866 } 867 868 ret = update_refcount(bs, offset, size, 1, false, QCOW2_DISCARD_NEVER); 869 } while (ret == -EAGAIN); 870 871 if (ret < 0) { 872 return ret; 873 } 874 875 return offset; 876 } 877 878 int64_t qcow2_alloc_clusters_at(BlockDriverState *bs, uint64_t offset, 879 int64_t nb_clusters) 880 { 881 BDRVQcow2State *s = bs->opaque; 882 uint64_t cluster_index, refcount; 883 uint64_t i; 884 int ret; 885 886 assert(nb_clusters >= 0); 887 if (nb_clusters == 0) { 888 return 0; 889 } 890 891 do { 892 /* Check how many clusters there are free */ 893 cluster_index = offset >> s->cluster_bits; 894 for(i = 0; i < nb_clusters; i++) { 895 ret = qcow2_get_refcount(bs, cluster_index++, &refcount); 896 if (ret < 0) { 897 return ret; 898 } else if (refcount != 0) { 899 break; 900 } 901 } 902 903 /* And then allocate them */ 904 ret = update_refcount(bs, offset, i << s->cluster_bits, 1, false, 905 QCOW2_DISCARD_NEVER); 906 } while (ret == -EAGAIN); 907 908 if (ret < 0) { 909 return ret; 910 } 911 912 return i; 913 } 914 915 /* only used to allocate compressed sectors. We try to allocate 916 contiguous sectors. size must be <= cluster_size */ 917 int64_t qcow2_alloc_bytes(BlockDriverState *bs, int size) 918 { 919 BDRVQcow2State *s = bs->opaque; 920 int64_t offset; 921 size_t free_in_cluster; 922 int ret; 923 924 BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_ALLOC_BYTES); 925 assert(size > 0 && size <= s->cluster_size); 926 assert(!s->free_byte_offset || offset_into_cluster(s, s->free_byte_offset)); 927 928 offset = s->free_byte_offset; 929 930 if (offset) { 931 uint64_t refcount; 932 ret = qcow2_get_refcount(bs, offset >> s->cluster_bits, &refcount); 933 if (ret < 0) { 934 return ret; 935 } 936 937 if (refcount == s->refcount_max) { 938 offset = 0; 939 } 940 } 941 942 free_in_cluster = s->cluster_size - offset_into_cluster(s, offset); 943 do { 944 if (!offset || free_in_cluster < size) { 945 int64_t new_cluster = alloc_clusters_noref(bs, s->cluster_size); 946 if (new_cluster < 0) { 947 return new_cluster; 948 } 949 950 if (!offset || ROUND_UP(offset, s->cluster_size) != new_cluster) { 951 offset = new_cluster; 952 free_in_cluster = s->cluster_size; 953 } else { 954 free_in_cluster += s->cluster_size; 955 } 956 } 957 958 assert(offset); 959 ret = update_refcount(bs, offset, size, 1, false, QCOW2_DISCARD_NEVER); 960 if (ret < 0) { 961 offset = 0; 962 } 963 } while (ret == -EAGAIN); 964 if (ret < 0) { 965 return ret; 966 } 967 968 /* The cluster refcount was incremented; refcount blocks must be flushed 969 * before the caller's L2 table updates. */ 970 qcow2_cache_set_dependency(bs, s->l2_table_cache, s->refcount_block_cache); 971 972 s->free_byte_offset = offset + size; 973 if (!offset_into_cluster(s, s->free_byte_offset)) { 974 s->free_byte_offset = 0; 975 } 976 977 return offset; 978 } 979 980 void qcow2_free_clusters(BlockDriverState *bs, 981 int64_t offset, int64_t size, 982 enum qcow2_discard_type type) 983 { 984 int ret; 985 986 BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_FREE); 987 ret = update_refcount(bs, offset, size, 1, true, type); 988 if (ret < 0) { 989 fprintf(stderr, "qcow2_free_clusters failed: %s\n", strerror(-ret)); 990 /* TODO Remember the clusters to free them later and avoid leaking */ 991 } 992 } 993 994 /* 995 * Free a cluster using its L2 entry (handles clusters of all types, e.g. 996 * normal cluster, compressed cluster, etc.) 997 */ 998 void qcow2_free_any_clusters(BlockDriverState *bs, uint64_t l2_entry, 999 int nb_clusters, enum qcow2_discard_type type) 1000 { 1001 BDRVQcow2State *s = bs->opaque; 1002 1003 switch (qcow2_get_cluster_type(l2_entry)) { 1004 case QCOW2_CLUSTER_COMPRESSED: 1005 { 1006 int nb_csectors; 1007 nb_csectors = ((l2_entry >> s->csize_shift) & 1008 s->csize_mask) + 1; 1009 qcow2_free_clusters(bs, 1010 (l2_entry & s->cluster_offset_mask) & ~511, 1011 nb_csectors * 512, type); 1012 } 1013 break; 1014 case QCOW2_CLUSTER_NORMAL: 1015 case QCOW2_CLUSTER_ZERO: 1016 if (l2_entry & L2E_OFFSET_MASK) { 1017 if (offset_into_cluster(s, l2_entry & L2E_OFFSET_MASK)) { 1018 qcow2_signal_corruption(bs, false, -1, -1, 1019 "Cannot free unaligned cluster %#llx", 1020 l2_entry & L2E_OFFSET_MASK); 1021 } else { 1022 qcow2_free_clusters(bs, l2_entry & L2E_OFFSET_MASK, 1023 nb_clusters << s->cluster_bits, type); 1024 } 1025 } 1026 break; 1027 case QCOW2_CLUSTER_UNALLOCATED: 1028 break; 1029 default: 1030 abort(); 1031 } 1032 } 1033 1034 1035 1036 /*********************************************************/ 1037 /* snapshots and image creation */ 1038 1039 1040 1041 /* update the refcounts of snapshots and the copied flag */ 1042 int qcow2_update_snapshot_refcount(BlockDriverState *bs, 1043 int64_t l1_table_offset, int l1_size, int addend) 1044 { 1045 BDRVQcow2State *s = bs->opaque; 1046 uint64_t *l1_table, *l2_table, l2_offset, offset, l1_size2, refcount; 1047 bool l1_allocated = false; 1048 int64_t old_offset, old_l2_offset; 1049 int i, j, l1_modified = 0, nb_csectors; 1050 int ret; 1051 1052 assert(addend >= -1 && addend <= 1); 1053 1054 l2_table = NULL; 1055 l1_table = NULL; 1056 l1_size2 = l1_size * sizeof(uint64_t); 1057 1058 s->cache_discards = true; 1059 1060 /* WARNING: qcow2_snapshot_goto relies on this function not using the 1061 * l1_table_offset when it is the current s->l1_table_offset! Be careful 1062 * when changing this! */ 1063 if (l1_table_offset != s->l1_table_offset) { 1064 l1_table = g_try_malloc0(align_offset(l1_size2, 512)); 1065 if (l1_size2 && l1_table == NULL) { 1066 ret = -ENOMEM; 1067 goto fail; 1068 } 1069 l1_allocated = true; 1070 1071 ret = bdrv_pread(bs->file, l1_table_offset, l1_table, l1_size2); 1072 if (ret < 0) { 1073 goto fail; 1074 } 1075 1076 for(i = 0;i < l1_size; i++) 1077 be64_to_cpus(&l1_table[i]); 1078 } else { 1079 assert(l1_size == s->l1_size); 1080 l1_table = s->l1_table; 1081 l1_allocated = false; 1082 } 1083 1084 for(i = 0; i < l1_size; i++) { 1085 l2_offset = l1_table[i]; 1086 if (l2_offset) { 1087 old_l2_offset = l2_offset; 1088 l2_offset &= L1E_OFFSET_MASK; 1089 1090 if (offset_into_cluster(s, l2_offset)) { 1091 qcow2_signal_corruption(bs, true, -1, -1, "L2 table offset %#" 1092 PRIx64 " unaligned (L1 index: %#x)", 1093 l2_offset, i); 1094 ret = -EIO; 1095 goto fail; 1096 } 1097 1098 ret = qcow2_cache_get(bs, s->l2_table_cache, l2_offset, 1099 (void**) &l2_table); 1100 if (ret < 0) { 1101 goto fail; 1102 } 1103 1104 for(j = 0; j < s->l2_size; j++) { 1105 uint64_t cluster_index; 1106 1107 offset = be64_to_cpu(l2_table[j]); 1108 old_offset = offset; 1109 offset &= ~QCOW_OFLAG_COPIED; 1110 1111 switch (qcow2_get_cluster_type(offset)) { 1112 case QCOW2_CLUSTER_COMPRESSED: 1113 nb_csectors = ((offset >> s->csize_shift) & 1114 s->csize_mask) + 1; 1115 if (addend != 0) { 1116 ret = update_refcount(bs, 1117 (offset & s->cluster_offset_mask) & ~511, 1118 nb_csectors * 512, abs(addend), addend < 0, 1119 QCOW2_DISCARD_SNAPSHOT); 1120 if (ret < 0) { 1121 goto fail; 1122 } 1123 } 1124 /* compressed clusters are never modified */ 1125 refcount = 2; 1126 break; 1127 1128 case QCOW2_CLUSTER_NORMAL: 1129 case QCOW2_CLUSTER_ZERO: 1130 if (offset_into_cluster(s, offset & L2E_OFFSET_MASK)) { 1131 qcow2_signal_corruption(bs, true, -1, -1, "Data " 1132 "cluster offset %#llx " 1133 "unaligned (L2 offset: %#" 1134 PRIx64 ", L2 index: %#x)", 1135 offset & L2E_OFFSET_MASK, 1136 l2_offset, j); 1137 ret = -EIO; 1138 goto fail; 1139 } 1140 1141 cluster_index = (offset & L2E_OFFSET_MASK) >> s->cluster_bits; 1142 if (!cluster_index) { 1143 /* unallocated */ 1144 refcount = 0; 1145 break; 1146 } 1147 if (addend != 0) { 1148 ret = qcow2_update_cluster_refcount(bs, 1149 cluster_index, abs(addend), addend < 0, 1150 QCOW2_DISCARD_SNAPSHOT); 1151 if (ret < 0) { 1152 goto fail; 1153 } 1154 } 1155 1156 ret = qcow2_get_refcount(bs, cluster_index, &refcount); 1157 if (ret < 0) { 1158 goto fail; 1159 } 1160 break; 1161 1162 case QCOW2_CLUSTER_UNALLOCATED: 1163 refcount = 0; 1164 break; 1165 1166 default: 1167 abort(); 1168 } 1169 1170 if (refcount == 1) { 1171 offset |= QCOW_OFLAG_COPIED; 1172 } 1173 if (offset != old_offset) { 1174 if (addend > 0) { 1175 qcow2_cache_set_dependency(bs, s->l2_table_cache, 1176 s->refcount_block_cache); 1177 } 1178 l2_table[j] = cpu_to_be64(offset); 1179 qcow2_cache_entry_mark_dirty(bs, s->l2_table_cache, 1180 l2_table); 1181 } 1182 } 1183 1184 qcow2_cache_put(bs, s->l2_table_cache, (void **) &l2_table); 1185 1186 if (addend != 0) { 1187 ret = qcow2_update_cluster_refcount(bs, l2_offset >> 1188 s->cluster_bits, 1189 abs(addend), addend < 0, 1190 QCOW2_DISCARD_SNAPSHOT); 1191 if (ret < 0) { 1192 goto fail; 1193 } 1194 } 1195 ret = qcow2_get_refcount(bs, l2_offset >> s->cluster_bits, 1196 &refcount); 1197 if (ret < 0) { 1198 goto fail; 1199 } else if (refcount == 1) { 1200 l2_offset |= QCOW_OFLAG_COPIED; 1201 } 1202 if (l2_offset != old_l2_offset) { 1203 l1_table[i] = l2_offset; 1204 l1_modified = 1; 1205 } 1206 } 1207 } 1208 1209 ret = bdrv_flush(bs); 1210 fail: 1211 if (l2_table) { 1212 qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table); 1213 } 1214 1215 s->cache_discards = false; 1216 qcow2_process_discards(bs, ret); 1217 1218 /* Update L1 only if it isn't deleted anyway (addend = -1) */ 1219 if (ret == 0 && addend >= 0 && l1_modified) { 1220 for (i = 0; i < l1_size; i++) { 1221 cpu_to_be64s(&l1_table[i]); 1222 } 1223 1224 ret = bdrv_pwrite_sync(bs->file, l1_table_offset, l1_table, l1_size2); 1225 1226 for (i = 0; i < l1_size; i++) { 1227 be64_to_cpus(&l1_table[i]); 1228 } 1229 } 1230 if (l1_allocated) 1231 g_free(l1_table); 1232 return ret; 1233 } 1234 1235 1236 1237 1238 /*********************************************************/ 1239 /* refcount checking functions */ 1240 1241 1242 static size_t refcount_array_byte_size(BDRVQcow2State *s, uint64_t entries) 1243 { 1244 /* This assertion holds because there is no way we can address more than 1245 * 2^(64 - 9) clusters at once (with cluster size 512 = 2^9, and because 1246 * offsets have to be representable in bytes); due to every cluster 1247 * corresponding to one refcount entry, we are well below that limit */ 1248 assert(entries < (UINT64_C(1) << (64 - 9))); 1249 1250 /* Thanks to the assertion this will not overflow, because 1251 * s->refcount_order < 7. 1252 * (note: x << s->refcount_order == x * s->refcount_bits) */ 1253 return DIV_ROUND_UP(entries << s->refcount_order, 8); 1254 } 1255 1256 /** 1257 * Reallocates *array so that it can hold new_size entries. *size must contain 1258 * the current number of entries in *array. If the reallocation fails, *array 1259 * and *size will not be modified and -errno will be returned. If the 1260 * reallocation is successful, *array will be set to the new buffer, *size 1261 * will be set to new_size and 0 will be returned. The size of the reallocated 1262 * refcount array buffer will be aligned to a cluster boundary, and the newly 1263 * allocated area will be zeroed. 1264 */ 1265 static int realloc_refcount_array(BDRVQcow2State *s, void **array, 1266 int64_t *size, int64_t new_size) 1267 { 1268 int64_t old_byte_size, new_byte_size; 1269 void *new_ptr; 1270 1271 /* Round to clusters so the array can be directly written to disk */ 1272 old_byte_size = size_to_clusters(s, refcount_array_byte_size(s, *size)) 1273 * s->cluster_size; 1274 new_byte_size = size_to_clusters(s, refcount_array_byte_size(s, new_size)) 1275 * s->cluster_size; 1276 1277 if (new_byte_size == old_byte_size) { 1278 *size = new_size; 1279 return 0; 1280 } 1281 1282 assert(new_byte_size > 0); 1283 1284 if (new_byte_size > SIZE_MAX) { 1285 return -ENOMEM; 1286 } 1287 1288 new_ptr = g_try_realloc(*array, new_byte_size); 1289 if (!new_ptr) { 1290 return -ENOMEM; 1291 } 1292 1293 if (new_byte_size > old_byte_size) { 1294 memset((char *)new_ptr + old_byte_size, 0, 1295 new_byte_size - old_byte_size); 1296 } 1297 1298 *array = new_ptr; 1299 *size = new_size; 1300 1301 return 0; 1302 } 1303 1304 /* 1305 * Increases the refcount for a range of clusters in a given refcount table. 1306 * This is used to construct a temporary refcount table out of L1 and L2 tables 1307 * which can be compared to the refcount table saved in the image. 1308 * 1309 * Modifies the number of errors in res. 1310 */ 1311 static int inc_refcounts(BlockDriverState *bs, 1312 BdrvCheckResult *res, 1313 void **refcount_table, 1314 int64_t *refcount_table_size, 1315 int64_t offset, int64_t size) 1316 { 1317 BDRVQcow2State *s = bs->opaque; 1318 uint64_t start, last, cluster_offset, k, refcount; 1319 int ret; 1320 1321 if (size <= 0) { 1322 return 0; 1323 } 1324 1325 start = start_of_cluster(s, offset); 1326 last = start_of_cluster(s, offset + size - 1); 1327 for(cluster_offset = start; cluster_offset <= last; 1328 cluster_offset += s->cluster_size) { 1329 k = cluster_offset >> s->cluster_bits; 1330 if (k >= *refcount_table_size) { 1331 ret = realloc_refcount_array(s, refcount_table, 1332 refcount_table_size, k + 1); 1333 if (ret < 0) { 1334 res->check_errors++; 1335 return ret; 1336 } 1337 } 1338 1339 refcount = s->get_refcount(*refcount_table, k); 1340 if (refcount == s->refcount_max) { 1341 fprintf(stderr, "ERROR: overflow cluster offset=0x%" PRIx64 1342 "\n", cluster_offset); 1343 res->corruptions++; 1344 continue; 1345 } 1346 s->set_refcount(*refcount_table, k, refcount + 1); 1347 } 1348 1349 return 0; 1350 } 1351 1352 /* Flags for check_refcounts_l1() and check_refcounts_l2() */ 1353 enum { 1354 CHECK_FRAG_INFO = 0x2, /* update BlockFragInfo counters */ 1355 }; 1356 1357 /* 1358 * Increases the refcount in the given refcount table for the all clusters 1359 * referenced in the L2 table. While doing so, performs some checks on L2 1360 * entries. 1361 * 1362 * Returns the number of errors found by the checks or -errno if an internal 1363 * error occurred. 1364 */ 1365 static int check_refcounts_l2(BlockDriverState *bs, BdrvCheckResult *res, 1366 void **refcount_table, 1367 int64_t *refcount_table_size, int64_t l2_offset, 1368 int flags) 1369 { 1370 BDRVQcow2State *s = bs->opaque; 1371 uint64_t *l2_table, l2_entry; 1372 uint64_t next_contiguous_offset = 0; 1373 int i, l2_size, nb_csectors, ret; 1374 1375 /* Read L2 table from disk */ 1376 l2_size = s->l2_size * sizeof(uint64_t); 1377 l2_table = g_malloc(l2_size); 1378 1379 ret = bdrv_pread(bs->file, l2_offset, l2_table, l2_size); 1380 if (ret < 0) { 1381 fprintf(stderr, "ERROR: I/O error in check_refcounts_l2\n"); 1382 res->check_errors++; 1383 goto fail; 1384 } 1385 1386 /* Do the actual checks */ 1387 for(i = 0; i < s->l2_size; i++) { 1388 l2_entry = be64_to_cpu(l2_table[i]); 1389 1390 switch (qcow2_get_cluster_type(l2_entry)) { 1391 case QCOW2_CLUSTER_COMPRESSED: 1392 /* Compressed clusters don't have QCOW_OFLAG_COPIED */ 1393 if (l2_entry & QCOW_OFLAG_COPIED) { 1394 fprintf(stderr, "ERROR: cluster %" PRId64 ": " 1395 "copied flag must never be set for compressed " 1396 "clusters\n", l2_entry >> s->cluster_bits); 1397 l2_entry &= ~QCOW_OFLAG_COPIED; 1398 res->corruptions++; 1399 } 1400 1401 /* Mark cluster as used */ 1402 nb_csectors = ((l2_entry >> s->csize_shift) & 1403 s->csize_mask) + 1; 1404 l2_entry &= s->cluster_offset_mask; 1405 ret = inc_refcounts(bs, res, refcount_table, refcount_table_size, 1406 l2_entry & ~511, nb_csectors * 512); 1407 if (ret < 0) { 1408 goto fail; 1409 } 1410 1411 if (flags & CHECK_FRAG_INFO) { 1412 res->bfi.allocated_clusters++; 1413 res->bfi.compressed_clusters++; 1414 1415 /* Compressed clusters are fragmented by nature. Since they 1416 * take up sub-sector space but we only have sector granularity 1417 * I/O we need to re-read the same sectors even for adjacent 1418 * compressed clusters. 1419 */ 1420 res->bfi.fragmented_clusters++; 1421 } 1422 break; 1423 1424 case QCOW2_CLUSTER_ZERO: 1425 if ((l2_entry & L2E_OFFSET_MASK) == 0) { 1426 break; 1427 } 1428 /* fall through */ 1429 1430 case QCOW2_CLUSTER_NORMAL: 1431 { 1432 uint64_t offset = l2_entry & L2E_OFFSET_MASK; 1433 1434 if (flags & CHECK_FRAG_INFO) { 1435 res->bfi.allocated_clusters++; 1436 if (next_contiguous_offset && 1437 offset != next_contiguous_offset) { 1438 res->bfi.fragmented_clusters++; 1439 } 1440 next_contiguous_offset = offset + s->cluster_size; 1441 } 1442 1443 /* Mark cluster as used */ 1444 ret = inc_refcounts(bs, res, refcount_table, refcount_table_size, 1445 offset, s->cluster_size); 1446 if (ret < 0) { 1447 goto fail; 1448 } 1449 1450 /* Correct offsets are cluster aligned */ 1451 if (offset_into_cluster(s, offset)) { 1452 fprintf(stderr, "ERROR offset=%" PRIx64 ": Cluster is not " 1453 "properly aligned; L2 entry corrupted.\n", offset); 1454 res->corruptions++; 1455 } 1456 break; 1457 } 1458 1459 case QCOW2_CLUSTER_UNALLOCATED: 1460 break; 1461 1462 default: 1463 abort(); 1464 } 1465 } 1466 1467 g_free(l2_table); 1468 return 0; 1469 1470 fail: 1471 g_free(l2_table); 1472 return ret; 1473 } 1474 1475 /* 1476 * Increases the refcount for the L1 table, its L2 tables and all referenced 1477 * clusters in the given refcount table. While doing so, performs some checks 1478 * on L1 and L2 entries. 1479 * 1480 * Returns the number of errors found by the checks or -errno if an internal 1481 * error occurred. 1482 */ 1483 static int check_refcounts_l1(BlockDriverState *bs, 1484 BdrvCheckResult *res, 1485 void **refcount_table, 1486 int64_t *refcount_table_size, 1487 int64_t l1_table_offset, int l1_size, 1488 int flags) 1489 { 1490 BDRVQcow2State *s = bs->opaque; 1491 uint64_t *l1_table = NULL, l2_offset, l1_size2; 1492 int i, ret; 1493 1494 l1_size2 = l1_size * sizeof(uint64_t); 1495 1496 /* Mark L1 table as used */ 1497 ret = inc_refcounts(bs, res, refcount_table, refcount_table_size, 1498 l1_table_offset, l1_size2); 1499 if (ret < 0) { 1500 goto fail; 1501 } 1502 1503 /* Read L1 table entries from disk */ 1504 if (l1_size2 > 0) { 1505 l1_table = g_try_malloc(l1_size2); 1506 if (l1_table == NULL) { 1507 ret = -ENOMEM; 1508 res->check_errors++; 1509 goto fail; 1510 } 1511 ret = bdrv_pread(bs->file, l1_table_offset, l1_table, l1_size2); 1512 if (ret < 0) { 1513 fprintf(stderr, "ERROR: I/O error in check_refcounts_l1\n"); 1514 res->check_errors++; 1515 goto fail; 1516 } 1517 for(i = 0;i < l1_size; i++) 1518 be64_to_cpus(&l1_table[i]); 1519 } 1520 1521 /* Do the actual checks */ 1522 for(i = 0; i < l1_size; i++) { 1523 l2_offset = l1_table[i]; 1524 if (l2_offset) { 1525 /* Mark L2 table as used */ 1526 l2_offset &= L1E_OFFSET_MASK; 1527 ret = inc_refcounts(bs, res, refcount_table, refcount_table_size, 1528 l2_offset, s->cluster_size); 1529 if (ret < 0) { 1530 goto fail; 1531 } 1532 1533 /* L2 tables are cluster aligned */ 1534 if (offset_into_cluster(s, l2_offset)) { 1535 fprintf(stderr, "ERROR l2_offset=%" PRIx64 ": Table is not " 1536 "cluster aligned; L1 entry corrupted\n", l2_offset); 1537 res->corruptions++; 1538 } 1539 1540 /* Process and check L2 entries */ 1541 ret = check_refcounts_l2(bs, res, refcount_table, 1542 refcount_table_size, l2_offset, flags); 1543 if (ret < 0) { 1544 goto fail; 1545 } 1546 } 1547 } 1548 g_free(l1_table); 1549 return 0; 1550 1551 fail: 1552 g_free(l1_table); 1553 return ret; 1554 } 1555 1556 /* 1557 * Checks the OFLAG_COPIED flag for all L1 and L2 entries. 1558 * 1559 * This function does not print an error message nor does it increment 1560 * check_errors if qcow2_get_refcount fails (this is because such an error will 1561 * have been already detected and sufficiently signaled by the calling function 1562 * (qcow2_check_refcounts) by the time this function is called). 1563 */ 1564 static int check_oflag_copied(BlockDriverState *bs, BdrvCheckResult *res, 1565 BdrvCheckMode fix) 1566 { 1567 BDRVQcow2State *s = bs->opaque; 1568 uint64_t *l2_table = qemu_blockalign(bs, s->cluster_size); 1569 int ret; 1570 uint64_t refcount; 1571 int i, j; 1572 1573 for (i = 0; i < s->l1_size; i++) { 1574 uint64_t l1_entry = s->l1_table[i]; 1575 uint64_t l2_offset = l1_entry & L1E_OFFSET_MASK; 1576 bool l2_dirty = false; 1577 1578 if (!l2_offset) { 1579 continue; 1580 } 1581 1582 ret = qcow2_get_refcount(bs, l2_offset >> s->cluster_bits, 1583 &refcount); 1584 if (ret < 0) { 1585 /* don't print message nor increment check_errors */ 1586 continue; 1587 } 1588 if ((refcount == 1) != ((l1_entry & QCOW_OFLAG_COPIED) != 0)) { 1589 fprintf(stderr, "%s OFLAG_COPIED L2 cluster: l1_index=%d " 1590 "l1_entry=%" PRIx64 " refcount=%" PRIu64 "\n", 1591 fix & BDRV_FIX_ERRORS ? "Repairing" : 1592 "ERROR", 1593 i, l1_entry, refcount); 1594 if (fix & BDRV_FIX_ERRORS) { 1595 s->l1_table[i] = refcount == 1 1596 ? l1_entry | QCOW_OFLAG_COPIED 1597 : l1_entry & ~QCOW_OFLAG_COPIED; 1598 ret = qcow2_write_l1_entry(bs, i); 1599 if (ret < 0) { 1600 res->check_errors++; 1601 goto fail; 1602 } 1603 res->corruptions_fixed++; 1604 } else { 1605 res->corruptions++; 1606 } 1607 } 1608 1609 ret = bdrv_pread(bs->file, l2_offset, l2_table, 1610 s->l2_size * sizeof(uint64_t)); 1611 if (ret < 0) { 1612 fprintf(stderr, "ERROR: Could not read L2 table: %s\n", 1613 strerror(-ret)); 1614 res->check_errors++; 1615 goto fail; 1616 } 1617 1618 for (j = 0; j < s->l2_size; j++) { 1619 uint64_t l2_entry = be64_to_cpu(l2_table[j]); 1620 uint64_t data_offset = l2_entry & L2E_OFFSET_MASK; 1621 int cluster_type = qcow2_get_cluster_type(l2_entry); 1622 1623 if ((cluster_type == QCOW2_CLUSTER_NORMAL) || 1624 ((cluster_type == QCOW2_CLUSTER_ZERO) && (data_offset != 0))) { 1625 ret = qcow2_get_refcount(bs, 1626 data_offset >> s->cluster_bits, 1627 &refcount); 1628 if (ret < 0) { 1629 /* don't print message nor increment check_errors */ 1630 continue; 1631 } 1632 if ((refcount == 1) != ((l2_entry & QCOW_OFLAG_COPIED) != 0)) { 1633 fprintf(stderr, "%s OFLAG_COPIED data cluster: " 1634 "l2_entry=%" PRIx64 " refcount=%" PRIu64 "\n", 1635 fix & BDRV_FIX_ERRORS ? "Repairing" : 1636 "ERROR", 1637 l2_entry, refcount); 1638 if (fix & BDRV_FIX_ERRORS) { 1639 l2_table[j] = cpu_to_be64(refcount == 1 1640 ? l2_entry | QCOW_OFLAG_COPIED 1641 : l2_entry & ~QCOW_OFLAG_COPIED); 1642 l2_dirty = true; 1643 res->corruptions_fixed++; 1644 } else { 1645 res->corruptions++; 1646 } 1647 } 1648 } 1649 } 1650 1651 if (l2_dirty) { 1652 ret = qcow2_pre_write_overlap_check(bs, QCOW2_OL_ACTIVE_L2, 1653 l2_offset, s->cluster_size); 1654 if (ret < 0) { 1655 fprintf(stderr, "ERROR: Could not write L2 table; metadata " 1656 "overlap check failed: %s\n", strerror(-ret)); 1657 res->check_errors++; 1658 goto fail; 1659 } 1660 1661 ret = bdrv_pwrite(bs->file, l2_offset, l2_table, s->cluster_size); 1662 if (ret < 0) { 1663 fprintf(stderr, "ERROR: Could not write L2 table: %s\n", 1664 strerror(-ret)); 1665 res->check_errors++; 1666 goto fail; 1667 } 1668 } 1669 } 1670 1671 ret = 0; 1672 1673 fail: 1674 qemu_vfree(l2_table); 1675 return ret; 1676 } 1677 1678 /* 1679 * Checks consistency of refblocks and accounts for each refblock in 1680 * *refcount_table. 1681 */ 1682 static int check_refblocks(BlockDriverState *bs, BdrvCheckResult *res, 1683 BdrvCheckMode fix, bool *rebuild, 1684 void **refcount_table, int64_t *nb_clusters) 1685 { 1686 BDRVQcow2State *s = bs->opaque; 1687 int64_t i, size; 1688 int ret; 1689 1690 for(i = 0; i < s->refcount_table_size; i++) { 1691 uint64_t offset, cluster; 1692 offset = s->refcount_table[i]; 1693 cluster = offset >> s->cluster_bits; 1694 1695 /* Refcount blocks are cluster aligned */ 1696 if (offset_into_cluster(s, offset)) { 1697 fprintf(stderr, "ERROR refcount block %" PRId64 " is not " 1698 "cluster aligned; refcount table entry corrupted\n", i); 1699 res->corruptions++; 1700 *rebuild = true; 1701 continue; 1702 } 1703 1704 if (cluster >= *nb_clusters) { 1705 fprintf(stderr, "%s refcount block %" PRId64 " is outside image\n", 1706 fix & BDRV_FIX_ERRORS ? "Repairing" : "ERROR", i); 1707 1708 if (fix & BDRV_FIX_ERRORS) { 1709 int64_t new_nb_clusters; 1710 1711 if (offset > INT64_MAX - s->cluster_size) { 1712 ret = -EINVAL; 1713 goto resize_fail; 1714 } 1715 1716 ret = bdrv_truncate(bs->file, offset + s->cluster_size); 1717 if (ret < 0) { 1718 goto resize_fail; 1719 } 1720 size = bdrv_getlength(bs->file); 1721 if (size < 0) { 1722 ret = size; 1723 goto resize_fail; 1724 } 1725 1726 new_nb_clusters = size_to_clusters(s, size); 1727 assert(new_nb_clusters >= *nb_clusters); 1728 1729 ret = realloc_refcount_array(s, refcount_table, 1730 nb_clusters, new_nb_clusters); 1731 if (ret < 0) { 1732 res->check_errors++; 1733 return ret; 1734 } 1735 1736 if (cluster >= *nb_clusters) { 1737 ret = -EINVAL; 1738 goto resize_fail; 1739 } 1740 1741 res->corruptions_fixed++; 1742 ret = inc_refcounts(bs, res, refcount_table, nb_clusters, 1743 offset, s->cluster_size); 1744 if (ret < 0) { 1745 return ret; 1746 } 1747 /* No need to check whether the refcount is now greater than 1: 1748 * This area was just allocated and zeroed, so it can only be 1749 * exactly 1 after inc_refcounts() */ 1750 continue; 1751 1752 resize_fail: 1753 res->corruptions++; 1754 *rebuild = true; 1755 fprintf(stderr, "ERROR could not resize image: %s\n", 1756 strerror(-ret)); 1757 } else { 1758 res->corruptions++; 1759 } 1760 continue; 1761 } 1762 1763 if (offset != 0) { 1764 ret = inc_refcounts(bs, res, refcount_table, nb_clusters, 1765 offset, s->cluster_size); 1766 if (ret < 0) { 1767 return ret; 1768 } 1769 if (s->get_refcount(*refcount_table, cluster) != 1) { 1770 fprintf(stderr, "ERROR refcount block %" PRId64 1771 " refcount=%" PRIu64 "\n", i, 1772 s->get_refcount(*refcount_table, cluster)); 1773 res->corruptions++; 1774 *rebuild = true; 1775 } 1776 } 1777 } 1778 1779 return 0; 1780 } 1781 1782 /* 1783 * Calculates an in-memory refcount table. 1784 */ 1785 static int calculate_refcounts(BlockDriverState *bs, BdrvCheckResult *res, 1786 BdrvCheckMode fix, bool *rebuild, 1787 void **refcount_table, int64_t *nb_clusters) 1788 { 1789 BDRVQcow2State *s = bs->opaque; 1790 int64_t i; 1791 QCowSnapshot *sn; 1792 int ret; 1793 1794 if (!*refcount_table) { 1795 int64_t old_size = 0; 1796 ret = realloc_refcount_array(s, refcount_table, 1797 &old_size, *nb_clusters); 1798 if (ret < 0) { 1799 res->check_errors++; 1800 return ret; 1801 } 1802 } 1803 1804 /* header */ 1805 ret = inc_refcounts(bs, res, refcount_table, nb_clusters, 1806 0, s->cluster_size); 1807 if (ret < 0) { 1808 return ret; 1809 } 1810 1811 /* current L1 table */ 1812 ret = check_refcounts_l1(bs, res, refcount_table, nb_clusters, 1813 s->l1_table_offset, s->l1_size, CHECK_FRAG_INFO); 1814 if (ret < 0) { 1815 return ret; 1816 } 1817 1818 /* snapshots */ 1819 for (i = 0; i < s->nb_snapshots; i++) { 1820 sn = s->snapshots + i; 1821 ret = check_refcounts_l1(bs, res, refcount_table, nb_clusters, 1822 sn->l1_table_offset, sn->l1_size, 0); 1823 if (ret < 0) { 1824 return ret; 1825 } 1826 } 1827 ret = inc_refcounts(bs, res, refcount_table, nb_clusters, 1828 s->snapshots_offset, s->snapshots_size); 1829 if (ret < 0) { 1830 return ret; 1831 } 1832 1833 /* refcount data */ 1834 ret = inc_refcounts(bs, res, refcount_table, nb_clusters, 1835 s->refcount_table_offset, 1836 s->refcount_table_size * sizeof(uint64_t)); 1837 if (ret < 0) { 1838 return ret; 1839 } 1840 1841 return check_refblocks(bs, res, fix, rebuild, refcount_table, nb_clusters); 1842 } 1843 1844 /* 1845 * Compares the actual reference count for each cluster in the image against the 1846 * refcount as reported by the refcount structures on-disk. 1847 */ 1848 static void compare_refcounts(BlockDriverState *bs, BdrvCheckResult *res, 1849 BdrvCheckMode fix, bool *rebuild, 1850 int64_t *highest_cluster, 1851 void *refcount_table, int64_t nb_clusters) 1852 { 1853 BDRVQcow2State *s = bs->opaque; 1854 int64_t i; 1855 uint64_t refcount1, refcount2; 1856 int ret; 1857 1858 for (i = 0, *highest_cluster = 0; i < nb_clusters; i++) { 1859 ret = qcow2_get_refcount(bs, i, &refcount1); 1860 if (ret < 0) { 1861 fprintf(stderr, "Can't get refcount for cluster %" PRId64 ": %s\n", 1862 i, strerror(-ret)); 1863 res->check_errors++; 1864 continue; 1865 } 1866 1867 refcount2 = s->get_refcount(refcount_table, i); 1868 1869 if (refcount1 > 0 || refcount2 > 0) { 1870 *highest_cluster = i; 1871 } 1872 1873 if (refcount1 != refcount2) { 1874 /* Check if we're allowed to fix the mismatch */ 1875 int *num_fixed = NULL; 1876 if (refcount1 == 0) { 1877 *rebuild = true; 1878 } else if (refcount1 > refcount2 && (fix & BDRV_FIX_LEAKS)) { 1879 num_fixed = &res->leaks_fixed; 1880 } else if (refcount1 < refcount2 && (fix & BDRV_FIX_ERRORS)) { 1881 num_fixed = &res->corruptions_fixed; 1882 } 1883 1884 fprintf(stderr, "%s cluster %" PRId64 " refcount=%" PRIu64 1885 " reference=%" PRIu64 "\n", 1886 num_fixed != NULL ? "Repairing" : 1887 refcount1 < refcount2 ? "ERROR" : 1888 "Leaked", 1889 i, refcount1, refcount2); 1890 1891 if (num_fixed) { 1892 ret = update_refcount(bs, i << s->cluster_bits, 1, 1893 refcount_diff(refcount1, refcount2), 1894 refcount1 > refcount2, 1895 QCOW2_DISCARD_ALWAYS); 1896 if (ret >= 0) { 1897 (*num_fixed)++; 1898 continue; 1899 } 1900 } 1901 1902 /* And if we couldn't, print an error */ 1903 if (refcount1 < refcount2) { 1904 res->corruptions++; 1905 } else { 1906 res->leaks++; 1907 } 1908 } 1909 } 1910 } 1911 1912 /* 1913 * Allocates clusters using an in-memory refcount table (IMRT) in contrast to 1914 * the on-disk refcount structures. 1915 * 1916 * On input, *first_free_cluster tells where to start looking, and need not 1917 * actually be a free cluster; the returned offset will not be before that 1918 * cluster. On output, *first_free_cluster points to the first gap found, even 1919 * if that gap was too small to be used as the returned offset. 1920 * 1921 * Note that *first_free_cluster is a cluster index whereas the return value is 1922 * an offset. 1923 */ 1924 static int64_t alloc_clusters_imrt(BlockDriverState *bs, 1925 int cluster_count, 1926 void **refcount_table, 1927 int64_t *imrt_nb_clusters, 1928 int64_t *first_free_cluster) 1929 { 1930 BDRVQcow2State *s = bs->opaque; 1931 int64_t cluster = *first_free_cluster, i; 1932 bool first_gap = true; 1933 int contiguous_free_clusters; 1934 int ret; 1935 1936 /* Starting at *first_free_cluster, find a range of at least cluster_count 1937 * continuously free clusters */ 1938 for (contiguous_free_clusters = 0; 1939 cluster < *imrt_nb_clusters && 1940 contiguous_free_clusters < cluster_count; 1941 cluster++) 1942 { 1943 if (!s->get_refcount(*refcount_table, cluster)) { 1944 contiguous_free_clusters++; 1945 if (first_gap) { 1946 /* If this is the first free cluster found, update 1947 * *first_free_cluster accordingly */ 1948 *first_free_cluster = cluster; 1949 first_gap = false; 1950 } 1951 } else if (contiguous_free_clusters) { 1952 contiguous_free_clusters = 0; 1953 } 1954 } 1955 1956 /* If contiguous_free_clusters is greater than zero, it contains the number 1957 * of continuously free clusters until the current cluster; the first free 1958 * cluster in the current "gap" is therefore 1959 * cluster - contiguous_free_clusters */ 1960 1961 /* If no such range could be found, grow the in-memory refcount table 1962 * accordingly to append free clusters at the end of the image */ 1963 if (contiguous_free_clusters < cluster_count) { 1964 /* contiguous_free_clusters clusters are already empty at the image end; 1965 * we need cluster_count clusters; therefore, we have to allocate 1966 * cluster_count - contiguous_free_clusters new clusters at the end of 1967 * the image (which is the current value of cluster; note that cluster 1968 * may exceed old_imrt_nb_clusters if *first_free_cluster pointed beyond 1969 * the image end) */ 1970 ret = realloc_refcount_array(s, refcount_table, imrt_nb_clusters, 1971 cluster + cluster_count 1972 - contiguous_free_clusters); 1973 if (ret < 0) { 1974 return ret; 1975 } 1976 } 1977 1978 /* Go back to the first free cluster */ 1979 cluster -= contiguous_free_clusters; 1980 for (i = 0; i < cluster_count; i++) { 1981 s->set_refcount(*refcount_table, cluster + i, 1); 1982 } 1983 1984 return cluster << s->cluster_bits; 1985 } 1986 1987 /* 1988 * Creates a new refcount structure based solely on the in-memory information 1989 * given through *refcount_table. All necessary allocations will be reflected 1990 * in that array. 1991 * 1992 * On success, the old refcount structure is leaked (it will be covered by the 1993 * new refcount structure). 1994 */ 1995 static int rebuild_refcount_structure(BlockDriverState *bs, 1996 BdrvCheckResult *res, 1997 void **refcount_table, 1998 int64_t *nb_clusters) 1999 { 2000 BDRVQcow2State *s = bs->opaque; 2001 int64_t first_free_cluster = 0, reftable_offset = -1, cluster = 0; 2002 int64_t refblock_offset, refblock_start, refblock_index; 2003 uint32_t reftable_size = 0; 2004 uint64_t *on_disk_reftable = NULL; 2005 void *on_disk_refblock; 2006 int ret = 0; 2007 struct { 2008 uint64_t reftable_offset; 2009 uint32_t reftable_clusters; 2010 } QEMU_PACKED reftable_offset_and_clusters; 2011 2012 qcow2_cache_empty(bs, s->refcount_block_cache); 2013 2014 write_refblocks: 2015 for (; cluster < *nb_clusters; cluster++) { 2016 if (!s->get_refcount(*refcount_table, cluster)) { 2017 continue; 2018 } 2019 2020 refblock_index = cluster >> s->refcount_block_bits; 2021 refblock_start = refblock_index << s->refcount_block_bits; 2022 2023 /* Don't allocate a cluster in a refblock already written to disk */ 2024 if (first_free_cluster < refblock_start) { 2025 first_free_cluster = refblock_start; 2026 } 2027 refblock_offset = alloc_clusters_imrt(bs, 1, refcount_table, 2028 nb_clusters, &first_free_cluster); 2029 if (refblock_offset < 0) { 2030 fprintf(stderr, "ERROR allocating refblock: %s\n", 2031 strerror(-refblock_offset)); 2032 res->check_errors++; 2033 ret = refblock_offset; 2034 goto fail; 2035 } 2036 2037 if (reftable_size <= refblock_index) { 2038 uint32_t old_reftable_size = reftable_size; 2039 uint64_t *new_on_disk_reftable; 2040 2041 reftable_size = ROUND_UP((refblock_index + 1) * sizeof(uint64_t), 2042 s->cluster_size) / sizeof(uint64_t); 2043 new_on_disk_reftable = g_try_realloc(on_disk_reftable, 2044 reftable_size * 2045 sizeof(uint64_t)); 2046 if (!new_on_disk_reftable) { 2047 res->check_errors++; 2048 ret = -ENOMEM; 2049 goto fail; 2050 } 2051 on_disk_reftable = new_on_disk_reftable; 2052 2053 memset(on_disk_reftable + old_reftable_size, 0, 2054 (reftable_size - old_reftable_size) * sizeof(uint64_t)); 2055 2056 /* The offset we have for the reftable is now no longer valid; 2057 * this will leak that range, but we can easily fix that by running 2058 * a leak-fixing check after this rebuild operation */ 2059 reftable_offset = -1; 2060 } 2061 on_disk_reftable[refblock_index] = refblock_offset; 2062 2063 /* If this is apparently the last refblock (for now), try to squeeze the 2064 * reftable in */ 2065 if (refblock_index == (*nb_clusters - 1) >> s->refcount_block_bits && 2066 reftable_offset < 0) 2067 { 2068 uint64_t reftable_clusters = size_to_clusters(s, reftable_size * 2069 sizeof(uint64_t)); 2070 reftable_offset = alloc_clusters_imrt(bs, reftable_clusters, 2071 refcount_table, nb_clusters, 2072 &first_free_cluster); 2073 if (reftable_offset < 0) { 2074 fprintf(stderr, "ERROR allocating reftable: %s\n", 2075 strerror(-reftable_offset)); 2076 res->check_errors++; 2077 ret = reftable_offset; 2078 goto fail; 2079 } 2080 } 2081 2082 ret = qcow2_pre_write_overlap_check(bs, 0, refblock_offset, 2083 s->cluster_size); 2084 if (ret < 0) { 2085 fprintf(stderr, "ERROR writing refblock: %s\n", strerror(-ret)); 2086 goto fail; 2087 } 2088 2089 /* The size of *refcount_table is always cluster-aligned, therefore the 2090 * write operation will not overflow */ 2091 on_disk_refblock = (void *)((char *) *refcount_table + 2092 refblock_index * s->cluster_size); 2093 2094 ret = bdrv_write(bs->file, refblock_offset / BDRV_SECTOR_SIZE, 2095 on_disk_refblock, s->cluster_sectors); 2096 if (ret < 0) { 2097 fprintf(stderr, "ERROR writing refblock: %s\n", strerror(-ret)); 2098 goto fail; 2099 } 2100 2101 /* Go to the end of this refblock */ 2102 cluster = refblock_start + s->refcount_block_size - 1; 2103 } 2104 2105 if (reftable_offset < 0) { 2106 uint64_t post_refblock_start, reftable_clusters; 2107 2108 post_refblock_start = ROUND_UP(*nb_clusters, s->refcount_block_size); 2109 reftable_clusters = size_to_clusters(s, 2110 reftable_size * sizeof(uint64_t)); 2111 /* Not pretty but simple */ 2112 if (first_free_cluster < post_refblock_start) { 2113 first_free_cluster = post_refblock_start; 2114 } 2115 reftable_offset = alloc_clusters_imrt(bs, reftable_clusters, 2116 refcount_table, nb_clusters, 2117 &first_free_cluster); 2118 if (reftable_offset < 0) { 2119 fprintf(stderr, "ERROR allocating reftable: %s\n", 2120 strerror(-reftable_offset)); 2121 res->check_errors++; 2122 ret = reftable_offset; 2123 goto fail; 2124 } 2125 2126 goto write_refblocks; 2127 } 2128 2129 assert(on_disk_reftable); 2130 2131 for (refblock_index = 0; refblock_index < reftable_size; refblock_index++) { 2132 cpu_to_be64s(&on_disk_reftable[refblock_index]); 2133 } 2134 2135 ret = qcow2_pre_write_overlap_check(bs, 0, reftable_offset, 2136 reftable_size * sizeof(uint64_t)); 2137 if (ret < 0) { 2138 fprintf(stderr, "ERROR writing reftable: %s\n", strerror(-ret)); 2139 goto fail; 2140 } 2141 2142 assert(reftable_size < INT_MAX / sizeof(uint64_t)); 2143 ret = bdrv_pwrite(bs->file, reftable_offset, on_disk_reftable, 2144 reftable_size * sizeof(uint64_t)); 2145 if (ret < 0) { 2146 fprintf(stderr, "ERROR writing reftable: %s\n", strerror(-ret)); 2147 goto fail; 2148 } 2149 2150 /* Enter new reftable into the image header */ 2151 cpu_to_be64w(&reftable_offset_and_clusters.reftable_offset, 2152 reftable_offset); 2153 cpu_to_be32w(&reftable_offset_and_clusters.reftable_clusters, 2154 size_to_clusters(s, reftable_size * sizeof(uint64_t))); 2155 ret = bdrv_pwrite_sync(bs->file, offsetof(QCowHeader, 2156 refcount_table_offset), 2157 &reftable_offset_and_clusters, 2158 sizeof(reftable_offset_and_clusters)); 2159 if (ret < 0) { 2160 fprintf(stderr, "ERROR setting reftable: %s\n", strerror(-ret)); 2161 goto fail; 2162 } 2163 2164 for (refblock_index = 0; refblock_index < reftable_size; refblock_index++) { 2165 be64_to_cpus(&on_disk_reftable[refblock_index]); 2166 } 2167 s->refcount_table = on_disk_reftable; 2168 s->refcount_table_offset = reftable_offset; 2169 s->refcount_table_size = reftable_size; 2170 2171 return 0; 2172 2173 fail: 2174 g_free(on_disk_reftable); 2175 return ret; 2176 } 2177 2178 /* 2179 * Checks an image for refcount consistency. 2180 * 2181 * Returns 0 if no errors are found, the number of errors in case the image is 2182 * detected as corrupted, and -errno when an internal error occurred. 2183 */ 2184 int qcow2_check_refcounts(BlockDriverState *bs, BdrvCheckResult *res, 2185 BdrvCheckMode fix) 2186 { 2187 BDRVQcow2State *s = bs->opaque; 2188 BdrvCheckResult pre_compare_res; 2189 int64_t size, highest_cluster, nb_clusters; 2190 void *refcount_table = NULL; 2191 bool rebuild = false; 2192 int ret; 2193 2194 size = bdrv_getlength(bs->file); 2195 if (size < 0) { 2196 res->check_errors++; 2197 return size; 2198 } 2199 2200 nb_clusters = size_to_clusters(s, size); 2201 if (nb_clusters > INT_MAX) { 2202 res->check_errors++; 2203 return -EFBIG; 2204 } 2205 2206 res->bfi.total_clusters = 2207 size_to_clusters(s, bs->total_sectors * BDRV_SECTOR_SIZE); 2208 2209 ret = calculate_refcounts(bs, res, fix, &rebuild, &refcount_table, 2210 &nb_clusters); 2211 if (ret < 0) { 2212 goto fail; 2213 } 2214 2215 /* In case we don't need to rebuild the refcount structure (but want to fix 2216 * something), this function is immediately called again, in which case the 2217 * result should be ignored */ 2218 pre_compare_res = *res; 2219 compare_refcounts(bs, res, 0, &rebuild, &highest_cluster, refcount_table, 2220 nb_clusters); 2221 2222 if (rebuild && (fix & BDRV_FIX_ERRORS)) { 2223 BdrvCheckResult old_res = *res; 2224 int fresh_leaks = 0; 2225 2226 fprintf(stderr, "Rebuilding refcount structure\n"); 2227 ret = rebuild_refcount_structure(bs, res, &refcount_table, 2228 &nb_clusters); 2229 if (ret < 0) { 2230 goto fail; 2231 } 2232 2233 res->corruptions = 0; 2234 res->leaks = 0; 2235 2236 /* Because the old reftable has been exchanged for a new one the 2237 * references have to be recalculated */ 2238 rebuild = false; 2239 memset(refcount_table, 0, refcount_array_byte_size(s, nb_clusters)); 2240 ret = calculate_refcounts(bs, res, 0, &rebuild, &refcount_table, 2241 &nb_clusters); 2242 if (ret < 0) { 2243 goto fail; 2244 } 2245 2246 if (fix & BDRV_FIX_LEAKS) { 2247 /* The old refcount structures are now leaked, fix it; the result 2248 * can be ignored, aside from leaks which were introduced by 2249 * rebuild_refcount_structure() that could not be fixed */ 2250 BdrvCheckResult saved_res = *res; 2251 *res = (BdrvCheckResult){ 0 }; 2252 2253 compare_refcounts(bs, res, BDRV_FIX_LEAKS, &rebuild, 2254 &highest_cluster, refcount_table, nb_clusters); 2255 if (rebuild) { 2256 fprintf(stderr, "ERROR rebuilt refcount structure is still " 2257 "broken\n"); 2258 } 2259 2260 /* Any leaks accounted for here were introduced by 2261 * rebuild_refcount_structure() because that function has created a 2262 * new refcount structure from scratch */ 2263 fresh_leaks = res->leaks; 2264 *res = saved_res; 2265 } 2266 2267 if (res->corruptions < old_res.corruptions) { 2268 res->corruptions_fixed += old_res.corruptions - res->corruptions; 2269 } 2270 if (res->leaks < old_res.leaks) { 2271 res->leaks_fixed += old_res.leaks - res->leaks; 2272 } 2273 res->leaks += fresh_leaks; 2274 } else if (fix) { 2275 if (rebuild) { 2276 fprintf(stderr, "ERROR need to rebuild refcount structures\n"); 2277 res->check_errors++; 2278 ret = -EIO; 2279 goto fail; 2280 } 2281 2282 if (res->leaks || res->corruptions) { 2283 *res = pre_compare_res; 2284 compare_refcounts(bs, res, fix, &rebuild, &highest_cluster, 2285 refcount_table, nb_clusters); 2286 } 2287 } 2288 2289 /* check OFLAG_COPIED */ 2290 ret = check_oflag_copied(bs, res, fix); 2291 if (ret < 0) { 2292 goto fail; 2293 } 2294 2295 res->image_end_offset = (highest_cluster + 1) * s->cluster_size; 2296 ret = 0; 2297 2298 fail: 2299 g_free(refcount_table); 2300 2301 return ret; 2302 } 2303 2304 #define overlaps_with(ofs, sz) \ 2305 ranges_overlap(offset, size, ofs, sz) 2306 2307 /* 2308 * Checks if the given offset into the image file is actually free to use by 2309 * looking for overlaps with important metadata sections (L1/L2 tables etc.), 2310 * i.e. a sanity check without relying on the refcount tables. 2311 * 2312 * The ign parameter specifies what checks not to perform (being a bitmask of 2313 * QCow2MetadataOverlap values), i.e., what sections to ignore. 2314 * 2315 * Returns: 2316 * - 0 if writing to this offset will not affect the mentioned metadata 2317 * - a positive QCow2MetadataOverlap value indicating one overlapping section 2318 * - a negative value (-errno) indicating an error while performing a check, 2319 * e.g. when bdrv_read failed on QCOW2_OL_INACTIVE_L2 2320 */ 2321 int qcow2_check_metadata_overlap(BlockDriverState *bs, int ign, int64_t offset, 2322 int64_t size) 2323 { 2324 BDRVQcow2State *s = bs->opaque; 2325 int chk = s->overlap_check & ~ign; 2326 int i, j; 2327 2328 if (!size) { 2329 return 0; 2330 } 2331 2332 if (chk & QCOW2_OL_MAIN_HEADER) { 2333 if (offset < s->cluster_size) { 2334 return QCOW2_OL_MAIN_HEADER; 2335 } 2336 } 2337 2338 /* align range to test to cluster boundaries */ 2339 size = align_offset(offset_into_cluster(s, offset) + size, s->cluster_size); 2340 offset = start_of_cluster(s, offset); 2341 2342 if ((chk & QCOW2_OL_ACTIVE_L1) && s->l1_size) { 2343 if (overlaps_with(s->l1_table_offset, s->l1_size * sizeof(uint64_t))) { 2344 return QCOW2_OL_ACTIVE_L1; 2345 } 2346 } 2347 2348 if ((chk & QCOW2_OL_REFCOUNT_TABLE) && s->refcount_table_size) { 2349 if (overlaps_with(s->refcount_table_offset, 2350 s->refcount_table_size * sizeof(uint64_t))) { 2351 return QCOW2_OL_REFCOUNT_TABLE; 2352 } 2353 } 2354 2355 if ((chk & QCOW2_OL_SNAPSHOT_TABLE) && s->snapshots_size) { 2356 if (overlaps_with(s->snapshots_offset, s->snapshots_size)) { 2357 return QCOW2_OL_SNAPSHOT_TABLE; 2358 } 2359 } 2360 2361 if ((chk & QCOW2_OL_INACTIVE_L1) && s->snapshots) { 2362 for (i = 0; i < s->nb_snapshots; i++) { 2363 if (s->snapshots[i].l1_size && 2364 overlaps_with(s->snapshots[i].l1_table_offset, 2365 s->snapshots[i].l1_size * sizeof(uint64_t))) { 2366 return QCOW2_OL_INACTIVE_L1; 2367 } 2368 } 2369 } 2370 2371 if ((chk & QCOW2_OL_ACTIVE_L2) && s->l1_table) { 2372 for (i = 0; i < s->l1_size; i++) { 2373 if ((s->l1_table[i] & L1E_OFFSET_MASK) && 2374 overlaps_with(s->l1_table[i] & L1E_OFFSET_MASK, 2375 s->cluster_size)) { 2376 return QCOW2_OL_ACTIVE_L2; 2377 } 2378 } 2379 } 2380 2381 if ((chk & QCOW2_OL_REFCOUNT_BLOCK) && s->refcount_table) { 2382 for (i = 0; i < s->refcount_table_size; i++) { 2383 if ((s->refcount_table[i] & REFT_OFFSET_MASK) && 2384 overlaps_with(s->refcount_table[i] & REFT_OFFSET_MASK, 2385 s->cluster_size)) { 2386 return QCOW2_OL_REFCOUNT_BLOCK; 2387 } 2388 } 2389 } 2390 2391 if ((chk & QCOW2_OL_INACTIVE_L2) && s->snapshots) { 2392 for (i = 0; i < s->nb_snapshots; i++) { 2393 uint64_t l1_ofs = s->snapshots[i].l1_table_offset; 2394 uint32_t l1_sz = s->snapshots[i].l1_size; 2395 uint64_t l1_sz2 = l1_sz * sizeof(uint64_t); 2396 uint64_t *l1 = g_try_malloc(l1_sz2); 2397 int ret; 2398 2399 if (l1_sz2 && l1 == NULL) { 2400 return -ENOMEM; 2401 } 2402 2403 ret = bdrv_pread(bs->file, l1_ofs, l1, l1_sz2); 2404 if (ret < 0) { 2405 g_free(l1); 2406 return ret; 2407 } 2408 2409 for (j = 0; j < l1_sz; j++) { 2410 uint64_t l2_ofs = be64_to_cpu(l1[j]) & L1E_OFFSET_MASK; 2411 if (l2_ofs && overlaps_with(l2_ofs, s->cluster_size)) { 2412 g_free(l1); 2413 return QCOW2_OL_INACTIVE_L2; 2414 } 2415 } 2416 2417 g_free(l1); 2418 } 2419 } 2420 2421 return 0; 2422 } 2423 2424 static const char *metadata_ol_names[] = { 2425 [QCOW2_OL_MAIN_HEADER_BITNR] = "qcow2_header", 2426 [QCOW2_OL_ACTIVE_L1_BITNR] = "active L1 table", 2427 [QCOW2_OL_ACTIVE_L2_BITNR] = "active L2 table", 2428 [QCOW2_OL_REFCOUNT_TABLE_BITNR] = "refcount table", 2429 [QCOW2_OL_REFCOUNT_BLOCK_BITNR] = "refcount block", 2430 [QCOW2_OL_SNAPSHOT_TABLE_BITNR] = "snapshot table", 2431 [QCOW2_OL_INACTIVE_L1_BITNR] = "inactive L1 table", 2432 [QCOW2_OL_INACTIVE_L2_BITNR] = "inactive L2 table", 2433 }; 2434 2435 /* 2436 * First performs a check for metadata overlaps (through 2437 * qcow2_check_metadata_overlap); if that fails with a negative value (error 2438 * while performing a check), that value is returned. If an impending overlap 2439 * is detected, the BDS will be made unusable, the qcow2 file marked corrupt 2440 * and -EIO returned. 2441 * 2442 * Returns 0 if there were neither overlaps nor errors while checking for 2443 * overlaps; or a negative value (-errno) on error. 2444 */ 2445 int qcow2_pre_write_overlap_check(BlockDriverState *bs, int ign, int64_t offset, 2446 int64_t size) 2447 { 2448 int ret = qcow2_check_metadata_overlap(bs, ign, offset, size); 2449 2450 if (ret < 0) { 2451 return ret; 2452 } else if (ret > 0) { 2453 int metadata_ol_bitnr = ctz32(ret); 2454 assert(metadata_ol_bitnr < QCOW2_OL_MAX_BITNR); 2455 2456 qcow2_signal_corruption(bs, true, offset, size, "Preventing invalid " 2457 "write on metadata (overlaps with %s)", 2458 metadata_ol_names[metadata_ol_bitnr]); 2459 return -EIO; 2460 } 2461 2462 return 0; 2463 } 2464